Power-Aware Markov Chain Based Tracking Approach for Wireless Sensor Networks

Transcription

1 Power-Aware Markov Chain Based Tracking Approach for Wireless Sensor Networks Hui Kang and Xiaolin Li Scalable Software Systems Lab, Department of Computer Science Oklahoma State University, Stillwater, OK 74078, USA Patrick J. Moran AirSprite Technologies, Inc. Marlborough, MA 01532, USA Abstract We propose a novel measure method of information utility for tracking and localization in wireless sensor networks (WSNs). The target moving arbitrarily in WSNs is modeled by Markov chains using a transition matrix. The proposed information utility measurement allows us to expect the next state of the target and identify the informative sensors. Further, compared with existing localization methods, the proposed poweraware sensor selection considers the energy constraint of WSNs. To conserve energy, a subset of sensor nodes are activated based on a combinative measurement including information utility, communication cost, and residual energy. We have implemented the proposed localization system on real motes and experimented in an obstacle-free environment. The experimental results demonstrate that the proposed method outperforms two popular baseline schemes, k-nearest-neighbor and stochastic schemes, at extending the network lifetime. In addition, it balances the energy level of sensors in the network so that energy consumption is spread uniformly over all the sensors. I. INTRODUCTION Wireless sensor network (WSN) emerges as a promising and significant technology to revolutionize our physical environment and lives, but posing many research challenges due to its special characteristics such as limited computing capability, constrained on-board energy, and unreliable radio links [2]. Because sensor nodes are powered by small batteries in most cases, energy will be depleted quickly without any conservation scheme. The extreme scarcity of energy requires WSNs to work in discrete dynamic energy level modes corresponding to different working states such as transmit, receive, and sleep. Usually nodes in sleep mode consume a large magnitude of energy less than in other modes, and some mechanism is employed to switch nodes between sleep mode and others. Motivated by this principle, WSNs wake up enough and appropriate number of sensor nodes around the phenomena to accomplish desired missions. Localization and tracking are two canonical applications in WSNs and closely related. We focus on developing an energyefficient collaborative processing approach in this scenario. Essentially, to reduce the energy consumption only a small number of sensors are activated to track and localize the target; while others are turned into sleep mode. In this paper, we propose a novel measurement of information utility based on Markov chains. Compared to existing 1 A preliminary version of this paper appears as Work-in-Progress and Poster [1] in IPSN 06. metrics, it is able to model the random moving trajectory of a target and also take the physical topology of the network into account. A newly defined combinative measurement, including information utility, communication cost, and residual energy level of each sensor, allows the proposed system to select informative sensors while considering the principal constraint in WSNs energy. Traditionally, however, intensive usage of either the most informative sensors or energy-efficient sensors leads to poor performance in network lifetime or data quality. Leveraging some recent technologies of ultra low power sensor devices, these parameters can be obtained without complicated computations. We develop the proposed system based on real motes and deploy it in an obstacle-free environment to show the improved network lifetime and localization results. II. RELATED WORK In this section, we give an overview of current and previous research efforts on energy saving and target tracking in wireless sensor networks. A. Tradeoff between energy consumption and information utility Energy saving is fundamental for system design in wireless sensor networks due to limited on-board resources of sensor nodes. An extensive amount of solutions have been developed to conserve energy in order to extend the lifetime of sensor networks. Stochastic sensor network in [3] is demonstrated to reduce energy consumption by unsynchronized low-duty-cycle operation of sensor nodes. However, the underlying assumption that energy consumption is dominated by computational costs rather than communication costs makes it not well suited for most sensor networks in which communications consume much more energy than other operations [4]. Other methods [5] [7] achieve energy conservation by either reducing the communication in the sensor network (turn unnecessary nodes into sleep mode or consume as little power as possible), or minimizing the radio transmission power. Essentially, these approaches consider communications dominate the energy cost in a sensor network and try to reduce the amount of communication and the number of active sensors. On the other hand, the number of deployed sensors in tracking applications is always larger than the required number to localize the target, attributing to other two important

2 properties of WSNs, i.e., data-centric and redundancy. The concept of information utility is thus proposed to measure the data a sensor can provide so that those informative sensors are preferred to be selected to join the task and others are in sleep mode. Several kinds of information utility metrics have been developed, such as Mahalanobis distance, nearestneighbor, and entropy [7], [8]. In particular, they are used to predict the data an unused sensor observes to update the status of the target. Therein they always choose the most informative sensors. These methods, however, will deplete the energy of the most informative sensors much more quickly than others, leading to the network partition and performance degradation of the sensor network in worse case. Moreover, Mahalanobis distance is not suitable for tracking application in that they are used to update the belief state, instead of physical location, and entropy is based on the accumulated probability of the target, which does not model random moving trajectory. The significance between energy consumption and information utility should thus be well traded; emphasizing one and ignoring the other result in expected poor performance of sensor networks. As such [9] advocates an energy-aware utility based approach by formulating utility function with energy constraints. However, it does not propose a concrete measurement of information utility, nor three special cases considered therein can be applied to tracking application. In this paper, specifically in Sect. III-A we derive the information utility based on Markov chains, which model the moving trajectory of a target. Afterwards, it will be used to select appropriate sensors in a power-aware manner. B. Tracking and localization Localization is one of the most important and hot topics in the community of WSNs; many algorithms have been proposed to improve the accuracy and robustness to noises existing in the surroundings. Tracking, which aims to obtain fine-grained or coarse-grained location of the specific target, significantly depends on the quality of localization. Most localization methods based on ranging estimations between sensor nodes can be classified into two categories: RF-based and time-of-flight (TOF). RF-based localization algorithm is widely employed due to its elimination of additional hardware equipments and other favorable properties with respect to energy consumption [10] [12]. The received signal strength indicator (RSSI), which becomes available in most representative wireless sensor modules, is used to estimate the distance between two sensors usually by a log-normal shadowing model. Since RSSI is a built-in value passing by the sender and receiver, no additional cost, energy, and configuration are required. This method, however, is sensitive to reflection, interference, multi-path, and so forth. The effects of different conditions on RF-based method are investigated in [13], [14]. Compared to RF-based, TOF-based localization algorithm provides more accurate results [15], [16]. The distance between two sensor nodes can be computed from the travel time from the sender to the receiver by some acoustic signal like audible sound or ultrasound, which has constant travel speed. As opposed to RF-based methods, some sophisticated hardware and software are required, e.g., powerful sound source and synchronized clocks at both ends. None of these proposed localization systems, however, considers energy consumption of sensor nodes; the performance can be degraded when the energy of sensor networks decreases. Since we base our tracking algorithm on the RFbased method, experiments are carried in a real-world wideopen area to avoid the interference and other noises. In general, the proposed sensor selection method can be readily applied to other localization systems. III. MARKOV CHAIN MODELING AND SENSOR SELECTION Typically in a tracking scenario, a target to be localized moves at random within the coverage of the sensor network. Although some previous work used the prior probability and probability distribution function (PDF) to predict the moving trajectory, there is no guarantee that the prior positions would influence the future status of the target. Further, because a target moving to its next position only depends on its current status, it can be modeled by Markov chains, from which information utility is derived (Sect. III-A). As mentioned above sensor nodes consume less energy in the sleep mode than the active mode and continuous radio communication drains energy quickly, it is desirable that the number of sensors to be activated should be necessary, but redundant. Therefore, information utility is further combined with energy consideration to select appropriate sensors in a power-aware fashion (Sect. III-B). In this section, sensor nodes and sensors are interchangeable and referred to deployed sensors to track and localize the target. We assume that the sink node has strong computational capability and is wired to a power supplier such as PC. A. Markovian utility measurement In our Markov chains model, a target moves to the next state from its current state with a probability, and this probability does not depend on which states the target was in before the current state. The state herein refers to a region where the target stays covered by a particular sensor node. As a result, the state is directly related to the location of the target and the state transition represents a moving step. For a certain topology of a sensor network, the whole area is divided into n nonoverlapping regions by n sensors. Each of them represents the range of a sensor the target could be in. Then we have a set of states for the target, S = {s 1, s 2,, s n }. The transition in the sensor network is described by a Markov chain as shown in Fig. 1. From this figure we see that a target can move from any state to any, but at each step it can only reach the neighboring states. Taking grid topology as an example, moving direction within the range of the sensor network can be described by Fig. 2(a). If the target is on the boundary, assuming that it will not be out of the range, Fig. 2(b) shows the moving direction on

3 Fig. 1. Markov chain describes behaviors of the target. (a) Fig. 2. Moving direction of the target (The current state is at T.) on, (a) non-boundary region of the sensor network (b) boundaries of the sensor network boundaries of the grid sensor network. In general, it can be extended to different physical topologies of sensor networks. Let P be the n by n transition matrix for the target in a network with n sensors, whose element p i,j denotes the probability with which the chain moves from state i to state j. n states are represented by n rows of P ; for each state, the columns are the states that it will move to. Evidently the higher probability indicates a higher possibility that the target will enter in its next step; the sensor on that column provides more informative sensing data. The powers, P (t), give transition matrix at discrete time sequence t. In addition, we define signal strength gain (SSG) as the difference of RSSI for a sensor node at different time. SSG is used to quantify whether the target is approaching the sensor node or not. A positive value of SSG for a sensor indicates that the target is moving towards it. Because both the transition matrix and SSG are related to the locations of the target and sensor nodes which are principle factors in the tracking context, they can be used to estimate the information utility of sensor nodes. Information utilities for deployed sensor nodes is denoted by a vector u, and formulated as u = [v T P (t) ] q, t = 1, (1) where v is the current state vector of the target and q is the SSG vector for all sensors in the network. The elementwise product ( ) determines the information utility of each sensor to contribute to the tracking task. Transition matrix as described above can be derived from the neighbor relation of specific topology of a sensor network. For example, in the n n grid sensor network, the corresponding initial transition matrix is (b) P (1) = p 11 p 12 p 1n p 21 p 22 p 2n p n1 p n2 p nn (2) p i,j is nonzero if sensor i and j are neighbors. The probabilities of moving from current state to its neighbors are equally divided after a sensor identifying its neighbors. This means that if one node has 8 neighbors, for that row each nonzero probability equals 1 8. Otherwise, a zero means the target can not move to the state. However, since the target can reach any state after several steps, the probability may not be zero even two sensors are not neighbors. Therefore, after every step we need to compute the powers of P so that the accumulated matrix is able to express the probability that a target can move to its non-neighboring states. Moreover, through some basic matrix computations, the probability from one state to another depends on their intermediate neighbors only (see the APPENDIX). For each row, the summation of p i,j satisfies p i,j = 1, i = 1 n; (3) j=1 then let C = v T P (t), and C should be normalized by, C i = C i n j=1 (C, i = 1 n. (4) j) The vector, q, in (1) is SSG for all the sensors; SSG for sensor i is calculated by { f(i, t), f(i, t) > 0 q i = 0, f(i, t) < 0, (5) where f(i, t) is the difference of RSSI value at discrete successive times, f(i, t) = RSSI(i, t k+1 ) RSSI(i, t k ). (6) The procedure of our tracking approach is summarized in Fig. 3. At phase a, the target is localized, and elements of its state vector v is set either 1 or 0; the position of 1 represents the current state of the target. Since the target belongs to exactly one range of a sensor node, we have only one element of the state vector equal to 1. At phase b, based on transition matrix and RSSI, the information utility vector, u, is determined using the algorithm given in Table I. Following that, at phase c these results will be further used with the proposed power-aware sensor selection scheme to activate informative sensors and balance the energy level. B. Power-aware sensor selection Since energy in WSNs is mainly consumed by communications between sensor nodes, it is critical that redundant information and unnecessary data transfer should be minimized. To achieve this, the information utility of each sensor node is used to evaluate the quality of data, and then we choose appropriate sensor nodes to update the status of the

4 Fig. 4. Localiztion of the target by three senosr nodes.! " Fig. 3. #"$ " " ""% Flowchart of the proposed tracking approach TABLE I ALGORITHM TO CALCULATE THE INFORMATION UTILITIES 1: q = 0; 2: C = 0; 3: If (initialization) Set P by the topology of the sensor work; Else P t = P (t 1) P (1) ; 4: C = v T P t ; 5: Normalize C ; 6: q i = f(i, t) > 0? f(i, t) : 0, i = 1 n ; 7: u = C q ; target. However, without considering the energy level of each sensor nodes, those most informative sensors will run out of energy much more quickly than other nodes, leading to a wide disparity in the energy levels of the nodes, and eventually disconnecting the network. To avoid intensive usage of certain nodes, the proposed power-aware method takes the energy cost and remaining energy as the parameters to select sensors. We define the following notations to ease the presentation: Cost(i, j): energy cost of the communication between i and j; µ i : the information utility for node i; e i : the remaining energy at node i. Given the above notations, the problem is considered to optimize the information utility and energy consumption, and formulated as a combinative measurement, λ i = αµ i + βcost(i, t) + γe i, i = 1 n. (7) Here α, β, and γ are the weighing factors adapted to coordinate the magnitude of the three factors. For the communication cost, because communication between sensors are the main energy consumption in the sensor network, they should ensure reliable transmission to avoid packet loss so that unnecessary retransmission is minimized. LQI (link quality indicator) implemented by IEEE is used in ChipCon2420 trans- ceiver, and adopted by most recent sensor modules, including Tmote sky/telos, micaz, and Intel Mote. LQI is an effective measurement of chip error rate and experimented to be an indicator of the link quality between two nodes [17]. Thus we use LQI to measure the communication cost between deployed sensors and the target. The remaining energy is one of the characteristics of each node. It can be stored and updated by the sink node. Here we ignore the energy cost to send the remaining energy from deployed sensors to the sink node. Then the combinative measurements are ordered so that λ 1 > λ 2 > λ 3 > λ n. By choosing the first k sensors, the sink node can collect informative data and avoid using those informative sensors intensively. During the data communication phase, deployed sensor nodes have two working modes: active and sleep. Sensor nodes in active mode receive signal emitted from the target, track it, and send data to the sink node; while in sleep mode, they only consume a small amount of energy to receive the beacon message from the sink node and do not talk with the target. The beacon message contains the command whether the specific sensor should be active or sleep based on the combinative measurement. Particularly, at phase c of Fig. 3, the activated sensor nodes form a subsystem of the sensor network to localize the target. The subsystem consists of at least three sensor nodes so that the position of the target can be finalized by their ranging distances shown in Fig. 4. IV. EXPERIMENTAL EVALUATION We implement the proposed localization system using Tmote sky motes [18], and compare its performance with other two baseline schemes, k-nearest-neighbor (k-nn) and stochastic selection method. K-NN method selects sensors that are closest to the target so that they can provide more reliable communication and sensing data. In contrast, stochastic selection method randomly chooses sensors avoiding intensive use of some nodes. A. Experimental settings The experiments are performed in an obstacle-free environment (Fig. 5) so that the effects of fading, path-loss, interference, and reflection are minimized for RF-based method. Before deploying the sensor nodes, the relation between RSSI and distance in the specific surrounding is sampled and plotted as shown in Fig. 6. In the experiment setup, 16 motes were placed on 50cmtall paper brackets in a 4 4 grid topology. The target was carried by a person who moved arbitrarily, and emitted radio

5 Fig. 5. Experiment environment (basketball court). 16 sensors standing on frames were deployed in a 4 4 grid pattern; sink node was connected to a PC as a base station and collected data from sensors. Fig. 7. Network lifetime comparison. TABLE II NUM OF EXISTING NODES AT TIME STEP 760 AND 920 Time step K-nearest neighbor 4 0 Stochastic selection 11 3 Proposed method Fig. 6. RSSI versus distance in the experiment. messages every three seconds. To facilitate the analysis and comparison of the results, every sensor nodes were given identical amount of initial energy. The parameters of energy consumption were set according to Tmote sky datasheet [18], i.e., the energy consumption in transmission, reception, and idle state is 52.2mW, 59.1mW, and 0.001mW respectively. We used other two motes, one as the sink node and one to display the data on a PC. Information utility, communication cost, and remaining energy are considered equivalent factors with reasonable parameters α = 100, β = 1, γ = 1 in (7). We are currently working on a middleware to address dynamic adaption of these parameters. When a sensor node receives messages signaled from the target, it reports the data, including RSSI and LQI, to the sink node. Then it turns into sleep mode to conserve energy. After the sink node collects data from all the sensors for one epoch time, it calculates information utilities by the transition matrix and construct combinative measurements for the 16 motes. Three nodes with the largest value are activated to track and localize the target until the next epoch time. B. Results and discussion We first illustrate that the network lifetime of the sensor network is improved by the proposed method. Here network lifetime is defined as the time till the network fails to serve its design purpose [19]. Because we need at least three nodes to position the target, lifetime should be counted until the number of nodes is less than three in the network. Fig. 7 shows the comparison results with respect to the number of nodes that run out of energy and the time step. The proposed method achieves longer lifetime than other two baseline schemes; while k-nearest-neighbor is the shortest. The improvements are 25% and 5%, respectively. Further, from this figure we observe that the number of depleted nodes in two baseline schemes increases much more quickly than the proposed method does. This means during the same time period, energy consumption spreads uniformly in the proposed network, instead of focusing on several nodes. For example, Table II gives the number of available nodes for three schemes after the network running for some time. Our method performs much better than the other two. Even at time step 920, which is very close to the end of lifetime, there are 10 nodes working using our method, while using stochastic selection there are 3 and none for k-nearest-neighbor method. In the second experiment, the mobile target follows a specified trajectory, and moves to its next position every five seconds. Sensor nodes selected by the proposed method use ranging estimation to localize the target. Fig. 8 shows the results plotted over ground truths. The average localization error is 4.7m, which is consistent with [14]. The localization algorithm used is less precise than TOF-based methods with some sophisticated hardware, but it is substantially cheaper and does not introduce additional computational cost since we use the inexpensive built-in RSSI values. In addition, the localization results can be improved by some RF-based Fig. 8. Localization results; ground truths and localized positions are linked by dash lines.

6 algorithms [11], [12], which is beyond the scope of this paper. V. CONCLUSION Stringent energy constraint of sensor networks requires sensor nodes to work collaboratively so that uninformative ones could be turned into sleep to conserve energy and extend the network lifetime. Specifically, for tracking application, we proposed a novel method to measure the information utility of sensor nodes by formulating a Markov chains model. Further, we presented the power-aware sensor selection to avoid continuously consuming those informative and energyefficient sensors by combining information utility, communication cost, and residual energy comprehensively. Experiments on real motes demonstrated that the proposed method extends network lifetime and efficiently balances the energy level of sensor nodes in the network, while not affecting the localization/tracking accuracy. Our future work on power-awareness in WSNs focuses on two areas: application-specific dynamic adaptation to balance the weights of different factors and middleware to abstract the underlying infrastructure. The middleware aims to ease the development of application programs regarding to power management, policy interpretation, tasking, collaborative information fusion, and message routing. APPENDIX Definition: A target moves from state s i to state s (k+1) through a set of intermediate nodes as s i d 1 d 2 d k s k+1. d k is called k-step-neighbor transferred by P (k), and is the intermediate node between d (k 1) and s (k+1). As such s k+1 is the (k+1)-step-neighbor of state s i. Since there are at least more than one neighbor for each state, d 1 d k are the set of immediate neighbors of its previous state, i.e., s d 1 are immediate neighbors of s i, s d 2 immediate neighbors of d 1, and so forth. The element of the transition matrix in (2) related to state s i and s k+1 is calculated step by step through intermediate nodes as follows, p (2) i,d 2 = p (k+1) i,s (k+1) = p (3) i,d 3 = i,d 1 = P r (8) i,l p(1) l,d 2 = i,d 1 d 1,d 2 (9) p (2) i,l p(1) l,d 3 = p (2) i,d 2 d 2,d 3 (10) p (k) i,l p(1) l,s k+1 = p (k) i,d k d k,s k+1. (11) Equation 11 can be rewritten by replacing previous equations as k p (k+1) i,s (k+1) = i,d 1 ( d l,d (l+1) ) d k,s (k+1). (12) Because d 1 d k are intermediate nodes between s i, and s (k+1), the information utility depends only on these intermediate nodes. REFERENCES [1] H. Kang and X. Li, Power-aware sensor selection in wireless sensor networks, 2006, ipsn06/wiplist.htm. [2] D. Culler, D.Estrin, and M. Srivastava, Overview of sensor networks, IEEE Computer, vol. 37, no. 8, pp , [3] S. C. Zhang, F. Koprulu, R. Koetter, and D. Jones, Feasibility analysis of stochastic sensor networks, in Sensor and Ad Hoc Communications and Networks, IEEE SECON First Annual IEEE Communications Society Conference on, no. 4-7, 2004, pp [4] L. Doherty, B. A. Warneke, B. E. Boser, and K. S. J. Pister, Energy and performance considerations for smart dust, Int. J. Parallel Distributed Systems and Networks, vol. 4, no. 3, pp , [5] G. Xing, C. Lu, Y. Zhang, Q. Huang, and R. Pless, Minimum power configuration in wireless sensor networks, in MobiHoc 05: Proceedings of the 6th ACM international symposium on Mobile ad hoc networking and computing. New York, NY, USA: ACM Press, 2005, pp [6] R. Shah and J. Rabaey, Energy aware routing for low energy ad hoc sensor networks, in Wireless Communications and Networking Conference, WCNC IEEE, ser , vol. 1. Wireless Communications and Networking Conference, WCNC IEEE, 2002, pp [7] H. Wang, K. Yao, G. Pottie, and D. Estrin, Entropy-based sensor selection heuristic for target localization, in IPSN 04: Proceedings of the third international symposium on Information processing in sensor networks. New York, NY, USA: ACM Press, 2004, pp [8] F. Zhao and L. Guibas, Wireless Sensor Networks: An Information Processing Approach, D. Clark, Ed. Morgan Kaufmann Publishers, [9] W.-P. Chen and L. Sha, An energy-aware data-centric generic utility based approach in wireless sensor networks, in IPSN 04: Proceedings of the third international symposium on Information processing in sensor networks. New York, NY, USA: ACM Press, 2004, pp [10] P. Bahl and V. Padmanabhan, Radar: an in-building rf-based user location and tracking system, in INFOCOM Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 2, 2000, pp [11] N. Patwari and I. Alfred O. Hero, Using proximity and quantized rss for sensor localization in wireless networks, in WSNA 03: Proceedings of the 2nd ACM international conference on Wireless sensor networks and applications. New York, NY, USA: ACM Press, 2003, pp [12] R. Stoleru and J. Stankovic, Probability grid: a location estimation scheme for wireless sensor networks, in Sensor and Ad Hoc Communications and Networks, 2004, IEEE SECON First Annual IEEE Communications Society Conference on, vol. 4, no. 7, 2004, pp [13] D. Lymberopoulos, Q. Lindsey, and A. Savvides, An empirical analysis of radio signal strength variability in ieee networks using monopole antennas, ENALAB, Yale University, Technical Report, , [14] K. Whitehouse, C. Karlof, A. Woo, F. Jiang, and D. Culler, The effects of ranging noise on multihop localization: an empirical study, in Information Processing in Sensor Networks, IPSN Fourth International Symposium on, 2005, pp [15] J. Sallai, G. Balogh, M. Maroti, and A. Ledeczi, Acoustic ranging in resource constrained sensor networks, Institute for Software Integrated Systems, Technical Reports ISIS , [16] L. Girod and D. Estrin, Robust range estimation using acoustic and multimodal sensing, in Proc. IEEE International Conference on Intelligent Robots and Systems, vol. 3, pp , [17] J. Polastre, R. Szewczyk, and D. Culler, Telos: enabling ultra-low power wireless research, in Information Processing in Sensor Networks, IPSN Fourth International Symposium on, 2005, pp [18] Tmote Sky Datasheet, [19] D. M. Blough and P. Santi, Investigating upper bounds on network lifetime extension for cell-based energy conservation techniques in stationary ad hoc networks, in MobiCom 02: Proceedings of the 8th annual international conference on Mobile computing and networking. New York, NY, USA: ACM Press, 2002, pp