Energy efficient strategies for object tracking in sensor networks: A data mining approach

(1)

Energy eﬃcient strategies for object tracking in sensor networks:

A data mining approach

q

Vincent S. Tseng

*

_{, Kawuu W. Lin}

Institute of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, ROC Received 8 March 2006; received in revised form 23 October 2006; accepted 14 December 2006

Available online 13 January 2007

Abstract

In recent years, a number of studies have been done on object tracking sensor networks (OTSNs) due to the wide applications. One important research issue in OTSNs is the energy saving strategy in considering the limited power of sensor nodes. The past studies on energy saving in OTSNs considered the object’s movement behavior as randomness. In some real applications, however, the object move-ment behavior is often based on certain underlying events instead of randomness completely. In this paper, we propose a novel data mining algorithm named TMP-Mine with a special data structure named TMP-Tree for efficiently discovering the temporal movement patterns of objects in sensor networks. To our best knowledge, this is the first work on mining the movement patterns associated with time intervals in OTSNs. Moreover, we propose novel location prediction strategies that utilize the discovered temporal movement pat-terns so as to reduce the prediction errors for energy savings. Through empirical evaluation on various simulation conditions and real dataset, TMP-Mine and the proposed prediction strategies are shown to deliver excellent performance in terms of scalability, accuracy and energy efficiency.

Keywords: Location prediction; Temporal movement patterns; Object tracking; Sensor networks; Data mining

1. Introduction

Energy eﬃcient tracking of objects in sensor networks is an emerging research ﬁeld attracting a lot of attention recently. Advances in wireless communication and micro-electronic device technologies have enabled the develop-ment of low-power micro-sensors and the deploydevelop-ment of large scale sensor networks. With the capabilities of perva-sive surveillance, sensor networks are applied in a lot of commercial and military applications, like the object

track-ing application and the environmental data collection. However, the intrinsic limitations such as power con-straints, synchronization, deployment, and data routing bring numerous research challenges (Akyildiz et al., 2002; WINS project).

In a sensor network, the deployed sensor nodes form ad hoc networks (Akyildiz et al., 2002; Hara et al., 2004) and the nodes can communicate with each other by RF radios without special infrastructure. Compared with the stan-dard ad hoc networks, a sensor network has the following characteristics: (1) the sensor nodes are static in terms of physical location; (2) the computing power is normally weak; (3) the energy carried in a sensor node is limited. Due to the environmental conditions that replenishing the battery charge is expensive or infeasible, the energy is one of the most important system resources that should be reserved (Carle and Simplot, 2004). In this paper, we focus on the problem of energy saving in the object track-ing sensor networks (OTSNs).

q

This paper is an extended version ofTseng and Lin (2005), entitled ‘‘Mining Temporal Moving Patterns in Object Tracking Sensor Networks’’, by V. S. Tseng and K. W. Lin, which appeared in Proceedings of the International Workshop on Ubiquitous Data Management (held with ICDE’05), April, 2005, Tokyo, Japan.

*

Corresponding author. Tel.: +886 6 2757575x62536; fax: +886 62747076.

E-mail address:tsengsm@mail.ncku.edu.tw(V.S. Tseng).

(2)

In an OTSN, each sensor node is composed of sensing, data processing, and communication components ( Raghu-nathan et al., 2002). Nevertheless, the power required by diﬀerent sensing components varies widely. For example, using a velocity-based strategy to track the moving objects requires the velocity sensing component, which is an energy expensive device and is not the necessary equipment for all sensor nodes. Hence, one of our research goals is to pro-pose energy eﬃcient strategies by using intelligent soft-ware mechanism instead of adding the energy expensive components.

A number of past studies tried to solve the energy saving issue from the hardware design. For instance, the optimiza-tion problem of the communicaoptimiza-tion cost by inactivating the RF radios of idle sensor nodes was widely discussed (Goel and Imielinski, 2001; Heinzelman et al., 2000). However, these studies did not consider the energy saving issues for these components (Xu et al., 2004) although the sensing and computing components consume relative less energy than radios (Raghunathan et al., 2002). Several researchers tried to save the energy through the software approach like scheduling of sensors. One of the novel ideas is to put a sensor node into sleeping mode when there are no objects in its coverage/sensing region, and a sensor node is acti-vated again whenever an object enters its sensing region. Based on this idea, the studies for energy saving in OTSNs can be further divided into two categories: non-prediction based tracking and prediction based tracking. The intuitive way of non-prediction based tracking method is periodi-cally turn the sensor nodes oﬀ and only activate the sensor nodes when it is time to monitor their sensing regions. Another non-prediction based tracking method is planting an agent onto the mobile device, named mobile agent. With the help of mobile agents, the communication and sensing overheads can be greatly reduced (Tseng et al., 2004). The prediction based methods use the information of a moving object like velocity or moving direction to predict the next location the object might visit.

Note that both of the non-prediction based and predic-tion based tracking methods neglected the event character-istic of objects. In some real applications, the behavior of the moving objects is often based on certain underlying events instead of randomness completely. For example, consider the bus tracking project inMani (2003), the route of each bus is pre-speciﬁed rather than being random. Cen-tral to this issue is the problem of discovering the move-ment behavior of objects. The wireless technologies nowadays have allowed the collection of large amount of movement logs (CRAWDAD; Reality Mining Project). Therefore, it is feasible to discover the hidden knowledge like movement behavior from the wireless log. Over the past few years a considerable number of studies have been done on using data mining techniques to discover this kind of interesting patterns/rules from World Wide Web (Pei et al., 2000), transactional databases (Agrawal and Srikant, 1995) and mobility databases (Huang et al., 2003; Kyriaka-kos et al., 2003; Tseng and Lin, 2006; Tseng and Tsui,

2004). Note that the discovered patterns in such applica-tions are implicitly assumed to be valid for some period until the mobility patterns change with time. To keep the patterns being updated, the data mining techniques may be applied on the most updated log periodically. Most of these past studies focused only on the aspect of path anal-ysis and only few of them (Wu et al., 2001) considered the temporal characteristic that is very critical in wireless net-works. Without considering the temporal information, the important knowledge may be overlooked (Roddick and Spiliopoulou, 2002).

Take a vehicle tracking application as example. Assume that each car is attached with a receiver that can receive the beacon of the sensor node the car visits. By collecting the log of cars, we may use the data mining method to discover temporal movement rules. Suppose the following rules are discovered: Rule1: (Station A! interval 10 min ! Station

B! interval 5 min ! Station C); Rule2: (Station A!

interval 20 min! Station B ! interval 5 min ! Station D). By dispatching these rules to the corresponding sensor nodes, the tracking can be mode in more energy eﬃcient way. For instance, if a car moves with the pattern as (Sta-tion A! interval 10 min ! Station B ! interval 5 min) that matches with Rule1, the node in Station B has only

to activate the node in Station C rather than that in Station D or those around Station B. As can be seen, the temporal clues can eﬀectively enhance the prediction accuracy in an OTSN. By integrating the temporal movement patterns (TMPs) into the prediction strategies, the number of sensor nodes that are incorrectly and unnecessarily activated is expected to be substantially reduced and more energy can be saved in an OTSN.

However, no studies have explored the issue of discover-ing objects’ temporal movement patterns in OTSNs so as to enhance the energy efficiency. In this paper, we propose a novel data mining method named TMP-Mine with a special data structure named TMP-Tree for efficiently discovering TMPs in OTSNs. To our best knowledge, this is the first work on mining the movement patterns with time intervals in OTSNs. Moreover, we propose two prediction strategies for predicting the location of a missing object in OTSNs by utilizing TMPs. The first prediction strategy named PTMP is capable of making prediction by employing TMPs with no need to detect the object velocity. Hence, it can be applied to the sensor networks with low-end sensor nodes. The second strategy, namely PES + PTMP, is a hybrid approach by integrating PTMP method with a popular velocity-based strategy named PES (Xu et al., 2004). This integrated strategy can further enhance the energy efficiency if the sensor nodes carry the velocity detection capability. Through empirical evaluation on various simulation condi-tions and real dataset, TMP-Mine and the proposed predic-tion strategies are shown to deliver excellent performance in terms of scalability, accuracy and energy efficiency.

The rest of this paper is structured as follows. We brieﬂy review the related work in Section 2. In Section 3, we describe the overall system architecture and workﬂow. In

(3)

Section4, we describe the problem deﬁnition and the pro-posed data mining algorithm, namely TMP-Mine, is pre-sented in Section 5. Section 6 gives the detailed description on the prediction strategies. The empirical eval-uation for performance study is made in Section 7. The conclusions and future directions are given in Section8. 2. Related work

In this section, we review the past studies on the three subjects closely related to this research, namely energy eﬃ-cient strategies for object tracking, behavior mining and behavior prediction.

For energy saving policies in sensor networks, a number of past studies tried to solve this issue from the aspect of hardware design. For instance, the optimization problem of the communication cost by inactivating the RF radios of idle sensor nodes was widely discussed (Goel and Imielinski, 2001; Heinzelman et al., 2000). There are also a lot of research efforts in energy efficient media access con-trol (MAC) (Shih et al., 2001; Woo and Culler, 2001; Ye et al., 2002). Several researches tried to save the energy through the software design approach. In Cerpa et al. (2001), the authors proposed the Frisbee scheme that keeps only a limited zone of the network called a Frisbee that is close to the moving object in the fully active state. How-ever, it is difficult to choose a good radius of the Frisbee. InLin et al. (2006), the authors developed some tree struc-tures for efficient object tracking by considering the physi-cal network structure. In Xu et al. (2004), Xu et al. proposed a network model, in which a sensor node is acti-vated only if there exist some objects in its coverage region. Besides, the activated node is scheduled to be active for X seconds and in sleeping mode for (T X) seconds during the T seconds periodically to save the energy. They also proposed a set of prediction based energy saving schemes named PES to conserve the scarce energy resource by using the latest detected or average velocity of a missing object to predict its current location. To select the object velocity and direction, three models named Heuristics INSTANT, Heuristics AVERAGE, and Heuristics EXP_AVG were also proposed. In the prediction phase, three mechanisms were proposed, namely Heuristics DESTINATION, Heuristics ROUTE, and Heuristics ALL_NBR. The Heuristics DES-TINATION utilizes only the velocity information for acti-vation while the Heuristics ROUTE activates all nodes on the route. The Heuristics ALL_NBR mechanism activates all neighboring nodes of the destination. However, the lat-est detected velocity of objects may be incorrect since the sensor node might lose the object in its periodical sleeping mode. Hence, the PES method still incurs the problem of incorrect prediction.

For the research on behavior mining, numerous studies have been done on mining users’ behavior patterns like association rules or sequential patterns in WWW (Pei et al., 2000) and transactional databases (Agrawal and Srikant, 1995; Heinzelman et al., 2000). In Pei et al.

(2000), the authors proposed a method named WAP-Mine for fast discovery of the web access patterns from web logs by using a tree-based data structure without candidate generation. Previous studies on the mining of temporal databases include (Agrawal and Srikant, 1995; Ale and Rossi, 2000; Guil et al., 2004; Li et al., 2003; Padmanabhan and Tuzhilin, 1996; Srikant and Agrawal, 1996). In Agra-wal and Srikant (1995), the authors proposed a method for mining the transactions to discover the time-ordered patterns named sequential patterns. InSrikant and Agra-wal (1996), the method using sliding window to restrict the time gap between sets of transactions in mining sequence patterns was proposed. The issue of using the temporal logic and related operators such as since, until and next was explored in Padmanabhan and Tuzhilin (1996). In the category of mobility mining, most of the existing researches focused only on the analysis of user movement behavior Lee and Wang (2003), Yavas et al. (2005). To discover the patterns from two-dimensional mobility data, the problem of mining location associated service patterns was ﬁrst studied by Tseng and Tsui (2004). A novel method for discovering users’ sequential movement patterns associated with requested services in mobile web systems was also proposed byTseng and Lin (2006).

In the area of behavior prediction, some researchers pro-posed variations of Markov models, such as Dependency Graph (DG) Padmanabhan and Mogul (1996), Predic-tion-by-Partial-Match (PPM) Palpanas and Mendelzon (1999)and N-gram model Su et al. (2000), for predicting the user behavior in WWW. Basically, these methods employ the last N page views of the user to predict the next view by using the patterns discovered. Yang et al. (2004)

studied the association-rule based sequential classiﬁers and considered features of association rules such as order, adjacency, and recency systematically to construct predic-tion models from web logs.

3. Problem statement

In this section, we ﬁrst state the problem. Afterwards, we describe the network environments and the behavior issues of moving objects. The performance metrics are also described in the end of this section.

In this work, we adopt a network model for OTSNs as proposed inXu et al. (2004), in which a sensor node is acti-vated only if there is object in its coverage/sensing region. Besides, the activated node is scheduled to be in active mode for X seconds and in sleeping mode for (T X) sec-onds during the T secsec-onds periodically to save the energy. Moreover, we assume the movement log of objects is col-lectable (Mani, 2003; Tseng and Lin, 2005) and the trajec-tory of each object is represented in the form of S =h(l1, t1)(l2, t2), . . . , (ln, tn)i, where lirepresents the sensor

node location at time ti. The log is considered as a valuable

resource since it contains the habitual patterns of objects. The targeted problem is to two fold: (1) eﬃcient discovery

(4)

of temporal movement patterns (TMPs) for objects, and (2) location prediction by utilizing TMPs for energy saving.

To solve the problem described above, we shall discover TMP in the form as P =h(l1, i1, l2, i2, . . . , ir1, lr)iwhere ik

semantically represents the time interval between two tra-versed locations. Moreover, we shall generate temporal movement rule (TMR) in the form of

Rt¼ hðl1; i1; l2; i2; . . . ; lm1; im1Þi ! hðlmÞi

for incorporation into the location prediction mechanisms so as to achieve low energy and low missing rate in the OTSNs.

Note that we assume the behavior of moving objects is often based on certain underlying events instead of ran-domness completely (Tseng and Lin, 2005, 2006; Tseng and Tsui, 2004; Yavas et al., 2005). An event is a stream of locations with time intervals. Note that the characteris-tics of events in OTSNs include not only location but also time interval. In our network model, the movement behav-ior of objects may be decided by certain events or be ran-dom. Detailed network model will be given in Section7.1. In solving the targeted problem, some important perfor-mance metrics should be considered. In this work, we adopt two popular metrics named Total Energy Consumed (TEC) (Xu et al., 2004) and missing rate (Xu et al., 2004). TEC indicates the total energy consumed by sensor nodes in the OTSN during data mining phase and object tracking phase. Missing rate denotes the number of erroneous pre-dictions in a speciﬁed time period in ratio of the total num-ber of movement of objects. Obviously, low TEC and low missing rate can beneﬁt the lifetime of the whole network, and this is the aimed goal for this research.

4. System architecture

Fig. 1 shows the proposed system architecture. We assume that the movement of objects in wireless sensor networks is recorded in the system logs (Mani, 2003; Tseng and Lin, 2005). In our proposed network model, each

object is able to record the sensor nodes it visited together with the arrival time at each node. To collect the movement log, several powerful sensor nodes equipped with storage devices are deployed over the outer of the network for retrieving the log of each object that is exiting from the net-work. The system workflow consists of two main phases: (1) data mining phase, and (2) object tracking phase. At first, the sensor network collects and integrates the move-ment log of moving objects. Then the integrated movemove-ment log is used as the input to the data mining method named TMP-Mine for discovering the TMPs. By performing the proposed TMP-Mine algorithm, the TMPs will be discov-ered and then the TMRs are generated for use by location prediction so as to track objects in energy efficient manner. The two phases are described in details in below:

1. Data mining mechanism: The data mining mechanism consists of three components, namely data integrator, TMP-Miner and TMR generator. Because the logs are distributed in the surrounding sensors of the network, a data integrator is required to integrate and preprocess the logs collected before the data mining algorithm is applied on the logs.Table 1is an example showing the prepared movement log with time intervals between visiting to sensor nodes. Take the fourth tuple, {(f, 0)(e, 5)(b, 13)}, as example, it means the object arrived in the region of sensor node f, e and b at time point 0, 5 and 13, respectively.Once the log is prepared, TMP-Mine algorithm will be applied to discover the TMPs from the integrated log. The functionality of rule generator is to generate the TMRs from discovered TMPs according to some parameters like conﬁdence. Afterwards, the TMRs are utilized by the location prediction strategies so as to achieve the goal for energy savings. Moreover, the rule generator will also evaluate the strength of each TMR for rule ranking (described in details in Section5.5). 2. Object tracking mechanism: The spirit of the proposed tracking mechanism is to predict the next location of each object by utilizing the TMRs. Before activating

(5)

the object tracking mechanism, we should dispatch the TMRs to appropriate sensor nodes. In considering the property that the storage associated with each sensor node is limited, we dispatch the TMRs to the sensor nodes according to the location-based criterion. That is, only the TMRs that are location related to a sensor node will be loaded into that node. Take the TMR, (f, 5)! (b), for example, we would deploy this TMR at only node f and its neighboring sensors rather than all the sensors in the network. Dispatching TMRs by the location-based criterion as described above will greatly decrease the demands of storage for the sensor nodes.The tracking mechanism is composed of the loca-tion predicloca-tion strategy and object recovering method. For location prediction strategies, we propose two approaches named PTMP and PES + PTMP. PTMP performs location prediction by employing TMRs with no need to detect the object velocity. Hence, it can be applied to the sensor networks with low-end sensor nodes. The approach PES + PTMP is a hybrid one by integrating PTMP method with a popular velocity-based strategy named PES (Xu et al., 2004). Recall that a sensor node is periodically activated when an object is in its coverage region according to the scheduling policy. Under such environments, the prediction mechanism will be triggered whenever a sensor node loses an object. If the prediction mechanism fails to recover the object within a speciﬁed deadline, the ﬂooding (Cerpa et al., 2001) strategy will be activated for recovering the miss-ing object.

5. Proposed data mining algorithm: TMP-mine

In this section, we ﬁrst formulate the mining problem and then propose a novel algorithm named TMP-Mine that can discover the TMPs eﬃciently. How the TMRs are generated is also described. We illustrate the discovery of TMPs by an elaborate example in the end of this section. 5.1. Formulation of mining problem

Let S =h(l1, rt1)(l2, rt2) . . . (lm, rtm)i be a temporal

move-ment sequence of an object with length equal to m, where li

represents the object’s location at time rti and rti<

rti+1"1 6 i < m. The ascending order of elements in a

sequence is decided by using the time as the key. In order to discover the temporal movement patterns, we use the time slot to uniformly segment the time dimension of a sequence. If the time slot is set to b, we will obtain a transformed sequence St=h(l1, t1)(l2, t2). . . . , (lm, tm)i, where ti¼ rtbi

. Deﬁnition 1. A temporal movement sequence S0¼ hðl0₁; t0₁Þðl0₂; t0₂Þ; . . . ; ðl0m; t0mÞi is a sub-pattern of another

access pattern S =h(l1, t1)(l2, t2), . . . , (ln, tn)i, written as

S0_{S, if m 6 n and there exists a strictly increasing}

sequence (k1, k2, . . . , km) of indices such that for all j¼

1; 2; . . . ; m; l0j¼ lkj and t 0

jþ1 t0j¼ tkjþ1 tkj. Here, S is called the super-pattern of S0.

Deﬁnition 2. Given a database D = {S1, S2, . . . , SN} that

contains N temporal movement sequences, the support of sequence S is deﬁned as

supðSÞ ¼jfSijS Siand 1 6 i 6 Ngj

N :

Deﬁnition 3. S =h(l1, t1)(l2, t2), . . . , (lr, tr)i is called a

frequent temporal movement sequence if sup(S) is greater than or equal to a speciﬁed support threshold d, and the corresponding TMP is written as P =h(l1, i1, l2, i2, . . . , ir, lr)i,

where iksemantically represents the time interval between

lkand lk+1visiting and ij= tj+1 tj.

With the above definitions, the problem of discovering TMPs is defined as follows. Given a database D containing the temporal movement sequences of objects and a speci-fied support threshold d, the problem is to discover all the TMPs existing in this database. In this research, we propose a new algorithm named TMP-Mine for discover-ing the TMPs. TMP-Mine uses a special data structure called TMP-Tree to achieve high efficiency in mining process.

5.2. TMP-Tree construction

In order to discover the TMPs efficiently, it is required to construct a TMP-Tree in advance. The purpose of con-structing TMP-Tree is to aggregate the temporal move-ment sequences into the memory in a compact form so that the mining of frequent patterns can be done efficiently. The main advantages of TMP-Tree are (1) only one phys-ical database scan is needed to mine all of the frequent pat-terns, and (2) the TMP-Tree is compact so that the huge amount of data can be handled efficiently.

Each node in TMP-Tree is termed as location node (LNode) since it semantically represents the location (i.e. the sensor node) an object traversed. Each LNode of TMP-Tree has the following structure:

LNode :¼ fl; c; parent-link; next-link; childrentable; ITreeg The label of LNode, namely LLabel, is stored in the var-iable l, and the number of traversed times for the LNode is stored in the variable c. The parent-link is a pointer linking Table 1

An integrated log of temporal moving sequences

Object ID Temporal moving sequence

1 h(a, 1)(e, 3)(c, 5)(b, 10)i

2 h(a, 3)(b, 5)(c, 7)(d, 12)i

3 h(a, 1)(e, 2)(c, 5)(b, 10)i

4 h(f, 0)(e, 5)(b, 13)i

5 h(a, 4)(b, 6)(c, 7)(d, 12)i

6 h(f, 0)(a, 4)(c, 6)(d, 10)i

7 h(a, 0)(b, 1)(c, 2)(d, 6)i

(6)

to the parent LNode and the next-link is a pointer linking to the next LNode with the same LLabel as that of the cur-rent LNode. All of the children LNodes of the curcur-rent LNode are tabulated in the children table. Each LNode is associated with an interval tree named ITree for record-ing the temporal information.

Fig. 2shows the TMP-Tree construction function. The input to this function is the temporal movement log and the function returns the TMP-Tree after inserting every tuple from the log into the TMP-Tree. In the beginning of the construction, TMP-Tree T is initialized (line 1). Then, the tuples are retrieved from the log one by one (line 2), and the location path named LPath and interval path name IPath are extracted from each tuple (line 3 and line 4). Afterwards, the extracted LPath is inserted into the TMP-Tree and the function returns the last LNode of T that corresponds to the last LLabel of the LPath. The func-tion then inserts IPath into the ITree on the returned LNode.

Besides, a LLabel table is maintained together with a TMP-Tree to record the total visited times for each LLabel/sensor node and the last-inserted-LNode for each

LLabel. The logical structure for each tuple of LLabel table is represented as below:

LLabel :¼ fl; c; last-inserted-LNodeg

Fig. 3shows the function for inserting an LPath into a TMP-Tree with specified count. The function first fetches the root node of T and stores it into a temporary node lnode, and LLabel table lt (line 1). Then, the function dis-cretizes the LPath into an array and inserts each label into TMP-Tree in order (line 2). For a LLabel l, if the lnode has a child with label = l (line 3), the children table will be looked up and the entry with label = l is assigned to lnode (line 4). The count of lnode is also increased by the specified count (line 5). If lnode has no child with LLabel = l, mean-ing that the insertmean-ing node is a new LNode on the LPath, a new LNode will be created with LLabel = l (line 7) and the count of lnode is set as the specified count (line 8). More-over, the last-inserted-LNode pointer and TMP-Tree struc-ture (line 9–line 11) will be updated so that we can keep track of all LNodes with a specified LLabel via the LLabel table. In the final step, the count of current LLabel in LLa-bel table is increased (line 13), and the function returns the last LNode on the LPath (line 15).

Fig. 4 shows the procedure for inserting an IPath into the ITree on LNode. The node of ITree is termed as inter-val node, namely INode, for it semantically represents the time interval between the traversed sensor nodes. The cor-responding label of INode is termed as interval label or ILabel. Each INode has the following structure:

INode :¼ fl; c; parent-link; peer-link; childrentableg

The insertion procedure is similar to the LPath inser-tion, except that this procedure uses peer-link structure instead of next-link for connecting each INode (line 9). The peer-link is a pointer that points to the next INode Input: the temporal movement log D

Output: TMP-Tree T Method: TMP_Tree_Construct(D) 1. T 2. FOREACH tuple in D 3. lp ExtractLPath ( ) 4. ip ExtractIPath ( ) 5. lnode InsertLPath(T, lp, 1) 6. InsertIPath(lnode, ip, 1) 7. END FOREACH 8. RETURN T Φ

Fig. 2. TMP-Tree construction function.

Input: a constructed TMP-Tree T, LPath (Location Path) lp, and the count for the LPath c Output: the last LNode on lp

Method: InsertLPath(T, lp, c)

1. lnode root(T), lt getLLabelTable(T)

2. FOR i = 1 to length(lp)

3. IF (getChildren(lnode, lp[i]))≠Φ) //the node already exists

4. lnode getChildren(lnode, lp[i]))

5. Increase the count of lnode by c 6. ELSE

7. lnode InsertLChild(lnode, lp[i])

8. set the count of lnode to c

9. tmpnode getLNode(lp[i], lt)

10. setNextLink(lnode, tmpnode)

11. Set the last-inserted-LNode to lnode

12. ENDIF

13. Increase the count of lp[i] in LLabel table lt by c 14. END FOR

15. RETURN lnode

(7)

on the same height of the ITree. By keeping track of the peer-link, it is easy to sum the counts of ILabels on the same height of ITree.

5.3. TMP-Mine algorithm

Fig. 5 shows the detailed algorithm for TMP-Mine, which takes a depth-ﬁrst search (DFS) approach like WAP-Mine (Pei et al., 2000). While WAP-Mine was designed for mining of single dimensional sequential pat-terns, the proposed TMP-Mine algorithm can manipulate two-dimensional patterns including location and time attri-butes simultaneously. The algorithm recursively constructs the TMP-Trees and mines the TMP-Trees until the termi-nation condition is met. First, we list all the LLabels with count greater than the support threshold d by scanning the LLabel table of current TMP-Tree, and the labels are stored into a temporary set (line 1). If the set is empty (line 2), meaning that no more reconstruction for TMP-Tree is needed, the preﬁx pattern of current TMP-Tree is output as one of the TMPs (line 3). Otherwise, for each frequent LLabel llabel, we fetch the LNodes with LLabel = llabel from current TMP-Tree into a temporary set LNode_tmp by tracking the last-inserted-LNode pointer (line 5 and line 6).

Under the preﬁx llabel, we accumulate the count of each distinct LLabel from the ancestor LPaths of LNodes in LNode_tmp (line 7). The function, getFrequentLIPair( ), returns the set of frequent location-interval pair (LIPair) where each LIPair in this set is with the count greater than the support threshold d. We use the set FreqLIPair to record the frequent LIPair (line 8). If the FreqLIPair is empty (line 9), meaning no more reconstruction is needed, the preﬁx pattern is output as one of the TMPs and the procedure returns (line 10). Otherwise, we reconstruct the TMP-Trees for each frequent LIPair in FreqLIPair (line 13), and the mining procedure is invoked recursively (line 14) to discover all of the TMPs.

5.4. TMP-Tree reconstruction

As described in Section5.3, the mining process of TMP-Mine requires recursive reconstruction for the TMP-Trees.

Fig. 6shows the TMP-Tree reconstruction algorithm. The algorithm begins by initializing a TMP-Tree T0_{(line 1). For}

each lnode in LNode_tmp (line 2), we get its cross-peer nodes by ipreﬁx from the ITree (line 3). In a TMP-Tree, an LLNode and an INode are in cross-peer relation if and only if they are of the same height. Note that the LNodes in LNode_tmp have the same LLabel (referred to line 12–line 15 in TMP-Mine algorithm) and the variable ipreﬁx is the interval part of the lipair. For example, the interval part of a LIPair (La, 10) is 10. All of the cross-peer

nodes with ILabel = ipreﬁx will be returned with the count of them summed up (line 4). Then, the function InsertL-Path( ) is invoked to insert the new LPath with the sum as the count, and the last LNode of the LPath is returned for later insertion (line 5). Afterwards, the IPaths, whose last INode’s ILabel is equal to the current ILabel, i.e. ipre-ﬁx, will be inserted into the returned LNode (line 6–line 8). After all of the LNodes in LNode_tmp are processed, the function returns the TMP-Tree T0_{(line 10).}

5.5. Temporal movement rules

For a discovered TMP, Pt=h(l1, i1, l2, i2, . . . , lm)i, the

form of the corresponding TMR Rtand the deﬁnitions of

conﬁdence conf(Pt) and strength strength(Pt) are given as:

Rt¼ hðl1; i1; l2; i2; . . . ; lm1; im1Þi ! hðlmÞi ð1Þ

confðPtÞ ¼

supðhðl1; i1; l2; i2; . . . ; lmÞiÞ

supðhðl1; i1; l2; i2; . . . ; lm1; im1ÞiÞ

100% ð2Þ We term the last location of antecedent, namely lm1, as

LLocation. Besides, in order to reveal the strength of each rule, each rule is ranked by the following formula that con-siders both of support and conﬁdence:

strengthðRtÞ ¼ supðPtÞ conf ðPtÞ ð3Þ

Input: LNode (Location Node) lnode, IPath (Interval Path) ip, and the count for the IPath c Method: InsertIPath(lnode, ip, c)

1. inode ITree_root(lnode), it getILabelTable(lnode)

2. FOR i = 1 to length(ip)

3. IF (getChildren(inode, ip[i]))≠ Φ) //the node already exists

4. inode getChildren(inode, ip[i]))

5. Increase the count of inode by c 6. ELSE

7. inode InsertIChild(inode, ip[i])

8. set the count of inode to c

9. tmpnode getINode(i, it)

10. setPeerLink(inode, tmpnode)

11. set the last-inserted-LNode to inode

12. ENDIF

13. Increase the count of ip[i] in ILabel table it by c 14. END FOR

(8)

Since a large number of rules could be generated, most of traditional data mining methods need a function utiliz-ing hashutiliz-ing tables (Su et al., 2000) or hashing trees ( Agra-wal and Srikant, 1995) to accelerate the rule access. However, we do not need any accelerating function for accessing the rules. As described in Section4, the rules will be deployed over the networks based on the location-related criterion. Therefore, dispatching TMRs to sensors by LLocation of each TMR requires only one scan over the physical rule repository. Take the antecedent h(l1, i1, l2, i2, . . . , lm1, im1)i of a TMR for example. Since

the LLocation of the antecedent is lm1, the sensors to load

this rule are those within the neighboring radius of lm1.

Considering that the rule that has been dispatched will not be used again in the future, no accelerating function is needed in our application.

In Section7, we will show through experimental results that ranking rules by strength instead of support or conﬁ-dence can save more energy. Moreover, if two or more rules have the same strength value, the rule with larger con-ﬁdence will be given higher priority over other rules. 5.6. An elaborate example

We illustrate the process of discovering TMPs by an elaborate example.Fig. 7shows the TMP-Tree constructed Input: a TMP-Tree T, the prefix pattern PfPtn, the LIPair (Location-Interval Pair) lipair, the

LNodes (Location Node) with the same LLabel (Location Label) LNode_tmp Output: a reconstructed TMP-Tree T’

Method: Reconstruct_TMP_Tree(T, PfPtn, lipair, LNode_tmp) 1. T’

2. FOREACH lnode in LNode_tmp

3. iprefix getIntervalPart(lipair)

INode_tmp getCrossPeerNodesByILabel(lnode, iprefix)

4. c sum the count

5. lp getLPath(lnode)

endnode InsertLPath(T’, lp, c)

6. FOREACH inode in INode_tmp

7. ip getIPath(inode), icount get the count of the inode

InsertIPath(endnode, ip, icount)

8. END FOREACH

9. END FOREACH 10. RETURN T’

Φ

Fig. 6. Algorithm for TMP-Tree reconstruction.

Input: a constructed TMP-Tree T, a specified support , and the prefix pattern PfPtn Output: all of the frequent TMPs

Method: TMP_Mine(T, , PfPtn) 1. lt LabelTable(T)

FreqL getFrequentLabel(lt, )

2. IF (FreqL == )

3. output prefix pattern PfPtn and RETURN 4. ENDIF

5. FOREACH llabel in FreqL

6. LNode_tmp getNodesByLLabel(llabel)

7. FreqAncestorLabels getFrequentAncestorLabels(LNode_tmp)

8. FreqLIPair getFrequentLIPair(LNode_tmp, FreqAncestorLabels)

9. IF (FreqLIPair == )

10. output prefix pattern PfPtn and RETURN 11. ENDIF

12. FOREACH lipair in FreqLIPair

13. T’ Reconstruct_TMP_Tree(T’, PfPtn, lipair, LNode_tmp)

14. newPfPtn generate new prefix pattern

TMP_Mine(T’, , newPfPtn) 15. END FOREACH 16. END FOREACH Φ Φ δ δ

(9)

in OTSNs. Through empirical evaluation and sensitivity analysis under various system conditions, TMP-Mine is shown to perform excellently in terms of execution eﬃ-ciency and scalability.

In the aspect of prediction strategies, we propose a pat-tern-based prediction strategy named PTMP and a hybrid strategy named PES + PTMP integrating the PES method with PTMP. The pure pattern-based prediction strategy works with no need to detect the object velocity; hence, it can be applied to the sensor networks with low-end sensor nodes. The hybrid strategy that exploits both the informa-tion of object velocity and movement patterns was shown to outperform PTMP and PES in terms of the energy con-sumption in an OTSN. Therefore, the hybrid strategy serves as an excellent mechanism for OTSNs in which the sensors are equipped with velocity detection ability. To adapt to the limited storage and weak computation ability of sensor nodes, a rule dispatching mechanism is also devised by complying the location-based criterion. Through experimental evaluation, it is shown that ranking rules by strength criteria delivers better results in terms of TEC and missing rate than that by using conﬁdence or support. For the future work, we will apply TMP-Mine on more real datasets and also evaluate the performance of the pro-posed prediction strategies. Besides, since the discovered TMPs can be exploited in wide applications, we will apply the TMP-Mine method on applications like data dissemi-nation and vehicle monitoring, with the aim to enhance the quality of new applications in sensor networks. Acknowledgements

This research was partially supported by Ministry of Economic Aﬀairs, R.O.C., under grant no. 92-EC-17-A-02-51-024 and by National Science Council, R.O.C., under grant no. NSC 93-2213-E-006-030.

References

Agrawal, R., Srikant, R., 1995. Mining sequential patterns. In: Proceed-ings of the 11th International Conference on Data Engineering, pp. 3– 14.

Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E., 2002. Wireless sensor networks: a survey. Computer Networks 38 (4), 393– 422.

Ale, J.M., Rossi, G.H., 2000. An approach to discovering temporal association rules. In: Proceedings of the 2000 ACM Symposium on Applied Computing. pp. 294–300.

Carle, J., Simplot, D., 2004. Energy-eﬃcient area monitoring for sensor networks. IEEE Computer 37 (2), 40–46.

Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M., Zhao, J., 2001. Habitat monitoring: application driver for wireless communications technology. In: Proceedings of the First ACM SIGMOMM Workshop on Data Communications in Latin America and the Caribbean. CRAWDAD Project.http://crawdad.cs.dartmouth.edu/index.php. Eagle, N., Pentland, A., 2005. Reality mining: sensing complex social

systems. Personal and Ubiquitous Computing 10 (#4), 2006. Goel, S., Imielinski, T., 2001. Prediction-based monitoring in sensor

networks: taking lessons from MPEG. ACM Computer Communica-tion Review 31 (5).

Guil, F., Bosch, A., Marin, R., 2004. TSET MAX: an algorithm for mining frequent maximal temporal patterns. In: Proceedings of the Workshop on Temporal Data Mining: Algorithms, Theory and Applications (TDM’04), pp. 71–77.

Han, J., Kamber, M., 2000. Data mining: Concepts and Techniques. Morgan Kaufman Publishers, ISBN 1-55860-489-8.

Hara, T., Murakami, N., Nishio, S., 2004. Replica allocation for correlated data Items in ad hoc sensor networks. SIGMOD Record 33 (1), 38–43.

Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H., 2000. Energy-eﬃcient communication protocol for wireless microsensor networks. In: Proceedings of the 33rd Hawaii International Conference on System Sciences.

Huang, J.L., Chen, M.S., Peng, W.C., 2003. Exploring group mobility for replica data allocation in a mobile environment. In: Proceedings of the ACM International Conference on Information and Knowledge Management, pp. 161–168.

Kyriakakos, M., Hadjiefthymiades, S., Frangiadakis, N., Merakos, L.F., 2003. Multi-user driven path prediction algorithm for mobile comput-ing. In: Proceedings of 14th International Workshop on Database and Expert Systems Applications (DEXA’03), pp. 191–195.

Lee, J.T., Wang, Y.T., 2003. Eﬃcient data mining for calling path patterns in gsm networks. Information Systems 28 (8), 929–948.

Li, Y., Ning, P., Wang, X.S., Jajodia, S., 2003. Discovering calendar-based temporal association rules. Data and Knowledge Engineering 44, 193–218.

Lin, C.Y., Peng, W.C., Tseng, Y.C., 2006. Eﬃcient in-network moving object tracking in wireless sensor networks. IEEE Transaction on Mobile Computing 5 (8).

Mani, M., 2003. Understanding the semantics of sensor data. ACM SIGMOD Record 32 (4).

Padmanabhan, V., Mogul, J., 1996. Using predictive prefetching to improve world wide web latency. ACM Computer Communication Review 26 (3).

Padmanabhan, B., Tuzhilin, A., 1996. Pattern discovery in temporal databases: a temporal logic approach. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pp. 351–354.

Palpanas, T., Mendelzon, A., 1999. Web prefetching using partial match prediction. In: Proceedings of the Fourth Web Caching Workshop. Pei, J., Han, J., Mortazavi-Asl, B., Zhu, H., 2000. Mining access patterns

eﬃciently from web logs. In: Proceedings of the Fourth Paciﬁc Asia Conference on Knowledge Discovery and Data Mining, pp. 396–407. Raghunathan, V., Schurgers, C., Park, S., Srivastava, M.B., 2002. Energy aware wireless microsensor networks. IEEE Signal Processing Mag-azine 19 (2), 40–50.

Reality Mining Project.http://reality.media.mit.edu/.

Roddick, J.F., Spiliopoulou, M., 2002. A survey of temporal knowledge discovery paradigms and methods. IEEE Transactions on Knowledge and Data Engineering 14 (4), 750–767.

Shih, E., Cho, S., Ickes, N., Min, R., Sinha, A., Wang, A., Chandrakasan, A., 2001. Physical layer driven protocol and algorithm design for energy-eﬃcient wireless sensor networks. In: Proceedings of Seventh ACM International Conference on Mobile Computing and Network-ing (Mobicom’01), pp. 272–287.

Srikant, R., Agrawal, R., 1996. Mining sequential patterns: generalizations and performance improvements. In: Proceedings of the Fifth Interna-tional Conference on Extending Database Technology (EDBT’06). Su, Z., Yang, Q., Lu, Y., Zhang, H., 2000. WhatNext: a prediction system

for web requests using n-gram sequence models. In: Proceedings of the First International Conference on Web Information Systems and Engineering (WISE’00), pp. 200–207.

Tseng, V.S., Lin, K.W., 2005. Mining temporal moving patterns in object tracking sensor networks. In: Proceedings of the International Workshop on Ubiquitous Data Management (held with ICDE’05), pp. 105–112. Tseng, V.S., Lin, K.W., 2006. Eﬃcient mining and prediction of user

behavior patterns in mobile web systems. Information and Software Technology 48 (6), in press.

(10)

Tseng, S.M., Tsui, C.F., 2004. Mining multi-level and location-aware associated service patterns in mobile environments. IEEE Transactions on Systems, Man and Cybernetics: Part B 34 (6).

Tseng, Y.C., Kuo, S.P., Lee, H.W., Huang, C.F., 2004. Location tracking in a wireless sensor network by mobile agents and its data fusion strategies. The Computer Journal 47 (4).

WINS project, Rockwell Science Center.http://wins.rsc.rockwell.com. Woo, A., Culler, D., 2001. A transmission control scheme for media access

in sensor networks. In: Proceedings of Seventh ACM Annual International Conference on Mobile Computing and Networking (Mobicom’01), pp. 221–235.

Wu, H.K., Jin, M.H., Horng, J.T., 2001. Personal paging area design based on mobiles moving behaviors. In: Proceedings of IEEE Infocom.

Xu, Y., Winter, J., Lee, W.C., 2004. Prediction-based strategies for energy saving in object tracking sensor networks. In: Proceedings of the Fifth IEEE International Conference on Mobile Data Management (MDM’04), pp. 346–357.

Yang, Q., Li, T., Wang, K., 2004. Building association rule based sequential classiﬁers for web document prediction. Journal of Data Mining and Knowledge Discovery 8 (3), 253–273.

Yavas, G., Katsaros, D., Ulusoy, O¨ ., Manolopoulos, Y., 2005. A data mining approach for location prediction in mobile environments. Data and Knowledge Engineering 54 (2).

Ye, W., Heidemann, J., Estrin, D., 2002. An energy-eﬃcient mac protocol for wireless sensor networks. In: Proceedings of the 21st IEEE Infocom, pp. 1567–1576.