Trade-off Between Coverage and Data Reporting Latency for Energy-Conserving Data Gathering in Wireless Sensor Networks

Transcription

1 Trade-off Between Coverage and Data Reporting Latency for Energy-Conserving Data Gathering in Wireless Sensor Networks Wook Choi and Sajal K. Das Center for Research in Wireless Mobility and Networking (CReWMaN) Department of Computer Science and Engineering The University of Texas at Arlington Abstract In this paper, we propose a novel energy-conserving data gathering strategy for wireless sensor networks. The proposed strategy is based on a trade-off between coverage and data reporting latency with an ultimate goal of maximizing the network s lifetime. The basic idea is to select in each round only a minimum of sensors as data reporters which are sufficient for a desired sensing coverage given by the users or applications. Such a selection of the minimum data reporters also reduces the amount of traffic flow to the data gathering point in each round, and thus avoids network congestion as well as channel interference/contention. The proposed strategy includes three schemes for the minimum -sensor selection. Using these schemes we evaluate such fundamental issues as event detection integrity and data reporting latency, which can be critical in deploying the proposed data gathering strategy. Simulation results demonstrate that the average data reporting latency is hardly affected and the real-time event detection ratio is greater than 8% when the desired sensing coverage is at least 8%. It is also shown that the sensors can conserve a significant amount of energy with a small trade-off, and that the higher the network density, the higher is the energy conservation rate without any additional computation cost. 1 Introduction Wireless sensor networks are task-specific information gathering platforms. They can be deployed both indoors and outdoors, substituting for our sensory organs in inaccessible or inhospitable areas. Depending on the deployment platform of sensor networks, there is a variety of applications such as environment or equipment monitoring [15], smart home/smart space [5], intrusion detection, and surveillance, etc. Such sensor networks can be characterized by high node density and highly limited resources such as bandwidth, energy, computational capability, and storage space. This distinguishes the sensor networks from the traditional ad hoc networks [1]. The sensors sense their vicinity (sensing coverage) and deliver sensed data to a data gathering point, consuming the limited energy resource which may not often be possible to replenish. Therefore, an important challenge in designing data gathering protocols for sensor networks is to make them highly energyefficient so as to maximize their lifetime. The frequency of data delivery depends on the models which can be classified into continuous, event-driven, on-demand, or hybrid, based on the application s or the user s interest [19]. The continuous model requests all sensors to transmit their sensed data periodically while they are alive. A cluster-based continuous data gathering scheme, called LEACH, is proposed in [9]. Further improvement on energy-conservation achieved by the LEACH is shown in [14] which connects all the sensors as a linear chain. Forming a chain requires sensors to have a global knowledge which makes the data gathering scheme unscalable. The clustering scheme requires sensors to consume a certain amount of energy while forming and maintaining clusters. Moreover, the role of cluster head leads to a relatively large amount of energy consumption as compared to an ordinary cluster member (i.e., non-cluster head). Thus rotating the role of cluster head is necessary to reduce the time variance of sensor failures caused by energy depletion. We proposed a two-phase clustering (TPC) in [4] which reduces the cluster head s workload and thus the cluster head rotation by requiring cluster members to maintain two types of paths to the cluster head: direct link (one-hop) and data relay link (multi-hop). Unlike the continuous model, in the event-driven data gathering model [11], sensors start reporting their sensed data only when a specific event occurs. Whereas, in the ondemand model [16] they report sensed data only at the users request. Due to high density of the network, it is common for multiple sensors to generate and transmit redundant sensed data which results in unnecessary power consumption and hence significantly decreases the network s lifetime. Among the sensors actions, such as data transmission and target sens /4/$2. 24 IEEE 53

2 ing, the energy consumption for wireless data transmission is the most critical. Therefore, minimizing the number of data transmissions between sensors by eliminating redundant data without losing data accuracy, or aggregating multiple sensed data saves a significant amount of energy. For this purpose, many protocols have been designed for routing and managing topological connectivity, as summarized below. Data-centric routing [12] attempts to reduce duplicate data transmissions by aggregating multiple packets cached for a certain amount of time (i.e., in-network data processing with some data transmission delay), thereby increasing the energy conservation. In [18], a sensing coverage preserving scheme is proposed which turns off the sensors having their coverage area overlapped with other sensors. More recently, two elegant algorithms, called connected sensor cover [8] and coverage configuration protocol [2] have been proposed considering coverage and connectivity problems simultaneously. The former selects a minimum number of sensors to cover a specified area for query execution, thus reducing unnecessary energy consumption from redundant sensing. The latter selects a minimum number of sensors to guarantee that any point within a monitored area is covered by at least sensors. These protocols find a relatively small set of (connected) sensors by running an algorithm with relatively high computational complexity, exchanging control information with local neighbors to cover the entire monitored area by 1%. The execution and implementation of such algorithms, however, are challenging because the sensors are basically under the highly limited resource environment. In fact, finding the smallest set of connected sensors that completely cover a given monitored area is an NP-hard problem [8]. 1.1 Our Contributions To this end, we propose a novel energy-conserving data gathering strategy for mainly the continuous data gathering model, based on a trade-off between coverage and data reporting latency with an ultimate goal of maximizing the network lifetime. The proposed strategy attempts to select at every data reporting round only a minimum of sensors as data reporters which are sufficient to cover as much of the monitored area as the user/application requests. Only these sensors transmit data to the gathering point while the others cache their sensed data waiting for the next reporting round, thus saving energy. All the sensors take turns in being selected as a data reporter. Thus, the parts of the area not covered by the first set of selected sensors will be covered by the next set of selected sensors with some delay. The lower the desired sensing coverage, the longer is the data reporting latency in each sensor; whereas the energy conservation rate is inversely proportional to the coverage. Besides the enhanced energy conservation, there is a subsequential benefit such as congestion avoidance and low channel interference/contention, which can be achieved by our proposed strategy using only sensors in each reporting round. This also contributes to energy savings, improving the overall network performance. The proposed strategy adopts three schemes for -sensor selection: nonfixed randomized selection (), non-fixed, and fixed disjoint randomized selections ( and F-). They differ from one another in terms of data reporting latency and implementation simplicity. The computational complexity of these three sensor selection schemes is constant (i.e., independent of network density and size), thereby providing a high scalability. In addition, they do not exchange (periodic) control information with local neighbors in selecting sensors. Thus, the proposed strategy is well suited for sensor networks which are required to run for a long time under highly limited resource constraints. Through intensive simulation studies, we evaluate fundamental issues such as event detection integrity and data reporting latency which are critical in deploying the proposed strategy. Simulation results demonstrate that ) the proposed schemes can meet the desired sensing coverage by making approximately sensors to report their sensed data in each reporting round and ) sensors can conserve a significant amount of energy with a small trade-off. More specifically, in a network field in which sensors have 3 circular sensing range, the real-time event detection ratio is more than 9% using only sensors, which are selected based on an 8% desired sensing coverage of the entire monitored area. It is also shown that the average sensed data reporting latency is hardly affected when the desired coverage is greater than. Furthermore, since the selection of sensors is not affected by the network density, the energy conservation rate increases without any additional computation cost as the network size grows. The remainder of this paper is organized as follows. Section 2 presents the motivation and problem under consideration. Section 3 introduces basic definitions and assumptions. Section 4 describes how to find the minimum sensors that meet the users desired sensing coverage. Section 5 presents three -sensor selection schemes for the deployment of the proposed data gathering strategy. Section 6 discusses simulation results and Section 7 concludes the paper. 2 Motivation and Problem Description In this paper, we focus on enhancing energy conservation while meeting the user/application s requirements such as data delivery latency and desired sensing coverage. We consider (uniformly distributed) randomly deployed sensor networks but the application of the proposed strategy is not limited to such networks only. Our motivation lies in the fact that depending on the type of applications used, the network lifetime can be much more critical than covering the entire monitored area at every data reporting round. The user may 54

3 &? B 8 8?.?? desire that only a certain portion of the area be covered at every data reporting round for the extended network lifetime if the sensed result for the entire monitored area can be acquired with a fixed delay. One example is: for a sensor network deployed for a statistical study of scientific measurement in a certain area, it may be accurate enough to monitor the status of the specific area s certain condition if the network covers approximately 8% of the field on an average in each round. Another example is: for a sensor network monitoring slowlymoving objects, it may be acceptable if the network covers only 5% of the area at every round on the condition that the sensed result covering the entire monitored area can be collected with a fixed delay. s 1 s 2 s 4 s 3 (a) s 5 s 6 Monitored Area s 1 s 2 s 3 s 4 Figure 1. Illustration of Coverage-Data Reporting Latency Trade-off Based Data Gathering The problem is associated with the selection of a minimum number of sensors, based on a desired sensing coverage specified by the user. Figures 1 (a) and (b) illustrate the problem. The black solid dots within the small circles (i.e.,!!! $ ) in both of the figures and the hollow dots in the figure (b) represent the currently selected sensors and the previously selected sensors, respectively. The large solid-line circle represents the sensing coverage of each sensor. Suppose that the first selected % sensors cover a desired portion of the area but not the entire sensing area, as shown in Figure 1 (a) (i.e., shaded area is not covered). The shaded area is being covered in the second set of selected sensors as shown in Figure 1 (b). Therefore, the user receives the sensed result for the entire monitored area with a fixed delay (i.e., two consecutive reporting rounds). We thus define the problem as follows: & ' ( & +, ,. 8 : ' 8 < Problem Definition: Given a set of sensors which are placed over a region such that each has sensing region. A minimum of sensors has to be chosen from such that desired sensing coverage. To solve this problem, we apply a geometric probability theory dealing with random circles on geometrical figures [7] which is the study of probabilistically measuring how much s 6 (b) s 5 area will be overlapped when a circle is randomly placed over a geometrical figure. 3 Basic Definitions and Assumptions A large number of sensors is densely-deployed in a twodimensional geographic space, forming a network. Although there is a feasible means to make the sensors aware of their location, such as global positioning system (GPS) or directional beaconing [2, 17], we do not assume that sensors are located by any specific coordination system because such localization mechanisms may not be available or practical in building lowcost and low-power sensors with small form factor. Formally, we shall define a sensor network as an undirected connected graph > 8, where? are the sets of nodes (sensors and data gathering points) and edges (bidirectional wireless links), respectively. A sequence of edges in > forms the path, A!!!. E. 8 G 8 8 for H H L L N, where. ( is a sensor and G ( is a data gathering point. Thus, A is considered as a multihop routing path and each node. on A acts as an individual router. A sensor node. generates a fixed-size data packet for a time unit as a sensed result. We call this time unit as a data reporting round and the interval between two consecutive data reporting rounds is denoted by Q R. All the sensor nodes are supposed to forward the generated data packet to the data gathering point using a routing path A, making the communication pattern many-to-one. Since our proposed scheme is considered as a data gathering protocol running on top of the routing layer, in this work we assume a non-geographical sensor routing protocol which connects all sensors at the deployment time [6, 1]. A control message from G is delivered to the sensor nodes through flooding [11]. Each node. ( has its specific radio and sensing ranges with radius V. Both of the ranges are denoted by W X Z. A sensor. can directly communicate with any nodes in its radio range W X Z. 4 Minimum ^ Data Reporters In this paper, the term desired sensing coverage (DSC) represents a probabilistic percentage for covering any point within the entire monitored area. The user specifies the DSC as the desired quality of service to be achieved by sensor data gathering. Thus, we define the desired sensing coverage as a trade-off factor for energy conservation. The DSC is proportional to the amount of sensed data traffic over the network and inversely proportional to both the energy conservation rate and data reporting latency. The question is: in order to meet the DSC specified by the user, how many sensors do we need to select at each data reporting round? To answer this question, let us first introduce the following basic definitions: Definition 4.1: A monitored area, denoted by _, is the actual 55

4 g j g b 4 Œ 4 j 4 4 area which has to be monitored by the sensors. We consider this area as an ` ` square. Definition 4.2: A sensor-deployed area, denoted by b, is a square area including all sensors which have an effect on covering_ such that the square will have rounded corners with distance less than or equal tov (radius of sensing range) from the boundary of_ c b (refer to Figure 2). Thus, the circular sensing range of a sensor residing inb N _ is not fully overlapped with the area_. Definition 4.3: A probabilistic sensing coverage, denoted by f, is the probability of any point in _ being covered by at least one of the selected sensors (residing in b ) circular sensing range with radiusv. This is given by either the user or the application as the desired sensing coverage. When sensors are randomly deployed over_, it is likely that they will be also placed a bit beyond the boundary of_ in order to completely cover the monitored area_ or due to the deployment inaccuracy. Thus, there is a separate definition of the sensor-deployed area,b, in measuring the probabilistic sensing coverage. Figure 2 illustrates the basic terms defined above. A circle centered at a point at the bottom-left corner in the sensor-deployed area is a sensing coverage of a sensor with radiusv. r Sensor-Deployed Area Λ Monitored Area Q Figure 2. Illustration of Monitored and Sensor- Deployed Areas Letg h _ be the part of_ which will be covered by the circular sensing ranges of H L? L N number of sensors residing inb. Then, the fraction kj is the user s desired sens- l ing coverage at each reporting round. Any point m 8 ( _ is considered to be covered if it is inside the circular sensing coverage of a selected sensor in the sensor-deployed area, b. To measure the probabilistic l sensing coverage, we first mea- m 8 l that a point m 8 ( _ will not sure the probabilitya j l be covered by a selected sensor,.. LetW l m 8 be a circular area centered at point m 8 with radius V. Then, the point will not be covered when l. ( b N W l m 8. Therefore, the probability that the point m 8 is not covered by a randomlyselected sensor is given by A j l m 8 B q s E t vx z{ q } l m 8~ l ~ m (1) where} l l m 8 B s is the probability that. is located on a point m 8 ( b. Eq. (1) represents the fraction ofb not covered by a randomly-selected sensor s circular sensing range. Thus, the probability that a point is not covered by randomlyselected sensors is obtained as A 4 l m 8 B.6 A j l m 8 8. Letg be the area of_ not covered. For randomly selected sensors, the expected value ofg can be given g ƒ B q k q A 4 l m 8~ l ~ m (2) As mentioned earlier, we consider how much area in_ can be covered by randomly-selected sensors. For this purpose, we first consider the fraction of_ not covered by these sensors within_. This can be obtained by dividing@ g ƒ by the area` of_ l, assuming all m 8 points are uniformly (randomly) distributed over _. Applying Eqs. (1) and (2), the fraction of_ not covered by selected sensors is given ` B b N W l m 8 B ` ` ` V ` V V Œ Finally, when sensors are randomly selected from_, the probabilistic sensing coverage (f ) that any point of_ will be covered by at least one of selected sensors circular sensing range is given by: f B ` B N ` ` ` V ` V V (3) Œ B N f 8 š š œ Ÿ œ Ÿ (4) Ÿ š 8 Therefore, the smallest integer which meets the desired sensing coverage,f, can be defined as: In order to verify the correctness in measuring, we simulate the analytical model and compare the simulation results with the numerical results measured from Eq. (3). Figures 3 (a) and (b) show the comparison results in covering a requested portion of the monitored area with varying network sizes and sensor s circular sensing ranges. The simulation results shown in each plot correspond to the average of 1 simulation runs. Regardless of the sizes of the network and sensing range, we observe in Figures 3 (a) and (b) that both the numerical and simulation results are found to match well. 5 ^ -Sensor Selection Schemes Based on the user s DSC, the size of the monitored area, and the sensor s sensing range, Eq. (4) gives the required minimum number ( ) of sensors which should be selected from 56

5 4? Monitored Area Coverage Ratio Sensing Range = 3m Analytical (1x1m 2 ) Simulation Analytical (2x2m 2 ) Simulation Analytical (3x3m 2 ) Simulation Number of Selected Sensors (a) Monitored Area Coverage Ratio Network Size = 2 x 2m 2 Simulationl (Sensing Range: 2m) Analytical Simulation (Sensing Range: 3m) Analytical Simulation (Sensing Range: 4m) Analytical Number of Selected Sensors Figure 3. Comparison of Simulation and Analytical Results for Covering a Monitored Area all the sensors uniformly distributed over the sensor-deployed area, b. Obviously, sensors will experience a certain latency in reporting their sensed data since they are allowed to report their sensed data only when they become one of the selected sensors. We use randomization techniques to select those reporting sensors in each reporting round. Depending on the randomized selection technique used, the reporting latency will vary. In the following we introduce three distributed - sensor selection schemes which are different from one another in terms of the data reporting latency and implementation simplicity. 5.1 Non-disjoint Randomized Selection () In, sensors elect themselves to be one of the sensors based on a probability, A Ÿ X B E. Each sensor draws a random number uniformly distributed within ƒ in each data reporting round; if the random number is less than or equal to A Ÿ X, the sensor becomes a data reporter. Thus, approximately data reporting sensors are selected in every reporting round. Since the DSC is a probabilistic percentage for covering the monitored area, the overall sensing coverage ratio is not much different from the one that has exactly sensors. More about this will be shown later while presenting experimental results. Theoretically speaking, a sequence of independent random selection trials to become a reporting sensor with the probability A Ÿ X can be modeled as a geometric distribution having A Ÿ X as success probability. Thus, the expected number of the trials for a success is. This implies that the data reporting (b) latency (i.e., elapsed time to be selected as a reporter again) could be very large when the DSC is small. In other words, the scheme can not guarantee gathering the sensed result for the entire monitored area with a fixed delay. Hence, this scheme is inappropriate for time-constrained monitoring for the entire area. Furthermore, the does not have any control operation for already-selected sensors before all sensors take their turn to report, which implies a set of selected sensors at the current reporting round may not be totally disjoint from the set selected in the previous round. However, due to the long term steady state behavior of the randomized selection, all the sensors in the sensor-deployed area b will approximately have the same number of chances in transmitting the sensed data. 5.2 Disjoint Randomized Selection () Unlike the, scheme covers the entire monitored area with a fixed delay by allowing all the sensors to have a chance to report their sensed data within the fixed number of data reporting rounds. Once the sensors report their sensed data, they do not elect themselves as a reporter again until all the sensors have had an equal chance to report. Based on and total number of sensors (L L N ), each sensor finds a cycle, denoted by «, of electing itself as a data reporter. The cycle«is composed of B E 4 reporting rounds. All the sensors select a data reporting round within reporting rounds at every «and become a data reporter for the selected round only, thus guaranteeing all the sensors report their sensed data with a fixed delay (i.e., within «). The users may consider some other constraints along with the desired sensing coverage while receiving sensed data from the sensors. For example, the entire monitored area is required to be monitored with a uniform delay even though the entire area is not covered in each round, or the monitoring pattern should not be learned by an adversary attempting to circumvent the sensing activity. That is, the uncovered part of the entire monitored area at every reporting round should be unexpected (i.e., random). In light of this we provide two selection schemes: non-fixed disjoint randomized selection (N- ) and fixed disjoint randomized selection (F-). Both guarantee a disjoint set of sensors in each round of «. ² Non-fixed disjoint randomized selection (): at every «, each sensor elects itself as a reporter by drawing a round randomly within reporting rounds of cycle «so that each set of selected sensors at each round is memoryless. Therefore, the monitoring pattern can not be known beforehand. The reporting latency of each sensor in ranges from Q R to N Q R (recall that Q R is the time interval between two consecutive reporting rounds). ² Fixed disjoint randomized selection (F-): similarly to, sensors choose a reporting round randomly from reporting rounds. However, in F-, sensors do not have a reporting round selection procedure at every «. They keep the initially selected reporting round. Therefore, at every cycle «, all the sensors have the same order of reporting round and hence a uniform fixed data reporting latency of Q R. 57

6 ? L Î Ï We first define a reporting sequence to introduce the above disjoint set selection schemes. Definition 5.1: A reporting sequence of a sensor., denoted by ¹ º X Z, is a sequence of bits. Each bit maps to each round of the cycle «and hence the number of bits is equal to the number of rounds in «. The sequence is initialized to zero (i.e., off) and only one bit will be flipped (i.e., on) depending on the selection result of a round within «. Thereby, ¹ º X Z indicates which round within «the sensor. should report its sensed data. For example, ¹ º X Z B ½ ¾ for B and L N B, represents that sensor. is a reporter in the third round of «. Algorithm Construct RS (Â, Ã Ä Ã ) 1: Å Æ È Ê Ë Ê Ì Í 2: Allocate Ñ Ò Å Ô and initialize Ñ Ò Å Ô with zero 3: Õ Æ Ö Ò Ù Å Ô 4: Ñ Ò Õ Ô Æ 5: return Ñ Ò Å Ô /* reporting bit sequence with one bit on */ End Algorithm Figure 4. Algorithm for Constructing Reporting Sequence (¹ º X Z ) in Sensor. Figure 4 shows an algorithm for constructing ¹ º X Z where W ƒ is a bit array of length and, ƒ is a function which returns an integer between and based on uniform random distribution. Sensors report their sensed data only when a bit corresponding to the current round in its W ƒ is equal to. Thereby, the selected sensors are disjointed at every round and all the sensors transmit their sensed data exactly once within a reporting cycle «. The actual delay time in acquiring the sensed result for the entire monitored area depends on DSC since is decided by measured from the DSC. Hence, the higher the desired sensing coverage, the shorter is the monitoring time of the entire area, whereas energy conservation rate is inversely proportional to DSC. After a reporting cycle «, if sensors are based on the scheme, they acquire a new reporting bit sequence ¹ º X Z by running the algorithm in Figure 4. Otherwise (i.e., for the F- scheme), they keep the initial reporting bit sequence ¹ º X Z until an update request is received. Due to the time-bounded data gathering characteristic of the entire monitored area, these two schemes can be used more widely than the scheme. 5.3 Discussions on Routing The routing is the most energy-consuming operation in a multihop wireless sensor network. Hence, the frequency of the routing service in sensors is a critical factor in determining their remaining lifetime. The amount of total traffic which has to be routed to a data gathering point over the network depends on the number of sensors wishing to report their sensed data in each round. Therefore, the routing load (RL) in the network is expressed as: ¹ Ý B Þ à X Z á â N ã 8 å X Z where å X Z is the sensed data generation rate in a sensor., æ is a set of sensors which have to report sensed data (L æ L B in our case), and ã H is the ratio of data redundancy, which depends on the sensor density and sensed data reporting interval. Then the total routing load, ¹ Ý, in each round is equivalent to the total number of sensed data packets generated by all. ( æ. Let ë be the number of packets a sensor has to forward as a router toward the data gathering point. Then the individual routing load of a sensor serving as a router is ì í â H ë H ¹ Ý if å X Z and ã are the same for all. ( æ, i.e., the individual routing load depends on how many routes the sensor is involved in as a router. Sensor failures may disconnect the network or reduce the quality of monitoring. Thus, the variance of ë among sensors is a critical factor which should be taken care of with a high priority for the network longevity. In the case that a minimum number of sensors is found to meet the desired sensing coverage, the additional number of sensors involved for the routing will be minimum if each of the data reporting sensors chooses the shortest path to the data gathering point and the jointedness of the chosen paths becomes maximum. However, this is clearly not an optimal routing for the network longevity. The optimal routing is to dynamically distribute RL over multiple sensors based on their remaining energy level even though more sensors are involved [3]. Such an optimal routing is needed for any data gathering scheme without regard to the optimality of the number of active sensors for sensing coverage and connectivity. In this paper, our main focus is to find a minimum number of sensors meeting the DSC and selection mechanisms of them. Currently, we are investigating a construction of data gathering tree which meets the DSC and simultaneously handles selected sensor connectivity, routing, and sensor state scheduling problems, as a part of our future work. 6 Performance Evaluation In this section we evaluate the performance of the proposed coverage-data reporting latency trade-off based energyconserving data gathering strategy using three -sensor selection schemes:,, and F-. A discrete time event simulator, called simlib [13], was used to implement the proposed sensor selection schemes and to collect statistical information. As mentioned earlier, we focus on fundamental issues of the proposed data gathering strategy in this study 58

7 î ² ² ² î B î z z î š z š Event Detection Ratio Sensing Range = 3m FRS (1x1m 2 ) FRS (2x2m 2 ) FRS (3x3m 2 ) Desired Sensing Coverage (a) Event Detection Ratio Sensing Range = 4m Desired Sensing Coverage (b) FRS (1x1m 2 ) FRS (2x2m 2 ) FRS (3x3m 2 ) Number of Selected Sensors (k) x1m 2 2x2m 2 3x3m 2 Sensing Range: 3m Desired Sensing Coverage Figure 5. Event Detection Ratio and Corresponding Number of Reporting Sensors ( ) (c) such as event detection integrity and sensed data reporting latency. We also present the energy conservation capability of the schemes. 6.1 Methodology We measure the performance of the proposed schemes by evaluating the following three metrics with desired sensing coverages varying from.1 to.9 with.1 interval: Real-time event detection ratio: The fraction of the number of event occurrences detected by at least one sensor. Events occur in the monitored area based on uniform distribution. Only selected sensors report detected events in each round. Thus, an event occurs and none of the sensors detects that event, it is considered as a failure of real-time event detection. Data reporting latency: The time difference between an event detection and its reporting to the data gathering point. All sensors are sensing the monitored area but only selected sensors report what they sensed to the data gathering point in each round and the others cache sensed result until they become a reporter, thus introducing the data reporting latency in reporting what they sensed. Energy conservation capability: The number of data transmissions during a time unit. The energy consumed for data transmission is the most critical in determining the sensor s lifetime. Thus, we consider the number of data transmissions in each sensor during a simulation time as an index of the energy conservation capability. We compare each metric measured from the proposed schemes and a common data gathering () scheme in which all the sensors transmit their sensed data to the data gathering point at every data reporting round. We generate three sensor network fields:,, and. Homogeneous sensors are scattered randomly and uniformly over the network field based on the density, 1 sensor/. Therefore, the total number of sensors for,, and î becomes 5, 2, and 45, respectively. We run 1 experiments, each with different sensor distributions, for each desired sensing coverage in each network field. Two different circular sensing ranges are used: V and. Events are occurring in the network based on uniform distribution and their occurrence interval is uniformly distributed between 5 and 1 seconds. Table 1 summarizes the parameters used. Note that the implementation does not include any feature of MAC layer and wireless channel characteristics since our main objective in this simulation study is to measure the real-time event detection capability and data reporting latency when the desired sensing coverage is specified by the user. Network Field Table 1. Simulation Parameter Values ï ï ï ï ð š œ ï ï ï ï ð š ñ ï ï ï ï ð Number of Sensors ï ï ð Sensor Density 1sensorò ï ð Sensing Range and œ ï ð Event Occurrence Event Occurrence Interval Uniform Distribution Uniformly distributed within [5,1] seconds Data Reporting Interval ó ô 1 seconds Simulation Time 16, seconds 6.2 Real-Time Event Detection Ratio Sensors are continuously monitoring the environment to detect specific events or to collect statistical information for the user. Figures 3 (õ ) and (ö ) showed that only sensors are sufficient to cover as much of the area as the user requests. What about detecting specific events occurring irregularly over the entire monitored area? We evaluated the realtime event detection ratio using only selected sensors. Intuitively, the event detection ratio will be approximately the same as the desired sensing coverage if the selected sen- 59

8 ý ý 4 Avg. Data Reporting Latency (second) x1m 2, Sensing Range = 3m F (a) Avg. Data Reporting Latency (second) x2m 2, Sensing Range = 3m F (b) Avg. Data Reporting Latency (second) x3m 2, Sensing Range = 3m (c) F- Max. Data Reporting Latency (second) x1m 2, Sensing Range = 3m (d) F- Max. Data Reporting Latency (second) x2m 2, Sensing Range = 3m (e) F- Max. Data Reporting Latency (second) x3m 2, Sensing Range = 3m (f) F- Figure 6. Data Reporting Latency (Avg. and Max.) in Simulated Network Fields sors meet the desired sensing coverage. We measure the realtime event detection ratio in two cases: fixed random selection (FRS) which has exactly ý sensors and the scheme, and compare the results. The FRS was only simulated as a counterpart to show the impact from having approximately sensors upon achieving the DSC. We count the number of events detected by ý sensors randomly selected based on FRS and separately. In all cases, Figures 5 (a) and (b) show slightly higher real-time event detection ratio than the desired sensing coverage in both FRS and. We conclude that this is due to the integer function applied to ý (see Eq. (4)). When is small or the sensor s sensing range is large, the detection ratio is more sensitively affected by the integer function. This explains the case of þ ÿ ÿ þ ÿ ÿ in both the figures. FRS shows almost the same event detection ratio as in two other network fields except for þ ÿ ÿ þ ÿ ÿ. In the case of þ ÿ ÿ þ ÿ ÿ, a relatively small number of sensors are selected to meet the desired sensing coverage, thus the ratio difference becomes large when does not have the exact ý sensors. This also explains that the event detection ratio difference between FRS and decreases as the DSC increases (i.e., ý becomes large). Figure 5 (c) shows the number of selected sensors to meet the desired sensing coverage. Since the trend that the number of selected sensors increases is almost identical in both cases of sensing range ÿ and ÿ, we present the result only for the case of ÿ. The number of sensors that has to be selected increases drastically for the DSC ( ) ÿ. In the case of þ ÿ ÿ þ ÿ ÿ, the event detection ratio is higher than using only 14 selected sensors (i.e., ÿ ) which is only of the total number of sensors in this experiment. This implies that the energy conservation can be further increased without overly decreasing the quality of service. 6.3 Data Reporting Latency Since sensors take turns in reporting their sensed data, there is a waiting period for a sensor to become a data reporter. In addition, as mentioned earlier, the scheme randomly selects ý data reporting sensors based on! "! $ % without having any control operation on the already-selected sensors so that the waiting period is not bounded. On the other hand, and F- schemes have control operations to bound the waiting period in each sensor. We measure the average and maximum data reporting latency until the sensors act as one of the selected ý reporting sensors and transmit their sensed data to the data gathering point. The number of reporting sensors (ý ) and the total number of sensors are main factors in deciding the latency since they determine ' and. Table 2 shows ý, ', and the ratio () )ofý to the total number of sensors, which are calculated based on the desired sensing coverage and the total number of sensors. Figures 6 (a), (b), and (c) show the average reporting latency of, 51

9 + / - Number of Data Reportings x 1m 2, Sensing Range: 3m Sensor Identifier Number of Data Reportings x 2m * 2, Sensing Range: 3m Sensor Identifier Number of Data Reportings x 3m 2, Sensing Range: 3m Sensor Identifier (a) (b) (c) Figure 7. Distribution of Number of Transmissions in Sensors (Index of Energy Conservation Capability) and F- in all the simulated network sizes, and Figures 6 (d), (e), and (f) show the corresponding maximum latency. We can observe that the F- scheme shows the lowest average data reporting latency with the help of the control operation which allows all the selected sensors to transmit exactly once within a reporting cycle with a uniform data reporting latency of ' 2 4. On the other hand, has a little larger average reporting latency compared to the F- and but it has the highest maximum latency which implies that the variance of data reporting time between sensors is large. This is because of the random selection without any control operation on the sensors which have been already selected. In particular, as mentioned earlier, when the desired sensing coverage is low the maximum latency of becomes very large. This is because sensors become a reporter after more election trials due to the relatively small success probability,. We can clearly observe in Figure 6, when compared to, that the average and maximum reporting latency difference decrease as the desired sensing coverage ( ) increases to.9. For ÿ 7, the average reporting latency is less than twice that of in all the cases. Particularly, when ÿ the average latency of F-,, and in all the simulated network fields is almost the same as the one in. As shown in Table 2, in the case of ÿ, ) 4%, 27%, and 23% for each consecutive network size. This implies that the proposed data gathering schemes can achieve a significant resource saving with a minimum trade-off. 6.4 Energy Conservation Capability Even though we did not measure the actual amount of energy consumed while the sensors are transmitting their sensed data using the proposed schemes, it is clear that the energy conservation rate is inversely proportional to the specified desired sensing coverage since only (or approximately) ý sensors report their sensed data while others are delaying their reporting duty conserving their energy. Here we denote both F- and schemes by without distinguishing them since both schemes make the sensors have the same number Table 2. Calculated Values for ý, ', and ) Based on and total number of sensors < % = % = = C D = = = C D G = G = = C D I ( 4 J L ) I ( 4 J L ) I ( 4 J L ).5 12% (6,8.3) 8.5% (17, 11.8 ) 7% (32, 14.1).6 16% (8,6.25) 11% (22, 9.1 ) 9% (42,1.7).7 2% (1,5.) 14% (28, 7.1) 12% (55, 8.2).8 28% (14,3.6) 19% (38, 5.3 ) 16% (73,6.16).9 4% (2,2.5) 27% (54, 3.7) 23% (14, 4.3) of sensed data transmissions. Given a desired sensing coverage, the energy conservation ratio in each sensor depends on ý and O P O Q þ as in the case of the data reporting latency. This is because ý and O P O Q þ determine the parameters ' and which affect the frequency of the sensed data reporting duty cycle of each sensor. That is, the energy conservation capability (number of data transmissions) is proportional to ' and inversely proportional to. Figures 7 (a), (b), and (c) show the distribution of the number of data reportings (transmissions) in each sensor with a desired sensing coverage ÿ in each network field along with the case of. We can observe that the number of data reportings in each sensor is noticeably small in all the figures as compared to. We observed in the previous sub-section that the data reporting latency is not significantly affected when the DSC ÿ 7. This implies that the network lifetime can be significantly increased with a small trade-off. Due to the steady state behavior property of random selection and the control operation, the data reporting chances in all the proposed schemes during 16, second simulations are well distributed among the sensors, implying that the variance of the amount of energy consumed for data reporting in each sensor is very small. Especially, in the case of, as we designed, all the sensors have the same number of data transmissions. We observe in Figure 7 that the number of data transmissions in each sensor depends on ' and the order of the number of data transmissions in each scheme is the same 511

10 as the descending order of ' in the case of ÿ in Table 2. Note that, in the case of, ' þ since all the sensors transmit their sensed data at every round. 7 Conclusion In this paper, we proposed a novel energy-conserving data gathering strategy in wireless sensor networks which is based on a trade-off between coverage and data reporting latency. The ultimate goal of the proposed strategy is to further extend the network lifetime by relieving the quality of service (i.e., sensing coverage of a monitored area) depending on the user s desire. The trade-off represents a negotiation between the user and the network in terms of sensing coverage of a monitored area and data reporting latency to extend the network lifetime. To achieve this goal, we first presented how to find a minimum of ý sensors as data reporters which can meet the desired sensing coverage using geometrical probability. Then we introduced three ý -sensor selection schemes with constant computational complexity: non-disjoint randomized selection (), non-fixed and fixed disjoint randomized selections ( and F-), in order to support the proposed energy-conserving data gathering strategy. Experimental results showed that our proposed strategy can meet the desired sensing coverage specified by the user, using only (or approximately) ý sensors. Moreover, it shows that the network lifetime can be significantly increased with a small trade-off and that the higher the network density, the higher is the energy conservation rate without any additional computation and communication costs. As a future work, we plan to integrate the proposed data gathering strategy with the routing and MAC layer features and measure the network lifetime and data reporting latency by running experiments with realistic data gathering scenarios. Acknowledgments This work was supported by NSF ITR grants IIS and IIS References [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless Sensor Networks: a Survey. Computer Networks Journal, vol. 38, no. 2, pp , 22. [2] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low cost outdoor localization for very small devices. IEEE Personal Communications, vol. 7, pp , 2. [3] W. Choi, S. K. Das, and K. Basu. Angle-based Dynamic Path Construction for Route Load Balancing in Wireless Sensor Networks. In Proceedings of IEEE Wireless Communications and Networking Conference (WCNC), pp , 24. [4] W. Choi, P. Shah, and S. K. Das. A Framework for Energy- Saving Data Gathering Using Two-Phase Clustering in Wireless Sensor Networks. In Proceedings of Int l Conference on Mobile and Ubiquitous Systems: Networking and Services (MobiQuitous), pp , 24. [5] S. K. Das, D. J. Cook, A. Bhattacharya, E. O. H. III, and T. Y. Lin. The Role of Prediction Algorithms in the MavHome Smart Home Architecture. IEEE Wireless Communications, vol. 9, pp , 22. [6] D. Ganesan, R. Govindan, S. Shenker, and D. Estrin. Highly- Resilient, Energy-Efficient Multipath Routing in Wireless Sensor Networks. ACM Mobile Computing and Communications Review, vol. 5, no. 4, pp , 21. [7] F. Garwood. The Variance of The Overlap of Geometrical Figures with Reference to a Bombing Problem. Journal of Biometrika, vol. 34, pp. 1 17, [8] H. Gupta, S. R. Das, and Q. Gu. Connected Sensor Cover: Self-Organization of Sensor Networks for Efficient Query Execution. In Proceedings of ACM Mobile Adhoc Network Symposium (MOBIHOC), pp , 23. [9] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In Proceedings of the 33rd International Conference on System Sciences (HICSS), 2. [1] X. Hong, M. Gerla, and H. Wang. Load Balanced, Energy- Aware Communications for Mars Sensor Networks. In IEEE Aerospace, vol. 3, pp , 22. [11] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks. In Proceedings of ACM Mobile Computing and Networking (MOBICOM), pp , 2. [12] B. Krishnamachari, D. Estrin, and S. Wicker. The Impact of Data Aggregation in Wireless Sensor Networks. In Proceedings of IEEE Int l Conference on Distributed Computing Systems Workshops (ICDCSW), pp , 22. [13] A. M. Law and W. D. Kelton. Simulation Modeling and Analysis, 3rd Edition. MacGraw-Hill, 2. [14] S. Lindsey, C. Raghavendra, and K. M. Sivalingam. Data Gathering Algorithm in Sensor Networks Using Energy Metrics. IEEE Transactions on Parallel and Distributed Systems, vol. 13, no. 9, pp , 22. [15] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson. Wireless Sensor Networks for Habitat Monitoring. In Proceedings of ACM Workshop on Wireless Sensor Networks and Applications (WSNA), pp , 22. [16] A. Manjeshwar and D. P. Agrawal. APTEEN: A Hybrid Protocol for Efficient Routing and Comprehensive Information Retrieval in Wireless Sensor Networks. In Proceedings of IEEE Int l Parallel and Distributed Processing Symposium (IPDPS), pp , 22. [17] A. Nasipuri and K. Li. A Directionality based Location Discovery Scheme for Wireless Sensor Networks. In Proceedings of ACM Workshop on Wireless Sensor Networks and Applications (WSNA), pp , 22. [18] D. Tihan and N. D. Georganas. A Coverage-Preserving Node Scheduling Scheme for Large Wireless Sensor Networks. In Proceedings of ACM Workshop on Wireless Sensor Networks and Applications (WSNA), pp , 22. [19] S. Tilak, N. B. Abu-Ghazaleh, and W. Heinzelman. A Taxonomy of Wireless Micro-Sensor Networks Models. ACM Mobile Computing and Communications Review, vol. 6, pp , 22. [2] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, and C. Gill. Integrated Coverage and Connectivity Configuration in Wireless Sensor Networks. In Proceedings of ACM Embedded Networked Sensor System (SENSYS), pp ,