Development of a Detection System using a Wireless Sensor ... - DiVA

Development of a Detection System using a Wireless Sensor Network

BERNARDO MACIEL

Masters’ Degree Project Stockholm, Sweden August 2008

XR-EE-RT 2008:017

Abstract Wireless sensor networks (WSNs) have recently been under the focus of the research community. The applications for this technology, as well as the problems it poses, are many. In this work, we consider using such technology to implement a detection system. Detection systems are based on statistical signal processing and provide the capability of deciding whether something has or has not happened. They found application in surveillance, monitoring and control systems, among others. This work is divided into two parts. First, we carry out the design, development and implementation of a WSN testbed where algorithms of all sorts can be tested in a real environment. Second, we concentrate on the design and development of a detection system that we implement on the testbed. We put into practice various approaches and architectures using a design approach outlined here. We analyze the algorithms in terms of detection performance. We address the issue of packet losses through the inclusion of heuristics. We analyze, from a theoretical point of view, the communication cost of the various schemes considered. In addition, we discuss other tradeoffs that stem from the use of WSN. Our results show closeness of the designed and measured detection performance for most of the cases, which leads us to the conclusion that we have a valid design procedure. The heuristics to counteract packet losses provide the algorithms with robustness against that issue. Variations on the detection physical scenario are also well stood. We conclude that using WSNs for implementing detection systems brings advantages but also many questions. It is relatively easy to conceive a detection system, using a WSN, that is robust, has low deployment and running costs and achieves good performance. Nevertheless, considerations on features such as energy, communication bandwidth and more have to be done so that we can obtain an optimized system.

I

II

Acknowledgements I would like to thank Prof. Kark Henrik Johansson and Prof. João Sousa for their invaluable mentorship and help throughout all the phases of this work. I especially acknowledge the immense help, guidance and patience that Maben Rabi provided me with. His advice was always a powerful boost in taking this work forward and I cannot thank him enough for that. I would also like to acknowledge the help of my laboratory fellows Erik Henriksson, Magnus Lindhé, Pan Gun Park and Piergiuseppe di Marco. Many thanks to José Ara´ ujo for accompanying me throughout this year in Stockholm in an unique experience. To my brother and parents, thank you for your neverending and strong encouragement and support at all times. Thanks to all the friends I made in Stockholm during this year, for all the fantastic times spent together. Thanks also to my friends in Portugal, especially Ana Azevedo, who was always available to listen, help and cheer, even in the rougher patches.

III

IV

Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . III 1 Introduction 1.1 Motivation . . . . . . . . . . 1.2 Problem formulation . . . . . 1.2.1 Scenario . . . . . . . . 1.2.2 Problem Formulation 1.3 Contributions . . . . . . . . . 1.4 Outline . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

Wireless Sensor Network Testbed 2 KTH Wireless Sensor Network Testbed 2.1 Background . . . . . . . . . . . . . . . . 2.1.1 Wireless Sensor Networks . . . . 2.1.2 State of the art . . . . . . . . . . 2.1.3 Hardware and Software . . . . . 2.2 Design . . . . . . . . . . . . . . . . . . . 2.2.1 System Breakdown Structure . . 2.2.2 Requirements . . . . . . . . . . . 2.3 Implementation . . . . . . . . . . . . . . 2.3.1 Solutions . . . . . . . . . . . . . 2.3.2 Deployment . . . . . . . . . . . .

1 1 2 2 2 5 5

9 . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

Detection System

9 9 9 11 13 14 14 14 16 17 20

25

3 Background 3.1 Detection Systems . . . . . . . . . . . . . 3.2 Detection Theory . . . . . . . . . . . . . . 3.2.1 Hypothesis Testing . . . . . . . . . 3.2.2 Approaches to Hypothesis Testing 3.2.3 Receiver Operating Characteristic V

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

25 25 27 27 29 34

3.3

3.4

Setups . . . . . . . . . . . . . 3.3.1 Centralized . . . . . . 3.3.2 Decentralized . . . . . 3.3.3 Fully decentralized . . Detection Systems and WSNs

. . . . .

4 Design and Implementation 4.1 Design approach . . . . . . . . 4.2 Design . . . . . . . . . . . . . . 4.2.1 Physical variable/Sensor 4.2.2 Following the approach 4.2.3 Packet Losses . . . . . . 4.2.4 Communication Cost . . 4.3 Implementation . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

37 37 39 42 43

. . . . . . . . choice . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

45 45 48 48 50 65 68 71

. . . . . . .

75 75 76 76 78 79 79 81

5 Experiments and Results 5.1 Experiments . . . . . . . . . . . . 5.2 Results and Analysis . . . . . . . 5.2.1 Designed vs. measured . . 5.2.2 Random vs. Deterministic 5.2.3 Various locations . . . . . 5.2.4 Heuristics . . . . . . . . . 5.2.5 More results . . . . . . . .

Conclusions

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

85

6 Final Remarks 85 6.1 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 85 6.2 Detection System and WSNs . . . . . . . . . . . . . . . . . . 85 6.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

Bibliography

89

*

VI

List of Figures 1.1

General detection system architecture. . . . . . . . . . . . . .

2.1

Survey of testbeds recently developed and successfully ployed and tested all over the world. . . . . . . . . . . . System Breakdown Structure. . . . . . . . . . . . . . . . Floor plan - KTH Q 6th (Automatic Control). . . . . . . Standard KTHWSNT power supply. . . . . . . . . . . . KTHWSNT deployment. . . . . . . . . . . . . . . . . . . KTHWSNT Network 2. . . . . . . . . . . . . . . . . . .

2.2 2.3 2.4 2.5 2.6 3.1 3.2 3.3 3.4

de. . . . . . . . . . . . . . . . . .

3 12 15 18 20 21 22

Block diagram of a detection system. . . . . . . . . . . . . . . Types of detection systems. . . . . . . . . . . . . . . . . . . . Choosing an approach to simple, binary hypothesis testing. . ROC for a NP test with two Gaussians with different means and equal variance as hypotheses. . . . . . . . . . . . . . . . .

26 27 30

Schematic for the source of light controlled by a Tmote Sky. . Photo of the source of light controlled by a Tmote Sky. . . . Histograms of the data collected from the network as a whole and discrete distributions fitted (Source of light OFF). . . . . 4.4 Histograms of the data collected from the network as a whole and discrete distributions fitted (Source of light ON). . . . . . 4.5 Histograms of the data collected from the network as a whole and continuous distributions fitted (Source of light OFF). . . 4.6 Histograms of the data collected from the network as a whole and continuous distributions fitted (Source of light ON). . . . 4.7 Histograms of the data collected from all sensors individually and Poisson distributions fitted (Source of light ON). . . . . . 4.8 Histograms of the data collected from all sensors individually and Poisson distributions fitted (Source of light OFF). . . . . 4.9 ROCs obtained for the centralized setup with NP approach. . 4.10 Contour plots for the centralized setup with sequential approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.11 ROCs for the decentralized setup with NP approach. . . . . .

51 51

4.1 4.2 4.3

VII

35

53 53 54 54 56 56 59 61 62

4.12 ROCs for the decentralized setup with NP approach. . . . . . 4.13 Contour plots for the decentralized setup with sequential approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.14 Communication cost if reporting decisions for H1 and H0 , for various Pr(H1 ). . . . . . . . . . . . . . . . . . . . . . . . . . . 4.15 Critical Pr(H1 ) for a given pair (PF , PD ). . . . . . . . . . . . 4.16 Flowchart for the centralized system algorithm. . . . . . . . . 4.17 Flowchart for the decentralized system algorithm. . . . . . . .

70 71 72 72

5.1 5.2 5.3 5.4 5.5 5.6

76 77 78 79 80 80

PD and PF , designed and measured. . . . . . . . . . . . . E(K|Hi ), designed and measured. . . . . . . . . . . . . . PD and PF , Random and Deterministic experiments. . . . E(K|Hi ), Random and Deterministic experiments. . . . . PD and PF under 3 different locations. . . . . . . . . . . . PD and PF , with the heuristics described in Section 4.2.3.

VIII

. . . . . .

. . . . . .

63 64

List of Tables 3.1

Correct and incorrect decisions for a binary hypothesis test. .

29

4.1 4.2

49

4.4

Details of sensors available. . . . . . . . Performance and thresholds for different (NP approach in a centralized setup). . Performance and thresholds for different (NP approach in a decentralized setup). Summary of the various design choices. .

5.1

Average fraction of tests failed and algorithm processing time. 81

4.3

IX

. . . . . . . . . . . . numbers of sensors . . . . . . . . . . . . numbers of sensors . . . . . . . . . . . . . . . . . . . . . . . .

58 63 65

Chapter 1

Introduction Recently, both the reduction in size of electronic devices and the advent of wireless communication have allowed the development of small, wirelesscapable sensors. This technological advance led to the emergence of a new kind of network, the Wireless Sensor Network (WSN) [57]. A WSN consists of sensors communicating between themselves and exchanging information towards ends such as monitoring, surveillance or automation. Due to its great potential and wide range of applications, nowadays there is a great deal of active research in this area. Furthermore, a WSN is a resource-constrained platform when it comes to communication bandwidth, energy and more, which raises many interesting problems.

1.1

Motivation

WSN Testbed A typical tool used in research departments is the testbed [55]. It allows the testing and development of (especially decentralized) algorithms and theories before deploying anything in the real world. In the WSN area, testbeds are particularly useful because research is very active today and also because they enable close contact with the real deployments. There is a high demand for a platform where researchers can test algorithms without a lot of effort. In the Royal Institute of Technology (KTH), there are people doing research in the WSN area. There are various topics where a testbed can be of great aid, e.g. control over wireless link, consensus filters and others. The need for a WSN testbed is therefore real and the work of thesis involves the development of such a system. Security systems The gearing of WSNs to security purposes is of great interest. Example applications include theft alarm (e.g. [9]), automatic surveillance and access control. WSNs allow the design of systems with very high reliability and low power consumption. However, the problem of 1

2

CHAPTER 1. INTRODUCTION

which algorithms to use in a WSN to obtain a security system with a given performance is an important one. There is a real need to make the best possible use of a WSN in terms of its limited resources. When designing a system based on a WSN, it is necessary to dig into the area of energy scavenging. Typically, we will have to accept tradeoffs between the amount of power consumed (thus system lifetime) and the performance of the system. In the case of security systems this is especially important because they are systems that need to have reliability as high as possible. Moreover, WSNs allow the implementation of important methods such as detection or estimation. The applications of these two branches of decision theory are many, including security, localization and tracking of personnel or objects, room ambience control and more. The applicability of these kinds of algorithms is then vast, making them good candidates for implementation in WSNs. The problem of designing a security system, while taking all the issues above discussed, can be looked at under the classical framework of design of detection systems [10, p. 17] with adaptations for WSNs [10, p.18]. Such an approach is the object of focus of this thesis as stated in the problem formulation given below in Section 1.2.

1.2 1.2.1

Problem formulation Scenario

We first describe the scenario envisioned for our problem and then state the problem more rigorously. Suppose the following scenario: a building floor needs to be secured against break-ins. We have a WSN deployed on that floor. How can we design a detection system, using the WSN, with a given probability of detection (and false alarm) while ensuring robustness, speed of response (delay in detection) and low communication? This general problem is more precisely stated as follows.

1.2.2

Problem Formulation

We begin by modeling the WSN as an undirected graph G = (V, E) where the vertex set is V = {F C}∪{1, . . . , N } and the edge set is E ⊆ V ×V. Node F C represents the Fusion Center (FC) i.e. a special node that is connected to every other node and thus we have a tree-like graph. We can have the case that the FC does not exist, that is, it exists at one or more sensors. In that case, the vertex set reduces to V = {1, . . . , N } and the graph has a mesh-like topology. Figure 1.1 shows a block diagram for the general system architecture.

3

1.2. PROBLEM FORMULATION

Figure 1.1: General detection system architecture. The figure shows the main components of a detection system that uses a WSN. There are nodes (sensors) that make observations (measurements) of an environment (process). They then communicate with the Fusion Center through a wireless channel. Taken from [10].

Break-ins are modeled as binary, stochastic events that are characterizable through a physical variable, typically scalar. That is, we have two hypotheses that describe whether a break-in has or has not occured, or H0 : “Break-in

has

occured”

H1 : “Break-in has not occured” At every sensor, we measure the physical variable of interest which is represented as y i , i ∈ V and that can be real or integer. Then, we have the full vector of observations at a discrete time instant j as yj1,...,N = ! "T yj1 , . . . , yjN . The discrete time instants are typically separated by an interval of time equal to the sensors’ sampling period. The hypotheses can also be described using the distributions of the observations or H0 : Y ∼ Distribution0 H1 : Y ∼ Distribution1 where Distributionk , k ∈ {0, 1} are discrete or continuous probability distributions of the physical variable under each hypothesis. The conditional

4


density functions (or mass functions, if the y i are discrete) are known and given by fY|Hk (y|Hk ). Moreover, the hypotheses have a priori probabilities Pr (“Break-in has occured”) = 1 − Pr (“Break-in has not occured”) Pr (H1 ) = 1 − Pr (H0 )

that are assumed unknown. To build a detection system, we need an algorithm capable of taking the measurements communicated by the sensors and producing a decision on what happened as an output. The procedure can be solely executed at the i FC or also at the sensors. In the latter case, we use # ia$ function of the y , i i typically local decisions made by the sensors d # =$γ y , i ∈ V \ {F C}, and in the former, we use the y i directly, that is, γ i y i = y#i , i ∈ V \{F C}. Thus $ we have, at the FC, an output or decision dF C = γ F C γ 1 (y 1 ), . . . , γ N (y N ) . Different algorithms will have different probabilities of committing an error, usually given by the probabilities of detection and false alarm i.e. Pr (Detection) = PD = Pr (“decide H1 when H1 is true”)

Pr (False alarm) = PF

= Pr (“decide H1 when H0 is true”)

We can have different communication amounts depending on which setup is used and that can be represented by a communication cost C which we define as the amount of data exchanged in the network in one algorithm run. Furthermore, in a WSN we will run into the problem of having packet losses and sensor failures. The algorithm must provide robustness against these problems with minimal performance loss. We summarize the assumptions considered 1. A break-in is a binary, stochastic event, characterizable through a physical variable. 2. We are able to determine the distribution of the physical variable that allows the characterization of the break-in. 3. We are not able to know anything about the frequency of the break-ins. The problem can then be formulated as Given a WSN, find detection algorithms with decision rules γ i , i ∈ V that yield a desired pair (PF , PF ) and analyze C and strategies for its reduction. Also, provide heuristics as add-ons to the algorithms so that they can stay robust against packet losses and sensor failures.

5

1.3. CONTRIBUTIONS

1.3

Contributions

Given the problem stated in Section 1.2, the work carried out can be divided in two parts. Thus, the contributions of this thesis are twofold and can be outlined as below. 1. Design and implementation of the KTH WSN Testbed.

1

• Analysis of several state of the art testbeds deployed around the world and the solutions employed. • Design and implementation of a testbed subject to the KTH Automatic Control research group requirements and expected utilization. 2. Development of a detection system based on statistical signal processing and using the KTH WSN Testbed. • Outline of a design procedure for detection systems.

• Design, implementation and analysis of different versions of a detection system with different setups and approaches. • Implementation and analysis of some heuristics to counteract packet losses. • Analysis of the communication cost for the setups and approaches implemented and design of reduction strategies. • Analysis of the tradeoffs and limitations of the use of WSN in building detection systems. This document roughly follows the list of contributions just presented, as we describe in the next section.

1.4

Outline

We begin by describing, in Chapter 2, the design, development and implementation of the KTHWSNT. We then proceed to discuss the work done on the detection system. In Chapter 3, we review the background theory of detection systems. We then explain the design and implementation of the detection system in Chapter 4. We propose a design approach and describe the work done while following that same approach. The issues of packet losses and communication cost are also object of discussion. The experiments run to evaluate the system and the consequent results obtained are shown and discussed in Chapter 5. Concluding remarks, including a discussion on some issues with the use of a WSN, follow in Chapter 6. 1

In collaboration with José Ara´ ujo.

6


Wireless Sensor Network Testbed

7

Chapter 2

KTH Wireless Sensor Network Testbed In this chapter we present the work carried out towards the design, development and implementation of a Wireless Sensor Network Testbed in KTH (KTHWSNT). First, it is given an overview about WSNs and WSNs testbeds. It is then presented the state of the art, which is followed by the description design approach taken and its implementation.

2.1

Background

In this section it is presented an overview of WSNs and WSNs testbeds. We will focus on the state of the art of WSN testbeds and show their features, architectures, communication characteristics as well as the hardware and software used.

2.1.1

Wireless Sensor Networks

Nowadays, in all systems used to perform a broad type of tasks – such as producing lines, airplanes, cars, electric plants, satellites, health-care, and many more – a feedback control is applied [25]. Since the early 1930’s we have witnessed developments in engineering towards better control of processes. The original theory is referred to as “classical control”, being the control made in hardware and continuous time. In the middle of the 1950’s “digital control”, where computers are the ones responsible for closing the loop (discrete control). After the 1980’s, due to communication links improvements, “networked control” began to be used, where a group of computational units are used in a cooperative and decentralized way to perform a certain task. In the beginning of the 21st century, novel control theories were introduced in order to cope with wireless technology, originating “wireless con9

10

CHAPTER 2. KTH WIRELESS SENSOR NETWORK TESTBED

trol” [24]. The main characteristic of these systems is that the link between sensors and controllers and controllers and actuators is wireless. In [17], it is presented the development path of WSN since 1994, when DARPA funded research on “Low power wireless integrated micro-sensor”. An announcement in 2003 from MIT followed which said that the WSN is one of the 10 technologies that will have the highest influence on the future. We are currently in 2008 and from all the research being made in this area (IEEE Signal Processing, Robotics, Communications) we can see great achievements but also that a lot of work remains to be done. As the forecast presented in [25] shows, the number of wireless sensors and embedded devices in the world will reach 1 trillion, fitting the idea that we can connect everything using wireless networks. These networks can be applied to solve or/and ease tasks that are usually performed with cable solutions and also to create novel applications. The applications can be, as seen in [25], [17], [36], and are for example • Wireless mining ventilation control. • Wireless control of flotation process (industrial monitoring and process control). • Vehicle fuel efficiency with networked sensing (automotive). • Disaster relief support using mobile sensors. • Surveillance with networked autonomous vehicles. • Environmental monitoring (terrestrial and aquatic monitoring). • Habitat monitoring. • Security of airports, stations, buildings, etc. • Military – information exchange, sniper detection, etc. • Domotics. • Health care. As advantages of the application of wireless networks we can easily recognize that the cost (wiring and installation work) is reduced and that flexibility (less physical design limitation, more mobile equipment, faster commissioning and reconfiguration) and reliability (no cable wear and tear and no connector failure) are improved. These characteristics place wireless technology on the spot and justify all the research efforts that are currently taking place. We also point out the disadvantages with this type of networks, which make the current challenges of the research community. It is shown in [25]

2.1. BACKGROUND

11

that security, reliability, low knowledge, lack of commercialized solutions are the main problems with this technology at the moment. A lot of these unpleasant features are due to the characteristics of the wireless communications i.e. the large variations in connectivity, low bandwidth, delays and packet losses and the not well developed, at the moment, communication theory [24]. Also presented in [24] and [23] are some solutions used in order to be able to perform control tasks in WSN, that is • Communication protocol suitable for control [59], [58], [60]. • Control application that compensates for communication imperfections [42], [40], [35]. • Cross-layer solution with integrated design of application and communication layers. We can see that in wireless control, communication and control are always together. Therefore, in order to improve system performance, solutions in both areas have to be sought.

2.1.2

State of the art

In order for the research community to test their control algorithms, power management solutions, protocols, among others, WSN testbeds were built. These testbeds are composed of an amount of wireless nodes which are deployed indoors or outdoors according to the needs of the test one wants to perform. Figure 2.1 shows some WSN testbeds recently developed and successfully deployed and tested, in several locations and with different features and settings. The references shown are websites and [3]. Various possibilities can be seen regarding the design of the testbeds. Features Most testbeds share the possibility of • Remote programming and debugging • Monitoring and real-time interaction • Logging • Network administration and management This subset of features allows running experiments with data collection and control over events. Additionally, batch mode, scheduling, quotas and support for multiple, simultaneous users are also common and interesting

:I18R,-)'0

M$>=&J#$*(84,>>&-(8 !$(3&-A

GC&-(&

1"-D"-2

G7H$-A$=$J

-&;-"##)*;89&"-2. $-"()&*?8-&9&(8>&.)()&*)*;?8=)D$8 D)2$&8)#";$. +&',.8&*8.,>>&-(85&-8(0$8[4!8=)5$'J'=$8"*28 ">>=)'"()&*8#&2,="-)(J8E$".J82$D$=&>#$*(8&58>=,;B)*.F 4$*.&-8'"=)9-"()&*8>-&'$2,-$.88 6">)28*$(3&-A8.("($8"*28$".J82"("8D).,"=)S"()&*8 4,>>&-(8(&82$>=&J#$*(8($"#.8 =$8.$*.&-8*$(3&-A.88 /=$-(8'&*5);,-"()&* !$(3&-A8-$>-&;-"##)*; [$98)*($-5"'$83)(08.'0$2,=)*;?8>-&;-"##)*;?8 ",(&#"()'8=&;;)*;8"*28"2#)*).(-"()&*8Eh,&(".F M)-$'(8"''$..8D)"84$-)"=+&-3"-2$-8E4+F H"('08,.$ C&3$-8#"*";$#$*( I$.(9$28#">.

$-*&2$. 4,>$-*&2$.8'&**$'(8(&8.$-D$-O'&*(-&=8.("()&*8D)"8:(0$-*$( G.$-.8'&**$'(8(&8>,9=)'8.$-D$-8(0-&,;08(0$8U*($-*$(

[$98)*($-5"'$83)(08-$.$-D"()&*8"*28h,&(". 4,>>&-(85&-8D"-)&,.8#&($.8(J>$. !$(3&-AB3)2$8>-&;-"##)*;8E-$#&($8"*28>"-"==$=F H"('08&-8)*($-"'()D$8,.$ 4,>>&-(85&-8#&($8(0"*8&*$8(J>$8&58#&($.8E5="(8&-8 0)$-"'0)'"=8.$(,>.F 4$=5B'&*5);,-"()&* /'()D$8>&3$-8.,>>=J8'&*(-&=8EG4H8&-89"(($-)$.i8 2J*"#)'8(&>&=&;J8.$=$'()&*i8*&2$85")=,-$.8)*T$'()&*F M"("8;"(0$-)*; !$(3&-AB3)2$8>-&;-"##)*;

>=)'"()&*?82"("9".$8"*283$98E5&-8,.$-.8(&8 "''$..F

4)#)="-8(&8$-)#$*(.8.$(,>8E9"('08&-8*&(F 6$#&($8#&($8)#";$8)*.("=="()&*

M$9,;;)*; 6$>-&;-"##)*;

4$*.&-8*&2$.@8.$*.&-8K8/OM8K8+C\/ H=,$(&&(0OR);9$$O"..O(0$-#&#$($-O;J-&.'&>$8 '&**$'($28D)"8-"2)&8(&8"8;"($3"J \"($3"J8'&**$'(.8D)"8:*($->-).$8!$(3&-A8(&8'=)$*(.8E,.$.8 >,9=).0O.,9.'-)9$8#&2$=?8-&(&'&=F

+&',.8&*8[4!8'&##,*)'"()&*O>$-5&-#"*'$8($.()*; 6$#&($8#$($-)*;@8D).,"=)S"()&*?8=&;;)*;

()*JN4 !$>(,.8B8#,=()>=$8D$0)'=$8&>$-"()&*8 .,>>&-(85-"#$3&-A?82$D$=&>$28"(8GP8 C&-(&

()*JN4 %)*,b 7"#$-".@8$-*&2$.@8%)*A.J.8!4%G` 4$-D$-@8U*($=8C$*(),#8]

I#&($84AJ 4("-;"($.@8ZB>&-(8G4H80,9O_Y`Paa 4$-D$-@8:(0$-*$(O_Y`Paa

@OO333P(3).(P(,B9$-=)*P2$O 1"*2S).A)Y^

0((>@OO$*=P,.'P$2,O>-&T$'(.O(,(&-*$(O

0((>.@OO#)-";$P9$-A$=$JP)*($=B-$.$"-'0P*$( 70,*Y^

M)#)(-)&,YZ

0((>@OO333P#)==$**),#P9$-A$=$JP$2,O.$*.&-*$(.O

0((>@OO#&($="9P$$'.P0"-D"-2P$2,O [$-*$-B/==$*Y^

C)*(&Yc

V&0*.&*Yc

0((>@OO'$()P'.$P&0)&B.("($P$2,OA"*.$)O :-()*Yc

I$.(9$2.[\8C-$.$*("()&*8YZ

+,--$-Yc

0((>@OO333P9(*&2$P$(0SP'0OC-&T$'(.OV"3. MJ$-YZ

0((>@OO&-$P#)*$.P$2,OQ#'&=";-&O$2;"-#)*$O

M$=,;$ ()*JN4

()*JN4 V"D"O