Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2011 International Conference on Parallel Processing Workshops

Optimal Multipath Planning for Neyman-Pearson Detection in Wireless Sensor Networks


Yung-Liang Lai and Jehn-Ruey Jiang National Central University Jhongli, Taiwan 945402011@cc.ncu.edu.tw jrjiang@csie.ncu.edu.tw

AbstractTarget detection is one of the most important services in wireless sensor networks (WSNs) for making decisions about the presence of specified targets by collecting sensed data from geographically distributed wireless sensors nodes. In this paper, we consider designing target detection systems in WSNs on the basis of the Neyman-Pearson Detector (NPD), a statistical decision making method of which accuracy depends on the amount of data collected within a limited time period. We propose the Optimal Multipath Planning Algorithm (OMPA) based on the maximum flow minimum cost algorithm for WSNs to set up paths to reliably deliver as many as possible data packets from data sources to the sink node. OMPA is optimal in the sense that it sets up the maximum number of node-disjoint paths composed of the links with the minimized expected transmission time (ETT). We also evaluate OMPAs decision quality with the help of the Receiver Operating Characteristic (ROC) curves and compare OMPA with the Minimum Cost Path Planning Algorithm (MCPPA) in terms of the detection decision quality and the number of available paths at the presence of node failures. Keywords- Neyman-Pearson Detector, Wireless Sensor Network, Maximum Flow Minimum Cost, Disjoint Path

I.

INTRODUCTION

The rapid progress of wireless communications and micro-electro-mechanical system (MEMS) technology has brought to us the wireless sensor network (WSN). A WSN consists of many tiny battery-powered sensor nodes equipped with microcontrollers, wireless radios, and analog/digital sensing modules to sense environmental conditions, such as light, temperature, sound, vibration, magnetic field, etc. Sensor nodes can communicate with each other and deliver the sensed data to the sink node through multi-hop wireless connections. The WSN is selforganizing in the sense that it can be formed without human intervention, and adapt to sensor failure and degradation. It can provide a variety of services like target detection [1], environmental surveillance [7][10] and so on. The WSN based target detection is to make decisions about the presence of pre-specified targets after collecting sensed data from geographically distributed wireless sensor nodes. The system for monitoring the active Volcano Reventador in northern Ecuador [12] is a representative example of WSN based target detection services. Natural

disasters like volcanic eruptions usually confront people with chaos and even threaten of death. For example, Simkin et al. [11] report at least 270,000 people have been killed by volcanic eruptions since AD 1. To prevent such threaten, the WSN based target detection systems need to be timely, reliable and accurate. Since sensor nodes are powered by batteries and the transmission links between nodes are wireless, the nodes and/or links may fail occasionally. It is therefore desirable to properly construct multiple paths to timely and reliably route sensed data from the data source sensor nodes to the sink node for making accurate decisions. In this paper, we consider designing target detection systems in WSNs on the basis of the Neyman-Pearson Detector (NPD) [6], a statistical decision making method of which accuracy depends on collecting an enough number of data in a limited time period. Thus, our design problem is how to timely and reliably route a certain number of sensed data from the source sensor nodes to the sink node. Aiming at this problem, we propose Optimal Multipath Planning Algorithm (OMPA) for a WSN to set up the maximum number of node-disjoint paths of the links with the minimized expected transmission time (ETT) [3] to transmit packets of sensed data from the source nodes to the sink. By the maximum number of node-disjoint paths, the number of simultaneous node failures tolerable by OMPA is maximized. Furthermore, by the minimized ETT, OMPA can deliver the maximum amount of data within a limited period of time so that NPD can make the most accurate decisions. The cost of running OMPA is low, since it is based on Edmonds-Karp maximum flow algorithm and Goldberg-Tarjan minimum cost flow algorithm [5], which are both of the polynomial time complexity. Fig. 1 shows an example of WSN target detection system for volcano eruption where sensor nodes near the crater are responsible for monitoring the magma eruption and transmitting sensed data packets, while other sensor nodes help forward the packets towards the sink node. On receiving the data packets, the Neyman-Pearson Detector at the sink node can accurately make decisions and then immediately take proper actions. The rest of this paper is organized as follows. In Section II, we introduce related work. In Section III, we present the decision model for target detection. Section IV describes the network model and the problem to solve. The proposed OMPA is described in Section V, and its performance simulation results are shown in Section VI. Finally, Section VII concludes the paper.
57 59

1530-2016/11 $26.00 2011 IEEE DOI 10.1109/ICPPW.2011.63

decision rules, and detection performance evaluation of NPD are shown below. S1 S2 S4 S5 S6 S8 S7 S3 [Statistical Assumptions] The target is assumed to emit physical signals, such as electromagnetic waves out of the noisy background, where the background noise signal follows Gaussian Noise Model with zero mean and variance of V. The signal emitted by the target is assumed to be of a characteristic value . As shown in Fig. 2, the samples generated by the sensor source nodes is a set of N elements, x[0], x[1],, x[N-1] with the background noisy signals being [0], [1],, [N-1]. . . .
ypothesis Test

Neyman-Pearson Detector Sk
Figure 1. A WSN target detection system of volcano eruption, where three disjoint paths (S1 S4 S5 SK, S2 S6 S8 SK, and S3 S7 SK) from source nodes S1, S2 and S3 to the sink node Sk are present

II.

RELATED WORK

x[0] x[1] . . . x[N-2] x[N-1] Noisy signals: [0], [1],, [N-1]

H0 / HA

In this section, we introduce some research results related to WSN based target detection. Arora et al. [1] proposed and implemented a system architecture for target detection, which relies on the routing protocol LGRP (Logical Grid Routing Protocol) to deliver packets to the sink node. In LGRP, each node selects a neighboring node closer to the sink node as its parent node to route sensed data towards the sink node. However, when the parent node fails, it requires a long and unpredictable time in selecting a new parent node. It is obvious that the performance of the target detection system is affected significantly by the underlying routing protocol. Wang et al. [9] gave a good survey of existing studies of wireless reliable routing protocols. Srinivas and Modiano [8] studied how to find k node-disjoint paths with minimum total energy consumption. However, it does not consider the case of multiple source nodes. Frank and Tardos [4] proposed an algorithm to construct minimum weight k node-disjoint paths based on the maximum cost submodular flow concept, where the paths are from a given node (called root) to the other nodes. Wang et al. [9] proposed algorithms to construct k node-disjoint paths from root to the other nodes, or from the other nodes to root. However, the algorithms do not consider how to increase the amount of data delivered within a limited period of time. III. DECISION MODEL OF TARGET DETECTION

Figure 2. Illurstration of Neyman-Pearson Detector (NPD) for target detection

[Decision Rules]: Based on the statistical interference theory [6], NPD has the following two statistical hypothesises: null hypothesis (H0) and alternative hypothesis (HA). Given the characteristic value of the target, NPD decides on H0 or HA according to the received sensed samples. H0: x[n] =[n], n=0,..,N-1 ; Target is absent ; Target is present

HA: x[n] = +[n], n=0,..,N-1

To test the above hypothesises to decide if the target is absent or present, NPD has to calculate Eq. (1). The NPD decides on HA (i.e., the target is present), if
( ( ) / ) / ( [ ]) ( [ ] )

>

(1)

In Eq. (1), is a threshold value to control the probability of the decision error to be under a certain level [6]. In contrast, if Eq. (1) is not satisfied, NPD decides on H0 (i.e., the target is absent). [Detection Performance]: The quality of decision making is evaluated by two metrics [6]: false alarm probability and accurate detection probability, shown below.

We present in this section the decision model of NeymanPearson Detector (NPD) to decide the existence of a target. Under statistical assumptions, NPD can make a decision based on validating decision rules and the detection performance can be evaluated. The statistical assumptions,

58 60

1. False Alarm Probability: The false alarm probability PFA is the probability of the case that NPD decides on HA but H0 is true, which is defined as: P = P(H ; H ) = Q(
/

B. Link Reliability Model

(2)

In Eq. (2), is derived from (the details of the derivation are referred to [6]) and Q() is the right-tail probability function defined as: Q(x) = e dt (3)

2. Accurate Detection Probability: The accurate detection probability PD is the probability of the case that NPD decides on HA and HA is true, which is defined as: P = P(H ; H ) = Q

We use the concept of the Expected Transmission Time (ETT) proposed in in [3] by Draves et al. for evaluating the expected time of transmitting a packet over a link. The ETT of a link is calculated by using the forward transmission ratio and the reverse transmission ratio of the link. For a link <i, j> from node i to node j, the forward transmission ratio R(i,j) is the measured probability that a packet sent by i is successfully received by j. On the other hand, the reverse transmission ratio R(j,i) is the measured probability that the acknowledge packet sent by j is successfully received by i. Let S and B denote the size of the sensed data packet and the transmission data rate, respectively. The expected transmission time ETT(i, j) for transmitting a packet over the link from s to t can be modelled as follows:

(4)

In Eq. (4), , , and are system parameters of constant values. Thus, PD depends on N, the number of samples. This implies the number of paths for delivering the sensed data is crucial for detecting the target with high quality. IV. NETWORK MODEL AND PROBLEM DEFINITION

ETT(i, j) =

( , ) ( , )

(5)

In this section, we introduce the network model, the link reliability model, and the problem to be solved. A. Network Model We represent a WSN as a directed graph GV(V, E), called virtual graph, where V is a set of nodes and E is a set of directed edges between the nodes. The nodes set V consists of one sink node and sensor nodes. To be more precise, V=VS{SK}, where SK is the sink node and Vs is the set of sensor nodes. The sensor nodes (i.e., the node in VS) are responsible for monitoring the target and/or transmitting packets to the sink node SK. They are further classified into two categories: source nodes and intermediate nodes. Source nodes are near the target for the purpose of monitoring. They can send out data packets embedded with sensed data, while intermediate nodes can forward the packets to other intermediate nodes or to the sink node. The edges set E represents the relationship between nodes. If a packet sent from a node i, iVS, can reach a node j, j VS or j=SK, then we say that there is a wireless link between i and j and there is an edge <i, j> in E. In the following context, the terms edge and link are used alternatively. In addition, each edge is assigned with a value to represent the cost in delivering a packet over the edge, which is measured by the model shown in the next subsection.

By Eq. (5), we can observe that lower transmission ratios lead to higher ETTs. That is, a link with a lower ETT is more reliable than a link with a higher ETT. For example, when transmitting packets in an open space without obstacles, we get links of low ETT values. Based on Eq. (5), the total expected transmission time of a path p is defined as follows:

ETT(p) =

( ) ETT(i, j)

(6)

C. Problem Definition

We are given a virtual graph GV=(V,E) with the node set V=VS {SK}, where VS is the sensor node set, SK is the sink node, and each edge is attached with a cost of the ETT value. VS is further divided into two partitions: the source node set VR and the intermediate node set VT. That is, VS=VRVT and VRVT=. There is a directed edge <i, j> in E if a packet sent by node i can be received properly by node j, where iVs and jVS SK. Based on GV=(V,E), we define the path set PS to be the set of paths traversable from a node in VR to the sink node SK. Fig. 3 is an instance of the virtual graph to illustrate a WSN for target detection. The network consists of 8 sensor nodes S1, , S8, and 1 sink nodes SK, where S1, S2, and S3 are source nodes near the target. The directed edges (links) represent the reachability relationship, and each edge is associated with a cost of the ETT value.

59 61

S1

2 2 S4 2 10 S6 6 8 7 S7 S8 8 2 S5 4

readings. The intermediate nodes are responsible for forwarding packets. The OMPA is described as follows.

Optimal Multipath Planning Algorithm (OMPA) 9 Sk Input: a virtual graph GV(V, E) Output: a set P of disjoint paths such that |P| is maximized and pP ETT(p) is minimized 1. [Transform GV into a flow network] Add an extra node NX into GV, add a directed edge <NX, i> with cost 0 going from node NX to each source sensor node i, and assign each edge with flow capacity 1. 2. [Transform nodes] For each node with multiple in-edges and multiple out-edges, transform it into the dual-node form. 3. [Derive the maximal-flow minimum-cost flow plan] Execute the Edmonds-Karp algorithm and the GoldbergTarjan minimum cost flow (MinCost) algorithm to derive the maximal-flow minimum-cost flow plan FP, in which each flow goes from NX to SK. 4. [Output] Output P, where P is the set of paths, each going from a source node to SK, as specified in the flow plan FP.

S2

S3

Figure 3. An instance of the virtual graph to represent a WSN for target detection, where the weight on each link denotes the expected transmission time (ETT) of the link.

Given a virtual graph, the problem to be solved is to find a maximum set P of disjoint paths in PS, while minimizing the sum of the total expected transmission times associated with all the paths in P. Note that two paths p1 and p2 are said to be disjoint if they have no common nodes except for the nodes at two ends of the paths. To be more precise, the problem has two goals. The first goal is of higher priority to achieve and is to find a maximized number of disjoint paths to form a path set P.

Goal 1:

Maximize |P|

(7) Now, we explain the details of OMPA step by step. In step 1, an extra node NX is added into GV. Furthermore, an edge <NX, i> going from NX to a source node i with cost 0 is also added. Finally, every edge is assigned with flow capacity 1 to transform GV into a flow network. In step 2, for each node with multiple in-edges and multiple out-edges, the algorithm transforms it into the dualnode form shown in Fig. 4. The purpose of the transformation is to restrict the generated path set to contain node-disjoint paths. As shown in Fig. 3, a node S with multiple in-edges and multiple out-edges will be transformed into the dual-node form containing two nodes S' and S'' connected by an edge with flow capacity 1 and cost 0. Node S' has the same in-edges as node S has and node S'' has the same out-edges as node S has.
Capacity=1

The second goal is to minimize the total ETT costs associated with the paths in the path set P determined in Eq. (7). Goal 2: Minimize pP ETT(p) (8)

V.

PROPOSED ALGORITHMS

In this section, we present the Optimal Multipath Planning Algorithm (OMPA) to solve the above-mentioned problem on the basis of the max-flow min-cost algorithm. The basic idea of OMPA is to obtain multiple disjoint paths for a given virtual graph GV=(V,E) such that the number of the paths is maximized and the total summation of the expected transmission time (ETT) of all the links on the paths is minimized. There are three classes of entities involved in OMPA: the sink node, the source nodes, and the intermediate nodes. The sink node is responsible for constructing and maintaining the routing paths. The source nodes are responsible for transmitting data packet with sensed

In-edges

S'

S''

Out-edges

Cost=0
Figure 4. Illustration of node transformation, in which node S is transformed into 2 nodes S' and S'' connected by an edge with flow capacity 1 and cost 0

60 62

No. of Available Paths

In step 3, OMPA determines the maximum flow minimum cost flow plan in the flow network derived from graph GV within two phases. In the first phase, the maximum flow value from NX to SK is determined by running the Edmonds-Karp algorithm [2]. In the second phase, the Goldberg-Tarjan minimum cost flow (MinCost) algorithm [5] is run to find the minimum cost flow plan FP based on the result of the first phase. In step 4, OMPA sets P as the set of paths specified in the flow plan FP, and then output P. It is noted that each path in P goes from a source sensor node to the sink node SK and does not contain the extra node NX and its out-edges. All the paths in P are disjoint due the assignment of the flow capacity 1 for each edge. By the Edmonds-Karp algorithm, P is the set of the maximum number of node-disjoint paths. And by the minimum cost flow (MinCost) algorithm, the total ETT cost of all paths is minimized. VI. PERFORMANCE EVALUATION

5 that OMPAs ROC curve is higher than MCPPAs ROC curve, which implies OMPA has better accuracy in target detection decision.

6 5 4 3 2 1 0 6 7 8 No. of Failing Nodes 9 OMPA MCPPA

In this section, we evaluate the performance of the proposed OMPA in terms of 1) the degree of fault tolerance, and 2) the accuracy of detection decision. The evaluation results are compared with those of the Minimum Cost Path Planning Algorithm (MCPPA) based on Dijkstras shortest algorithm [2] that finds the minimum cost path for each source node to the sink node without the restriction that paths should be node-disjoint. We take MCPPA for comparison since it returns the set of paths with the minimum cost (the minimum summation of expected transmission time). We develop a simulator on the MATLAB platform to evaluate the performance of OMPA and MCPPA. In the simulation experiments, the network consists of only one group of 10 source nodes, 40 intermediate nodes, and 1 sink node. Moreover, some sensor nodes are turn off arbitrarily to simulate the case of node failures. To evaluate the degree of fault tolerance for OMPA and MCPPA, we measure the number of available paths under arbitrary 6, 7, 8, and 9 failing sensor nodes. Fig. 5 shows the numbers of available paths of the two algorithms for different number of node failures. As can be seen, OMPA has higher fault tolerance degree (i.e., more available paths) than MCPPA. In signal detection theory, a Receiver Operating Characteristic (ROC) curve [6] is a popular tool to evaluate the detection decision quality, which illustrates a graphical plot of the accurate detection probability (PD) vs. false alarm probability (PFA) for a true or false detection decision system. The curve far from the diagonal axis is considered to have high accuracy of decision. Fig. 6 shows the ROC curves of both OMPA and MCPPA. We can observe by Fig.

Figure 5. Comparsions of OMPA and MCPPA in terms of the number of avabaile paths for different number of node failures

1 0.9 0.8 0.7 0.6 PD 0.5 0.4 0.3 0.2 0.1 0 0

RMPA MCPPA Diagonal Axis

0.2

0.4

PFA

0.6

0.8

Figure 6. Comparsions of ROC curves between OMPA and MCPPA

VII. CONCLUSION In this paper, we have studied how to make a target detection decision based on Neyman-Pearson Detector (NPD) [6] using multiple reliable disjoint paths in a WSN. We have proposed the Optimal Multipath Planning Algorithm (OMPA) to obtain the maximum number of disjoint paths with the minimum summation of expected transmission time (ETT) values [3] by running the Edmonds-Karp algorithm [2] and the minimum cost flow (MinCost) algorithm [5]. We have also evaluated OMPAs decision quality with the help of the Receiver Operating Characteristic (ROC) curves [6] and compare OMPA with the Minimum Cost Path Planning Algorithm (MCPPA) in

61 63

terms of the detection decision quality and the number of available paths at the presence of node failures. For the critical applications like volcano eruption detection, how to accurately and timely detect the presence of the target is very important. Our study provides a way to make accurate decisions for the detection of a specified target by using NPD and ETT. In the future, we plan to extend our proposed path planning approach with other metrics of the wireless networks. REFERENCES
[1] A. Arora, P. Dutta, S. Bapat, V. Kulathumani, H. Zhang, V. Naik, V. Mittal, H. Cao, M. Demirbas, M. Gouda, Y. Choi, T. Herman, S. Kulkarni, U. Arumugam, M. Nesterenko, A. Vora, and M. Miyashita, B, A line in the sand: A wireless sensor network for target detection, classification, and tracking, Computer Network., vol. 46, no. 5, pp. 605634,2004. T. Cormen, C. Leiserson, R. Rivest, and C. Stein, Introduction to algorithms, MIT Press, 2001. R. Draves, J. Padhye, and B. Zill, Routing in multi-radio, multi-hop wireless mesh networks, in ACM MobiCom, Sept. 2004, pp. 114 128. A. Frank and E. Tardos, An application of submodular flows, Linear Algebra and its Applications, Vol. 114-115, pp. 329-348,1989.

[5]

[6] [7]

[8] [9]

[10]

[2] [3]

[11] [12]

A. Goldberg and R. Tarjan, Solving minimum-cost flow problems by successive approximation, Journal of Algorithms, 22, pp. 7-18. 1987. S. Kay, Fundamentals of statistical signal processing: detection theory, Prentice-Hall, 1998. K. Romer, F. Mattern, and E. Zurich, The design space of wireless sensor networks, IEEE Wireless Communications, vol.11, no.6, pp.54-61, Dec. 2004. A. Srinivas and E. Modiano, Minimum energy disjoint path routing in wireless ad-hoc networks, Proc. of ACM MobiCom03, 2003. F. Wang, M. T. Thai, Y. Li, X. Cheng, and D.-Z. Du, Fault tolerant topology control for all-to-one and one-to-all communication in wireless networks, IEEE Transaction on Mobile Computing, vol. 7, no. 3, pp. 322-331, 2007. G. Werner-Allen, K. Lorincz, M. Ruiz, O. Marcillo, J. Johnson, J. Lees, and M. Welsh, "Deploying a wireless sensor network on an active volcano, IEEE Internet Computing, Vol. 10, no. 2, pp. 18-25, March-April 2006. T. Simkin, L. Siebert, and R. Blong, Disasters: volcano fatalities-lessons from the historical record, Science, vol. 291, 2001, pp. 255. G. Werner-Allen, K. Lorincz, M. Ruiz, O. Marcillo, J. Johnson, J. Lees, and M. Welsh, Deploying a wireless sensor network on an active volcano, IEEE Internet Computing, vol. 10, 2006, pp. 18-25.

[4]

62 64

You might also like