Digitally Modulated Signals Identification Based on ...

3 downloads 0 Views 333KB Size Report
Monteporzio Catorne, Roma, Italy. Prof. Ercole M. Gloria. Via Giunta Pisano 2, Pisa, Italy. Prof. Franco Gori. Dip. di Fisica, Università Roma III. Roma, Italy. Prof.
FONDAZIONE GIORGIO RONCHI http://ronchi.isti.cnr.it

SATTAR B. SADKHAN, ASHWAQ Q. HAMEED, HADI A. HAMED

Digitally Modulated Signals Identification Based on Artificial Neural Network

Estratto da: Atti della Fondazione Giorgio Ronchi Anno LXX, n. 1 - Gennaio-Febbraio 2015

Tip. L’Arcobaleno s.n.c. - Via Bolognese, 54 - Firenze 2015

ANNO LXX

GENNAIO-FEBBRAIO 2015

N. 1

AT T I D E L L A « F ONDA Z ION E GI O RG IO RO NCHI» EDITORIAL BOARD Prof. Roberto Buonanno Osservatorio Astronomico di Roma Monteporzio Catorne, Roma, Italy

Prof. Riccardo Pratesi Dipartimento di Fisica Università di Firenze, Sesto Fiorentino, Italy

Prof. Ercole M. Gloria Via Giunta Pisano 2, Pisa, Italy

Prof. Adolfo Pazzagli Clinical Psychology Prof. Emerito Università di Firenze

Prof. Franco Gori Dip. di Fisica, Università Roma III Roma, Italy Prof. Vishal Goyal Department of Computer Science Punjabi University, Patiala, Punjab, India Prof. Enrique Hita Villaverde Departamento de Optica Universidad de Granada, Spain Prof. Irving Kaufman Department of Electrical Engineering Arizona State University, Tucson Arizona, U.S.A. Prof. Franco Lotti I.F.A.C. del CNR, Via Panciatichi 64 Firenze, Italy

Prof. Edoardo Proverbio Istituto di Astronomia e Fisica Superiore Cagliari, Italy Prof. Andrea Romoli Galileo Avionica, Campi Bisenzio Firenze, Italy Prof. Ovidio Salvetti I.ST.I. del CNR Area della Ricerca CNR di Pisa, Pisa, Italy. Prof. Mahipal Singh Deputy Director, CFSL, Sector 36 A Chandigarh, India Prof. Marija Strojnik Centro de Investigaciones en Optica Leon, Gto Mexico

Prof. Tommaso Maccacaro Direttore Osservatorio Astronomico di Brera, Via Brera 28, Milano

Prof. Jean-Luc Tissot ULIS, Veurey Voroize, France

Prof. Manuel Melgosa Departamento de Optica Universidad de Granada, Spain

Prof. Paolo Vanni Professore Emerito di Chimica Medica dell’Università di Firenze

Prof. Alberto Meschiari Scuola Normale Superiore, Pisa, Italy

Prof. Sergio Villani Latvia State University, Riga, Lettonia

Pubblicazione bimestrale - Prof. LAURA RONCHI ABBOZZO Direttore Responsabile La responsabilità per il contenuto degli articoli è unicamente degli Autori Iscriz. nel Reg. stampa del Trib. di Firenze N. 681 - Decreto del Giudice Delegato in data 2-1-1953 Tip. L’Arcobaleno - Via Bolognese, 54 - Firenze - Febbraio 2015

Atti della “Fondazione Giorgio Ronchi”

Anno LXX, 2015 - N. 1

INDEX

Announcement

A. MESCHIARI, Microscopi Amici nella ricerca scientifica (Amici microscopes in scientific research), Fondazione Giorgio Ronchi, Firenze 2014.

Pag. 1

Crystals

K.N. CHOPRA, Technical note on mathematical modeling and short review on the » optics of Uniaxial Crystals

3

Geothermal Energy

K.N. CHOPRA, Optimization of the conversion of the geothermal energy into electricity. A short note

»

17

History of Science

»

27

Letter to the editor

»

35

»

37

»

43

Neural Networks

»

59

Optical Testing

»

77

Statistics

» 103

Thin films

» 113

M.T. MAZZUCATO, Cartesio G. NEBBIA, Luce e Sole

Materials

K. YADAV, S. KUMAR, N. JAGGI, M. GIRI, Preparation and Characterization of Cobalt-doped CdSe Nanoparticles K.N. CHOPRA, Mathematical Modeling and some Novel Studies on Er3+ and Er3+/ Yb3+ co-doped Phosphate Glasses based Fiber Optics Amplifiers - A short Note S.B. SADKHAN, A.Q. HAMEED, H.A. HAMED, Digitally Modulated Signals Identification Based on Artificial Neural Network K.N. CHOPRA, Optical Testing of Optical Elements. Technical Analysis and Overview MEENAKSHI, M.S. SAROA, V. KUMAR, Lacunary Statistical Limit Superior and Limit Inferior on Probabilistic Normed Spaces K.H. ABASS, Effect of the Cobalt Additive on the Optical Properties of CdO Thin Film

Atti della “Fondazione Giorgio Ronchi”

Anno LXX, 2015 - N. 1

NEURAL NETWORKS

Digitally Modulated Signals Identification Based on Artificial Neural Network SATTAR B. SADKHAN (*), ASHWAQ Q. HAMEED (**), HADI A. HAMED (**) SUMMARY. – Modulated signals identification is a technique for classifying the modulation scheme of the modulated signals, possibly noisy signals whose modulation scheme is unknown. Signal identification is divided into two subsystems, namely the feature extraction and the classifier. The challenge of classification is to find a vector of features that distinguishes one type of modulated signal from others, as well as an appropriate neural network architecture that gives the best identification. The classifiers process those features and identify the modulation type. ANN is one of the important soft computing techniques used efficiently in digital modulation identification systems. This paper presents a family of strategies using artificial neural networks (ANN) for automatic modulation identification to observe how the classification performance varied based on number and types of modulated signal schemes, features extraction, neural network architectures, hybrid systems, and learning rules.

Key words: Digital modulation types, Modulation identification, Artificial neural network, neural network learning algorithms, feature extraction, and hybrid classification system.

1. Introduction The process of identifying the modulation format of the modulated signals without prior knowledge of the modulation characteristics is either modulation classification or recognition. Automatic modulation recognition (AMR) or classification (AMC) is an intermediate step between signal detection and demodulation, and play a key role in civilian applications like: spectrum management, interference identification, software defined radio (SDR), and cognitive radio (CR). As well as military applications like: Electronic warfare, Target acquisition, Jamming and Homing.

(*) SMIEEE, University of Babylon, Iraq; e-mail: [email protected] (**) University of Technology, Babylon, Iraq; e.mails: [email protected]; hady. [email protected]

60

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

In the past decades, many AMR algorithms have been developed for various applications. The earliest work, pattern recognition techniques to identify the type of modulation on a high-frequency (HF) signal may be that by Weaver at 1969 (1). Most AMR algorithms are developed after the mid of 1980s, there are two primary methods: decision theoretic (DT) and pattern recognition (PR). The DT approaches use probabilistic and hypothesis arguments to formulate the recognition problem. The major drawback of this approach are the difficulties of forming the right hypothesis and difficulties to set the correct threshold values (2). In the PR approach, the classifier is composed of two subsystems. The first is called the feature extraction subsystem and its role is to extract useful information from the raw data, the second is a pattern recognizer subsystem whose function is to indicate the membership of the modulation type (3). The different modulation types were characterized by extracted features. These features were used by the classifier to identify the modulation types. There are various classifiers of PR such as support vector machines (SVM) classifiers (4), decision-tree classifiers (5), and neural network classifiers (6, 2). In contrast to the DT method, the PR methods are non-optimal, but they are robust and simple to implement, but if the PR approaches are carefully designed, they can achieve nearly optimal performance. ANNs have been successfully applied to a modulation classification. The main characteristics of neural networks are based on the ability to learn complex nonlinear input-output relationships, using different learning rules. In 1998 Nandy and Azzouz in (7) declared that the success rates obtained from the ANN approach are better than those obtained by the DT approach, and the results are very encouraging and point toward the adoption of ANN. The objective of this paper is to present a compact overview of available features and ANN classifiers used in digital AMR. This will assist designers to choose the appropriate algorithm with respect to their intended applications. Furthermore it helps the newcomers to this field to determine the limitations associated with the ANN based digital modulated signal identification. The paper is organized as follows. In sect. 2, digital modulation techniques are presented. In sect. 3, feature extraction based ANN classifiers are presented. In sect. 4, ANN based Identifiers for digital modulated signals are reviewed. Conclusion and future work are presented in sect. 5. 2. Digital Modulation Techniques In modern communication systems, analog modulation techniques, such as AM, FM, and PM, were gradually replaced with digital communications systems. Digital modulation techniques offered several advantages over analog modulation techniques. Some of these advantages are: better spectral efficiency, easier multiplexing, better noise and fade rejecting capability, easier for error detection and correction, better data encryption. Some of the most widely used digital modulation techniques are summarized as follows:

Digitally Modulated Signals Identification Based on Artificial Neural Network

61

2.1 – Amplitude Shift Keying (ASK) Amplitude shift keying ASK transmits digital data by varying the amplitude of the sin wave. Both frequency and phase remain constant while the amplitude changes. The general expression for M-ASK or (M-ary ASK) is: [1]

si (t) =

2Ei (t) cos ( 2π f c t + Φ) ; 0 ≤ t ≤ T; i = 1, 2, 3,..., M T

The parameters Ei and T are the symbols of energy and time duration respectively, the phase F is an arbitrary constant, and fc is the carrier frequency (8, 9). 2.2 – Phase Shift Keying (PSK) Phase shift keying (PSK) transmits data by varying the phase of the carrier wave. The simplest form of PSK is called BPSK. In BPSK system, the phase of the carrier signal is switched between two values a 180° phase shift is usually used. Quadriphase-shift keying (QPSK) uses four phase shifts, QPSK is a special case of MPSK, in MPSK the phase of the carrier takes on one of M possible values, during each signaling interval of duration T, one of the M possible signals [2]

si (t) =

⎡ 2E 2π ⎤ cos ⎢( 2π f c t ) + (i −1) ⎥; ⎣ T M ⎦

i = 1, 2, 3,..., M

Here E is the signal energy per symbol (10). 2.3 – Quadrature Amplitude Modulation (QAM) QAM is a combination of ASK and PSK. It is design to transmit two quadrature carriers cos(2πfct) and sin(2πfct). This structure of QAM allows for M discrete amplitude level (M-QAM). The transmitted QAM signal for symbol k, is defined by: [3]

s k (t) =

2E0 2E0 ak cos ( 2π f c t ) − bk sin ( 2π f c t ) , 0 ≤ t ≤ T, k = 0,±1,± 2,.... T T

The signal sk(t) consists of two phase quadrature carriers with each one being modulated by a set of discrete amplitudes, hence the name quadrature amplitude modulation (10). 2.4 – Frequency Shift Keying (FSK) FSK is depending on how the frequency variations are imparted into the carrier signal, in binary FSK the frequency of the carrier is varied to represent binary 1 or 0, both amplitude and phase remain constant. In MFSK the modulated signal are defined by:

62

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

[4]

si (t) =

⎡ π ⎤ 2E cos ⎢ ( nc + i ) t ⎥; i = 1, 2, 3,..., M ⎣ T ⎦ T

where the carrier frequency fc = nc/2T, nc is some fixed integer (8,10). 2.5 – Minimum Shift Keying (MSK) Minimum shift keying (MSK) is a continuous-phase frequency shift keying (FSK) with a minimum modulation index (h = 0.5) that will produce orthogonal signalling. The MSK signal is: πt πt cosωc − AQ (t)sin sinωc 2T 2T where AI(t) and AQ(t) encode the even and odd information respectively (11). MSK exhibits a constant envelop signal which reduces problems caused by non linear distortion. [5]

s(t) = AI(t)cos

3. Features Extraction used in Identification of Digital Modulated Signals In a typical AMR system, raw datasets are reduced in dimension before they are fed into the recognizer. Such reduced data, often called key features, carry distinctive information about the raw data. The advantage of feature extraction is to allow the work to be done with smaller datasets therefore resources are needed (12). 3.1 – Time Domain Features Various feature extracting techniques exist. There are no exact rules as to which feature should be used. The key features proposed for identification are derived from instantaneous amplitude a(t), the instantaneous phase F(t) and instantaneous frequency f(t). These features are: Maximum value of the spectral power density of the normalized centered instantaneous amplitude: [6]

γ max =

max DFT ( acn (i))

2

Ns

where Ns is the number of samples per segment, and acn(i) is the normalized centered amplitude defined by: acn (i) = an (i) −1 where an(i) = a(i)/ma and ma is the average of the amplitude over one segment. Standard deviation of the absolute value of the centered non linear component of the instantaneous phase:

Digitally Modulated Signals Identification Based on Artificial Neural Network

[7]

⎛ 1 1 2 σ ap = Φ NL (i) − ⎜⎜ ∑ Φ NL (i) ∑ C an (i )> at ⎝ C an (i )> at

⎞ ⎟ ⎟ ⎠

63

2

where FNL(i) is the value of the normalized centered component of the instantaneous phase, C is the number of samples in FNL(i) for which an(i) > at, and at is the threshold for a(i) below which the estimation of instantaneous phase becomes noise sensitive. Standard deviation of the direct instantaneous phase, given as: [8]

⎛ 1 ⎞ 1 2 σ dp = Φ NL (i) − ⎜⎜ ∑ Φ NL (i) ⎟⎟ ∑ C an (i )> at ⎝ C an (i )> at ⎠

2

Standard deviation of the absolute value of the normalized centered instantaneous amplitude: [9]

N ⎛ 1 Ns ⎞ 1 s 2 ⎜⎜ ∑ acn (i) ⎟⎟ σ aa = a (i) − ∑ cn N s i=1 ⎝ N s i=1 ⎠

2

Standard deviation of the absolute value of the normalized centered instantaneous frequency: [10]

⎛ 1 ⎞ 1 σ af = f N2 (i) − ⎜⎜ ∑ f N (i) ⎟⎟ ∑ C an (i )> at ⎝ C an (i )> at ⎠

2

The spectral features mentioned above were used in previous works by Nandi and Azzouz (13, 14). Maximum value of the power spectral density of the normalized centered instantaneous frequency (15): [11]

γ max f = max FFT ( f n (i))

2

Standard deviation of the normalized centered instantaneous amplitude in the nonweak segment of a signal (6,7) [12]

⎛ 1 ⎞ 1 σa = acn2 (i) − ⎜⎜ ∑ acn (i)⎟⎟ ∑ L an (i )> at ⎝ L an (i )> at ⎠

2

where L is the number of samples in acn(i) for which an(i) > at. The kurtosis of the normalized instantaneous amplitude defined by (7): [13]

µ a42 = E { An4 (t)}

2

{E ⎡⎣ A (t)⎤⎦} 2 n

64

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

The kurtosis of the normalized instantaneous frequency defined by (7): f µ 42 = E { f n4 (t)}

[14]

{E ⎡⎣ f

2 n

2

}

(t)⎤⎦

The advantage of time domain features is that they are easy to implement, but the disadvantage is that carrier frequency would need to be estimated in practical applications. 3.2 – Statistical Features 3.2.1 – Higher order Statistics Higher order statistics (HOS) is divided into high order moments (HOMs) and high order cumulants (HOCs) (16): •  High order moments Probability distribution moments is a general concept of the expected value, and can be used to illustrate the probability density function, for a received signal x(n). A particular HOM, Mp+q,p, is defined as: q ⎤ ⎡ M p+q,p = E ⎢x(n) p x(n)* ⎥ ⎣ ⎦

(

[15]

)

where p and q represent the number of the non conjugated terms and conjugated terms respectively, and p + q is called the moment order •  High order cumulants The symbolism of nth order cumulant is similar to that of the nth order moment, HOC can be derived from HOM (16,17). Let be a signal vector Xi = xi1, xi2, …, xiN and < > denote the statistical expectation. The second, third and fourth order cumulants at zero lag are: [16]

C x1 ,x2 =< X1 , X 2 >=

[17]

C x1 ,x2 ,x3 =< x1 ,x 2 ,x 3 >=

[18]

1 N n n ∑ x1 x2 N n=1 1 N n n n ∑ x1 x2 x3 N n=1

C x1 ,x2 ,x3 ,x4 = = < x1 ,x 2 ,x 3 ,x 4 > − < x1 ,x 2 >< x 3 ,x 4 > − − < x1 ,x 3 >< x 2 ,x 4 > − < x1 ,x 4 >< x 2 ,x 3 > =

1 N n n n n ∑ x x2 x3 x4 − Cx1x2 Cx3x4 − Cx1x3Cx23x4 − Cx1x4 Cx23x3 N n=1 1

Digitally Modulated Signals Identification Based on Artificial Neural Network

by

65

•  Feature set of complex envelope Let Hy be the complex envelope of the sampled signal y(t) which is defined H y = [ y(t)+ yˆ (t)] exp (− j 2πf c t )

[19]

where ŷ is the Hilbert transform of y(t) and fc is the carrier frequency. Let R be the real part of Hy and I its imaginary part of Hy. Thus the following feature set is proposed: CR,R , CR,I, CI,I CR,R,R, CR,R,I, CR,I,I , CI,I,I CR,R,R,R, CR,R,I,I, CR,I,I,I, CI,I,I,I which are the second, third and fourth order cumulants and cross cumulants of the real and imaginary parts of the signal (2) 3.2.2 – Cyclostationarity Cyclostationarity of modulated signal is due to the periodic repetition invoked by the symbol rate. Important parameters to measure the spectral correlation of a time series x(t) are spectral correlation density (SCD) and spectral coherence function (SCF). The SCF of a function x(t) with frequency α is defined as: [20]

C Xα ( f ) =

S Xα ( f )

1/2

⎡ 0 ⎛ α ⎞ 0 ⎛ α ⎞⎤ ⎢ S X ⎜ f + ⎟ ∗S X ⎜ f − ⎟⎥ ⎝ 2 ⎠ 2 ⎠⎦ ⎣ ⎝

SX∝(f) is SCD, it is a cross correlation between two frequency components separated by α, namely (f – α/2) and (f + α/2). The following properties make cyclic spectral analysis a very useful tool for signal classification: different types of digital modulated signals have highly distinct SCD/SCF, and stationary noise exhibits no spectral correlation (5,18). 3.2.3 – Wavelet Transform The digitally modulated signals are decomposed as a linear superposition of the analyzing wavelet at a variety of scales. The properties of the Wavelet transform are used to extract the necessary features for AMR. The features extracted from the wavelet transform contain time domain and frequency domain. The wavelet transform has several properties like linearity, shift property, scaling property and localization property. The scale value is chosen such that narrow peaks

66

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

arise at the points of phase change, frequency change or amplitude change. Scale analysis is the main reason that wavelet transform is used for feature extraction (19, 20). The important types of wavelet are the Mexican Hat Wavelet, the Haar Wavelet, and the Morlet Wavelet. 3.3 – Zero-Crossing Counting the number of zero-crossing of the received modulated signals has been employed for modulation classification (16). The zero-crossing sampler has the advantage of providing accurate phase transition information over a wide dynamic frequency range. Many signal parameters such as zero-crossing variance, carrier to noise ratio (CNR) and carrier frequency can be estimated. Signal parameters such as phase difference and zero crossing interval histograms have important roles of features for AMR (21, 22). 3.3.1 – Temporal Features Temporal features are derived from the general digital IF signal represented in Eq. [21]: [21]

s(t) = Re { A(t)g(t)exp [ j 2πf (t)+ ϕ(t)]}

where A(t) is the amplitude, g(t) is the response of the symbol pulse shaping filter, f(t) is the carrier frequency, and φ(t) is the phase. All temporal features are essentially extracted from these parameters(23, 24). Beside these features there are many other features including modulations, type of channels, and comments presented in (25,26). 4. ANN Based Identifier for Digital Modulated Signals 4.1 – Theory ANN is an important information system that processes data in the same manner as biological Neural Network, based on the assumptions that: •  Information processing occurs at many elements called neurons. •  Signals are passed between neurons over connection links. •  Each connection link has an associated weight, which multiplies the signal transmitted. •  Each neuron applies an activation function (usually nonlinear) to its net input to determine its output signal. A neural network is characterized by: •  Its pattern of connections between the neurons (called architecture).

Digitally Modulated Signals Identification Based on Artificial Neural Network

67

•  Its method of determining the weights on the connection (called its training or learning algorithm). •  Activation function. In computational networks, the activation function of a node defines the output of that node given an input or set of input, activation function having types, Identity function, Binary step function, binary sigmoid function and Bipolar sigmoid function (27). Classification and recognition are important forms from the basic forms of neural information processing (28). ANNs are important in modulation identification because of: • Input-output mapping: Input-output mapping is learned from training data, the synaptic weights are modified to optimize the criterion. • Nonlinearity: Nonlinearity allows modeling of nonlinear functions and processes. •  Adaptivity: weights can be retained with new data, the network can adapt to nonstationary environment (29). 4.2 – Neural network architectures: The important neural network architectures are: •  Multi-layer perceptrons (MLP): the multi-layer feed forward neural network (MLNN) called also multi-layer perceptron (MLP), it is one of the most popular neural network architectures. Its basic unit, the neuron, is composed of a linear combiner and an activation function. The activation function may be a linear or nonlinear function, the choice of the activation function depends on the application aimed at by the neural network (30). •  Radial basis function (RBF) networks: a function is said to be a radial basis function (RBF) if its output depends on the distance of the input from a given stored vector, the RBF neural network has an input layer, a hidden layer and an output layer, each hidden neuron computes the distance from its input to the neuron’s centered point, and applies the RBF to that distance, and the output of all these hidden neurons are combined linearly at the output node (31). •  Self-organizing maps (SOM): the self-organizing map (SOM) was first introduced by Kohonen in 1982. Now this algorithm is very popular, and widely applied in signal processing applications. Kohonen self organizing maps, otherwise called a topology preserving maps, assume a topological structure among the cluster unit. Kohonen network follows the ‘’winner-takes all’’ policy. The network cluster unit whose weight vector matches more closely with the input pattern is considered the winning neuron (32)(33). • Adaptive resonance theory (ART) network: adaptive resonance theory (ART) was developed by Carpenter and Grossberg (1987). One form, ART1, is designed for clustering binary vectors, another, ART2 accepts continuous valued vectors. Details of the operation of ART1 and ART2 are presented in (27). •  Principal component analysis (PCA): principle component analysis (PCA)

68

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

is a data analysis technique used for dimension reduction and feature extraction. PCA is a non-parametric method of extracting relevant information from data sets. The transformed data is able to reveal hidden dynamics in more accurate way than the original data set. It provides a useful procedure to reduce a complex data set to a lower dimension (34). •  Hopfield networks: the Hopfield network is normally used with binary inputs. Hopfield network can be used as an associative memory or to solve optimization problems. When the Hopfield network is used as a classifier, the output after convergence must be compared to the exemplars which its number equals to the number of clusters, to determine if it matches an exemplar exactly. If it does, the output is that class whose exemplar matched the output pattern. If it does not then a “no match” result occurs (35). •  Learning vector quantization (LVQ): learning vector quantization (LVQ) is a pattern classification method in which each output unit represents a particular class or category. The architecture of an LVQ neural network, is essentially the same as that of a Kohonen self organizing map, each output unit has a known class that it represents (27). •  Probabilistic neural network (PNN): The probabilistic neural network is constructed using ideas from classical probability theory, such as Bayesian classification and classical estimation for probability density function, to form a neural network for pattern classification (27). PNN has the advantage of much faster training over other neural networks (34). • Hierarchical neural network: a Hierarchical classifier based on multivariate classification method (33). Some neural networks are classified as feed-forward while others are recurrent (i.e implement feedback) depending on how data is processed through the network. A popular family of ANN is the feed-forward networks which include the MLP and the RBF networks (2). More details about neural network architectures can be found in (27)(30). 4.3 – Neural network learning rules Another basis for classification is to differentiate neural networks by their learning mode, supervised and unsupervised. A supervised ANN is used when the desired output is known while an unsupervised ANN is used when there are no target outputs. But in some situations fewer details are available about the correctness of the network’s response, reinforcement learning has been so-called (28). Most of the ANNs that have been used in the AMC field use the supervised learning techniques including MLP and the RBF. SOM neural network is an unsupervised ANNs techniques that has been used in designing AMC (16). It was recently observed that learning or training algorithms has significant impact on the performance of ANN, also they are modified to reduce the training time, there are several algorithms like back propagation (BP), resilient propagation (RPROP)

Digitally Modulated Signals Identification Based on Artificial Neural Network

69

(36), scaled conjugate gradient (SCG) (37), conjugate gradient descent (CONJGRAD) algorithm (37, 38), quick prop (QP), extended delta-bar-delta (EDBD), and super self adaptive back propagation (SUPERSAB) (39), and Levenberg Marquardt (40, 41). 4.4 – Hybrid identification systems The hybrid identification system that recognizes a variety of digital communication signals. The hybrid system includes three main modules: feature extraction module, classifier module and optimization module. In the optimization module, it is optimized the classifier for selection the best features that are fed to the classifier (41), to reduce ANN complexity and the training time (42). It has high recognition accuracy (41). There are many method for optimization, stochastic search method using Genetic Algorithm (GA) (2, 43). GA adopts Darwin’s theory of survival of the fittest. Another methods for optimization are implemented by PCA (34). Bees algorithm (BA) (39), and particle swarm optimization (PSO) (41) . 4.5 – Identification Approaches with different learning rules, different feature extraction, different ANN architectures, and hybrid systems. ANN based AMC system which can be used to classify digital modulated signals. There are three different steps in developing an automatic modulation identification. The first is the extraction features and selection. The second step involves the development of the automatic modulation classifier based on the learning rules. The third step involves the performance evaluation of the developed automatic classifier with a related study from ANN architectures. The direct comparison between different works is difficult in modulated signal recognition. This is mainly due to the fact that there is no single unified data set available (39). Table 1 shows selected detailed information about the different types of identification using ANN from years (1997-2014) and shows the effect of three steps above, the table includes types of modulated signals, feature extraction and selection, hybrid system, ANN architectures with learning rules, lower SNR, and recognition performance (RP). All the researches aim to fasten learning speed, heighten the recognition rate, decreasing the signal to noise ratio (SNR) required for identification, and increasing the types of modulated signals. Additive white Gaussian noise (AWGN) channels are proposed, except (34) the channel is flat fading channel with phase offset. 4.5 – Table 1 review In the proceeding of the Canadian conference of electrical and computer engineering, CCECE’97, Kremer and Shiels (1997) pointed out no existing classifiers were able to classify an accuracy of more than 90% for SNR below 10 dB. In (40) their research aims to classify the modulated signals with SNR as low as 5

Ref. No.

(40)

(44)

(13)

(39)

(2)

(32)

(41)

(23)

(34)

(45)

No.

1

2

3

4

5

6

7

8

9

10

Table 1

0 dB 0 dB

Time based features and statistical MLP- BP with momentum and features. adaptive weight adjustment Time domain features, HOM and MLP- BP with momentum, HOC. BA for features selection. adaptive learning rate and RPROP.

-5 dB

1 dB

10 dB

0 dB

15 dB

0 dB

10 dB

σdp, σap,γmax, δa, δda, σfn and γmaxf

MLP- BP (Levenberg Marquardt) algorithm

SNR 5 dB

ANN Architecture-learning rules MLP- BP

Time domain features

Feature extraction and selection

γmax,δap, δdp, δaa, δaf, MLP- BP with momentum, Cumulants and complex envelop by adaptive learning rate and RPROP. Hilbert transform. GA for selection. Modified SOM- Kohonen Self2FSK, 4FSK, 2PSK, 8PSK, MSK, Single frequency, instantaneous organized learning algorithm and and 16QAM. amplitude max at center etc [32] neighborhood function 4ASK, 8ASK, 2PSK, 4PSK, 8PSK, HOMs, HOCs. and instantaneous 2FSK, 4FSK, 8QAM, 16QAM, RBF features. PSO for optimization 32QAM, 64QAM, and V32 1. σa, σF, σδF, σ|δF| BPSK, QPSK, BFSK, 8QAM, 2. Histogram extracted from Hierarchical classifier 32QAM, 16QAM and 64QAM. instantaneous amplitude. 16QAM, 64QAM, 4ASK, 8ASK, 1.PNN Continuous Wavelet transform 2PSK, 4PSK, 8PSK, 2FSK, 4FSK 2. MLP (CWT). PCA for feature selection and MSK. for comparative study 2ASK, 4ASK, 2FAK, BPSK, and MLP- BP, Python programming γmax, σap, σdp and σaa, QPSK language rather than Matlab.

Types of modulated signals ASK, BPSK, FSK1, FSK2, QPSK, analog AM, FM and CW ASK2, ASK4 MSK, FSK2, FSK4, PSK2, and PSK4 ASK2, ASK4, BPSK, QPSK, FSK2, FSK4, QAM16, V29, V32 and QAM64 2FSK, 4FSK, 2PSK, 4PSK, 4ASK, 8ASK, 8QAM, 16QAM, 32QAM, 64QAM, and 128QAM ASK2, ASK4, BPSK, QPSK, FSK2, FSK4, QAM16, V29, V32, and QAM64

Summary of selected automatic modulation identification techniques

Hybrid system, and efficient Learning rules

Hybrid system

Previous work. Low success rate, high SNR

Notes

Better in high order modulation types

98%

Using Python programming

1.99.8% 1.Hybrid 2.99.2% 2.Fading channels

95%

98.45% RBF, and hybrid system

97.5% Using SOM

99%

92%

98%

99%

80%

RP

70 S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

Digitally Modulated Signals Identification Based on Artificial Neural Network

71

dB, such development is introduced by selecting 21 features for AMR, and MLP classifier with one hidden layer. The classification processes had not achieved the accuracy of 90%, it shows a success rate of 80%. In (44), MLP classifier was used, with two hidden layers, some changes were made to increase the performance, one of them was to use tan-sigmoid activation function in the hidden layers. In the output layer a log-sigmoid activation was used. By also using the LevenbergMarquardt back propagation algorithm. Thus the training time and network size were reduced. The classifier was able to achieve accuracy of 99% for SNRs ranging of 10-20 dB, there is a degradation in performance at 5 dB due to that ANN was not trained at 5 dB SNR. The performance can be improved by training the network on signals with 5 dB SNR, but this may result in a degradation of performance at higher SNRs. In (13) MLP with one hidden layer, back-propagation learning with momentum as well as adaptive weight adjustment are used for fast network training. New statistical feature set in addition to the time base features were proposed, the addition of statistical feature set might have to get adequate performance with one hidden layer. One of the disadvantage fc is known, was assumed, but in practice applications it would need to be estimated. In (39), 11 different modulation types including high level QAM (128QAM) were proposed for classification, good RP and low training time were obtained by using hybrid system, which includes, BA for selection of best combination of features, MLP classifier with single hidden layer, the activation functions of tan sigmoid and logistic were used in the hidden and the output layers respectively. MLP was trained with different learning algorithms, it was pointed that QP is the fastest learning algorithms among the tested algorithms, thus it reduces the training time. In (2), hybrid system, QA was used to select the best feature set, MLP classifier with single hidden layer, and RPROP learning rule to reduce the training time and improve the performance. In (32) SOM neural network was adopted, the traditional training algorithm of SOM was improved in terms of learning efficiency and classification performance. It was approved that SOM neural network has higher recognition rate than the method based on BP neural network with the same SNR and same types of modulated signals. In (41), hybrid system was proposed to classify 12 different modulation types, RBF was recommended as a classifier, PSO for feature selection, the proposed classifier as well as the optimization method, generally improve the performances of identification at low SNR. In (23), a hierarchical approach was used to first make separations into intermediate subclasses, then a second classifier was used to discriminate between higher order modulation schemes, in general the principle of hierarchical model is dividing the modulation schemes into subclasses, then the hierarchical neural network required fewer weights compared to large MLP and hence reduce the training time, also the hierarchical neural network is efficient at higher order modulation schemes. In (34) for flat fading channel with a phase offset a hybrid system was proposed, PCA for feature selection, PNN as a classifier. PNN was compared with MLP, it was found that PNN perform better in terms of classification accuracy, as well as training time than MLP. In (45) MLP with one hidden layer, test signal genera-

72

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

tion and feature keys extraction simulation using Matlab while the classifier was developed using Python programming language, the developed classifier has successfully recognized the modulation formats of interest with success rate greater than 98% at SNR below 0 dB. 5. Conclusion and future work 5.1 – Conclusion Digitally modulated signals identification is a tool generally used to classify the modulation types of the modulated signals without a priori knowledge about the received signals. AMR has become one of the active areas of applications for ANN. Digitally modulated signals, features extraction, ANN architectures and automatic classification approaches are investigated. Most of the previous techniques before 1997 can only recognize a few kinds of digital signals, they need high level of SNR for classification of the digital modulated signals. These problems are mainly due to features, classifiers, and learning rules. In this paper we have demonstrated a multiple types of feature extraction keys as well as the feature selection methods to reduce the complexity of the recognizer and improve the performance such as GA, PCA, BA and PSO. Different approaches for classification with different architectures of ANN like MLP, SOM, RBF, Hierarchical neural network and PNN, different types of training algorithms QP, EDBD, Super SAB, and CG, and a hybrid classifiers are demonstrated. Most of works have been MLP used because it’s simplicity, and it is considered to be more flexible, it’s disadvantage is that it is quite computationally intensive compared to other traditional techniques. Using Python programming language rather than the MATLAB to reduce the computational complex. According to the mentioned survey, PNN exhibits better performance than MLP in terms of accuracy and computational complexity. It is clearly shown that the performance of the classification is improved when a feature selection and optimization are introduced, hybrid techniques are used, using other efficient ANN types based classifiers, and modified learning rules. 5.2 – Future Works 1.  Introduce new feature extraction method using S-domain by applying Laplace Transform to the modulated signals that may be less sensitive to noise or improve the performance of the classifiers. 2.  The channel environment for modulation recognition is taken as AWGN channels in the most cases, any other signal impairments that degrade the signal quality must be taken into account. The effect of another types of channels need to be carried out. 3.  MLP, RBF and SOM are the popular family of ANN, they are chosen as

Digitally Modulated Signals Identification Based on Artificial Neural Network

73

classifiers, the other types can be used for classification or feature selection. 4.  It is very important to reduce the training time, new method is requested to remedy this matter. 5. ANN can be combined with other soft programming techniques like fuzzy, Genetic and Swarm, the combination must be chosen to offer high performance at low SNR, optimum feature selection, and low training time.

REFERENCES (1) C.S. Weaver, C.A. Cole, R.B. Krumland, M.L. Miller ,’ The automatic classification of modulation types by pattern recognition, Techn. Rep. Accession No. ADO691069, (Stanford University, Calif. Stanford Electronic Labs, April 1969). (2) M.L.D. Waong, A.K. Nandi, Automatic digital modulation recognition using artificial neural network and genetic algorithm, Elsevier Signal Processing, 84, 351-365, 2004. (3) S.S. Soliman, S. Hsue, Signal classification using statistical moments, IEEE Trans. on Commun., 40 (5), 908-916, 1992. (4) Z. Wu, X. Wang, Z. Gao, G. Ren, Automatic digital modulation recognition based on support vector machines, Proc. of 2005 Intern. Conf. on Neural Networks and Brain, pp. 10251028, 2005. (5) L. Qian, C. Zhu, Modulation classification based on cyclic spectral features and neural network, Proc. 3rd Intern. Congress (CISP2010) on Image and Signal Processing, pp. 3601-3605, 2010. (6) M. Richterova, D. Juracek, A. Mazalek, Modulation classifier of analogue modulated signals based on method of artificial neural networks, Proc. Intern. Conf. on applied electronics, pp. 145-147, 2006. (7) A.K. Nandi, E.E. Azzouz, Algorithms for automatic modulation recognition of communication signals, IEEE Trans. on Commun., 46 (4), 43-436, April 1998. (8) J.J. Popoola, Sensing and detection of a primary Radio signal in a cognitive Radio environment using modulation identification technique, Ph-D thesis, Univ. of the Witwatersrand, Johannesburg, 2012. (9) B.A. Forouzan, Data communications and computer networking, (MacGraw Hill Higher Education, 2006). (10) S. Haykin, Communication Systems, (John Wiley and Sons, 2001). (11) L.W. Cough, Digital and Analog communication systems, 8th ed., (Prentice Hall, 2012). (12) W. Juan-Ping, H. Ying-Zheng, Z. Jin-Mei, W. Hua-Kui, Automatic Modulation Recognition of digital communication signals, Proc. 1st Intern. Conf. on Signal Processing, pp. 590593, 2010. (13) M.L.D. Wong, A.K. Nandi, Automatic digital modulation recognition using spectral and statistical features with multi-layer perceptron, Proc. ISSPA 2001, Malaysia, pp. 390-393, 2001. (14) E.E. Azzouz, A.K. Nandi, Automatic identification of digital modulation types, (Elsevier Science), pp. 55-69, 1995. (15) H. Wang, B. Zhang, J. Wu, Y. Han, X. Wu, R. Jia, A research on automatic modulation recognition with the combination of the rough sets and neural networks, Proc. 1st Intern. Conf. on Pervasive computing, signal processing and applications, pp. 807-810, 2010. (16) A. Hazza, M. Shoaib, S. A. Alshebeili, A. Fahed, An overview of feature- based methods for digital modulation classification, Proc. 1st Intern. Conf. on Communications, signal processing, and their applications (ICCSPA), Sharjah, pp. 1-6, 2013. (17) A. EbrahimZadeh, Automatic modulation recognition using RBFNN and efficient fea-

74

S.B. Sadkhan - A.Q. Hameed - H.A. Hamed

tures in fading channels, Proc. 1st Intern. Conf. on Networked digital technologies, pp. 485-488, 2009. (18) A. Fehske, J. Gaeddert, J.H. Reed, A new approach to signal classification using spectral correlation and neural networks, Proc. 1st IEEE Intern. Symp. on New frontiers in dynamic access network, pp. 144-150, 2005. (19) R. Hippenstiel, H. El-kishky, C.F. and S. Dataprasad, Modulation identification using neural networks and wavelet domain based approaches, Proc. 2004, conf. record of the 38th Asilomar Conf. on Signals, systems and computers, pp 2116-2120, 2004. (20) Z. Chaozhu, Y. Lianbai, Discrete wavelet neural network group system for digital modulation recognition, Proc. IEEE 3rd Intern. Conf. on Communication software and networks, pp. 603-606, 2011. (21) S.Z. Hsue, S.S. Soliman, Automatic modulation recognition of digitally modulated signals, Proc. IEEE MILCOM, Vol. 3, pp. 645-649, 1989. (22) S.Z. Hsue, S.S. Soliman, Automatic modulation classification using zero crossing, in: IEE Radar and Signal Processing, 137, 459-464, 1990. (23) Z. Ye, G. Memik, J. Grosspietsch, Digital modulation classification using temporal waveform features for cognitive radios, Proc. 18th Annual IEEE Intern. Symp. (PIMRC’07) on Personal, indoor and mobil radio communications, 2007. (24) A. El Rube, N. El-Madany, Cognitive digital modulation classifier using artificial neural networks for NGNS, Proc. 2010 7th Intern. Conf. (WOCN) on Wireless and optical communication networks, (IEEE publ., 2010), pp. 1-5, (25 ) M.W. Aslam, Pattern recognition using genetic programming for classification of Diabetes and modulation data, Ph-D thesis, Univ. of Liverpool, 2013. (26) O.A. Dobre, A. Abdi, Y. Bar-Ness, W. Su, A survey of automatic modulation classification techniques, classical approaches and new trends, IET Commun., 1 (2), 137-156, Apr. 2007. (27) L Fausett, Fundamental of neural networks, (Englewood Cliffs, N. J. Prentice Hall, 1994). (28) J.M. Zurada, Introduction to artificial neural systems, (West Publ. Co., 1992). (29) S. Haykin, Neural networks. A comprehensive foundation, (Macmillan, 1994). (30) M. Ibnkahla, Application of neural networks to communications. A survey, Signal Processing (Elsevier), 80, 1185-1215, 2000. (31) A. Ebrahimzadeh, S.E. Mousavi, Classification of communication signals using an advanced techniques, Applied Soft Computing (Elsevier), 11 , 428-435, 2011. (32) Y. Xu, L. Ge, B. Wang, Digital modulation recognition method based on self organizing MAP neural network, Proc. of 4th Intern. Conf. on Wireless communications, networking and mobile computing, Wicom (IEEE publ.), 418, 1-4, 2008. (33) S.N. Sivanandam, M. Paulraj, Introduction to artificial neural network, (Vikas Publ. House, 2003). (34) H.C. Dubey, Nandita, A.K. Tiwari, Blind modulation classification based on MLP and PNN, Proc. of 2012 Students Conf. on engineering and systems (SCES), Allahabad, pp. 1-6, 2012. (35) R.P. Lippmann, An introduction to computing with neural nets, IEEE ASSP Magazine, 4 (2), 4-22, 1987. (36) M. Riedmiller, H. Braun, A direct adaptive method for faster backpropagation learning: The RPROP algorithm, Proc. of 1993 IEEE Intern. Conf., San Francisco, CA on Neural network, Vol.1. pp. 586-591 1993 (37) J.J. Popoola, R.V. Olst, Effect of training algorithms on performance of a developed automatic modulation classification using artificial neural network, Proc. of Conf. AFRICON 2013, on Telecommunications, pp. 583-588, 2013. (38) Ai-Sheng, Z. Qi, Automatic modulation classification based on the combination of clustering and neural network, (Elsevier), J. of China Universities of Posts and Telecommun., 18 (4), 13-19, August 2011. (39) A.E. Shrme, Hybrid intelligent technique for automatic communication signals recognition using Bees Algorithm and MLP neural networks based on the efficient features, (Elsevier),

Digitally Modulated Signals Identification Based on Artificial Neural Network

75

Expert Systems with Applications, 38, 6000-6006, 2011. (40) S.C. Kremer, J. Shiels, A tested for automatic modulation recognition using artificial neural networks, Proc. of IEEE 1997 Canadian Conf. on Electrical and computer engineering, Vol.1, pp. 67-70, 1997. (41) A. Ebrahimzadeh, R. Ghazalian, Blind digital modulation classification in software radio using the optimized classifier and feature subset selection, (Elsevier), Engin. Applic. of Artificial Intelligence, 24, 50-59, 2011. (42) J.J. Popoola, R.V. Olst, A novel modulation-sensing method, IEEE Vehicular Technology, 6 (3), 60-69, 2011. (43) A. Ebrahimzadeh, A. Ranjbar, Intelligent digital signal-type identification, (Elsevier), Engin. Applic. of Artificial Intelligence, 21, 569-577, 2007. (44) G. Arulampalam, V. Ramakonar, A. Bouzerdoum, D. Habibi, Classification of digital modulation schemes using neural networks, Proc. of the 5th Intern. Symp. (ISSPA’99) on Signal processing and its applications, (IEEE publ.), Vol.2, 649-652, 1999. (45) J.J. Popoola, Automatic recognition of both inter and intra classes of digital modulated signals using artificial neural networks, J. Engin. Sci. Techn., 9 (2), 273-285, 2014 (School of Engineering, Taylor’s University).