Adaptive Sink Mobility in Event-driven Clustered ...

Proc., 6th Int. Network Conference (INC 2006), pp. 315-322, Plymouth, UK, 11-14 July, 2006

Adaptive Sink Mobility in Event-driven Clustered Single-hop Wireless Sensor Networks Z. Vincze, D. Vass, R. Vida and A. Vidács1 Dept. of Telecommunications and Media Informatics, Budapest University of Technology and Economics, Magyar tudósok krt. 2., Budapest, Hungary, e-mail: {vinczez, vass, vida, vidacs}@tmit.bme.hu

Abstract Energy ecient operation is of paramount importance in wireless sensor networks. In this paper we propose a novel solution to decrease energy consumption by adaptively moving the sink node. We assume an event driven scenario, where only a reduced number of activated sensors report to the sink node. We also assume that these events have to be reported in real-time, a requirement that makes the existing random or controlled sink mobility solutions inappropriate. We dene three dierent strategies to adaptively move the sink node; the goal is to reduce the communication distance, and thus the energy consumption, of the reporting nodes. The eciency of the strategies is evaluated through simulations.

Keywords Mobile Sink, Wireless Sensor Network, Clustering

1 Introduction and Related Work Wireless sensor networks (WSN) became recently one of the hottest topics for the networking research community, both because of the large number of possible application scenarios and the wide range of related open questions still to be answered through robust and ecient communication solutions. One of the most important features of a WSN is its energy-eciency, as sensors usually dispose of small batteries that are impossible or impractical to change or recharge. In such conditions it is of paramount importance to develop dedicated communication solutions that handle sensor data gathering in an energy-sparing manner. Sensor networks consist of many small, low-cost devices with sensing, processing, and communication capabilities. They monitor a specic region, and subsequently gather and relay information to one or more sink node(s). An ecient data gathering solution should aim at minimizing the communication distance. A straightforward solution to do so is relying on multi-hop communication: instead of communicating directly with a possibly distant sink node, data is relayed hop-by-hop by the intermediate sensors. However, while multi-hop routing spares the energy of distant sensors, it increases the load on those located near the sink. Moreover, complicated routing algorithms, in an error-prone wireless environment might not yet be protable and reliable enough for application and service providers; the majority of the deployed, real-world sensor networks we are aware 1 A.

Vidács is with the Research Group for Informatics and Electronics of the Hungarian Academy of Sciences.

of are based on single hop communication. Thus, in this paper we address only the case of direct communication between sensors and the sink. Another way to reduce the communication distance is to introduce mobility of either the sensors or the sink. There are some approaches that rely on sensor mobility, not for energyeciency but mostly for coverage problems (Wang, et al. 2004)(Butler & Rus 2003). However, using mobile sink nodes to get near the reporting sensors is a more common solution. The approaches based on sink mobility can be classied in three categories: random, predictable and controlled mobility. In (Shah, et al. 2003) randomly moving mobile agents, called Data MULEs (Mobile Ubiquitous LAN Extension) are used to collect data in sparsely populated sensor networks. In (Tong, et al. 2003) a similar solution is presented, but for dense networks; a mobile agent is randomly ying above the sensor eld and gathers data from sensors that are triggered based on the estimated fading state of their communication with the agent. A predictable movement of the sink is also proposed in several papers. In (Chakrabarti, et al. 2003) the sink moves along a predened path, and pulls data from the sensors when it arrives close to them. In (Luo & Hubaux 2005) the predened path is the periphery of the covered circular region itself, multi-hop communication being used as opposed to the previous cases. There are also solutions based on controlled mobility. In (Kansal, et al. 2004) a mobile micro-server moves along a specic trail; its mobility is controlled in the sense that it spends extra time (e.g., it stops or slows down) in areas where there is a large amount of data to send, or where the channel capacity requires so. Controlled mobile nodes are used for message ferrying also in (Zhao & Ammar 2003). What we propose in this paper is a solution signicantly dierent from all the above. We dene an adaptive strategy for sink mobility in an event-driven scenario, so as to continuously move the sink towards an optimal position as far as energy-consumption is concerned, taking into account the current events in the network. Most of the previous solutions addressed time-driven scenarios. In such cases all the sensors measure and send data periodically to the sink. Therefore, moving the sink node has only a load-balancing role. The event-driven scenario that we address is a totally dierent case: only those sensors start to communicate with the sink, which have detected a specic event, and communication lasts only until that event persists. There are several possible applications (e.g., intrusion detection, seismic activity monitoring or wildlife habitat monitoring), that would require such a solution. In such an event-driven scenario the adaptive mobility of the sink could result in energy-saving benets that reach far-beyond simple load balancing. If at a given moment only those nodes are active which have sensed a specic event, it might worth moving the sink node towards them. In the single-hop communication case that we address, even a small reduction of the reporting distance might result in signicant energy saving. Thus, we propose adaptive sink mobility solutions to support energy-ecient single-hop communication in event-driven applications. According to the current events, the sink moves to a neighboring location (within a circle of a given radius) or stays in its current position so as to minimize the energy consumption for the next reporting period. We propose three strategies to choose the optimal position of the sink: in minavg we minimize the average energy consumption of the reporting sensors, in minmax we minimize the maximum energy consumption, while in minrel we take into account the current energy level of the sensors as well, in order to spare the energy of the sensors that are more severely

depleted. There are several approaches in the literature to address the problem of optimal sink placement. In (Bogdanov, et al. 2004) the authors assume a time-driven scenario, and solve the problem of placing n sink nodes so as to optimize energy-consumption. A similar problem is addressed in (Pan, et al. 2005). As opposed to these, we propose to continuously optimize the position of the sink in function of the specic events, not only at the deployment phase. Energy can be spared also by reducing the amount of data to send. A widely used procedure to do so is to employ clustering techniques. The network is divided into small clusters, a cluster head (CH) being responsible for aggregating and relaying to the sink the information gathered from the sensors of its clusters (Kawadia & Kumar 2003). This ts well our scenario; if the same event is sensed by several neighboring sensors at the same time, they send the data only to their CH, which aggregates the packets and sends a single alert message to the sink. Such a solution would result in signicant energy saving. Therefore, we will use clustering in our adaptive mobility strategies; the sink node is moved so as to minimize the energy consumption of the reporting cluster heads. Optimizing cluster formation and maintenance was not among our goals; any ecient clustering technique could be adapted to be used with our sink mobility strategies. The rest of this paper is organized as follows. In section 2 we describe the details of our proposed mobility strategies, through which the sink can adapt to the current events in the network. In section 3 we present some simulation results regarding the eciency of these strategies when compared to the cases of a xed or a randomly moving sink node. Finally, section 4 concludes the paper.

2 Adaptive Sink Mobility In this section we dene adaptive mobility strategies for the sink node, so as to reduce the distance nodes have to send reports to. The solution perfectly ts the event-driven scenario, as the mobile sink can adapt and get closer to the nodes that are currently active, signicantly reducing their energy consumption. It is also dierent from the existing controlled mobility solutions, as sensors do not wait to report until the sink gets close to them; thus, it can support delay-sensitive real-time applications. constraints, and is able to move anywhere inside the network. We also assume that nodes are able to adjust their radio power depending on their distance d from the sink. Although sensing also requires energy, this is far less than the energy used for communication; thus, we neglect it.

2.1 Minimizing average energy consumption Let A be the set of active cluster heads (CHs) that have data to send to the sink. Let (x0 , y0 ) denote the coordinates of the sink, and (xi , yi ) the coordinates of the ith CH. Let di denote the distance between the sink and the ith CH. In the most commonly accepted attenuation model, signal power falls as d−α i ,where α is the attenuation exponent, with values ranging from 2 to 5. Thus, the energy needed for the ith CH to transmit data is

Ei = E0 · dαi ,

(1)

α ∈ [2, 5], P

where E0 is constant. The energy consumed by all the active cluster heads is E = i∈A Ei . To minimize the total (or average) energy consumption, the sink needs to be placed where this sum is the smallest, i.e., (x0 , y0 ) = arg min E. (2) (x,y)

The idea here is that the network always spends the minimal energy for the communication between the active CHs and the sink. The total energy is minimal when ¯

¯

∂E ¯¯ ∂E ¯¯ ¯ = 0 and ¯ = 0, ∂x ¯x=x0 ∂y ¯y=y0 where

Ã

(3)

!

h i α−2 X X ∂E ∂ = E0 dαi = E0 α (x − xi ) (x − xi )2 + (y − yi )2 2 . ∂x ∂x i∈A i∈A

(4)

The partial derivative ∂E/∂y can be calculated similarly. Unfortunately, there is no closed form solution for (3); thus, it has to be solved using optimisation methods (for example some kind of gradient-based search (Carson & Maria 1997)). In the following we will refer to this strategy as minavg.

2.2 Minimizing maximum energy consumption The drawback of the minavg approach is thatalthough the total energy consumption is minimizedit can happen that the energy contributions of the sensors are rather uneven. In order to avoid this problem, the strategy introduced here minimizes the transmission energy for the most remote cluster head in the network. In other words, the maximum transmission energy is minimized, i.e., (5)

max Ei → min . i∈A

Hence, energy consumption will be more balanced. As the transmission energy depends on the distance between the sensor and the sink, this strategy is equivalent with minimizing the maximum distance between the sink and every active CH in the network. Thus, the optimal location of the sink in this case is given by µ (x,y)

¶

q

(x0 , y0 ) = arg min max i∈A

(x − xi )2 + (y − yi )2 .

(6)

The optimisation task is equivalent to the Minimal Enclosing Circle Problem, where the task is to nd the minimum radius circle that encloses all points of a point set on the plane. There are several algorithms to solve this problem. E.g., it has been shown that it can be solved in O(n) time using the prune-and-search techniques for linear programming (Megiddo 1983). In the following we will refer to this strategy as minmax.

2.3 Minimizing relative energy consumption Neither of the two previous strategies take into account the current energy levels of the sensor nodes; thus, they are not able to protect from depletion those cluster heads that have already sensed and reported many events and their batteries are getting exhausted. Nodes with little remaining battery power could be spared if the sink would move closer to them, while moving away from nodes with more remaining energy. One such possible strategy is when the maximum relative energy that a node has to spend on transmission is minimized, i.e., Ei → min, (7) max i∈A Erem,i where Erem,i denotes the average remaining energy per node within the ith cluster. Note that we do not take into account the remaining energy of the current cluster head, but

the total energy within the cluster. By doing so we ensure a more balanced energy consumption in the network. The sink will not get close to a cluster full of energy, just because its CH is near of depletion; the CH should be changed instead. On the other hand, a cluster that is running out of energy will be spared even if its current CH is still in good shape. The CH will continuously monitor the energy level of the sensors in the cluster. An aggregated report on the cluster's energy level can then be passed to the sink, e.g., by piggybacking it on the normal reports sent by the active cluster heads; the sink can then use the information and adapt its movement accordingly. When the CH role is passed to a new sensor, the information about the remaining energy in the cluster is transmitted as well. The minimization problem is equivalent with the following: 

q



(x0 , y0 ) = arg min max (x,y)

i∈A

α

(x − xi )2 + (y − yi )2  . Erem,i

(8)

For example, if there are two active CHs, with average remaining energies Erem,1 = 60 and Erem,2 = 4 units within the clusters respectively, the sink is at the optimal location if the rst CH spends E1 = 15 units of energy for transmission while the second one uses only E2 = 1 unit. This is because in this case both of them would consume 25% of their average remaining energy. There is no closed form solution for nding the solution of (8); thus, it has to be determined using optimisation methods. We will refer to this strategy as minrel.

3 Simulations To evaluate the performance of the three mobile sink strategies, we compared them with the case when the sink is xed and is deployed in the center of the network, and with the case when the sink moves randomly using the Random Waypoint Mobility model (originally proposed in (Johnson & Maltz 1996)). We simulated the proposed sink relocation strategies using MATLAB. The network was modeled to be circle-shaped with radius (R) of 400 m, in which we randomly distributed 8000 sensors, using a uniform distribution model. The initial energy of every sensor was 2 J, E0 was 1 nJ. The attenuation exponent α was chosen to be 3. Clusters were formed using the LEACH (Heinzelman, et al. 2000) algorithm, where in the initializing phase each sensor node declared itself as cluster head (CH) with a probability of 0.05. The operation of the network is event driven. In our event radius model, an event is located in a single point in the network area, and all nodes within a distance of 20 m (called the sensing range, r) of this event are considered to be active. An event can occur in a uniformly distributed manner in the area. The time was split into equal periods (i.e., 10 sec) and we assumed that an event can be reported only at the beginning of the time period. The number of new events within a period was modeled as a Poissondistributed random variable, with intensity parameter (λ) of 0.3. The duration of the event was geometrically distributed; thus, an existing event persisted in the next round with probability 0.9, having an average lifetime of 100 seconds. Every active sensor sends the same amount of data in a round, including the sensed attribute values as well as its ID and remaining battery power, and communicates with its cluster head directly. The sink can be relocated in each round; after determining the best position for the next step, according to the dierent strategies using eqs. (2), (6), (8), it moves there. Note, that the

4

2

x 10

98 1.5

time(round)

area coverage(%)

100

95

92

89

86 0

minmax minavg minrel fix rwp

1

0.5

10 000

20 000

time(round)

30 000

0

minmax

minavg

minrel

fix

rwp

Figure 1: (a): area coverage (left) and (b): average lifetime (right). sink mobility is limited, i.e., the mobile sink can only move with nite maximum velocity (4 m/s). To inform the cluster heads about its current position, our solution assumes a periodic update message broadcasted by the mobile sink. An important power-saving factor is that only those CHs that have data to report listen to the update messages. The area covered by operating sensors can be an important measure of eciency from the application's point of view. Fig. 1(a) shows the area coverage as a function of time, in the case of all introduced strategies. Here we dened coverage as follows: if the energy level of all sensors within a cluster falls below a certain critical threshold, it can happen that the elected cluster head will not be able to send its message to a far away sink. Thus, coverage of that area is lost, i.e., it is not assured that the sink node can be notied when an event appears in that particular area. Fig. 1(a) reveals that the randomly moving sink case is far less ecient than the other strategies. The xed sink case is also considerably less eective than the adaptive strategies. For high area coverage values the dierence between the three mobile strategies is not signicant. However, the advantage of minimizing the average energy becomes more apparent at lower coverage values. Fig. 1(b) shows the time elapsed until the area coverage falls below 100%. This can be seen as the lifetime of the network, if the application's requirement is so strict that even a single event loss cannot be tolerated. The network lifetime is the highest in case of minrel, while minavg and minmax have approximately the same performance, both well exceeding the lifetime obtained with a xed or a randomly moving sink. It is also interesting to see how the ve strategies aect the sensors located at dierent parts of the covered area. Fig. 2(a) shows the average energy consumption of a sensor in function of its distance to the center of the area. An obvious result is that the xed strategy depletes the sensors in the least balanced manner. One can see that moving the sink randomly results in more balanced, but rather high energy consumption in the network. The most homogenous energy usage can be achieved by using the minrel strategy. Even if it causes higher energy usage compared to the other two mobile sink strategies, it still results in longer network lifetime, as it prevents sensors with low remaining energy from depletion. Fig. 2(b) shows the amount of the total energy consumed by the network. Recall that the minavg strategy minimizes the total energy in every round; therefore, the total energy used by the network during its entire operation is also minimal. Minrel uses the most energy among the three adaptive strategies, but still less than the xed or the randomly moving sink case. The latter consumes almost twice as much energy as the xed

2

minmax minavg minrel fix rwp

total used energy(J)

used energy(J)

12 000 1.5

1 fix minrel minmax minavg rwp

0.5

0

0

50

100

150

200

250

300

350

8 000

4 000

0

400

0

10 000

20 000

30 000

time(round)

distance from the center(m)

Figure 2: (a): The energy consumption in function of the distance from the

percentage(%)

percentage(%)

center (left), and (b): the total energy consumption (right).

0.8 0.6 0.4 0.2 0 800

800

400

y

0.8 0.6 0.4 0.2 0 800

0 0

800

400

400

y

x

0 0

x

percentage(%)

(b) Minavg

percentage(%)

(a) Minmax

400

0.8 0.6 0.4 0.2 0 800

800

400

y

400 0 0

(c) Minrel

x

0.8 0.6 0.4 0.2 0 800

800

400

y

400 0 0

x

(d) Rwp

Figure 3: Distribution of the sink location within the area sink case. In Figure 3 we present the distribution of the sink coordinates after 500.000 simulation rounds. We can see that in the case of the minmax strategy the sink often resides near to the center, while in the minrel and the minavg case the distribution is more homogeneous. In the case of the rwp strategy the sink can be practically anywhere within the area.

4 Conclusions In this paper we presented the idea of adaptively moving the sink of a clustered sensor network, in order to decrease the amount of energy required for communication, and hence prolong the lifetime of the network. We introduced three dierent strategies for moving the sink: minavg, minmax, and minrel. The rst one minimizes the average

energy required for the communication, the second one minimizes the maximum energy consumption among active cluster heads, while the third one minimizes the relative energy consumption by taking into account the remaining battery power of the nodes in the active clusters. We presented simulation results and a performance evaluation. The results have shown that all the three proposed adaptive sink relocation strategies perform signicantly better than the xed or randomly moving sink case. Among them, the minrel strategy turned out to be the most ecient one, as long as the network lifetime is concerned.

References Aazhang, B., Chakrabarti, A. & Sabharwal, A. (2003). `Using Predictable Observer Mobility for Power Ecient Design of Sensor Networks'. In Proc., 2nd Int. Workshop on Information Processing in Sensor Networks (IPSN), pp. 129145, Palo Alto, CA, USA. Also in Lecture Notes in Computer Science, Vol. (2634), pp. 129-145. Adireddi, S., Tong, L. and Zhao, Q. (2003). `Sensor networks with mobile agents'. In Proc., IEEE Military Communications Conference 2003, vol. 22, pp. 688693, Boston, MA, USA. Ammar, M. and Zhao, W. (2003). `Message ferrying: Proactive routing in highly-partioned wireless ad hoc networks'. In Proc. of the 9th IEEE Workshop on Future Trends in Distributed Computed Systems (FTDCS 2003), pp. 308314, San Juan, Puerto Rico. Balakrishnan, H., Chandrakasan, A. & Heinzelman, W.R. et al. (2000). `Energy-Ecient Communication Protocol for Wireless Microsensor Networks'. In Proc.,of 33rd Hawaii International Conference on System Sciences. Bogdanov, A., Maneva, E. and Riesenfeld, S. (2004). `Power-aware base station positioning for sensor networks'. In Proc., IEEE INFOCOM 2004, Hong Kong. Brunette, W., Jain, S., Roy, S. and Shah, R.C. (2003). `Data MULEs: Modeling a Three-tier Architecture for Sparse Sensor Networks'. In Proc., IEEE Workshop on Sensor Network Protocols and Applications (SNPA), pp. 3041, Anchorage, alaska, USA. Butler, Z. & Rus, D. (2003). `Event-based Motion Control for Mobile-Sensor Networks'. IEEE Pervasive Computing 2(4):3442. Cai, L., Hou, Y.T., Pan, J., Shen, S.X. and Shi, Y. (2005). `Optimal Base-Station Locations in Two-Tiered Wireless Sensor Networks'. In IEEE Transactions on Mobile Computing 4(5):458473 . Cao, G., Porta, T.L. and Wang, G. (2004). `Movement-assisted sensor deployment'. In Proc., IEEE INFOCOM 2004, Hong Kong. Carson, Y. & Maria, A. (1997). `Simulation optimization: Methods and applications'. In Proc. 9th conference on Winter Simulation, pp. 118126, Atlanta, USA. Estrin, D., Kaiser, W.J., Kansal, A., Pottie, G.J., Rahimi, M. and Srivastava, M.B. (2004). `Controlled Mobility for Sustainable Wireless Networks'. In Proc., IEEE Sensor and Ad Hoc Communications and Networks (SECON), Santa Clara, CA. Hubaux, J.P. and Luo, J. & (2005). `Joint Mobility and Routing for Lifetime Elongation in Wireless Sensor Networks'. In Proc., IEEE INFOCOM 2005, Miami, FL, USA. Johnson, D.B. and Maltz, D.A. (1996). `Dynamic source routing in ad hoc wireless networks'. Mobile Computing 353:153181. Chapter 5. Kawadia, V., and Kumar, P. (2003). `Power control and clustering in ad hoc networks'. In Proc. IEEE INFOCOM 2003, pp. 459469, San Francisco, CA, USA. Megiddo, N. (1983). `The weighted Euclidean 1-center problem'. Mathematics of Operations Research 8(4):498504.