An Object Tracking Scheme for Wireless Sensor Networks Using Data Mining Mechanism Wen-Hwa Liao
Kuo-Chiang Chang
Sital Prasad Kedia
Dept. of Information Management Tatung University Taiwan
[email protected] Dept. of Information Management Tatung University Taiwan
Dept. of Computer Science Purdue University USA
Abstract—Recently, object tracking application of sensor networks has drawn significant attention of the researchers due to its wide application. However, most of these studies cannot deal with the trade-off between energy efficiency and accuracy of the tracking. In object tracking sensor networks (OTSNs), the movement of the object generally follows some definite patterns. The moving object location, arrival time, path are likely to hide some useful association rules, which can be excavated by applying suitable data mining algorithm. In this paper, we have proposed an object tracking scheme for OTSNs using data mining approach. We have improved the Apriori algorithm for mining association rules and made it applicable to the OTSNs. The data mining algorithm is applied to the past movement information of the object and useful association rules are excavated, which are then used to predict the next location of the object. Our scheme predicts the next location of the object more accurately and increases the network lifetime. Experimental results have been conducted to evaluate the performance of our proposed scheme for OTSNs and they show that our scheme outperforms the existing schemes in terms of energy efficiency and accuracy of tracking. Keywords-data mining; object tracking sensor networks (OTSNs); wireless sensor network (WSN).
I.
INTRODUCTION
Wireless sensor network (WSN) is composed of large number of small sized, inexpensive and computable sensors, which are limited in power, memory and computation. Normally, large numbers of tiny sensors are deployed randomly to monitor one or more phenomena, to collect and process the sensed data, and to send the data back to the sink. The important applications of WSN include environmental monitoring, object tracking application, personal healthcare, and enemy monitoring [1][2]. These kinds of applications introduce new challenges for WSN, among which tracking of objects in sensor networks has attracted extensive attentions [3]. In such applications, users are interested in events instead of the identity of a particular sensor. Energy consumption is the primary consideration in designing an object tracking sensor network and most related research works [5][6] focus on reducing energy consumption. The use of a sleeping mechanism is one of the most effective energy saving methods. In object tracking sensor networks (OTSNs), power consumption is a major issue, which directly affects the
c 978-1-4673-0269-2/12/$31.00 2012 IEEE
lifetime of the network. Energy can be conserved in OTSNs by putting a sensor node to sleep when there is no object in the node’s sensing region and turn on a sensor node when an object is to enter the node’s sensing region. Based on this idea, the research for energy conservation in OTSNs can be classified into two types: non-prediction tracking and prediction tracking [4]. In non-prediction tracking, the sensor nodes are periodically put to sleep in order to conserve energy and they wake up after some time to monitor the sensing region and collect the object movement information. On the other hand, prediction based tracking uses the latest detected location of the object, its average speed and direction of travel to predict the next location of the object. By this way, only the sensor nodes at the predicted location of the objects can be active and thus saving significant amount of energy. However, these predictions may not be entirely correct, which leads to loss of object, and large amount of energy may be wasted in recovering the object. In real-world applications, the behavior of the moving objects is often based on certain underlying events, instead of complete randomness. We observe that many creatures such as animals, birds usually form mass organization and migrate together for food. Many recent works have focused on application of data mining technology in OTSNs to discover useful association rules from past movement information of the object. The discovered association rules are then compared with the movement behavior of the object under observation to predict its next location. Use of association rules in predicting the next location of the object helps to improve the performance of prediction and increase the accuracy of tracking. The authors proposed an object tracking algorithm by mining object movement data from its temporal movement patterns [4]. They construct a temporal movement pattern tree from the temporal movement log of the object, which is used to identify the useful association rules and forecast the next location of the object. In another research work, the authors proposed a multi-level wireless sensor networks construction based on K-mean clustering technique [5]. They first conduct the hierarchical clustering to form a hierarchical model for the sensor nodes. Then the movement logs of the objects are analyzed by a data mining algorithm to obtain the movement rules, which are used to predict the next position of the object. The authors proposed a technique for mining region movement patterns based on the three kinds of movements pattern i.e.,
526
node-to-node, node to region and region to region movement pattern [6]. First, a data mining methodology is proposed for discovering the region-based movement patterns of moving objects in OTSNs. Moreover, they also proposed the corresponding prediction strategies for tracking objects in an energy-efficient way. The authors proposed a seamless data mining algorithm named STMP-Mine to efficiently discover the temporal movement patterns of objects in sensor networks without predefining the time segment [7]. They also proposed a location prediction strategy that employs the discovered temporal movement patterns to reduce prediction errors to save energy. The rest of this paper is organized as follows. Section II introduces the network environment of OTSNs. Section III presents our association rule mining for OTSNs. Section IV presents the results of performance evaluation. Section V concludes this paper. NETWORK ENVIRONMENT
II.
The network architecture consists of a large number of sensor nodes deployed in the monitoring region and each node is equipped with sensor memory to record the object movement information as shown in Fig. 1. The object movement information is transmitted to the sink and at the same time each node also maintains its own record table containing the object movement information. Table I shows an example of the table of sensor node 17. From the table we can see that the frequency count of the node 17 is 6, because it has detected six object movements. The probability that an object moves from node 17 to node 23 is 50%, node 21 is 33% and node 20 is 16%. After gathering the object movement information at the sink node for a certain period of time, object movement sequence table is formed at the sink node as shown in Table II. This table contains the sequence of movement of all the objects and it serves as the transaction database for discovering association rule to predict the next location of an object.
TABLE II. Object 1 2 3 4 5 6
III.
TABLE I.
THE TABLE OF SENSOR NODE 17.
Node ID: Node 17 Object ID Time Object 1 13:10-13:30 Object 2 13:50-14:10 Object 3 14:10-14:30 Object 4 14:45-15:15 Object 5 15:30-15:50 Object 6 15:55-16:20
Count 1 1 1 1 1 1
Next node Node 23 Node 20 Node 23 Node 21 Node 23 Node 21
Sequence 1ĺ6ĺ10ĺ14ĺ17ĺ23ĺ26ĺ27 3ĺ8ĺ11ĺ17ĺ20ĺ24ĺ31 2ĺ7ĺ10ĺ14ĺ17ĺ23ĺ26ĺ27 2ĺ6ĺ10ĺ14ĺ17ĺ21ĺ22 3ĺ8ĺ11ĺ17ĺ23ĺ26ĺ27 1ĺ6ĺ10ĺ14ĺ17ĺ21ĺ22
ASSOCIATION RULE MINING FOR OTSNS
In OTSNs, the movement of the object generally follows some definite patterns instead of complete randomness. The moving object location, arrival time, path are likely to hide some useful association rules which can be excavated by applying suitable data mining algorithm. The authors have proposed a data mining algorithm called Apriori algorithm for discovering association rules from a large data base of transaction [8]. The algorithm makes several passes over the transaction database and in each pass it finds the set of items (or itemsets) whose support count is more than the minimum support count. Let’s consider the transaction database D given in Table III. In the first iteration of the algorithm, each item is a member of the set of candidate 1-itemsets, C1. Suppose that the minimum support count required is 2. The set of frequent 1 itemsets L1 is the sets of items in C1, whose number of occurrence in D is more than 2 as shown in Table IV(A). In next pass, the algorithm performs a join of L1 on L1 to obtain the candidate 2-itemsets C2 and the set of frequent 2 itemesets L2 can be constructed by collecting all the items in C2, whose support count is more than 2 as shown in Table IV(B). This procedure is repeated until no more frequent k-item sets can be found. TABLE III. TID T1 T2 T3 T4 T5 T6 T7 T8 T9 TABLE IV.
Figure 1. Network environemnt.
OBJECT MOVEMENT SEQUENCES.
ITEM TRANSACTION DATABASE. Item IDs list I1, I2, I4 I2, I5 I2, I3 I1, I2, I5 I1, I3 I2, I3 I1, I3 I1, I2, I3, I4 I1, I2, I3
APRIORI ALGORITHM. (A) L1 FROM C1. (B) L2 FROM C2.
Item-set I1 I2 I3 I4 I5
C1 Sup_count 6 7 6 2 2
Item-set I1, I2 I1, I3 I1, I4 I1, I5 I2, I3 I2, I4
C2 Sup_count 4 4 1 2 4 2
Item-set I1 I2 I3 I4 I5 (A) Item-set I1, I2 I1, I3 I1, I5 I2, I3 I2, I4 I2, I5
2012 IEEE Network Operations and Management Symposium (NOMS): Short Papers
L1 Sup_count 6 7 6 2 2 L2 Sup_count 4 4 2 4 2 2
527
L5’
2 0 1 0
I2, I5 I3, I4 I3, I5 I4, I5
(B)
We will apply the Apriori algorithm to the past movement logs of objects to find out association rules for movement of the objects. In OTSNs, an association rule is useful only when it follows the object movement sequence. The association rules, which are not part of the object movement sequence, can be discarded as they might not be feasible. Based on the above observation, we can prune the frequent item sets in each iteration, so that it contains only useful entries. For example, the candidate 2-itemsets C2 can be obtained for the transaction data base of Table II. Suppose that the minimum support count required is 2. All the entries in C2, whose support count is less than the minimum support are removed and the remaining entries of C2 form the frequent 2 items-set (L2) which has been shown in Table V(A). All the entries in L2 is then searched in the object movement log in the Table II and the entries which are not found are discarded to form the pruned frequent 2 itemset L2’ as shown in Table V(B). For example, L2 includes the entries 1ė6, 1ė10, 1ė14, and 1ė17. However, since only 1ė6 exists in Table II, L2’ does not includes the entries 1 ė10, 1ė14 and 1ė17. We can apply the similar procedures to get L3’, L4’, L5’, and L6’ as shown in Table VI. TABLE V.
L2 AND L2’ OF OUR ASSOCIATION RULE MINING FOR OTSNS.
Item-set
L2 Sup_count 2 2 2 2 2 2 2 2 2 . . . (A)
1ė6 1ė10 1ė14 1ė17 2ė10 2ė14 2ė17 3ė8 3ė11 . . .
Item-set 1ė6 3ė8 6ė10 8ė11 10ė14 11ė17 14ė17 17ė21 17ė23 21ė22 23ė26 26ė27
Sup_count
1ė6ė10ė14ė17 6ė10ė14ė17ė21 10ė14ė17ė21ė22 10ė14ė17ė23ė26 14ė17ė23ė26ė27
2 2 2 2 2 L6’
Item-set
Sup_count
6ė10ė14ė17ė21ė22 10ė14ė17ė23ė26ė27
2 2
Fig. 2 illustrates how the node 1 decides to predict the next node to track the moving object. Node 1 selects the potential node 6 of L2’ to track the next object move path. The other nodes keep in sleep mode to save energy. The Fig. 3 and Fig. 4 show the node 6 and node 17 to predict the next node in OTSNs, respectively.
Figure 2. Illustration of the node 1 to predict the next node in OTSNs.
L2’ Sup_count 2 2 3 2 4 2 4 2 3 2 3 3 (B)
Figure 3. Illustration of the node 6 to predict the next node in OTSNs.
L3’, L4’, L5’, AND L6’ OF OUR ASSOCIATION RULE MINING FOR OTSNS.
TABLE VI.
L3’
528
Item-set
L4’
Item-set
Sup_count
Item-set
Sup_count
1ė6ė10 3ė8ė11 6ė10ė14 8ė11ė17 10ė14ė17 14ė17ė21 14ė17ė23 17ė21ė22 17ė23ė26 23ė26ė27
2 2 3 2 4 2 2 2 3 3
1ė6ė10ė14 3ė8ė11ė17 6ė10ė14ė17 10ė14ė17ė21 14ė17ė21ė22 14ė17ė23ė26 17ė23ė26ė27
2 2 3 2 2 2 2
Figure 4. Illustration of the node 17 to predict the next node in OTSNs.
IV.
SIMULATION RESULTS
In this section, we evaluate the performance of our proposed method under varying conditions by altering the parameters such as number of sensor nodes and number of
2012 IEEE Network Operations and Management Symposium (NOMS): Short Papers
objects. The simulation parameters have been given in Table VII. We design a simulation model with an object generator, which randomly produces the distribution of sensors. We use x-y coordinate axis to represent the location of a sensor and assume that no distinct sensors are present in the same location. TABLE VII. Symbol N X×Y r E Ptx Preceive Pidle Psleep E Q
SIMULATION PARAMETERS AND SETTING.
Definition No. of sensor nodes Size of the sensing region Transmission range Energy Pw consum. in Tx Pw consum. in Rx Pw consum. in idle listing Pw consum. in sleeping No. of moving object No. of query
Setting 100/200/500/1000 100×100 m2 10 m 4000 mW 50 mW/sec 50 mW/sec 20 mW/sec 5 mW/sec 100 150/250/350/450
Fig. 5(a) shows the comparison of average remaining energy for varying number of sensor nodes from 100 to 1000. We compare our object tracking using data mining scheme (OTDM) with the MLS scheme [5]. As shown in Fig. 5(a), the average remaining energy for our scheme is more than that of MLS scheme. This is because of the fact that lesser number of sensor nodes is active at any time in our scheme than MLS. Fig. 5(b) shows the result of missing rate for varying number of sensor nodes from 100 to 1000. As shown in Fig. 4(b), the missing rate of our scheme is less than that of MLS. The reason is that our scheme uses the data mining approach to predict the next location of the tracking object. Hence, the accuracy of prediction is more in case of OTDM.
(a)
V.
CONCLUSION
In this paper, we have proposed an object tracking scheme for WSNs. We have improved the Apriori algorithm for mining object moving association rules and applied it to the OTSNs. Data mining algorithm is applied to the past movement logs of the object and the useful association rules are excavated to increase the accuracy of the tracking. Experimental results show that our approach outperforms the existing MLS scheme in both of energy-efficiency and accuracy of tracking. Our method can reduce the number of active sensor nodes at any time thus energy consumption in OTSNs can be reduced. In addition, the missing rate of our scheme is less than that of MLS because our approach considers past movement behavior of the object in predicting its next location. ACKNOWLEGEMENTS This work was supported in part by the National Science Council, Taiwan, under grants NSC 99-2221-E-036-036-MY2 and Tatung University, under grants B100-N01-052. REFERENCES [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A Survey on Sensor Networks,” IEEE Communications Magazine, Vol. 40, No. 8, pp. 102-114, 2002. [2] I. F. Akyildiz and E. P. Stuntebeck, “Wireless Underground Sensor Networks: Research Challenges,” Ad Hoc Networks, Vol. 4, No. 6, pp. 669-686, 2006. [3] W.-C. Peng, Y.-Z. Ko, and W.-C. Lee, “On Mining Moving Patterns for Object Tracking Sensor Networks,” IEEE International Conference on Mobile Data Management, 2006. [4] V. S. Tseng, K. W. Lin, “Mining Temporal Moving Patterns in Object Tracking Sensor Networks,” IEEE International Workshop on Ubiquitous Data Management, 2005. [5] V. S. Tseng, E. H. C. Lu, and K. W. Lin, “An Energy-Efficient Approach for Real-Time Tracking of Moving Objects in Multi-Level Sensor Networks,” IEEE International Conference on Embedded and Real-Time Computing Systems and Applications, 2005. [6] V. S. Tseng, M. H. Hsieh, and K. W. Lin, “Mining Region-based Movement Patterns for Energy-Efficient Object Tracking in Sensor Networks,” IEEE International Conference on Intelligent Systems Design and Applications, 2008. [7] K. W. Lin, M. H. Hsieh, and V. S. Tseng, “A Novel Prediction-Based Strategy for Object Tracking in Sensor Networks by Mining Seamless Temporal Movement Patterns,” Expert Systems with Applications, Vol. 37, No. 4, pp. 2799-2807, 2010. [8] R. Agrawal and R. Srikant, “Fast Algorithms for Mining Association Rules,” IEEE International Conference on Very Large Data Bases, 1979.
(b) Figure 5. Simulation results. (a) Average remaining energy v.s. number of sensor. (b) Missing rate v.s. number of sensors.
2012 IEEE Network Operations and Management Symposium (NOMS): Short Papers
529