Proceedings of The International Conference on ...

29 downloads 378904 Views 12MB Size Report
third party present (conference calls), selecting only calls originating from the mobile ...... S2 = 15.706 kN, S4 = 401.30 kN, S6 = 18.642 kN, S7 = 412.89 kN, S9 ...
Proceedings of The International Conference on Modelling and Simulation (MS’08 JORDAN)

PETRA ( Jordan) 18-20 November, 2008

TABLE OF CONTENTS

Table of Contents Table of Contents Conference committees Sponsors Welcome address

III VII VIII IX

Electrical Engineering - Power Evaluation of Transformer Dielectric Response at Different Equilibrium Conditions and Temperatures A.Setayeshmehr, J. Abdalla, H. Bosri, and E. Gockenbach

1

Modelling & Design of a Linear Variable Differential Transformer Sinan Taifour, Lutfi Al-Sharif, Mohammad Kilani

7

A Multi-objective Genetic Algorithm for a Rapid and Efficient Load Flow Solution for Electrical Power Systems Hassan Kubba, Rosli Omar and Jafar Soltani

14

Electrical Engineering - Control HMM Based Voice Command System, A Way to Control a Manipulator Arm Mohamed Fezari, Hadj Ahmed Abbassi, Brahim Boulebtateche, and Mohamed Boughazi

20

An Adaptive Steering System for a Ship Using an Neural Network and Pole Placement Based Self-Tuning Composite Controller Ali Zayed, Mahmood Elfandi, M.A. Hamza, and Amir Hussein

26

Speed Control for Different Cylindrical Materials of Centrifugal Casting Machine Abdulbaset Al-Emam, Ali Zayed and Mahmoud Elfandi

30

Constrained Predictive Control for Motion Platform M.L. Saidi and A. Debbeh

33

HOS-Based ARMA Model Order Estimation Using the Determinant of Sub-Matrices of the Covariance Matrix Adnan M. Al-Smadi and Husam A. Hamad

37

Electrical Engineering - Electronics Design of Sub-threshold Comparator Using the ACM Model for Biomedical Wireless Sensor Application Leila Koushaeian and Aladin Zayegh

41

A Novel Current to Voltage Converter for a Portable Blood Gas Analyser Jaideep Chandran, Robin Kalia, Alex Stojcevski, Thinh Nguyen and Aladin Zayegh

45

Modeling and Simulation of Micro-Electro-Mechanical System Pressure Sensor for Biomechanical Application MS’08 Jordan

III

TABLE OF CONTENT

Y. Wahab, A. Zayegh, R.K. Begg, and R. Velganovski

49

Electrical Engineering - Communications Electromagnetic Computation Methods for very small Particles: Part II S.O.Bashir

55

Nonparametric Estimation of the Mixture of Signal (Interference) Probability Densities Emir Aminovich Ibatoulline

57

Modeling and Simulation of Pulse Ratio Modulation System Hind B. Bouraoui, Amer R. Zerek and Othman A. Soltan

60

Computer-Based Investigation for Side-lobes Accompanying the Pulse Compression Radar Mostafa B. Abuitbel, Amer R. Zerek and Ali M. Saleh

64

Computer Engineering & Computer Science Monte Carlo Simulation of a 2D Ising Model Bouamra Mohamed, Ait Abdelmalek Djamila

67

An Efficient Text Compression Technique Using Lempel-Ziv Algorithm Ahmed Musa, Ayman Al-Dmour, Osama Al-Khaleel, and Mansour Irshid

71

Self-Organizing Maps for User Communities Sennaike O.A., Ojo A. K., and Sofoluwe A.B.

76

A Novel Technique in Multimedia Information Hiding using Quantization Level Based Visual Cryptography Rand A. Al-Dallah, Moussa H. Abdallah, and Rola I. Al-Khalid

80

Design of Service-Oriented Architecture Systems Using Identity Management Concepts Gaber Elsharawy

87

A 90 nm Low-Power High-Speed Encoder Design for UWB Flash Analog to Digital Converter Anand Mohan, Aladin Zayegh and Alex Stojcevski

93

Fast K-MEANS Clustering Algorithms M. B. Al-Zoubi, N. B. Venkateswarlu and S. A. Roberts

98

Civil Engineering Optimal Structural and Mechanism Design using Topology Optimization G.E. Stavroulakis, N. Kaminakis, Y. Marinakis, M. Marinaki, N. Skaros

101

Fuzzy-logic-based Definition of Chen Model of Plasticity Nataliya Pokorná* and Petr Štemberk

106

Fuzzy-logic-based Expert System for Proportioning of Concrete Composition MS’08 Jordan

IV

TABLE OF CONTENTS

Alena Kohoutková and Petr Štemberk

111

Mathematical modeling of cascade areas in constructional materials A A. Togambayeva, A. Kupchishin, T. Shmygaleva

115

Simulation of the Pull-Out Phenomenon of a Bolt from a Steel Plate Connection D.N. Kaziolas, E. Efthymiou, M. Zygomalas and C. C. Baniotopoulos

120

Modelling and Numerical Investigation of Flow and Contaminant Removal in Horizontal Wetlands Konstantinos A. Liolios, Konstantinos N. Moutsopoulos, Vassilios A. Tsihrintzis and Christos S. Akratos

125

Modeling of Evolution of Concrete Stress-strain Diagrams with Respect to Various Parameters Petr Štemberk, Pavel Tomek, and Alena Kohoutková

131

A Numerical Approach for Structures Environemtally Damaged and Strengthened by Cable-Elements Asterios A. Liolios, Angelos A. Liolios, Konstantinos A. Liolios, and Khairedin M Abdalla

136

Chemical Engineering Numerical Simulation of Flows in a Clinker Cooler Demagh Yassine and Noui Samira

140

Development of Model for Crude oil Degradation in a Simplified Stream System C. P. Ukpaka, V. G. Nnadi, S.A. Amadi, and N. Umesi

145

Thermodynamic Study of the Ternary Systems (Al-Ga-M) (M=As, P, Sb) Y. Djaballah and A. Belgacem-Bouzida

155

Mechanical Engineering Two Dimensional fatigue crack growth model of roller burnished 2024 T351 aluminium alloy H. Hamadache, A. Amirat, and K. Chaoui

161

Development of Distinct Element Method (DEM) for Modeling Non-spherical Particles Feras Y. Fraige, Paul A. Langston and Laila A. Al-Khatib

169

Atmospheric Modeling of the Stellar Binary System 9Cyg M. Al-Wardat and H. Wedyan

175

E-Learning A suggested Generic Intelligent Tutoring Framework Fares Fraij and Victor Winter

178

Application of Matlab with Simulink for Teaching the Principles of Modulation Techniques Amer R. Zerek, Meftah M. Almrabet, and Hind B. Bouraoui

182

MS’08 Jordan

V

CONFERENCE COMMITTEES

Organizing Committee Ahmad Nuseirat, Chairman [email protected] Fares Fraij, Co-Chairman [email protected] Ahmed Musa [email protected] Ayman Al-Dmour [email protected] Laila Al-Khatib [email protected] Marwan Batiha [email protected] Abdalla Maharmeh [email protected] Mowafak Fathi [email protected] Belal Al-Salameen Program Committee Fares Fraij, Chair Marwan Batiha Mowafak Fathi Feras Freige Reception Committee Akef AlBdour, Chair Emad AlFanatseh Mohamed AlNasarat Waleed AlAlayah Abdallah Ihmedat Ashour AboElzayt Ismail AlFalahat Public Relations Committee Mowafak Fathi, Chair Ayman Al-Dmour Abdalla Maharmeh Social Program Committee Laila Al-Khatib, Chair Belal Al-Salameen Ayman Hamadeen Publications Committee Marwan Batiha, Chair Ahmed Musa Abdalla Maharmeh Scientific Committee Jaime Gil Aluja, President Ahmad Nuseirat, Vice-President MS’08 Jordan

Abdelhamid Fanatsa, Jordan Abdulmalik Al-Salman, Saudi Arabia Ahmad Abu Elhaija, Jordan Ahmed Oulad Said, Morocco Aladin ZAYEGH, Australia Anna-María GIL LAFUENTE, Spain Basel M. Al-Eideh, Kuwait Bernard BALLAZ, France Carlo Francesco Morabito, Italy Ch. Baniogtoboulos, Greece Christian Berger-Vachon, France D. P. Kothari, India Emmanuel PERRIN, France Fandi Alwaked, Jordan Federico G. SANTOYO, Mexico G. Stavroulakis, Greece Houcine CHAFOUK, France Ibraheem Ghareeb, Jordan Iskander YAACOB, Malaysia Jaime Gil Lafuente, Spain Jaime T.-ARANDES, Venezuela Jihad Mohamad ALJA'AM, Qatar K. Mustafa , India Kedi HUANG, China Khair Jadaan, Jordan Khairedin Abdalla, Jordan Li Ge, China Ludwig Simone, Canada Maha Shadaydeh, Syria Mansour Abbadi, Jordan Mehdi Shadaram, USA Mousa Mohsen, Jordan M. Jamshidi, USA Nabeel Alrousan, Jordan Nashaat EL-KHAMEESY, Egypt Paul Langston, UK Petr Stemberk, Czeck Prabhat Kumar MAHANTI, India P. KOTMARI, India Qiang Ye , Canada Reyad El-Khazali, UAE R. Abu-Zitar, Jordan Said ALLAKI, Spain Salamah Salamah, USA Sandip Ray, USA Turki Obaidat, Jordan Vasyl TCHABAN, Ukraine Vladimir S. NERONOV, Kazakstan Victor Winter, USA Waleed Joher, Jordan

VII

SPONSORS

Sponsors Al-Hussein Bin Talal University (AHU) Association for Modelling and Simulation Technique in Enterprise (AMSE) Microsoft Indo-Jordan Chemical Company Jordanian Islamic Bank, Jordanian Engineers Association Abdulhamid Shoman Foundation ZeeDimension Jordanian Construction and Contractors Association Aqaba Railway Corporation Nestle Al-Own Advanced for Contracting Co.

MS’08 Jordan

VIII

WELCOME ADRESS

Welcome Address It is my pleasure and special honor to welcome all participants and guests of the International Conference on Modelling and Simulation (MS’08 Jordan), which has been organized by Al-Hussein Bin Talal University(AHU) in Cooperation with International Association for Modelling and Simulation Technique in Enterprise (AMSE). The conference aims to provide a forum for researchers, experts and engineers to discuss research ideas, exchange knowledge and practices, and review advancements in related topics. The review process was conducted in two phases. In the first phase, abstracts were initially reviewed. Then, in the second phase, full-papers were reviewed. The two phases were performed solely by the members of the members of the scientific committee. The final number of accepted papers was 70 from 20 different countries. In the proceeding included only the presented papers during the conference sessions. The papers were divided into many sessions where each session was chaired by a colleague to encourage discussions among participants. Two keynote speakers state-of-the-art advancements in engineering and applied sciences. The support to the conference activities provided by Al-Hussein University, AMSE, Microsoft, IndoJordan Chemical Company, Jordanian Islamic Bank, Jordanian Engineers Association, Abdulhamid Shoman Foundation, ZeeDimension, Jordanian Construction and Contractors Association, Aqaba Railway Corporation, Nestle and Al-Own Advanced for Contracting Co.. On behalf of the organizing committee I wish to express my deep appreciation to all members of the Scientific Committee for the evaluation of authors contributions. I would like to express my sincere thanks to all members of the Organizing Committee and all SubCommittees and to persons involved in the organization of the conference activities for their effort to ensure successful conference. Finally, I welcome all participants and guests and I wish you all a successful conference and a pleasant stay in Jordan.

Prof. Ahmad M. Nuseirat Conference Chairman

MS’08 Jordan

VII

EVALUATION OF TRANSFORMER DIELECTRIC

Evaluation of Transformer Dielectric Response at Different Equilibrium Conditions and Temperatures A. Setayeshmehr *, J. Abdallah **, H. Borsi * and E.Gockenbach * *Institute of Electric Power Systems, Division of High Voltage Engineering, Schering-Institute Leibniz Universität Hannover, Germany E-mail:[email protected]

**Tafila Technical University, Jordan E-mail:[email protected] Abstract: This paper presents an experimental investigation of transformer dielectric response on the base of Hybrid Frequency Dielectric Spectroscopy and Polarisation Current measurements method (FDS and PC). Reference measurements were performed at equilibrium conditions of water content in oil and paper of a transformer for different stable temperatures (25, 50, 60 and 70°C) to get references for the evaluation of the insulation behaviour at non-equilibrium conditions. Some measurements performed at the different temperatures simulate normal service condition of a transformer at this temperature. The obtained results show that when the transformer temperature is much higher than its ambient temperature, the transformer temperature decreases immediately after disconnecting the transformer from the network and this temperature reduction influences the transformer insulation condition during the measuring process. In addition to the oil temperature near the places of the sensors, the temperature uniformity in a transformer will be strongly influenced by the changes in the load of the transformer before the measurement and this will also strongly influence the measuring results. Keywords: Transformer Monitoring, Frequency Domain Spectroscopy, Polarization and Depolarization Current.

1 Introduction Transformers are very significant components of distribution and transmission systems both in terms of cost and system performance. They are the most expensive items at the substations; it is about 60% of the total investment. It is very important to monitor and diagnosis the efficiency of the transformers operating system, due to the increasing demand of electric energy. The cost of premature and unexpected failure of a power transformer can be several times its initial cost. There are not only the refurbishment or replacement costs but also possible costs associated with clean-up, loss of revenue, and deterioration in quality of power delivery. With increasing age, there are potential risks of extremely high monetary losses due to unexpected failures and outages. Insulation is the major component, which plays an important role in the life expectancy of the transformer, and the degradation of transformer insulation is one of the major causes of transformer breakdown. Most of the transformers in a system, around the world are exceeding their designed life. In the absence of insulation assessment, good number of transformer failed due to insulation problems, before reaching to their designed technical life. It is important to investigate the causes of the insulation degradation with respect to age. The average age of the transformers that failed due to insulation issues was 17.8 years – far from the expected life of 35 to 40 years [1]. The monitoring of power system plant is a complex problem with many different strategies and engineering solutions. The transformer monitoring system has to provide continuous on-line information, analysis of transformer operation and off-line performance analysis, and periodically MS’08 Jordan

monitors various parameters related to transformer load and condition. Therefore condition monitoring of the transformers insulation is an important issue, since many transformers in electrical industries around the world are approaching the end of their design life. Indeed, condition monitoring can be utilized to attempt the prediction of the insulation condition and the remaining life time of a transformer. The most important aging indicator is the water content in the cellulose. One of the non-destructive, off-line monitoring methods is the Interfacial Polarization Spectra (IPS), measured by the Recovery Voltage Measurement (RVM), Frequency Domain Spectroscopy (FDS) and Polarization and Depolarization Current Measurements (PDC) [2]. These methods became only recently available as user-friendly methods. Off-line field measurements on transformers of both the FDS and PDC measurements are generally performed under nonequilibrium conditions. The main contribution of this work is to experimentally investigate the dielectric responses (FDS and PDC) of a transformer at different operating temperatures and moisture content changes according to the oil temperature. For the transformer windings paper/pressboard insulation system is very hygroscopic and can still contain moisture, even if the oil has been dried and has different temperature. There are indirect methods available to estimate moisture content of paper. Some of the transformers measurements can be only done off-line, by disconnecting the transformer from the electrical system for a period of time. This disconnection cause changes in the temperatures of the oil and the insulations inside the transformer. The temperature and the water content 1

A. SETAYESHMEHR, J. ABDALLAH, H. BORSI AND E.GOCKENBACH

are very important parameters, which determine the insulations condition and behaviour. Any change in the temperature, which is often decreasing after disconnection leads to changes in the oil water solubility and water content in the insulation paper of the transformer. As the temperature changes consequently the measuring results will differs from the actual when the transformer is in operation condition. Therefore moisture analysis is helpful in detecting some types of failures in a transformer. The measurement of partial discharges is probably the most effective method to detect pending failure in an electrical apparatus [3]. Measurements have been performed in the laboratory of the Schering- Institute, Leibniz University of Hanover, Germany, on a distribution transformer. The temperature effect on the characteristic of the transformer insulation system (oil - paper) was investigated on a 400 V / 10 kV, 100 kVA oil filled distribution transformer, which has been in service for 20 years. It has been removed from service two years ago and stored in the laboratory at room temperature that guarantees water equilibrium in the insulation. The measured water content 18 ppm at room temperature at the beginning of the measurements.

After that the voltage is removed and the object is shortcircuited at t = tC, enabling the measurement of the depolarization current (or discharging, or de-sorption) idpol(t) in the opposite direction, without contribution of the conductivity. The polarization current measurement can usually be stopped if the current becomes either stable or very low. According to the superposition principle the sudden reduction of the voltage UC to zero is regarded as a negative voltage step at time t = tc. Neglecting the second term in (1) we get for t = (t0 + TC) [5, 6]:

idepol (t ) = −C oU c [ f (t ) − f (t + Tc )]

(2)

where Tc is the polarization time (charging time) of the test object. The schematic diagram of the PDC measuring technique is shown in Fig. 1 and the typical nature of these currents due to a step charging voltage UC in Fig. 2.

idepol

ipol Test object

Uc

2 Theoretical Background The off-line insulation condition transformer measurements of the dielectric response and analysis of water content in oil paper insulations of power and instrument transformers can be achieved with different methods [2]. Recently, those methods are grouped as: Time domain, Frequency domain, analytical Fourier transform.

Electrometer

Fig. 1: Principle of test arrangement for the “PDC” measuring technique

2.1 Time Domain Measurements The time domain model is based on measurements of the polarization and depolarization current “PDC”. The (PDC) following a dc voltage step is one way in the time domain to investigate the slow polarization processes [4]. The dielectric of the test object must be cleared or discharged from any charges which can be remaining in before the PDC measurement. The voltage source should be stable and without any ripple and noise in order to detect any small polarization current with sufficient accuracy. The test object is stressed for a long time (e.g., 10000 s) with a dc charging voltage of magnitude Uc. The polarization current ipol(t) through the test object is measured during this time. The ipol(t) arising from the activation of the polarization process with different time constants corresponds to different insulation materials and to the conductivity of the object, which has been previously carefully discharged. Then the polarization (or absorption, or charging) current ipol(t) through the test object can be expressed by the equation: σ  (1) i pol (t ) = C o U c  o + ε ∞ δ (t ) + f (t ) εo  where: Co -the geometrical capacitance of the test object, Uc - the step voltage (charging voltage), σo- the DC conductivity of the dielectric material, εo - the vacuum permittivity, ε∞ - the high frequency component of the permittivity, δ(t) - the delta function arising from the suddenly applied step voltage at t = t0, f(t) - the response function of the dielectric material.

MS’08 Jordan

Uc

ipol (t)

to

TC

tc

t idepol (t)

Fig. 2: Principle of polarization and depolarization current (PDC) The insulation between transformer windings is charged by a dc voltage step of 200 V. A long charging time is required (10,000 s) in order to assess the interfacial polarization and paper condition. The initial time dependence of the polarization and depolarisation currents ( 1.

in this work for its simplicity, ease of programming and gives us the required accuracy. Here, (w) is chosen to be (0.5), since the two objective functions (f1) and (f2) have the same degree of importance [7].

3.7 Mutation Random mutations alter a certain percentage of the genes in the list of chromosomes. If care is not taken, the genetic algorithm can converge too quickly into one region of the cost surface. If this area is in the region of the global minimum, that is good. However, some functions, such as the one we are modeling, have many local minima. If nothing is done to solve this tendency to converge quickly, it may end up in a local rather than a global minimum. To avoid this problem of overly fast convergence (premature convergence), the routine is forced to explore other areas of the cost surface by randomly introducing changes, or mutations, in some of the variables. Mutation points are randomly selected from the (Nind×Nvar), total number of genes in the population matrix. Increasing the number of mutations increases the algorithm’s freedom to search outside the current region of variable space. It also tends to distract the algorithm from converging on a popular solution. With the process of the crossover and mutation taking place, there is a high chance that the optimum solution could be lost as there is no guarantee that these operators will preserve the fittest string. To counteract this, elitist models are often used. In an elitist model, the best individual in the population is saved before any of these operations take place. After the new population is formed and evaluated, it is examined to see if this best structure has been preserved. If not, the saved copy is reinserted back into the population. The genetic algorithm then continues on as normal [6].

4. Multiple objective optimization (MOO) In many applications, the cost function has multiple, often times, conflicting objectives. The most important approach to MOO is: sum of weighted cost functions. The most straightforward approach to multi-objective optimization is to weight each function and add them together. N Cost = ∑wifi (8) i=1 Where: fi is the cost function (i). N wi is the weighting factor and ∑ wi = 1. i=1 Implementing this multiple objective optimization approach in a genetic algorithm only requires modifying the cost function to fit the form of equation (8) and does not require any modification to the genetic algorithm. Thus, cost=wf1+(1-w)f2

(9)

In load flow problem, (f1) is the mismatch active power (eqn.3) and (f2) is the mismatch reactive power (eqn.4) for each bus except the slack busbar. This approach is adopted MS’08 Jordan

5. Sparsity techniques Sparse matrices are a special class of matrices that contain a significant number of zero-valued elements. This property allows to: • Store only the nonzero elements of the matrix, together with their indices, to reduce the storage requirements. • Reduce the computation time for any arithmetic operation by eliminating operations on zero elements [8]. 5.1 Sparse matrix storage For full matrices, any software package stores internally every matrix element. Zero-valued elements require the same amount of storage space as any other matrix element. For sparse matrices, however, the sparsity technique stores only the nonzero elements and their indices. For large matrices with a high percentage of zero-valued elements, this scheme significantly reduces the amount of memory required for data storage. The implementations of sparsity technique, for example in MATLAB uses three arrays internally to store sparse matrices with real elements. Consider an (m-by-n) sparse matrix with (NNZ) nonzero entries (NNZ is number of nonzero elements): • The first array contains all the nonzero elements of the array in floating-point format. The length of this array is equal to (NNZ). • The second array contains the corresponding integer row indices for the nonzero elements. This array also has length equal to (NNZ). • The third array contains integer pointers to the start of each column. This array has length equal to (n). This matrix requires storage for (NNZ) floating-point numbers and (NNZ+n) integers. At 8 bytes per floatingpoint number and 4 bytes per integer, the total number of bytes required to store a sparse matrix is: Grand total of bytes=8*NNZ+4*(NNZ+n)

(10)

Sparse matrices with complex elements are also possible. In this case, it uses a fourth array with (NNZ) elements to store the imaginary parts of the nonzero elements. An element is considered nonzero if either its real or imaginary part is nonzero.

5.2 Creating Sparse Matrices Every software package never creates sparse matrices automatically. Instead, we must determine if a matrix contains a large enough percentage of zeros to benefit from sparse techniques. The density of a matrix is the number of nonzero elements divided by the total number of matrix elements. Matrices with very low density are often good candidates for use of the sparse format. In contrast, the matrix sparsity is the number of zero elements divided by the total number of 17

H. KUBBA, R. OMAR AND J. SOLTANI

matrix elements. Matrices with very high matrix sparsity are often good candidate for use of the sparse format.

power systems by using the proposed method. The reduction in computational time and storage requirements increase as the matrix density decreases or in other words matrix sparsity increases.

5.3 Viewing sparse matrix We can provide a number of functions that let us get quantitative or graphical information about Bus Active power sparse matrices. The MATLAB’s commands No. mismatch(p.u) provides high-level information about matrix 1 Slack storage, including size and storage class. For 2 0.000329 example, the following list shows information 3 0.000131 about sparse and full versions of the same 4 0.000484 matrix: 5 0.000890 6

0.000798

Reactive power mismatch(p.u) Slack PV PV PV PV 0.000484 0.000602 0.000779 0.000605 0.000321 0.000221 0.000536 0.0005215 0.000765

Voltage magnitude(p.u)

Voltage angle(deg.)

No. of generations

1.06 1.00 1.00 1.00 1.00 0.957131 0.973818 0.925362 0.956395 0.95163 0.97696 0.961725 0.982889 0.988537

0.00 3.2117 -4.35826 -6.14362 -12.4235 6.30252 -4.6541 -1.71203 1.44081 -9.00316 -5.48283 7.67754 -11.0288 -3.34466

‫ـــــ‬ 15 6 19 40 75 87 101 72 18 90 43 29 47

7 0.000365 Illustration example: 8 0.000222 Name Size Bytes Class 9 0.000185 double M_full (1100x1100) 9680000 10 0.000273 array 11 0.000950 12 0.000411 M_sparse(1100x1100) 4404 sparse 13 0.000770 array 14 0.000209 Grand total is (1210000) elements using *Total Computational Time: 4.15 sec. (9684404 bytes). Notice that the number of bytes used is much less in the sparse case, because zerovalued elements are not stored. In this case, the density of Table 1. Power flow solution of 14-bus IEEE test system by the RCGA with sparsity technique method with a the sparse matrix is (4404/9680000), or approximately standard accuracy (costfunction≤0.001 p.u.). 0.00045 (0.045%).

6. The proposed method: Multi-objective realcoded genetic algorithm with sparsity technique (Implementation and results) Three test systems were used to demonstrate the performance of the proposed method, namely: 1. 14 busbar IEEE International test system which consists of: 1 slack bus, 4 generator busses (PV) and 9 load busses (PQ) [3]. 2. 40 busbar practical test system [3], which consists of:1 slack bus, 9 generator busses (PV) and 30 load bus (PQ). 3. Iraqi National Grid (ING), which consists of 362 busbar.1 slack bus, 29 generator bus (PV) and 332 load bus (PQ). The load flow solution using real-coded genetic algorithm programs with and without sparsity technique have been developed by the use of MATLAB version7, and tested with a Pentium 4, 3GHz (Cache 2M) PC with 2GB RAM. Table(1) illustrates the power flow solution for 14-bus IEEE test system using RCGA with sparsity technique with two objective functions which are the mismatch active and reactive powers at each bus according to its constraints except the slack bus. The sum of weighted cost functions is used. Because of the stochastic nature of the genetic algorithm process, each independent run will probably produce a different number of generations and consequently the computation time and the best amongst these should be chosen. The best of the 10 implementation runs is shown in the table. The total computation time was 4.15 sec while the total computational time without using sparsity technique was 7.156 sec with the same accuracy (cost function≤0.001). Table (2) illustrates the reduction in computational time and storage requirements for different MS’08 Jordan

*Total Computational Time using RCGA without Sparsity Technique = 7.156 sec.

Type of power system

Matrix density of [Y]

14-bus IEEE

6.83%

40-bus

5.37%

%Reduction in Computational Time (RCGA with sparsity technique) 41% 59%

%Reduction in Storage Requirement (RCGA with sparsity tec.) 60% 75%

3620.628% 90% 95% bus(ING) Table 2. Comparison of reduction in computational time and storage requirement for different power systems using the RCGA with sparsity technique method.

7. Conclusion The proposed method (RCGA with sparsity) presented in this paper is much faster and has less storage requirements than the simple genetic algorithm. Thus it can be concluded that the proposed method can be implemented on-line for small and medium-scale power systems and it can be used for planning study for large-scale systems. The proposed method has reliable convergence and high accuracy of solution. Whereas the traditional numerical

18

A MULTI-OBJECTIVE GENETIC ALGORITHM FOR A RAPID AND EFFICIENT LOAD FLOW SOLUTION FOR ELECTRICAL POWER SYSTEMS

techniques (Gauss-Seidel, Newton-Raphson, Fast decoupled,…etc.) use the characteristics of the problem to determine the next sampling point (e.g. gradient, linearity and continuity), genetic algorithm makes no such assumptions. Instead, the next sampled point is determined based on stochastic sampling or decision rules rather than on a set of deterministic decision rules. Also, whereas the traditional numerical techniques mentioned above use a single point at a time to search the problem space, genetic algorithm uses a population of candidate solutions for solving the problem. Thus, reducing the possibility of ending at a local minima.

Acknowledgment The authors thank the Department of Electrical Engineering, Faculty of Engineering, University of Malaya, Kuala Lumpur, Malaysia for their support for making this research. The corresponding author Ass. Professor Hassan Kubba thanks the Department of Electrical Engineering, Engineering College, Baghdad University, Baghdad, Iraq.

References [1] N. R. Dhar,″Computer Aided Power System Operation and Analysis″, Jadavpur University, Calcutta, 1982. [2] L. Ippolito, A. Cortiglia, and M. Petrocelli, ″Optimal Allocation of Facts Devices by Using Multi-Objective Optimal Power Flow and Genetic Algorithms″, International journal of emerging electric power systems, Vol. 7, No. 2, 2006. [3] H. A. Kubba, "A improved and more reliable decoupled load flow method", Engineering, Scientific Journal of Engineering college/ Baghdad University, No.3 , Vol.7 , sept. 2001 pp. 25-37. [4] M. Mitchell,″An Introduction to Genetic Algorithms″, MIT Press, Cambridge, MassachusettsLondon,England,5th printing, 1999. [5] H. T. Yang, and L.C. Huang, ″A Parallel Genetic Algorithm Approach to Solving the Unit Commitment Problem″ IEEE Transactions on power systems, Vol. 12, No. 2, May 1997. [6] S.B.M. Ibrahim,″The PID Controller Design Using Genetic Algorithm″,A dissertation submitted to University ofSouthern Queensland, Faculty of engineering and surveying, Electrical and Electronics Engineering, October 2005. [7] A. Talib, ″An Optimization Approach of Robot Motion Planning Using Genetic Algorithm″, M.Sc Thesis, Mechatronics Department, AL-Khwarizmi Engineering, University of Baghdad, 2007. [8] Brameller, "Sparisty", Pitman Publishing, 1976.

MS’08 Jordan

19

MOHAMED FEZARI, H. A. ABBASSI, B. BOULEBTATECHE, M. BOUGHAZI

HMM BASED VOICE COMMAND SYSTEM, A WAY TO CONTROL A MANIPULATOR ARM Mohamed FEZARI, H. A. ABBASSI, B. BOULEBTATECHE, M. BOUGHAZI Faculty of Engineering, Department of Electronics, University of Annaba Laboratory of Automatic and Signals, Annaba, BP.12, Annaba, 23000, ALGERIA [email protected], [email protected],[email protected],[email protected]

Abstract: A voice command system for a manipulator arm is designed as a part of a research project. The methodology adopted is based on a spotted words recognition system based on a robust HMM (Hidden Markov Model) technique with cepstral coefficients as parameters used in automatic speech recognition system. To implement the approach on a real-time application, a Personal Computer parallel port interface was designed to control the movement of a set of stepper motors. The user can control the movements of four degree of freedom (DOF) for a robot arm using a vocal phrase containing spotted words. Keywords: Human-machine interaction, hidden Markov Model, voice command, stepper motors and Robotics.

1. Introduction Many of the most promising applications for manipulator robots are those in which robots can reduce or eliminate the need for humans to perform tasks in dangerous environments. Examples include space exploration, mining, and toxic waste cleanup [1][2]. Human-robot voice interface has a key role in many application fields and various studies made in the last few years have given good results in both research and commercial applications [3-4] and [11]. This paper proposes a new approach to the problem of the recognition of spotted words within a phrase, using a statistical approach based on HMM [5] and [7]. The increase in complexity as compared to the use of only traditional approach is quite acceptable but not negligible, however the system achieves considerable improvement in the recognition phase, thus facilitating the final decision and reducing the number of errors in decision taken by the voice command guided system. Speech recognition systems constitute the focus of a large research effort in Artificial Intelligence (AI), which has led to a large number of new theories and new techniques. However, it is only recently that the field of robot and AGV navigation has started to import some of the existing techniques developed in AI for dealing with uncertain information. HMM is a robust technique developed all applied in pattern recognition. Very interesting results were obtained in isolated words speaker independent recognition system, especially in limited vocabulary. However, the rate of recognition is lower in continuous speaking system. The approach proposed here is to design a system that gets specific words within a large or small phrase, process the selected words (Spots) and then execute an order [6][7]. As application, a set of four stepper motors were installed via a PC parallel port interface. The application uses a set of twelve commands in Arabic words, divided in two subsets one subset contains the names of main parts of a robot arm ( base, arm, fore-arm, wrist (hand), and gripper), the second subset contains the actions that can be MS’08 Jordan

taken by one of the parts in subset one ( left, right, up, down, stop, open and close). The application should be implemented on a DSP or a microcontroller in the future in order to be autonomous [8] and has to be robust to any background noise confronted by the system. The aim of this paper is therefore the recognition of spotted words from a limited vocabulary in the presence of background noise. The application is speaker-independent. Therefore, it does not need a training phase for each user. It should, however, be pointed out that this condition does not depend on the overall approach but only on the method with which the reference patterns were chosen. So by leaving the approach unaltered and choosing the reference patterns appropriately (based on speakers), this application can be made speaker-dependent [9]. As application, a vocal command for a set of stepper motors is chosen. There have been many research projects dealing with robot control and tele-operation of arm manipulators, among these projects, there are some projects that build intelligent systems [10-12]. Since we have seen human-like robots in science fiction movies such as in “I ROBOT” movie, making intelligent robots or intelligent systems became an obsession within the research group. Voice command needs the recognition of spotted words from a limited vocabulary used in Automatic Guided Vehicle (AGV) system [13] and in manipulator arm control [14].

2. Designed Application Description The application is based on the voice command for a set of four stepper motors. It therefore involves the recognition of spotted words from a limited vocabulary used to recognise the elements and control the movement of a robot arm. The vocabulary is limited to twelve words divided into two subsets: object name subset necessary to select the part of the robot arm to move and command subset necessary to control the movement of the arm example: turn left, turn right and stop for the base ( shoulder), Open close and stop for the gripper. The number of words in the vocabulary was kept to a 20

HMM BASED VOICE COMMAND SYSTEM, A WAY TO CONTROL A MANIPULATOR ARM

minimum both to make the application simpler and easier for the user. The user selects the robot arm part by its name then gives the movement order on a microphone, connected to sound card of the PC. The user can give the order in a natural language phrase as example: “Yade, gripper open execute”. A speech recognition agent based on HMM technique detects the spotted words within the phrase, recognises the main word “Yade” witch is used as a keyword in the phrase, it recognises the spotted words, then the system will generate a byte where the four most significant bits represent a code for the part of the robot arm and the four less significant bits represent the action to be taken by the robot arm. Finally, the byte is sent to the parallel port of the PC and then it is transmitted to the robots via a radio frequency emitter. The application is first simulated on PC. It includes three phases: the training phase, where a reference pattern file is created, the recognition phase where the decision to generate an accurate action is taken and the appropriate code generation, where the system generates a code of 8 bits on parallel port. In this code, four higher bits are used to codify the object names and four lower bits are sued to codify the actions. The action is shown in real-time on parallel port interface card that includes a set four stepper motors to show what command is taken and a radio Frequency emitter.

3. The Speech Recognition Agent The speech recognition agent is based on HMM. In this paragraph, a brief definition of HMM is presented and speech processing main blocks are explained. However, a pre-requisite phase is necessary to process a data base composed of twelve vocabulary words repeated twenty times by twenty persons. So before starting in the creation of parameters, 20*20*12 “wav” files are recoded in a repertory. The training phase will, each utterance ( saved wav file) is converted to a Cepstral domain (MFCC features, energy, and first and second order deltas) which constitutes an observation sequence for the estimation of the HMM parameters associated to the respective word. The estimation is performed by optimisation of the likelihood of the training vectors corresponding to each word in the vocabulary. This optimisation is carried by the BaumWelch algorithm [7]

compute the most probable state sequence, the Viterbi algorithm is the most suitable. A HMM model is basically a stochastic finite state automaton, which generates an observation string, that is, the sequence of observation vectors, O=O1,..Ot ,… ,OT. Thus, a HMM model consists of a number of N states S={Si} and of the observation string produced as a result of emitting a vector Ot for each successive transitions from one state Si to a state Sj. Ot is d dimension and in the discrete case takes its values in a library of M symbols. The state transition probability distribution between state Si to Sj is A={aij}, and the observation probability distribution of emitting any vector Ot at state Sj is given by B={bj(Ot)}. The probability distribution of initial state is Π={ πi}.

aij = P(qt +1 = S j qt = Si )

(1)

aij = P(qt +1 = S j qt = Si )

(2)

π i = P(q0 = Si )

(3)

Given an observation O and a HMM model λ=(A,B,∏), the probability of the observed sequence by the forwardbackward procedure P(O/λ) can be computed [10]. Consequently, the forward variable is defined as the probability of the partial observation sequence O1O2 ,....Ot (until time t) and the state S at time t, with the model λ as α(i). and the backward variable is defined as the probability of the partial observation sequence from t+1 to the end, given state S at time t and the model λ as β(i). the probability of the observation sequence is computed as follow: N

N

i =1

i =1

P (O / λ ) = ∑ α t (i ) * βt (i ) = ∑ α T (i )

(4)

and the probability of being in state I at time t, given the observation sequence O and the model λ is computed as

follow:

π i = P(q0 = Si )

(5)

The flowchart of a connected HMM is an HMM with all the states linked altogether (every state can be reached from any state). The Bakis HMM is left to right transition HMM with a matrix transition defined as:

3.1 HMM basics A Hidden Markov Model (HMM) is a type of stochastic model appropriate for non stationary stochastic sequences, with statistical properties that undergo distinct random transitions among a set of different stationary processes. In other; words, the HMM models a sequence of observations as a piecewise stationary process. Over the past years, Hidden Markov Models have been widely applied in several models like pattern [6], or speech recognition [6][9] . The HMMs are suitable for the classification from one or two dimensional signals and can be used when the information is incomplete or uncertain. To use a HMM, we need a training phase and a test phase. For the training stage, we usually work with the Baum-Welch algorithm to estimate the parameters ( Π i ,A,B) for the HMM [7, 9]. This method is based on the maximum likelihood criterion. To Fig. 1 : Presentation of left-right ( Bakis) HMM MS’08 Jordan

(6)

21

MOHAMED FEZARI, H. A. ABBASSI, B. BOULEBTATECHE, M. BOUGHAZI

3.2 Speech processing phase Once the phrase is acquired via a microphone and the PC sound card, the samples are stored in a wav file. Then the speech processing phase is activated. During this phase the signal (samples) goes through different steps: pre-emphasis, frame-blocking , windowing, feature extraction and MFCC analysis. a) Pre-emphasis step In general, the digitized speech waveform has a high dynamic range. In order to reduce this range preemphasis is applied. By pre-emphasis [1], we imply the application of a high pass filter, which is usually a first order FIR of the form H ( z) = 1 − a × z −1. The pre-emphasis is implemented as a fixed- coefficient filter or as an adaptive one, where the coefficient a is adjusted with time according to the autocorrelation values of the speech. The pre-emphasis block has the effect of spectral flattening which renders the signal less susceptible to finite precision effects (such as overflow and underflow) in any subsequent processing of the signal. The selected value for a in our work is 0.9375.

extraction is the suppression of information irrelevant for correct classification, such as information about speaker (e.g. fundamental frequency) and information about transmission channel ( e.g. characteristic of a microphone). The feature measurements of speech signals are typically extracted using one of the following spectral analysis techniques: MFCC Mel frequency filter bank analyzer, LPC analysis or discrete Fourier transform analysis. Currently the most popular features are Mel frequency Cepstral coefficients MFCC [7]. e)

MFCC Analysis

The Mel-Filter Cepstral Coefficients are extracted from the speech signal . The speech signal is pre-emphasized, framed and then windowed, usually with a Hamming window. Melspaced filter banks are then utilized to get the Mel spectrum. Figure 2.b shows the Mel-spaced filter banks that are used to get the Mel-spectrum. The natural Logarithm is then taken to transform into the Cepstral domain and the Discrete Cosine Transform is finally computed to get the MFCCs. as shown in the block diagram of Figure 3.

b) Frame blocking Since the vocal tract moves mechanically slowly, speech can be assumed to be a random process with slowly varying properties [1]. Hence, the speech is divided into overlapping frames of 20ms every 10ms. The speech signal is assumed to be stationary over each frame and this property will prove useful in the following steps. c)

Windowing

To minimize the discontinuity of a signal at the beginning and the end of each frame, we window each frame frames [1]. The windowing tapers the signal to zero at the beginning and end of each frame. A typical window is the Hamming window of the form:

 2π n  W (n) = 0.54 − 0.46*cos   0 ≤ n ≤ N −1  N −1  N  π k (i − 1/ 2)  Ck = ∑ log( Ei ) *cos   N  i =1

(7)

Fig. 2.b : Mel-spaced filter Bank output MFCC

Input Speech

(8) PE-FB-W

FFT

LOG

DCT

Sentence with spotted words

Fig. 3: MFCC block diagram

Fig. 2.a Windowing

Where the acronyms signify: -PE-FB-W: Pre-Emphasis, Frame Blocking and windowing. - FFT: Fast Fourier Transform - LOG: Natural Logarithm - DCT: Discrete Cosine Transform

d) Feature extraction

4. Parallel Interface Circuit

In this step, speech signal is converted into stream of feature vectors coefficients which contain only that information about given utterance that is important for its correct recognition. An important property of feature

The speech recognition agent based on HMM will detect words, and process each word. Depending on the probability of recognition of the object name and the command word a code will be transmitted to the parallel port of the PC. The

MS’08 Jordan

22

HMM BASED VOICE COMMAND SYSTEM, A WAY TO CONTROL A MANIPULATOR ARM

vocabulary to be recognized by the system and their meanings are listed as in Table 1. it is obvious that within these words, some are object names and other are command names. The code to be transmitted is composed of 8 bits, four bits most significant bits are used to code the object name and the four least significant bits are used to code the command to be executed by the selected object. Example: “yade diraa fawk tabek” wich means “Arm upper limb up execute”. A parallel port interface was designed to display the realtime commands. It is based on the following TTL IC (integrated circuits): a 74LS245 buffer, a microcontroller PIC16F84 and a wireless transceiver system from RADIOMETRIX TX433-10 ( modulation frequency 433 Mhz and transmission rate 10 Kbs). Tx Antenna Set of Green Yelow and Red LEDs

Personal Computer ( HSR)

TXM433-10

Microphone

transmission rate is 10 Kbs) [16]. Each motor in the robot arm performs the corresponding task to a received command (example: “yamin”, “kif” or Fawk”) as in Table 1. Commands and their corresponding tasks in autonomous robots may be changed in order to enhance or change the application. In the recognition phase, the application gets the word to be processed, treats the word, then takes a decision by setting the corresponding bit on the parallel port data register and hence the corresponding LED is on. The code is also transmitted in serial mode to the TXM-433-10. Reception Antenna RF

Receiver

M4

H

M1

H

M2

H

M3

Power Supply +12 volts DC.

Fig. 5.a. Robot Arm block diagram

Fig. 4. Parallel interface circuit Table 1. The meaning of the vocabulary voice commands, assigned code and motor controlled. 1) Yade (1) Name of the manipulator 2) Diraa (2) Upper limb motor (M1) 3) Saad (3) Limb motor (M2) 4) Meassam(4) Wrist (hand) motor (M3) 5) Mikbath(5) Gripper motor (M4) 6) Yamine (1) Left turn (M0) 7) Yassar (2) Right turn (M0) 8) Fawk (3) Up movement M1, M2 and M3 9) Tahta (4) Down movement M1, M2 and M3 10) Iftah (5) Open Grip, action on M4 11) Ighlak (6) Close grip, action on M4 Stop the movement, stops M0,M1, M2, 12) Kif (7) M3r or M4

Pic16f84 With Quartz= 4Mhz

H

M3: Meassam M2 saad

M4: mikbadh

M1 : Diraa

However, a simulation card was designed to control the set of four stepper motor directly by MATLAB software. It is based on a buffer 74LS245, a 74LS138 3 to 8 decoder, and four power circuits for the motors.

5. Manipulator Arm Interface As in Figures 5.a and 5.b, the structure of the mechanical hardware and the computer board of the robot arm in this paper is similar to MANUS [10-12]. However, since the robot arm needs to perform simpler tasks than those in [13] do, the computer board of the robot arm consists of a PIC16F876 , with 8K-instruction EEPROM (Electrically Programmable Read Only Memory) [15], four power circuits to drive the stepper motors and one H bridges driver using BD134 and BD133 transistors for DC motor to control the gripper, a RF receiver module from RADIOMETRIX which is the SILRX-433-10 (modulation frequency 433MHz and MS’08 Jordan

Fig. 5.b. Overview of the Robot arm and Parallel interface

6. Experiments Results The developed system has been tested within the laboratory of L.A.S.A, there were two different conditions to be tested: 23

MOHAMED FEZARI, H. A. ABBASSI, B. BOULEBTATECHE, M. BOUGHAZI

The distance of the microphone from the speaker, and the rate of recognition in periodic noise (PN) environment. The system, first, had been tested in the laboratory and outside in order to detect the environment effect on the recognition rate. After testing the recognition of each word 25 times in the following conditions: a) outside the Laboratory (LASA) with PN and b) inside the LASA with PN. The results are shown in figure 6 where the numbers in abscess axe corresponds to the order of voice command word as they appear in Table1.

ra te

effect of environment

f Ki

am i kb a Ya t h m in Ya e ss ar Fa wk Ta ht a I ft ah Ig hl ak

ad M

M

ea

ss

ra a

Sa

Di

Ya

de

100 90 80 70 60 50 40 30 20 10 0

Vocab

outdor-PN

indor

Fig. 6. The effect of SN or NSN in and out the laboratory

7. Conclusion A voice command system for robot arm is proposed and is implemented based on a HMM model for spotted words. Since the designed electronic command for the robot arm consists of a microcontroller and other low-cost components namely RF transmitters, the hardware design can easily be carried out. The results of the tests shows that a better recognition rate can be achieved inside the laboratory and especially if the phonemes of the selected word for voice command are quite different. However, a good position of the microphone and additional filtering may enhance the recognition rate. Several interesting applications of the proposed system different from previous ones are possible, such as command of a set of autonomous robots or a set of home electronic goods. The HMM based model gives better results than DTW (dynamic time warping) or Crossing zero and extremums detection based approach. It is speaker independent. However by computing parameters based on speakers’ pronunciation the system can be speaker dependant. The increase in computational complexity as compared with a traditional approach is, however, negligible. Spotted words detection is based on speech detection then processing of the detected. Once the parameters were computed, the idea can be implemented easily within a hybrid design using a DSP with a microcontroller since it does not need too much memory capacity. Finally we notice that by simply changing the set of command words, we can use this system to control other objects by voice command such as an electric

MS’08 Jordan

wheelchair movements or a set of autonomous robots [17] [18].

References [1] Beritelli F., Casale S., Cavallaro A., (1998). A Robust Voice Activity Detector for Wireless Communications Using Soft Computing, IEEE Journal on Selected Areas in Communications (JSAC), special Issue on Signal Processing for Wireless Communications, Vol. 16, n. 9, [2] Bererton, C. & Khosla, P., (2001). Towards a team of robots with reconfiguration and repair capabilities, Proceedings of the 2001 IEEE International Conference on Robotics and Automation, pp. 2923– 2928. [3] Rao R.S., Rose K. and Gersho A., (1998). Deterministically Annealed Design of Speech Recognizers and Its Performance on Isolated Letters, Proceedings IEEE ICASSP'98, pp. 461-464. [4] Gu L. and Rose K.,(2001). Perceptual Harmonic Cepstral Coefficients for Speech Recognition in Noisy Environment. Proc ICASSP 2001, Salt Lake City. [5 Djemili R. , Bedda M. , Bourouba H. (2004). Recognition Of Spoken Arabic Digits Using Neural Predictive Hidden Markov Models. International Arab Journal on Information Technology, IAJIT, Vol.1, N°2, pp. 226233. [6] Renals, S., Morgan, N., Bourlard, H., Cohen, M. & Franco, H. (1994), Connectionist probability estimators in HMM speech recognition, IEEE Transactions on Speech and Audio Processing 2(1), pp. 161-174. [7] Rabiner L. R. Rabiner. ( 1 98 9) . Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition. Readings in Speech Recognition, chapter A, pp. 267-295. [8] Hongyu L. Y. Zhao, Y. Dai and Z. Wang, (2004). A secure Voice Communication System Based on DSP, IEEE 8th International Conf. on Cont. Atau. Robotc and Vision, Kunming, China, 2004, pp. 132-137. [9] Ferrer M.A. , I. Alonso, C. Travieso, (2000). Influence of initialization and Stop Criteria on HMM based recognizers , Electronics letters of IEE, Vol. 36, pp.1165-1166 [10] Kwee Hok, (1997). Intelligent control of Manus Wheelchair. In proceedings Conference on Reabilitation Robotics, ICORR’97, Bath 1997, pp. 91-94. [11] Yussof, H.; Yamano, M.; Nasu, Y.; Mitobe, K. & Ohka M. (2005). Obstacle Avoidance in Groping Locomotion of a Humanoid Robot, International Journal of Advanced Robotic Systems, Vol. 2, No. 3, pp. 251 – 258. [12]Buhler C., Heck H., Nedza J. and Schulte D.,(1994). MANUS wheelchair-Mountable Manipulator- Further Devepolements and Tests. Manus usergroup Magazine, Vol. 2 ,no.1 , pp. 9-22. [13] Heck Helmut, (1997). User Requirements for a personal Assistive Robot, In proc. Of the 1st MobiNet symposium on Mobile Robotics Technology for Health Care Services, Athens, pp. 121-124. [14] larson Mikael, (1999). Speech Control for Robotic arm within rehabilitation. Master thesis, Division of Robotics, Dept of mechanical engineering Lund Unversity, 1999.

24

HMM BASED VOICE COMMAND SYSTEM, A WAY TO CONTROL A MANIPULATOR ARM

[15] Data sheet PIC16F876 from Microchip inc. User’s Manual, 2001, http://www.microchip.com. [16] Radiometrix components, TXm-433 and SILRX-433 Manual, HF Electronics Company. http://www.radiometrix.com. [17] Kim W. J., et al., (1998). Development of A voice remote control system. In Proceedings of the 1998

MS’08 Jordan

Korea Automatic Control Conference, Pusan, Korea, , pp. 1401-1404 [18] Fezari M., M. Bousbia-Salah and M. Bedda,(2005). Hybrid technique to enhance voice command system for a wheelchair, in proceedings of Arab Conference on Information Technology ACIT’05, Jordan, 2005.

25

ALI ZAYED, MAHMOOD ELFANDI AND AMIR HUSSAIN

An Adaptive Steering System for a Ship Using an Neural Network and Pole Placement Based Self-Tuning Composite Controller ALI ZAYED*, MAHMOOD ELFANDI**, AND AMIR HUSSAIN*** *Department of Electrical & Electronic Engineering, 7Th of April University, Saprata, Libya. **Department of Electrical& Electronic Engineering, Elfateh University, Tripoli, Libya. *** Department of Computing Science and Mathematics, University of Stirling, FK9 4LA, Scotland, UK. Abstract: The paper describes a new composite control method combining a neural network estimator with a conventional pole-placement based adaptive controller. The neural network estimation technique [1,2] is particularly effective when there is no complete plant information, or when considering a controlled plant as a 'black box'. In the proposed composite controller, the neural network estimator weights are adapted on-line to minimise the identification error, and these weights are fed into a robust self-tuning PID controller [3,4,5], which provides an adaptive mechanism to ensure that the closed loop poles are placed at the desired positions. Finally, the Algorithm is used for control of a ship steering system dynamical model in order to demonstrate the effectiveness of the proposed controller. Simulation results show that the proposed method applies to general linear or non-linear control systems. Key words: pole placement control, PID control, self-tuning control, Adaptive steering system for a ship. 1. Introduction Over the last decade or so, there has been much progress in modelling and controlling nonlinear systems with neural networks [5,6]. This is due to their proven ability to learn arbitrary non-linear mappings [5,7] which has been effectively utilized to represent the controller non-linearity. Other advantages inherent in neural networks such as their robustness, parallel architecture and fault tolerant capability act as further incentives. In particular, the neural network technique is highly effective for controlling complex non-linear systems when we have no complete model information, or even when considering a controlled plant as a black box [6]. On the other hand, significant work to-date has also been done on the conventional linear, generalised minimum variance controller (GMVC), which was originally developed by Clarke and Gawthrop [8]. It was extended to achieve poleplacement by Allidina and Hughes [9] and modified to have a PID structure by Cameron and Seborg [10] and Yusof et al. [11]. More recently, in work by Zayed [3-5], the GMVC has been modified to a robust controller which can be used either as a PID controller or as an adaptive pole-placement controller through the use of a switch. In this paper, we present a composite control method combining a neural network estimator with a conventional pole-placement based adaptive controller. The neural network estimator weights are adapted on-line to minimise the identification error, and these weights are fed into a robust self-tuning PID controller [3,4], which provides an adaptive mechanism to ensure that the closed loop poles are placed at the desired positions. Preliminary simulation results show that the proposed method applies to general linear or nonlinear control systems. The paper is organised into four sections as follows: the derivation of the control law and the neural network MS’08 Jordan

employed are described in section 2. In section 3, various simulation case studies are carried out in order to demonstrate the effectiveness of the proposed controller in the performance of the closed loop system. Finally, some concluding remarks are presented in section 4. 2. Derivation Of The Control Law The generalised minimum variance control strategy has been discussed elsewhere [8-10]. In this section only the key equations necessary for the pole-placement compensator are presented. The process used in the pole-placement controller, is described by a linear controlled auto-regressive moving average (CARMA) model of the form:

A( z −1 ) y (t ) = z − k B( z −1 )u (t ) + C ( z −1 )ξ (t ) (1) where y (t ) and u (t ) are respectively the measured output and the control input at the sampling instant t , k is the integer-sample dead time of the process, and ξ (t ) is a zero-mean Gaussian noise sequence. The polynomials A( z −1 ) and C ( z −1 ) are respectively of orders n a and nc . They are assumed to be monic and C ( z −1 ) is strictly −1 stable. In what follows, the z notation will be omitted from the various polynomials to simplify the presentation. The above polynomials are estimated using a simple neural network estimator as discussed in section 2.2. Note that the neural network estimator is applicable to arbitrary (linear or non-linear) plant models, which can be assumed to have a black-box structure. The generalised minimum variance controller minimises the following cost function:

J = E{( Py (t + k ) + Qu (t ) − Rw(t )) 2 }

(2) 26

AN ADAPTIVE STEERING SYSTEM FOR A SHIP

−1

−1

Where w(t ) is the set point and P ( z ), Q ( z ) and R ( z −1 ) are user-defined transfer functions in the backward shift operator z −1 . E{.} is the expectation operator. The control law which minimises J is [12]:

We can now define the identity:

F [CRw (t ) − ( ) y (t )] Pd u (t ) = ( EB + CQ )

(q n A + z − k q d B ) = TC

(3)

E is of order (k − 1), and the order of F is P (n a + n pd − 1) where ( P = n ) . Pd where

The polynomials E and following identity [13]:

F are obtained from the

CPn = APd E + z − k F

(4)

By multiplying both sides of the equation (3) by obtain:

u (t ) = where

[ Hw(t ) − Fy (t )] q

Pd we (5)

H = Pd CR and q = ( EBPd + CQPd )

(6)

H and q are still user defined transfer functions since they depend on the transfer functions R and Q . We further assume that q can also be expressed as [3,4]:

∆ qn [ ] v qd

q=

where

If we assume B = vBF and A = ∆A , then equation (10) becomes: −k −k (q n A + z q d B )y(t) = z vq d BHw(t) + ∆q n Cξ (t ) (11)

(7)

∆ = (1 − z − 1 ) , v is a gain and ( q n and q d ) are

pole-placement compensator polynomials. The design uses a velocity-form PID controller such as that given in [5]:

∆u (t ) = K I w(t ) − [ K P + K I + K D ] y (t ) − [− K P − 2 K D ] y (t − 1) − K D ( y − 2)

(12)

(3)

where T is the desired closed loop poles. For equation (12) to have a unique solution, the order of the regulator polynomials and the number of the desired closed loop poles have to be [5,9]:

  n qn = nb + k − 1  nT ≤ n a + nb + k − nc − 1 n f = na − 1

(13) (4)

Where n a , nb , nc , nqd , nqn , nT and k are the orders of the polynomials A , B , C, q d , qn , number of desired closed loop poles and the time delay respectively; where nb = n f + nb and n a = n a + 1. By using equation (5) (9) and (10) we obtain:

y (t ) =

z − k vq d BH q ∆C w(t) (6) + n ξ(t) T T

(14)

Clearly, the closed loop poles and zeros are in their desired locations. If the polynomial F is of first order then we have a PI controller. A PID controller occurs if F is of second order. In both cases we can switch the compensator on (for poleplacement) or off. If the plant is of order greater than three, the designs proposed in [10,11] can not achieve a PID controller since the order of the polynomial F which is computed from the Diophantine equation given by equation (4) depends on both the order of the plant and the order of the pre-filter polynomial Pd . This limitation is overcome here through the use of the compensator for pole-placement. The control law in equation (5) is illustrated in the figure (1).

(8)

q n = q d = 1 . Using equations (5) and (6) and setting H as follows: With the compensator turned off,

2.2 Neural Network Estimator In general, neural networks can be divided into feedforward and recurrent nets [1].

nf

H = ∑ f i z −i i =0

(9) z =1

K P = −v[ f 1 + 2 f 2 ] , K I = v[ f 0 + f 1 + f 2 ] and K D = vf 2 . It

is

easily

shown

that

2.1 Pole Placement Design (Compensator Switched On) If we substitute for u (t ) (given by equation (5)) into the process model given by equation (1) and make use of equations (6) and (7), we obtain

(q n ∆A + z − k vq d BF ) y (t ) = z − k vq d BHw(t ) + ∆q n Cξ (t ) (10) MS’08 Jordan

The neural network used(9)for estimation of the plant parameters in this paper is a simple Single-Layered Perceptron (SLP) network adapted using the Delta Rule (DR) [1]. The DR is the basis for the widely used Generalized Delta Rule or Back Propagation algorithms used for adapting the Multi-Layered Perceptron neural network. The output X j (t ) of the neuron j at the output layer of an SLP is given by: p

v 'j (t ) = f (∑ w 'ji (t ) x i (t ))

(15)

i =1

X j (t ) = f (v 'j (t ))

(16)

27

ALI ZAYED, MAHMOOD ELFANDI AND AMIR HUSSAIN

th

'

where x i (t ) is the i input of the SLP, w ji (t ) denotes th th the synaptic weight connecting the i input to j output, with p representing the total number of neurons in the input layer, and f denoting the non-linear activation function of th the j neuron. The correction or change in weight of any output layer in the SLP is given by:

w 'ji (t ) = w 'ji (t − 1) + ηe 'j (t ) f ' (v 'j (t )) x i (t ) '

where e j (t ) is the error and ' of f at v j (t ) .

(17)

f ' (v 'j (t )) is the gradient

Equations (15) - (17) above are collectively known as Delta Rule (DR) updating algorithm. 2.3 Overall Control Algorithm It is clear from figure (1) that the algorithm can be summarised as: step 1. Select the pre-filters polynomials Pn and Pd and also the gain v and desired closed-loop system poles polynomial T .

y (t ) and w(t ). ˆ , Bˆ , and Cˆ step 3. Estimate the process parameters A ˆ using ANNs then compute F using (4). After this compute qˆ n and qˆ d using (12). ˆ , Bˆ , and Cˆ step 3. Estimate the process parameters A ˆ using ANNs then compute F using (4). After this compute qˆ n and qˆ d using (12). step 2. Read the new values of

Step 4. Compute the control input using (5). Step 1 to 4 are to be repeated for each sampling instant. 3. Simulation Results The objective of this section is to study the performance and the robustness of the closed loop system using the technique proposed in section 2.1. A simulation example will be carried out in order to demonstrate the ability of the proposed algorithm to locate the closed poles and zeros at their desired locations under set point changes. The simulation example is performed over 600 samples with the set point changing every 100 sampling instants and the compensator was switched on from the 150th sampling instant to the 350th sampling interval when it is again switched off. The proposed controller is used to control the ship steering. The ship steering dynamics, between the actual rudder angel u (t ) and the ship's heading angel y (t ) angle can be described by the following discrete transfer function [14,15]:

MS’08 Jordan

(1 + a1 z −1 + a 2 z −2 ) y (t ) = (b0 z −1 + b1 z −2 )u (t )

a1 = −1.6061 , a 2 = 0.6061 , and b1 = 3.6161 . Where

b 0 = 4 . 2715

The desired closed loop poles polynomial T was selected as follows: T = 1 − 0.5 z −1 . The controller parameters were also selected as follows: Pd = 1 + 0.4 z −1 , Pn = 1 + 0.3z −1 and (17) v = 1 . The neural network estimator was chosen as a 2-input SLP using Delta Rule (with y (t − 1) and u (t − 1) inputs) and a convergence factor η =0.1. The output and the control input are shown in the figures (1a) and (1b). It is clear from the figures (2a) and (2b) that transient response is shaped by the choice of the polynomial T when pole-placement compensator is used. 4. Conclusions In this paper, a composite control method combining a neural network estimator with a conventional PID/poleplacement based adaptive controller has been presented. The neural network estimator weights are adapted on-line to minimise the identification error, and these weights are fed into a robust self-tuning PID controller [3,4,5], which provides an adaptive mechanism to ensure that the closed loop poles are placed at the desired positions. The proposed controller has been used to control a ship steering system simulated model. Preliminary simulation results show that the proposed method applies to general linear or non-linear control systems. Note that the parameter estimates (for the A, B and C polynomials) obtained for both the linear and non-linear plants above (for the CARMA process model used by the PID/pole placement controller), were approximated by the weights of the SLP-DR neural network (which assume no prior knowledge about the modelled plant dynamics). However, as is usually the case with neural network based black-box modelling, the estimated neural network weight values differed from the actual parameter coefficient values specified in the original plant models. Nevertheless, the composite PID/pole-placement controller was still found to work highly satisfactorily for the considered (automatic steering system for a ship) case. For future work, as in [9], correlation and chi-squared statistic based model validity tests will be used to further confirm that the identified neural network model is indeed an adequate representation of the plant model.

28

AN ADAPTIVE STEERING SYSTEM FOR A SHIP

C A

ξ (t ) w( t )

+

qˆ d qˆ n

1 ∆

ˆ v H

u( t )

process

z −k

B A

+

+

y (t )



ANN Estimator

ˆ = Fˆ (1) H qˆ n , qˆ d Compensator

Aˆ F$





parameters calculation

v Fˆ

Figure(1): Self -Tuning Pole-Placement Controller with the ANN estimator. 0.3

0.08

0.25

0.06

0.2

0.04

0.15

0.02

0.1

0

0.05

-0.02

0

-0.04

-0.05 0

-0.06

100

200

300

400

500

600

Figure (2a): the output REFERENCES [1] A. Hussain, “Real-time Adaptive Non-linear Prediction using a new class of Locally-Recurrent Neural Networks,” Journal of Control and Intelligent Systems, Vol.28, No.2, pp.65-71, 2000. [2] A. Hussain, J. J.Soraghan and T. S. Durrani, “A new Adaptive Functional-Link Neural Network Based DFE for Overcoming Co-channel Interference,” IEEE Transactions on Communications, Vol.45, No.11, pp.1358-1362, 1997. [3] A. S. Zayed, “Minimum Variance Based Adaptive PID Control Design,” M.Phil Thesis, Industrial Control Centre, University of Strathclyde, Glasgow, U.K., 1997. [4] A. S. ZAYED, L. PETROPOULAKIS AND M. R. KATEBI, “AN EXPLICIT MULTIVARIABLE SELF-TUNING POLE-PLACEMENT PID CONTROLLER,” 12TH INTERNATIONAL CONFERENCE ONSYSTEMS ENGINEERING (ICSE’97), COVENTRY, UK., PP 778-785, 9-11 SEPTEMBER 1997. [5] A. S. Zayed, “Novel linear and non-linear minimum variance techniques for adaptive control engineering,” PhD Thesis, Stirling University, Stirling, U.K., 2005. MS’08 Jordan

0

100

200

300

400

500

600

Figure (2b): the control input [6] W. Yu and A. Poznyak, “Indirect adaptive control via parallel dynamic neural networks” IEE Proc. Contol theory Appl., vol. 146, pp.25-30, 1999. [7] A. Hussain, “Novel Artificial Neural Network Architectures And Algorithms For Non-linear Dynamical System Modeling And Digital Communications Applications,” PhD Thesis, University of Strathclyde, Glasgow, U.K., 1996 [8] D. W. Clarke and P. J. Gawthrop. “Self-tuning control,” Proc. Inst. Electr. Engineering, Part D, vol. 126, pp. 633-640, 1979 [9] A. Y. Allidina and F. M. Hughes, “Generalised self-tuning controller with pole assignment,”. Proc. Inst. Electr. Engineering, Part D, vol. 127, pp. 13-18, 1980. [10] F. Cameron and D. E. Seborg, “A self-tuning controller with a PID structure” Int. J. Control, vol. 38, pp. 401-417, 1983. [11] R. Yusof, S. Omatu and M. Khalid, “Self-tuning PID: a multivariable derivation and application,” Automatica, vol. 30, pp. 1975-1981, 1994.

29

ABDULBASET M ALEMAM, ALI ZAYED AND MAHMOOD ELFANDI

Speed Control for Different Cylindrical Materials of Centrifugal Casting Machine ABDULBASET M ALEMAM*, ALI ZAYED**, MAHMOOD ELFANDI*** *Department of Mechanical Engineering, the Higher Institute of Mechanical Professionals, Tripoli, Libya **Department of Electrical & Electronic Engineering, 7Th of April University, Saprata, Libya. ***Department of Electrical& Electronic Engineering, Elfateh University, Tripoli, Libya. Corresponding author’s E-mail: [email protected] Abstract: This paper presents the description of the centrifugal casting machine which designed and constructed in the Higher Institute of Mechanical Professionals (HIMP). In this machine, the metal is poured into a permanent mold, which rotated around its axes at a certain speed, in order to produce pre-specified cylindrical shape products. However, the products’ properties may be influenced if another type of material is poured instead. To overcome this major drawback, a speed control is suggested. In this original work a PID control is used to regulate the rotating mold speed, in order to achieve satisfactory closed loop performance. Simulation results demonstrate the effectiveness of the proposed control design. Key words: Centrifugal casting machine, mold, PID controller, speed control, simulation.

1. Introduction Centrifugal casting involves rotating or revolving a mold which filled with molten metal about an axis at a known speed between 300 to 3000 rpm, and solidified at a proper rate then extracted from the mold. The centrifugal pressure forces the metal against the interior mold wall or cavity. The speed of the rotation and metal pouring rate vary with size and shape of the alloy being cast. In centrifugal casting the axis of rotation can be vertical, horizontal, or at any angle. There are advantages in both vertical and horizontal machines; one drawback in horizontal molds is that the castings have lower properties because the metal can not be poured into the mold as quickly as in vertical molds. However, mold's length in vertical is shorter then that in horizontal. In general, walls of molds must be of proper thickness for chilling of the molten metals, the average of the mold wall thickness 2.5 times the casting thickness. The most preferable shapes for centrifugal casting are cylindrical and it's also suitable for hollow parts, such as pipes. Centrifugal casting has greater reliability than static castings. They are relatively free from gas and shrinkage porosity. Since most of the impurities and inclusions are concentrate closer to the inner diameter of the castings and that would be machined away easily. The technique of employing centrifugal force to make castings had been known for a long time, it was A. G. Eckhardt’s original patent of 1809 which revealed understanding the basic principles involved [1]. Centrifugal casting can be divided into three categories: True centrifugal casting, semi centrifugal casting and centrifuging [2]. In true centrifugal casting, the mold is rotating about its own axis during the solidification. No central cores are used. Molten metal is thrown to the mold wall and held there until MS’08 Jordan

it solidifies. A casting may have a hole through the center, the size determined by the amount of metal poured. An important should be considered in this method is that solidification is directional from the outside toward the center of rotation. In semi centrifugal casting, central cores are used to form irregular contours inside the central cavity, and may be made with or without gate. Wheel or dish shaped parts are commonly cast by this process because the mold is rotating on the vertical axis. Centrifuge casting is the most of the three processes and is often used for nonsymmetrical shapes which can not be rotated around their own axis. In centrifuge and semi centrifugal casting, mechanical properties of castings can be higher than those of other casting methods, but generally lower than forgings [2]. The control of the rheological properties in the centrifugal casting process of thin walled ceramic tubes is achieved by the high rotational speed (2000 rpm). It was found that the slow rotational speed at 1000 rpm did not lead to a homogeneous distribution of the wall thickness during crosslinking, a homogeneous tube with uniform wall thickness. And the ceramic tubes were characterized with respect to their shape stability, phase formation, microstructure, and bending strength [3]. Kyung-Hee Kima, et al. have reviewed and discussed the control of the porosity, pore size distribution and the permeability of the microstructure of alumina tubes of membrane by varying particle size of starting alumina powders and controlling sintering temperature. This investigation demonstrates that centrifugal casting is a promising technique to fabricate different tubes [4].

30

SPEED CONTROL FOR DIFFERENT CYLINDRICAL MATERIALS OF CENTRIFUGAL CASTING MACHINE

2. Machine description and calculation of rotational speed The horizontal centrifugal casting machine is consisting of three main parts, the first part is the electrical motor whose maximum speed rotating at 1400 rpm and its power 3/4 hp. The motor is constructed on moving table where the mold was mounted and connected with the motor as well. Then the mold which is the second part of the machine is made of alloy steel, and is used to produce cylindrical shape for only one size. The diameter and the length of the rotating cylinder are 135 mm and 250 mm. The mold designed to form aluminum alloys and cast irons and capable of resisting the temperature at more than 1000 º c. The third main part is the moving table. This table is made to carry the mold when its feeding by the melted metal, is moving to release the tube (melted metal channel) up to 50 cm, which is the length of the trolley. The whole of the machine components are constructed on the trolley that made of steel [5]. The machine constructed rigidly to prevent excessive vibration. Figure (1) shows illustration of centrifugal casting machine.

Table (1) shows rotational speed of different materials. Material Type ρ (g/cm³) N (RPM)

Aluminu m 2.7 1346.3

Cast Iron 6.8

Zinc

Tin

Steel

Cooper

7.1

7.3

7.2

9

847.9

829

818

823.8

737.9

3. Simulation results The speed control of a discrete DC motor is the one of the commonly occurring control problem in many process industries. As described above, the permanent mold rotated around its axes in order to get a pre-specified cylindrical shape. As explained above that the properties of the products can be influenced if either the speed of the mold is not accurate or another of material poured instead (i.e. operation conditions changes). Therefore, a digital DC control is required to overcome this considerable limitation. In this simulation example, the PID pole-placement proposed by Ali Zayed et al [7] is used to control the speed of the permanent mold. The estimated discrete open loop transfer function of a DC motor can be written as follows:

y (t ) = 1.088 y (t − 1) − 0.2369y (t − 2) + 0.009201u (t − 1) + 0.05709u (t − 2)

(2)

The simulation was performed over 400 samples under set point w(t ) changes every 100 sampling instants (changes four times depending on the type of material poured into the permanent mold), to achieve a required speed for each type of material. Therefore the set point is selected as follows:

Fig (1) illustration of centrifugal casting machine.

2.1 Rotational speed calculation As mentioned above that centrifugal casting involves revolving mold around an axis at a known speed in order to produce the pre-specified cylindrical shape. To overcome this major problem Konstantin equation [6] could be used to obtain the minimum rotational speed of the mold according to the required material. The suggested materials to be formed by this machine are illustrated on table (1): The inner radius of the mold of the machine is constant equal 6.25 cm. The rotational speed of each material may be obtained by equation (1).

N =

1) w(t ) = 1346.3 RPM is selected in the first 100 sampling instants for aluminum. 2) w(t ) = 847.9 RPM is selected in the second sampling instants for cast iron. 3) w(t ) = 823.88 RPM is selected in the third sampling instants for steel. 4) w(t ) = 737.9 RPM is selected in the fourth sampling instants for copper. The output and the control input are shown in the figures (2a) and (2b). 1.5

1

5520

ρ×R

(1)

Where: N Minimum rotational speed of the mold (rpm) ρ Density of the metal (g/cm³) R Radius of the castings (cm)

0.5

Aluminum

Castiron

Steel

Copper

No 0

0

50

100

150

200

250

300

of 350

400

Fig(2a ): The output signal

MS’08 Jordan

31

ABDULBASET M ALEMAM, ALI ZAYED AND MAHMOOD ELFANDI

15

10

5

Cast iron

Copper

Aluminum 0

0

50

Steel

No of samples

100

150

200

250

300

350

400

Fig( 2b): The control input signal

It is clear from the figures (2a) and (2b) that, the transient response is shaped by the choice of the polynomial T when either PID zero-pole placement controller is used. It can also clearly be seen from figures (2a) and (2b) that control action varied in response to the change in the process dynamics.

[3] R. Melcher, P. Cromme, M. Scheffler, P. Greil, “centrifugal casting of thin- walled ceramic tubes from preceramic polymers,” J. Am. Ceram Soc, 86 [7] 121113 (2003). [4] K. Kim, S. Cho, K. Yoon, J. Kim, J. Ha, D. Chun, “ centrifugal casting of aluminum tube for membrane application,” Journal of membrane science 199 (2002) 69-74. [5] A. Alemam, Y. Deep, “Design and manufacturing of centrifugal casting machine of producing aluminum cylinder piston,” national conference general program on diversified vocations, towards advanced vocations. Janzor, Tripoli- Libya, 17-18 dec 2007. [6] S. Sagbeni, Metal Casting, faculty of engineering, Alepp University, Syria [with out date]. [7] A. Zayed, A. Hussain, “A new multivariable non- linear multiple- controller Incorporating a neural network learning sub-model”, The first international conference on brain inspired cognitive systems, Stirling, Scotland, Uk., 29 Aug.-1 Sep, 2004. [8] A. Zayed, A. Hussain and L. Smith, “A new multivariable generalized minimum-variance stochastic self-tuning with pole-zero placement, international journal of control and intelligent systems, 32 (1), 2004, 35-44.

4. Conclusion Different rotation motor speed for a centrifugal casting machine is calculated by using Konstantin equation. A PID pole-placement controller designed for speed control of this machine. The results presented indicate that the proposed PID pole-placement controller best tracks set point and operation conditions changes with the desired speed of response, penalizes the excessive control action, and can deal with nonminimum phase systems. The transient response is shaped by the choice of the pole polynomial T ( z −1 ) , while the zero

~

−1

polynomial H ( z ) can be used to reduce the magnitude of control action or to achieve better set point tracking [7],[8]. As a future-work, this paper can be extended to implement practically using an intelligent adaptive control system.

Acknowledgment The authors gratefully acknowledge the financial support of the Higher Institute of Mechanical Professionals, TripoliLibya, which partly funded this study. Special thanks are due to Dr. Amir Hussain, Stirling University, Scotland, UK for his undying assistance and consultancy.

References [1] W. Heine, R. Loper, C. Rosenenthal, Principles of Metal Casting. Tota McGraw-Hill publishing company Ltd, New Delhi, 1983. [2] American foundary mens, Society, Inc. Aluminum Casting Technology. Despainers, illions, USA, 1986.

MS’08 Jordan

32

CONSTRAINED PREDICTIVE CONTROL FOR MOTION PLATFORM

CONSTRAINED PREDICTIVE CONTROL FOR MOTION PLATFORM M.L.SAIDI *, A.DEBBEH ** *

Department of Electronics, University of Annaba, Algeria [email protected]

**

Department of Electronics, University of Annaba, Algeria [email protected]

Abstract: In the last decades, driving simulators are used for vehicle system development, human factor studies and as a tool for driving training in a safe controlled environment. The vehicle driving simulators are virtual reality devices from which a driver can obtain realistic driving feelings through movements, visual displays, and sound effects in virtual driving situations. They use a motion system in order to provide drivers with the appropriate inertial, proprioceptive and tactile motion cues. The important aspect in motion simulators as the human feel sensations must be close to real as possible. The real vehicles move through large distances, while the motion platforms have restrictions on the range of motion. Then the motion platform must return to reference position in anticipation of the next movement. This movement back to neutral position must occur without the driver realise what is happening. In this paper, we are especially interested by the longitudinal movement control. Because of the difficulty of the classical algorithms as the washout to take into account the limitation of the workspace of the platform and the ability of generalized predictive control strategy, to drive the process output more closely to the reference trajectory in the presence of constraints on the input and output signals, it is very interesting to investigate its utility in driving situations in order to resituate a correct effect of longitudinal accelerations. This strategy of control is applied to a platform designed conjointly by the French National Institute for Transport and Safety (INRETS) and Laboratory of Complex Systems (LSC) – France. The platform is composed of two independent systems linked mechanically, a motorised rail for the longitudinal movement and a motorised seat allowing pitch movement to improve human perception during driving situations. Key words: A driving simulator, motion platform, constrained generalized predictive control.

1. Introduction Driving simulators are developed for investigating human control strategies. The driving simulators have been used effectively for vehicle system development, safety improvement and human factor study. The driving simulators, having their roots on flight simulators applied since the early 1900s, have begun to appear in primitive forms [6], [11]. Many factors should be carefully considered in developing and applying full-scale driving simulators effectively, such as construction costs, application areas and target performance. The driving simulator can not reproduce the vehicle accelerations, then we have to provide an extra degree of realism for driver using it. In the case of continuous accelerations the illusion is generally produced by tilting the driver forward or backward. Such tilt can be interpreted by his/her vestibular system, as either a positive or negative acceleration, depending on the direction of the tilt [9]. In the transient acceleration case, the platform is linearly moved in the same acceleration direction and come back when the acceleration is continuous [5], [9]. The implementation of this technique depends strongly on the architecture of the motion platform, the limits of its workspace and its band-width capacity as well as on the dynamic characteristics of the actuators used to move the platform [5]. MS’08 Jordan

In this presentation, a low-cost motion platform designed and built by INRETS-LSC is briefly presented which is able to animate the simulator’s cabin with a longitudinal movement. The objective is the study of normal driving situations. Here, the constrained generalized predictive control is used to realize the longitudinal movement of the platform inside the limited workspace.

2. Platform modelling The INRETS-LSC driving simulator motion is considered as two independent systems mechanically linked, the rotating driving seat and the longitudinal motion platform. Each of them is driven by a single actuator, d.c. motor. The motion platform can realize translational motions corresponding to one direction (front or back) which correspond to driver's acceleration and deceleration.

ρ

y

x

l

Fig 1: Moving Platform 2.1 The linear motion platform The linear motion of the cabin's set (seat, the vehicle board 33

M.L.SAIDI AND A.DEBBEH

and the driver) is made thanks to a balls screw/nut transmission mechanism driven by a DC actuator. The system components equations are as follows: The actuator's electric equation is: −

u

e

=

R

1

i +

L

1

di dt

(1)

Where u is the armature applied voltage in Volt, e is the back electromotive force in Volt, R1 is the actuator resistance in Ohms, L1 is the armature inductance in Henry and i is the armature current in Amperes. The mechanical equation of the actuator pulling the cabin is:

T

a1

= J

a1

dω dt

a1

+

f

a 1ω a 1

+ T al 1 N 1

feedback disturbances can be anticipated and eliminated. Consider the following locally linearized controlled autoregressive and moving average (CARIMA) time discrete model [2]:

A(q −1) y (k ) = B(q −1)u (k − 1) + e(k ) / ∆

Where u(t), y(t) and e(t) are respectively the control input, the controlled variable, and uncorrelated random sequence at time k; q −1 is the backward shift operator, ∆ is the differencing operator ( ∆ = 1 − q −1 ); and A(q −1), B(q −1) are the polynomials obtained by instantaneous linearization method:

A(q −1) = 1 + a1 q −1 + a 2 q −2 + ... + a n q − n (2)

Where the indexes a and l state respectively for actuator and load, T is the torque in N.m, J is the rotational inertia in kg.m², ω is the rotational speed in rad/sec, f is the rotational armature friction in N.m.sec/rad, and N1 the reduction factor. Where the torque Ta1 and the back emf are given by: e = ke1 a1 (3) T a1 = k t1i , i : armature current; ωa1 : rotational velocity; kt1,ke1 : constants

ω

k t1

The objective of the generalized predictive control strategy is to minimize a cost function based on error between the predicted output of the process and the reference trajectory. The cost function is minimized in order to obtain the optimal control input that is applied to the linear plant (Fig.2). The cost function has the following quadratic form: N2

J=

(4)

2

Nu

∑ [ yˆ(k + j)−r(k + j)] + λ∑∆u (k + j −1) N1

N1 N2

2π ( J 1 s + f 1 )( L1 s + R1 ) + N k e1 k t1 p1 X: cabin’s position U: voltage command signal

(6)

B( q −1) = b0 + b1 q −1 + b2 q −2 + ... + bm q − m

i=

After some mathematical treatment of the physical equations, the following transfer function is obtained:

X 1 = U s

(5)

N

j r

λ ∆

2

i =1

(7) : the minimum prediction horizon; : the maximum prediction horizon; : Control horizon; u : the order of the predictor; : the reference trajectory; : weight factor; : the differentiation operator N2

For more details see [5]. The second degree of freedom (seat pitching) is not discussed in this paper, for interesting person please see [14].

3. Generalized predictive control Predictive control is now widely used in industry and a large number of implementation algorithms has been presented in literature such as Generalized Predictive Control – GPC [2]. Most of these control algorithms use an explicit process model to predict the future behaviour of a plant and because of this, the term Model Predictive Control – MPC is often utilized. The most important advantage of the MPC technologies comes from the process model itself which allows the controller to deal with an exact replica of the real process dynamics, implying a much better control quality [1],[2],[13]. Also, the constraints with respect to input and output signals are directly considered in the control calculation, resulting in very rare or even no constraints violation. The inclusion of the constraints is feature that most clearly distinguishes MPC from other process control techniques. Another important characteristic, which contributes to the success of the MPC technology, is that the MPC algorithms consider plant behaviour over a future horizon in time. Thus, the effects of both feed forward and MS’08 Jordan

NU

Reference trajectory W Past output y

Predicted output Y

……………………..

Control signal U

K-1

K K+1 …

K+Nu…………

K+N2

Time

Fig.2: Generalized predictive control strategy principle Thus, the goal is to drive the future outputs y ( k + j ) close to r ( k + j ) . For N1 = 1 and N 2 = N , the prediction vector: …

Yˆ = [ yˆ (k + 1), yˆ (k + 2),..., yˆ k + N )]T is given by

Yˆ = G∆U + F

(8)

Where 34

CONSTRAINED PREDICTIVE CONTROL FOR MOTION PLATFORM

∆U = [∆u (k ), ∆u (k + 1),..., ∆u (k + N − 1)]T G is an N * N lower triangular matrix and

F = [ f (k + 1), f (k + 2),..., f (k + N )]T

are the predictions of the output by assuming that future control increments are all zero. Then, the control law is given by:

∆U = (GT G + λI ) −1 GT ( R − F ) Where R

After having filtered the acceleration, the signal produced is integrated twice in order to obtain the desired position profile. This is filtered by high-pass filter. This second filter is used for bringing the platform back to its neutral position in order to allow the generation of the following acceleration. The desired position is used as reference trajectory for the platform control (Fig.3).

(9)

= [r (k + 1), r (k + 2),..., r (k + N )]T

Vehicle dynamics

If after a certain horizon N u , control horizon, the increments are assumed to be zero:

Longitudinal Acceleration

Washout filter

∆u (k + j − 1) = 0, 1 ≤ N u < j ≤ N 2 The control law becomes:

∆U = (G1T G1 + λI ) −1 G1T ( R − F )

Desired position

(10)

Predictive Controller

 g ,0,...,0,..............,0   0   g , g ,...,0,...........,0  0  (11) G1 =  1 .,......,.......,......,.......,0     g N −1 , g N −2 ,..., g N −  N u 

Voltage

Direct current motor

Fig.3: Motion platform control diagram

Gj

= EjB

(12)

and E j results from the recursive solution of the Diophantine equation:

1 = E j (q −1) A(q −1)∆ + q − j F j (q −1) (13) deg( E j )

= j − 1 , deg( F j ) = n

As the cost function used in GPC is quadratic then quadratic programming (QP) techniques are well suited for solving the problem of constraints on the control signal, the output signal or on the increments of the control signal.

U min

≤ U ≤ U max

Y min

≤ Y ≤ Y max

(15)

∆U min ≤ ∆U ≤ ∆U max

(16)

(14)

4. Movement restitution algorithm Commands will drive the platform for the short displacements and will drive it back to its neutral position in order that the workspace limit won’t be met and the driver still able to get the sensation of a new acceleration or deceleration operation without he actually realising that it is happened. Transitory Acceleration is obtained by filtering the simulated acceleration signal through a high-pass filter in order to isolate the high frequency component. . MS’08 Jordan

5. Simulation The acceleration signal contains acceleration, deceleration and continuous acceleration phases (Fig. 4). The washout algorithm transforms this acceleration signal into a desired position profile with the ability to drive the platform back to its neutral position during the continuous acceleration phase. As the simulator workspace is limited to – 60 cm and +60 cm then the constrained generalized predictive control is used to avoid the collision. After several tests of selecting values of the design parameters of the GPC corrector, the best choice is: N1=1, N2=7, Nu=4, λ =0.001 which insures that the platform position follows exactly the desired position (Fig. 5). 2

Acceleration (m/s2)

And the matrix (G1T G1 + λI ) is N u * N u . The coefficients of the matrix G1 can be obtained from polynomials G j given by

0 -2 -4 -6 -8 -10 10

20

30

40

50

60

70 Time [s]

Fig. 4: Vehicle longitudinal acceleration

35

M.L.SAIDI AND A.DEBBEH

Consign position Platform position

1

Position [m]

0.5 0

-0.5 -1

-1.5 10

20

30

40

50

N1=1, N2=7, Nu=4, λ=0.001

60

70 Time [s]

Fig.5: Consign position, platform position 6. Conclusion In this paper, we have presented a predictive control of a motion platform designed in order to resituate a correct effect of longitudinal accelerations. The vehicle move through large distances, while simulator motion base has a restricted workspace. To give the driver the impression to do a real motion in the presence of constraints on the position due to the workspace limitation, the constrained generalized predictive control is used. The advantage of the generalized predictive control is to be able to drive the process output closely to the reference trajectory with an acceptable control signal by selecting the best choice of the design parameters: the minimum prediction horizon, the maximum prediction horizon, the control horizon and the weight factor. The results obtained are very satisfactory and showed that the GPC control strategy is very efficient for this type of problem.

[6] B. Repa, W. Wierwille, Driver performance in controlling a driving simulator with varying vehicle response characteristics, SAE Paper 760779, 1976. [7] G. Reymond, A. Kemeny, 2000, Motion cueing in the Reneault Driving Simulator, Vehicle System Dynamics, pp.249-259. [8] G. Reymond, A. Kemeny, J. Droulez, A.Berthoz, Role of Lateral Acceleration in Curve Driving: Driver Model and Experiments on a Real Vehicle and a Driving Simulator, Human Factors, Vol. 43 No. 3, pp. 483-495. [9] Reymond G., Kemeny A, Droulez J., Berthoz A., 1999, Contribution of motion platform to kinesthetic restitution in a driving simulator, DSC2000, Paris, France, pp 33-5 [10] G. Reymond, 2000, contribution respective des stimuli visuels, vestibulaires et proprioceptifs dans la perception du mouvement du conducteur, Paris VI University thesis (in French). [11] B. Richter, Driving simulator studies : the influence of vehicle parameters on safety in critical situations, SAE paper 741105, 1974. [12] I. Seigler, A. Kemeny, 2001, Etude sur la pertinence de la restitution physique du mouvement en simulation de conduite en fonction des caractéristiques physiologiques et psychophysiques de la perception du mouvement propre. [13] R. Soeterboek, 1992, Predictive control. A unified Approach, Prentice-Hall, 1992. [14] L. Nehaoua, A. Amouri, and H. Arioui, Classic and adaptive washout comparison for a low cost driving simulator 13th mediterranean conference on control and automation (MED), IEEE, 27-29 Juin 2005.

Acknowledgment The first author would gratefully like to acknowledge the support from the French Complex Systems Laboratory of Evry enabling his stage visit at the Val d’Essonne University.

References [1] E. F. Camacho, and C. Bordons, 1999, Model Predictive Control, Springer-Velag, London. [2] D. W. Clarke, C. Mohtadi, and P. S. Tuffs, 1987, Generalized Predictive Control, Part 1 and Part 2, Automatica, 2 3 pp.137-160. [3] A. Kemeny, 1999, Simulation et perception du movement, DSC 99 ‘driving simulation conference’, Paris, France, pp33-p5. [4] A. Kheddar, P. H. Garrec, 2002, Architectures de plates– formes mobiles pour simulateurs de conduite automobile, CRIIF . [5] H. Mohellebi, S. Espié, H. Arioui, A. Amouri and A. Kheddar, 2004, Low cost motion platform for driving simulator, 5th international conference on machine automation, ICMA’04, Osaka, Japan

MS’08 Jordan

36

HOS-BASED ARMA MODEL ORDER ESTIMATION

HOS-Based ARMA Model Order Estimation Using the Determinant of Sub-Matrices of the Covariance Matrix Adnan M. Al-Smadi*, IEEE Senior Member and Husam A. Hamad ** *Department of Computer Science, Prince-Hussein Bin Abdullah College for Information Technology Al Al-Bayt University, Al-Mafraq, Jordan [email protected]

**Department of Electronics Engineering, Hijjawi College for Engineering Technology Yarmouk University, Jordan [email protected] Abstract: A new approach to the problem of model order estimation of an autoregressive moving average (ARMA) model using third order cumulants is presented. The proposed technique is based on the determinant of sub-matrices of a data covariance matrix derived from the observed data sequence. The observed sequence is modeled as the output of an ARMA system that is excited by an unobservable input, and is corrupted by zero-mean Gaussian additive noise of unknown variance. The system is driven by a zero-mean independent and identically distributed (i.i.d.) non-Gaussian sequence. Examples are given to demonstrate the performance of the proposed algorithm. Key words: Model order, cumulants, covariance matrix.

1. Introduction Model order selection of autoregressive moving average (ARMA) models is an area of research to which many efforts have been devoted in the past. This problem has been of considerable interest for some time and it has along and continuity history [1]. The reason for this interest is twofold: relevance of the issue in many practical applications and the unsatisfaction got by the users of the existing methods. This unsatisfaction comes from the fact that the problem of order selection is an ill-posed problem; that is, desirable features of an algorithm can hardly be written in mathematical form. In most practical cases, the model order is not known. This vital and crucial step is ignored, chosen rather arbitrarily, or assumed to be available in many of the commonly employed ARMA modeling algorithms. For example, in spectrum analysis and modeling, the problem of model order selection is of most importance [2]. That is because the accuracy of the frequency estimates depends on the estimated order of the prediction filter [3]. Various techniques have been reported in the technical literature for estimating the order of ARMA models [4-9]. One method by Liang et. al. [9] is shown to yield a level of performance for a general ARMA model order estimation never before achieved. This method is derived from the minimum description length (MDL) principle [6, 7]. It is based on the minimum eigenvalue (MEV) of a family of covariance matrices computed from the observed data. The MEV method does not require prior estimation of the model parameters that means fewer computations than the other MDL-based algorithms. Liang et. al. showed that the MDL did not work well at low signal-to-noise ratio (SNR) and is computationally expensive. This is due to the prediction error used in computing MDL that is directly affected by the

accuracy of the parameter estimates. Al-Smadi and Wilkes [10] extended the MEV method in [9] to (EMEV) using third order cumulants. In addition, they extended the original results of Liang to the case of colored Gaussian noise. Higher order statistics (HOS), or cumulants, have received attention in signal processing (see [11], [12] and reference therein). Research in HOS has been in existence for almost four decades. Several papers have been published over the past 30 years dealing with the applications of HOS and especially that of the bispectrum. Examples are geophysics, biomedicine, telecommunications, speech processing, and economic time series [12]. The growth of research in digital signal processing with HOS has been explosive during the past 15 years. That is because Cumulants are generally nonsymmetrical functions of their arguments. Hence, cumulants carry phase information about the ARMA transfer functions. Therefore, cumulants are capable of determining the order of ARMA models that contain all-pass (i.e., phase only) factors inherent in ARMA models. Also, cumulants are capable of identifying non-minimum phase systems and reconstructing non-minimum phase signals if the signals are non-Gaussian. In addition, cumulants of order greater than 2 of a Gaussian process vanish. Hence, cumulants provide a measure of non-Gaussianity. In this paper, we present a new approach to the problem of ARMA model order estimation by utilizing theoretical ideas. The proposed algorithm is based on the determinant of sub-matrices of a data covariance matrix derived from the observed data sequence using third-order cumulants. The observed sequence is modeled as the output of an ARMA system that is excited by an unobservable input, and is corrupted by zero-mean Gaussian additive noise. A comparison will be presented between the new algorithm and the EMEV method [10] for different SNRs on the output

MS’08 Jordan

37

* Professor Adnan M. Al-Smadi is currently on sabbatical leave from Yarmouk University, Jordan.

AL-SMADI AND HAMAD

Rpq = [Dpq]TDpq

signal.

Liang’s et. al. MEV method [9] is based on the MDL [6, 7] and leads to the criterion

2. Problem Formulation Let x(t) denote a real-valued stationary ARMA(p,q) signal given by

x(t ) + a1 x(t − 1) + L + a p (t − p ) = b0 w(t ) + b1w(t − 1) + L + bq w(t − q ) or p

q

∑ai x(t − i) = ∑bi w(t − i) i=0

(1)

i=0

where w(t) is the excitation sequence and x(t) is the noiseless output signal. The excitation signal w(t) is assumed to be zero-mean, non-Gaussian, independent and identically distributed (i.i.d.) process. The parameters a0,…, ap are the AR parameters; the number of AR parameters is the order p. The parameters b0,…,bq are the MA parameters; q is the MA order. Equation (1) can be written symbolically in the more compact form

Ap ( z −1 ) x(t ) = Bq ( z −1 ) w(t )

x(t ) = x(t − k ) ], A p ( z −1 ) = 1 + ∑ ak z −k q

= 1 + ∑ bi z −i

(3) −1

The roots of the polynomial A p ( z ) are denoted the poles −1 of the ARMA process. The roots of Bq ( z ) are the zeros. Processes are called stationary if all poles are within the unit circle, and they are invertible if all zeros are within the unit circle [13]. We model the noisy output as x0(t) = x(t) + v(t)

(4)

where v(t) is additive Gaussian noise. This paper addresses the problem of estimating the orders of a general ARMA model. The method looks for a corner in the cost function obtained from the third order cumulants covariance matrix. Now, the system in Equation (1) can be written in matrix format as Dpq θ = v

(5)

where Dpq is a composite data matrix such that Dpq = [Dp Dq]

where λ min is the minimum eigenvalue of Rpq, p is the number of AR parameters, q is the number of MA parameters, N is the length of the observed noisy sequence. The MEV criterion calculates a table of JMEV(p,q) for all values of p and q. The table is organized so that p increases from left to right while q increases from top to bottom down the table. The search method utilizes row- and column-ratio tables. The tables are formed by dividing each row/column of the JMEV(p,q) by the previous row/column. An estimate of the AR order, p, is set equal to the column number that contains the minimum value of column ratio table. Similarly, the MA order, q, is set equal to the number of the row having the minimum value of the row ratio table. Recently, Al-Smadi and Wilkes [10] extended the MEV criterion in [9] to (EMEV) using third-order cumulants. The extension was made by multiplying both sides of Equation (1) by x(t+m)x(t+n) and taking the expectation which results in

The system in (10) can be represented in a matrix form analogous to (5), that is

and

i =1

(9)

,

k =1

Bq ( z −1 )

λmin ( N 1 / N ) ( p + q )

Cxxx(m,n) + a1 Cxxx(m+1,n+1) + … +apCxxx(m+p,n+p) =Cwxx(m,n) + b1 Cwxx(m+1,n+1)+…+ bq Cwxx(m+q,n+q) (10)

p

−k

JMEV(p,q) =

(2)

where z-1 is the unit delay operator [z

(8)

(6)

( 3) C pq

θ = v(3)

(11)

where v(3) represents modeling error in the cumulant domain, and ( 3) C pq

( 3)

= [ C xxx

( 3) C wxx ]

(12)

( 3) with C xxx containing the cumulants of the observed output ( 3) sequence and C wxx containing the cross-cumulants of the input and output sequences. Hence, the data covariance matrix of third order cumulants is ( 3)

( 3)

RTOC =[ C pq ]T C pq

(13)

The EMEV criterion becomes J(p,q) = λMIN TOC ( N

1/ N

) ( p+ q)

(14)

where λMIN TOC is the minimum eigenvalue of the third order cumulants covariance matrix RTOC. In the ideal case, if p and q are chosen such that p ≥ pt and q ≥ qt (where pt and qt are the true orders), then v(3) = 0 and from (11) and (13) it follows that λMIN TOC will be zero [10]. That is,

θ is the coefficients vector, θ = [1 a1 … ap –1 -b1 … -bq]T

(7)

EMEV(p,q) = J(p,q) =

(15)

and v represents the modeling error. The data covariance matrix is obtained as MS’08 Jordan

38

HOS-BASED ARMA MODEL ORDER ESTIMATION

for p = 0,1,…, pmax and q = 0,1,.., qmax. Hence, the J matrix forms an infinite flat plane of zeros in the pq-plane. Therefore, the correct model order lies in the corner of this flat plane. To search for the corner where J(p,q) drops sharply, the EMEV method utilizes the row/column ratio tables as in the MEV method. The row/column ratio tables’ method was observed in [9] as a method that works without any kind of mathematical proof. Even though this method provides good estimates of the true model order [9,10], it has no justification of why it works. In what follows, a method that uses the determinant of submatrices of the EMEV(p,q) matrix to search for the corner that defines the true order (p,q) will be described. The method is based on determination the singularity of principal sub-matrices through determinants. The algorithm is divided into two parts: MA order and AR order.

2.1 Moving Average (MA) Order To find the estimate of the MA order q, we test a sequence of sub-matrices of increasing orders. The sequence starts at the upper right corner of the EMEV(p,q) matrix; namely at J(1, pmax). Each determinant is a scalar associated with that submatrix. As the determinant of each sub-matrix will give one value, the determinants of all of the sub-matrices of J(p,q) will result in a row vector. In theory, the determinant of a singular matrix is equal to zero. Hence, the row vector will have an entry of zero for each singular sub-matrix. That is, gq = [det(EMEV1) … 0… det(EMEVn-1) det(EMEVn)] (16)

where det(EMEVl) is the determinant of a sub-matrix of order l × l for l =1,2,… ,n. Therefore, in theory, the order qt is estimated to be the location of the first change from non-zero to zero in the gq vector. Notice that the first time the determinant of a sub-matrix will be zero is when a whole row of zeros appears. This is the row that is located below the qt row. Hence, the determinant of this sub-matrix will define the MA order q. 2.2 Autoregressive (AR) Order To find the AR order p, a similar approach is described. Again we test a sequence of sub-matrices of increasing orders. The sequence starts at the lower left corner of the EMEV(p,q) matrix; namely J(qmax ,1). The determinants of all of the sub-matrices of EMEV(p,q) will result in a row vector. Notice that the determinant of a sub-matrix of the EMEV matrix will be zero if all the elements in any row or column vector are zeros (i.e., the sub-matrix will not be full rank). Examining the EMEV(p,q) matrix in (15), it can be seen that the first several values for the determinants in the resulting row vector come from the smaller order sub-matrices such as 1 × 1 and 2 × 2. Also the last several values for the determinants come from the larger order sub-matrices such as (n × n) and (n-1) × (n-1) will not be zeros either. That is, gp=[det(EMEV1)...0…det(EMEVn-1)det(EMEVn)]

(17)

Now, the first time the elements of gp vector changes from non-zero to zero is considered to be the estimate of the true order. Notice that the first time a sub-matrix will be singular is when a whole column of zeros appears. This column of MS’08 Jordan

zeros is located at the right side of the pt column. Hence, this column will define the AR order p. Notice that the computational complexity of the proposed method can be reduced by performing a recursive computation of sub-determinants using the following equation. det(A)=

, j=1,2,…, n

(18)

where Mjk is the determinant of a sub-matrix of order n-1 which is obtained by deleting the row and column of the entry ajk (i.e., the jth row and the kth column). In this way, det(A) is defined in terms of n determinants of order n-2, and so on. The final determinant is of order 2 in which those submatrices consist of single elements whose determinant is defined to be the element itself. Note that Mjk is called the minor of the entry ajk . In practice, most of the determinants will not be zeros. Hence, several corners can be detected. In order to locate the true corner and to facilitate the model order selection process, the largest drop between successive values will be used to define the corner of the estimate of the true model order. That is, each value in the row/column vector will be divided by the previous value and the location of these ratios is used as the estimates of the true orders; i.e., gq(i)/gq(i-1) and gp (i)/ gp (i-1)

(19)

The order q is selected as being the number of the value that has the largest drop in the gq vector; while the order p is selected as being the number of the value that has the largest drop in the gp vector.

3. Simulation Example In this section, we present simulation results concerning the proposed approach to model order selection from only the observed noisy output data. To study the robustness of the algorithm, a number of experiments were performed. In these experiments, the proposed method has been compared with the EMEV method. The computations were performed in MATLAB. A finite length of N=1500 points was considered in each experiment. The driving input sequence is not observed. However, it is needed for computing third order ( 3)

cross-cumulants to construct the matrix C wxx . Therefore, the technique in [10] was used to estimate the input sequence. Example: The time series to be considered is given by x(t) – x(t–1) + 0.5x(t–2) = w(t) – 2w(t–1) + 2w(t–2) (20)

This model has two poles and two zeros. The poles are located at 0.5 ± j0.5, and the zeros at 1 ± j. Note that this model contains an inherent all-pass factor. The noisy output was generated with Gaussian measurement additive noise at different SNRs. The ARMA model order was then estimated by performing 100 independent simulations for both techniques. Each simulation trial has noise with different seeds. The results of both techniques are displayed in Table I.

39

AL-SMADI AND HAMAD

Table I. Model order estimation results for the example Number of correct estimates SNR(dB) EMEV method Proposed method 2 8 6 3 28 30 4 32 36 5 60 64 6 84 86 7 96 98 8 100 100

[12] C.L. Nikias and M.R. Raghveer,” Bispectrum estimation: A digital signal processing framework,” Proceedings of the IEEE, vol.75, no. 7, pp. 869-891, July, 1987. [13] P. T. Broerson, “Automatic spectral analysis with time series models,” IEEE Trans. Instrum. Meas. vol. 51, pp. 211-216, 2002.

4. Conclusion A new approach for selecting the model order for ARMA models has been presented. The method presented is an extension to the results by Al-Smadi and Wilkes. As in the EMEV method, we look for a corner in the tabulation of the cost function EMEV. The corner is detected using the determinant of sub-matrices of the EMEV criterion. Numerical examples show that the two methods perform about the same even though the methodology and the derivations of both techniques are different.

5. References [1] S. Phillai, T. Shim, and D. Youla,” A new technique for ARMA system identification and rational approximation,” IEEE Trans. On Signal Processing, vol. 41, no. 3, pp. 1281-1304, 1993. [2] X. Zhang any Y. Zhang, ”Determination of the MA order of an ARMA process using sample correlation,” IEEE Trans. On Signal Processing, vol. 41, no. 6, pp. 22772280, June 1993. [3] D. W. Tufts and R. Kumaresan, “Accuracy of frequency estimation and its relation to prediction filter order”, Proc. ICASSP, pp. 975-989, 1984. [4] H. Akaike, “Statistical predictor identification,” Ann. Inst. Statist. Math., pp. 203-217, 1970. [5] H. Akaike,” A new look at statistical model identification,” IEEE Trans. Automat. Contr. vol. AC-19, pp. 716-723, 1974. [6] J. Rissanen,” Modeling by shortest data description,” Automatica, vol. 14, pp. 465-471, 1978. [7] G. Schwarz,” Estimating the dimension of a model,” Ann. Statist., pp. 461-464, 1978. [8] E. Parzen,” Some recent advances in time series modeling,” IEEE Trans. Automat. Contr., vol. AC-19, pp. 723-730, December 1974. [9] G. Liang, D.M. Wilkes, and J.A. Cadzow,” ARMA model order estimation based on the eigenvalues of the covariance matrix,” IEEE Trans. On Signal Processing, vol. 41, no. 10, pp. 3003-3009, October 1993. [10] A. Al-Smadi & D. M. Wilkes, “Robust and accurate ARX and ARMA model order estimation of nonGaussian processes”, IEEE Trans. On Signal Processing, vol. 50, no. 3, pp. 759 –763, March 2002. [11] J. Tugnait,” Fitting MA models to linear non-Gaussian random fields using higher order cumulants,” IEEE Trans. On Signal Processing, vol. 45, no. 4, pp. 10451050, 1997.

MS’08 Jordan

40

DESIGN OF SUB-THRESHOLD COMPARATOR USING THE ACM MODEL FOR BIOMEDICAL WIRELESS SENSOR APPLICATION

Design of Sub-threshold Comparator Using the ACM Model for Biomedical Wireless Sensor Application Leila Koushaeian*, Aladin Zayegh * *

Faculty of Heath, Science, and Engineering, Victoria University, PO Box 14428, MCMC, Melbourne, VIC 8001, Australia [email protected] [email protected]

Abstract: One of the solutions to decrease power consumption in wireless sensor devices for biomedical applications is to design an analog circuit in weak inversion region. The conventional model of MOS transistor does not emphasize the different operating regions of the transistor, hence the accuracy and capability of model prediction becomes problematic. Since the transistor behavior in moderate and weak inversion becomes so critical for analog designer, the advanced compact current based MOS model (ACM) has been introduced with less number of parameters and continuity. This Paper presents a different approach in design of a comparator circuit in sub-threshold region by using ACM model, which is valid in all regimes of operation. The methodology is presented here has some advantages such as providing a hand calculation of transistor sizing, reducing the design time without trial-and-error approach to find the optimum dimensions of each transistor in CMOS designed comparator. The comparator consumes 750nW in a standard 0.25µm CMOS process with 1V power supply. The proposed comparator is intended for use in the state of the art biomedical wireless sensor, where it is essential to have lower power consumption. Key words: Sub-threshold, CMOS technology, Comparator, BSIM model, and ACM model.

1. Introduction Power consumption is a major issue in wireless sensor devices used in biomedical applications, as it accounts for a significant proportion of their power budget. Nowadays one of the aims is to reduce power consumption as low as possible in order to be able to provide the required power from ambient energy, such as body s own heat or movement, for the chips, especially for implantable medical devices. The low-power devices, which were designed for biomedical application mostly has been focused on the weak inversion region. In addition, this is not only limited to the low power circuit design. At present, analog circuit design and even RF design demand shifting the transistor’s operating region into weak and moderate inversion region, since it offers a good compromise between the power consumption, linearity, matching, noise, and bandwidth [1]. Therefore, understanding the transistor behavior and providing a consistent transistor model between the different regions of operation becomes an essential for analog circuit designers. The analog circuit design automation is far behind of its digital counterpart. In the design of an analog integrated circuit, the mathematical expression of transistor behavior will help to predict the circuit behavior especially for hand calculation purpose. The transistor square law equation that is valid only in saturation region in large devices becomes linear in short channel also Channel Length Modulation (CLM) and mobility could not be considered constant anymore. Thus, the demand to provide a robust solution for hand-calculation becomes substantially critical for short channel transistor. The transistor modeling for the simulator is slightly different from the one used in hand calculation that means the



MS’08 Jordan

simulator model has to be more complex in order to provide more accuracy. Furthermore, modeling is always a trade-off between the complexity and the convergence. Considering a robust CMOS model and its constraints can cut the cost, time-to-market and the risk of prototype and measurement verification. Berkeley Short-Channel IGFET Model (BSIM model) has been selected as the world s first industry standard model since 1997 [2]. By increasing the demand for accuracy of the BSIM model, it gets more complex by adding the extra parameters to obtain a better and precise model. However, in the BSIM model, transition between different operating regions of the MOS transistor is not consistence. The reason of the inconsistency of the conventional model is due to using the threshold-based model. This threshold-voltage based model divides the MOS operating region into pieces, each one was described with a set of its own equation and there is not unified equation for the whole different region of operation. For this reason, the advanced compact current based (ACM) MOS model was introduced with less number of parameters and more consistency comparing to the BSIM model. The evaluation of the strengths and weakness between the BSIM model and the ACM model is out of the perspective of this paper and more detailed in [3]. The focus of this work is on the design of the comparator in the weak inversion region based on ACM model that allows an accurate and consistent prediction of different operation region of the transistor. However, this method could apply to any other analog circuit in sub-threshold region. This paper is organized as follows: section 2 presents a brief review of the ACM model and its expression. Section 3 discusses the parameter extraction of the ACM model for 0.25 µm CMOS technology. In section 4, the transistor sizing for the comparator in weak inversion region based on ACM



41

LEILA KOUSHACIAN, ALADIN ZAYEGH

model will be presented. The simulation results and conclusion will be discussed in more details in section 5 and 6 respectively.

For weak inversion region the transconductance of transistor is approximately based on the following equation (5)

2. ACM Model Besides the BSIM model, other innovative model is EnzKrummenacher-Vittoz (EKV) model that was derived specially for low power (micro- power) analog circuits by analog circuit designer. As mentioned before, the attempt was done to substitute the threshold voltage base model because of its disadvantage; one approach for this substitution is using the surface potential-based model and another one is chargebased model [1]. The charge-based model offers better expression for hand calculation while a surface-potential model individually cannot offer it [1] Another model, which is similar to EKV model, is the advanced compact current based (ACM) model. It is based on EKV model with a specific difference that, it avoids the use of non-physical interpolating equations for linking the weak and strong inversion mode [3,4]. The main drawback of this model is availability and acceptability in comparison with BSIM model [5]. In the ACM model, the drain current is expressed as a difference between the forward current and the reverse current [4].

The current to transconductance ratio is given by [4]

By substituting the equation (2) in (6), we could obtain the expression for aspect ratio of transistor (7) [4]

3. The Parameter Extraction of ACM Model Where i is forward normalized current and depends only on source-gate which is independent of drain voltage, i is a reverse normalized current that, related to drain-gate voltage which is independent from source, and I the specific current [4]. Different modes of transistor operation could be obtained from the value of i and i . For instance, in saturation region, we could ignore the reverse current and state that the drain current is roughly equal to forward current. If i < 1 and i < 1, then transistor is in weak inversion region, and the whole channel is only weakly inverted [1]. IS which is also called the normalized coefficient is equal to (2) [4] f

r

s

f

r

f

r

The most reliable solution to extract the ACM model is to obtain it from the foundry. Since, it provides the model parameter through its access to extraction software, hardware, and knowledge of process. Otherwise, the only option, despite the fact that it is not accurate, is to extract the ACM model from BSIM model. The important parameter extraction will be discussed more in the following subsection; also, more detail could be found in [6]. 3.1 Specific current (I ) parameter extraction One method to extract the specific current is explained in [6, 7] which is through the calculation of the strong inversion slope of the square root of drain current versus source voltage for a given gate voltage. However, the more accurate method is to determine the specific current from the transconductance efficiency. Transconductance efficiency is defined as the ratio of the transconductance (g ) to drain current [5] .It could be used as an indicator of the mode of transistor operating region; Figure.1 is a plot of g -efficiency versus normalized drain current. The level of inversion is defined as the boundary between the weak and strong inversion when normalized drain current is become equal to one that is called moderate inversion. The center of moderate inversion is the intersection of the weak and strong inversion asymptotes. From the simulation as was shown in Figure.1 the specific current for 0.25µm CMOS technology is approximately around the 10 µA. One of the most important characteristics of the g -efficiency plot is that, there are roughly two decades of drain current between the weak and strong inversion regions [2, 5]. Thus, the boundary of each region could be estimated as follows; for large current (i >10), the transistor obviously operates in strong inversion. For very small currents (i , along with their statistical uncertainties , and , respectively. Our program has been validated on a system by computing exactly and < > using the derivative of the partition function and using the Onsager formula. Our results are in excellent agreement with the theory - absolute errors are for and , and for < > - in the temperature range and 40,000 MC runs. Key words: Glauber Sampling Algorithm, Ising Model, Monte Carlo Simulation, Software.

1. Introduction The Ising model [1-4] is used not only in physical statistics but also in areas of science as different as biology, sociology, economics etc. It is also one of the very few models solvable exactly in some cases and the simplest on displaying a phase transition for dimensions higher than one. The Ising Model can represent not only magnetic systems as iron and nickel but also numerous other physical situations. It is a model where a great number of results are known. Indeed, in two dimensions, Onsager [7] in 1944 has obtained the exact solution in zero magnetic field and Zamolodchikov [8] in 1989, in a nonzero field. However, in three dimensions, no known analytic solution has been obtained to date. The Ising model at zero magnetic field is defined by the Hamiltonian: (1) where the sum means that the interaction is limited to two distinct nearest neighbors. J is the bonding constant: if J>0, the interaction is called ferromagnetic and if J as well as their statistical uncertainties , and , respectively. The exact computation of and < > using the partition function [9] is very memory and time consuming even for small sized lattices hence our choice of the system to validate our program. The lattice being of finite size, we have assumed periodic boundary conditions (PBC) which means that the system is completely translationally invariant. Before carrying out the measurements of the thermodynamic quantities of interest, we have let our system equilibrates (i.e. after about 1,000 MC cycles where one cycle corresponds to iterations). We assumed an initial configuration where the spins are oriented randomly up (+1) and down (-1) in an uncorrelated fashion (i.e. infinite temperature). In all our simulations, we have performed 1,000 MC cycles of equilibration and 40,000 more MC cycles to carry out measurements, and the mean CPU time of one simulation is about 6 hours. Finally, we have varied adaptively the temperature in the range to which we have included the critical temperature . The remainder of this paper presents the thermodynamic quantities of interest, the sampling method used, a discussion of the results obtained, and a summary section in which conclusions are drawn.

2. Simulated thermodynamic quantities With the advent of cheap and powerful PC’s, computer experiments are considered nowadays as tools of choice to explore the nature of the ferromagnetic-paramagnetic phase transition. In this work, we have developed a program to simulate the dynamics of the spins in a square Ising Model using the Monte Carlo method [5] with Glauber sampling scheme [6]. The program was written in C programming language and implemented on PC with MS’08 Jordan

In statistical physics, at equilibrium and a given temperature T, we are interested in the following mean quantities: (3) where is a given configuration of spins. The double summation runs over all the possible configurations ( possible states for an N spins Ising system). The quantity 68

MONTE CARLO SIMULATION OF A 2D ISING MODEL

is computed for a configuration . The denominator in the Eq. (3), , is the partition function of the Ising model: (4) For the Ising model, the interesting thermodynamic quantities are: • the magnetization : (5) •

4. 5. 6. 7.

, choose a random number if then set , else set , update various expectations values: , and < >, the steps (2)-(6) are then repeated in order to obtain a sufficient good representation of states .

This algorithm relaxes the state of the Ising model towards thermal equilibrium at a given . On average it takes steps to update the state of spins on the lattice. These steps define one cycle.

the energy:

4. Results and discussion (6) Other quantities could be deduced from the energy and the magnetization. The specific heat per spin is given by the fluctuations of the energy: (7) and the susceptibility per spin is given by the fluctuations of the magnetization : (8) In this paper, we have not included for the validation of our code the average susceptibility per spin because to our knowledge no analytic solution exists yet for it [10]. We also note that the computation of the mean thermodynamic quantities could begin only after the system has reached its thermal equilibrium. Hence, in a Monte Carlo simulation we have two phases: the first phase, where given an initial configuration, we perform a dynamic of spins such that we bring the system near its equilibrium; and the second phase where the system evolves in the vicinity of its equilibrium and the computation of the mean quantities are performed.

In this section, we look in more details at the thermodynamic quantities outputted by our program. The program reads in input the size of the lattice , the inverse temperature and the number of Monte Carlo simulation , the mean runs , and outputs the mean energy per spin magnetization per spin , and the mean specific heat per spin < > along with their statistical uncertainties , and , respectively. Each site having four neighbors, we have stored , for efficiency reasons, the nine possible selection probabilities in a table to avoid recomputing them each time in a simulation step and allocated one byte for the orientation of the spin. In all our simulations, we have considered a lattice of size with periodic boundary conditions . In all our plots, our numerical measurements are given with the error bars and each dot represents an independent run of 1,000 MC cycles of equilibration and 40,000 more MC cycles of data taking. Our normalized units are such that .

3. Glauber sampling algorithm To study phase transitions we need to supplement the Ising model with dynamics. The most frequently used ones are the Metropolis [11] and Glauber [6] stochastical sampling algorithms. In our simulations, we have found that for this problem, both algorithms give almost identical results hence our choice in this paper of the latter, also known as Monte Carlo with a heat bath. In the Glauber Algorithm every time step consists of the following sub-steps: 1.

2. 3.

establish an initial configuration state with for example the spins randomly oriented up (+1) and down (-1) , choose a random spin from the lattice , calculate a probability where site i is the center site and the summation is over the four nearest neighbors of the center site,

MS’08 Jordan

Figure 1. Mean energy per spin as a function of

temperature. Thick: theory, Dot: simulation. 4.1 Mean system energy versus temperature In Monte Carlo simulations, the energy is the most accurately computed quantity because among all the thermodynamic quantities introduced in section 2, it is the one that fluctuates the least about its mean and reaches rapidly its asymptotic value i.e. has a small transient (see Fig. 1). Indeed, the absolute error is about almost everywhere in the temperature range and it equals 69

M. BOUAMRA AND D. AIT ABDELMALEK

0.00012 at where the theoretical energy per spin was computed using the first derivative with respect to the temperature of the logarithm of the partition function [9]. In the other hand, Fig. 4 shows that the statistical errors are small i.e. for and for

We should also emphasize that the divergence around is characteristic of all phase transitions. In infinite sized systems, the fluctuations themselves actually become infinite. Of course, they are limited by the size of the system in a finite size simulation.

4.2 Mean system magnetization versus temperature The Ising model presents for at zero magnetic field a phase transition from a paramagnetic phase at high temperature to a ferromagnetic phase at low temperature.

Figure 2. Mean magnetization per spin as a function of temperature. Thick: theory, Dot: simulation. Indeed, from Fig. 2 where we have represented the magnetization per spin as a function of the temperature, we note a spontaneous magnetization below (i.e. most of the spins align on the same direction up or down with equal probability), significant fluctuations near , and oscillations around beyond . More precisely, the absolute error for the magnetization is about in the temperature range excepted at where it is equal to 0.008, and Fig. 4 shows that the statistical uncertainties have a sharp peak at i.e. for and for . The theoretical value for the magnetization is given by the Onsager formula [7]: (9)

Figure 3. Mean specific heat per spin as a function of temperature. Thick: theory, Dot: simulation

4.3 Mean specific heat

versus temperature

In Fig. 3 a run of simulations that closely covers the phase transition show that the behavior of the specific heat as a function of the temperature is a sharp and tall peak. We also note that the simulated mean values lie on the exact curve - absolute error less than - where the was computed using the second theoretical derivative with respect to the temperature of the logarithm of the partition function [9]. In the other hand, as it is shown in Fig. 4, the statistical uncertainties tend to increase when we approach from below or above the phase transition temperature i.e. for

, our result is It is important to note that near consistent with the theory which predicts that in the vicinity of a phase transition, the fluctuations in the system blow up. More precisely, the mean magnetization per spin behaves as:

for

and . Once

more, as in the case of the magnetization per spin, our result is consistent with the theory which predicts a power law behavior for the specific heat per spin :

(10)

MS’08 Jordan

70

MONTE CARLO SIMULATION OF A 2D ISING MODEL

(11)

with

logarithmically.

We have also determined numerically that the peak of is at , a value slightly different from i.e. because of the finite size effects.

uncertainties of and < >, and respectively, diverge at the transition temperature as predicted by the theory of critical phenomena [4]. Therefore our code, written in the C programming language, is a reliable tool to study and explore the paramagnetic-ferromagnetic phase transition in a 2D Ising model. Finally, the work is ongoing to generalize our program to the 3D case where no exact solution exists to date.

References [1] E. Ising, Z. der Physik, 31, 253, (1925) [2] M.E.J. Newman and G.T. Barkema, Monte Carlo Methods in Statistical Physics (Clarendon Press, Oxford, 1999) [3] J.M. Yeomans, Statistical Mechanics of Phase Transitions (Clarendon Press, Oxford, 1992) [4] H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena (Clarendon Press, Oxford, 1987) [5] N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth , A.H. Teller, and E. Teller, 153 J. Chem. Phys. 21, 1087 [6] R.J. Glauber, J. Math. Phys. , 4, 294 (1963) [7] L. Onsager, Phys. Rev. , 65, 117 (1944) [8] A.B. Zamolodchikov, Int. J. Mod. Phys. , A4, 4235 (1989) [9] A.E. Ferdinand and M.E. Fisher, Phys. Rev., 185, 832 (1969) [10] W.P. Orrick, B. Nickel, A.J. Guttmann, and J.H.H. Perk, J. Stat. Phys., 102, N°. ¾, 2001 [11] N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H Teller and E. Teller, J. Chem. Phys., 21, 1087, 1953 Figure 4. Statistical uncertainties of
. versus temperature .

5. Conclusion The average thermodynamic quantities of a 2D Ising model : the energy per spin , the magnetization per spin and the specific heat per spin < > are accurately determined by our Monte Carlo simulation program supplemented with the Glauber sampling scheme. The absolute errors for a system with 40,000 MC cycles are about 10-6 for and , and 10-4 for < >. We have also shown that the statistical

MS’08 Jordan

71

AHMED MUSA, AYMAN AL-DMOUR, OSAMA AL-KHALEEL, AND MANSOUR IRSHID

An Efficient Text Compression Technique Using LempelZiv Algorithm Ahmed Musa1, Ayman Al-Dmour2, Osama Al-Khaleel3 and Mansour Irshid4 1 2

Department of Computer Engineering Department of Information Technology Al-Hussein Bin Talal University Ma’an – Jordan {shorman, d.ayman}@ ahu.edu.jo 3

Departmet of Computer Engineering Department of Electrical Engineering Jordan University of science and Technology Irbid- Jordan 4

{oda, mabbadi}@just.edu.jo

Abstract: This paper presents a new idea to compress text files based on adaptive source encoding scheme using Lempel-Ziv algorithms. In the encoding scheme, a pre analysis is performed to count frequent of occurrences of each character in the original source. According to this analysis, each character is replaced by a weighted fixed-length code (e.g., eight bits/character) in lieu of arbitrary codes such as ASCII code. This replacement generates an equivalent binary source with two symbols, zero and one. The eighth-order extension of binary source possesses entropy closer to that of the original one. Afterwards, the bitwise Lempel-Ziv algorithms can be applied to the nth-order extended binary source that contains 2n symbols. The compression technique acquires a high compression ratio when the adaptive encoding scheme is used. Furthermore, the hardware implementation of the compression technique becomes easier. This is because the number of symbols in the generated binary source (two symbols) is drastically reduced in compare with those ones available in the original source (ninety six symbols). Key words: Source encoding, Lempel-Ziv algorithm, Text compression, ASCII code, Source entropy.

1. Introduction Several applications in digital systems such as videoon-demand, teleconferencing, etc., have generated large amount of data. This gigantic amount of data required extra high capacity storage devices and transmission bandwidth. In order to handle such amount of data, new technologies have been emerged. Data compression is one of those promising technologies that are used to handle data storage along with information transfer problems [1], [2]. Needless to say, an efficient data compression technique will save the cost of adding infrastructure which is required to enhance higher data store and transfer capabilities [2], [3]. There are four different metrics used to measure the compression technique performance. Firstly, compression and de-compression speed. Secondly, the amount of memory size that is consumed in the compression and decompression processes. Thirdly, implementation complexity. Finally, the compression ratio which can be measured by the number of bits needed to represent one character generated from the original source. All studies and researches in the field of data compression look forward to improving the aforementioned metrics. Earlier compression techniques tended to use small amount of memory and CPU time. Recently, both of these metrics

MS’08 Jordan

become cheaper because of the advancement in VLSI technology and thus later compression techniques have concentrated on achieving better compression ratio. In this paper, we present an adaptive source encoding scheme to achieve better compression metrics (i.e., implementation complexity, compression ratio, etc.). In the adaptive, the given text is analyzed to count the frequency of occurrence of each available character. In the light of this analysis, all characters available in the text are sorted in ascending order and saved in a header file. Thereafter, the character is mapped into a unique fixed-length weighted binary codeword based on its order. The weighted codeword is akin to that one where an arbitrary eight-bit code (e.g., ASCII code) is given to each character. Thus, non-binary information source can be converted to the binary one. Afterward, the bitwise Lempel-Ziv coding algorithms such as LZ-77, LZ-78, LZMW, etc., are applied on the nth-order extension of the resulting binary source [4], [6]-[8], [10]. The experimental results show that the proposed technique has outperformed the other bitwise compression techniques that use static and arbitrary character encoding scheme [1]. Moreover, compressing the source based on its binary equivalent source will profoundly improve the compression metrics. Here, the bitwise compression process can use bit extension of 1, 2, 3, etc., in lieu of bit extension of 8, 16, 24, etc., character 72

AN EFFICIENT TEXT COMPRESSION TECHNIQUE USING LEMPEL-ZIV ALGORITHM

wise extension [1], [5]. In fact, the significant saving in compression metrics is due to the reduction in the number of symbols in the alphabet of the new extended binary source.

2. Adaptive encoding rule In this study, each character is mapped into a unique fixed-length weighted binary codeword according to its probability of occurrence. The weighted codewords are assigned and ordered based on the following rule: The character in the input text with the largest frequency of occurrence is replaced by a codeword with the largest Hamming weight. For instance, the all ones eight-bit binary codeword, whose Hamming weight is eight, is assigned to the character that has the largest probability of occurrence in the given text. This assignment is repeated until a codeword with a minimum Hamming weight is assigned to the character with the rare frequency of occurrence. In general, codewords with a Hamming weight of (N-k) are assigned to the characters:

N  k

 N ! ( N − k )!  = k! 

(1)

Assigning a codeword for each character using the aforementioned rule will minimize the difference between the first-order entropy of the resulting binary source, H(B), multiplied by the length of the code, N, and the entropy of the original source, H(S). This means that the difference in entropy, ∆H, between the original source and the nth-order extended binary source, is less when the adaptive scheme is used.

∆H = H(S) - NH(B)

(2)

A header file, which includes all characters found in the text sorted in descending order, the extension order (n) and the codeword length (L) of the binary encoded representation of the data stream, must be successfully received by the de-compression process, see Figure 1. The codeword length L is calculated in the following section. The transmission of such information from the compression procedure to the de-compression one will add an extra cost to the entire compression/de-compression algorithm. However, the additional cost of sharing such key information between the compression and decompression procedures is amortized over the compression savings.

Where N is the length of the assigned binary codeword and k is ranged from 0 to N.

Compression process Source Encoder

Source

Source Analyzer

Compression Algorithm

De-compression process

Compressed File

Decompression Algorithm

Source Decoder

Original Source

Header File: - Available characters - Extension order (n) - Binary representation Length (L)

Figure 1: Compression and de-compression process based on adaptive source mapping scheme.

3. Example: Mapping rule utilization Naturally, the frequency of occurrence of each character in English text varies from one file to another [8]. It is a matter of fact that it is difficult to find typical files that can be used to evaluate the performance of compression methods. For this reason, in this study we select a text file at random, which is a novel written by William Morris entitled “The Well at The World End” (WWEND.txt) [9], and it is comply with the criteria identified in [8]. Table 1 lists the frequency of occurrence of some characters available in the WWEND.txt file along with the code assigned for each character based on each encoding scheme. This table shows that the Carriage Return ‘CR’

MS’08 Jordan

character will be coded as ‘00111110’ if the global identification and characteristics of English text is used (i.e., static encoding scheme). On the other hand, if the adaptive scheme is used the CR character will be coded as ‘10111110’. This means that the Hamming weight of the given code is higher when the adaptive scheme is used. This is because the adaptive scheme is coded characters based on their frequency of occurrences in the input source. Thus, compressing a given file using the encoding scheme described in [1] might not lead to an optimal compression ratio. It should be mentioned that the code assigned to each character based on using the adaptive scheme varies from one text file to another. As a result, additional complexity is added to the compression/decompression process.

73

AHMED MUSA, AYMAN AL-DMOUR, OSAMA AL-KHALEEL, AND MANSOUR IRSHID

Table 1: Probabilities of some English characters exist in WWEND10.txt along with their corresponding adaptive, static and ASCII code. Character Space `•` e a ….. Carriage Return `CR` ….. ; r …..

Probability of occurrence 0.181 0.097 0.065 ….. 0.019 ….. 0.0026 0.0022 …..

4. Bitwise Lemple-Ziv algorithms 4.1 Compression process In the compression process one of the well-known compression techniques such as Lempel-Ziv, Huffman, etc., applies to the resulting binary source. Lempel-Ziv compression algorithms offer a significant compression performance for a wide range of data types. The high performance is achieved mainly by processing the data stream and then adaptively constructing a dictionary at both encoding and decoding tasks [11], [12]. Fixed-length codes generated out of these variants algorithms make the decoding process easier. Here, Compression algorithms such as bitwise LZ-78 algorithm can be used to manipulate the nth-order extension binary source on a bitwise basis. It is a matter of fact that applying LZ-78 algorithm on a bitwise basis simplifies the software and the hardware implementation of such algorithm [1]. This is because the number of symbols in the resulting binary file is drastically reduced comparing with the original source. LZ-78 algorithm constructs a string-based dictionary without limitation on how far back in the dictionary the match can be achieved with the previous strings [11]. This enlarges the probability of finding a match with the previous strings in the evolving dictionary. Furthermore, we start the compression procedure using a pre-filled compression dictionary. This dictionary has been filled with subsequences common in the data being compressed [13]. In the bitwise LZ-78 technique, subsequences are determined based on the extension-order at which the binary file will be manipulated. As an example, if the bitwise LZ-78 algorithm manipulates the binary file at an extension-order of four (i.e., n= 4), the dictionary will be pre-filled with the following subsequences 0000, 0001, … , 1111. Afterward, the bitwise LZ-78 algorithm starts manipulating the binary data stream. Upon the completion of the bitwise LZ-78 algorithm, a uniform block of bits of length L is generated. The last n bits in the block represent the innovation symbol. The remaining (L-n) bits provide the equivalent binary representation of the pointer to the root subsequence that match the one in question except the MS’08 Jordan

ASCII Code 00100000 11100101 01100001 ….. 00001101 ….. 00111011 01010010 …..

Static Code 11111111 11111110 11101111 ….. 00111110 ….. 01011110 01011111 …..

Adaptive Code 11111111 11111110 11111011 ….. 10111110 ….. 11010111 10110111 …..

innovation symbol. To have an in-depth understanding on how Lempel-Ziv algorithms manipulate the resulting binary source it is recommended to see the example illustrated in [14]. 4.2 De-compression process De-compression process is basically a table-lookup and can be done by reversing the compression task. Prior to the beginning of the de-compression process, information available in the header file must be successfully extracted. The information include the characters available in the original source sorted in descending order, extension order (n) at which the binary file is manipulated, and the length of the binary encoded representation (L) of the subsequence obtained out of the bitwise LZ-78 algorithm, see Figure 1. Here, a lookup table is also pre-filled with the same subsequences that are used in the compression process. As was mentioned previously, these subsequences are determined based on the extension order (n). Thereafter, the binary data stream stored in the compressed file is parsed into uniform block of bits of length L. The last n bits of each uniform block are excluded. The remaining Ln bits provide the equivalent binary representation of the index of the root subsequence that matches the subsequence under consideration except for the innovation bits. Thus, the decoder simply uses the index to identify the root subsequence and appends the innovation bits. Here the decompression process starts expanding the compressed data without receiving any information about the dictionary format. Eventually, the resulting equivalent binary file retrieves back. In turn, each weighted codeword will be remapped into its corresponding character. The remapping can be simply accomplished by parsing the retrieved binary file into eight-bit codewords. These codewords are replaced by characters based on the received header file.

5. Results and Evaluation of bitwise LempelZiv algorithm To illustrate the performance of the proposed compression technique a corpus of texts has been adaptively encoded and compressed using bitwise LZ-78 74

AN EFFICIENT TEXT COMPRESSION TECHNIQUE USING LEMPEL-ZIV ALGORITHM

algorithm. As an example, a text file WWEND10.txt of size 1.186 MB, is encoded using ASCII code, static scheme as well as the adaptive scheme. Table 1 lists the codewords given to each character available in the file based on the aforementioned encoding schemes. One can compute the entropy of the original source, WWEND10.txt, based on the frequency of occurrence of each character and it is found to be 4.47 bits/character. However, if the same source is encoded using the adaptive scheme, the probabilities of symbol “one” and symbol “zero” are 0.855 and 0.145, respectively, in the generated binary source. Thus, the entropy of the first-order binary source is found to be 0.598 bits/symbol. In turn, the entropy of the nth-order extended binary source is n times that of the first-order one. For example, the entropy of the eighth-order extended binary source, which has the same number of symbols as that of the original source, is 4.78 bits/symbol.

On the other hand, if the static encoding scheme is used, the probabilities of symbol “one” and symbol “zero” are 0.696 and 0.304, respectively. Therefore, the entropy of the resulting first-order binary source is 0.886 bits/symbol and the entropy of the eight-order extended binary source is 7.088 bits/symbol. In comparison, the use of the ASCII code given in [15], which is an arbitrary encoding scheme, to encode the same file leads to entropy of 0.99 bit/symbol and 7.9 bits/symbol for first-order binary source and its correspondence eighth-order extended binary source, respectively. Table 2 summarizes the above entropy computations. From the listed figures one can observe that the difference in entropy between the original source and the nth-order extended binary source is less when the adaptive encoding scheme is used. Therefore, better compression metrics are achieved based on using this scheme in compare with other encoding schemes.

Table 2: Summary of entropy computations based on the first-order binary source entropy. Entropy (bits/Symbol) Source Entropy H(S) (bits/Character) 4.47 To examine the idea of the adaptive encoding scheme, a JAVA program is written to implement the bitwise LZ-78 algorithm. This program mainly manipulates the resulting binary file using the bitwise LZ-78 algorithm. The compression ratio as well as the compression/decompression speed at any extension order (n) can be measured. Figure 2 displays the results obtained based on manipulating the resulting binary sources generated based on using the adaptive and static encoding schemes as well as the conventional ASCII encoding scheme at different extension orders (e.g. n = 1, 2, 4, and 8) using bitwise LZ-78 algorithm. In Figure 2, one can observe that the adaptive

ASCII

Static

Adaptive

7.9

7.088

4.78

compression technique outperforms the other compression techniques that manipulates the binary file generated based on using the static encoding technique described in [1] and the conventional ASCII code [15]. Furthermore, it is easy to observe that the bitwise LZ-78 algorithm of the fourthorder extended binary source (i.e., a source with 16 symbols) achieves a compression ratio close to that of the conventional LZ-78 algorithm when it is applied directly into the original source which includes 256 characters regardless the mapping technique that has been used. It is worth mentioning that at the extension order of eight (e.g., one character at a time) the file is condensed to less than half of its original size.

Compression ratio (bit/character)

8 7

CR_ASCII CR_Static CR_Adaptive

6 5 4 3 2 1 1

2

4

8

Extension order (n)

Figure 2: Compression performance of different mapping schemes using LZ-78 algorithm (WWEND10.txt).

MS’08 Jordan

75

AHMED MUSA, AYMAN AL-DMOUR, OSAMA AL-KHALEEL, AND MANSOUR IRSHID

6. Conclusion The non-binary information source is mapped into an equivalent binary information one according to the probability of occurrences of each character. Working on an equivalent binary source with low number of symbols (i.e., zero and one) drastically improves the compression/decompression techniques metrics. Applying bitwise Lempel-Ziv algorithms to nth-order extended binary source using the adaptive mapping scheme outperforms those compression techniques, which

use static or arbitrary encoding schemes in terms of compression ratio. In this work, compression values attained by adopting the adaptive encoding scheme based on using bitwise LZ78 algorithm at an extension order of eight are in the range of 3.8 to 4.3 bits/character. Furthermore, at the extension order of four, the performance of the compression technique is close to that one if the extension order of eight is used despite the source encoding scheme.

References [1]

[2] [3] [4]

[5]

[6]

[7]

[8]

Elabdalla, A., and Irshid, M. (2001). An efficient bitwise Huffman coding technique based on source mapping. Journal of Computers and Electrical Engineering, 27, 265 – 272. Held, G., and Marshall, T. (1991). Data compression. New York: John Wiley and Sons. Weiss, J., and Shremp, D. (1993). Putting data on diet. IEEE Spectrum, 30, 36-39. C. Zeeh, The Lempel Ziv algorithm [online]. Available: http://tuxtina.de/files/seminar/LempelZiv.pdf, last accessed February 2008. Elabdalla, A., Irshid, M., and Nassar, T. (2006). A file splitting technique for reducing the entropy of text files. International Journal of Information Technology, 3, 109 – 113. Ziv, J., and Lempel, A. (1977). A universal algorithm for sequential data compression. IEEE Trans. Information Theory, 23, 337-343. Ziv, J., and Lempel, A. (1978). Compression of individual sequences via variable-rate coding. IEEE Trans. Information Theory, 24, 530-536. Powell, M., Evaluating lossless compression methods [Online]. Available:

MS’08 Jordan

[9] [10]

[11] [12]

[13]

[14] [15]

http://corpus.canterbury.ac.nz/research/evaluate.pdf, last accessed March 2008. http://www.gutenberg.org/browse/authors/t#a53, last accessed February 2008. M. Arimura, Bookmarks on source coding/data compression [Online]. Available: http://www.hn.is.uec.ac.jp/~arimura/compression_li nks.html, last accessed March 2008. M. Nelson, the data compression book, Prentice Hall, 1992. T. Kida, M. Takeda, A. Shinohara, S. Arikawa, Shift-And approach to pattern matching in LZW compressed text, in Proceeding of CMP’99, in : Lecture Notes in Computer Science, vol. 1645, Springer-Verlag, Berlin, 1999, PP. 1-13. J. Reynar, F. Herz, J. Eisner, L. Ungar, Lempel- Ziv data compression technique utilizing a dictionary pre-filled with frequent letter combinations, words and/or phrases, US Patent Issued on September 1999. Available: http://www.patentstorm.us/patents/5951623.html. Haykin, S., (2001). Communication systems. New York: John Wiley and Sons. ASCII character code reference [Online]. Available: http://nemesis.lonestar.org/reference/telecome/codes/as cii.html, last accessed March 2008.

76

SELF-ORGANISING MAPS FOR USER COMMUNITIES

SELF-ORGANISING MAPS FOR USER COMMUNITIES Sennaike O. A., Ojo A. K. and Sofoluwe A. B. Department of Computer Sciences, Faculty of Science University of Lagos, Lagos, Nigeria. [email protected] [email protected] [email protected] Abstract: User communities are usually constructed from user data. User data are fairly static and are explicitly acquired from the user. Usage data, on the other hand, is generated as users interact with a system, thus very dynamic and can be acquired unobtrusively. Usage data, however, present a number of challenges including processing of sequential and variable length interaction data. We propose the use of Self Organising Maps (SOM) in constructing soft bounded user communities based on usage data. We introduce use of the transition matrix for the representation of usage data in order to capture the temporal nature of the data. We further show the applicability of our approach by applying it to call data from a mobile telecommunications network operator. Key words: Self-organising maps, User communities, User modeling.

1. Introduction User models attempt to provide a model of a user. The different kinds of data in a user model can be classified as user data, usage data and environment data [8]. User data comprise the various characteristics of the user. This includes demographic data, user knowledge, user skills and capabilities, user interests and preferences, user goals and plans [8]. Usage data is related to data about user interaction with the system. Environment data relates to data about the user’s environment that are not related to users themselves. User modeling [6, 7] can be viewed as the process of constructing and applying (often computer based) models of individuals and (or) groups of users. It is desirable that initial user groups in the domain in question be identified. These groups share the same interests according to a set of criteria [1] and are referred to as stereotypes. Stereotypes are organized in a single-rooted hierarchical structure, with a stereotype being able to inherit information from several immediate subsumers as in a Lattice [5]. An individual user model is thus represented as a leaf node in the hierarchy. Users can be assigned to one or more stereotype with the users inheriting the characteristics of these stereotypes [5]. Some problems associated with stereotypes [13] include the fact that it is difficult to determine what stereotypes to define for a certain applications; it is also difficult to establish the boundaries between stereotypes; the information created needs to be constantly revised in order to maintain the consistency of the models and the inferences are too general to be used in fine-grained models. Further, the user classifications of the stereotyping method are ad hoc and un-principled and they can be exploited by the adaptive system only after a large number of trials by various kinds of users. The concept of user communities refers to explicitly clustering users with similar behaviour through the users’ MS’08 Jordan

interaction with the system. The idea of user communities was introduced by Jon Orwant in the user modeling system called Doppelganger [Error! Reference source not found.]. The idea of user communities is similar to stereotypes [Error! Reference source not found.], [Error! Reference source not found.] in that they permit prediction of default values for the user model. The Doppelganger system employed the ISODATA clustering algorithm to generate its communities. Other approaches, COBWEB and ITERATE, based on conceptual clustering technique has also been applied [Error! Reference source not found., Error! Reference source not found.]. The Self-Organising Map (SOM) was developed by Teuvo Kohonen in the early 1980's [9]. This artificial neural network tries to emulate the development of topological maps in the brain using locally interconnected networks and an algorithm based on local neighborhoods. A full description of the basic SOM algorithm, a number of ways of improving the performance of the SOM algorithm and a number of variants of the SOM is presented in [10].

2. Construction of User Communities from Usage data A user classification serves as a basis for an adaptive system; it saves and analyzes the data pertaining to each particular user and makes available information relevant to the program’s adaptation to the user in each successive stage. User communities basically capture generalizations about large classes of users. Thus, given incomplete information on a user, the user community the user belongs to will help in eliciting and filling the missing information. Our proposed approach is based on SOM, an unsupervised learning technique which automatically discovers hidden (implicit) relationships in data and is also able to identify clusters in data making it a natural candidate for the 77

SENNAIKE O. A., OJO A. K. AND SOFOLUWE A. B.

automatic construction of user communities. The fact that SOM preserves topological order inherent in the data makes it a particularly attractive approach since users with similar characteristics automatically become neighbours on the SOM grid.

Black

2.1 Methodology

Red

Black

Red Hot

Hot B

Cold

Hot

The first stage is the data collection stage. Usage data may be extracted from an existing database, through some empirical research, sensors, or from some other means. Usage data is collected in an unobtrusive manner. Our data set thus consists of usage data of different users. We refer to each user’s usage data as data item. Each data item is a series of ordered vectors. Data items may have varying lengths. Next is the feature extraction stage. With usage data available, we need to decide on a representation of the data items and subsequently the metric to be used. The metric used will depend on (or sometimes inform) our representation decision. A number of metrics for different representations is presented in [10]. It may also be important to scale the selected features before applying the SOM algorithm. If knowledge of the relative importance of the components of the data items is available, the corresponding dimensions of the input space can be scaled according to this information [4]. The third stage is the transition matrix representation stage. We employ the use of the transition matrix for representing data. This matrix keeps a count of the number of transitions between states. The definition of the states for each variable will usually depend on the nature of the data. With the states defined, each data point is categorised into its constituent state(s) and the transition matrix is constructed. For a problem with one variable and N states, we have a two dimensional N X N matrix. An element a(i,j) in the matrix will indicate the number of transitions from state i to state j. The definition of the states for each variable will usually depend on the nature of the data. With the states defined, each data point is categorised into its constituent state(s) and the transition matrix is constructed. To give a visual example of a problem with more than one variable, we give a hypothetical example for a 2 variable problem, say, colour and temperature. We define the states as follows: Color = {Black, Red, White}; Temperature = {Hot, Cold} We further give the following ordered data as observed in a hypothetical experiment: (Red, Hot), (Black, Cold), (Black, Cold), (White, Hot), (Red, Hot) The transitions can be outlined as follows: Transition A: (Red, Hot), (Black, Cold); Transition B: (Black, Cold), (Black, Cold); Transition C: (Black, Cold), (White, Hot); Transition D: (White, Hot), (Red, Hot)

Cold

Cold

Hot White

Cold

Hot

Hot

Hot

Cold

Cold

Cold A

Hot

White Hot

Hot

Cold

Hot

Hot

Cold

Cold

Cold

Hot

Hot

Hot

Cold

Cold

Cold

D

Cold

C

Hot

Cold

Hot

Cold

Hot Cold

Table 1: Four dimensional Transition matrix The labels (A, B, C, D) in the example above are for illustration only. In practice, the entries will be a count of the number of transitions from one state to the other. Each transition matrix constructed is subsequently transformed into a vector which is used as input for the construction the SOM. This transformation can simply be achieved by listing the elements of each dimension of the transition matrix in order. The construction of the maps follows from the SOM algorithm. A number of maps can be generated for the same data set. Since the SOM algorithm is a stochastic process, the maps generated will be different. A number of techniques that can be used to determine the best map is presented in [4, 3, Error! Reference source not found., Error! Reference source not found.]. During the construction of the map, each data point is labeled so that the training data example can easily be referenced. On visual inspection of the resulting map, clusters are usually noticed. Data points that are close on the map are more similar than ones that are far apart. In many applications, it may be necessary to identify and label these clusters. A number of techniques for cluster visualization have been presented [11]. However, automated approaches to cluster identification have also been researched into. See [Error! Reference source not found., 2]. One of the problems identified in [13] is that it is difficult to establish boundaries between stereotypes. Rather than trying to define distinct boundaries between our user communities, we propose soft boundaries between our user communities. This means that we do not strictly define boundaries for our user communities. Since the main aim of identifying a user’s community is to be able to predict default values for missing information in the user model, defining hard boundaries between our user communities is of little or no value. In fact, allocating users to hard bounded user communities tends to limit the generality of predicted values in the user model. Doppelganger tried to introduce some generality by assigning a user probabilistic membership, matching some communities better than others.

3. The Call Data Example We collected call detail data from a mobile telecommunications operator that consists of calls made by 500 prepaid subscribers over a period of six months totaling 808610 records.

The transitions are visualized below: MS’08 Jordan

78

SELF-ORGANISING MAPS FOR USER COMMUNITIES

Call Time 09/04/2003 11:55 21/04/2003 14:18 23/04/2003 09:30 24/04/2003 18:12 29/04/2003 09:30 30/04/2003 09:26 05/05/2003 13:48 08/05/2003 15:05 09/05/2003 14:25 21/05/2003 18:02 27/05/2003 18:31 28/05/2003 09:24 29/05/2003 10:21 29/05/2003 13:03 29/05/2003 14:00 29/05/2003 16:08 29/05/2003 16:34

Dur 7 57 109 9 5 53 112 94 26 7 6 26 64 11 66 19 7

Call Time 04/06/2003 18:46 06/06/2003 09:13 10/06/2003 13:56 11/06/2003 09:24 19/06/2003 18:25 19/06/2003 18:31 22/06/2003 13:52 25/06/2003 17:38 26/06/2003 17:38 06/07/2003 16:55 07/07/2003 10:06 07/07/2003 10:07 14/07/2003 13:42 16/07/2003 17:37 22/07/2003 12:44 23/07/2003 08:35 23/07/2003 08:44

Dur 52 7 55 20 242 25 6 33 1 10 6 79 44 48 16 27

Call Time 29/07/2003 12:32 29/07/2003 17:40 04/08/2003 09:13 04/08/2003 10:34 04/08/2003 19:09 04/08/2003 19:09 06/08/2003 14:44 06/08/2003 20:50 06/08/2003 20:59 06/08/2003 21:07 20/08/2003 16:04 22/08/2003 10:53 17/09/2003 09:56 18/09/2003 18:05 26/09/2003 16:18 28/09/2003 15:46

Dur 60 17 62

13 64 43 228 202 21 2 89 28 30 29 17

6

Preprocessing of the data involved normalizing calls with third party present (conference calls), selecting only calls originating from the mobile phone operator in question and removing records that had no other party field, reducing the number of call records to 225292. For simplicity, we selected one variable, the duration of the call. The date and time the call was made provided us with information about the order in which the calls were made. The resultant data, after preprocessing stage, were records with variable lengths. We generate a fixed length feature vector for each subscriber by employing the use of a state transition matrix. We define the states (heuristically) for call duration as follows: Call duration less than 10 seconds; Call duration from 10 to 39 seconds Call duration from 30 to 59 seconds Call duration from 60 to 120 seconds Call duration from 120 seconds and above

We then construct a 5 X 5 transition matrix to represent the each user’s call record. The transition matrix is then transformed into a vector with 25 features for each user. A MS’08 Jordan

The maps were trained using the standard SOM algorithm [6].

47

Table 2: Sample subscriber record showing the time the call was made (call time) and the duration in seconds (Dur)

State 1: State 2: State 3: State 4: State 5:

sample normalized feature vector is as follows: 0.16667, 0.16667, 0.41667, 0.25, 0.00, 0.42857, 0.14286, 0.07143, 0.28571, 0.07143, 0.2, 0.4, 0.1, 0.2, 0.1, 0.1, 0.5, 0.3, 0.1, 0.00, 0.00, 0.66667, 0.00, 0.00, 0.33333

Table 3: Sample Map constructed from the subscriber training data using a rectangular topology. The visualization method used employs the use of color to indicate the similarity between adjacent nodes [6]. For the rectangular grid network, the square difference between neighboring units on the trained map is calculated and the value is used to color the edged separating the units. Dark lines are used to indicate strong difference, and light lines to indicate strong resemblance. The results show that the fact that nodes are close together does not necessarily indicate strong similarity. Given an exemplar caller with incomplete information, we can easily identify the community as the winning node and nodes within its neighbourhood. This community is used to predict the missing values. The closer a node is to the winning node, the stronger the influence it has in predicting the missing information of the exemplar caller.

4. Conclusions and future work This paper presented a novel methodology for constructing user communities using usage data. In order to preserve important temporal information on the usage data, we employed the use of a transition matrix to represent usage data. This matrix representation was then used as an input to the Kohonen SOM to successfully generate soft bounded user communities by clustering users with similar usage patterns close together on the generated map. With the map generated, it is possible to identify the user community a user belongs to even when given incomplete data on the user, and thus make some predictions on the user’s usage pattern. We demonstrated the applicability of our methodology by successfully applying it to call data from a mobile 79

SENNAIKE O. A., OJO A. K. AND SOFOLUWE A. B.

telecommunications network operator. For the mobile telecommunications network operator, predicting the usage pattern of its mobile subscribers will assist in planning and subsequently improving the quality of service provided to its customers. There are a number of issues that still need attention in our proposed use of SOM in discovering user communities. In its current form, our approach does not associate raw data directly with its identified user community. A more robust approach will have to provide a way in which raw data can be directly associated with the identified user community without having to go through the transition matrix transformation process and the learning/clustering process. Further, there is need to investigate and establish general factors and techniques that can be used in eliciting salient features in a given data set.

Acknowledgment This research was partly supported by CRC No 99/2000/02 of University of Lagos, Akoka, Nigeria.

References [1] E. Benaki, V. A. Karkaletsis and C. D. Spyropoulos, “Integrating User Modeling Into Information Extraction: The UMIE Prototype,” In Anthony Jameson, Cécile Paris, and Carlo Tasso (Eds.), User Modeling: Proceedings of the sixth International Conference, User Modeling (UM97) , 1997, pp. 55-57. [2] T. Galliat, W. Huisinga and P. Deuflhard, “SelfOrganizing Maps Combined with Eigenmode Analysis for Automated Cluster Identification,” Proceeding of the ICSC Symposia on Neural Computation (NC'2000), Berlin, Germany, 2000. [3] T. Honkela, “Comparisons of self-organized word category maps,” In Proceedings of WSOM'97, Workshop on Self-Organizing Maps, Espoo, Finland, 1997, pp. 298-303. [4] S. Kaski, “Data exploration using self-organizing maps,” Acta Polytechnica Scandinavica, Mathematics, Computing and Management in Engineering Series No. 82. DTech Thesis, Helsinki University of Technology, Finland, 1997. [5] R. Kass and T. Finin, “A general User Modeling Facility, In Proceedings of Computer Human Interaction,” Association of Computing Machinery (ACM), New York City, 1998, pp. 145-150. [6] http://odur.let.rug.nl/~kleiweg/kohonen/kohonen.html, 2001. [7] A. Kobsa, “Generic User Modeling Servers”, User Modeling and User-Adapted Interaction, Vol. 11, 2001, pp. 49-63. [8] A. Kobsa, J. Koenemann, and W. Pohl, “Personalised hypermedia presentation techniques for improving online customer relationships,” The Knowledge Engineering Review, 16(2), Cambridge University Press, 2001, pp. 111-155. [9] T. Kohonen, “Self-Organising Maps of Massive Document Collections,” Proceedings of the IEEE-INNSENNS International Joint Conference on Neural Networks (IJCNN'00), 2000, Vol. 2, pp. 3-12.

MS’08 Jordan

[10] T. Kohonen, Self-Organising Maps, Springer-Verlag, Berlin, 2001. [11] D. Merkl. and A. Rauber, “Alternative ways for cluster visualization in self-organizing maps,” In Proceedings of WSOM'97, Workshop on Self-Organizing Maps, Espoo, Finland, 1997, pp. 106-111. [12] J. Orwant, Dopelgänger Goes to School : Machine Learning for User Modeling, M.Sc. Thesis, MIT, 1993. [13] A. M. Paiva, About User and Learner Modeling – an Overview, 1995. [14] G. Paliouras, C. Papatheodorou, V. Karkaletsis, C. Spyropoulos and V. Malaveta, “Learning User Communities for Improving the Services of Information Providers,” Lecture Notes in Computer Science, Springer-Verlag, 1998, Vol. 1513, pp. 367-384. [15] G. Paliouras, V. Karkaletsis, C. Papatheodorou and C. D. Spyropoulos, “Exploiting learning techniques for the acquisition of user stereotypes and communities,” in J. Kay (ed.) UM99 User Modeling: Proceedings of the Seventh International Conference, Springer-Verlag, 1999, pp. 45–54. [16] M. Siponen, J. Vesanto, O. Simula and P. Vasara, “An approach to automated interpretation of SOM,” Advances in Self-Organising Maps, Springer, 2001, pp. 89-94. [17] J. Vesanto, Using SOM in Data Mining. Licentiate Thesis, Helsinki University of Technology, Finland, 2000. [18] A. Ypma and R. P. W. Duin, “Novelty detection using self-organizing maps,” In Nikola Kasabov, Robert Kozma, Kitty Ko, Robert O'Shea, George Coghill, and Tom Gedeon, editors, Progress in Connectionist-Based Information Systems, Springer, London, 1997, Vol. 2, pp. 1322-1325.

80

INFORMATION HIDING

A Novel Technique in Multimedia Information Hiding using Quantization Level Based Visual Cryptography Randa A Al-Dallah*, Moussa H Abdallah** and Rola I Al-Khalid*** *

Information Technology Department, Al-Balqa Applied University, Salt, Jordan [email protected]

**

Department of Electronics Engineering, Princess Sumaya University for Technology, Amman, Jordan [email protected] ***

Department of Computer Information System, University of Jordan, Amman, Jordan [email protected]

Abstract: The idea of stego-cryptography is to simplify the embedding technique by having more layers of cryptic shares. A novel approach is presented for multimedia information hiding using quantization level based visual cryptography. The technique proves simplicity and retrieved information quality. The experiments done on various images and the results obtained showed the robustness of the technique. Key words: Information hiding, Quantization level, visual cryptography, color images.

1. Introduction The secret sharing technique (SST) was first introduced by Naor and Shamir [1]. The SST is a method to protect a master key by breaking it to a set of participants, and only qualified subsets of participants can retrieve the master key by combining their shares. The method is also called a threshold scheme such that for (k,n) threshold schemes, the master key is divided into n different shares. We can recover the master key by combining any k shares where k < n, but k-1 or fewer shares will get no information [2]. There have been many published studies [3]–[8] of visual cryptography. Most of them, however, have concentrated on discussing black-and-white images, and just few of them have proposed methods for processing gray-level and color images. Rijmen and Preneel [9] have proposed a visual cryptography approach for color images. In their approach, each pixel of the color secret image is expanded into a 2×2 block to form two sharing images. Each 2×2 block on the sharing image is filled with red, green, blue and white (transparent), respectively, and hence no clue about the secret image can be identified from any one of these two shares alone. Rijman and Preneel [9] claimed that there would be 24 possible combinations according to the permutation of the four colors. Because human eyes cannot detect the color of a very tiny sub-pixel, the four-pixel colors will be treated as an average color. Chang Hou [10] achieved a certain degree of sharing color image information, the drawback is that secret images must be decrypted with heavy computation, which would violate the principle of visual cryptography that uses human eyes to decrypt secrete images. Our technique combines both steganographic and visual cryptographic effects to improve the security of the data.

MS’08 Jordan

Using this technique (stego-cryptography) we embed multimedia intelligent information such as pictures, voices, biometric fingerprints, personal profiles, medical records and family history in a cover image. We provide three levels of security: First, using a new method based on visual cryptography to encrypt the personal picture and produce three shares, then embed the other personal information in a random share using any steganography technique such as the least-significant-bits (LSBs) method. In the LSBs, we use a permutation of pixel locations in which to embed the bits, so that the attacker may not be able to locate the secret data. Finally we embed the encrypted image which carries the embedded data in a cover image using also any steganography technique.

2. The proposed technique 2.1 Hiding phase We propose a new method “Quantization Level (QL)” based on visual cryptography to encrypt the secret image. The QL method will produce three shares and a private key to decrypt these shares. The size of the shares will depend on the encryption method. The private key of the QL method is an integer number. The QL method uses the subtractive CMY model. A CMY color is equal to a set of three intensity values, one for each color component. The intensity of a color component can be defined as the gray level of 8 bits. Each single color based on C, M, and Y can represent 0-255 variations of scale. In the (C, M, Y) representation, (0, 0, 0) represents full white and (255, 255, 255) represents full black.

81

RANDA A AL-DALLAH, MOUSSA H ABDALLAH AND ROLA I AL-KHALID

2.1.1 QL method based on color visual cryptography: The color visual cryptography method [10] used digital halftoning technique. The purpose of digital halftoning technique is to convert the grayscale image into a bi-level image, where each pixel has only two possible values: blank or not blank. Then apply the original visual cryptography technique. In the color image, each color components (C, M and Y) is treated as a grayscale image to which halftoning is applied independently. The digital halftoning technique results in some downgrading of the secret image quality and this leads to loss in contrast of the decrypted image from its halftone version. A new method based on color visual cryptography is introduced to overcome the effect of the halftone, enhance the quality of the decrypted image and to allow visual cryptography to be directly applied on color images. In our method, we enhance the quality of the decrypted image by saving 2l of color levels for each pixel instead of bi-levels that are produced by halftoning technique (where l is the number of bits to represent each pixel). This method uses modulus operation for encryption and decryption. However, the modulus operation needs computation but the perfect reconstruction of the image is possible. To encrypt the secret color image, transform it into three color components C, M, and Y. Then reduce the number of the color levels for each color components from 2k to 2l according to the following equation:  S ij • (2 l − 1)   .............................................. (1) S ij = Round  (2 k − 1)    Where 1 ≤ i ≤ N , 1 ≤ j ≤ M and k > l

This will reduce the number of the encrypted bits that represent each pixel (from k to l). If we increase the number of the bits that represent each pixel, then the image quality will increase but the payload will decrease. Thus if k = l then the method can reconstruct exactly the same original image but that will affect the payload drastically. In our

work we will reduce the color levels from 28 to 24 (where k=8 and l=4). To encode each pixel in each obtained color components, we generate a random integer rnd according to a private key l where 0 ≤ rnd ≤ 2 − 1 , and then we apply the following equation: l Sij = ( Sij + rnd ) mod 2 .............................................. (2) Where 2l is the number of levels. Equation (2) uses a random number to increase the security of encryption. We use a private key to guarantee the regeneration of the same random numbers in the decryption method. The private key is used as a signature for encryption and decryption. The previous steps produce three shares, a share for every color components. Each encrypted pixel expands into nxm block where the block size nxm depends on the number of bits l that represents the pixel value. The block size must satisfy the following equations:

l  ………………..……..………………….. (3) 2 l m = 1 − Round   ….……………………...……...…….. (4) 2 n = Round 

To expand the pixel, convert each pixel into l bits and decompose the l bits to l pixels. After that, arrange the l pixels as nxm block. When the original secret image size is NxM and the block size is nxm, then the produces shares size will be WxH, where WxH satisfies the following equations: W = n * N ............................................................... (5) H = m * M ............................................................... (6)

Figure 1 shows the detailed steps to encrypt the colored secret image. Figure 2, shows diagrammatically how the QL encryption method is applied to a colored secret image

Encryption Algorithm using QL method: Input: Colored Secret Image SI of size NxM Output: Three shares ShareC, ShareM ShareY of size WxH Steps:

 C  1  R 

1.

Transfer the colored secret image SI into CMY image S   M  =  1  −  G 

2.

Decompose

3.

Reduce the number of levels for each color components of the image S from 2

4.

To encode the value for each pixel, do the following:

       Y  1  B 

S into its color components S C , S M , S Y . k

l

to 2 , according to equation (1).

∀S Cij ∈ S C : l

a) Generate a random integer rnd according to the private key K1, Where 0 ≤ rnd ≤ 2 − 1 b) Apply equation (2) to the pixel value c) Convert SCij to l bits binary string denoted as b = (b1b2 .....bl −1bl ) 2 d) Decompose the l bits into l pixels. e) Arrange the l pixels as nxm block, where the value of n and m must satisfy equations (3) and (4) respectively. f) Store the outcomes in ShareC 5.

Repeat step 4 for S M and SY except 4.f

6.

Store the outcomes in ShareM and ShareY respectively.

Figure 1. Encryption Algorithm using QL method. MS’08 Jordan

82

INFORMATION HIDING

Figure 2. QL method 2.1.2 First Level of Embedding Technique In this level of embedding technique, we embed the multimedia secret personal information such as personal voice, biometric fingerprint, personal profile and so on in the three shares. Each type of information is stored with its original format, for example the personal voice is stored in wav file, the biometric fingerprint is stored in bmp file and the personal profile is stored in text file. We use the LSB method to embed personal information in the three shares because it is simple and well known technique. To increase the security in the LSB method, we embed the personal information in random share after permutating all the pixels randomly in the share using a private key. The pixel accuracy for the shares is one bit; we will expand the pixel from 1 bit to 4 bits. The three least significant extra bits is used to embed other personal information. Set the three extra bits that will be used to embed other personal information to zero. Before embed the personal information, we must also embed some details about the personal information such as the number of files to be embedded, the type and the size of each file. These

details are called header data. The process used to embed the header data is as follows: Convert the header data into sequence of bits then select one of the three shares randomly to embed the header data in the first least significant bit by apply the OR operation between the header data bit and the first least significant bit of the selected share. Figure 3 shows the embedding operation in the first least significant bit of ShareY. After embedding the header data there are 2 bits available in the share where the header data are stored, and 3 bits free in each of the other two shares, totally we have 8 bits free distributed in the three shares. Use these bits to embed the secret files contents by reading every file byte by byte and convert each byte to a sequence of bits then embed the sequence of bits in the free bits in the shares. After embedding the header data and the secret files, restore the original pixel positions according to the private key by rearranging the pixels using inverse random permutations. The technique is presented in figure 3; and figure 4 shows the technique diagrammatically. The technique shown is generalized to n secret files.

Hiding secret data (Personal Information) into shares Input: Three shares ShareC, ShareM ShareY and the n secret files ( f1 , f 2 ,...., f n ) Output: Shares after hiding data ShareC, ShareM ShareY Steps:

(

)

1.

Generate the header data hd of size (W/8 x H/8) that contains: number of secret files (n), the file type and size for each file f1 , f 2 ,...., f n .

2.

Convert each byte in the header data

3. 4.

Decompose the 8 bits into 8 pixels and store the outcomes in h of size WxH. Generate the random permutation order for the ShareC, ShareM and ShareY according to the private key K1

hdij into 8 bits binary string denoted by b = (b1b2 ....b8 ) 2 Randomize

6.

Randomize Randomize ShareM → ShareM , Share C   → Share C , Share Y   → Share Y Set the value of the first three least significant bits from ShareC, ShareM and ShareY to zero as follows: ∀Share Cij ∈ ShareC , Share Mij ∈ Share M and Share Yij ∈ Share Y , do the following: ShareCij = ShareCij • 248 , ShareMij = ShareMij • 248 , Share Yij = ShareYij • 248 Where • is a bitwise operation Hide the header data (h) into ShareY as follows: Share Y = Share Y + h , Where + is a bitwise operation

7.

Convert each byte in each secret file into 8 bits binary string denoted by bij = (bij1bij 2 ....bij 8 ) 2

8.

Hide the secret files as follows:

5.

9.

Share Cij = Share Cij + 00000 bij1bij 2 bij 3 , Share Mij = Share Mij + 00000bij 4 bij 5 bij 6 , Share Yij = Share Yij + 00000 bij 7 bij 8 0 Where + is a bitwise operation Rearrange the image using inverse random permutations according to the private key K1 Inverse Inverse Inverse Share C   → Share C , ShareM   → ShareM , Share Y   → Share Y

Figure 3. Hiding secret data in the shares.

MS’08 Jordan

83

RANDA A AL-DALLAH, MOUSSA H ABDALLAH AND ROLA I AL-KHALID

Figure 4. The Share Embedding Technique. 2.1.3 Second Level of Embedding Technique In this level, we embed the three shares obtained into color cover image using LSB. We embed the bits of the shares into the four least-significant bit plane of the cover image. Before embedding the three shares, we must first prepare the three shares and the cover as follows: First, expand the number of bits for each pixel in the shares from 4 bits to 8 bits and set the value of the 4 most significant bits to zero. Second, choose the cover image at least of size WxH to be large enough to fit the share where the size of shares is WxH. Once, both the shares and the cover image have been

prepared, embed each color components of the encrypted image into the corresponding color components of the cover image by using OR operation between encrypted and cover image. The technique is presented in Figure 5 and Figure 6. Further, to make the LSB method more robust against casual attacks, apply a random permutation for the bits in each pixel and a random permutation for each pixel in the image according to a private key.

Hiding shares into a Cover Image Input: Three shares ShareC, ShareM ShareY, the Cover image C of size WxH Output: Color Stego Image CI of size WxH Steps:  C  1  R 

1.

Transfer the colored cover image C in to CMY image CI   M  = 1 −  G   Y  1  B       

2.

Decompose CI into its color components: C, M, and Y. CI    →[CI C CI M CI Y ] Hide ShareC in the corresponding color components as follows: a) CI C = CI C • 240 b) ShareC = ShareC • 15 c) CI C = CI C + ShareC Where • and + is a bitwise operation Repeat step 3 for ShareM and ShareY then store the outcomes in CIM and CIY respectively Transfer the shares from CMY to RGB Combine the three shares CIR, CIG and CIB. splitCMY

3.

4. 5. 6.

[CI R

CI G CI B ]    → CI Combine RGB

Figure 5. Hiding shares into a Cover Image.

Figure 6. Cover Image Embedding Technique MS’08 Jordan

84

INFORMATION HIDING

2.2 Extraction Phase

images using the QL method with l equal to 4 and 2x2 block size. It is worth noting that all the images, shown in this paper, have been scaled down to the same size so as to fit the page requirements. As a result, there could be a loss in quality. The secret personal information and cover image used in the experiments are shown in Figure 10. Note the size of the cover image WxH satisfies the equations (5) and (6). Figure 11 shows the results after applying QL method. Figure 11(a) is the secret image after applying equation (1) with l is equal to 4 to reduce number of color levels, figure 11(b), 11(c) and 11(d) are the generated shares with hidden secret information, figure 11(e) is the result of stacking the three shares, figure 11(f) is the stego image that hides the stacked image. Figure 11(g) is the output of the algorithm. Note that the reconstructed secret image in figure 11(g) is the same as the image in figure 11(a), and the other secret information are the same as original secret information.

In the hiding phase, the multimedia secret information's about the person is hidden on the meaningful cover image by using two keys; key K1 to encrypt the secret image and key K2 to embed the secret information in the encrypted image. The same two keys must be used in the receiving side to extract the secret image and information from the stego image. Figure 7 shows the steps to extract the shares that contain all the hidden data from stego image. After that, extract the secret information from the shares according to the private key K2, and save the information in the files with suitable types to the information as shown in figure 8. Then, extract the secret image from the shares according to the encryption method by using the private key K1. Figure 9 shows the steps needed to decrypt the secret image using the QL method.

3. Experimental Results We have conducted several experiments with various Extracting the shares with hidden data from the Stego image Input: Color Stego Image CI of size WxH Output: Shares with hidden data CI C , CI M , CIY of size WxH Steps: 1.

 C  1  R  Transfer the colored cover image CI in to CMY image CI   M  =  1  −  G         Y  1  B 

2.

Decompose CI into its color components: C, M, and Y.

3.

To retrieve the hidden data from the shares do the following: CI C = CI C • 15 , Where



CI   →[CI C splitCMY

CI M

CI Y ]

CI M = CI M • 15 ,

CIY = CIY • 15

is a bitwise operation

Figure 7. Extracting shares with hidden data.

Extracting n Secret files from the Shares Input: Shares with hidden data CI C , CI M , CIY of size WxH Output: n secret files ( f1 , f 2 ,...., f n ) Steps: 1. Generate the random permutation order for each color components of the image CI according to the private key K2 Randomize CI C   → CI C ,

2.

3.

Randomize CI M → CI M ,

Randomize

CI Y   → CI Y

To retrieve the header data hd from CI Y do the following: a) Convert each pixel in CI Y into 8 bits string denoted by b = (b1b2 ....b8 ) 2 b) Retrieve the b8 bits (right-most bits) from all pixels in CI Y and store them in hd c) Combine each 8 bits from hd into one byte To retrieve the file contents for the n secret files do the following: a) Extract each byte bij in each secret file and store the outcomes in fd as follows: bCij = CI Cij • 7 bYij = CI Yij • 6

,

bMij = CI Mij • 7

, Note: multiply by 6 not 7 to remove the header data bit from the CI Y

' ' bCij = bCij > 1

Where : > is shift to right , 1 ≤ i ≤ W , 1 ≤ j ≤ H b) Depending on the header data, split fd into n files according to the size for each file and save each file in the suitable type.

Figure 8. Extracting secret information from the shares.

MS’08 Jordan

85

RANDA A AL-DALLAH, MOUSSA H ABDALLAH AND ROLA I AL-KHALID

Extracting the Secret Image from the Shares using QL method Input: Shares with hidden data CI C , CI M , CI Y of size WxH Output: Color Secret Image S of size NxM Steps:

1.

To decode the value for each pixel in the Secret image, do the following: a)

∀CI Cij ∈ CI C

:

For every nxm block of CI C , retrieve the 4th least significant bit that contains the secret image from the l pixels and convert it into one pixel, where nxm should satisfy the equations (3) and (4) where 1 ≤ i ≤ W , 1 ≤ j ≤ H and

b) Store the outcomes in S Crc .

1 ≤ r ≤ N, 1 ≤ c ≤ M

c) Generate a random integer rnd according to the private key K2, Where 0 ≤ rnd ≤ 2l − 1 d) Apply the following equation to the pixel value if ( S Crc < rnd )

l

and 2 is the number of levels

S Crc = S Crc − rnd + 2 l else S Crc = S Crc − rnd e)

l

k

Change the number of levels of the image S C from 2 to 2 , according to the following equation:

SC =

S C • (2 k − 1)

where l < k

(2 l − 1)

2.

Repeat step 2 for CI M and CIY except 2.b and store the outcomes in S Mrc and SYrc respectively

3.

Combine S C

S M S Y together.

[S

C

]

Combine CMY

S M SY → S

Figure 9. Extracting Secret Image using QL method.

(a)

(b)

(c)

Figure 10. Inputs: personal secret information's and cover image (a) Color secret image of 300 x 300. (b) Other secret information's such as biometric fingerprint (Bmp file), sample of personal voice (wav file) and personal profile (txt file). (c) Color cover image of 600 x 600 pixels.

(a)

(b)

(d)

(c)

(e)

(f)

(g)

Figure 11. Example of XOR or QL method for 16 color levels. (a) Color secret image with 16 color levels. (b) ShareC with secret information. (c) ShareM with secret information. (d) ShareY with secret information. (e) Stacked (encrypted) image (f) Stego image. (g)Secret image after decryption MS’08 Jordan

86

INFORMATION HIDING

4. PSNR The visual cryptography leads to the degradation in the quality of the decrypted images; because using the digital halftoning techniques that results in some downgrading of the original image quality due to loss some information permanently when convert the pixel value from 256 levels to bi-level. To overcome the effects of halftoning techniques, we proposed a new method “QL”. The major contribution of this method is to improve the quality of the decrypted images by preserve 2l levels instead of bi-level for pixel value. The quality of the decrypted image is measured in terms of the Peak Signal-to Noise Ratio (PSNR). This ratio is often used as a quality measurement between the original and the decrypted image. The higher PSNR, the better quality of the decrypted image. To compute the PSNR, we first calculate the meansquared error using the following equation: N

M

∑∑ (Original Im g MSE =

− Decrypted Im g ij )

2

ij

i =1 j =1

N ∗M , Where M and N are the input image size. Then we compute the PSNR using the following equation:  Max 2   PSNR = 10 log10   MSE  , Where Max is the maximum pixel value of the input image. We do several experiments using different inputs and calculate the PSNR for the proposed method. The QL method with l equal to 4 achieve a PSNR of about 37.5dB, but using visual cryptography method[10] with 2x2 block achieve a PSNR of about 28.1dB. The results show that our proposed “QL” method can improve significantly the quality for the decrypted image compared with visual cryptography method.

5. Security Analysis Our proposed technique has perfect security, because we use multilevel of security. One of the security levels is encrypt the image randomly to produce three shares. The second level is hide secret data in the random share and in random position. The third level is hidden secret data has different type such as wav, bmp and txt. The final level is hiding encrypted image and secret data in meaningful cover image in random position. More levels can be implemented depending on initial specifications and keys.

embed multimedia personal information such as color personal image, personal voice, biometric fingerprint, personal profile in a cover image. We provide three levels of security for the personal information: First encrypt the personal image using a QL method and produce three shares, then embed the other personal information in a random share using any steganography technique such as least-significant-bits (LSBs) method. In the embedding technique, we use a permutation of pixel locations at which to embed the bits for more robustness and security. Finally we embed the encrypted image which carries the embedded data in a cover image using also any information hiding technique. The QL approach shows high level of security with the simplest steganographic technique, without the degradation of the quality of the image. The QL technique kept the embedding simplicity and preserves the quality.

References [1] M. Naor and A. Shamir, Visual Cryptography. Proc. Eurocrypt'94, lecture notes in Computer Science, Perugia, Italy, 1994, 1–12. [2] C.N. Yang, New visual secret sharing schemes using probabilistic method, Pattern Recognition Letters, 25(4), 2004, 481-494. [3] M. Nakajima and Y. Yamaguchi, Extended Visual Cryptography for Natural Images, Journal of WSCG, 10(2), 2002, 303–310. [4] C.C. Lin and W.H. Tsai, Secret image sharing with steganography and authentication, Journal of Systems and Software, 73(3), 2004, 405-414. [5] R. Youmaran, A. Adler and A. Miri , An Improved Visual Cryptography Scheme for Secret Hiding. Proc. 23rd Biennial Symposium on Communications (QBSC), Kingston, 2006, 340–343. [6] C. C. Lin and W. H. Tsai, Visual cryptography for graylevel images by dithering techniques, Pattern Recognition Letters, 24(1-3), 2003, 349-358. [7] G. Ateniese, C. Blundo, A. De Santis, and D. R. Stinson, Visual Cryptography for General Access Structures, Information and Computation, 129(2), 1996, 86–106. [8] Z. Zhou, G. R Arce, and G. Di Crescenzo, Halftone Visual Cryptography. Proc. IEEE International Conference on Image Processing, Barcelona, Spain, Sept 2003, 521–524. [9] V. Rijmen and B. Preneel, Efficient colour visual encryption for shared colors of Benetton. Proc. Eurocrypto’96, Rump Session, Berlin, 1996. [10] Y.C.Hou, Visual cryptography for color images, Pattern Recognition, 36(7), 2003, 1619-1629.

6. Conclusion and future work The approach of stego-cryptography focuses on combining the effect of the steganography and quantization level based visual cryptography. Using this approach we

MS’08 Jordan

87

GABER ELSHARAWY

Design of Service-Oriented Architecture Systems Using Identity Management Concepts Gaber Elsharawy Faculty of Science, Al_Azhar UniversityCairo, Egypt. [email protected] Abstract: Service-oriented architecture SOA is regarded as a style of information systems architecture that enables the creation of provisioning of different services that may be under the control of different domains. These services inter-operate based on a formal definition that is independent of the underlying platform and programming language. The interface definition hides the implementation of the language-specific service. SOA-based systems can therefore be independent of development technologies and platforms (such as Java, .NET, … etc). Services written in C# running on .Net platforms and services written in Java running on Java EE platforms, for example, can both be consumed by a common composite application. Applications running on either platform can also consume services running on the other as Web services, which facilitates reuse. The need for precise control over information access has increasingly been met by identity management (IdM) solutions which solve the problem of keeping corporate data safe from both external and internal threats. IdM is not new, it is generally defined as a set of technologies including password management, user management, access control, and user provisioning, Adoption of SOA and identity management enhance and simplify business process workflow, on the other hand it brings with it increased security threats. Similarly, Web services introduce new security concerns which, if not properly addressed, threaten the success of any SOA project. In many business process workflows, one Web service may call other Web services that, in turn, may call multiple other Web services. To ensure that proper privileges and policies are applied across the entire workflow, the concept of identity management must be extended beyond users, to also include Web services, devices and other entities. Key words: Business process reengineering, Computer security, Information system, Modeling.

1. Introduction SOA is a concept for organizing and utilizing distributed applications that may be under the control of different domains. It provides a uniform means to offer, discover, interact with and use capabilities to produce desired effects consistent with measurable preconditions and expectations. SOA modularize application functions and present them as loosely coupled services that support specific business needs. These services are platform independent, and can be used and reused by different departments and different applications. SOA closes the gap between IT and business execution, enabling rapid response to changing market conditions, competition and new customer requirements. SOA is not tied to a specific technology. It may be implemented using a wide range of technologies. It may use a file system mechanism and/or database management system to communicate data to a defined interface specification between processes conforming to the SOA concept. The key is independent services with defined interfaces that can be called to perform their tasks in a standard way, without the service having knowledge about the calling application, and without the application having or needing knowledge of how the service actually performs its tasks. One of the main objectives of SOA is to reuse the functionality of existing systems rather than building them from scratch. A real dependency is a state of affairs in which one system depends on the functionality provided by another. MS’08 Jordan

The problem is that we create artificial dependencies along with real dependencies [4]. If you travel overseas, you know that you must bring a mobile phone charger along with you. The real dependency is that you need power; the artificial dependency is that your plug must fit into the local outlet. Looking at all the varying sizes and shapes of those plugs from different countries, you would notice that some of them are small and compact while many others are big and bulky. Obviously, we cannot remove artificial dependencies, but we try to reduce them. If the artificial dependencies among systems have been reduced, to the minimum, we have achieved loose coupling. In that sense, we should work to reduced artificial dependencies to the minimum but real dependencies should not be altered.

2. SOA in real life SOA is actually everywhere. Let's look at an example of SOA which is likely to be found in your living room. Take a CD player for instance. If you want to play it, you put your CD into a CD player and the player plays it for you. The CD player offers a CD playing service. This is nice because you can replace one CD player with another. You can play the same CD on a portable player or on your expensive stereo. They both offer the same CD playing service, but the quality of service is different.

88

DESIGN OF SERVICE-ORIENTED ARCHITECTURE SYSTEMS USING IDENTITY MANAGEMENT CONCEPTS

Software Agent

Software Agent

End Result

Software Agent

Loose coupled Interfaces

Consume

Service Request

Software Agent Service provider

Figure (1) Structure of SOA interfaces

3. SOA vs. Object Oriented paradigm The idea of SOA departs significantly from that of object oriented paradigm, which strongly suggests that you should bind data and its functions together. In object oriented style, every CD would come with its own player and they are not supposed to be separated. The results of a service are usually the change of state for the consumer but can also be a change of state for the provider or for both. After listening to the music played by your CD player, your mood has changed, for example, from "depressed "to "happy". The reason that we want someone else to do the work for us is that they are experts. Consuming a service is usually cheaper and more effective than doing the work ourselves. Most of us are smart enough to realize that we are not smart enough to be expert in everything. The same rule applies to building software systems. We call it "separation of concerns", and it is regarded as a principle of software engineering [7].

4. SOA constrains Descriptive messages constrained by an extensible schema delivered through the interfaces. Few system behaviors are prescribed by messages. A schema limits the vocabulary and structure of messages. An extensible schema allows new versions of services to be introduced without breaking existing services [10]. As illustrated in the CD player example, interfacing is fundamentally important. If interfaces do not work, systems do not work. Interfacing is also expensive and error-prone for distributed applications. An interface needs to prescribe system behavior, and this is very difficult to implement MS’08 Jordan

across different platforms and languages. Remote interfaces are also the slowest part of most distributed applications. Instead of building new interfaces for each application, it makes sense to reuse a few generic ones for all applications. Since we have only a few generic interfaces available, we must express application-specific semantics in messages. Theoretically, we can send any kind of message over SOA generic interfaces. Interfaces

Application

Message passing

Application

Figure (2) SOA message passing

5. SOA rules There are rules to follow before we can say that architecture is service oriented. These rules should be applied on SOA in order to improve its scalability, performance and, reliability: 1.

The messages must be descriptive, rather than instructive, because the service provider is responsible for performing such service. This is like going to a restaurant; you tell your waiter what you would like to order, but you don't tell their cook how to cook your dish.

2.

The messages should be written in a format, structure, and vocabulary that is understood by all 89

GABER ELSHARAWY

3.

4.

parties. Limiting the vocabulary and structure of messages is a necessity for any efficient communication. The more restricted a message is, the easier it is to understand the message, although it comes at the expense of reduced extensibility.



Service abstraction - Services hide logic from the outside world



Service reusability - Logic is divided into services with the intention of promoting reuse

Extensibility is very important. If messages are not extensible, consumers and providers will be locked into one particular version of a service.. Restriction and extensibility are opposing each others. We need both, and increasing one comes at the expense of reducing the other. The problem is to have a right balance.



Service compensability - Collections of services can be assembled to form composite services



Service autonomy – Services have control over the logic they encapsulate



Service statelessness – Services minimize retaining information specific to an activity

SOA must have a mechanism that enables a consumer to discover which service provider may provide a service needed by the consumer [7].



Service discoverability – This let service be found and assessed via available discovery mechanisms[5]

8. SOA Environment 6. Stateless and Stateful Services Each message that a consumer sends to a provider must contain all necessary information for the provider to process it. This constraint makes a service provider more scalable because the provider does not have to store state information between requests. This is effectively "service in mass production" since each request can be treated as generic. It is also claimed that this constraint improves visibility because any monitoring software can inspect one single request and figure out its intention. There are no intermediate states to worry about, so recovery from partial failure is also relatively easy. This makes a service more reliable Stateful service is difficult to avoid in a number of situations. One situation is to establish a session between a consumer and a provider. A session is typically established for efficiency reasons. For example, sending a security certificate with each request is a serious burden for both any consumer and provider. It is much quicker to replace the certificate with a token shared just between the consumer and provider. Stateful services require both the consumer and the provider to share the same consumer-specific context, which is either included in or referenced by messages exchanged between the provider and the consumer. The drawback of this constraint is that it may reduce the overall scalability of the service provider because it may need to remember the shared context for each consumer. It also increases the coupling between a service provider and a consumer and makes switching service providers more difficult [15].

7. SOA principles The following specific architectural principles for design and service definition focus on specific themes that influence the behavior of a system and the style of its design: •

Service Encapsulation



Service Loose coupling - This minimizes dependencies and only requires that they maintain an awareness of each other



Service contract - Services adhere to a communications agreement, as defined collectively by one or more service description documents

MS’08 Jordan

There are several types of components in SOA environment to consider. In order of increasing consolidation, these can be grouped into data services, business services, and business processes. •

Data services provide consolidated access to data.



Business services contain business logic for specific, well-defined tasks and perform business transactions via data services.



Business processes coordinate multiple business services within the context of a workflow.

Data within SOA generally falls into one of two categories: •

Conversational state: The conversational state is managed by business services and processes and corresponds to currently executing operations, processes, sessions, and workflows.



Persistent data: The persistent data is managed by data services and is usually stored in databases. The role of data services is to provide centralized control to the access of enterprise data by expressing the data in terms of the business needs without requiring any external knowledge of how the data is actually managed. This centralization brings significant scalability and performance challenges. Scalability issues arise when many business services depend on a single data service. Poor scaling data services will become bottlenecks and requests those services to queue. Data services built around a specific use case will provide too much data for simple use cases, and more-complex use cases will need more data, resulting in more service invocations. In either case, performance will be affected [20]. When business services are integrated into complex workflows, the added dependencies decrease availability. If a business process depends on several services, the availability of the process is actually the product of the weaknesses of all the composed services. For example, if a business process depends on six services, each of which achieves 99 percent uptime, the business process itself will have a maximum of 94 percent uptime, meaning more than 800 hours of unplanned downtime each year (94% of hours in a year).

90

DESIGN OF SERVICE-ORIENTED ARCHITECTURE SYSTEMS USING IDENTITY MANAGEMENT CONCEPTS

Business Process

Business Service

Data Service

Resources Figure (3) The Structure of SOA Architecture

In conversational state, such as the hypertext transfer protocol (HTTP) session state utilized by Web services, is often short-lived, rapidly modified, and repeatedly used. The life span of the data may be a matter of seconds, spanning a dozen requests, each of which may need to read or update the data. Moving from traditional user-centric applications to an SOA environment means that, in addition to users, machines are now accessing services at machine speed. The result is that technologies which were capable of handling traditional user loads are not suitable because of the increased load associated with an SOA deployment. Ensuring the reliability and integrity of conversational state is critical, and difficult to manage by traditional means. Using database servers is the traditional solution for scalable data services, but they cannot cost-effectively meet the throughput and latency requirements of modern large-scale SOA environments. Most in-memory solutions depend on compromises such as queued (asynchronous) updates, master/slave high-availability (HA) solutions, and static partitioning. Most SOA vendors strongly recommend avoiding stateful services if possible, due to these scaling and performance challenges.

8. The need of IdM

password management, user management, access control, and user provisioning. IdM is not new; security controls has been applied for years in many enterprise applications. Recently, organizations and companies have expanded their use of internet-based distribution channels, that needs a precise control over information access; that may be met using IdM solutions. IdM is an important issue in the IT industry and is now associated as the management of a user's credentials and how they might log onto an online system. Identities may be managed by either the entities themselves or by other parties, which may be private parties, for example employers or shops, or public parties like personal records offices and immigration services. The IdM problem is presented and should be solved in the following areas [ 9 ]:•

Access control: Authorization, the ability to manage access on different applications and platforms.



Authentication: The process by which someone proves they are actually who they claim to be. Authentication problem may be solved using smart cards, biometrics or digital signatures.

IdM is defined as a set of technologies including MS’08 Jordan

91

GABER ELSHARAWY



Automatic provisioning: Granting access of specific applications and systems to employees.

This includes creating user IDs and passwords and can include provisioning physical items such as cell phones, computers and key cards. •

Directory: The storage area for user IDs and passwords. It offers one place for a company to view system access across the company.



Federated IdM: The ability to grant system access to parties outside the company’s firewall, such as supplier's and outsourcing partners.



Single sign-on and self-service: The ability to sign on to a system once and then move through the company’s networks without having to repeatedly re-authenticate. Also includes the ability to reset passwords without the assistance of the IT help desk.

In the real world context of engineering online systems, IdM can be given three perspectives [14]: •

The user identification process: creation, management and deletion of identities without regard to access or entitlements.



The user access rights: a smart card and its associated data that a customer uses to log on to a service or services.



The available services to the user: a system that delivers personalized, role-based, online, ondemand, presence-based services to users and their devices.

IdM in the user access perspective would be an integrated system of business processes, policies and technologies that enable organizations to facilitate and control their user access to critical online applications and resources, while protecting confidential personal and business information from unauthorized access. It represents a category of interrelated solutions that are employed to administer user authentication, access rights, access restrictions, account profiles, passwords, and other attributes supportive of user roles/ profiles on one or more applications or systems administrative costs. Enterprises embarking on the SOA path must seriously consider laying out IdM infrastructure first. Selecting the right IdM solution vision support the business today and in the future.

and policies are applied across the entire workflow, the concept of IdM must be extended beyond users, to also include Web services, devices and other entities. Securing applications within an SOA environment should

consider not only the typical threats include message integrity, confidentiality, and availability but also the issues unique to the SOA environment itself, such as: •

Services are not always user-initiated; there is also a substantial amount of application-to-application communication.



Unlike applications, services have multiple points of entry.



Web services operate in heterogeneous environments, so authentication and authorization must be interoperable with all platforms and applications.



Web services strung together can form a complex workflow, which necessitates authentication and authorization at every step, as well as auditing and management of all the affected nodes.

For these reasons, security in an SOA world is complex, and critical. Unfortunately, it’s not realistic to expect the typical Web services developer to understand and implement all of the security functionality needed for a sound SOA implementation. Instead, SOA security needs to be part of a centralized, integrated offering that can be woven into Web services and applications by developers and easily reused. Achieving this goal requires that IdM functions be delivered as an integrated standard Web services

IdM Control

SOA Software

Application Software

9. Integration of IdM and SOA While SOA promises a new level of IT agility, it also brings with it increased security threats. Web services introduce new security concerns which, if not properly addressed, threaten the success of any SOA project. Web services are inherently open and easily accessible; anyone with a computer can call a Web service. For this reason, Web services must be protected by authentication and authorization processes similar to those invoked by users when they access applications today. Also, in any given business process workflow, one Web service will call other Web services that, in turn, might call multiple other Web services. To ensure that proper privileges MS’08 Jordan

Fig 4 Integration of` IdM and SOA with Application Software At the intersection of IdM and SOA is middleware. Indeed, today’s SOA platforms derive many of their core services directly from middleware including security, and by extension, IdM. That’s why middleware solutions have evolved from proprietary tools that join disparate systems and applications together to one of the most important and strategic areas of business today.

92

DESIGN OF SERVICE-ORIENTED ARCHITECTURE SYSTEMS USING IDENTITY MANAGEMENT CONCEPTS

10. Conclusion The increasing rate of change in the modern business environment required greater agility in an organization’s technology infrastructure. This has a direct impact on data management systems. SOA offers a solution of less interdependence between information systems and, thus, responds faster to business needs. SOA can help businesses respond more quickly and cost-effectively to the changing market conditions. This style of architecture promotes reuse at the macro (service) level rather than micro levels (e.g. objects). It can also simplify interconnection and usage of different IT assets. SOA promotes the goal of separating users from the service implementation details. Services can therefore be run on various distributed platforms and be accessed simply across networks. This can also maximize reuse of services and workflow. But it also raises many challenges for SOA designers that should be taken into consideration before any SOA system is deployed such as: •

How will data access services be affected by the increasing number of services and applications that depend on them?



How can we ensure that services don’t fail when underlying services fail?



What happens when the database server reaches full capacity? And how can I ensure the availability of reliable data services even when the database becomes unavailable?

Generally, SOA strategy must ensure data availability, reliability, performance, and scalability. Securing applications within SOA environment should consider not only the typical threats include message integrity, confidentiality, and availability but also the issues unique to the SOA environment itself. The scope of IdM should include all the resources of the company that are used to deliver online services. This includes devices, network equipment, servers, portals, applications and products as well as a user's credentials, address books, preferences, entitlements and even telephone numbers. This may needs a major changes in many organizations in order to reorganize their systems to include SOA and IdM that bring identity coherence to their world. This coherence is required to deliver unified services to very large numbers of users on demand, cheaply and with a successful security solution

Acknowledgment The author would like to thank DR Afaf Abo El Fotoh and Dr Nashaat El Khamessy for their valuable comments and support. The author is also grateful to all his family members for their support and encourage. Thanks also go to all of the staff` in the chair of` computer science, faculty of` science, Al Azhar University for their suggestions that have made the article more enjoyable to read .

[2] Bieberstein, Norbert; Sanjay Bose, Marc Fiammante, Keith Jones, Rawn Shah," Service-Oriented Architecture Compass - Business Value, Planning and Enterprise Roadmap",Upper Saddle River: Pearson,2006 [3] Brian L. Hawkins ," What Higher Ed Leaders Need to Know about IdM" EDUCAUSE Review Articles ,2007 [4] Bloomberg, Jason; Ronald Schmelzer ,"Service- orient or Be Doomed", Hoboken, New Jersey: WILEY,2006. [5] Christopher Koch, "A New Blueprint For The Enterprise", CIO Magazine, Mar 1 2005 [6] Dion Hinchcliffe, "Is Web 2.0 The Global SOA", SOA Web Services Journal, 28 October 2005 [7] Erl, Thomas ,"Service-Oriented Architecture: A Field Guide to Integrating XML and Web Services", Upper Saddle River: Prentice Hall PTR. 2004. [8] Erl, Thomas ,"Service-Oriented Architecture: Concepts, Technology, and Design",Upper Saddle River: Prentice Hall PTR, 2005. [9] Giff Davis “Identity Management Practice” December Media Custom Publishing,2005 [10] Hurwitz, Judith; Robin Bloor, Carol Baroudi, Marcia Kaufman ,"Service Oriented Architecture for Dummies", Hoboken: Wiley,2006 [11] Jones, Steve , "Toward an acceptable definition of service ", IEEE Software, 2005 [12] Krafzig, Dirk; Karl Banke, Dirk Slama , "Enterprise SOA Service Oriented Architecture Best Practices", Upper Saddle River: Prentice Hall PTR, 2004. [13] Mittal, Kunal ,"Requirements process for SOA projects, Part 1 of 3: Capturing requirements for an SOA application - Initial requirements to build out your SOA (HTML)", IBM Developerworks,2006 [14] Norma B. Holland and Steven L. Worona , "Building a Balanced Identity Management Infrastructure, Pennsylvania State University,(EDUCAUSE) 2006 [15] Paul Krill," Make way for SOA and IdM", InfoWorld , May 17, 2006 [16] Pulier, Eric; Hugh Taylor , "Understanding Enterprise identity", Greenwich: Manning Publications, 2006. [17] Richard N. Katz and Ted Dodds , International Study of Identity Management and IT Security in Higher Education ,The University of British Columbia, 2007 [18] Shan, Tony; Hua, Winnie ,"A Service-Oriented Solution Framework for Internet Banking ",International Journal of Web Services Research, Vol. 3, Issue 1, pp 29-48, 2006 [19] Tim McCarthy "Is There Real Business Value Behind the Hype of SOA", Computerworld, June 19, 2006 [20] Wada, Hiroshi; Suzuki, Junichi , "A Service-Oriented Design Framework for Secure Network Applications ", Proc. of the 30th IEEE International Conference on Computer Software and Applications Conference , 2006. [21] Wada, Hiroshi; Suzuki, Junichi ,"A Model-Driven Development Framework for Non-Functional Aspects in Service Oriented Grids ", Proc. of 2nd IEEE International Conference on Autonomic and Autonomous Systems, 2006.

References [1] Barry, Douglas K. "Web Services and Service-Oriented Architectures: The Savvy Manager's Guide", San Francisco: Morgan Kaufmann Publishers,2003 MS’08 Jordan

93

ANAND MOHAN, ALADIN ZAYEGH AND ALEX STOJCEVSKI

A 90 nm Low-Power High-Speed Encoder Design for UWB Flash Analog to Digital Converter Anand Mohan*, Aladin Zayegh* and Alex Stojcevski* *Victoria

University P.O. Box 14428 Melbourne Vic 8001 [email protected] [email protected] [email protected]

Abstract: This paper presents an encoder design for high speed flash analog to digital converters. The paper presents the different design methodologies and techniques. Two of the most popular techniques for the designs including Wallace tree and Priority line based have been discussed in some detail. The two techniques have been designed and implemented to test their functionality for high speed performance. It is seen that priority encoder is the faster of the two techniques, consuming 5 milliWatts of power without pipelining and about 7 milliWatts with two stage pipelining at 4 Gsps. The designs have been implemented using ST-Microelectronics 90 nanometer CMOS technology.

1. INTRODUCTION Ultra Wideband as a communication medium entails the use of a large frequency spectrum from 3.1GHz to 10.6GHz for the purpose of secure high speed low power indoor data transfer. The deregulation of the spectrum has meant that there is effectively 7.5GHz of free spectrum available for data communication. Data transmission in the range of tens to hundreds of gigabits per second are realisable effectively with minimal loss. Transmission of data is in the form of very short sub-nanosecond pulses over short distances. Typical UWB signals are composed of pulses having an effective bandwidth of 550MHz or greater. To effectively receive and digitise the data requires a very high speed and efficient analog to digital converter [1,2]. The following sections detail out the converter topology used and discuss the type of encoding logic.

2. CONVERTER TOPOLOGY There are many different analog to digital converter designs, however to match the requirements of high speed UWB communication the most effective are the Flash architectures. Flash architectures comprise of a number of comparators which compare the input signal with a set of reference voltages generated by a reference ladder. The outputs of each of the comparators are then fed into a high speed encoder to obtain a digital output. The ADC consists of 2 number of comparators for a N bit output. Although the parallel nature of the Flash ADC enables very high speed data conversion it also increases the amount of power consumed by 2 . Therefore the design of two main parts of the Flash ADC, the comparator and the encoder are crucial to meet the requirements for high speed but low power operation. A typical design of a high speed Flash ADC is as shown in figure 1 [3]. N

N

MS’08 Jordan

Figure 1. High Speed Flash ADC Architecture The speed of the encoder is one of the limiting factors affecting the overall performance of the converter. In order to obtain good performance, the architecture of the encoder is crucial.

3. ENCODER ARCHITECURES A typical encoding methodology for Flash architecture is shown in figure 2. It is composed of diodes and a pull down resistor network. This type of topology, although being the simplest and easiest to design has its limitation in terms of speed and functionality. It is well known that the overall performance of the converter in terms of linearity, Signal to Noise Ratio (SNR), Signal to Noise-Distortion Ratio (SNDR) is co-dependent on the performance of the encoder [3,4] In order to design an encoder effectively it is essential that the total number of output bits be set correctly. In analysis performed previously it is estimated that for UWB converters 4 bits output are sufficient to realise a digital code [4] 94

A 90 NM LOW-POWER HIGH-SPEED ENCODER DESIGN FOR UWB FLASH ANALOG TO DIGITAL CONVERTER

3.1 Wallace Tree Structure A popular method of implementing the thermometer to binary encoding is to use a Wallace Tree Structure as shown in figure 4. The Wallace Tree is typically composed of a number of full adders that are linked to obtain an N bit output. Traditionally Wallace Tree encoder structures are built using multiplexers [6]. The Wallace Tree shown in figure 4 was designed for a 4 bit flash ADC and is made up of 11 full adders which encode the 15 bit input to a 4 bit binary logic [6].

Figure 2. Typical ROM encoder In the flash architecture, the outputs of the comparators are linearly increasing or decreasing. This linear variation leads to what is called as a Thermometer encoding methodology. Most modern encoder designs are based on the thermometer to binary conversion scheme. The following set of equations demonstrates the transition from thermometer code to binary output with an intermediate gray code. [5]

(1) Figure 4. Wallace Tree Encoder 3.2 Encoder Sub-Circuit Gate Design (2)

The intermediate gray coding is used to prevent any unnecessary code skipping or metastability issues. Figure 3 shows the implementation of this scheme by the use of universal logic gates. The encoder makes use of 11 NAND/AND gates and 3 XOR gates [5].

The design of each of these individual components was performed using two methodologies. A static implementation was carried out by using only 2:1 Multiplexers on both the topologies. A dynamic implementation using Common Mode Logic (CML) gate design was also performed on both structures shown in figures 3 and 4 and their respective outputs compared for performance. 3.2.1 CML implementation

Figure 5. CML Gate

Figure 3. Thermometer to Binary Encoding MS’08 Jordan

Figure 5 shows the design of a NAND/AND gate using a fully differential CML technique. The input pair transistors are fully differential to provide good immunity against supply 95

ANAND MOHAN, ALADIN ZAYEGH AND ALEX STOJCEVSKI

noise and mismatch. The gate makes use of a differential clocking mechanism too which offers good immunity against linearity errors [5,7]. 3.2.2 MUX Implementation The multiplexer was used to design the full adder circuit. Each of the multiplexers was designed using simple transmission gate logic as shown in figure 6.

Figure 7. CML Latch The other type of latch (Figure 8) used was an inverter chain based flip-flop. The speed of the inverter chain along with the single phase clock effectively determined the total gate delay present in the circuit [9]

Figure 6. 2:1 MUX based Gate From figure 6, the traditional architecture for the multiplexer can be modified to work as a typical gate. The select lines of the multiplexer was used as the secondary input and one of the regular inputs were grounded so as to feed a constant zero to the input logic. The output tapped can then be inverted so as to obtain a AND as well as a NAND gate.

4. PIPELINING Pipelining is an important performance enhancer in terms of the overall delay reduction in the circuit. However pipelining a circuit means that there is an additional power and area consumption. There are few ways to pipeline a circuit including adding buffers, inverters for sharper roll-off but the most effective methodology is by the use of latches and flip-flops. To pipeline a circuit it is necessary to split the circuit into sections and induce the latches where in gate fanin is the maximum. In this paper pipelining was performed on both the encoder architectures and the results compared. Two different types of latches were used to implement the pipelining. One of the latches is a CML based fully differential latch. The advantage of this latch is its very low swing operation, resistance to input jitter and use of a fully differential clocking mechanism. The clocks regulate the speed of operation of the latches and inturn also have an effect in the total gate delay of the encoder. Figure 7 shows the circuit of the CML latch. It is noted that the structure is based on a track and hold mechanism. The load, typically composed of resistors was replaced by N transistors biased carefully in the triode region. This was done so as to maximise the speed of the latch to enable low swing and faster slewing. The only drawback was the increased possibility of supply noise injection but there was enough voltage headroom available to minimise this effect [5,8].

Figure 8. Inverter chain D-Flipflop The following section discusses the implementation results of the two architectures before pipelining and after one and two stage pipelining using the aforementioned latch and flip-flops.

5. RESULTS The propagation delay sets the maximum switching frequency of the encoder. It is seen that however the power consumed by the encoder is proportional to the switching frequency. The following graphs give an overview of the total power consumed for different switching frequencies for both the priority and the Wallace tree configuration. Figures 9 and 10 show the performance of the encoders without pipelining, in terms of their total power consumption and their Probability of Missing Code (PMC).

Figure 9. Power Performance vs Switching Frequency for No Pipelining MS’08 Jordan

96

A 90 NM LOW-POWER HIGH-SPEED ENCODER DESIGN FOR UWB FLASH ANALOG TO DIGITAL CONVERTER

overall with the least probability of missing codes. At lower frequencies the Wallace tree encoder had reasonably good performance in terms of the total power consumed but is susceptible to erroneous code problem.

REFERENCES

Figure 10. Probability of Error vs Switching Frequency for No Pipelining It can be observed from figures 9 and 10 that the power consumed by the encoder increases significantly at higher frequencies. It is also observed that there is a greater chance of missing codes without the use of pipelining due to the increase in the overall propagation delay of the gates. To mitigate these effects one stage and two stage pipelining was performed on the encoder. The results of the pipelining are shown in figures 11 and 12 respectively for power performance and erroneous code. From the two graphs it is observable that pipelining decreases the probability of missing codes significantly. Moreover multistage pipelining does provide a less probability of missing codes, however at an additional cost of increased power consumption.

6. CONCLUSION This paper presented the design techniques for high speed encoders using CMOS 90 nm technology. The designed encoders were tested for their performance on a high speed flash ADC architecture for UWB. From the results it can be observed that pipelining increases the overall performance of the encoder with however an increase in the amount of power consumed. Some critical observations from the graphs conclude that CML designed encoder performed the best

MS’08 Jordan

[1] Federal Communications Commission, “First Report and Order: Revision of Part 15 of the commissions rules regarding ultra wideband systems”, ET Docket No. 98153, FCC, 2002. [2] Australian Communications and Media Authority, “Use of ultra wideband approved for the first time”, Media Release, Media Release No. 24, April 2004. Website http://www.acma.gov.au/ [3] P. P. Newaskar, R. Blazquez, and A. P. Chandrakasan, “A/D precision requirements for digital ultra-wideband radio receivers”, Journal of VLSI Signal Processing Systems for Signal, Image, and Video Technology, vol. 39, pp. 175-188, 2005. [4] A. Stojcevski, H.P. Le, J. Singh and A. Zayegh, “Flash ADC Architecture”, IEEE Electronic Letters, vol. 39, March 2003, pp. 501-502. [5] A. Mohan, A. Zayegh, A. Stojcevski, “A High Speed Analog to Digital Converter for Ultra Wide Band Applications”, Emerging Directions in Embedded and Ubiquitous Computing, Lecture Notes in Computer Science, p.p 169-180. ISBN 978-3-540-77089-3. [6] J. Fernandes and M. Silva, “A Very Low-Power CMOS Parallel A/D Converter For Embedded Applications”, Proceedings of 2005 Intl. Sym. On Circuits and Systems, ISCAS 2004, 2004 vol.1, p.p 1056-9. [7] Razavi, B.: Princi Principles of Data Conversion System Design, IEEE Press, 1995.

[8] Mohan, A., Zayegh, A., Stojcevski, A., Veljanovski, R.: High Speed Ultra Wide Band Comparator in Deep SubMicron CMOS, Int. Sym. On Integrated Circuits, ISIC2007. [9] G.C Wu, “Silicon-on-insulator dynamic d-type-flip-flop”, Patent Number 6737900, May 2004. http://www.freepatentsonline.com/6737900.html.

97

AUTHORS M. B. AL-ZOUBIN. B. VENKATESWARLU AND S. A. ROBERTS

Fast K-MEANS Clustering Algorithms M. B. Al-Zoubi*, N. B. Venkateswarlu ** and S. A. Roberts** *

CIS Dept., University of Jordan, Amman, Jordan [email protected]

**

School of computer studies, Leeds University, UK

Abstract: K-MEANS is one of the most popular clustering algorithms. The CPU time required by K_MEANS is often unacceptable, particularly for large problems. In this article, some new techniques are presented to reduce CPU time. Experiments on two data sets gave very good CPU time savings. Key words: Clustering, K-MEANS, Classification, PCA.

1. Introduction

Tij =

( M iT M i )[ M iT ( M i − 2 M j )] − ( M Tj M j )[ M Tj ( M j − 2 M i )]

(2)

2( M iT M i + M Tj M j − 2 M iT M j )

Clustering techniques have received attention in many areas such as image processing applications for data compression. For large clustering problems such as Vector Quantization [1], the time required by the K-MEANS algorithm [2] is unacceptable, due to the amount of time required to compute nearest neighbors [3], [4].

Since Tij is symmetric, only the upper triangle needs to be computed.

Zaki et al. [5] developed a clustering method called Ensemble Average (EA) method. Venkateswarlu et al. [6] applied the EA method for classifying remote sensing images and reported that the method demands less CPU time than Euclidean classifier method, although both methods gave the same classification.

2.2 Modified Ensemble Average (MEA) Algorithm

In this article, first we investigate the efficiency of the EA method to clustering problems and propose using it with KMEANS. Further, we propose some variants of the EA-KMEANS. The results are compared with the K-MEANS algorithm using two real data sets.

2. K-MEANS Algorithm The K-MEANS algorithm [2] is based on minimizing the sum of squared distances di(X) from all input vectors X in the cluster domain to their cluster centers. Let X1, X2, …, XN be the input vectors to be clustered; let Mi be cluster centers involved. Thus, each input vector X is assigned to cluster (class) ci if di (x) < dj (x) for all j ≠ i, j = 1… K.

The EA method is a new non-parametric classification procedure in which the ensemble average (mean) of the input vectors in each cluster is computed. Then, an input vector X is assigned to cluster ci if

where Tij is a threshold value defined as MS’08 Jordan

In this algorithm we propose a new logic (termed PNND) which is based on the Nearest Neighboring Distance (NND) [7] and Expanded Distance (ED) [6]. The NND of a cluster is on e-half of the distance to is nearest cluster in D-space. In the ED method, di(X) is expressed as:

d i (X) = (X T X - W T M i + M i T M i )

(1)

(3)

where W = 2X. The PNND logic is as follows: if input vector X is assigned to cluster q (in a previous iteration ) and the distance between X and q is less than the NND of q the X is assigned to q; otherwise, apply the EA algorithm . Thus, an input vector X is assigned to cluster ci if

( X T X - W T M q < NND(q) - M q T M q )

2.1 Ensemble Average (EA) Algorithm

X T ( M i - M j ) ≥ Tij , ∀j ≠ i

To assign an input vector to its nearest class, the EA method requires (K – 1)D multiplications, D is the dimensionality, while the original K-MEANS requires K D multiplications.

(4)

otherwise, apply the EA algorithm . 2.3 PCA-MEA Algorithm In this algorithm we use Principal Component Analysis (PCA) with MEA. Given a data set with D variables, it is possible to construct a new set of p variables, p < D which are a linear

98

FAST K-MEANS CLUSTERING ALGORITHMS

Table 1. CPU time for the image data set (N=2048, K=256) with different dimensions, D, first 20 iterations

transformation of the original dimensions [8]. The PCA-MEA algorithm is as follows: 1. Conduct a linear transformation to reduce the original dimension to a smaller one with 95% preserved information. 2. Run the MEA algorithm, with the reduced dimension. 3. Reconstruct the generated cluster conducting an inverse transformation.

centers

by

3. Experimental Results and Discussion

D

4

8

16

32

K-MEANS

45

98

183

343

EA

45

98

186

357

MEA

33

76

125

242

PCA-MEA

5

15

33

55

Table 2. CPU time for the Speech data set (N=2048, K=256) with different dimensions, D, first 20 iterations

To evaluate the proposed algorithms two data sets have been used. The first set represents the widely used Baboon image; the second contains a data extracted from one minute of speech.

D

4

8

16

32

K-MEANS

163

348

650

1214

EA

153

345

643

1225

All algorithms were implemented in C++ programming language and executed on a Sun work station. The CPU time is measured in seconds. The number of dimensions, D, varies s between 4 and 32 and the number of clusters, K, varies between 22 and 210. The approach of Katsavounidis et al. [9] was used to initialize the clusters for all methods.

MEA

48

114

276

448

PCA-MEA

42

91

220

358

Figure 1 shows the performance of each algorithm when applied on the Images data, with varying number of clusters (k).

The clustering obtained from the PCA-MEA algorithm is very close to those of the K-MEANS algorithm. This small difference is likely to be acceptable as long as are seeing 90% savings in CPU time. It is expected that one or two further iterations may be needed to achieve the same results as those of the K-MEANS. This process will be cheap if the MEA algorithm is used.

4. Conclusion In this article, new strategies have been incorporated into the K-MEANS clustering algorithm. These strategies were tested on two data sets. The results show that the percentage of CPU time savings varies between 60 to 90%. The new strategies represent efficient tools to clustering problems. Figure 1. Image data with D = 8, N = 8192, No. PCs = 2 Figure 2 shows the performance of each algorithm when applied on the Speech data, with varying number of clusters (k). The figures show that the MEA and PCA-MEA algorithms performed better than K-MEANS in all cases. Also the EA algorithm performed better than the original KMEANS when K < 512 and D < 8, although this improvement was marginal. We had also tested the algorithm by varying dimensionality, while the number of clusters and samples (input vectors) ware fixed. The results (tables 1 and 2) show that the performance of the MEA and PCA-MFA algorithm continued to perform the best.

Figure 2. Speech data with D = 8, N = 29000, No. PCs = 6

MS’08 Jordan

References [1] A. Gersho and R. Gray, Vector Quantization and Signal Compression. Boston: Kluwer, 1993. [2] J. Mac Queen, “Some methods for classification and analysis of multi-variant observations,’’ Proc. 4th Berkeley Symposium on Mathematics, Statistics and Probability, 1967, pp. 281-297. [3] M. Soleymani and S. Morgera, “An efficient nearest neighbor search method,” IEEE Transactions on Communications, vol. COM-35, 1987, pp. 677-679. [4] N. B. Venkateswarlu and P. S. K. Raju, “ Fast isodata clustering algorithms , pattern Recognition, vol. 25, no. 3, 1992, pp. 335-342. [5] F. Zaki, A. Abd El-Fattah, Y. Enab and S. El-Konyaly, “ An ensemble average classifier for pattern recognition machines,’’ Pattern Recognition, vol. 21, no. 4, 1988, pp. 372 – 332. [6] N. B. Venkateswarlu and P. S. K. Raju “A new fast classifer for remotely sensed imaginary,’’ International Journal of Remote Sensing, vol. 14, no. 2, 1993, pp. 383-389. [7] M. Hodgson, “Reducing computational requirement of the minimum distance classifier, “Remote Sensing of Environments, col.25, 1988, pp. 117-128. 99

AUTHORS M. B. AL-ZOUBIN. B. VENKATESWARLU AND S. A. ROBERTS

[8] C. Chatfield and A. Collins, Introduction to Multivariate Analysis. London, Chapman and Hall, 1980. [9] I. Katsavounidis, C. Kuo and Z. Zhang. “A new initialization technique for generalized Lioyd iteration,

MS’08 Jordan

IEEE Signal Processing Letters, vol. 1, no.10, 1994, pp. 144-146.

100

STRUCTURAL AND MECHANISM DESIGN USING TOPOLOGY OPTIMIZATION

OPTIMAL STRUCTURAL AND MECHANISM DESIGN USING TOPOLOGY OPTIMIZATION G.E. Stavroulakis* N. Kaminakis* Y. Marinakis* M. Marinaki* and N. Skaros* *

Technical University of Crete Department of Production Engineering and Management GR-73100 Chania, Greece [email protected] [email protected] [email protected] [email protected] Abstract: Topology optimization extends the applicability of classical optimal structural design and can be used for the design of compliant mechanisms. Single and multiple criteria optimization problems arise in this context, and are presented in the present paper. They can be solved numerically with several methods (iterative, local optimization, global optimization, or hybrid techniques). A novel hybrid algorithm is proposed and tested here. Representative numerical experiments are presented. The results of this work can be used for the design of optimal structures or microstructures respectively optimal mechanisms and micromechanisms for elastic mechanical problems or in multiphysics. Key words: Global optimization, Mechanism design, Structural optimization, Topology optimization.

1.

Introduction

Optimal structural design is an old discipline which, nevertheless, requires the usage of theory and numerical methods from various areas: structural analysis, optimization, parametric modeling, scientific computing. Even with these tools and the availability of high performance computers, the classical optimal structural design can not get rid of the restrictions posed by the adopted parameterization. Topology optimization allows us find the optimal form of the structure without the previously mentioned restriction. A typical topology design using classical structural optimization is shown in Figure 1, where the optimal design is considered for the thin background truss (structural universe) and the optimal topology is extracted from the nonzero sizes of elements at the solution.

Figure 1. Finding the optimal truss structure from a larger number of potential members (structural universe). First two pictures show optimal design for one vertical loading at the second (resp. forth) node of the bridge, the third example considers both loading cases. Pixel-based topology optimization works similarly for the continuum. We consider in details the topology optimization of compliant mechanisms which leads to nonconvex optimization problems with several local minima. Single and multi-criteria optimization problems are formulated. A novel

MS’08 Jordan

hybrid algorithm is proposed, which is based on the combination of classical, iterative, solution algorithms, with evolutionary algorithms and particle swarm methods. The results of this work can be used for the design of optimal structures or microstructures respectively optimal mechanisms and micromechanisms for elastic mechanical problems or in multiphysics. The novel global optimization algorithms used here are briefly described in Section 2. The formulation of the optimal design problem is outlined in Section 3. Representative numerical experiments are presented in Section 4.

2.

Evolutionary and Nature Inspired Algorithms

2.1 Particle Swarm Optimization Particle swarm optimization (PSO) is a population-based swarm intelligence algorithm. It was originally proposed by Kennedy and Eberhart [1] as a simulation of the social behaviour of social organisms such as bird flocking and fish schooling. PSO uses the physical movements of the individuals in the swarm and has a flexible and well-balanced mechanism to enhance and adapt to the global and local exploration abilities. Most applications of PSO have concentrated on the optimization in continuous spaces while recently, some work has been done to the discrete optimization problem. Recent complete surveys for the Particle Swarm Optimization can be found in [1],[3],[7]. First a number (swarm) of particles is randomly initialized, where a particle is a solution to the problem. The position of each particle is represented by a d-dimensional vector in problem space si = (si1, si2,..., sid), i = 1, 2,..., N (N is the population size), and its performance is evaluated on the 101

G.E. STAVROULAKIS, N. KAMINAKIS, Y. MARINAKIS, M. MARINAKI AND N. SKAROS

predefined fitness function. Thus, each particle is randomly placed in the d-dimensional space as a candidate solution. The velocity of the i-th particle vi = (vi1, vi2 ,..., vid) is defined as the change of its position. The flying direction of each particle is the dynamical interaction of individual and social flying experience. The algorithm completes the optimization through following the personal best solution of each particle and the global best value of the whole swarm. Each particle adjusts its trajectory toward its own previous best position and the previous best position attained by any particle of the swarm, namely pid and pgd. The velocities and positions of particles are updated using the following formulas:

where β∈ (0,∞) is the scale factor. The upper bound of β is usually the value 1 because as it has been proved if the β >1 there is no improvement in the solutions [7] and the most usually utilized value is β = 0.5. The base vector xi1(t) can be determined in a variety of ways and the two most known ways are by selecting a random member of the population or by selecting the best member of the population. The differences vector xi2 and xi3 are selected usually at random. Except of the classic mutation operator (Equation 3), there are a number of different mutation operators that have been proposed [7]. Here we have used a mutation operator in which the parent is mutated using two different vectors:

vid (t + 1) = vid (t ) + c1rand1( pid − sid (t)) + c2rand2 ( pgd − sid (t)) (1)

ui (t) = xi1(t) + β(xopt(t) − xi1(t))+ β∑(xi2,l (t) − xi3,l (t)) (4)

sid (t + 1) = sid (t ) + vid (t + 1)

(2)

where t is the iteration counter; c1 and c2 are the acceleration coefficients; rand1, rand2 are two random numbers in [0, 1]. The acceleration coefficients c1 and c2 control how far a particle will move in a single iteration. Typically, these are both set equal to a value of 2.0, although assigning different values to c1 and c2 sometimes leads to improved performance. 2.2 Differential Evolution Differential Evolution (DE) is a stochastic, population-based algorithm that was proposed by Storn and Price [5]. Recent books for the Differential Evolution can be found in [4],[7]. Differential Evolution has the basic characteristic of the evolutionary algorithms as it is an evolutionary algorithm. However, it has a number of differences compared to them, like the fact that this method focuses in the distance and the direction information of the other solutions. In the evolutionary algorithms, if a crossover operator is used, it is applied initially and, then the generated offspring are mutated. Mutation operators are sampled from some probability distribution function. There are two basic differences in the differential evolution algorithms: 1. mutation is applied first to generate a trial vector, which is then used within the crossover operator to produce one offspring, and 2. mutation step sizes are not sampled from an a priori known probability distribution function but they are influenced by differences between individuals of the current population. The mutation operator produces a trial vector for each individual of the current population by mutating a target vector with a weighted differential. This trial vector will, then, be used by the crossover operator to produce offspring. For each parent, xi(t), the trial vector, ui(t), is generated as follows: a target vector, xi1(t), is selected from the population, such that i≠i1. Then, two individuals, xi2 and xi3, are selected randomly from the population such that i≠i1≠i2≠i3. Using these individuals, the trial vector is calculated by perturbing the target vector as follows:

ui (t ) = xi1 (t ) + β ( xi 2 (t ) − xi 3 (t ))

MS’08 Jordan

(3)

2

l=1

After the completion of the mutation phase of the algorithm a crossover operator is, usually, applied. We used here binomial crossover where the points are selected randomly for the trial vector and for the parent. Initially, a crossover operator number (Cr) is selected [7] that controls the fraction of parameters that are selected from the trial vector. The Cr value is compared with the output of a random number generator, randi(0,1). If the random number is less or equal to the Cr the corresponding value is inherited from the trial vector, otherwise it is selected from the parent: ui (t ), if rand i( 0,1 ) ≤ Cr (5) xi ' ( t ) =   xi (t ), otherwise Thus, the choice of the Cr is very significant because if the value is close or equal to 1, then, the most of the values in the offspring are inherited from the trial vector (the mutant) but if the value is close to 0, then, the values are inherited from the parent [7]. After the crossover operator, the fitness function of the offspring xi'(t) is calculated and if it is better than the fitness function of the parent, then, the trial vector is selected for the next generation, otherwise the parent survives for at least one more generation.

3.

Topology Design of Compliant Mechanisms

3.1 Formulation of the Problem One of the current research areas in topology optimization is the design of compliant mechanisms. A compliant mechanism is a single-piece flexible structure with mobility of a conventional mechanism and stiffness of a conventional structure. A compliant mechanism utilizes the structural deformation, induced by an input actuation, to transmit force or deliver motion. Due to the absence of joints, a compliant mechanism can also be seen as a structure that is stiff enough to bear loads. In other words a compliant mechanism is a combination of a structure and a mechanism, since the jointless feature resembles a structure, while the function of the structure resembles a mechanism. It is a single piece flexible structure that gains its mobility from structural deformation due to an input actuation. In order to find the best structure that utilizes the above demands (flexibility and stiffness), topology optimization is used.

102

STRUCTURAL AND MECHANISM DESIGN USING TOPOLOGY OPTIMIZATION

Figure 2. A horizontal force inverter. Topology optimization of a compliant mechanism can be performed based on continuum as well as truss and frame discretizations (cf. Figure 1 for the optimal structure). In this paper we will focus on the continuum discretization. As an example of a compliant mechanism design problem we consider the displacement inverter in Figure 2. The goal of the topology optimization problem is to design a structure that converts an input displacement at point (1) to the opposite direction at point (2) minimizing at the same time the vertical displacement at point 2 (Uout@Y). We define as positive X direction rightwards and positive direction downwards. Although for real applications of compliant mechanisms the usage of large deformation theory is necessary, the basic concepts as well as (in most cases) a useful topology can be extracted from the linear theory which is used in this paper. We assume that the input actuator is modeled by a spring with stiffness kin and a force fin. The goal of the optimization problem is to maximize the displacement uout at point 2, performed on the workpiece modeled by a spring with stiffness kout. By specifying different values of kout we can control the displacement amplification. If we specify a low value of kout we get large displacement and vice versa. In order to maximize the work on the output spring, the available material must be distributed in the structurally most efficient way. An optimization problem incorporating these ideas can be written as M.P. Bendsoe and O. Sigmund suggests [8]:

Figure 3. Different inital and final material distributions indicating multiplicity. First row the loading case of Figure 2. Second row a different, vertical load case (cf. Figure 4). The “99 line topology optimization code” [8] was converted according to [9] in order to simulate the topology optimization of a compliant mechanism [10]. Different final material distributions occurred from different initial points, and this justifies the usage of global optimization techniques like Evolutionary strategies and Particle Swarms. 3.3. Extension for Multiple Loading Cases We assume that two load cases are applied. The objective is to find the material distribution that satisfies two different load cases. This is shown in the following figure.

max{uout } ρ

s.t. : Ku = f in

Figure 4. Two loading cases (the horizontal one of Figure 2 and a vertical one).

N

∑ρ v

e e

≤ Vmax

e =1

0 < ρ min ≤ ρ e ≤ 1 Where uout is the displacement at the output point (2), ρe is the density of the element e and ve is the volume of the element.

{

x y max u 1 out + u 2 out

ρ

The goal is to maximize the displacement at point (2) in the horizontal direction and at the same time minimize the displacement at the vertical direction. The problem is modified as follows:

s .t . : K ( ρ ) u 1 = F 1 in and K ( ρ ) u 2 = F 2 in N



ρ e v e ≤ V max

  

0 < ρ min ≤ ρ e ≤ 1 ve = 1

The design intent is to maximize displacement in the directions where the different load are applied, and minimize the displacement in the perpendicular directions:

{ } { } {u 2 }, min {u 2 }

x y max u 1 out , min u 1 out ,

max

MS’08 Jordan

}

e =1

3.2. Output Control

x  u out max  y  u out

In this case the iterative local search problem is set as follows:

y out

x out

103

G.E. STAVROULAKIS, N. KAMINAKIS, Y. MARINAKIS, M. MARINAKI AND N. SKAROS

The objectives in this case are similar with the case of one load case. x  u 1 out max  y  u 1 out

y   u 2 out  and max  x   u 2 out

  

This is a multiobjective problem which can be transformed to single objective by maximizing the minimum of the two objectives:

 max  min 

x y  u 1 out u 2 out ,  y x  u 1 out u 2 out

    

3.4. The Hybrid Solution Method

population was initialized, between [ρmin ,1], where ρmin =0.001. Due to the large mutation ratio, and the two base vectors that were chosen, the new member might be out of bounds. Every single density of the material distribution that is out of bounds was initialized. This might be called as an extra mutation-chromosome-repair operator. The length of the evaluation was 60 iterations. For the PSO algorithm the following constants have been used: c1=c2=2.0, wmax=0.9, wmin=0.1, 20 particles and 50 iterations.

4.

Numerical Results

4.1 One Loading Case, Differential Evolution

Differential Evolution was used as the base mechanism for the optimization. The “topology optimization for compliant mechanisms” was used to evaluate each population member. By this way every single member that started from a initial “random” material distribution, after the initialization and the crossover & mutation operators. This procedure is explained in Figure 5.

Figure 6. Compliant mechanism for the horizontal force inverter (cf. Figure 2). Final material distribution, initial and deformed displacements demonstrating the function of the mechanism, evolution of the fitness function (DE) and detail. 4.2 One Loading Case, Particles Swarm Optimization

Figure 5. Flow Chart of proposed hybrid algorithm. For all results we used the following DE settings: Number of populations: 50, members per population: 20, Crossover ratio Cr =0.9, and mutation ratio β =1.5. The DE strategy is DE/rand/2/bin, meaning that two random vectors were chosen as base vectors for the mutation, and the crossover operator was binomial. At the beginning of the procedure the MS’08 Jordan

Figure 7. Compliant mechanism for the horizontal force inverter (cf. Figure 2). Final material distribution, initial and deformed displacements demonstrating the function of the mechanism, evolution of the fitness function (PSO) and detail.

104

STRUCTURAL AND MECHANISM DESIGN USING TOPOLOGY OPTIMIZATION

4.3 Two Loading Cases, Differential Evolution

stiffness and the compliance. The two-loading mechanism design is a multiobjective optimization problem as well. The hybrid algorithm proposed and tested in this paper, due to the global optimization nature of its ingredients, addresses some of the arising difficulties. Further research is needed in order to exploit all previously mentioned aspects of the problem. Extension to nonlinear problems and multidisciplinary (multiphysics) applications will make the problem much more difficult.

References

Figure 8. Compliant mechanism for the multifunctional force inverter (cf. Figure 4). Final material distribution, evolution of the fitness function (DE) initial and deformed displacements for load case one and load case two, demonstrating the two different functions of the mechanism. 4.4 Two load cases, Particles Swarms Optimization

[1] J. Kennedy, and R. Eberhart, “Particle swarm optimization,” in Proceedings of 1995 IEEE International Conference on Neural Networks, Vol. 4, 1995, pp. 1942-1948. [2] A. Banks, J. Vincent, and C. Anyakoha, “A review of particle swarm optimization. Part I: background and development,” in Natural Computing, Vol. 6, No. 4, 2007, pp. 467-484. [3] A. Banks, J. Vincent, and C. Anyakoha, “A review of particle swarm optimization. Part II: hybridisation, combinatorial, multicriteria and constrained optimization, and indicative applications,” in Natural Computing, Vol. 7, 2008, pp. 109-124. [4] R. Poli, J. Kennedy, and T. Blackwell, “Particle swarm optimization. An overview,” in Swarm Intelligence, Vol. 1, 2007, pp. 33-57. [5] R. Storn, K. Price, “Differential evolution - A simple and efficient heuristic for global optimization over continuous spaces,” in Journal of Global Optimization, Vol. 11, No. 4, 1997, pp. 341-359. [6] V. Feoktistov, Differential Evolution - In Search of Solutions. Springer, 2006. [7] K.V. Price, R.M. Storn, J.A. Lampinen, Differential Evolution: A Practical Approach to Global Optimization, Springer, Berlin, 2005. [8] M.P. Bendsoe, O. Sigmund, Topology Optimization. Theory, Methods and applications, Springer, Second Edition, 2004, pp. 94-98. [9] O. Sigmund, “A 99 line topology optimization code written in Matlab”, In Struct. Multidisc. Optim., Vol. 21, 2001, pp. 120-127. [10] M.P. Bendsoe, O. Sigmund, Topology Optimization. Theory, Methods and applications, Springer, Second Edition, 2004, pp. 269-270.

Figure 9. Compliant mechanism for the multifunctional force inverter (cf. Figure 4). Final material distribution, evolution of the fitness function (PSO) initial and deformed displacements for load case one and load case two, demonstrating the two different functions of the mechanism.

5.

Conclusions

The topology design problem for compliant mechanisms is a multiobjective optimization problem with two objectives, the

MS’08 Jordan

105

AUTHORS

Fuzzy-logic-based Definition of Chen Model of Plasticity Nataliya Pokorná* and Petr Štemberk * *

Department of Concrete and Masonry Structures Faculty of Civil Engineering Czech Technical University in Prague Thákurova 7, 166 29 Prague 6, Czech Republic [email protected] [email protected]

Abstract: Numerical simulations, which are used in design of civil engineering structures, require realistic material models. Most of the material models are derived theoretically and later verified with experimental data. Due to the mathematical description, which is mostly based on analytical formulae, the material models do not correspond with the real material behavior, which is amplified when a large structure is analyzed. This paper proposes an alternative definition of material models which is based on the fuzzy logic. This approach provides a flexible tool allowing the material model to cover even very detailed features in behavior, which are often neglected in analytically derived models. The proposed method is explained in the example of the Chen model of plasticity of concrete.

Key words: Chen model of plasticity, fuzzy logic, fuzzy sets, material modeling.

1. Introduction In the construction industry, the emphasis is put on reliability. Therefore, any material or a construction technology is verified with numerical simulations at the design stage. In order to obtain acceptable results with the numerical simulations, such as those based on the finite element method, the material models need to describe the material behavior as close to the real behavior as possible. As most of the material models are derived theoretically, the often focus on expressing the general behavior and lack the capability of covering the detailed features, which is mostly caused by the mathematical apparatus commonly applied. However, the detailed features decide the result of the numerical simulation and thus the ultimate reliability of the structure. An example can be represented by the transition from elastic to plastic region in multi-dimensional analyses, when it is necessary to evaluate the exact direction of the plastic damage, but the Chen model of plasticity [1], due to its definition, cannot provide this information for some stress distributions. Recently, several papers were published which used the fuzzy logic, e.g. [2], for definition of the relations among several material parameters, such as the definition of the relation among stress, strain and amount of reinforcing fibers in [3] or the evolution of the stress-strain curve with respect to the progressing hydration of concrete in [4]. In all cases, the fuzzy logic processed only triangular fuzzy sets, which are on the one hand very easy to work with, but on the other hand the triangular fuzzy sets cannot provide a differentiable curve, unless an almost infinite number of the fuzzy sets is used. MS’08 Jordan

This paper presents an alternative approach to obtain a relation among material parameters. The method makes use of various non-linear shapes of the membership functions of fuzzy set and the Sugeno definition of fuzzy relations to provide a differentiable curve. The method is explained in an example of the Chen model of plasticity. It should be also noted that this approach further broadens the range of application of the Chen model of plasticity by allowing to include the effect of hydration on the Chen model of plasticity presented in [5] and to include the effect of possible uncertainties in the material parameters, which was shown in [6].

2. Method based on definition of shapes of membership functions Since 2004, new alternative methods of defining the material models have been published, e.g. [2], [3]. However, all these models use only the fuzzy sets with triangular membership functions, which are very easy for processing. The triangular membership function, on the other hand, cannot describe highly non-linear behavior, unless a large number of fuzzy sets is used. Moreover, the first derivation is impossible to obtain in this way, which in the case of the Chen model of plasticity is essential. 2.1 Definition of shapes of input fuzzy sets The following criteria were selected for selection of the suitable shapes of the fuzzy sets: - the value of the membership function is at the center of the fuzzy set equal to 1, - the value of the membership function for quantities outside the fuzzy set are equal to 0, 106

FUZZY-LOGIC-BASED DEFINITION OF CHEN MODEL OF PLASTICITY

-

the upper and lower bound of the fuzzy set coincides with the center of the neighboring fuzzy set, the shapes of the membership functions at both sides of the maximum are independent of each other, the shape always depends on only one parameter.

ExpL>1

µx

ExpR 0. Therefore based on the assumption of λ2 > 0, equation (9) is expressed as:

T′ = λ2 T

(10)

Rearranging equation (10) yields: 2

T′- λ T = 0 The general solution to equation (11) becomes:

(11)

147

C. P. UKPAKA, S.A. AMADI AND V. G. NNADI

L µ  − max ⋅  C e µ µ   n   

2

T = C1e λ t

(

(12)

Z′ = λ2 Z

−µ

µ L − max2   Cn µmax L   Cn e µ  (26) S L =  e   C2  

2

-µZ′+ λ Z = 0 (14) Therefore, the general solution to equation (14) becomes: λ2 Z µ

Z = C2 e Also, µmax =

(15)

λ2

(16)

Therefore,

λ2

= µmax (17) Substituting equation (12) and (15) into equation (4) yields: λ  − Z  C e µ   2   

Substituting equation (21) into equation (26) yields: µ L − max2   Cn µmax L   Cn e µ  (27) S L =  e   C2 



(

S = C1e

)

 − µmax2 L  e µ     

(

(

S = C1 e − µ max t

)

   

S L = Cn e

(19)

µ L ln S L = − µ max L ⋅ max2 Cn µ 2

(

µ 2 ln

µ  − max (0 )  C e µ   2   

)

Cn = C1C2

µ 2 = − µ 2 max

(20)

Cn C2

ln St = − µ 2 max t 2 Cn

(21)

2.1 For Case 2: at z = L, t= t Therefore substituting the boundary condition in case II into equation (19) becomes:

(

S = C1e

− µ max L

µ  − max t  C e µ   2   

)

(22)

Since,

L t

(23)

Rearranging equation (23) yields:

L

µ

(24)

⇒ L = µt (24a) Substituting equation (24) into equation (22) yields:

MS’08 Jordan

ln C n 2 L SL

(30)

Similarly, in terms of time equation (22) reduces to:

C1 =

t=

SL = − µ 2 max L2 Cn

Therefore,

Therefore,

µ=

(29)

Rearranging equation (29) becomes:

For Case 1: at z = 0, t = 0, S = Cn

Cn = C1e

(28)

2

µ L ln S L = − max 2 Cn µ

Substituting the boundary conditions into equation (19) yields:

− µ max ( 0 )

)

µmax L

(18)

Since λ2 = µmax, therefore equation (18) becomes: µ  − max Z C e µ  2 



Simplifying equation (27) reduces to:

2

− λ2 t

(25)

(13)

Rearranging equation (13) yields:



)

S L = Ci e µmax L

Similarly,

(31)

Biokinetic Model

The role of biodegradation in the chemical evolution of the residual hydrocarbon mixture has given rise to a new trend of technology in the petroleum industry. The fundamental principle is to create conditions under which micro-organism grow and use the petroleum hydrocarbon as substrate. The result of this is the transformation of the residual petroleum hydrocarbons to carbondioxide, biomass, heat released and water. The material balance for the degradation rate of various enzyme kinetic of individual hydrocarbon mixture for the system is given as:

[S1] +[E1]

K1  →[E1S1]

K4

[S2] +[E2]

5  →[E2S2]

[S3] + [E3]

9  →[E3S3]

K K8 K

K1

K3  → [E1] +[P1]

K2

7  → [E2] +[P2]

K

K6

11 K  → [E3] + [P3]

K10

(31)

(32)

(33) 148

DEVELOPMENT OF BIOKINETIC MODEL FOR CRUDE OIL DEGRADATION IN A SIMPLIFIED STREAM SYSTEM

Similarly for the other species we have

[S4] +[E4]

[

]

K13  → E4S4 K16

 → [E4] +[P4] K15

K14

(34)

Since the reaction is proceeding at a uniform rate for each of the stage, the rates of formation and dissociation of the intermediate are equal, although the concentrations of reactants and products are continuously changing. This situation often occurs, particularly when the enzyme concentration is low compared to the concentration of the 3 reactants. Assuming normal first-order reactions with respect to each entity apply, the following equation describes the balance for [E1S1], [E2S2], [E3S3] and [E4S4]. Then for [E1S1] we have;

[E1 S1 ]

= K2[E1][S1] – K1[E1][S1] + K3[E1][P1] – K4[E1S1]

dt = 0

(35)

Since the rate of formation of [E1S1] from product and enzyme is normally negligible, i.e., K3 ≈ 0. Therefore K7, K11 and K15 is also negligible i.e. K7 ≈ K11 ≈ K15 ≈ 0. The overall velocity of the reaction for stage 1 depends on the rate of product formation. V1 =

d [P1 ] dt

= K4[E1S1]

(36)

The following mass balance also applies at any time for stage 1. [E]0 = [E1] + [E1S1] (37) It is difficult to measure the amount of [E1S1] present for use in equation (37); therefore equation (35) and (37) are used to develop a formula in terms of more readily assessed parameters. The following expression for [E1S1] will be found [E1S1] = (38) K 2 [E ]0 [S 1 ] [E ]0 [S 1 ] = K + K4 K 2 [S 1 ] + K 1 + K 4 [S 1 ] + 1 K2 The determination of the individual rate constants involved is difficult and an overall constant, K, known as the Michael’s – Menten (they made the original development) constant or half-velocity constant, is defined as;

K' =

K1 + K 4 K2

(39)

Substituting equation (38) and (39) into equation (36) yields V1 =

K 4 [E ]0 [S1 ] K ' + [S1 ]

(40)

The maximum rate of reaction Vmax will occur when [E1S1] = [E]0. Therefore, V = K4[E]0 = V1max

 V( pae ) max [S1 ]  V1 max [S1 ] or V = (42) pae  I  K I + [S1 ]  K pae + [S1 ] pae

MS’08 Jordan

V2max[S2 ] or KII +[S2 ]

V3 =

V( pme) max[S3] V3max[S3] or V =  III  pae KIII +[S3]  Kpme+[S3]  pme

(44)

V4 =

V ) max[S4] V4max[S4] or Vpae =  ( pve  IV IV K +[S4]  Kpve +[S4] 

(45)

(43)

pve

Also the constants in these equations KIpae, KIIpst, KIIIpme, KIVpve and V(pae)max, V(pst)max,, V(pme)max, V(pve)max are determined by using the experimental data obtained for substrate concentration and specific rate of reaction for single enzyme reaction at the various stages. The constants were determined using Line Weaver Burk

3. Materials and Method 3.1 Part A: Stream System Biodegradation The investigation was conducted in studying the crude oil degradation in a simplified stream system. Crude oil from Niger Delta Nigeria: Ibewa was used for this investigation, while carrying out this investigation precaution was taken to ensure non-disturbance of the stream by any object. The substrate concentration of the individual petroleum hydrocarbon was measured. The investigation was conducted by monitoring the degradation process with respect to distance of spreading and time.

Procedure Sample of the crude oil was collected and transferred to sterilized sample bottles for analysis of individual hydrocarbon composition. The pH meter was standardized with buffer pH7 and checked with buffer pH4. The temperature of the buffer solution was also obtained. The pH values of the various samples were then measured after the standardization. Gas chromatography was used for hydrocarbon analysis. The composition of the crude oil was determined for C6, C7, C8, C9, and C10 only; and the pH at zero point. The crude oil was introduced into the stream and the change in substrate concentration was monitored with respect to distance and time. The substrate concentration on the stream system was measured linearly with time at t to tn. The time interval considered during the investigation was 10sec; to enable one determining the substrate concentration of crude oil composition and the distance attended.

3.2 Part B: Experimental Set-up to investigate crude oil degradation

(41)

Substituting equation (41) into equation (40) yields V1 =

V( pst)max[S2 ] Vpae =  II   Kpst +[S2 ]  pst

V2 =

The various microorganisms isolated, and identified were cultured and inoculated indifferent reactors set-up to determine the biokinetic of each species of microorganism in the degradation of crude oil. The biodegradation of petroleum hydrocarbon in aqueous environment provided with aeration represents suitable aerobic biodegradation process. With the 149

C. P. UKPAKA, S.A. AMADI AND V. G. NNADI

aeration system, adequate mixing is obtained. The composition of substrate and metabolites can be obtained using suitable analytical techniques. Similarly suitable specie of the microbes is required. Major Equipment: Detector (infrared detector), Thermometer, Gas chromatography. Materials: Sample collected and Gas cylinder

3.2.1

Sample Collection

Samples were collected from a stream system in Niger Delta area, Nigeria; with 20litres capacity containers. In studying the degradation kinetics of enzyme reaction of petroleum crude oil at specific concentration. Crude oil mixture collected in an stream system was used for this research, while carrying out this research precaution was taken to ensure no additional effluent was introduced into the system. The substrate concentration of the individual petroleum hydrocarbon was measured.

Procedure The research was conducted by monitoring the degradation process with respect to time. Analysis was conducted on samples for the determination of temperature, pH, hardness, dissolved oxygen and total dissolved solid particles. The cultured microorganisms were introduced into the different experimental set up (labeled A-D). The experimental analyses on the physicochemical properties of the samples were measured. The initial substrate concentrations of the samples were measured, before microbes were introduced into the system. The samples were analysed with time intervals of 5 days. The stream water was used as the reaction medium for the system.

3.2.3

Hydrocarbon Analysis

Sample of the individual mixtures were analyzed using gas chromatography to detect the concentration and the composition of individual hydrocarbon in the sample.

3.2.2 Microbial Culture The crude oil mixture from the stream system was collected from Niger Delta area of Nigeria. The microbes from the crude oil mixture were isolated and identified according to Buchannan and Gibbon [19]. The various microbial cultured were then prepared as recommended by Buchannan and Gibbon (19).

3.2.4

Microbial Analysis

Total microbial counts were measured by a standard plate count technique using difco plate counter agar.

4. Result and Discussion The following results were obtained from the investigation which are presented in Tables and Figures

Table 1: Analysis results on individual hydrocarbon concentration, spreading rate and pH on the stream system

MS’08 Jordan

L2

µ2

Time (sec)

Distance (meter)

Substrate Concentration (mol %)

Specific rate (velocity) m/sec

pH

t 0

L 0

µ -

0

10

50

5.00

250

25.00

7.42

20

130

6.50

16900

42.25

7.29

30

223

7.40

49729

54.76

7.24

40

370

9.25

136900

85.56

7.08

50

496

Cn C6 = 0.051 C7 = 0.260 C8 = 2.00 C9 = 6.13 C10 =11.62 C6 = 0.030 C7 = 0.19 C8 = 1.65 C9 = 5.27 C10 = 11.02 C6 = 0.022 C7 = 0.14 C8 = 1.40 C9 = 4.68 C10 = 10.50 C6 = 0.018 C7 = 0.11 C8 = 1.31 C9 = 4.57 C10 = 9.99 C6 = 0.011 C7 = 0.08 C8 = 1.25 C9 = 4.06 C10 = 9.39 C6 = 0.010 C7 = 0.060 C8 = 1.00 C9 = 3.98 C10 = 9.02

9.92

246016

98.41

7.03

7.48

150

DEVELOPMENT OF BIOKINETIC MODEL FOR CRUDE OIL DEGRADATION IN A SIMPLIFIED STREAM SYSTEM

Table 2: Analysis results on some physicochemical parameters at S/No Parameters 1 pH Hardness Ca2+ 2 2+ mg Dissolved oxygen SO423 NO324 Total dissolved solid Table 3: Analysis results on Microbial concentration S/No microorganisms 1 Pseudomonas aeruginosa 2 Pseudomonas stuturi 3 Pseudomonas mendociar 4 Pseudomonas vesicularis 5 Bacillus Substitis 6 Pseudomonas sp.

MS’08 Jordan

temperature of 18-400C Unit mg/L mg/L mg/L

Values 7.48 138.0 32.25 17.0 8.0 4.7

Bacterial Count (jrs) 2.7E7 2.3E7 2.0E8 1.6E8 3.6E5 4.8E8

151

C. P. UKPAKA, S.A. AMADI AND V. G. NNADI

Table 4:

Analysis Results on Biokinetics Substrate Concentration

Component C6

C7

C8

C9

C10

C11

MS’08 Jordan

Specific Rate

[S]pae (mol%)

[S]pst (mol%)

[S]pme (mol %)

[S]pve (mol %)

1/[S]pme (mol %)-1

1/[S]pst (mol%)-1

1/[S]pme (mol%)-1

1/[S]pve (mol%)-1

Vpae mol/cm3

Vpst mol/cm3

Vpme mol/cm3

Vpre mol/cm3

1/Vpae mol/cm3

1/Vpst mol/cm3

1/Vpme mol/cm3

1/Vpre mol/cm3

0.0691 0.0540 0.0400 0.0300 0.0250 0.2100 0.1900 0.1700 0.155 0.145 2.492 2.150 1.841 1.672 1.623 7.293 7.020 6.697 6.457 6.262 12.300 12.010 11.500 11.160 11.051 13.140 13.050 12.991 12.939 12.880

0.073 0.056 0.048 0.035 0.027 0.211 0.170 0.162 0.150 0.130 2.488 2.270 2.180 2.097 2.034 7.286 7.030 6.820 6.650 6.432 12.303 12.130 11.741 11.404 11.247 13.141 13.081 13.000 12.980 12.951

0.0691 0.051 0.039 0.033 0.028 0.212 0.181 0.163 0.152 0.144 2.493 2.211 1.999 1.799 1.632 7.290 7.030 6.820 6.631 6.395 12.301 12.151 11.790 11.492 11.301 13.142 13.121 13.104 13.094 13.050

0.0685 0.049 0.039 0.030 0.026 0.212 0.170 0.156 0.143 0.133 2.501 2.010 1.680 1.650 1.610 7.285 6.911 6.741 6.581 6.384 12.321 11.690 11.227 10.892 10.770 13.138 12.872 12.010 11.871 11.591

14.472 18.519 25.000 33.333 40.00 4.762 5.263 5.882 6.667 6.897 0.401 0.465 0.543 0.598 0.616 0.137 0.142 0.149 0.155 0.160 0.081 0.083 0.087 0.090 0.090 0.076 0.076 0.077 .077 0.78

13.699 16.949 22.222 28.571 37.037 4.739 5.882 6.178 6.667 7.692 0.402 0.441 0.458 0.478 0.492 0.137 0.140 0.145 0.146 0.155 0.081 0.082 0.0852 0.0877 0.0889 0.076 0.076 0.077 0.077 0.077

14.718 19.608 25.641 30.303 35.714 4.717 5.525 6.135 6.579 6.944 0.401 0.452 0.501 0.556 0.613 0.137 0.142 0.147 0.151 0.156 0.081 0.092 0.0848 0.087 0.088 0.076 0.076 0.076 0.076 0.076

14.599 20.408 25.641 33.333 52.632 4.717 5.882 6.410 6.993 7.518 0.400 0.498 0.595 0.606 0.621 0.137 0.145 0.148 0.152 0.157 0.081 0.0855 0.0891 0.0918 0.0929 0.076 0.078 0.083 0.084 0.086

0.00302 0.00280 0.00200 0.0010

0.0034 0.0028 0.0026 0.0016

0.0036 0.0024 0.0012 0.0010

0.0039 0.0020 0.0018 0.0010

331.13 357.14 500.00 1000.00

294.12 357.14 384.62 625.00

277.78 416.67 833.33 1000.00

256.41 500.00 555.56 1000.00

0.0048 0.0040 0.0030 0.0020

0.0082 0.0040 0.0032 0.0024

0.0062 0.0036 0.0022 0.0016

0.0084 0.0028 0.0026 0.0020

208.33 250.00 333.33 500.00

121.95 250.00 321.50 416.67

161.29 277.78 454.55 625.00

119.05 357.14 384.62 500.00

0.0684 0.0618 0.0338 0.0294

0.0436 0.06100 0.0380 0.0300

0.0564 0.0424 0.0398 0.0334

0.0982 0.066 0.026 0.024

14.62 16.18 29.59 34.01

22.94 25.00 26.32 33.33

17.73 23.58 25.13 29.94

10.18 15.15 38.46 41.67

0.0546 0.0460 0.0358 0.0302

0.0512 0.0420 0.0340 0.0328

0.0520 0.0420 0.0378 0.0356

0.0748 0.034 0.032 0.0300

18.32 21.74 29.93 33.11

19.53 23.81 29.41 30.49

19.23 23.04 26.46 28.09

13.37 29.41 31.25 33.33

0.0601 0.042 0.030 0.025

0.0346 0.0321 0.0287 0.0238

0.0300 0.0297 0.0253 0.0212

0.0512 0.0396 0.0298 0.0242

16.64 23.81 33.33 40.00

28.90 31.15 34.84 42.02

33.33 33.67 39.53 47.17

19.53 25.25 33.56 41.32

-

-

-

-

-

-

-

-

152

DEVELOPMENT OF BIOKINETIC MODEL FOR CRUDE OIL DEGRADATION IN A SIMPLIFIED STREAM SYSTEM

Table 5: Theoretical computation on maximum specific rate of degradation

S/No 1 2 3 4 5 6 7 8

Parameters

Unit

C6 0.0067 0.0067 0.0067 0.0111 0.0609 0.1111 0.1259 0.1452

Mol/cm3 mol/cm3 mol/cm3 mol/cm3 -

V(pae) max V(pst) max V(pme) max V(pve) max KIpae KIIpst KIIIpme KIVpve

Component C8 0.1250 0.1429 0.1333 0.1667 3.8551 2.6587 2.2677 2.4094

C7 0.0167 0.0167 0.0133 0.02 0.3886 0.3022 0.3627 0.4946

0.06

C9 0.1250 0.1429 0.2222 0.1818 7.5117 8.3464 10.7311 17.348

C10 0.0625 0.0714 0.0769 0.1111 8.5085 11.08565 14.6349 20.3252

2.5

C7

2 Specific rate (velocit) (m /s)

S p e c ific r a t e ( v e lo c it ) (m /s )

C8

0.05 0.04 0.03 0.02

1.5

1

0.5

0.01 0

0

0

100

200

300

400

500

0

600

100

200

300

400

500

600

Distance (m )

Distance (m) Figure 2: Substrate concentration ve rsus distance for C7 and C8

Figure 1: Substrate concentration versus distance for C6

120

14 C9

S p e c if ic r a t e ( v e lo c it ) ( m /s )

12

C10

100

10

80

8

60

6

40

4

20

2

0 0

0 0

100

200

300

400

Distance (m)

500

600

50000

100000

150000

200000

250000

300000

L2 Figure 4: µ2 versus L2 for the determination of intercept and slope

Figure 3: Substrate concentration versus distance for C9 and C10

MS’08 Jordan

153

C. P. UKPAKA, S.A. AMADI AND V. G. NNADI

The developed model for the crude oil degradation in a stream system was investigated on the influence of environmental factor. The degradation pattern of this system is slightly different as most of the hydrocarbon concentration increased. The rate of change in substrate concentration varies with distance and is different for different hydrocarbons. The graphs for substrate concentration versus distance for C6, C7, C8, C9, and C10 are presented in Figures 1, 2 and 3. These results were used for determination of the specific rate (velocity). The specific rate with respect to the degree of degradation was determined only for components of C6, C7, C8, C9, and C10. The result presented in Table 1 indicates the behaviour pattern of the Ibewa Crude Oil at specific condition.

L2 as shown in Figure 4, was used in ln Cn the determination of µ2max and . SL

The graph of µ2 versus

The results obtained in Table 4 were used to determine the specific substrate uptake for the individual hydrocarbon. The various values of the specific rate for the components C6 to C11 were obtained as presented in Table 4. The values for C6, C7, C8, C9, C10 and C11 seem to agree with predicted results while C11 is non-degradable. The values of Vmax and K were determined using the most widely and accurate Line Weaver Burk Plot. Moreover this method is easy to evaluate visually, Irwin [10] and as showed in equations 42, 43, 44, and 45 for various microorganism. The results presented in Table 4 indicates the behaviour pattern of the biochemical reaction since the specific rate increased with increase in degradation rate and the decrease in rate with time and concentration also depleted. The results of theoretical determined values of the maximum specific rate of degradation for the reaction using various microorganism isolated and identified as presented in Table 3.

5.

petroleum hydrocarbon. For petroleum related contaminants various investigation work have been done in developing models for the biodegradation of petroleum hydrocarbon. In this work, comprehensive models for biodegradation of petroleum hydrocarbon were developed. This includes the various microorganisms which have been formulated for an aerobic degradation of petroleum hydrocarbon mixture in aqueous medium. Various results obtained from the experiment are shown in Table 4. Similarly graph of concentration against time are presented in Table 4. Values obtained from the Table 4 were used in determining the specific rate (slope of curve). Values for the reciprocal of the substrate concentration were plotted with values for the reciprocal of the specific rate for every individual composition of the petroleum hydrocarbon and this lead in the determination of constants (Vmax and K) for the microorganism as showed in Table 5. The values of these constants are as shown in Table 5. The comparison of these constants shown a close range, therefore, any of the four microorganisms can be use in degradating the petroleum hydrocarbon mixture. The practical significance on the application of the specific rate of biodegradation for the reaction using microorganism is showed on this paper. The formulated model for the biodegradation of petroleum hydrocarbon could be applied in solving various environmental and other technical problems in the petroleum and chemical industries. Since in practice, industrial effluents are composed of mixtures of different hydrocarbons. These models could be useful in: 1. 2. 3. 4. 5.

Conclusion

In modeling, the degradation rates of crude oil in simplified stream system were influenced by the degree of environmental factors. The degradation rate of the crude oil was observed to have increase with distance and time. The investigation was carried out to study the effect of substrate concentration with distance and time on stream system. The developed model could be applied in monitoring the rate of degradation of the individual petroleum hydrocarbon, estimating the degree of effect and the affected areas, estimating the spreading rate and residence time for each hydrocarbon component in the stream system. The mathematical model developed in this work is presented in equation (3) and (31); by using method of separation of variable and the necessary boundary conditions. Increase in substrate concentration will affect the velocity of the stream system. Since the community depend on the stream for their source of water for domestic purpose, a high concentration of the hydrocarbon, will be a potential danger to the health of the community and the ecosystem. The model developed from this research work is found useful in the determination of degradation rate of crude oil component. The investigation was carried out to study the importance of various pseudomonas species on the degradation of

MS’08 Jordan

Monitoring of petroleum hydrocarbon based effluent Monitoring of bioremediation of polluted area Estimating the biodegradation period for each specie Design of treatment plant for petroleum based effluent Identification of microbial specie that can be capable of biodegrading the non-degradable components of the petroleum hydrocarbon.

Nomenclature

∂s ∂t

= Change in substrate concentration per unit time

(mol.%/sec)

∂s ∂z

= Change in substrate concentration per unit distance

(mol.%/m) µ = Specific rate or velocity (m/s) µmax = Maximum specific rate or maximum velocity, (m/s) Cn; S = Substrate concentration (mol%) t, T = time (sec) z = distance (z – coordinate) (m) C1, C2 = Constant [S1], [S2], [S3], and [S4] [mol. %] = substrate concentration [E1] [mm/min] = Enzyme concentration (pseudomonas aeruginosa) [E2] [mm/min] = Enzyme concentration (pseudomonas stuturi) [E3] [mm/min] = Enzyme concentrations (pseudomonas mendociar)

154

DEVELOPMENT OF BIOKINETIC MODEL FOR CRUDE OIL DEGRADATION IN A SIMPLIFIED STREAM SYSTEM

[E4] [mm/min] = Enzyme concentration (pseudomonas vesicularis) [P1] --- [Pn] [g/cm3] = products V1(Vpae) = [mol/cm3] = specific rate of reaction for single enzyme reaction (pseudomonas aeruginosa)

[7]

James, E. B. and David, F. O. (1977). Biochemical Engineering Fundamentals. McGraw – Hill Book Company Publishers, New York, London, pp. 488 – 489.

V2 (Vpst)[mol/cm3] = specific rate of reaction for single enzyme[8] reaction (pseudomonas stutzeri)

Dasappa S.M., Loehr R.C., (1991). Toxicity reduction in contaminated soil bioremediation processes. Water research oxford, 25:9, pp. 11211130.

V3 (Vpme)[mol/cm3]= specific rate of reaction for single enzyme reaction (pseudomonas mendociar) [9]

Scow, K.M., Merica, R.R., Alexander, M., (1990). Kinetic Analysis of enhanced biodegradation of carbofuran. Journal of Agricultural and food Chemistry, 38, 3. Pp. 908.-912

[10]

1Ong, S.K. and Bowers, A.R. (1990). Steady State analysis for Biological Treatment of inhibitory Substrate. Journal of Environmental Engineering 116 (6), 101-108.

[11]

Stewart, J.E.; Finnerty, W.R.; Kallio R.E. and Stevenson D.P. (1960). Esters from Bacterial Oxidation of Olefins. Journal of Science 132, pp. 1254-1255.

V4max( V(pre)max ) [mol/cm3] =maximum specific rate of reaction for single enzyme reaction (pseudomonas vesicularis)

[12]

Evans, W.C. (1963). The Microbiological Degradation of Aromatic Compounds. Journal of Gen. Microbial 32, pp.177-184.

[E]0 [mol/cm3] = total concentration of enzyme KIpaeKIIpstKIIIpme and KIVpre [dimensionless] = equilibrium constant of the various substrate, pseudomonas specie can biodegradate the bonny light crude oil.

[13]

Parekh, B.R.; R.W. Traxler and Sabek, J.M. (1977). n-Alkane Oxidation Enzymes of a pseudomonas. Appl. Environment Microbial 33, pp.881-884.

[14]

Odu, C.T.I. (1982). Microbiology of Soils Contaminated with petroleum Hydrocarbon. Extent off Contamination and Some Soil and Microbial properties after Contamination. Journal of Inst. Petroleum 58, pp.201-207.

[15]

Livingston, A.G. and Chase, H.A. (1989). Modelling phenol Degradation in a Fluidized-bed Bioreactor. AICHE Journal 13 (12), pp.190-192.

[16]

Odudu, W. O. (1992). Environmental Pollution and Property Values. Paper presented during NIESV Conference at Hotel Presidential, Port Harcourt.

Abowei, M.F.A. and Wami, B.N. (1988). Mathematical Modelling of Dissolution rate of Crude oil Particles in Water. Journal of Modelling, Simulation and control. Vol. 13, No. 3, pp. 1-7.

[17]

Abowei, M. F. N. and Suss, A. A. (1989). Crude Oil Diffusion in Spherical Polar-Co-ordinate Water System. Journal of Modeling, Simulation and Control, 26 (1). pp. 17 – 23.

Irwin H. Segel (1975). Enzyme Kinetics: Behaviour and analysis of rapid equilibrium and steady state enzyme systems. Wiley Interscience Pub., pp. 18240.

[18]

Bernard Alkinson and Ferda Mavintuna, (1983). Biochemical Engineering and Biotechnology Hand Book. The Nature Press, pp. 205-305.

[19]

Buchannan, R.E. and Gibbons, N.E. (eds). (1974) Bergey’s Manual of Determinative Bacteriology (7th ed.). Baltimore. The William and Whisking Co.) p. 274.

V4 (Vpve)[mol/cm3] = specific rate of reaction for single enzyme reaction (pseudomonas vesicularis) V1max (V(pst)max) [mol/cm3] = maximum specific rate reaction for single enzyme reaction (pseudomonas aeruginosa) V2max (V(pst)max) [mol/cm3] = maximum specific rate of reaction for single enzyme reaction (pseudomonas stutzeri) V3max (V(pme)max) [mol/cm3] = maximum specific rate of reaction for single enzyme reaction (pseudomonas mendociar)

References [1]

Gandy, A. F. and Gandy, E. T. (1988). Element of Bio-environmental Engineering. Engineering Press Inc. San Jose Clafil. Pp. 73 – 81.

[2]

Amadi, A. and Antai, S. P. (1991). Degradation of Bonny Medium Crude Oil by Microbial Species Isolated from Ohshika – Oyekama Oil Polluted Area. International Journal of Biochemphysics, Vol.1, pp. 10 – 14.

[3]

[4]

[5]

[6]

Abowei, M. F. N. and Wami, B. N. (1988). Mathematical Modeling of Dissolution Rate of Crude Oil Particles in Water. Journal of Modeling, Simulation and Control. 26 (1), pp. 37 – 51. Ogoni, H. A. (2003). Topic on Biochemical Engineering. Pearl Publishers, Nigeria, pp. 48 – 59.

MS’08 Jordan

155

Y. DJABALLAH and A. BELGACEM-BOUZIDA

THERMODYNAMIC STUDY OF THE TERNARY SYSTEMS (AL-GA-M) (M=AS, P, SB) Y. Djaballah and A. Belgacem-Bouzida Laboratoire d’Etude Physico-Chimique des Materiaux Physics Department, Faculty of Science, University of Batna 05000, Batna, Algeria [email protected] [email protected] Abstract: Ternary phase diagrams of (Al-Ga-As), (Al-Ga-P) and (Al-Ga-Sb) solid solutions have been investigated. A set of thermodynamic parameters for liquid and compound phases were evaluated by optimizing previous experimental data on phase diagrams and thermochemical proprieties. The binary systems were optimized, using CALPHAD method and the polynomial Redlich-Kister model to describe the excess Gibbs energy of the liquid phase. Together, with published data, the (AlAs GaAs), (AlP-GaP) and (AlSb-GaSb) pseudo-binary systems and liquidus lines of the ternary systems have been calculated. Key words: Gibbs energy - Phase diagrams - Semiconductor - Thermodynamic assessment.

Gallium, Phosphorus and antimony pure elements are taken from the compilation published by Dinsdale [3].

1. Introduction Epitaxial layers of AlxGa1-xAs, AlxGa1-xP and AlxGa1-x Sb solid are used in the production of semiconductor lasers of various types, light emitting diodes and photodetectors. Changes of composition of the ternary solid solution make it possible to control the energy gap width and, consequently, the radiation wavelength. Epitaxial layers of the above solid solutions may be grown from the gaseous as well as the liquid phase. Monocrystalline layers of solid solutions are most widely produced by liquid phase epitaxy [1]. The present work was performed in order to optimize the (AlGa-As), (Al-Ga-P) and (Al-Ga-Sb) ternary alloys. A set of thermodynamic data for liquid and binary compounds were evaluated by optimizing previous experimental data on phase diagrams and thermodynamic proprieties. These ternary systems were calculated using CALPHAD method [2]. The binary systems which were needed for the thermodynamic extrapolation must be assessed initially.

2.2. Liquid solution In general the total Gibbs energy of a phase ϕ is given by the sum of three contributions:

Gϕ = ref Gϕ + id Gϕ + excGϕ The first term,

third term,

Giϕ = a + bT + cTLnT + dT 2 + eT 3 + fT -1 + gT 7 + hT -9

(1)

Gϕ is the so-called excess term. In the binary

Gϕ = x AG Aϕ + xB GBϕ

(3)

Gϕ = RT [ x A ln( x A ) + xB ln( xB )]

(4)

ref

id

Gϕ = x A x B



ν

LϕA ,B ( x A − x B )ν

(5)

ν

Where G Aϕ and GBϕ are the Gibbs energies of the pure elements in the ϕ structure and the

ν

LϕA ,B functions are the

Redlich-Kister [4] coefficients given by: ν

Gibbs energy of the pure elements is referred to the enthalpy of this element in the standard element reference state. The temperature dependence of G(T) is usually expressed as a power series of T :

exc

case (A-B) the three contributions to the Gibbs energy are:

exc

2.1. Unary phase

Gϕ corresponds to the reference Gibbs

energy, the second term, id Gϕ the ideal Gibbs energy and the

2. Thermodynamic models For the calculation of phase equilibria in multicomponent systems, it is necessary to minimize the total Gibbs energy (G) of all the phases that take part in this equilibrium [2]. Thus the thermodynamic study of a system requires a thermodynamic description of each existing phase in this system.

ref

(2)

LϕA ,B = aνϕ + bνϕT

(6)

For the ternary liquid solution the following expression is used to represent the Gibbs energy ref

id

Gϕ = x AG ϕA + xB GBϕ + xC GCϕ

(7)

Gϕ = RT [ x A ln( x A ) + xB ln( xB ) + xC ln( xC )] (8)

The Gibbs energy functions of the Aluminum, Arsenic, MS’08 Jordan

156

THERMODYNAIC STUDY OF THE TERNARY SYSTEMS (Al-Ga-M) (M=As, P, Sb)



the (A-B-C) ternary system . The first three terms in the last equation represents the excess Gibbs energy of the phase from the three boundary binaries using the Muggianu extrapolation [5]. The last term represents the ternary interaction.

range of accuracy associated with this datum. The choice of the weighting factor is completely free. If its value is equal to zero then the datum is ignored in the calculation. In the other cases, the corresponding equation is multiplied by this number. In order to obtain an optimized set of coefficients of the Gibbs energy function, it is desirable to take into account all types of experimental data. The thermodynamic descriptions of the ternary liquid phase was obtained by a combination of the corresponding Gibbs energy functions from the assessments of the binary systems using Muggianu interpolation of binary excess terms. The ternary interaction parameters were fixed to zero as a result of lack of information regarding the solidification behavior and the homogeneity ranges of ternary solution phases. The excess parameters of the pseudo-binary solid solution are obtained by simultaneous optimization with respect to the thermodynamic data of the ternary liquid solution and pseudo-binary phase diagrams.

2.3 Pseudo-binary section

4. Experimental information

The Gibbs energy of the (AlM-GaM) pseudo-binary solid solution is given by:

4.1. Binary systems

exc

Gϕ = x A x B

ν

LϕA ,B ( x A − x B )ν

ν

+ x A xC



ν

LϕA ,C ( x A − xC )ν

(9)

ν

+ x B xC



ν

ϕ

ν

LB ,C ( x B − xC )

ν

+ x A x B xC ( 0LϕA ,B ,C x A + 1LϕA ,B ,C x B + 2LϕA ,B ,C xC )

Where ν LA ,B , ν LA ,C and ν LB ,C are the interaction parameters in the binary (A-B), (A-C) and (B-C) systems respectively, 0 ϕ L A ,B ,C , 1 LϕA ,B ,C and 2 LϕA ,B ,C are the interaction parameters in

S G AlM −GaM = G AlM y AlM + GGaM y GaM + RT [ y AlM Ln( y AlM )

(10)

+ yGaM Ln( yGaM )] + y AlM yGaM LALM −GaM

M = As, P or Sb, Gij is the formation Gibbs energy of the binary compound ij, LAlM-GaM represent the excess parameters and yij the molar fraction of the compound ij. 2.4. Binary compound All intermediate binary phases were treated as stoichiometric compounds. The Gibbs energy per mol of atoms is given by following expression:

G φ = xG Aφ + ( 1 − x )G Bφ + ∆H φf − ∆S φf T

(11)

In the six semiconductor binary systems (Al-As), (Al-P), (AlSb), (Ga-As), (Ga-P) and (Ga-Sb) two type of phase exist: primary solution phases and stoichiometric compounds. The experimental data necessary for the optimization are drawn from the compilation of Massalski [7]. In all assessment find in the literature [8]-[11] of these systems the enthalpy and the entropy of formation of the binary compounds are considered as temperature independent. In the current research we take into account this temperature dependency. The (Al-Ga) system was thermodynamically modelled by several authors [12]-[14] using the Redlich-Kister polynomial model. A comparison between the results of those authors shows a general agreement of the liquid phase description. Therefore, the present work used thermodynamic description of Watson [14].

The enthalpy and entropy of formation ( ∆H φf and ∆S φf ) of

4.2. Ternary systems

the φ phase were considered as temperature independent optimizing parameters. But in order to study the influence of the temperature on the enthalpy and entropy, Gibbs energy of formation can be expressed by the following equation:

4.2.1. (Al-Ga-As) system The (Al-Ga-As) phase diagram exhibits a solid solution AlxGa1-xAs. Foster et al. [15] determined this pseudo-binary section in the composition range 0.2-0.92. A complete phase diagram of this system has not been experimentally investigated. Determination of the liquidus and solidus has been made in the Ga rich corner only [16],[17]. Muszynski [1] used a simplex lattice method to determine the liquidus surface of this ternary system.

G φ = xG Aφ + ( 1 − x )G Bφ + [ Aφf + B φf T + C φf TLn( T )] (12) The enthalpy and entropy of formation are given by: H φ = Aφf − C φf T (13)

S φ = − B φf − C φf [ 1 + ln( T )]

(14)

3. Optimization procedure The optimization was made using a computer program, which allows the simultaneous consideration of various types of thermodynamic data (enthalpy of formation, partial enthalpies, Gibbs energy, etc) and phase diagram information (temperature of liquidus and solidus) [6]. The program works by minimizing an error sum where each piece of information used is given a certain weight according to the experimental MS’08 Jordan

4.2.2. (Al-Ga-P) system The (Al-Ga-P) phase diagram has not experimentally determined in sufficient detail, it has only studied at low temperature region below 1373 K [18], [19]. No thermodynamic data of this system has been found in the literature. In the database of Ishida et al. [20], calculation of the (Al-Ga-P) phase diagram is performed using literatures assessment of the binary and pseudo-binary systems. The Gibbs energy of the solid solution in (AlSb-GaP) section is modelled as ideal solution. 4.2.3. (Al-Ga-Sb) system In the ternary (Al-Ga-Sb) system, there is complete 157

Y. DJABALLAH and A. BELGACEM-BOUZIDA

2000

1600 1400

GaP

1200 1000

852.3 K 800 600 400

302 K

200

5. Results of optimization

0.0

0.1

0.2

0.3

0.4

Ga

2200

2063 K

X0.5 P

0.6

0.7

0.8

0.9

1.0

P

(d) 1500 1400

1332.9 K

1300

2000

1200 1100 1000

929.5 K

896 K

900

AlSb

TEMPERATURE (K)

The optimization of the thermodynamic parameters was carried out using our computer program. We first optimized the (Al-As), (Al-P), (Al-Sb), (Ga-As), (Ga-P) and (Ga-Sb) binary systems by using different experimental data. Optimized parameters of the excess Gibbs energy of binary liquid phases and the formations Gibbs energy of the binary compounds AlAs, AlP, AlSb, GaAs, GaP and GaSb are presented in table 1. From these results we calculated the corresponding binary phase diagrams (Fig. 1).

800 700 600 500

1800

400 0.0

1600 1400

0.1

0.2

0.3

1200

0.5

0.6

0.7

0.8

0.9

XSb

1.0

Sb

(e)

1089.5 K

1000

0.4

Al

AlAs

TEMPERATURE (K)

1768.4 K

1800

TEMPERATURE (K)

miscibility in the liquid phase as well as between AlSb and GaSb compounds. A pseudo-binary diagram is formed between AlSb and GaSb. (AlxGal-x)Sb plus liquid equilibrium phase boundaries and tie lines have been measured [21]-[26] The AlSb-GaSb pseudo-binary phase diagram has been determined by several authors [27]-[29]. Girard et. al. [30] have determined the enthalpy of mixing in ternary Al-Ga-Sb alloys at various compositions and temperatures, while, Gerdes and Predel [31] have measured the enthalpy of mixing of AlSb-GaSb liquid compositions at 1345K;

933.4 K 1100

0.1

0.2

0.3

0.4

0.5

Al

0.6

0.7

0.8

0.9

1.0

As

XAs

(a) 1600

1514 K 1400

980.1 K

1000

864.3 K

900 800 700

GaSb

0.0

TEMPERATURE (K)

800

600 500

302 K 1072 K

1000

300

GaAs

TEMPERATURE (K)

400 1200

800

200 0.0

Ga

0.1

0.2

0.3

0.4

X0.5 Sb

0.6

0.7

0.8

0.9

1.0

Sb

600

(f)

400

302.5 K

200 0.0

Ga

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

XAs

(b) Error! Not a valid link. (c)

1.0

As

Figure 1. Calculated binary phase diagrams a) (Al-As) b) (Ga-As) c) (Al-P) d) (Ga-P) e) (Al-Sb) f) (Ga-Sb) Figures 2 to 4 gives the (AlAs-GaAs), (AlP-GaP) and (AlSbGaSb) pseudo-binary sections respectively calculated from this optimization. A comparison between the liquidus projections of the (Al-GaAs) system calculated at various temperatures (Fig. 5) and those established by Muszynski [1] (Fig. 6), shows a reasonable agreement between the two diagrams except at 1100 °C and in the As corner, the agreement is not very satisfactory. This is due to the assessment of the (Al-As) binary system where we have take into account a new experimental data for this binary system.

Table 1. Thermodynamic parameters of the binary and MS’08 Jordan

158

THERMODYNAIC STUDY OF THE TERNARY SYSTEMS (Al-Ga-M) (M=As, P, Sb)

(Al-As) (Ga-As)

(Al-P)

(Ga-P)

(Al-Sb)

(Ga-Sb)

(AlAsGaAs) (AlPGaP) (AlSbGaSb)

Liquid

Lliquid =-8672+8.498T 1

GaSb

Lliquid =-10052+15.635T 2 GGaSb=-23152+8.779T–0.100TLn(T)

Solid

Lsolid =5000 0

Solid

2800

Liquid 2855.7 K

2600 2400

Liquid+Solid

2200 2000 1800

1768.4 K

1600

Solid

1400 1200 1000 800 0.0

0.1

0.2

L

0.4

0.5

0.6

0.7

0.8

0.9

1.0

GaP

XGaP

Figure 3. Calculated (AlP-GaP) pseudo-binary section. 1500 1400

Liquid

1300 1332.9 K 1200

Liquid+Solid 1100 1000

980.1 K

900 800

Solid

700 600 500 400 0.0

Solid

solid 0

0.3

AlP

Lsolid =0 0

0.1

0.2

0.3

0.4

AlSb

=10500+25,078T

0.5

0.6

0.7

0.8

0.9

1.0

GaSb

XGaSb

Figure 4. Calculated (AlSb-GaSb) pseudo-binary section.

2200

2063.4 K

2100

TEMPERATURE (K)

3000

TEMPERATURE (K)

system

TEMPERATURE (K)

pseudo-binary systems. Phase Parameters =-15566–33.998T Liquid Lliquid 0 AlAs GAlAs=-34744–12.342T+4.221TLn(T) =-25500–4.235T Liquid Lliquid 0 GaAs GaAs G =-44419+25.621T–2.446TLn(T) Lliquid =30712-14.255T 0 Liquid liquid L0 =-40458+18.087T AlP GAlP=-47419+5.035T Lliquid =-13906+13.514T 0 Liquid liquid L1 =26933-24.918T GaP GGaP=-38334-3.270T+1.597TLn(T) Lliquid =1714–9.916T 0 Liquid liquid L1 =4422–11.951T AlSb GAlSb=-24777–10.002T+2.162TLn(T) Lliquid =-13946+7.470T 0

Liquid

2000 1900

Al

Liquid+Solid

0.0

1800 0.1

1700

1.0 0.9

0.2

1600

0.8

Solid

1500

0.3

1514.8 K

0.7

0.4

1400

0.6

0.5

1300

2000 K

0.6

1200 0.0

AlAs

0.1

0.2

0.3

0.4

0.5

XGaAs

0.6

0.7

0.8

0.9

0.4

1.0

GaAs

0.7

1900 K

Figure 2. Calculated (AlAs-GaAs) pseudo-binary section. 1.0 0.0

Ga

0.2

1700 K

0.9

The (Al-Ga-P) liquidus lines calculated from this optimization at seven values of temperature are presented in figure 7. Figures 8 and 9 compares the liquidus data of Koster and Thoma [26] with our calculated values in the temperature range of 1800 K to 2800 K. The calculated liquidus agree well with these data.

0.3

1800 K

0.8

MS’08 Jordan

0.5

1400 K 0.1

0.2

0.3

0.1

1600 K

1500 K

0.0 0.4

0.5

0.6

0.7

0.8

0.9

1.0

As

Figure 5. Calculated liquidus projection of the (Al-Ga-As) system

159

Y. DJABALLAH and A. BELGACEM-BOUZIDA

Figure 6. liquidus projection of the (Al-Ga-As) system from reference [1]

0.0

P

6. Conclusion

1.0

0.1

0.8

0.3

0.7

0.4

0.6

K 28 00

0.5 0.4

27 00

K

0.6 0.7

0.3

2600 K

0.8

0.2

2400 K

0.9

2200 K

0.1

1800 K

2000 K

1.0

The experimental thermodynamic data of the liquid phases and compounds in the binary and ternary semiconductors systems are very limited from where the need of modelling of these functions. In this present work a thermodynamic data set for the calculation of phase equilibria in the binaries and ternary systems (Al-Ga-As), (Al-Ga-P) and (Al-Ga-Sb) is obtained by optimising the experimental information according to the CALPHAD method. With the thermodynamic description, one can now make various calculations of practical interest.

0.9

0.2

0.5

0.0

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Ga

Al

Figure 7. Calculated liquidus projection of the (Al-Ga-P) system

0.0

Sb

0.1

1.0 0.9

0.2

0.8

0.3

0.7

0.4

0.6

0.5 0.6

0.5

1300 K

0.4

1200 K

0.7

1100 K 0.8

0.3

1000 K

0.9

0.2

900 K

0.1

1.0 0.0

Al

0.0 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Ga

Figure 8. Calculated liquidus projection of the (Al-Ga-Sb) system MS’08 Jordan

Figure 9. liquidus projection of the (Al-Ga-Sb) system from reference [26].

References [1] Z. Muszynski and N. G. Riabcev, J. Crystal Growth 36, 1976, pp. 335-341. [2] M. Hillert, “Phase equilibria, phase diagrams and phase transformations, their thermodynamic basis”. Cambridge University Press, 1998. [3] A. T. Dinsdale, Calphad 15, 1991, pp. 317-425. [4] O. Redlich, A. T. Kister, Indust. Eng. Chem., 40, 1948, pp. 345-352. [5] Y. M. Muggianu, M. Gambino and J. P. Bros, J. Chim. Phys. 72, 1975, pp. 83-90. [6] Y. Djaballah , PHD Thesis, University of Batna, Algeria, 2005. [7] T. B. Massalski, Binary Alloy Phase Diagrams, ASM International, 1990; [8] I. Ansara and D. Dutartre, Calphad, 8, 1984, pp.323342. [9] K. Yamaguchi, K. Itagaki and Y. A. Chang, Calphad, 20, 1996, pp. 439-446. [10] K. Ishida, T. Shumiya, T. Nomura, H. Ohtani and T. Nishizawa, J. Less-Common Metals 142, 1988, pp. 135144. [11] R. C. Sharma and M. Srivastava, Calphad, 16, 1992, pp. 387-408. [12] I. Ansara, J. P. Bros and C. Girard, CAFE, 2, 1978, pp. 187-192. [13] J. Murray, Bull Alloy Phase Diagrams, 4, 1983, pp. 183. [14] A. Watson A: Calphad, 16, 1992, pp. 207-217.

160

THERMODYNAIC STUDY OF THE TERNARY SYSTEMS (Al-Ga-M) (M=As, P, Sb)

[15] L. M. Foster, J. E. Scardefield and J. F. Woods, J. Electrochem. Sot. 119, 1972, pp. 765-771. [16] M. B. Panish and S. Sumski, J. Phys. Chem. Solids, 30, 1969, pp. 129. [17] W. G. Rado, W. L. Johnson and R. L. Crawley, Solid State Science and Technol. 119, 1972, pp. 652-661. [18] A Tanaka, T. Sugiura, and T. Sukegawa, J. Crystal Growth, 60, 1982, pp. 120-122. [19] M. Ilegems and M. B. Panish J. Crystal Growth, 20, 1973, pp. 77-81. [20] K. Ishida, H. Tokumaga, H. Ohtani and T. Nichizawa, Journal of Crystal Growth, 98, 1989, pp. 140-147. [21] K. Y. Cheng and G. L. Pearson, J. Electrochem. Soc., 124, 1977, pp. 753-759. [22] S. M. Bedair, J. Electrochem. Soc., 122, 1975,pp. 11501158. [23] C. Girard, J. M. Miane, J. Riou, R. Baret and J. P. Bros, Ber. Bunsenges. Phys. Chem., 92, 1988, pp. 132-139.

MS’08 Jordan

[24] K. Osamura, K. Nakajima and Y. Murakami, J. Electrochem. Soc., 126, 1979, pp. 1992. [25] A. N. V. Mau, C. Ante, G. Bougnot, J. Crystal Growth, 36, 1976, pp. 273-277. [26] V. W. Koster W. and B. Thoma, Z. Metallkunde, 46, 1955, pp. 291. [27] J. L. Murray, Bull. Alloy Phase Diagrams, 4, 1983. [28] A. S. Borschevski, I. I. Burdiyan, E. Yu. Lubenskaya and E. V. Sokolova, Zh. Nauch. Khim., 4, 1959, pp. 2824. [29] J. F. Miller, H. L. Goering and R. C. Himes, J. Electrochem. Soc., 107, 1960, pp. 527. [30] C. Girard, J. M. Miane, J. Riou, R. Baret and J. P. Bros, J. Less Common Metals, 128, 1987, pp.101-115 . [31] F. Gerdes and B. Predel, J. Less Common Metals, 64, 1979, pp. 285-294.

161

H. HAMADACHE, A. AMIRAT AND K.CHAOUI

TWO DIMENSIONAL FATIGUE CRACK GROWTH MODEL OF ROLLER BURNISHED 2024-T351 ALUMINUM ALLOY H. Hamadache*, A. Amirat ** and K. Chaoui* * Mechanics of Materials and Plant Maintenance Research Laboratory (LR3MI), Mechanical Department, Faculty of Engineering,Badji Mokhtar University, Annaba 23000, Algeria [email protected] [email protected] **

Mechanics of Materials and Plant Maintenance Research Laboratory (LR3MI) [email protected]

Abstract: Fatigue occurs as a result of fluctuating loading and may lead to different failure mechanism for given structure depending on the characteristics of loading, the characteristics of the material and the significance of the crack for the structure seen as a system. However, manufacturing process of aircraft structures calls upon mechanical surface treatments (MST) such as roller burnishing. This process is based on the elastic-plastic cold working in the near-surface. Results show that in initiating phase, the grain boundaries constitute obstacles for cracks whose size is twice lower grain 3 size. Indeed, crack can stop during approximately 10 cycles. These blockings of cracks cause falls with zero crackgrowth rate. For a remote uniform-tension stress aiming a given fatigue life, crack will cross the second grain boundary after approximately 110000 cycles for the unburnished specimen whereas it is necessary for him up to 120000 cycles for the nd roller burnished state. Beyond the 2 grain boundary (crack propagating phase) and until fracture, roller burnishing reduce the number of cycles from approximately 18%. General two-dimensional model to predict the fatigue crack growth for the special case of semi elliptical surface cracks at hole is presented. Fatigue tests have been run on in service representative case to validate the approach. Key words: 2024T351Al alloy, Crack Initiation, Crack propagation, Fatigue, Roller burnishing, Stress Intensity Factor

1. Introduction In spite of their relatively low volume compared to the core of the matter, the surface layers by the aspect which they present play a fundamental role to characterize the total resistance of mechanical structures. Because of their quasi permanent exposure to the aggression of the environment and because of the mobility of dislocations which they generate, the surface layers are the first to undergo the various mechanical requests. Thus, it appeared that an improper surface quality as well physical as geometrical of these layers can be responsible for approximately 85% of failure damage of parts. The most frequent and most serious incident likely to occur to a mechanical structure is its failures as result of fatigue. This phenomenon constitutes alone 80% of these failures [1], [2]. In several fields of the mechanical engineering industry, the best results for the machinery parts at the risk of fatigue damage are obtained by carrying out a treatment which combines a good surface quality to a deformation strengthening associated at residual compressive stresses. These results are often found by means of mechanical surface treatments (MST) such as roller burnishing [3],[4]. This process is based on the elastic-plastic cold working in the near-surface. Applying burnishing optimum parameters, some works showed that in MS’08 Jordan

addition to its positive effect on the surface roughness, the roller burnishing treatment work hardened surface layers [5], and induces residual compressive stresses in the surface region [6]. According to material and roller burnishing conditions, the roller burnished surfaces provide amongst other things a higher resistance to fatigue crack initiation and/or propagation [7]-[9]. Some works showed improvement of fatigue strength from 110% to 300% [10]. However, for aeronautic structural components, some technological constraints such as holes or quite simply the associated incidence of chromic anodic oxidation (CAO) [11] made that the treatment does not affect the totality of the structure. Fatigue failures usually occurs from the initiation and propagation phases of cracks [12]-[14] where the crack front can be propagate from a semi elliptical surface crack at center of hole and /or a quarter-elliptical corner crack at hole. The present work is an attempt to study the effect of partially roller burnishing on fatigue life of 2024 Al. A fatigue crack-growth pattern of corner crack at hole is developed in order to distinguish the two main fatigue phases

2. Material and experimental techniques In this work, 2024-T351 aluminium alloy was used as the work-piece. Aluminium alloys are particularly well 162

Two Dimensional Fatigue Crack Growth Model of Roller Burnished 2024 T351 Aluminum Alloy

suited for parts and structures requiring high strengthto- weight ratio and probably the best known materials used extensively in aircraft. A scanning electronic microscope (SEM) analysis revealed a chemical composition listed in table 1.

about 30Hz. Taking into account the treatment applying conditions quoted above and especially along the bore of the hole, it is worth to find the initiation phase from the propagation phase throughout prediction of crack growth and fractured crack fronts (Fig. 2).

Table 1. Chemical composition in % without Al Si Fe Cu 0.5 0.5 3.8÷4.9

Mn 0.3÷0.9

Mg Cr Zn Ti 1.2÷1.8 0.1 0.25 0.15

b)d)

c) a)

In order to characterize material in traction, the testtubes are taken in panels of standard quality according to the longitudinal direction (L). Corresponding tensile properties in this direction are listed in table 2: Table 2 Mechanical characteristics σe0,2 (MPa) σr (MPa) A (%) Direction L 334 ÷ 339 463 ÷ 467 14.5 ÷ 16

Figure 2 Fractography of fractured crack fronts: a) quarter –elliptical corner crack at hole, b) semi –elliptical surface crack at hole

3. Results

CSD Treated zone

TD

L

3.1 Wöhler curves (S-N) Industrial fatigue tests made it possible to plot curves of Wöhler (Fig. 3) and to reproduce the fractography of the final faces after rupture (Fig.2). Two thicknesses (4.5 and 12 mm) were retained. 300 UB - CAO UB + CAO RB (90b, 1p) + CAO RB (90b,3p) + CAO RB(140b, 1p) - CAO RB (1410b, 1p) + CAO

280 260 240 Stress, Smax (MPa)

The microstructure of 2024-T351 Al alloy has been investigated by metallographic etching using Keller reagent on polished specimen. A microscopic examination revealed the average grain size in longitudinal (LD), transversal (TD) and cross section (CSD) directions respectively as follow: dmoy = 0.316 x 0.220 x 0.050 mm. The fatigue test specimens (Fig. 1) have according to case's, undergone a treatment of partial roller burnishing under some conditions of pressure and number of passes.

220 200 180 160 140 120 100 1,0E+04

1,0E+05 1,0E+06 Number of cycles, N (cycles)

300

200±1

3,2 - 0,04

UB + CAO RB (120b,1p) + CAO RB (200b,1p) + CAO RB (200b,3p) + CAO

Stress, Smax. (MPa)

275

=

100±1

=

1,0E+07

250 225 200 175 150 125

4,5±0,1 12±0,1

Figure 1 Specimen for tension fatigue loads test Fatigue tests were performed in axial loading (R = 0.1) under constant stress amplitude load using a servo-hydraulic testing machine. Tests were performed at an ambient temperature. The test frequencies were MS’08 Jordan

100 1,0E+04

1,0E+05 1,0E+06 Number of cycles, N (cycles)

1,0E+07

Figure 3 Wöhler curves for some treatment conditions: a) thin specimen (t=4,5mm), b) thick specimen (t=12mm)-RB=Roller Burnished specimen; UB=Unburnished specimen

The analysis of these curves makes it possible to

163

H. HAMADACHE, A. AMIRAT AND K.CHAOUI

release two types of endurance: One traditional which relates to 107 cycles to failure, the other more specific and relates to 105 cycles (retained fatigue-life for the aircraft manufacturers). It comes out from a first analysis of these curves that whatever is, the surface treatment or the thickness, the endurance with 105 cycles is definitely higher than that with l07 cycles (Fig. 4). 250

NR= 1E5 cycles NR= 1E7 cycles

Thin Specimen: t-4,5 mm 215

213

204

202

195

200

193

175

Stress (MPa)

165 140

150

130

130

125

3passes) specimens lose up to 100000cycles of their fatigue life is a difference in 55 %. Table 3 summarizes fatigue strength aiming 105 cycles to failure and toughness (KIC) for various applying conditions of roller-burnishing. Table 3 Fatigue strength and toughness for various applying conditions of roller-burnishing Fatigue Roller strength Toughness burnishing to 105 KIC [MPa√m] Conditions cycles P (bars) and i σD Loss Loss Loss (passes) KIC (a) KIC (c) [MPa] % % % Thin specimen, t=4,5 mm

100

Unburnishing

213

-

X

62,00

X

25,79

X

50

Unburnishing +

195 -8,5 81,08* +30,77 33,16 +28,57

R.B (140b, 1p) -

215 +0,9 61,22 -1,25 54,02 +10,94

0 UB-CAO

UB+CAO

RB(140b;1p)-CAO RB(90b;1p)+CAO RB(90b;3p)+CAO RB(140b;1p)+CAO

Roller burnishing conditions

a) 200 180 180

Thick specimen: 179 t-12mm

177

R.B (90b, 1p)

176

+

160 135

Stress (MPa)

140

RB (90b, 3p)

132

+

**

**

29,10 +12,83

202 -5,2 44,26 -28,61 31,80 +23,30

116

120

100 100 80 60

R.B (140b, 1p)+

193 -9,4 47,93 -22,69 33,39 +29,46

Thick specimen, t=12 mm + CAO

40

NR =1E5 cycles

20

NR= 1E7 cycles

0

b)

204 -4,2

UB+CAO

RB(120b;1p)+CAO RB(200b;1p)+CAO RB(200b;3p)+CAO Roller burnishing conditions

Figure 4 Effect of surface treatment parameters on fatigue strength: a) thin specimen (t=4,5mm), b) thick specimen (t=12mm) This can be interpreted by the number of cycle which the specimen underwent which is directly related to the applied remote uniform-tension stress. 3.2. Effect of roller burnishing on fatigue strength Figure 4a shows that for the thin specimens, the best treatment favourable to the fatigue endurance remains machining of reference i.e. without roller burnishing. The conditions of rolling selected currently deteriorate the fatigue strength owing to the fact that one notices a fall of the endurance being able to reach up to 9.4 %. For a loading aiming 105 cycles to failure of the unburnished specimens (σmax =195 MPa); one loses up to 30000 cycles of fatigue life which represents 30% (Fig. 3a). However a low improvement of 0,9% is offered by a roller burnishing (140bars-lpasse) The same influence is also observed on thick specimens (Fig. 4b) where the fall of the endurance reaches up to 2.2 %. In this case, Wöhler curve (Fig. 3b) shows that for an applied remote uniform-tension stress σmax = 180 MPa, roller-burnished (200barsMS’08 Jordan

Unburnishing+ 180

X

41,67

X

30,70

X

R.B (120b, 1p)+

179 -0,6 47,13 +13,10 31,89 +3,87

R.B (200b, 1p)+

177 -1,7 55,17 +32,39 33,13 +7,91

R.B (200b, 3p)+

176 -2,2 59,42* +42,59 31,43* +2,37 +

-

With CAO; Without CAO

Obviously for the whole of these conditions, one does not find the beneficial effects of the process since it was recorded losses of fatigue strength from -0.6 to 9.4% compared to the reference state (machining) except for roller burnishing state (140bars, 1passe) without OAC where a modest profit from + 0,9% was recorded. However the computation results show that in the majority of the cases roller burnishined state is tougher compared to the reference state. 3.3. Evolution of crack front The initiation phase can be en general defined as microscopic and macroscopic initiation phase according as the crack size is lower or higher than (1 to 3)d grain size [9]. For a loading aiming 105 cycles to failure, in spite of its blockings, a crack present in an unburnished specimen will cross the 2nd grain boundary after 164

Two Dimensional Fatigue Crack Growth Model of Roller Burnished 2024 T351 Aluminum Alloy

approximately 110 kilocycles whereas it can go up to 120 kilocycles for a roller-burnished specimen (fig. 5). Beyond the 2nd grain boundary and rupture (phase of propagation), one counts approximately 29 kilocycles for the unburnished specimen whereas in the presence of roller burnishing this phase is reduced to 24 kilocycles; that is to say a fall of 18%. With imposed loading (σmax = 270 MPa), the

unburnished state resists tiredness overall better than the burnished state. Fatigue life having fallen of approximately 36 %. In spite of the blocking of the crack at once of the unburnished specimen (fig. 5), one notices that to cross the 2nd grain boundary (initiation phase), it is necessary to consume a number of cycles Na to starting such as Na/NR = 78%.

2,0

1,4

13GP-RB (219MPa) w a

1,2

UB 7NGP-NRB (270MPa)

t

1,6 Crack size c;a (mm)

16GP- RB (270MPa)

a-N (7NGP) c-N (7NGP) a-N (13GP) c-N (13GP) a-N (16GP) c-N (16GP)

1,8

1,0

c

0,8 0,6

2nd GB

0,4 1st GB 0,2 0,0 20000

40000

60000

80000

100000

120000

140000

Number of cycles N (cycles)

Figure 5 Evolution of crack size during fatigue tests: RB = Roller Burnishing; UB = Unburnishing cycles In the presence of roller burnishing, this report/ratio is reduced à.70%. In this case, one loses up to Na 44%.

During the initiation phase of roller burnished specimen crack front evolves with ratio c/a>1. However this ratio progressively decrease for tending towards 1 little before failure i.e. On the other hand, the opposite phenomenon (c/a < 1) was observed on the unburnished specimens where the c/a ratio believes until more than 120 kilocycles after which the crack front is reversed (c/a > 1) during all the remainder of the propagation phase.

Between the crossing of the 2nd grain boundary following transversal direction (TD) and the failure (phase of propagation), the losing number of cycles assigned with this phase is (the difference 11%).The effectiveness of roller burnishing thus is appreciated in this phase. In the case of roller burnishing, increasing stress does not influence the initiation phase. For a loading aiming a given fatigue life of 105 kilocycles), crack front evolution (fig.6) seems to be different between the unburnished specimen and that having undergone the treatment. w

R atio c /a

16GP (270 Mpa)

2,5

7NGP (270 Mpa)

a

3

t

3,5

2

3.4. Crack growth ratio

c

13Gp (219 Mpa)

1,5 1 0,5 0 20000

40000

60000

80000

100000 120000 140000

Number of cycles N [cycles]

For an unburnished specimen, the principal crack is propagated with a crack growth ratio (da/dN) which increases up to 5.10-5 mm/cycle. On arrival to the first grain boundary the crack decreases until approximately 3.10 -6 mm/cycles. Little before the 2nd grain boundary, the crack can more not advance and its crack growth ratio can fall to zero (fig.7). However, beyond a crack size of 0,45mm, the slope of the crack growth curve strongly decreases; it is necessary to judge that roller burnishing delays the propagation phase. The loading aiming 105 cycles to failure of the burnished specimen makes that in spite of the importance of its size inside matter, the crack is propagated more quickly at once than on the surface at hole. In this case, the crack growth ratio (dc/dN) is about 10 -5 to 10 -4 mm/cycles (fig.7).

Figure 6 Evolution of ratio c/a versus number of MS’08 Jordan

165

H. HAMADACHE, A. AMIRAT AND K.CHAOUI

Raju [12]. To take account of the hole, several boundary correction factors in terms of (r/t) an (r/w) have completed the SIF formulations. The equation for the SIF for a quarter-elliptical corner crack in three-dimensional finite bodies subjected to remote tension loads is:

1,E-04 da/dN : 7NGP da/dN : 13GP dc/dN : 13GP da/dN : 16GP

6,E-05

dc/dN : 16GP

a a c r r  K I = σ π a ⋅ F jn  ; ; ; ; ; θ  c t w t w 

4,E-05

(1)

In this, solution gives the SIF at any point along the crack front (0 ≤ θ ≤ 90°).

2,E-05

0,E+00 0

0,2

0,4 0,6 Crack size c; a (mm)

0,8

1

4.2. Fatigue crack-growth pattern For a great number of materials, it was shown that the crack propagation could be predicted by assuming the Paris and Erdogan relationship:

Figure 7 Crack growth ratios versus crack size

However a fall speed is observed for a crack size of c = 0.45 to 0.55 mm probably because the blocking of the crack on the level of the grain boundary after which the crack growth curve presents a less stiff slope.

4. Prediction of fatigue crack – growth 4.1. Stress intensity factor equations in 3D In aircraft structures, fatigue failures usually occur from the initiation and propagation of cracks from notches or defects in the material that are either on the surface or at a corner. The crack configuration considered in this study (Fig. 8a) is assumed to be a quarter elliptical corner crack at hole (Fig. 8b).

da = C ⋅ (∆ K )m e dN

(2)

To account the crack-closure phenomena [14], the crack-growth rates were calculated by assuming the Elber relationship between the crack-growth rate and the effective stress-intensity factor range. Thus, the rates da/dN and dc/dN against ∆Ka and ∆Kc respectively at both points A and C are given by equations (4) and (5): da = C e ⋅ (∆K eff dN

)

dc = Ce ⋅ (∆K eff dN

m a

)

m

c

= C e ⋅ (U dp ⋅ ∆K a )

(3)

= Ce ⋅ (U cp ⋅ ∆K c )

(4)

m

m

m and Ccp = Ce ⋅ (U cp )m , Defining: C dp = Ce ⋅ (U dp ) Equations (3) and (4) are rewritten as:

At point A:

( )

da = C ⋅ ∆K m cp a dN

a)

(5)

2w

At point C:

( )

dc = C ⋅ ∆K m dp c dN

2t

A

θ

a

Crack growth rate [mm/cycle]

dc/dN : 7NGP 8,E-05

b)

2r

c

This two dimensional defects evolving in massive parts represent a very frequent practise case [7]. To predict the crack-propagation and fatigue life from elastic-plastic fracture mechanics, accurate stressintensity factor (S.I.F) solutions are needed for this crack configuration. S.I.F equations for a corner crack located along the bore of a hole in a finite plate are given by Newman & MS’08 Jordan

Ratio of equation (5) to (6) gives the crack front evolution:

C

Figure8 Quarter elliptical corner crack at hole: Configuration, a) photography of crack front, b) geometric model.

(6)

K dc  =  βr ⋅ c da  Ka

  

m

(7)

The material parameters Ccp and Cdp were estimated from crack front [15]. To release a history of cracking, it is enough to integrate the coupled differential equations (6) and (7) then: dc (8a) = C ∫ ∆K m ∫ cp c dN

( )



( )

da = C ∫ ∆K m dp a dN

(8b)

166

Two Dimensional Fatigue Crack Growth Model of Roller Burnished 2024 T351 Aluminum Alloy

cycles was given by: Crack-growth rate will be calculated from these equations. The propagating zone on the surface i.e. the crack extension (cf -c0) from the initial crack length "cO" to final crack length "cf" (failure), is divided into several equal increments “∆c”. The crack-growth rate is calculated from equation (7). A computer program was developed. The crack evolution is calculated from equations (8) using crack length (c), crack depth (a), applied load (σmax) and the crack closure factor ratio (βr). Because “∆c” is known, the number of incremental

∆N =

∆c

(dc dN )

(9)

Figure 9 show the experimental and predicted crack front evolution under various conditions of roller burnishing. The crack-growth pattern established from full stress intensity factor range (∆Keff) deduced from Newman-Raju formula seems valid only for an isotropic body.

Crack (mm)

size

c

Newman !

Crack

size

a

a) Number of Cycles, N (cycles) Crack Growth Rate, (mm/cycles)

However, these results show that the relationship between lifetime and crack size is not unique as it was suggested since the crack-closure ratio (βR) appears to be a function of the remote tension stress (σmax), treatment conditions and some others parameters. In fact the single solution (βR = 0.9) given by Newman seems true only for an applied maximum stress and a given initial crack size. This factor ratio seems specific for each treatment condition (burnishing parameters). Otherwise, when the crack size is lower than the grain size, the predicted growth pattern does not fitted well to the measurements because of different blockings and changing direction of micro crack at grain boundaries. In roller-burnishing state, the predicted growth patterns are in good agreement with the measurements according to βr ratio about 0.82 (Fig. 10a). Plots of crack-growth rate against ∆Keff (Fig. 10b) show that at given stress intensity factor range, crackgrowth rate (da/dN) is faster than that (dc/dN). Throughout cracking one finds that SIF at point A is about 10% higher that at point C. This seems reasonable on the basis of analytical calculations by Newman & Raju.

Crack Size a ; c (mm)

(mm) and predicted crack-front evolution for a corner crack at hole Figure 9 Experimental

b)

Stress Intensity Factor Range, ∆K (MPa.m -1/2)

Figure 10 Fatigue crack growth in roller burnished 2024 T-351AL a) Experimental and predicted fatigue growth, b) Crack-growth rates. MS’08 Jordan

167

H. HAMADACHE, A. AMIRAT AND K.CHAOUI

5. Conclusion Roller burnishing, a mechanical surface treatments (MST) is based on elastic-plastic cold working in the near-surface region that modifies basic properties of 2024-T351 Al industrial. Due to roller burnishing industrial conditions and the associated incidence of a chromic anodic oxidation (OAC), the burnishing is partially applied limiting the beneficial effect of the mechanical treatment and even so can be detrimental. Taking into account the chosen treatment conditions, the two phases were predicted from fatigue crack growth pattern under tension. The crack growth rate was calculated according the stress intensity factor range in 3D cracked bodies and along the crack front as given by Newman-Raju formula. The results are analyzed in term of effective stress intensity factor (∆Keff). They allowed to obtain the cracking history and to find from final crack fronts the associated defect size at initiation. However, these results show that the relationship between lifetime and crack size is not unique as it was suggested since the crack-closure ratio (βR) appears to be a function of the remote tension stress (σmax), treatment conditions and some others parameters.

Acknowledgment The authors would like to thank the steel company Arcelor Mittal Annaba for the furniture of the material. This work was supported by the ministry oh higher education and scientific research of Algeria under the research project J2301/03/52/06bis.

Nomenclature A dmoy t w r Fjn k keff U N NR c a R da/dN Ce, m Ccp Cdp

Ultimate elongation % Average dimensions of grains mm Thickness of cracked specimen mm Width of cracked specimen mm Radius of hole mm Boundary-correction factor on stress intensity Stress intensity factor MPa.m-1/2 Full stress intensity factor MPa.m-1/2 Crack closure factor Number of cycles cycles Number of cycles to failure cycles Length of crack mm Depth of crack mm Stress ratio Crack growth rate mm/cycle Material parameters in Paris relationship Effective crack-growth coefficient in point A Effective crack-growth coefficient in point C

Greek Symbols Tensile yield strength σe0.2 Ultimate tensile strength σr Remote uniform tension stress σ Maximum stress σmax Stress intensity factor range ∆K Parametric polar angle of the ellipse θ σ

Remote uniform tension stress

MS’08 Jordan

MPa MPa MPa MPa MPa.m-1/2

σmin σmax βr ∆Κ θ

Minimum stress Maximum stress Crack closure factor ratio Stress intensity factor range Parametric polar angle of the ellipse

MPa MPa MPa.m

-1/2

References [1] H.P. Lieurade,

“La pratique des essais de fatigue”, Pyc éditions, Paris, 1982. [2] O. Lauvide, “Les fissures sous hautes surveillance, l’usine Nouvelle”, N° 2385, Nov.1992. [3] H. Hamadache, A. Amirat, K. Chaoui, “Effect of diamond ball burnishing on surface characteristics and fatigue strength of XC55 steel“. International Review of Mechanical Engineering, Vol.2, N°. 1, pp. 40-48, January 2008. [4] L. Wagner, “Mechanical surface treatments on Ti, Al, and Mg alloys”, Materials science and Engineering, Vol. A263 pp. 210-216, 1999. [5] P.I. Kudryavstsev, “Surface work-hardening delays fatigue development”, Russian Engineering journal, Vol. LII N°1, pp.61-65, 1983. [6] M.0. Fattouh, M.M. EL Khabeery, “Residual stress distribution in burnishing solution treated and aged 7075 Al alloy”, Int. J. Mach. Tools Manufact. Vol. 29 N°1, pp. 153-160, 1989. [7] H. Hamadache, “Influence du galetage sur la tenue en fatigue d’un alliage d’Al2024-T351”, Thèse de doctorat de l’université de Annaba, 2007. [8] M.H. Faber, “Lecture notes on risk and safety in civil engineering”, Swiss federal of technology, ETHZ Switzerland, 2002. [9] H. Hamadache, A. Amirat, K. Chaoui, “Roller Burnishing effect on fatigue life of 2024-T351 Al alloy. Proceeding of Best off of The International Conference on Modeling and Simulation, MS’07, pp. 63-70, January 2008. [10] J. Lemercier, “Emploi rationnel du galetage, Métaux et déformation”, N° 41, pp.5-18, 1976. [11] H. Hamadache, K. Chaoui, “Effet d’un Galetage Partiel combiné à l’O.A.C sur les fronts de rupture dans un alliage Al 2024”, 1è Séminaire de la Maintenance Industrielle, Souk-Ahras, 18-19 Mai, 2004. [12] J.C. Newman, I.S. Raju, “Stress intensity factor equation for cracks in 3D finites bodies subjected to tension and bending loads”, NASA, Technical Memorandum 85793, April 1984. [13] A.C. Pickard, “Stress intensity factors for cracks with circular and elliptic cracks fronts determined by 3D finite elements methods”, Numerical Methods in fracture mechanics, pp. 559-619, 1980. [14] W. Elber, “The significance of Fatigue Crack Closure. In Damage Tolerance in Aircraft”, ASTM STP 486, pp.230-242, 1977.

MPa 168

Two Dimensional Fatigue Crack Growth Model of Roller Burnished 2024 T351 Aluminum Alloy

[15] A. Clerivet, C. Bathias, “Influence of some mechanical parameters on the crack closure effect in fatigue crack propagation in aluminum alloys”, ASTM STP 982, pp. 583-597, Philadelphia 1988.

MS’08 Jordan

169

FRAIGE, LANGSTON AND AL-KHATIB

Development of Distinct Element Method (DEM) for Modeling Non-spherical Particles Feras Y. Fraige*, Paul A. Langston ** and Laila A. Al-Khatib * *

College of Mining and Environmental Engineering, Al-Hussein Bin Talal University, Jordan, Ma'an, P.O. Box 20.

[email protected] , [email protected] "Corresponding author" [email protected] **

School of Chemical and Environmental Engineering, Nottingham University, University Park, Nottingham, NG7 2RD. UK. [email protected]

Abstract: Granular materials are important in nature and industry. They are present in applications such as pharmaceuticals, chemical and process plants, agriculture, mining, and energy production. Despite their simplicity, granular materials can exhibit behavior different from the other traditional states of matter: solid, liquid, or gas. Hence understanding the dynamics and physics involved is important for improved design and optimization. Distinct Element modeling (DEM) is a powerful technique to model such complex behavior at the particle level. Non-spherical particle modeling is one of the challenges facing the DEM community. Contact detection, “overlap” and direction of contact force calculation are more complex and hence consume more time. This work reviews the development of DEM to model non-spherical particles. The model presented is based on multi-contact principles. It investigates the effect of particle shape on solid packing structure and discharge dynamics from hoppers. It shows the effect of point source vibration in enhancing flow. Key words: Contact detection, Distinct element modeling (DEM), Granular Material, Hopper discharge, Non-spherical particle, Packing structure.

1. Introduction Processing of granular and powder materials is important in many engineering applications. These encompass operations such as storage, conveying, mixing and sizing from small scale pharmaceutical or food processing operations, where composition control may be critical, to large scale minerals industry storage where wall stress and silo-quake may be important. Bulk solids behaviour is generally more unpredictable than for gases and liquids and problems such as unsteady flows often occur in the course of handling and processing. The design of hoppers to achieve a smooth and reliable mass flow rate for a specified material has long been a subject of interest to both researchers and process engineers, such as [1]-[3] and, more recently [4],[5]. Although this has been greatly advanced by the introduction of pre-measuring the various flow properties of the material encountered [6], the determination of a range of flow parameters for a bulk solid can be an expensive exercise [7]. Moreover, the classic shear testers often suggest larger hopper outlets than is actually required [2] and most practical design methods are based on theoretical-empirical approaches. Conventional mass-flow hopper design tends to give tall hoppers with steep sides and large outlets. This gives problems in areas of MS’08 Jordan

limited space and in conditions where small or modest flowrates are required. Vibration is often used as a means of initiating and/or controlling flow. It is relatively inexpensive and can be fitted as a “bolt-on” to existing hoppers. However, the mechanics of vibration are complex and there is much confusion as to how vibration actually works [8]. Indeed, in some circumstances, vibration is used to compact and consolidate materials rather than dilate and induce flow. Some workers see vibration as a means of ensuring flow in situations that are on the limit of conventional flow, and of modest effect has provided the most complete body of work on the use of vibration in hoppers [9],[10]. They developed a modified, vibrating, Jenike shear cell and a Jenike-type method of analysis. He also reported modest improvements in flow with the application of vibration. Matsusaka was able to get cohesive materials to flow through very small, capillary tubes by use of vibration – well beyond the limits of standard design [11]-[13]. Matchett used a continuum approach to model limiting states during the application of vibration to a hopper wall [14]. He assumed a circular arc principal stress orientation, originally proposed by Enstad [2], modified to operate in principal stress space. This provided a rational method for positioning a vibrational device in the hopper and was based upon standard continuum material properties with no need for sophisticated vibrational cells. The model was a 170

DEVELOPMENT OF DISTINCT ELEMENT METHOD (DEM) FOR MODELING NON-SPHERICAL PARTICLES

pseudo-static, limit analysis with no dynamic terms. Other approaches to modelling have developed in recent years. With increasing computer power simulation is becoming important in understanding particulate processing using techniques such as Finite Element or the Discrete Element Method (DEM). Langston et al. [15] used a DEM model of spheres to compare with the continuum model approach [14].

2. Non-spherical particle modelling

El Shourbagy et al. [26] use 2D polygons to investigate the effect of particle shape and friction on bulk stress-strain behaviour. Elongated particle systems showed twice the shear strength of non-elongated particles. It concludes that rounded particles and particles with vanishing Coulomb friction cannot reproduce the behaviour of granular materials. Azema et al. [27] use contact dynamics and implicit numerical integration to model the vibration of irregular polygons in an enclosed vessel. The method used a nonsmooth formulation of mutual exclusion and dry friction instead of a repulsive potential.

2.1 Background Džiugys and Peters [16] and Langston et al. [17] review general methods of modelling non-spherical particles in DEM. In [17],[18] they show how the method of sphere intersection (not union) was used to model sphero-disc flow, and Fraige et al. [19] was used to model cubic shaped particles. These models were closely validated by experiment. This section briefly reviews some methods used to model polygons and polyhedra. In polygons, polyhedra, and irregular shapes in general, the contact detection can be quite cumbersome, because the boundaries for these shapes cannot be represented by a single function as the case with spherical particles. Nezami et al. [20] reviewed the ‘common plane’ (CP) technique for contact detection introduced by Cundall. CP is a plane that bisects the space between the two contacting particles. The particles will be in contact if both intersect with CP. And if neither intersects with the CP then they are not in contact. [20] proposed a ‘Fast Common Plane’ (FCP) approach. They stated that the FCP approach recognises that a common plane has identifying characteristics, which reduce the search space for the CP to 5 candidate planes in 2D, while in 3D, the candidate planes fall within 4 types related to the geometry of the particles and their relative positions. Their numerical experiments revealed that FCP algorithm could be up to 40 times the available search methods for finding the CP in 3D. Kohring et al. [21] simulated non-spherical particles flowing through a hopper. Four flow regimes were applied in this simulation depending upon the rate of particles flowing into the hopper. They used convex polygons to model particle shape to overcome the limitations of spherical particles. The contact point was set between the two intersection points of two overlapping polygons. The force was perpendicular to the contact surface and its magnitude was proportional to the overlap area. Matuttis et al. [22] also used polygons in 2D to model heaps formation and the resulting stresses which were dependent on particle shape. Feng and Owen [23] proposed an energy-based polygon to polygon normal contact model in 2D. They derived a normal contact law for a corner-corner contact, where the overlap is based on area. However, their model does not consider the computation of tangential friction and the difficulties to extend it to 3D. In [24],[25] they have used a combined FEM-DEM model to simulate the gravitational depositions of cubical shapes. Encouraging results were obtained and the authors highlighted some of the variation between simulation and experiment results due to the sensitivity of initial conditions. The potential application to rock blasting is discussed. MS’08 Jordan

2.2 Objective of this paper This paper develops a DEM technique for 2D polygon shaped particles to investigate the potential for hopper vibration to promote and control flow. The method and results are compared with the previous study [15] which used a DEM simulation of spherical particles and compared the results with a continuum Stress Arc Model.

3. Distinct Element Modeling The DEM technique uses an explicit time stepping approach to numerically integrate the motion of each particle from the resulting forces acting on them at each timestep. The particle flow model here follows a fairly standard DEM approach for spheres. Cohesion is also included here. The inter-particle and particle wall contacts are modelled using the spring–dashpot–slider analogy. Contact forces are modelled in the normal and tangential directions with respect to the line connecting the particles centres. For more details, the reader is refereed to [15],[19]. The principal data used in this paper is shown in Table 1.

4. Results and discussions 4.1 Critical orifice size The model is used to estimate the critical orifice size (Bc). It is defined as the minimum orifice size to allow slow flow and below which, blockage and arching is observed. Four different particle sets are tested and their critical orifice sizes are shown in Table 2. It shows that Bc is smallest for circles. For the sets with polygons Bc is smaller for polygons with more vertices, i.e. the ‘rounder’ the particle the easier the flow. Fig. 1 shows the blockage and flow limits for the different particle sets. It illustrates that a small increase in orifice size from Bc allows flow. Note that some of the dimensions in this paper are given in units of D the maximum particle diameter since this is a meaningful scaling parameter in hopper flow. 4.2 Effect of composition of particle shape It is observed that set b, triangles and quadrilaterals, has the poorest flow characteristics. The effect of adding circles on the critical orifice size is shown in Fig. 2 in terms of mass percentage of circular particles. (Note mass proportion is equal to volume and 2D area proportion here.) As the proportion of circular particles increases Bc decreases as expected. This indicates that the addition of circles enhances 171

FRAIGE, LANGSTON AND AL-KHATIB

the flow of polygons reducing the effect of their angularity or ‘blockiness’. The main point to note is that the relationship is linear. Table 1: Principal Data in DEM simulation Hopper half-angle, α (degrees) 30 Orifice size, B (cm & D) Up to 0.018 (18D) Number of particles 500 Particles diameter range, d (cm) 0.009 – 0.01 Particle stiffness, (dyne/cm) 10,000 Particle cohesive stiffness 250 (dyne/cm) Particle cohesive radius 0.01 particle radius Particle coefficient of friction 0.53 Normal damping coefficient (dyne 0.05 s/cm) 3 Particle density (g/cm ) 0.9 Vibrator radius 1D Vibration amplitude, A 0.3D Vibration frequency, f (Hz) 60 1 Time step (µsec) Notes D is maximum particle diameter Particle size random in range shown. For polygons each radial line randomly assigned in same range. Typical colloidal particle force measured with Atomic Force Microscopy [29] is 7.6e-5 dyne for d=8 micron. Assuming surface forces scale as d2 then 0.01cm particle gives 0.011 dyne. This was used to set above cohesion data. Table 2: Critical orifice size determination. Particle set a Circles b Triangles and quadrilaterals c Circles, triangles and quadrilaterals d Pentagons and hexagons D is maximum particle diameter = 0.01 cm

N 500 853 697 609

Bc ( D ) 10 17 14 16

4.3 Hopper flow under gravity and vibration (Vibration case). In order to enhance flow at a smaller orifice size, localised vibrators are attached to the hopper walls at different location. The effect of vibration on flow characteristics of polygonal particle sets are investigated varying vibrator location (height above orifice) h, frequency f and amplitude A. In order to examine the effect of vibration on polygon flow, an orifice size of B = 11 D is chosen. It is smaller than the smallest critical orifice size of particle sets b to d. It is considered as a challenging case. 4.3.1 Effect of vibrator height (h) The general trend of varying vibrator height agrees with the previous study on spheres [15], h is optimal at a few particle diameters above the orifice, too high or too low it is not so effective. For set b, triangles and quadrilaterals (Bc = 17 D) it is optimal at h = 1 D, although it does not result in total discharge. For set c, circles, triangles and quadrilaterals (Bc = 14 D), vibration generally enhances flow. It allows full discharge at most locations, however, at the orifice (h = 0) the percentage of material discharged is small (12%). For set d, pentagons and hexagons (Bc = 16 D), material flow is enhanced but to a lesser extent than set c. MS’08 Jordan

4.3.2 Effect of vibrator frequency (f) The effect of varying vibrator frequency has been monitored at three locations. Halving the frequency has little effect on particle discharge, while doubling the frequency slightly enhances the discharge in some cases. 4.3.3 Effect of vibration amplitude (A) Increasing vibration amplitude has had a more significant effect. Any blockages are more readily disturbed by the increased movement. However, this is a more challenging situation practically. It is observed that for this orifice size, the increased amplitude maximises material discharged. As discussed above, vibrator position is important. For h = 2 D and 5 D full discharge is observed. At h=0, flow is enhanced but partial discharge is recorded (for example 50% of particles discharged at A = 1 D). This again shows that a single source vibration should not be placed at the orifice.

5. Conclusion This paper reviews the development in modeling nonspherical particles used in DEM simulations. It develops a 2D DEM model of convex cohesive polygons. This is used to investigate the effect of shape on flowability and the use of vibration to promote flow in wedge shaped hoppers. The critical orifice size to allow gravity flow as a function of particle shape is determined. The potential benefit of vibration is demonstrated. It is concluded that: 1- The ‘rounder’ the particle the easier the flow. 2- Addition of circular particles promotes flow of polygons in a linear manner. 3- Location of vibrator is important. Optimal location here is generally about 2-3D above the orifice. Lower or higher it will be less effective. This supports the results of a previous continuum analysis [14] and DEM sphere model [15]. 4- Increasing vibration amplitude has a significant effect on flow through a small orifice whereas varying frequency has a negligible effect for the cases here. Increasing amplitude may not be so easy practically. In future work, a 3D DEM model of polyhedra will be developed. It will be used to investigate factors such as bed structure and voidage in flow and vibration. Also, it would be useful to quantify system rheology against particle shape.

References [1] Jenike AW. Quantitative design of mass-flow bins. Powder Technology 1967; 1(4):237-244. [2] Enstad G. On the theory of arching in mass flow hoppers. Chemical Engineering Science 1975; 30:1273-1283. [3] Williams JC. The rate of discharge of coarse granular materials from conical mass flow hoppers. Chemical Engineering Science 1977; 32:247-255. [4] Langston PA, Nikitidis MS, Tuzun U, Heyes DM, Spyrou NM. Microstructural simulation and imaging of granular flows in two- and three-dimensional hoppers. Powder Technology 1997; 94(1):59-72. [5] Kozichi J, Tejchman J. Application of a cellular automation to simulations of granular flow in silos. Granular Matter 2005; 7:45-54.

172

DEVELOPMENT OF DISTINCT ELEMENT METHOD (DEM) FOR MODELING NON-SPHERICAL PARTICLES

[6] Kamath S, Puri VM, Manbeck HB, Hogg R. Flow properties of powders using four testers – Measurement, Comparison and Assessment. Powder Technology 1993; 76(3):277-289. [7] Arnold PC. Flow properties for mass-flow hopper geometry determinations – what range of shear testing is required? Powder Handling & Processing 2003; 15(5):315-317. [8] IMechE conference, Hopper & Silo Discharge: Successful Solutions, IMechE, London, 27 November 1998. [9] Bradley, M., “Strategy for Selecting Solutions”, paper 13, Hopper & Silo Discharge: Successful Solutions, IMechE, London, 27 November 1998. [10] Roberts A.W., Chapter 5, Handbook of Powder Science and Technology, 2nd Ed, M.E.Fayed, L.Otten (Eds), Chapman & Hall, New York, 1997 [11] Matsusaka S., M.Urakawa, H.Masuda, ‘Micro-feeding of a fine powder using a capillary tube with ultrasionic vibration’, Advanced Powder Technology, 1995, 6, no.4, 283-293 [12] Matsusaka, S., K.Yamamoto, H.Hiraoki, ‘Micro-feeding of a fine powder using a vibrating capillary tube’, Advanced Powder Technology, 1996, 7, no.2, 141-151 [13] Matsusaka S., M.Urakawa, M.Furutate, H.Masuda, ‘Micro-feeding of fine powders using a capillary with ultrasonic vibration’, Third World Congress on Particle Technology, paper 343, Brighton, England, June 1998 [14] Matchett AJ. A theoretical model of vibrationally induced flow in conical hopper systems. Chemical Engineering research & Design 2004; 82(A1), 85-98. [15] Langston PA, Matchett AJ, Fraige F, Dodds J. Vibration induced flow in hoppers: Continuum and DEM model approaches, Granular Matter, submitted. [16] Džiugys A, Peters B. An approach to simulate the motion of spherical and non-spherical fuel particles in combustion chambers. Granular Matter 2001; 3(4):231266. [17] Langston P.A, Al-Awamleh M.A, Fraige F.Y, Asmar B.N., 2004. Distinct Element Modelling Of NonSpherical Frictionless Particle Flow, Chem. Eng. Sci., 59, 425-435. [18] Li J, Langston PA, Webb C, Dyakowski T. Flow of sphero-disc particles in rectangular hoppers--a DEM and experimental comparison in 3D. Chemical Engineering Science 2004; 59(24):5917-5929.

MS’08 Jordan

[19] Fraige F.Y., Langston P.A., Chen G.Z., 2008, Distinct Element Modelling of Cubic Particle Packing and Flow. Powder Technology, 186, 224-240. [20] Nezami EG, Hashash YMA, Zhao D, Ghaboussi J. A fast contact detection algorithm for 3-D discrete element method. Computers and Geotechnics 2004; 31(7):575587. [21] Kohring GA, Melin S, Puhl H, Tillemans HJ, Vermohlen W. Computer simulations of critical, non-stationary granular flow through a hopper. Computer Methods in Applied Mechanics and Engineering 1995; 124(3):273281. [22] Matuttis HG, Luding S, Herrmann HJ. Discrete element simulations of dense packings and heaps made of spherical and non-spherical particles. Powder Technology 2000; 109(1-3):278-292. [23] Feng YT, Owen DRJ. A 2D polygon/polygon contact model: algorithmic aspects. Engineering Computations 2004; 21(2/3/4):265-277. [24] Munjiza A, Latham JP. Comparison of experimental and FEM/DEM results for gravitational deposition of identical cubes. Engineering Computations 2004; 21(2/3/4):249-264. [25] Latham JP, Munjiza A. The modelling of particle systems with real shapes; Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 2004; 362(1822): 1953 - 1972. [26] El Shourbagy SAM, Morita S, Matuttis HG, Simulation of the dependence of the bulk-stress-strain relations of granular materials on the particle shape. Journal Of The Physical Society Of Japan 2006; 75 (10): Art. No. 104602 [27] Azema E, Radjai F, Peyroux R, et al., Vibrational dynamics of confined granular materials, PHYSICAL REVIEW E 74 (3): Art. No. 031302 Part 1 SEP 2006 [28] 28. Asmar BN, Langston PA, Matchett AJ, Walters JK. Validation tests on a distinct element model of vibrating cohesive particle systems. Computers & Chemical Engineering 2002; 26(6):785-802. [29] Joseph Anthony S, Hoyle W and Yulong Ding (Eds), Review of Granular Materials – Fundamentals and Applications, 2004, The Royal Society of Chemistry, Cambridge, UK

173

FRAIGE, LANGSTON AND AL-KHATIB

Particle Set

No Flow Condition

Slow Flow Condition

B=9

B = 10

B = 16

B = 17

B = 13

B = 14

(a) circles

(b) triangles & quadrilaterals

(c) circles, triangles & quadrilaterals

(d) pentagons & hexagons

B = 15 B = 16 Fig. 1: Critical orifice size for different particle sets. It shows cases of no flow (arching and blockage on left) and slow flow on right under gravity.

MS’08 Jordan

174

JIEEEC 2005 Fig. 2: Effect of composition on critical orifice size. Composition defined in terms of proportion of mass of circles in circle, triangle & quadrilateral mix. (Triangles & quadrilaterals are in equal number proportion.)

20 18 16

R² = 0.97

14

Bc (D)

12 10 8 6 4 2 0 0

20

40

60

80

100

Composition (%)

(i) Particle set b: Triangles and quadrilaterals

(ii) Particle set c: Circles, triangles and quadrilaterals

(iii) Particle set d: Pentagons and hexagons.

Fig. 3 Comparison of vibration discharge for B = 11D, f = 60Hz. For each set: 1st fig shows snapshot at t=0.11s at start of vibration, no material has discharged; 2nd fig shows t=0.15s, 3rd figs shows when near empty or time when discharge stopped.

MS’08 Jordan

175

ATMOSPHERIC MODELING OF THE STELLAR BINARY SYSTEM 9CYG

Atmospheric Modeling of the Stellar Binary System 9Cyg Al-Wardat M.* and Widyan H. * *

Department of Physics, Al-Hussein Bin Talal University, P.O.Box 20, 71111, Ma'an, Jordan [email protected] [email protected]

Abstract: Aim: To estimate the physical and geometrical parameters of the components of the visually close stellar binary system 9Cyg. Procedures: Depending on the comparison and best fit between the entire observational spectral energy distribution of the system, and the synthetic ones created by atmospheric modeling of the components of the binary system using a grid of Kurucz's blanketed models. a b Results: The parameters of the components of the system were derived as Teff = 4700 ± 150 K , Teff = 10000 ± 500 K , log g a = 2.50 ± 0.15 , log g b = 4.50 ± 0.15 , R a = 22 ± 2R and R b = 3.2 ± 0.2R with K3 spectral type for the primary and A0 for the secondary. Key words: 9Cyg, Atmospheric modeling, Stellar Parameters, Spectrophotometry.

1. Introduction In addition that they play an important role in determining several key stellar parameters, they form more than 50% of the galactic stellar systems. These facts raised up the importance of the stellar binary systems. The case of visually close binary systems (VCBS) is more complicated since they are not resolved as binaries in seeing-limited images, but can be resolved in space based observations, or by using modern techniques of ground based observations, like speckle interferometry and adaptive optics, which have greatly improved the situation. The high resolution imaging techniques are not sufficient to determine the individual physical parameters of the systems' components, that is why we use atmospheric modeling overcome this problem. A solution came up with an accurate determination of the effective temperature, radius, spectral type and luminosity class for each component of such systems. This method (the Complex Method) was successfully applied to two binary systems Cou1289 and Cou1291 [1]. The system 9Cyg has been selected for this study since it fulfills the requirements of the method which are: (1) it has a precise magnitude difference measurements, (2) observational SED covers the optical range and (3) photometrical magnitude measurements. This is in addition to the fact that studies of high-eccentricity binaries with composite spectra are of special interest, since such studies provide information about the internal structure of these stars, whose apsidal motion is due primarily to the mass distribution in the volumes of the stars forming the binary. 9 Cyg represents a good example of this kind of stars, it consists of a G giant and a hot A dwarf [2]. Table 1 contains the SIMBAD data of the system (http://simbad.u-strasbg.fr/simbad/), and Table 2 contains data from Hipparcos and Tycho Catalogues [3].

MS’08 Jordan

Table 1. Data from SIMBAD. α2000 δ2000 Tyc HD Spectral Type

19h 34m 50s.9287 +29o 27' 46''.636 2150-5010-1 184760 A0V

Table 2. Data from Hipparcos and Tycho Catalogues. VJ 5m.39 BT 6 m.085 VT 5 m.469 (B-V) J (Tycho) 0 m.581 Hip Trig. Parallax (mas) 6.06 ± 0.58 Tyc Trig. Parallax (mas) 4.4 ± 2.1

2. Atmospheric modeling The reference observational SED was taken from [4] (Fig. 1), note that there are some strong telluric lines and depressions in the spectrum, especially in the red part of the spectrum (around λ6867Å , λ7200Å, λ7605Å), whichare H2O and O2 lines and depressions. In order to start atmospheric modeling, we need the magnitude difference between the two components of the system as estimated from different kind of observations. We used Sheller et al. measurement [5], as the most accurate speckle-interferometry measurement with the 6-m telescope of the Special Astrophysical Observatory, which gives ∆m = 1m.31 in an 80 Å band at 6560 Å. Other references give ∆mV = 1m.1 using far-ultraviolet IUE spectra (Parsons and Ake [6]), ∆mV = 0m.75 depending on computed photometric model of the binary (Griffin et al. [7]), and ∆mV = 0m.0 near 175

AL-WARDAT M. AND WIDYAN H.

4000 Å as reporte ∆mV d by Martin et al. [8] .

atmospheres of each component using grids of Kurucz’s 1994 blanketed models (ATLAS 9) [10], where we used solar abundance model atmospheres, from which the spectral energy distributions in the continuous spectrum for each component were computed. The total energy flux from a stellar binary system results from the net luminosities of the components a and b located at a distance d from the Earth. So we can write:

3.50E-011

Log

F

(erg/cm

2.s.Å)

3.00E-011 2.50E-011 2.00E-011

Fλ .d 2 = H λa .R a2 + H λb .R b2 ,

1.50E-011

(4)

1.00E-011

from which

5.00E-012 0.00E+000

4000

5000

6000

7000

2 2 R 2   R   Fλ =  a   H λa + H λb .  b ,  d   Ra      

8000

Wavelength (Å)

Figure 1. Spectral energy distributions (SED) of the binary

system 9Cyg, (Al-Wardat 2002b). Using the above value, along with apparent visual magnitude mV = 5m.41 [4] and the standard relation: Fa = 2.512 − ∆ m Fb

(1)

we get mVa = 5m.69 and mVb = 7m.00. Using Hipparcos trigonometric parallax from Table 2 , the absolute magnitudes were calculated as MVa = -0m.39 and MVb = 0m.92, and using Lang 1992 bolometric corrections, the bolometric magnitudes were calculated as Mbol.a = 1m.29 and Mbol.b = 0m.69. Hence, the individual luminosities follow as La = 261 ± 34L and Lb = 42 ± 5L . The preliminary estimations of the individual effective temperatures were done using the empirical Teff-Mbol and Sp– M relations for main sequence stars (Lang 1992 [9]), since Balega et al. [2] show that both components are main sequence stars. So

where

(5)

H λa and H λb ¸ are the fluxes from a unit surface of

the corresponding component. F λ here represents the entire SED of the system. These values were used to build model atmospheres for each component, and to compute the energy distributions in a b the continuous spectrum H λ and H λ ¸ and the total energy flux from the whole star F λ . Many attempts were made to achieve the best fit between the observed flux and the total computed ones using the iteration method of different sets of parameters. Within the criteria of the best fit, which are the maximum values of the fluxes, the inclination of the spectra, and the profiles of the absorption lines, the best fit was found using the following set of parameters (Fig. 2):

Teffa = 4700 ± 150 K ,

Teffb = 10000 ± 500 K ,

log g a = 2.50 ± 0.15 ,

log g b = 4.50 ± 0.15 ,

R a = 22 ± 2R

R b = 3.2 ± 0.2R

,

T effa = 12500K , T effb = 9230K and using equation (2), the luminosities follow as 212 ± 28L and Lb = 91 ± 12L .

The radii were calculated using the following relation:      log  R  = 0.5 log  L L  − 2 log  T T  ,  R      as

(2)

La =

According to these parameters, the spectral types of the components of the system estimated as K3 for the primary cool star and A0 for the secondary hot star.

R a = 3.45R and R b = 2.45R .

In a similar way, the gravity acceleration at the surface of each component were calculated using the following relation: log g = log  M M 

    − 2 log  R R  + 4.43,   

(3)

as log g a = 3.93 , log g a = 4.07 , where the masses have been taken from Lang 1992 as Ma=3.8M and Mb=2.8M . The derived values of the effective temperatures and gravity acceleration allowed us to construction the model

MS’08 Jordan

8. Conclusion Atmospheric modeling of the stellar close binary system 9Cyg was used to estimate the individual parameters of its components. The models were constructed depending on Kurucz’s 1994 blanketed models (ATLAS 9) [10]. The spectral types of the components of the system were concluded to be K3 for the primary cool star and A0 for the secondary hot star, which gives a big difference in the spectral type of both components. This requires more investigations and observations in order to conclude the formation and evolution process of such a system. 176

ATMOSPHERIC MODELING OF THE STELLAR BINARY SYSTEM 9CYG

References 3.50E-011 Entire flux (observational)

Log

F

(erg/cm

2.s.Å)

3.00E-011

Entire flux (theoretical)

2.50E-011 2.00E-011 1.50E-011

Component a

1.00E-011 5.00E-012 0.00E+000

Component b

4000

5000

6000

7000

8000

Wavelength (Å)

Figure 2. Dotted line: observational SED in the continuous spectrum of the system 9Cyg. Solid lines: the entire computed SED of the two components, the computed flux of the primary component with Teff=4700 ± 150K, logg=2.50 ± 0.15, R=22 ± 2R and the computed flux of the secondary component with Teff=10000 ± 500K, logg=4.50 ± 0.15, R=3.2 ± 0.2R .

MS’08 Jordan

[1] M.A. Al-Wardat, AN, 2007, 328, pp. 63-69 [2] Yu. Yu. Balega, V.V.Leushin, M. K. Kuznetsov, and V. Tamazian, Astronomy Reports, 2008, Vol. 52, No. 3, pp. 226–236. [3] ESA: 1997, The Hipparcos and Tycho Catalogues, ESA SP-1200, European Space Agency [4] M.A. Al-Wardat, Bull. Spec. Astrophys. Obs. 53, 2002a pp. 51-58 [5] M. Sheller, I. Balega, Yu. Balega, et al., Pis’ma Astron. Zh. 1998, 24, pp. 337 [Astron. Lett. 1998, 24, pp. 283] [6] S. B. Parsons and T. B. Ake, Astrophys. J., Suppl. Ser. 1998,119, pp. 83-88 [7] R. Griffin, R. Griffin, and D. W. Beggs, Mon. Not. R. Astron. Soc. 1994, 270, pp. 409-414 [8] C. Martin, F. Mignard, W. I. Hartkopf, and H. A. McAlister, Astron. Astrophys., Suppl. Ser. 1998, 133, pp. 149-155 [9] K. R. Lang, 1992, Astrophysical Data: Planets and Stars, Springer-Verlag, New York [10] R. Kurucz, 1994, Solar Abundance Model Atmospheres for 0,1,2,4,8 km/s, Smithsonian Astrophysical Observatory, Cambridge, Mass.

177

A SUGGESTED GENERIC INTELLIGENT TUTORING FRAMEWORK

A Suggested Generic Intelligent Tutoring Framework Fares Fraij* and Victor Winter** *

Department of Information Technology Al-Hussein Bin Talal University P.O.Box 20 Ma’an, Jordan [email protected] **

Department of Computer Science University of Nebraska at Omaha 6001 Dodge Street, PKI 175A Omaha, NE 68182-0116 USA [email protected]

Abstract: This paper presents an intelligent tutoring framework that can be effectively utilized to assist teaching courses and therefore to achieve pedagogical goals. The courses generated using the framework are adaptive, i.e., they adjust their behavior to overcome the individual differences among students. The architecture of the framework provides three modules for an administrator, an instructor and a student. Furthermore, students explore the material of the course through two modes, namely non-interactive and interactive (or adaptive). To achieve the goals of the framework, it is recommended to employ an agile software development process such as extreme programming. Furthermore, the development team of the framework must involve students and therefore proceeds in a user-centered fashion. Key words: Extreme Programming, Interaction, Intelligent Tutoring Systems, User-Centered Approach.

1. Introduction An Intelligent Tutoring System (ITS) is a software application for delivering course material. Its purpose is to improve learning outcomes. To be effective, such systems must be adaptable to individual differences among students. Adaptability can be achieved by empowering an ITS with the ability to adjust its behavior in order to provide the students with a flexible and efficient learning environment. ITS can be either static or dynamic. A static ITS is course specific, i.e., it is designed and implemented to teach a specific course. As a result, the benefits of a static ITS is somewhat limited. In contrast, a dynamic ITS is course independent, i.e., such a system is designed to teach any course. Such ITS are advantageous as they can be loaded with the material of any course and automatically generate a learning environment. In this paper, we will use the term dynamic ITS and intelligent tutoring framework (ITF) interchangeably. In general, an ITS can be employed in blended learning [1] or e-learning [2]. Blended learning is characterized by introducing course material with a mix of traditional face-toface approach and technology [1]. Employing technology may take several forms such as presenting visual activities to elaborate on some concepts of the material using a PowerPoint presentation, a computer, and a data show, or using a software application. The goal of e-Learning is to MS’08 Jordan

provide education at a distance. It is worth mentioning that fully-taught courses through e-Learning may not be accredited in some countries. Although commercial computer supported instruction environments are available such as WebCT [3] and Blackboard [4, 5], these environments are costly to afford. Furthermore, such environments are equipped with many functions that may not be fully utilized by the instructors. Adaptive computer-aided Systems have been investigated in the literature [7, 8, 9]. In [7] an adaptive e-Learning framework is introduced, which dynamically generates suitable courses for each student. In [8], a framework for adaptive e-Learning based on distributed reusable learning activities was investigated by utilizing a Knowledge Tree. In [9] a preliminary assessment of the adequacy of existing eLearning standards for supporting the introduction of adaptation techniques in e-Learning systems was discussed. Interesting shortcomings identified in the available eLearning standards were insufficient coverage of the model of the individual learners and the need to enrich the methods of adaptation of e-Learning systems. This paper presents an intelligent tutoring framework intended to be employed in either blended learning or elearning. The paper is organized as follows. Section 2 introduces the pedagogical underpinnings of the framework. In section 3, the architecture of the framework is discussed. Section 4 introduces the modes of interaction supported by the framework. Section 5 presents the software development 179

FARES FRAIJ & VICTOR WINTER

process to be employed in constructing the framework. Section 6 concludes and discusses future directions.

the users of the system are connected to a central database.

2. Pedagogical Underpinning Introductory courses offered at a university may attract students with a relatively broad range of background. This background diversity presents challenges to instructors. On one hand, effort must be made to provide students with a common foundation upon which the more advanced topics of the course will be built. On the other hand, if too much time is spend on foundational (prerequisite) material, the quality of the course will suffer and the more advanced students will not be sufficiently challenged. A long standing rule-of-thumb in teaching is to address individual differences among students in order to maximize learning opportunities. To achieve this goal, educators try to adapt learning material and curriculum to meet individual differences among the students [10]. Adapting the material to a particular student’s background requires measuring the student’s knowledge of the material. This can be achieved by interacting with the student which provides a studentscentered collaborative environment. Historically, Socratic Method employed questions as a mean to interact with students to achieve educational goals [11]. An evaluation test (or pre-test) [10] represents a convenient mechanism for measuring a student’s grasp of a particular subject matter. Within the framework of a course, such pre-tests are useful for evaluating a student’s knowledge prior to presenting specific material. The benefits of such tests are that they represent foundation of instructional decisions, help an instructor how best to present a given subject matter and influence the nature of interaction between students and the instructor. Pre-tests also represent efficient, yet unbiased means to know what the students are capable of knowing or doing. The instructor can plan more effective material and the students are helped to learn more [12]. It is also crucial for instructors to ensure that students have achieved the learning objectives of a course. In this regard, post-tests or summative evaluation [10] are used to measure what the students have learned after the instruction is completed. Recently, higher levels of interaction can be achieved due to the employment of technology in teaching. It is important to realize that recent advances in ICT has relieved educators concerns such as keeping records of preand post-tests [10] and preparing and conducting comprehensive pre- and post-tests [12]. It is worth mentioning that incorporating visual activities during presenting the course material help in increasing students’ interaction in a teaching session.

3. Framework Architecture Figure 1 presents the architecture of a framework which consists of three modules: an administrator, an instructor and a student. The administrator is responsible for creating usernames and passwords for instructors and is in-charge of the evolvement of the framework. Instructors feed-in the course material and plan the lessons of the course. Finally, students access and interact with the material of the course and submit comments and questions to their instructors. All MS’08 Jordan

Figure 1. Overview of the Framework Architecture 3.1 Administrator Module The administrator of the framework receives feedback from users, studies them, and performs the necessary modifications to the framework. The administrator is also responsible for detecting and fixing programming bugs. Finally, the administrator creates accounts for instructors interested in using the framework to deliver their courses. 3.2 Instructor Module The instructor performs major activities such as preparing the material of a course by dividing it into lessons where each lesson consists of sections [13]. Each section is atomic in the sense that it only represents one concept and is associated with well-defined set of skills a student is required to achieve by the end of the section. The lesson planner is a tool that is used by the instructor to determine the structure and the sequence of the lessons of the course. The test editor module is used by the instructor to construct a database of questions to be used in tests. The result analyzer tool is mainly built to obtain statistics from student’s performance database. The tool considers the students’ profile to study their performance and to identify difficulties faced by students belonging to a variety of majors. 3.3 Student Module Students interact with the framework through the display engine. The students have a username and a password to login in to the engine. Through the engine, the students view and interact with the material of the course and perform the needed exams.

180

A SUGGESTED GENERIC INTELLIGENT TUTORING FRAMEWORK

4. Modes of Interaction with the User The courses developed by the framework can be presented in two modes: non-interactive and interactive. The algorithm of the two modes were thoroughly discussed in [14]. In addition, a course specific implementation of a simplified version of the framework was presented [15]. However, the suggested framework introduced in this paper is course independent and designed to present any course of choice. For completeness, the two modes are briefly discussed in this section.

The team will also have a team leader, an analyst, a designer, a programming specialist, a design specialist, and educational specialist. The framework will be developed using an agile software development process, namely extreme programming. This method is appropriate for interactive software design, has been proven to be successful expediting the time-to-market [20], and reduces risk associated with the conformance between implementation and user requirements.

6. Conclusion and Future Directions 4.1 Non-Interactive Mode This mode presents the course’s material in a predetermined, recommended order similar to the traditional teaching method. The material’s concepts are presented in a menu so that a student can choose the concepts s/he wants to consider. By the end of each concept, a set of exercises is introduced, and finally a post-test is proposed. The grade of the post-test is recorded and the process is repeated until all concepts are covered. If the student performance is good in all the post-tests of the concepts, the student is introduced to a final post-test to measure his/her performance. Otherwise, a menu is formulated and presented to the student containing the concepts s/he faces challenges with again. 4.2 Interactive Mode This mode encourages a student to focus on the concepts s/he does not know and the student is given the opportunity to skip to the concepts s/he is already familiar with. To achieve this goal, the student’s grasp of the course concepts is measured before introducing him/her to the course concepts through a pre-test. Afterward, a menu is formulated and presented to the student. This menu focuses on strengthening the concepts the student found challenging.

5. Development Process In spite of the substantial development costs and development time associated with producing quality eLearning applications [16], a recent study [6] has shown that the development cost of such applications can be dramatically decreased by the availability of well-trained IT staff co-operate to develop in-house software and employ a rapid software development process such as extreme programming [15]. The proposed framework will be developed using Microsoft.NET Framework [17]. More specifically, VisualBasic.NET [18] will be employed as the main software development tool. The reasons for choosing VB.NET is the availability of VB.NET programmers and its adequacy and support for implementing web-based applications. A software development team will be developed using a user-centered approach [19]. This approach directly studies the cognitive, behavioral, and attitudinal characteristics of users. Furthermore, the users’ reactions and performance to scenarios, manuals, simulations & prototypes are observed, recorded and analysed. The user will be a full-time, long-term member of the project team which will be introduced in the next paragraph. MS’08 Jordan

An adaptive computer-supported framework for generating courses has been sketched in this paper. The framework is oriented toward adjusting its behavior to fit student’s needs. The framework also is intended to enable instructors with limited IT skills to develop e-courses. The framework is course independent and is recommended to be employed in teaching courses within a university. When used in conjunction with traditional face-to-face teaching, the framework is expected to improve the learning outcomes as it handles individual differences among students in the classroom.

References [1] The Ohio State University (OSU) “E-Learning Implementation Strategy and Plan,” 2003. http://telr.osu.edu/resources/ITelearning.pdf accessed January, 12 2008. [2] M. Ramshirish and P. Singh, “E-Learning: Tools and Technology,” DRTC Conference on ICT for Digital Learning Environment, Bangalore, India, January, 2006. [3] M. W. Goldberg and S. Salari, “An Update on WebCT (World-Wide-Web Course Tools) A Tool for the Creation of Sophisticated Web-Based Learning Environments,” Proceedings of NAUWeb, June 12-15, 1997, Flagstaff, Arizona. [4] Blackboard Inc., Blackboard Learning System, 2004. [5] Blackboard Inc., Educational Benefits of Online Learning, 2000, http://www.backboard.com. [6] A. AlDmour and F. Fraij, “Developing In-House Software: Seeking for Excellence, Facing Challenging, and Exploiting Opportunities”, ICITE, Paris, France, 2008. [7] M. Sasakura and S. Yamasaki, “A Framework for Adaptive e-Learning Systems in Higher Education with Information Visualization,” 11th International Conference Information Visualization (IV'07), 2007. [8] P. Brusilovsky and H. Nijhavan “A framework for adaptive e-learning based on distributed re-usable learning activities”, Proceedings of World Conference on E-learning, 2002, pp.154–161, 2002. [9] A. Paramythis and S. Loidl-Reisinger “Adaptive learning environments and e-learning standards”, Electronic Journal of eLearning, 2(1), March, 2004. [10] R. M. Thorndike, Measurement and Evaluation in Psychology and Education, 7th Edition, Prentice Hall, 2004. [11] R. E. Lopez-Herrejon, M. Schulman, “Using Interactive Technology in a Short Java Course: An Experience 181

FARES FRAIJ & VICTOR WINTER

Report,” ACM SIGCSE Bulletin, Volume 36, Issue 3, September 2004. [12] J. H. McMillan, Classroom Assessment: Principles and Practice for Effective Instruction, Third Edition, Allyn & Bacon publishing, 2003. [13] S. Ritter, J. Anderson, M. Cytrynowicz, and O. Medvedeva, Authoring Content in the PAT Algebra Tutor, (1998) Authoring Content in the PAT Algebra Tutor. Journal of Interactive Media in Education, 98 (9) [www-jime.open.ac.uk/98/9], Published 8 Oct., 1998. [14] F. Fraij, “Visual and Interactive Computer-Based Software for Teaching a Mathematics Course: An Approach”, Proceedings of the IEEE Conference on AI Tools in Engineering, Pune, India, 2008. [15] F. Fraij and I. Al-awamleh, “SFRL: Intelligent, WebBased Tutoring Software for Teaching a Mathematics

MS’08 Jordan

Course” International Journal of Computing and Information Sciences, submitted for publication. [16] D. Dagger, V. P. Wade, and O. Conlan, Developing Active Learning Experiences for Adaptive Personalised eLearning, Third International Conference on Adaptive Hypermedia and Adaptive Web-Based Systems (AH2004) Proceedings, Eindhoven, The Netherlands (2004) [17] Microsoft.NET Framework: www.microsoft.com/NET/ [18] VisualBasic.NET: msdn.microsoft.com/en-us/vbasic/ [19] H. Sharp, Y. Rogers, and J. Preece, Interaction Design beyond Human-Computer Interaction, Wiley, 2007, 9780-470-01866-8 [20] D. Rogers, D. Lambert, and A. Knemeyer, “The Product Development and Commercialisation Process”, The International Journal of Logistics Management, Vol. 15, No.1, 2004.

182

APPLICATION OF MATLAB WITH SIMULINK FOR TEACHING THE PRINCIPLES OF MODULATION TECHNIQUES

Application of Matlab with Simulink for Teaching the Principles of Modulation Techniques Amer R. Zerek *, Meftah M. Almrabet ** and Hind B. Bouraoui *** *

Associate Professor, Electric and Electronic Engineering Department, Academy of Engineering, Tajura, Libya. **

Email [email protected]

Assistant Professor , Faculty of Engineering 7th April University Alzawia- Libya E-mail [email protected]

***

Assistant Lecturer, Electric and Electronic Engineering Department Academy of Maritime Studies, Tripoli , Libya. Email [email protected]

Abstract: Now days , computer technologies have become very powerful and their software is available for performing many tasks with ease way. Most useful software packages available a round the world in the market are the simulation software such as Matlab with Simulink, EWB, MC3Ss, MathCAD, Multisim, Labview, … etc. Their packages are widely used in universities, Institutes and factories . This paper presents some examples of simulated experiments using Matlab and Simulink V. 7.2 as demonstration to students to make them take hold of the basics of communication systems. The simulated experiments deal with CW modulation techniques to present how signals are transformed by the modulation process in both time and frequency domains. Students can easily learn how to experiment with these software packages in order to learn at their own time and fast. This give helps to visualize the fundaments of communication techniques as well as any other physical system approaches. Key words: Simulation, Matlab with Simulink, Modulation, Communication Systems.

1. Introduction Computers technologies have become very powerful and their software is available in the market for performing numerous tasks conveniently. Some of the most useful software available in the market are the simulation software such as Matlab with Simulink, MathCAD, Electronic Work Bench (EWB) , Microcap ( MS3s) and so on. [3, 5]. Several examples available in most of the text books do not convey some of the important characteristics of complex signals besides being too complex for students with insufficiency in the English language. Another advantage of simulation technique is that it can be used complementary to the hardware laboratory equipments. One of the most and very important principle that students must grasp in different communication areas such as electromagnetic, microwave, antenna, … etc is visualization of the basic information before making further progress. It is very difficult to comprehend the basic physical phenomena by mere derivation or stating of the mathematical equation as most teachers do. Even laboratory equipment may give measurement but do not show physical quantities. Hence, visualization software can be used to make-up for these inherited deficiencies. Software gives students the opportunity to experiment in their own time and according to each one's pace of learning in order to verify the behavior of MS’08 Jordan

the real system i.e. physical system. Simulation techniques can be used in addition to visualization as well as to analyze and observe the behavior of its model. It is very useful as a teaching aid. Simulation techniques have some advantages such as optimization of the design as well as developing and understanding the behavior of complicated physical systems without having to build and test them. To verify simulation work, experimental approach is required to get a clear understanding of the difficulties involved in the measurements and to confirm the results obtained from simulation techniques [3, 4, 5]. This paper is organized as follows. Section 1 presents a general introduction on simulation software In section 2, the concepts of Matlab with Simulink package is introduced. Some examples of Matlab applications in CW modulation are represented in section 3. Finally in section 4 it conclude the achieved simulation results.

2. Matlab With Simulink Package MATLABTM, invented by MathWorks Inc., is an interactive program for high performance numerical computation and data visualization. The abbreviation of MATLABTM comes from MATrix Laboratory, therefore the fundamental structure of MATLABTM is the matrix and the basic type of the data is

183

AMER R. ZEREK, MEFTAH M. ALMRABET

AND HIND B. BOURAOUI

an array, although there are special cases of the basic type of the data such as scalars, vectors, real and complex matrices. MATLABTM has become the technical computing environment of choice for many interested engineers because it is an interactive package that incorporates numeric, symbolic computations and scientific visualization. It allows expressing the entire algorithm in a few lines, and computing the solution with high accuracy in a few minutes on a computer . Then displays the achieved results in friendly format preferred by the programmer. MATLABTM is widely used implementing different analogue and digital systems because it is very easy to use and the commands simple and straight forward. MATLABTM has more flexibility and more control. Also it is compacted but in the same time, it is easy to use. It has functions ready to use in the form of mathematical and matrix operations, colors, graphics, and sound control; also, it is suitable for data acquisition applications, and low-level file I/O Figure (1) illustrated the MATLABTM environment. [1, 2, 5].

…etc), signal routing (Mux, Demux, Switch,… etc) and so on. [1, 2].

Figure 2. Simulink library.

3. Examples of Matlab Applications in CW Modulation The objective of simulation of the CW modulation is to achieve the CW modulated signals in both time and frequency domains. It is divided into two types such as linear and non-linear modulation

Figurre 1. The Matlab command window. MATLABTM is associated with software known as a SIMULINK. In the last few years, SIMULINK has become the most widely used software package in academia and industry for modeling and simulating dynamic systems. SIMULINK encourages any one to try things out. It can easily build models from scratch or take an existing model and add to it. Simulations are interactive, so it can change the simulation parameters on the fly and immediately see what happens. The main purpose of the SIMULINK is to give the user sense of fun of modeling and simulation, through an environment that encourages one to pose a question, model it and see what happens. Figure 2 illustrated the Simulink library which includes block library of sources (Constant, Signal Generator, Ground, Sine wave,…etc), sink (Scope, To File, To Workspace, Out1, …etc), Continuous ( Integrator, Derivative, Transfer Fun., …etc), discrete (Unity Delay, Discrete Time Integrator, Discrete Transfer Function, … etc), Math Operations (Sum, Product, Sign, Gain, Math Function,

3.1 Linear modulation (AM ) system The AM modulated signal can be obtained by multiplying the baseband signal (message signal) with peak amplitude of 0.5volt and modulating frequency 1KHz into sinusoidal of signal (carrier signal) with peak amplitude of 1 volt and carrier frequency of 10KHz, then the product output is added to the carrier signal. The linear modulation system can be achieved by implementing the simulation circuit using Matlab with Simulink as shown in figure (3). The simulation circuit of the AM modulator/demodulator (MODEM) consists of two: Sine Wave (Sources). , sum, Product (Math). To Workspace, Scope (Sinks) and Math Function, (Math Operations) and butterworth low pass filter (Filter Designs)

Figure 3. AM modulator/demodulator simulation circuit.

MS’08 Jordan

184

APPLICATION OF MATLAB WITH SIMULINK FOR TEACHING THE PRINCIPLES OF MODULATION TECHNIQUES

AM MODEM Parameters Sine Wave 1 (0.5 Volt, 2*pi*1000 rad/sec , pi/2 rad) Sine Wave 2 (1 Volt, 2*pi*10000 rad/sec , pi/2 rad) To Workspaces {(Variable names: - mt, vt and vto), (Limit data points to last=2^20) and (format: Array)} Analog Filter Design {Design method (Butterworth), filter type (low pass), filter order (2) and cutoff frequency (3*pi*2650 rad/sec)} Simulation Parameters {Solver (start time 0.0 sec, stop time=0.002sec, Min step size=0.002/2^21 and Max step size=0.002/2^20), Data Import/Export (Time =t and Limit data points to last=2^20)} After implementing the AM modulator circuit and setting all the parameters of the equipment as well as the simulation parameters. The simulation results are obtained in both time and frequency domains. To represent the achieved results sin time and frequency domains, a Matlab program is written in M-file, and through it some parameters are defined, and the necessary commands are written to draw the simulated signals in time and frequency domains as shown in figure (4)

Figure 5. FM modulator/demodulator simulation circuit. FM MODEM Parameters Modulation index, β = 5. Sine Wave 1 1 Volt, 2*pi*1000 rad/sec , pi/2 rad) Sine Wave 2 (10 Volt, 2*pi*10000 rad/sec , pi/2 rad) To Workspaces {(Variable names: - mt, vt and vto), (Limit data points to last=2^20) and (format: Array)} Analog Filter Design {Design method (Butterworth), filter type (low pass), filter order (2) and cutoff frequency (3*pi*2650 rad/sec)} Simulation Parameters it set same as in linear AM system.

10

5

M a g . (v o lt )

A m p . (v o lt )

After implementing the FM modulator circuit and setting all the parameters of the equipment as well as the simulation parameters. The simulation results are obtained in both time and frequency domains. To represent the achieved results of the FM MODEM in time and frequency domains, a Matlab program is written in Mfile, and through it some parameters are defined, and the necessary commands are written to draw the simulated signals in time and frequency domains as shown in figure (6)

0

-10 0

-2

0 Freq. (KHz)

2

3 M a g . (v o lt)

A m p . (v o lt )

2

0

1

-10 0

0.5 1 1.5 Time (msec.)

1

0

5

10 15 Freq. (KHz)

20

0.4

0

-1

0

2

M a g . (v o lt)

3.2 Non-linear modulation (FM ) system The non- linear modulation system can be achieved by implementing the simulation circuit using Matlab with Simulink as shown in figure (5). The simulation circuit of the FM modulator consists of two: Sine Wave (Sources)., sum, two Product, two Trigonometric Function (Math Operations). To Workspaces, Scope (Sinks). In addition, the FM demodulator consists of Math Function, Gain (Math Operations). Derivative (Continuous) and butterworth low pass filter (Filter Designs)

0

2

10

A m p . (v o lt )

Figure 4. Amplitude modulation waveforms in both time and frequency domain: (a) Modulating signal (b) AM Modulated signal (c) Detected signal

0.5 1 1.5 Time (msec.)

0.2

0

0.5 1 1.5 Time (msec.)

2

0

-2

0 Freq. (KHz)

2

Figure 6. Frequency modulation waveforms in both time and frequency domain: (a) Modulating signal (b) AM Modulated signal (c) Detected signal

MS’08 Jordan

185

AMER R. ZEREK, MEFTAH M. ALMRABET

AND HIND B. BOURAOUI

4. Conclusion This paper concludes that the application of Matlab with Simulink for teaching the principles of modulation techniques are not challenging and that simulation technique provides a good approach for learning and making the way of teaching very simple. The modern software can be used to simulate almost any physical system and to experiment it without having to build it.

References [1] R. Brian , L. R. Hunt , L. G. Jonathan and M. Rosenberg, "A Guide to Matlab for beginners and experienced Users", Cambridge University Press. 2001. [2] R. Ziemer, R., and W. H. Tranter, . "Principles of communications, system modulation and noise"., John Wiley & Sons, INC 2002.. [3] M. C. Jeruchim, B. Philip, and K. Sam Shanmuga, "Simulation of communication system", New York , Boston , 2nd Ed. , 2000. [4] M. P. Fitz, "Fundamental of communication systems", McGraw Hill companies, 2007. [5] F. M. Gardner and J. D. Baker, "Simulation Techniques", Wiley, New York , 1997.

MS’08 Jordan

186