IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
VOL. 21,
NO. 9,
SEPTEMBER 2010
1373
Delay-Bounded and Energy-Efficient Composite Event Monitoring in Heterogeneous Wireless Sensor Networks Yingshu Li, Member, IEEE, Chunyu Ai, Student Member, IEEE, Chinh T. Vu, Student Member, IEEE, Yi Pan, Senior Member, IEEE, and Raheem Beyah, Senior Member, IEEE Abstract—Wireless sensor networks can be used for event warning applications. Till date, in most of the proposed schemes, the raw or aggregated sensed data are periodically sent to a data consuming center. However, with those schemes, the occurrence of an emergency event such as a fire is hardly reported timely, which is a strict requirement for event warning applications. In wireless sensor networks, it is also highly desired to conserve energy so that network lifetime can be maximized. Furthermore, to ensure the quality of surveillance, some applications require that if an event occurs, it needs to be detected by at least k sensors, where k is a user-defined parameter. In this work, we examine the Timely Energy-efficient k-Watching Event Monitoring (TEKWEM) problem and propose a scheme, which involves an event detection model and a warning delivery model, for monitoring composite events and delivering warnings to users. Theoretical analysis and simulation results are shown to validate the proposed scheme. Index Terms—Composite event detection, wireless sensor networks, energy efficiency, bounded delay.
Ç 1
INTRODUCTION
D
and threat brought by natural and man-made disasters can cause profound and long-lasting impact on human lives and the economy. If the appropriate authorities are not immediately notified with accurate sensed data in an emergency situation, many lives can be lost. Therefore, delaybounded and energy-efficient emergency warning systems, where the occurrence of an event is of more interest to users rather than the detailed event information, are in need to reduce the impact of disasters. As communication technology, embedded computing technology and sensing technology become more mature, affordable wireless sensor networks that have the capabilities of sensing, computation and communication make an ideal solution for emergency warning systems. In this paper, we propose a wirelesssensor-network-based scheme for monitoring composite events and delivering warnings to users in a delay-bounded and energy-efficient manner. This scheme can be employed by applications such as meteorological hazard detection, bacteriological threat detection in public areas, earthquaketsunami alerting, fire detection in forests, people locating in disasters, and enemy detection in battlefields. In this work, the following requirements of event monitoring and special characteristics of Wireless Sensor Networks (WSNs) are taken into account. AMAGE
1.1 Composite Event Detection In disaster warning scenarios, WSN users are more interested in the monitored events, rather than the sensors themselves or . The authors are with the Department of Computer Science, Georgia State University, 34 Peachtree Street, Suite 1450, Atlanta, GA 30303. E-mail: {yli, chunyuai, chinhvtr, pan, rbeyah}@cs.gsu.edu. Manuscript received 31 Mar. 2008; revised 7 Oct. 2008; accepted 29 July 2009; published online 4 Aug. 2009. Recommended for acceptance by X. Zhang. For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TPDS-2008-03-0123. Digital Object Identifier no. 10.1109/TPDS.2009.129. 1045-9219/10/$26.00 ß 2010 IEEE
the large amount of irrelevant readings from sensors. For example, for the event fire that occurred at 10:00 am in area R, users may not issue such a query as “What is the temperature sensed by node No. 70?” but queries such as “Which events happened in region R from 9 to 11 am?” or “Where was fire detected from 9 to 11 am?” Here, the interested information is an answer to a question, which can be derived by a set of predicates, not from the raw data in the form of numerical values. Users are not interested in how hot the monitored area is (temperature) or how dense the smoke is (smoke density). Instead, users expect a quick answer for the concise question “Is there fire in the monitored area?” To guarantee accuracy and reliability for emergency alerting, a conclusion indicating the happening of an event should not be decided only based on one property of the event. For example, the event fire is a fusion of multiple sensed values of multiple different attributes, i.e., the occurrence of fire should satisfy some conditions such as “temperature > 100 C AND smoke > 100 mg=L AND light > 500cd,” rather than a simple condit i o n “ temperature > 100 C, ” “ smoke > 100 mg=L, ” o r “light > 500cd” alone. Any change in either temperature or smoke density that makes temperature > 100 C or smoke > 100 mg=L true can be understood as an atomic event. The event that is a combination of several atomic events is a composite event, e.g., the event of fire is represented as “temperature > 100 C AND smoke > 100 mg=L AND light > 500cd.” This work handles composite event monitoring and we formally define a composite event in Section 3.
1.2 Bounded Delay For event monitoring applications, the most popular scenario in the literature is the following: all the sensors periodically send their data to an information processing center, e.g., a base station (BS). A conclusion is then made at the BS to decide whether a predefined event has happened based on the reported data. We name this scenario as “data collection.” Another more efficient scenario is “data Published by the IEEE Computer Society
1374
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
aggregation,” where data are processed at some nodes innetwork before being forwarded. However, these methods are not suitable for real-time applications, especially for emergency warning applications, where the alarm is stringently required to be announced timely, e.g., a missile defense system, where seconds/milliseconds count. A blip on the radar, noise at a certain frequency, and disturbance in the air stream can imply that a missile is coming quickly toward a target. The drawbacks of most of the existing methods are that the real-time requirement is not taken into account, and the amount of the exchanged data may be huge, which causes large energy consumption. In addition, at a BS, the received data need to be further analyzed to obtain a conclusion, which delays the timely announcement of the alarm. The proposed scheme treats bounded delay as the first priority goal in order to notify users of the occurrence of any event timely.
1.3 Energy Efficiency Energy conservation is one of the primary concerns for WSNs since each node is battery powered and a WSN must operate for at least a required mission time or as long as possible. Sensors could be left in an unattended environment for months or even years [38] and charging/recharging of the batteries at sensors is difficult or even impossible. Therefore, an energy-efficient way of operation is critical. A sensor node consumes more energy for transmission than that for computation. It is shown that the power to transmit 1 bit is enough for executing at least 3,000 instructions [47]. Accordingly, sensors should have the processing capabilities to locally conduct simple computations and route only the required data back to the sink [1]. In this work, preprocessing of data is performed in-network to reduce sensor transmissions. Furthermore, the amount of transmitted data is reduced as much as possible by sending 1 bit to indicate the detection of an event instead of sending longer packages containing raw data. 1.4 Heterogeneous Architecture of WSNs In this work, we take advantage of heterogeneous WSNs, where there are two kinds of nodes in a network: resourceconstrained sensor nodes and resource-rich gateway nodes. Resource-constrained sensor nodes are normal sensors with limited energy, size, and weight. These sensor nodes may have short sensing and transmission ranges, low data rate, small memory, and limited computation ability to trade off cost. It is expected that each sensor node costs less than $1 [46]. Conversely, gateway nodes are rich in energy and memory. Also, they have much stronger computation and communication abilities compared with sensor nodes. However, the cost of each gateway node is more expensive than that of a sensor node, therefore, fewer gateway nodes are employed in a WSN. A gateway node has two transceivers, one connected to the WSN and the other connected to the upper layer gateway node network as shown in Fig. 1. To send a packet from A to B, instead of using the longer path involving many sensor nodes, the shorter path involving three gateway nodes is preferred to deliver the packet timely and to conserve energy for the sensor nodes. Such gateway nodes are now commercially available [9]. It is shown by Intel [14] that heterogeneous
VOL. 21,
NO. 9,
SEPTEMBER 2010
Fig. 1. A heterogeneous WSN.
WSNs can provide better QoS. In the proposed scheme, sensor nodes are mainly responsible for collecting raw data from the monitored region and gateway nodes are used to shorten the delay of delivering warnings as much as possible. The proposed user-centric scheme in this paper focuses on the user and allows the user to specify a tolerable amount of delay and pose high-level questions in a heterogeneous WSN, which is a new emerging architecture for WSNs. The network translates this information to facilitate optimal routing to ensure that the delay is met as well as breaks down the composite questions and has the ability to answer these composite questions by fusing together different atomic events. Additionally, the user can specify the level of fault tolerance required (the number of nodes that must observe the event). All this is done while providing the normal energy efficiency that is required in sensor networks. Furthermore, the decision making is done in network, which allows you to send 1 bit as opposed to the actual value, saving transmission power. This significance was verified via analysis and simulation. Our novel scheme has the following contributions and characteristics: 1.
2.
3.
4.
5.
6.
No significant amount of data is sent to users in deciding whether an event occurred, thus, each sensor can naturally conserve more energy to extend network lifetime. Furthermore, reducing network traffic has a good effect on lowering radio interference. Decisions are locally made at gateway nodes, then only particular conclusions, e.g., the ones that specify the occurrence of an interested event, are reported to users through the upper level gateway node network, such that users can obtain valuable information timely. This framework is used for monitoring composite events involving multiple properties and users can define composite events relevant to them. By sending a warning from a gateway node to users, users can know where the event happened (with the assumption that all the gateway nodes’ positions are known to users). Fault tolerance is addressed, therefore, even if a limited number of sensors concurrently fail, users can still be properly warned if an event of interest happens. The energy consumption among sensors is well balanced, i.e., the more exhausted sensors have a greater possibility to go to sleep than the energized ones.
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
The connectivity is guaranteed among sensors in the current detection set (defined in Section 3) by using a topology and routing control scheme. The rest of the paper is organized as follows: Section 2 presents some related works. Some preliminaries and the problem definition are introduced in Section 3. The detailed working scenario of the proposed framework is illustrated in Section 4. The simulation results are shown in Section 5. Finally, Section 6 ends the paper with a conclusion and some future research work. 7.
2
RELATED WORK
The event monitoring scheme for heterogeneous WSNs proposed in this paper mainly involves two technologies: event detection and event notification. In [15], a framework for both simple event and composite detection was developed using distributed collaboration of sensors, which may have different sensing capabilities. An application subscribes an event of interest with a corresponding location, and the proposed protocol builds an event-based tree (EBT) and the data in the form of predicates will then be collected along this tree. The protocol for atomic events is quite straightforward. For a composite event, the protocol maintains a counter for each atomic event to count the number of nodes that can detect it and such nodes are added to a cluster until the counter for each atomic event is greater than a predefined threshold. This work does not take into account the real-time necessity for emergency warning systems. In [21], the purpose is to construct a middleware, which provides real-time detection service for WSNs. Essentially, the authors propose a mechanism to handle the unreliability of sensor error reports, to correlate among different sensors’ observations, and to utilize partial detection when failure rate is high. Again, the authors do not consider how data are gathered, how data are routed to a data processing center timely. They consider fault tolerance by correlating data. We also take fault tolerance into account based on other aspects of WSNs. We guarantee that an atomic event is monitored by multiple sensors. In [33], the issue of how to make a decision from predicates was addressed. Since an atomic event frequently changes states, automaton can be used to combine the state changes of atomic events to form the final states and those final states help decide if an event occurs. The drawback of automaton is its complexity, i.e., the number of states may be the exponential of the number of the atomic events and there are probably infinite ways to reach a state. Moreover, the time overlap of state changes is hard to handle. The authors only consider the issue of making decision from collected predicates without clearly specifying how the predicates are gathered. In other words, the work in [33] focuses on system issues rather than network issues, which are considered in this work. Event detection is also extensively investigated in the distributed systems research field [11], [22], [31], [34]. However, those works cannot be employed in WSNs due to the special features of WSNs mentioned in Section 1. In this work, we investigate the k-watching problem. If we simplify the network model as follows: all the sensors have only one sensing component, all the sensing components are the same, and a composite event consists of only one atomic event that can be detected by each sensor, then the k-watching problem becomes the k-coverage problem.
1375
The works in [48] and [49] propose protocols to construct connected cover sets (sets of sensors that can k-cover the monitored area). The k-coverage problem is a special case of the k-watching problem, which is our focus in this work. The issue of event notification relates to the routing protocols at the network layer. Basically, there are four types of network layer routing protocols: cluster-based routing protocols, location-based routing protocols, data-centric routing protocols, and energy-aware routing protocols. Cluster-based routing protocols [12], [23], [24], [29], [40] divide sensors into clusters according to some rules. The cluster head for each cluster is responsible for routing arrangement and every other node communicates with its cluster head. Location-based routing protocols [2], [16], [17], [30] assume that each node knows its location and the locations of the receivers. Data are transmitted based on that information. In [6], a coordinate system is constructed where location information is not available. In [4], network topology is abstracted based on medial axis and a routing protocol without using location information is presented. In [13], directed diffusion, a data-centric protocol, is developed. First, the base station broadcasts the interest under the form of attribute-value pairs to all the sensors. Each sensor stores the interest, time stamp, gradient, and some other information in its cache. If the sensor’s sensing data match the interest, the sensor uses the gradient field to send the data back to the base station. The broadcast of interest is expensive in both energy consumption and time complexity. The work in [3] overcomes this drawback. The work in [27] makes use of the work in [3] to design an acquisitional query processor for data collection. TTDD [43] is another data-centric protocol that supports mobile sinks. A similar protocol was given in [25]. Buragohain et al. [5] study the energy-aware routing problem for sensor network databases (e.g., TinyDB [26]). Energyaware routing protocols [7], [8], [20], [26], [36], [44] can help with balancing energy consumptions among sensors and extending network lifetime. The above works are devoted to communication and routing. However, event notification proposed in this paper involves event monitoring, which is not considered in the above-mentioned works. The above works do not consider some of the special characteristics of WSNs. We should be aware that WSNs have limited communication ability, data-centric feature, limited power, limited computation ability, large number of nodes, huge deployment area, and infinite sensing data streams. To the best of our knowledge, no research work in WSNs has been conducted for both event detection and event notification considering the stringent requirement of delivering a timely warning. Different from all the above works, we partition the set of sensors into a number of nondisjoint subsets, each of which is referred as a “detection set,” such that each subset can solely detect atomic events for all the predicates of a composite event. Atomic events are reported to a gateway node, which will then decide whether a composite event occurs and quickly deliver a warning to users through the gateway node network. The most relevant work to ours is [15]. Nonetheless, the work in [15] concentrates on system issues rather than network issues. It does not consider the energy consumption and it discovers only one detection set for each subregion. Besides, the proposed protocol in [15] requires the sensors’ locations, while our proposed scheme does not require normal sensors’ locations. Conversely, energy is one of our foremost concerns
1376
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
since it is a key issue in WSNs. We further consider other strict requirements such as fault tolerance, connectivity, topology, and routing control for the resulting detection sets. The proposed scheme takes advantage of the benefits brought by heterogeneous WSNs. It is shown in [42] that heterogeneity may increase average delivery rate and network lifetime. A significant amount of research on heterogeneous WSNs has been conducted in [10], [14], [28], [35], [39]. This work provides an additional application example for heterogeneous WSNs.
3
PRELIMINARIES AND PROBLEM DEFINITION
In this section, we formally define the TEKWEM problem and the following are some preliminary definitions and notations. We define a composite event as a set of predefined observation attributes and the corresponding predicates defined on the attributes. Then a composite event is defined as ðA1 ; A2 ; . . . ; Aj Þ ¼ F ðP1 ; P2 ; . . . ; Pj Þ; where Aj is a monitored attribute involved in an event , Pj is the predicate defined on Aj , and F is a function of Boolean algebra operators such as “^,” “_,” or “:.” An atomic event is a special composite event involving only one observation attribute and its corresponding predicate. As shown by the example in Section 1, any change in either temperature, smoke density, or light intencisy that makes temperature > 100 C, smoke > 100 mg=L, or light > 500 cd true is an atomic event. The composite event fire is then represented as “temperature ^ smoke ^ light.” We define a Detection Set as a subset of sensors, which jointly accomplish the event detection task. Then, a kwatched atomic and a k-watched composite event are defined as follows: Definition 1 (k-watched atomic event). An atomic event is said to be k-watched by a detection set D if at any time, this event occurs at any point within the interested area, at least k sensors in D can detect this occurrence. Definition 2 (k-watched composite event). A composite event is said to be k-watched by a detection set D if every atomic event involved in that composite event is k-watched by D. We introduce the k-watching concept for the purpose of fault tolerance so that all the atomic events of a composite event are ensured to be detected even when any k 1 sensors monitoring the same property concurrently fail. Additionally, k-watching helps a gateway node to more quickly be aware of the occurrence of a composite event. In a primitive and standard form, the k-watching problem can be defined as follows: Definition 3 (k-watching problem). Given a set of sensors S, a monitored area A, and a composite event involving r atomic events l ; l ¼ 1::r, find a subset D of S such that every l is kwatched by the set D. Practical applications always require several other network constraints. In event monitoring, the most important issue is to bound the warning delivery delay as much as possible. In a heterogeneous WSN, data transmission from sensor nodes to the BS usually consists of two steps. First, a sensor node sends the packet to its dominating gateway
VOL. 21,
NO. 9,
SEPTEMBER 2010
node. Second, the gateway node relays the packets to the BS through the upper layer gateway node network as shown in Fig. 1. Accordingly, in our scheme, warning delivery contains two steps. First, the sensor nodes that detect an atomic event inform a gateway node, then if the gateway node decides that a composite event occurred, it delivers a warning to the BS through the upper layer gateway node network. Obviously, to ensure that the messages are properly routed, a path from every sensor node which detects an atomic event to a gateway node must be maintained; and a path from the gateway node which identifies a composite event to the BS must be maintained as well. Additionally, conserving energy while accomplishing tasks is important. Thus, to bound the warning delivery delay and minimize energy consumption are our primary concerns in this work. Since an atomic event is a special composite event, we only consider the latter. Formally, the proposed scheme concentrates on the following TEKWEM problem: Definition 4 (Timely Energy-efficient k-Watching Event Monitoring—TEKWEM). Given a set of sensors S, a monitored area A, and a composite event involving r atomic events l ; l ¼ 1::r, find a set of nondisjoint connected detection sets Dj ; j ¼ 1::m of S, and decide their corresponding active duration such that: 1. 2. 3.
4
The composite event is k-watched by Dj ; j ¼ 1::m at any time. The warning delivery delay is minimized. The network lifetime is maximized.
WORKING SCENARIO
In this section, we introduce the entire event monitoring scheme. This scheme consists of two models: event detection model and warning delivery model. Section 4.1 illustrates how the sensor nodes within each observation zone coordinate to ensure that each event is k-watched. This section also demonstrates an algorithm for constructing warning delivery paths from all the gateway nodes to the BS so that warnings can be delivered in time. Both the event detection model and the warning delivery model take into account energy conservation. The monitored area of a WSN is divided into m n observation zones as shown in Fig. 2. To monitor an event ðA1 ; A2 . . . ; Aj Þ within each observation zone, we need at least one gateway node and some sensor nodes to observe the attributes A1 ; A2 ; . . . ; Aj . We assume a uniform deployment. The work in [19] investigates the conditions for a network to be able to provide k-coverage service. It provides ideas for node deployment so that any atomic event can be k-watched. The sensor nodes serve as the event observation nodes for event . In order to capture the event , a gateway node serves as the event fusion node for . The compound statement for event is only disseminated to gateway nodes. The program code that will be executed at a node is burned into the chip of this node before it is deployed to the monitored region. The thresholds for events are sent to sensors by the base station so that users may define new events during network lifetime. The threshold values of the monitored attributes are preinstalled on sensor nodes. Once a sensor node detects that the current sensed value reaches the threshold of its monitored attribute, it sends one bit “1” instead of the sensed value to its
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
1377
In the network deployment phase, each sensor node subscribes to a gateway node in the same observation zone. Event definitions are diffused to all the gateway nodes in the deployment phase. Thresholds for the monitored attributes are preinstalled in the sensor nodes. Note that gateway nodes may also carry out sensing tasks. Furthermore, we have the following assumptions: More than enough sensors are deployed to monitor all atomic events in each observation zone. . At least one gateway node is deployed to each observation zone. . Each node has different sensing abilities. That is, each node may be equipped with more than one sensing component and both the numbers and types of those sensing components may be different among nodes. For example, a node can monitor light intensity and/or smoke density, while its neighbor monitors temperature and/or pressure. . There is only one of each type of sensing component on each node. For example, a node has only one temperature sensing component and/or one pressure sensing component. . All of a node’s sensing components turn on or off simultaneously. . Nodes may have different communication ranges and different initial battery supplies. . Gateway nodes have much more energy than sensor nodes. Within each observation zone, the gateway node is responsible for constructing a set of detection sets by executing the algorithm shown in Algorithm 1. The input includes the set of sensors S in this observation zone and the user-specified fault tolerance level k. The output gives a series of Detection Sets (DSs) Dj with their activation durations tj . The gateway is in charge of building BreadthFirst-Search-like (BFS-like) trees rooted at itself connecting sensors belonging to a DS. The gateway greedily adds into a DS sensors whose sensing components can help to kmonitor atomic events. For example, assume that the desired coverage level is 2 (k ¼ 2), and so far only one sensor in the current DS is equipped with a temperature sensing component. The gateway will add into the current DS any sensor equipped with a temperature sensing component. These sensors are said to have nonzero contribution because the number of its current helpful sensing components (which will be defined later) is bigger than 0. For sensors which have no contribution to the current DS, it may be added to the current DS later to provide connectivity for routing paths among sensor nodes. After executing this algorithm, the gateway node has knowledge of how many detection sets are constructed; for each detection set, which sensor nodes are involved in this detection set and their topology and routing information; and the duration during which this detection set should remain active to perform the detection task. Based on this information, the gateway node can decide a working schedule for the set of obtained detection sets. This working schedule is then broadcasted to all the sensor nodes so that each of them can know when it needs to be active. During the active time of a detection set, the sensor nodes in this detection set route their messages to the gateway node based on the provided topology and routing information. .
Fig. 2. A heterogeneous WSN.
dominating gateway node. If a gateway node receives a “1,” it checks if any compound statement defined for an event derives a T RUE value. If so, it immediately sends a warning to the BS through the upper layer gateway node network. In this work, delay includes two parts, detection delay and warning delivery delay. Detection delay is the duration from the occurrence of an event to the time point it is detected by a gateway node. It is decided by the longest path along which the gateway receives reports. On this path, the sender and forwarders are regular sensors. A gateway node is in charge of processing reports from regular sensors and judging whether an event happened. Once an event is detected at a gateway node, a warning is sent from this gateway node to the Base Station. Only gateway nodes participate in warning delivery. Since detection and warning delivery are two different phases and the data transmit rates of regular sensors and gateway nodes vary a lot, detection delay and warning delivery delay are studied separately in this work. The proposed scheme can be extended to the case where different gateway nodes are capable of detecting different events. Users may predefine many composite events at each gateway node. Once a gateway node receives the reports that some monitored attributes exceed their thresholds, it checks all the event definitions involving these attributes to decide whether these events occur.
4.1 Event Detection Model At a gateway node G, it has the compound statement definition to a composite event. For example, event fire is defined as “temperature ^ smoke ^ light,” where temperature, smoke, and light are binary parameters. Each sensor knows the threshold for its monitored attribute. Once a sensor detects that its reading exceeds this threshold, it sends a bit “1” to a gateway node. Otherwise, nothing will be sent. For example, temperature sensor S1 sends 1 to G if it detects that the current temperature is greater than 100 C, smoke sensor S2 sends 1 to G if it detects that the current smoke level is greater than 100 mg/L, and a light sensor S3 sends 1 to G if it detects that the current light intensity is greater than 500 cd. Gateway node G then can determine the occurrence of a composite event fire based on the received 1s. Since this scheme is designed mainly for warning delivery applications, users are less interested in the exact readings of sensors. The detection of the occurrence of an event is of more interest. This scheme can reduce the communication cost, conserve energy at sensor nodes, and alleviate workload at gateway nodes.
1378
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
Essentially, the messages are routed to the gateway node along the constructed BFS-like tree for the current DS. The use of BFS-like trees is motivated by the requirements of topology control and routing paths maintenance for data delivery to the gateway node. Also, a BFS-tree has a similar characteristic as that of an SPT, so the total delay of sending data from active sensors (i.e., sensors belonging to a DS) to the gateway can be minimized. Algorithm 1. Detection-Sets-Construction(S, k) 1: m=0. 2: while S 6¼ do 3: T ¼ . Set all the counters cl to k. 4: Color all the nodes WHITE. 5: gw.Color = BLACK. /*gw denotes the gateway node.*/ 6: T ¼ fgwg. 7: /* Discover detection set */ 8: while at least one counter > 0 do 9: L = Construct-Leaves(S, T ) 10: (Re)Calculate the contribution of each node in L. 11: Color all the nodes with contribution of 0 RED and remove them out of L. 12: if L ¼¼ then break; 13: Sort L in descending order of contributions. 14: while L 6¼ do 15: Remove the node from the top of list L. 16: Add node into T . 17: .Color = (.Parent is BLACK)? BLACK : GREEN 18: Decrease all the ’s correlated counters by 1. 19: if a counter == 0 then 20: all counters == 0? goto line 25: goto line 10 21: end if 22: end while 23: end while 24: /*Update the subset and the sensors’ energy*/ 25: if any counter > 0 then 26: Remove T from S. /* T is isolated */ 27: else 28: if there exists a GREEN node then 29: for each GREEN node do 30: ¼ .Parent 31: while is RED do 32: .Color = BLACK 33: Add to T. 34: ¼ .Parent 35: end while 36: end for 37: end if 38: m=m+1 39: gwm ¼ gw; Dm ¼ T /* A new detection set */ 40: Assign tm the smallest lifetime of a node in Dm . 41: Recalculate residual energy of sensors in Dm . 42: Remove from S the sensors who ran out of energy. 43: end if 44: end while 45: return m; fDj ; tj gj¼1::m
VOL. 21,
NO. 9,
SEPTEMBER 2010
The detection set’s construction algorithm starts by constructing a set of connected BFS-like trees rooted at the gateway node. The sensor nodes in each of these trees form a detection set. For each atomic event l , we maintain a counter cl recording the number of the currently needed sensors, which can detect l for the current detection set to provide kwatching for l . The initial value of cl is k for each l . A sensor may be equipped with several sensing components that can monitor different l . For a sensor si , a sensing component componenti;l is called a helpful sensing component for counter cl if it can monitor l and the current value of cl is greater than 0, and cl is called a correlated counter for component componenti;l . At any point of time, a color in the following list is used to represent a sensor node’s state: W HIT E: This node has not been considered. BLACK: This node has already been added to a detection set and is connected to the gateway through other sensor nodes in the current detection set. . RED: This node is useless (the one whose all correlated counters cl are 0) and has been considered, but it has not been added to a detection set. . GREEN: This node has been added to a detection set but its parent is not a BLACK node, thus, it is not connected to the gateway through other sensor nodes in the current detection set. At the beginning of a detection set construction, all the nodes are W HIT E except the BLACK gateway node gw, thus, the tree T contains only gw. Denote L as a list of all of T ’s leaves’ WHITE neighbors, which intuitively is the set of nodes in the next level in the BFS tree of the current T ’s leaves’ level. Algorithm 1 works in a greedy manner on a sensor node’s attribute named contribution to find a sensor node that can be added to the tree. For a sensor node si , its contribution i can be determined depending on some parameters such as its residual energy, the energy needed to transmit a message to its parent node, and the number of its helpful sensing components. i can be formulated as follows: . .
i ¼ fðei ; di Þ
hi ; sci
ð1Þ
where: fðei ; di Þ is a function to calculate si ’s lifetime depending on its current residual energy ei and its current communication range di . This function is further discussed in detail in Section 5. . hi is the number of si ’s helpful sensing components. . sci is the number of all the sensing components that si is equipped with. Equation (1) can easily be extended to take any other parameters into account. Note that the contribution of a sensor node becomes 0 when hi ¼ 0, i.e., when all the correlated counters for all its sensing components are 0, it becomes a useless sensor (for the current detection set). The algorithm tries to construct as many detection sets as possible. At each iteration of the main loop, a temporary variable T is used to store the current tree being constructed. When the tree is completely built, it becomes a new detection set. T is gradually constructed by adding the node in L that has the biggest contribution. For other nodes in L having contribution of 0, they are colored RED. Each time a node is added to T , it is removed from L and all .
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
1379
Fig. 3. The construction of a detection set with k ¼ 3, r ¼ 3. Node 1 is the gateway node. (a) Initial graph, (b) graph after coloring nodes, and (c) a detection set consisting of BLACK and GREEN nodes.
of its correlated counters are decreased by 1. Also, the node added to T is colored BLACK if it can connect with the gateway node through other BLACK nodes. If it cannot, i.e., its parent is a RED or GREEN node, it becomes GREEN. Each time any counter becomes 0, all the remaining sensors in L need to recalculate their contribution i , since their hi values may change. T ’s construction process finishes when 1) all the counters reach 0, i.e., sensors in T can now provide k-watching for the composite event or 2) there exist no T ’s neighbors. For the latter case, remove all the sensors in T from S (line 1, Algorithm 1) since T is isolated from the other sensors in S. For the former case, to guarantee connectivity, some RED nodes need to be added to make GREEN nodes connected to the gateway, which is accomplished by the block of code from line 28 to line 37. When T is completely built, it contains only BLACK and GREEN nodes. It is then assigned an active time t, which is the smallest lifetime of a sensor in T . The algorithm keeps constructing more trees using the above process until no more detection sets can be discovered. Finally, the algorithm returns the constructed detection sets with their active durations. While constructing L (Algorithm 2), we explicitly specify the parent for each node in L. Thus, a topology for each detection set is as well established. Furthermore, when a sensor node needs to report an atomic event, it only needs to forward that report to its assigned parent. By that mechanism, no addition routing protocol is needed. Algorithm 2. Construct-Leaves(S, T ) Input: A set of sensors S and a tree T . Output: L - the list of all the children of T ’s leaves 1: Construct a list L consisting of all the WHITE neighbors of T ’s leaves. 2: /* Assign parent for each node in L*/ 3: for each node in L do 4: if any neighbor of is BLACK or GREEN then 5: .Parent = the BLACK/GREEN node with the least number of children. 6: else 7: .Parent = the RED node with the longest lifetime. 8: end if 9: end for 10: return L. Theorem 1. The algorithm ensures that a composite event is kwatched by every detection set and all the detection sets are connected sets.
Proof. When a counter for an atomic event reduces to 0, the number of the sensors in the current detection set being able to detect that event is at least k, i.e., that event is currently being k-watched by the set. In our algorithm, a detection set is claimed to be discovered when all the counters become 0. Thus, the composite event is kwatched by any detection set. If a detection set D contains only BLACK nodes, D is a connected set since all the BLACK nodes are connected to the gateway node through some sensors in D. If D contains several GREEN nodes, the block of code from line 28 to line 37 adds RED nodes to connect GREEN nodes in the lower level to the GREEN or BLACK nodes in the upper level of the BFS tree. The GREEN nodes in the upper level will also be or have already been connected to the BLACK=GREEN nodes in the higher layer of the BFS tree. Thus, all the GREEN nodes, and hence, all the nodes in the current detection set eventually are connected to a gateway node. u t As mentioned, a composite event can be understood as the combination of some atomic events, e.g., high temperature, dazzling light, and dense smoke indicate a fire event. Fig. 3 illustrates an example of constructing one detection set in an observation zone for a fire alarming system. Each node is equipped with one or several sensing components, where 1 is for temperature, 2 is for light, and 3 is for smoke sensing component. The bold number above each node is its ID and the list beside it is the list of its equipped sensing components. Note that in Fig. 3b, each node is assigned a unique parent, and in Fig. 3c, node 5 is changed to BLACK and is added to the detection set with the purpose of connecting GREEN nodes 7 and 10 to BLACK node 3. Thus, if node 10 wants to report an event to the gateway node 1, it can send that report along the path: 10-7-5-3-1.
4.2 Warning Delivery Model Once a gateway node decides that a composite event occurs, it needs to deliver a warning to the BS. At this point, the route along which this warning is delivered is of great concern. On one hand, even though the gateway nodes are rich in resources, energy conservation is still necessary for maintaining a long-lived network backbone. On the other hand, users expect to receive warnings as soon as possible. Therefore, minimizing energy consumption and warning delivery delay are two optimization targets here. Warning delivery delay refers to the time for a warning to be delivered to the BS across the network. It is highly
1380
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
VOL. 21,
NO. 9,
SEPTEMBER 2010
P total weight of a graph G ¼ ðV ; EÞ is W ðGÞ ¼ e2E wðeÞ. In a wireless network, all the neighbors within a sender’s transmission range can receive the packets from the sender, unlike in wired networks where a sender needs to send to every neighbor separately. This is called the Wireless Multicast Advantage (WMA) [41]. Then for each node u, if it transmits at power level pðuÞ ¼ maxfwuv : ðu; vÞ 2 Gg, this transmission can be received by all of its neighbors. Let T be a spanning tree of graph GðV ; EÞ and the total P energy consumption of T , denoted by P ðT Þ, is P ðT Þ ¼ u2V pðuÞ. Clearly, among all the spanning trees of G, a Minimum Spanning Tree (MST) T has the smallest W ðT Þ, and a Shortest Path Tree (SPT) T 0 has the smallest DðT 0 Þ. We now formally define the warning delivery problem as follows: Definition 5 (Energy-efficient wArning Delivery with bOunded Delay (EADOD) Problem). Given a set of gateway nodes consisting the upper layer gateway node network of a WSN, construct a warning delivery tree T so that both P ðT Þ and DðT Þ are as small as possible.
Fig. 4. An example ð; Þ-dLAST. (a) Gateway node network. (b) An MST TMST . (c) An SPT TSP T . (d) An (; )-dLAST T .
dependent on the length of the traversed route. Therefore, to obtain a small warning delivery delay, it is desired that the route from the source to the BS be as short as possible. In this section, we illustrate the warning delivery model with the primary concern of minimizing the total energy consumption while bounding the warning delivery delay. Only the gateway nodes constituting the upper layer gateway node network are considered in this section. The sensor nodes can communicate with its dominating gateway node within an observation zone in a constant time, so sensor nodes are ignored in this model. We assume that gateway nodes periodically report their remaining energy and positions to the BS so that the BS can know the topology information of the gateway node network. According to the updated topology information, the BS constructs a warning delivery tree, which involves all the gateway nodes. This warning delivery tree should be optimized at both total energy consumption and delivery delay. The root of the tree could be the BS itself or a gateway node with more remaining energy that is close to the BS. In this work, we use the latter. The BS then notifies each gateway node which other gateway node is its parent node to which it delivers a warning. We model the upper layer gateway node network as an undirected graph G ¼ ðV ; EÞ, where V includes all the gateway nodes and E is the edge set of the gateway node networks. The euclidean distance between node u and node v is represented by duv . To send a packet from u to v, the required transmission power of u, denoted by puv , should be greater than the signal detection threshold at v, which is Cv duv [32], where Cv is a constant number determined by v’s resident environment and is a number between 2 and 4 depending on the communication medium. We assign each edge e ¼ ðu; vÞ a weight wðu; vÞ ¼ maxðCu duv ; Cv duv Þ. As shown in Fig. 4a, for each edge, the number without parentheses is the length of this edge (denoted by dðeÞ) and the number in parentheses is the weight of this edge (denoted by wðeÞ). wðeÞðe ¼ ðu; vÞÞ can be regarded as the energy needed to transmit a message from u toPv. Then the total length of a graph G ¼ ðV ; EÞ is DðGÞ ¼ e2G dðeÞ and the
The (; )-Light Approximate Shortest path Tree ((; )LAST) is a concept in wired networks introduced in [18]. Based on (; )-LAST, we define the (; )-deliver LAST ((; )-dLAST) in WSNs as follows: Definition 6 ((; )-dLAST). For 1 and 1, a spanning tree T of a graph G meeting the following two requirements is called an (; )-dLAST rooted at r such that: .
.
(Delay) For every node u, the distance between r and u in T is at most times the shortest distance from r to u in G, where is a user-specified parameter. (Energy) The total energy consumption of the nodes in T is at most times the total energy consumption of an optimal solution to the EADOD problem.
The weight of an edge in [18] is the distance between a pair of nodes. In this paper, the weight is defined as the required energy for a pair of nodes to communicate in WSNs. The idea of constructing an (; )-dLAST is similar to the one for constructing (; )-LAST in [18]. First, an MST TMST and an SPT TSP T both rooted at r are constructed based on the reported topology information. Next, TMST is traversed in a depth-first manner. When visiting a node u, if the total length of the path from r to u in TMST is larger than times the total length of the path from r to u in TSP T , then this path is replaced by the path from r to u in TSP T . is a user-specified parameter, which indicates the user tolerance level for warning delivery delay. This process is applied to every gateway node. The obtained new tree is the desired (; )dLAST. The detailed algorithms are provided in [18]. Fig. 4 shows an example of constructing an (; )-dLAST. Based on the topology and weight information of the gateway node network (Fig. 4a), an MST TMST (Fig. 4b) rooted at node 2 and an SPT TSP T (Fig. 4c) rooted at node 2 are constructed. We can regard node 2 as the base station, where all the warning should be delivered to. The total weight of TMST is 29 and the total weight of TSP T is 36. The total length of TMST is 11 and the total length of TSP T is 9. The larger the total weight of a tree, the more energy consumption it indicates if messages are sent from any node in the tree to the root along the paths in this tree. The larger the total length of a tree, the longer delay it indicates if messages are sent from any node in the tree to the root along the paths in this tree. Therefore, TMST
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
incurs smaller energy consumption while longer delay compared to TSP T . After traversing all the gateway nodes in TMST , the paths from root 2 to nodes 5, 6, 7, and 8 are replaced by the corresponding ones in TSP T and the new tree shown in Fig. 4d is the desired (; )-dLAST T . Whether to replace a path in TMST with a path in TSP T depends on the user input . In this example, we set to 2. When visiting node 5, the length of the path from root 2 to node 5 in TMST is 6, which is more than twice of the length of the path from root 2 to node 5 in TSP T (which is only 1). Thus, the path from node 5 to root 2 is replaced with the one in TSP T . Similar replacements happen to nodes 6, 7, and 8. In the new constructed Y , once a gateway node u needs to send a warning, this warning is delivered along the path from u to node 2 in T . T balances energy consumption and transmission delay. Users can control to show their preference for small energy consumption or small transmission delay. is a parameter that affects warning delivery delay and it is decided by users, while evaluates the total energy consumption of all the gateway nodes and it is decided by the constructed (; )-dLAST. In [18], is obtained under the constraints of wired networks. However, WSNs are so different from wired networks and now we derive for WSNs. To derive , we have the following lemmas: Lemma 2. For an undirected graph G ¼ ðV ; EÞ, 12 P ðGÞ W ðGÞ 2 P ðGÞ, where is the maximum node degree in G. Proof. For an undirected graph G ¼ ðV ; EÞ, X X X X pðuÞ ¼ maxðu;vÞ2E wðu; vÞ wðu; vÞ P ðGÞ ¼ u2V
u2V
u2V ðu;vÞ2E
¼ 2W ðGÞ; X 1X X wðu; vÞ ¼ wðu; vÞ W ðGÞ ¼ 2 u2V ðu;vÞ2E ðu;vÞ2E
1381
Let p1 ¼ ¼ pn ¼ 1, we have ðd1 þ þ dn Þ d1 þ þ dn : n n Assume that Cl ¼ minðC1 ; . . . ; Cn Þ, then ðd1 þ þ dn Þ d1 þ þ dn C1 d1 þ þ Cn dn : n n nCl Therefore, we have D ðT Þ ¼ ðd1 þ þ dn Þ
n1 C1 d1 þ þ Cn dn Cl
n1 W ðT Þ: Cl t u
For a graph representing a WSN, the total weight of this graph may be different from the total length of this graph. This lemma shows the relationship between these two parameters. We introduce C1 ; . . . ; Cn since each node may reside in a different environment and have a unique signal detection threshold. Lemma 4. Given an MST TMST and an SPT TSP T , W ðTSP T Þ 2 1 Ck ð1 Þ n W ðTMST Þ. Proof. It has been proved in [18] that DðTSP T Þ 2 1 DðTMST Þ. Then by Lemma 3, we have W ðTSP T Þ Ck D ðTSP T Þ 2 D ðTMST Þ Ck 1 Ck 2 n1 W ðTMST Þ; Cl 1 where Ck ¼ maxðC1 ; . . . ; Cn Þ and Cl ¼ minðC1 ; . . . ; Cn Þ. t u
1X pðuÞ ¼ P ðGÞ: 2 u2V 2 t u
For a graph representing a WSN, the total weight of this graph may be different from the total energy consumption of all the transmitting nodes due to the wireless transmission pattern and wireless broadcast advantage. This lemma shows the relationship between these two parameters. 1
Lemma 3. For a tree T , D ðT Þ nCl W ðT Þ, where 2 ½2; 4 and Cl ¼ minðC1 ; . . . ; Cn Þ. Proof. Let F ðxÞ ¼ x . For 1, F ðxÞ is a convex function defined in ðþ1; 1Þ. Jensen’s Inequality [45] is as follows: if F ðxÞ is defined in an open or close interval ða; bÞ and is a convex function, then for any positive numbers p1 ; . . . ; pn , and for any points x1 ; . . . ; xn in ða; bÞ, p1 d1 þ þ pn dn p1 F ðd1 Þ þ þ pn F ðdn Þ : F p1 þ þ pn p1 þ þ pn Then by Jensen’s Inequality, ðp1 d1 þ þ pn dn Þ p1 d1 þ þ pn dn : ðp1 þ þ pn Þ p1 þ þ pn
This lemma compares the weight of a shortest path tree with the weight of a minimum spanning tree in a WSN. Based 2 on Lemmas 2-4, we derive that ¼ ð1 þ CCkl n1 ð1 Þ Þ from the following theorem: Theorem 5. Let T be the obtained (; )-dLAST, Topt be the optimal solution to the EADOD problem. Then P ðT Þ 2 ð1 þ CCkl n1 ð1 Þ ÞP ðTopt Þ. Proof. Let TMST be an MST and TSP T be an SPT. From Lemmas 2-4, we have P ðT Þ 2W ðT Þ 2ðW ðTMST Þ þ W ðTSP T ÞÞ Ck 2 1 n 2 1þ W ðTMST Þ Cl 1 Ck 2 n1 W ðTopt Þ 2 1þ Cl 1 Ck 1 2 1þ n P ðTopt Þ: 1 Cl t u We believe that the EADOD problem is NP-hard with the constraint to bound warning delivery delay and the
1382
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
VOL. 21,
NO. 9,
SEPTEMBER 2010
TABLE 1 Simulation Settings
optimization goal of minimizing the total energy consumption. Theorem 5 evaluates the proposed solution through showing its performance ratio, which is ð1 þ CCkl n1 2 ð1 Þ Þ. Here, alpha is a user-specified value that bounds the warning delivery delay. Theorem 5 shows that the energy consumption of the proposed solution is ¼ 2 Þ Þ times of the energy consumption of ð1 þ CCkl n1 ð1 the optimal solution.
5
SIMULATION RESULTS
Since energy consumption and event detection delay are primary concerns in this work, we conducted simulations to evaluate the proposed scheme. Delay involves two parts: event detection delay incurred by event detection within a grid and warning delivery delay incurred by the delivery of a warning to the base station through gateway nodes. In the simulations, energy consumption and event detection delay of the event detection model are shown first. Then, we evaluate energy consumption and warning delivery delay of the warning delivery model. Table 1 shows the basic settings of the simulations. We evaluated delay and energy consumption for different parameters including network size, , k, and the number of atomic events r. For each measurement, we run the simulations on 100 different completely randomized networks and report the average results.
5.1 Performance of Event Detection Model Since the event detection model is employed in all the grids independently, it suffices to evaluate it in a single grid. One gateway node and a number of normal sensor nodes are randomly deployed in a grid. The network lifetime was measured for different values of k and network sizes. The function to calculate sensor si ’s lifetime mentioned in (1) can be computed as follows: fðei ; di Þ ¼
ei ; T xi þ Sxi
ð2Þ
where: 1. 2.
ei is the current residual energy (in millijoules (mJ)). Sxi is the energy needed to sense an event. Since we do not consider the coverage problem in this work, for the simulation part, we assume that Sxi is proportional to the number of the sensing components sci equipped on si , and each sensing component spends 10 (mJ/unit time) when it is turned on. Thus, Sxi ¼ 10 sci ðmJ=unit of timeÞ.
Fig. 5. Network setting impact. (a) Network lifetime evaluation. (b) Event detection delay evaluation. (c) Network lifetime evaluation.
T xi is the energy needed to transmit a message, which is a function of the sensor’s communication range di . The following model is widely adopted in the literature: T xi ¼ a di þ , where a, , and are constants, 2 4, and a is usually set to 1 [37]. Fig. 5a shows the network lifetime for a grid. We assume that a composite event continuously happens, so the sensors in each detection set have to keep reporting events all the time, which is the worst case in practice. This assumption makes it easier to illustrate our scheme’s performance. Accordingly, the practical network lifetime resulting from our scheme will be longer than what is shown in the figure. As can be observed, the network lifetime is relatively N . This effect is understandable proportional to the ratio rk since the bigger the value of the level of fault tolerance k or the number of the atomic events r, the more sensors need to be involved in the event detection task. Therefore, the more energy that the network consumes in a unit of time, hence, the smaller the network lifetime. On the other hand, the larger the number of the sensors, the bigger the number of detection sets, hence, the longer the network lifetime. We evaluated how fast a gateway node can determine that an event happens after the event’s actual occurrence moment and the results are shown in Fig. 5b. We name the time interval from an event’s actual occurrence moment to the 3.
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
1383
Fig. 6. Comparisons of total energy consumptions. (a) Total energy pðT Þ consumption. (b) pðMST Þ.
Fig. 7. Comparisons of transmission delays. (a) Delivery delay. dðT Þ (b) dðSP TÞ .
moment when a gateway node detects this event as the event detection delay. Note that Fig. 5b shows the total time for all the members of a detection set to route their messages to the gateway node by using the routing path created by our algorithm without any aggregation. As can be seen in Fig. 5b, the event detection delay can be reduced if more sensors are employed or if k is smaller. With a larger number of sensors, the height of the tree could be smaller, thus, the time for a packet to be delivered to the gateway node is smaller resulting in a smaller event detection delay. With a higher fault tolerance level k, a gateway node has to wait for a longer period of time, until it collects k reports of the occurrence of all atomic events, consequently the event detection delay becomes larger. The event detection delay is actually the time that the gateway node needs to know for sure that a composite event indeed occurs, i.e., the time for the gateway node to receive k reports for each atomic event. However, the gateway node can be aware of the occurrence of a composite event right after it receives only one report for each atomic event, meaning that the gateway node may be notified of the occurrence of the event in a much shorter time. To measure this kind of time, we introduce a new measurement criterion named average notification time, time . This value is less than which is defined as Notification k 5.7 for all the simulations. It is small enough for an event to be warned timely as required by TEKWEM. Fig. 5c shows the network lifetimes for different values of r and k. As can be seen, with the same value of k, the network lifetime decreases when the value of r increases. Similarly, with the same value of r, the network lifetime also decreases when the value of k increases. The effect can be reasoned by the same explanations we have made for Fig. 5a.
Fig. 6a compares the total energy consumptions by the three routing trees for warning delivery. It shows that an MST has the smallest energy consumption and an SPT has the most energy consumption. On average, an SPT’s energy consumption is 2.5 times of that of an MST, and a dLast’s energy consumption is 1.7 times of that of an MST. Fig. 6b pðSP T Þ pðdLastÞ shows the ratios of pðMST Þ and pðMST Þ , where pðT Þ denotes the total energy consumption for warning delivery if T is used as the routing tree. It shows that, on average, a dLast consumes 39 percent less energy than an SPT does. Fig. 7a compares the total warning delivery delay of the three routing trees. It shows that an SPT has the shortest delivery delay and an MST has the longest delivery delay. On average, an MST’s delivery delay is 1.9 times of that of an SPT, and a dLast’s delivery delay is 1.3 times of that of an Þ dðdLastÞ SPT. Fig. 7b shows the ratios of dðMST dðSP T Þ and dðSP T Þ , where dðT Þ denotes the delivery delay if T is used as the routing tree. It shows that, on average, a dLast takes 27 percent less time to deliver a warning than an MST does. The simulations whose results are shown in Fig. 8 are for different values of . limits the upper bound of warning delivery delay to trade off energy consumption. Þ In Fig. 8a, it shows that as increases, pðdLAST pðMST Þ converges Þ to 1. In Fig. 8b, it shows that as increases, dðdLAST dðSP T Þ ðMST Þ converges to dðSP T Þ . The reason is that given a large , the delivery delay of any path in an MST does not exceed the user-specified upper bound, therefore, a dLAST may adopt most paths of an MST. A dLAST is constructed through traversing an MST, based on the user-specified , a path in this MST may be replaced by a path in an SPT. In Fig. 8b,
5.2 Performance of Warning Delivery Model In the proposed scheme, when a gateway node detects an event, it deliveries an event warning to the base station through the routing paths obtained from the warning delivery model. In this set of simulations, we evaluate the delivery delay and energy consumption of the warning delivery model. An MST provides an optimal solution for energy consumption and an SPT provides an optimal solution for warning delivery delay. Since the derived routing tree dLAST takes the advantages of both MSTs and SPTs with a trade-off, we compare a dLast with an MST and an SPT here. In the simulations whose results are shown in Figs. 6 and 7, is set to 2.
Fig. 8. Simulations based on different . (a) Impact of to the energy consumption. (b) Impact of to the transmission delay.
1384
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,
[9]
it shows that when is greater than 6, dLAST performs exactly the same as MST does. This is because no paths in the MST are replaced by the paths in the SPT. In this case, dLAST is actually the MST and it indicates users only choose energy consumption as the metric.
[11]
6
[12]
CONCLUSION
In this paper, we propose a wireless-sensor-network-based scheme for monitoring composite events and delivering warnings to users. This scheme consists of two models: event detection model and warning delivery model. The event detection model guarantees that each event can be kwatched and the warning delivery model ensures that warnings can be delivered to users timely. Both of these models take into account energy conservation. Theoretical analysis and simulation results show that the proposed scheme can k-watch events and deliver warnings in a delaybounded and energy-efficient manner. We are interested in further investigating new representation methods for composite events to improve efficiency and accuracy. For warning delivery, we plan to study some other routing structures, which may well balance energy consumption and delivery delay. In this work, only warnings are delivered once an event is detected, while event information is not recorded. We will develop an event storage scheme based on the scheme proposed in this paper. It is assumed that there are more than enough sensors and at least one gateway node in each observation zone. We will further investigate how to deal with the case where these assumptions cannot be made in hostile environments.
ACKNOWLEDGMENTS This work is supported by the US National Science Foundation (NSF) under grants No. CCF-0545667 and CCF 0844829, the Key Program of NSFC under grant No. 60533110, the National Grand Fundamental Research 973 Program of China under grant No. 2006CB303000, and the NSFC-RGC of China under grant No. 60831160525.
[10]
[13]
[14] [15]
[16]
[17]
[18]
[19] [20] [21]
[22]
[23]
[24]
[25]
[26]
REFERENCES [1] [2] [3] [4] [5] [6] [7] [8]
I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless Sensor Networks: A Survey,” Computer Networks, vol. 38, pp. 393-422, 2002. P. Bose and P. Morin, “Online Routing in Triangulations,” Proc. 10th Int’l Symp. Algorithms and Computation, pp. 113-122, 1999. D. Braginsky and D. Estrin, “Rumor Routing Algorithm for Sensor Networks,” Proc. First Workshop Sensor Networks and Applications (WSNA ’02), Oct. 2002. J. Bruck, J. Gao, and A. Jiang, “MAP: Medial Axis Based Geometric Routing in Sensor Networks,” Proc. ACM MobiCom, 2005. C. Buragohain, D. Agrawal, and S. Suri, “Power Aware Routing for Sensor Databases,” Proc. IEEE INFOCOM, 2005. A. Caruso, S. Chessa, S. De, and A. Urpi, “GPS Free Coordinate Assignment and Routing in Wireless Sensor Networks,” Proc. IEEE INFOCOM, 2005. A. Cerpa and D. Estrin, “ASCENT: Adaptive Self-Configuring Sensor Networks Topologies,” Proc. IEEE INFOCOM, vol. 3, pp. 23-27, June 2002. B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris, “Span: An Energy-Efficient Coordination Algorithm for Topology Maintenance in Ad Hoc Wireless Networks,” ACM Wireless Networks J., vol. 8, no. 5, Sept. 2002.
[27]
[28]
[29]
[30]
[31]
[32] [33]
VOL. 21,
NO. 9,
SEPTEMBER 2010
“Crossbow MIB and Stargate Gateway Nodes,” http:// www.xbow.com, 2009. O. Gnawali, B. Greenstein, K.-Y. Jang, A. Joki, J. Paek, M. Vieira, D. Estrin, R. Govindan, and E. Kohler, “The Tenet Architecture for Tiered Sensor Networks,” Proc. ACM Conf. Embedded Networked Sensor Systems (SenSys), Nov. 2006. D. Haban and D. Wybranietz, “A Hybrid Monitor for Behaviour and Performance Analysis of Distributed Systems,” IEEE Trans. Software Eng., vol. 16, no. 2, pp. 197-211, Feb. 1990. W.R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-Efficient Communication Protocol for Wireless Microsensor Networks,” Proc. Hawaii Int’l Conf. System Sciences, Jan. 2000. C. Intanagonwiwat, R. Govindan, and D. Estrin, “Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks,” Proc. ACM MobiCom, pp. 56-67, 2000. “Heterogeneous Networks with Intel XScale,” http://www.intel. com/research/exploratory/heterogeneous.htm, 2008. D. Janakiram, A.V.U.P. Kumar, and A.M. Reddy, V, “Component Oriented Middleware for Distributed Collaboration Event Detection in Wireless Sensor Networks,” Proc. Int’l Workshop Middleware for Pervasive and Ad-Hoc Computing (MPAC ’05), Nov./Dec. 2005. E. Kranakis, H. Singh, and J. Urrutia, “Compass Routing on Geometric Networks,” Proc. 11th Canadian Conf. Computational Geometry, pp. 51-54, Aug. 1999. F. Kuhn, R. Wattenhofer, and A. Zollinger, “Worst-Case Optimal and Average-Case Efficient Geometric Ad Hoc Routing,” Proc. Fourth ACM Int’l Symp. Mobile Ad-Hoc Networking and Computing, pp. 267-278, 2003. S. Khuller, B. Raghavachari, and N. Young, “Balancing Minimum Spanning and Shortest Path Trees,” Algorithmica, vol. 120, no. 4, pp. 305-321, 1994. S. Kumar, T.H. Lai, and J. Balogh, “On k-Coverage in a Mostly Sleeping Sensor Network,” Proc. ACM MobiCom, 2004. Q. Li, J. Aslam, and D. Rus, “Online Power-Aware Routing in Wireless Ad-Hoc Networks,” Proc. ACM MobiCom, 2001. S. Li, Y. Lin, S.H. Son, J.A. Stankovic, and Y. Wei, “Event Detection Services Using Data Service Middleware in Distributed Sensor Networks,” Telecomm. Systems, vol. 26, 351-368 2004. C. Liebig, M. Cilia, and A. Buchmann, “Event Composition in Time Dependent Distributed Systems,” Proc. Int’l Conf. Cooperative Information Systems (COOPIS), Sept. 1999. S. Lindsey, C. Raghavendra, and K. Sivalingam, “Data Gathering in Sensor Networks Using the Energy Delay Metric,” Proc. 15th Int’l Parallel and Distributed Processing Symp., Apr. 2001. S. Lindsey and C.S. Raghavendra, “PEGASIS: Power-Efficient Gathering in Sensor Information Systems,” Proc. IEEE Aerospace Conf., vol. 3, Mar. 2002. X. Liu, Q.F. Huang, and Y. Zhang, “Combs, Needles, Haystacks: Balancing Push and Pull for Discovery in Large-Scale Sensor Networks,” Proc. Conf. Embedded Networked Sensor Systems (SenSys ’04), Nov. 2004. S. Madden, S. Szewczyk, M.J. Franklin, and D. Culler, “Supporting Aggregate Queries over Ad-Hoc Sensor Networks,” Proc. Workshop Mobile Computing and Systems Application, 2002. S. Madden, M.J. Franklin, J.M. Hellerstein, and W. Hong, “The Design of an Acquisitional Query Processor for Sensor Networks,” Proc. ACM SIGMOD, June 2003. A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson, “Wireless Sensor Networks for Habitat Monitoring,” Proc. Int’l Workshop Wireless Sensor Networks and Applications, 2002. A. Manjeshwar and D.P. Agrawal, “TEEN: A Routing Protocol for Enhanced Efficiency in Wireless Sensor Networks,” Proc. 15th Int’l Parallel and Distributed Processing Symp., Apr. 2001. J. Newsome and D. Song, “GEM: Graph Embedding for Routing and Data-Centric Storage in Sensor Networks without Geographic Information,” Proc. First ACM Conf. Embedded Networked Sensor Systems (Sensys ’03), Nov. 2003. B. Oki, M. Pfluegl, A. Siegel, and D. Skeen, “The Information Bus—an Architecture for Extensible Distributed Systems,” Proc. ACM SIGOPS, 1993. T.S. Rappaport, Wireless Communications: Principles and Practice. Prentice Hall, 1996. K. Ro¨mer and F. Mattern, “Event-Based Systems for Detecting Real-World States with Sensor Networks: A Critical Analysis,” Proc. DEST Workshop Signal Processing in Sensor Networks at ISSNIP, pp. 389-395, Dec. 2004.
LI ET AL.: DELAY-BOUNDED AND ENERGY-EFFICIENT COMPOSITE EVENT MONITORING IN HETEROGENEOUS WIRELESS SENSOR...
[34] M. Mansouri-Samani and M. Sloman, “A Generalized Event Monitoring Language for Distributed Systems,” IEE/IOP/BCS Distributed Systems Eng. J., vol. 4, no. 25, Feb. 1997. [35] T. Schoellhammer, B. Greenstein, and D. Estrin, “Hyper: A Routing Protocol to Support Mobile Users of Sensor Networks,” Technical Report 63, Center for Embedded Network Sensing, Univ. of California, Jan. 2006. [36] R.C. Shah and J.M. Rabaey, “Energy Aware Routing for Low Energy Ad Hoc Sensor Networks,” Proc. Wireless Comm. and Networking Conf. (WCNC ’02), vol. 1, pp. 17-21, Mar. 2002. [37] I. Stojmenovic and X. Lin, “Power-Aware Routing in Ad Hoc Wireless Networks,” IEEE Trans. Parallel and Distributed Systems, vol. 12, no. 11, pp. 1122-1133, Nov. 2001. [38] A.Y. Wang, S.H. Cho, C.G. Sodini, and A.P. Chandrakasan, “Energy Efficient Modulation and MAC for Asymmetric RF Microsensor Systems,” Proc. Int’l Symp. Low Power Electronics and Design (ISLPED), pp. 106-111, 2001. [39] H. Wang, D. Estrin, and L. Girod, “Preprocessing in a Tiered Sensor Network for Habitat Monitoring,” EURASIP J. Applied Signal Processing, vol. 4, pp. 392-401, 2003. [40] Z.Z. Wei, W. Zhao-hui, and L. Huai-zhong, “An Event-Driven Clustering Routing Algorithm for Wireless Sensor Networks,” Proc. Int’l Conf. Intelligent Robots and Systems, 2004. [41] J.E. Wieselthier, G.D. Nguyen, and A. Ephremides, “On the Construction of Energy-Efficient Broadcast and Multicast Trees in Wireless Networks,” Proc. IEEE INFOCOM, pp. 585-594, 2000. [42] M. Yarvis, N. Kushalnagar, H. Singh, A. Rangarajan, Y. Liu, and S. Singh, “Exploiting Heterogeneity in Sensor Networks,” Proc. IEEE INFOCOM, 2005. [43] F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang, “A Two-Tier Data Dissemination Model for Large-Scale Wireless Sensor Networks,” Proc. ACM MobiCom, Sept. 2002. [44] Y. Yu, R. Govindan, and D. Estrin, “Geographical and Energy Aware Routing: A Recursive Data Dissemination Protocol for Wireless Sensor Networks,” Technical Report UCLA/CSD-TR-010023, UCLA Computer Science Dept., May 2001. [45] A. Zygmund, Trigonometric Series, third ed., pp. 21-22. Cambridge Univ. Press, 2002. [46] J. Rabaey, J. Ammer, J.L. da Silva, Jr., and D. Patel, “PicoRadio: Ad-Hoc Wireless Networking of Ubiquitous Low-Energy Sensor/ Monitor Nodes,” Proc. IEEE CS Ann. Workshop VLSI, pp. 9-12, Apr. 2000. [47] G.J. Pottie and W.J. Kaiser, “Wireless Integrated Network Sensors,” Comm. ACM, vol. 43, no. 5, pp. 51-58, 2000. [48] G. Xing, X. Wang, Y. Zhang, C. Lu, R. Pless, and C.D. Gill, “Integrated Coverage and Connectivity Configuration for Energy Conservation in Sensor Networks,” ACM Trans. Sensor Networks, vol. 1, no. 1, pp. 36-72, 2005. [49] Z. Zhou, S.R. Das, and H. Gupta, “Connected K-Coverage Problem in Sensor Networks,” Proc. 13th Int’l Conf. Computer Comm. and Networks, Oct. 2004. Yingshu Li received the BS degree in computer science from Beijing Institute of Technology, China, and the MS and PhD degrees in computer science from the University of Minnesota at Twin Cities. She is currently an assistant professor in the Department of Computer Science at Georgia State University. Her research interests include wireless networks, optimization theory, and data management in sensor networks. She is the recipient of the US National Science Foundation (NSF) CAREER Award. She is a member of the ACM and the IEEE. Chunyu Ai received the BS and MS degrees in computer science from Heilongjiang University, China, in 2001 and 2004, respectively. She is currently working toward the PhD degree in the Department of Computer Science at Georgia State University, Atlanta. Her current research interests include wireless sensor networks and data stream. She is a student member of the IEEE.
1385
Chinh T. Vu received the BS degrees in electronics and information technology from Hanoi University of Technology and Hanoi Open University, Vietnam, in 2000, and the MS and PhD degrees from the Department of Computer Science at Georgia State University, Atlanta, in 2007 and 2009, respectively. His current research interests include wireless sensor networks, approximation algorithms, and distributed systems. He is a student member of the IEEE.
Yi Pan received the BEng and MEng degrees in computer engineering from Tsinghua University, China, in 1982 and 1984, respectively, and the PhD degree in computer science from the University of Pittsburgh, Pennsylvania, in 1991. He is the chair and a professor in the Department of Computer Science and a professor in the Department of Computer Information Systems at Georgia State University. His research interests include parallel and distributed computing, networks, and bioinformatics. He has published more than 100 journal papers of which more than 40 papers have been published in various IEEE journals. In addition, he has published more than 150 papers in refereed conferences. He has also authored/edited 34 books (including proceedings) and contributed many book chapters. He has served as an editor-in-chief or editorial board member for 15 journals including six IEEE Transactions, and a guest editor for 10 journals. He has organized many international conferences and workshops and has also served as a program committee member for several major international conferences such as INFOCOM, GLOBECOM, ICC, ICDCS, IPDPS, and ICPP. He has delivered more than 10 keynote speeches at many international conferences and is a speaker for several distinguished speaker series. He is listed in Men of Achievement, Who’s Who in Midwest, Who’s Who in America, Who’s Who in American Education, Who’s Who in Computational Science and Engineering, and Who’s Who of Asian Americans. He is a senior member of the IEEE. Raheem Beyah received the bachelor of science degree in electrical engineering from North Carolina A&T State University in 1998, and the master’s and PhD degrees in electrical and computer engineering from the Georgia Institute of Technology in 1999 and 2003, respectively. He is an assistant professor in the Department of Computer Science at Georgia State University, where he leads the Georgia State Communications Assurance and Performance Group (CAP). He is also an adjunct professor in the School of Electrical and Computer Engineering at the Georgia Institute of Technology. Prior to joining Georgia State in 2005, he was a research faculty member with the Georgia Institute of Technology’s Communications Systems Center (CSC) for four years and remains a part of the Center. He also worked as a consultant in Andersen Consulting’s (now Accenture) Network Solutions Group. He is an associate editor of several journals including the (Wiley) Security and Communication Networks Journal and the (Wiley) Wireless Communications and Mobile Computing Journal. His research interests include network security, wireless networks, network traffic characterization and performance, and security visualization. He received the US National Science Foundation CAREER award in 2009. He is a member of the ACM, the NSBE, and a senior member of the IEEE.
. For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.
