1 Introduction

Interconnecting devices to generate a “smart” environment is the basic concept in the Internet of Things (IoT), and the sensor-actuator-internet is the framework to achieve this smart environment [5]. Wireless sensor networks (WSNs) have been widely used in many IoT applications due to ubiquitous sensor devices [3, 5, 19]. Extensive research activities in the areas of security [18, 20], topology [24], synergies with other technologies [17], and energy consumption [5] in WSNs for IoT have been conducted recently.

WSN data collection and event reporting are important research topics in the IoT. The IoT is a worldwide network where all devices are interconnected and information has to be available fast and efficiently, thus redundancies and useless information have to be eliminated. A key approach for efficient interconnection is to give devices a “smart” behavior where they can communicate, process information, and take decisions without human intervention [13].

Designing energy-efficient communication protocols in WSNs is vital. Most of the sensors are battery powered, and sometimes it is impractical or infeasible to replace or recharge the nodes. In addition, WSNs may be deployed in areas where access is difficult and monitoring has to be done over a long period of time. Due to the power constraint characteristic of WSNs, the applications using WSNs employ different mechanisms to minimize the energy consumption in order to prolong the network lifetime [1].

Depending on the application, events can change their shape, can appear/disappear (e.g., fire events), and can move (e.g., storms such as hurricanes, tornadoes). These events are called spatio-temporal events or dynamic events [23]. Spatio-temporal events have the characteristic that their shape, size, and speed of movement can change over time. Different mechanisms have been developed by researchers in order to save energy in the network. One mechanism is using clustering for data collection. Cluster heads are in charge of collecting data from its cluster members, aggregate it, and send it to the sink [2, 4, 6]. Also, if events are sparse and expected to occur sporadically, then a reactive protocol could be used. In this case, when there is no event, no energy is wasted to maintain an updated path to the sink [7, 8, 16].

In this paper, we improve the solution proposed in [1], where the authors address the problem of detecting and reporting a composite mobile event to a mobile sink. We design a new event-based clustering mechanism and we revised the dynamic clustering technique. This improvement results in reduced energy consumption and a higher number of packets processed successfully by the sink compared to [1]. The rest of the paper is organized as follows. Section 2 describes the design methods of our routing protocol and summarizes the experiments. Section 3 introduces the state of the art in monitoring spatio-temporal events. Section 4 presents the event model. Section 5 describes the problem definition. In Section 6, we present our anchor-based routing protocol for spatio-temporal event detection and reporting. The performance of our algorithm is illustrated in Sections 7 and 8, where we conduct simulations using WSNet [22]. The conclusions are stated in Section 9.

2 Methods/experimental

In this paper, we propose an energy-efficient and distributed routing mechanism for event detection and reporting to a mobile sink. We consider that events can be mobile, thus the set of sensor nodes which detect a certain event change over time. To deal with the sink and event mobility, we use several important techniques in our algorithm design. First, we use a dynamic anchor mechanism. A set of anchors may be active at a certain time, and this set may change over time depending on the sink movement. Another important technique is the use of a dynamic, event-based clustering, which allows the collection of information about the moving events. The cluster (or clusters) monitoring a certain event may change over time, according to the movement of the event. This dynamic clustering mechanism reduces the number of event-reporting messages sent to the sink, thus the overall energy consumed by the network is decreased.

We conduct simulations using the WSNet [22] network simulator and measure various performance metrics such as the average residual energy, the number of clusters which are active, and the percentage of composite events processed successfully by the sink. Simulation results show that the newly proposed mechanism has the best performance in terms of energy efficiency and percentage of events detected successfully by the sink.

3 Related works

According to [23], spatio-temporal events can be diffuse. This means that the events can spread in space similar to a fire or they can move similar to a tornado. The authors use extended linear-chain conditional random fields, which are undirected graphs that use a specific group of features to encode a conditional probability distribution. The authors incorporate temporal constraints to a spatial field in order to determine the spatio-temporal dependencies among observations and events in the network. In this way, the sensors can determine if the event is diffuse or whether it is moving. In our work, the events can be fixed or moving, and we address the problem of reporting the event detection to the sink.

An example of using WSNs for IoT is illustrated in [12]. An indoor distributed sensor system with an intelligent processing unit is used to track a fire. When sensor reading exceeds a given threshold, the sensor node sends its reading and location to the base station (or processing unit). The base station calculates a function f (x, y, z) and its gradient using the coordinates x, y, and z of the node. The results are plotted as vectors. The readings closer to the center of the event will have bigger vectors. The base station sends a “start sensing process message” to the 1-hop neighbors of the border of the event nodes, to continue tracking the movement of the fire. The base station can also calculate the direction of the event movement using the relative weighting of the i, j, and k components of the gradient. Our work is different in several ways. We are not restricted to indoor detection, both the sink and the event are mobile in our case, and we use energy-efficient techniques such as clustering.

Article [11] presents an IoT mechanism for detecting a forest fire using artificial neural network. A multi-layer back-propagation artificial neural network algorithm is used to detect multiple criteria that depend on specific attributes of a forest fire (e.g., temperature, radiation, and light). The neural network is feeded with environmental information collected by heterogeneous sensor nodes, and the correct alarm decisions. With this information, the system is capable of learning and making decisions when new information is feeded in the system. In our work, we assume that the detection is performed by the sensor nodes, and we use energy-efficient techniques to reduce the energy consumed by the WSNs.

WSNs can be used in IoT applications for monitoring environmental changes, people, and animal movements. Article [15] presents an IoT application developed to monitor the Australian Great Barrier Reef using big data analytics and a hierarchical heterogeneous sensor network. Sensor nodes collect environmental data from the ocean such as salinity, temperature and level of chlorophyll. The sensor nodes collect information, then aggregate, and send it to a buoy. The buoys are connected to poles which are in charge of sending the aggregated data to the base station. The elliptical summaries anomaly detection (ESAD) algorithm uses artificial intelligence to detect anomalous patterns based on the data collected. The algorithm measures data dissimilarities, forms clusters of similar information, and extracts the anomalous clusters. This article has a very important practical applicability, and it is different from our work since it assumes a certain hierarchical infrastructure. Works [9, 10] also deal with monitoring diffuse events. In this case, the sensor network is used to detect the boundaries of the event. A sensor detecting an event (also called “affected” node) checks whether its neighbors detect the event too. If all neighbors are affected as well, then the node proclaims itself a non-boundary node. But if at least one neighbor did not detect the event, then the node broadcasts the result of a function computed based on the event detected and the number of neighbors affected. The node then waits to receive the evaluation of the same function from its neighbors. The node with the smallest function value becomes a boundary node. Then, the boundary node sends the information to the sink using a multi-hop routing protocol. The process is repeated periodically to track the changes in the size of the event. In our work, the algorithm continuously monitors and reports the event, not only the changes in the boundary of the event.

Article [14] proposes an algorithm that uses probabilistic graphical models to detect spatio-temporal events. The authors use a graph representation of the nodes, where every node is a random variable and the edges are the probabilistic interaction between them. The process of detection and reporting has two phases. In the first phase, a temporal dependency modeling of the events is done using a first-order Markov chain model. The nodes predict the next state of their sensed values (cold, mild, hot) based on their own information. The second phase is the spatial dependency model of the event that uses Markov random fields. In this case, the nodes predict their sensing values using the information of its neighbors. Every time a node senses an event, it informs its neighbors so that they can start the process of detecting the event. The detection mechanism used in our case is different, see Section 4, and we are also concerned with employing energy-efficient techniques.

To track a diffuse event in IoT, article [21] uses a Bayesian trust model and a network consisting of a grid of nodes. Each node has a virtual cluster, the node being the center of the cluster. Its members are the neighbors within communication range. The node calculates a trust index with a probability density function. If the value of this trust index is less than some predefined threshold, then the node readings do not match one or more readings of its neighbors. To adjust the trust index, the node uses a Bayesian filtering procedure to incorporate neighbors’ readings to its own readings. The result of this procedure is e = 1 or e = 0, where e = 1 means that the node is detecting the event and e = 0 means that the event is not detected. Every time period T, the node sends to the sink its Id and the value of e. Based on the information received from the nodes, the sink reconstructs the map of the network and the progression and size of the event. In our case, we assume that the nodes are randomly deployed rather than in a grid. We developed an event-based clustering, where only the nodes detecting the event are involved in reporting the event.

In this paper, we propose a reactive anchor-based routing protocol with constrained flooding and dynamic clustering for spatio-temporal event detection and reporting. Our work improves on the solution presented in [1] and which will be detailed in Section 6.

4 Description of the event model

We define an event as an observable occurrence of a phenomenon during some specific time in a specific area. We use the concepts of atomic events and composite events, given in [4]. An atomic event is triggered when a single sensing value (or attribute) exceeds some given threshold and it is denoted by e(t, s, R) where t is the time when the event occurs, s is the location of the event, and R is a logical expression defining the conditions when the event occurs. To detect complex events, we use composite events, where variations in different attributes are detected. A composite event is denoted by the following:

$$ E\left(\left({e}_1,{\delta}_1\right),\left({e}_2,{\delta}_2\right),\dots, \left({e}_k,{\delta}_k\right),{C}_t,{C}_s,\delta \right)=\left({R}_1\wedge {R}_2\wedge \dots \wedge {R}_k\wedge {C}_t\wedge {C}_s,\delta \right) $$

where ei, i = 1 to k, is an atomic event. δi with 0 ≤ δi ≤ 1 is the confidence of ei, indicating the probability of E occurring when ei occurs. Ct is the constraint on atomic events’ times t1, t2, …, tk. Cs is the constraint on atomic events’ locations s1, s2, …, sk. The confidence δ of the composite event is defined as δ = δ1 + δ2 + … + δk, and it is expected to satisfy the property δ1 + δ2 + … + δk = 1 [4].

We consider spatio-temporal events similar to [1, 23]. Our event moves with a certain speed, but it maintains the same circular shape and moves on a random path, see Fig. 1.

Fig. 1
figure 1

Spatio-temporal event

Full size image

5 Problem definition

We consider a heterogeneous WSN consisting of n nodes N1, N2, …, Nn and one mobile sink S. The nodes are randomly deployed in an area A. The nodes have the same communication range Rc and the same initial energy Einit. We consider a mobile sink with a communication range Rc and infinite energy. Each node has one or multiple sensing components from the set {s1, s2, …, sm}.

Each sensing component can be used to detect an atomic event for that attribute [1]. Table 1 shows the main notations in the paper. We consider the following assumptions for our work:

  • The nodes have synchronized clocks, so we do not deal with node synchronization in this work.

  • The network is heterogeneous, that means sensor nodes may be equipped with different sensing components. This occurs since nodes may be manufactured with different sensing capabilities, some nodes may have purposely turned off some sensing components to save energy, and some sensing components may fail over time.

  • Events are fixed or mobile.

  • Nodes are resource constrained in terms of power, bandwidth, memory, and computing capabilities.

Table 1 Notations
Full size table

5.1 Problem definition—spatio-temporal composite event detection and reporting (STCEDR) in mobile-sink WSNs

Given a WSN deployed in an area A, consisting of n nodes with different sensing components from the set {s1, s2, …, sm} and a mobile sink S, designing an energy-efficient and distributed algorithm for detecting and reporting a spatio-temporal composite event E inquired by the sink S, the composite event E is defined using atomic events corresponding to the sensing components {s1, s2, …, sm}.

6 ARER—anchor-based routing protocol for event reporting in mobile-sink WSNs

In this section, we propose ARER, anchor-based routing for event reporting, an improvement of the anchor-based routing protocol with constrained flooding and dynamic clustering presented in [1]. The improvement has two components: a new event-based clustering mechanism and a new rule for dynamic clustering, which will be described later in this section. These mechanisms result in a reduced energy consumption and an increased percentage of composite events processed successfully by the sink, which will be illustrated in Section 8.

Figure 2 shows the main phases of the protocol. In phase 1 the sink S selects the first anchor A1 using the following mechanism. S broadcasts FindClosestNode and the nodes in range reply with their ID and residual energy after a small delay. The sink chooses the closest node based on the signal strength, and in case of a tie the residual energy and the smallest ID are used. The sink S then sends the request for monitoring the composite event E to A1, who is in charge of flooding the CompositeEventRequest(S, A1, E, δth, hops) through the network, where E is the composite event and δth is the threshold parameter for the composite event.

Fig. 2
figure 2

Main phases of the ARER protocol

Full size image

As the message floods the network, a convergecast tree T is formed, where A1 is the root. Each node Nj that receives the message will set-up its parent Nj .tp to the sending node, will increment the hops field of the message, and will resend CompositeEventRequest(Nj, E, δth, hops). At the end of this phase, each node has a pointer to its parent in the convergecast tree T rooted at A1.

In phase 2, the nodes that satisfy the location requirement Cs and are equipped with sensing components needed to detect one or more atomic events e1, …, ek, start monitoring the event for a time duration Ct.

In phase 3, one or more nodes start detecting the event. The nodes form one or more clusters. Each cluster selects a cluster head (CH) and initiates the mechanism for event reporting. Our event-based clustering is described in Section 6.1.

6.1 Event-based clustering

A node can become CH only if it detects at least one atomic event and if its residual energy is larger than a predefined threshold. Based on its residual energy and ID (used to break ties between nodes with the same residual energy), a node proclaims itself CH and sends a message NewCluster over hcluster hops. This message contains the ID of the CH and event ID. All the nodes within the hcluster hops distance join the cluster and resend the message as long as the number of hops is less than hcluster. As the NewCluster message is forwarded, a cluster tree Tcluster rooted at the CH is formed. The nodes will not send any ack message back to the CH.

Depending on the size of the event and the number of hops in the NewCluster message, more than one cluster may form. The CH receives atomic events from its cluster members, and as messages are sent from cluster members to the CH, aggregation is performed.

6.2 Event reporting

The event report messages flow from the nodes to the CH, from the CH to A1 along T, and from A1 to S. Since S is mobile, A1 sends beacons (or data) periodically to the sink to maintain the communication with S. If S does not hear a beacon (or data) from A1 for α periods, where α is an input parameter, then S selects a new anchor A2 as follows. S broadcasts a message NewAnchorRequest(S, A1). If a node Nj receives both A1’s beacons (or data) and S’s message, then it waits a time based on the signal strength of the message NewAnchorRequest and sends a message NewAnchorReply(S, A1, Nj) to the sink. The node Nj from which the sink receives the first reply becomes the second anchor.

If S moves out of the range of A2, then the process repeats and a new anchor A3 is selected, see Fig. 3a. In this case, the data travels from the node to A1, A2, A3, and then the sink S, respectively. After the maximum number of anchors β is reached, the anchor selection process resets, that means a new anchor A1 is selected.

Fig. 3
figure 3

Example showing a the new anchor selection and b the shortcut mechanism

Full size image

To save energy and to reduce the event-reporting delay, we implement a shortcut mechanism. If a node Nj in T receives beacons from the last anchor, then Nj stores this anchor as its parent Nj .tp. Then Nj sends data directly to the last anchor instead of sending it through the rest of the path. This mechanism is illustrated in Fig. 3b.

To address the case when events cease to exist, a time-out procedure is implemented. If no data (event reports) are sent for a duration γ, then the fields Nj .tp are marked obsolete. Similarly, the anchor status of a node becomes obsolete if no data is forwarded toward the sink for a duration γ. Thus, our data collection mechanism is reactive, where the data collection infrastructure forms and is maintained only when there are events to report. If the parent field is obsolete, then a CH has to find a path to reach the sink S. The CH broadcasts RouteRequest. When S receives the message, it selects an anchor A1 using the mechanism described previously. A1 then floods the whole network with a RouteReply message which contains the parameters of the composite event. As the RouteReply message is sent, a convergecast tree rooted at A1 is formed. A1 floods the reply message since often times more clusters are formed, and in this way we avoid other RouteRequest messages being initiated by other CHs.

Note that some nodes might have the tp attribute active, while others may have this attribute obsolete. To reduce the energy consumption, a “constrained flooding” algorithm is proposed, which is basically an incremental ring-search mechanism. Rather than flooding the whole network, first, the CH broadcasts the RouteRequest message to h hops. If a node Nj with an active tp attribute receives the message, then it sends back a RouteReply message to the CH which contains the number of hops to A1 or the number of hops to the last anchor if the shortcut mechanism was applied. If the CH receives more than one RouteRequest, then it sets its tp attribute to the node that has the smallest number of hops to the sink. If no RouteRequest is received, then the CH increases the number of hops, and after a number of unsuccessful iterations it floods the network to find a path to the sink.

6.3 Dynamic clustering

When spatio-temporal events move, additional clusters are formed to report the event to the sink S. As the event moves, some of the clusters (or parts of clusters) may not detect the event, while new nodes start detecting the event. These new nodes will form new clusters.

When a new cluster is formed, the CH sends a NewCluster message, as described previously in phase 3. The nodes which do not belong to any cluster and detect atomic events join the new cluster.

To reduce the number of clusters active in the network, a dynamic clustering mechanism is proposed. A node Nj which belongs to a cluster and receives a NewCluster message for an event with the same ID, but from another CH, joins the newer cluster. Actually, if more NewCluster messages are received by a node Nj for the same event ID, then Nj joins the newest cluster using a time-stamp field in the NewCluster message.

A node Nj belongs to a cluster and acts as a relay node for prior clusters in which it belonged. Nj has a table containing the parent node for both current and prior clusters. Let us assume that Nj was a member of the cluster C1, and it is now a member of C2. Nj sends its data reports to the CH of C2. At the same time, Nj acts as a relay node for the nodes transmitting data to the CH of C1. The data reporting messages are sent to the CH along the paths in the cluster tree, as explained previously.

The dynamic clustering technique is expected to reduce the number of event-reporting messages circulating from CHs to the sink, since older clusters will have no event reporting sooner. Then, less energy is consumed.

Message complexity is discussed next. The number of messages exchanged for phase 1 (see Fig. 2) is O(n), where n is the number of sensor nodes. In phase 2, no messages are exchanged. In phase 3, each cluster formation has message complexity O(n), and as the event moves, more clusters may form. Thus the overall message complexity for clustering is O(n2). CHs report data to the sink following the parent path in the convergecast tree T, with message complexity O(n). The overall message complexity of the protocol is O(n2).

7 Simulation specifications

We compare the performance of three algorithms:

  • ASC(CF), which is the protocol proposed by Aranzazu-Suescun and Cardei in [1], the option that uses only the constraint flooding (CF)

  • ASC(CF&DC), which is the protocol proposed in [1], the option that uses both the constraint flooding (CF) and the dynamic clustering (DC) mechanisms

  • ARER (CF&DC), which is the protocol presented in Section 6

We conduct simulations using WSNet [22], an open-source event-driven simulator for WSNs. WSNet uses the object-oriented C + + language, Linux operating system, and provides support for energy model and event modeling. Tables 2, 3, 4, and 5 present the main parameters used in the simulations.

Table 2 Simulation parameters
Full size table
Table 3 Types of sensors used
Full size table
Table 4 Sink speed
Full size table
Table 5 Event speed
Full size table

Our WSN is deployed into a square area with side length A.L = 1100 m, with a sink S located in the middle of the right side at the beginning of the simulation.

The WSN has n = 3125 nodes randomly deployed in the area A. Each node is equipped with a random number between 1 and 5 of sensing components. We define a composite event with five atomic events shown in Table 3. The composite event that the sensor network monitors is defined as: E((e1, 0.35), (e2, 0.1), (e3, 0.15), (e4, 0.3), (e5, 0.1), ts, A, δ), where ts is the simulation time after the request is sent by the sink S and A is the deployment area.

The initial energy of each node is Einit = 1 Joule. To measure the energy consumption in the simulations, we use the energy model from LEACH [6]:

\( {\displaystyle \begin{array}{c}{E}_{Tx}\left(l,d\right)={E}_{\mathrm{elec}}\times l+{E}_{\mathrm{amp}}\times l\times {d}^2\\ {}{E}_{Rx}(l)={E}_{\mathrm{elec}}\times l\end{array}} \)

where Eelec = 50 nJ/bit, Eamp = 100 pJ/bit/m2, and d is the distance between nodes.

A node can become a CH if its residual energy exceeds the threshold of 900 mJ. In the simulations, we assume that events have a circular area. Events move using the random walk model with speeds indicated in the Table 5.The center of the event is generated randomly. We consider three types of events

  • small events, where event radius is 45 m

  • medium events, where event radius is 200 m

  • large events, where event radius is 400 m

The simulation time is 1 h. Nodes detecting an event send data messages to the sink S periodically, every 5 s. In the ASC algorithms, cluster membership becomes obsolete after 10 periods. A composite event occurs if the confidence exceeds the threshold value 0.75. We run each simulation scenario 10 times using different seed values and report the average values in the graphs.

8 Results and discussion

Figure 4 shows the average residual energy of the network for n = 3125 nodes, A.L = 1100 m, medium size events, and average sink speed 5 m/s. We vary the average event speed to 1 m/s, 9 m/s, and 18 m/s, respectively. There is only one continuous event in all the experiments and the duration of the event is 25% and 75% of the total time of the simulation. For example, ASC(CF)(25%) means that the duration of the event is 25% of 1 h, which is 15 min.

Fig. 4
figure 4

Average residual energy of the network with different speeds of the event. a Average event speed 1 m/s. b Average event speed 9 m/s. c Average event speed 18 m/s

Full size image

In all the experiments, ARER(CF&DC) consumes less energy than the ASC protocols. Both algorithms with dynamic clustering (DC) consume less energy than ASC(CF). This points out the benefits of dynamic clustering. Since nodes join new clusters as events move, fewer clusters are involved in data reporting. ARER(CF&DC) is more energy-efficient than ASC(CF&DC) due to the new clustering mechanism which omits acknowledgment messages. A higher speed of the event results in more clusters, thus the residual energy decreases.

Figure 5 shows the residual energy of the network for n = 3125 nodes,

Fig. 5
figure 5

Average residual energy of the network with different speeds of the sink. a Average sink speed 1 m/s. b Average sink speed 5 m/s. c Average sink speed 12.5 m/s

Full size image

A.L = 1100 m, medium-size events, and average event speed of 9 m/s. The average sink speed is 1 m/s, 5 m/s, and 12.5 m/s, respectively. When the sink moves faster and events are present, new anchors have to be selected to maintain the communication between the clusters and the sink. When the maximum number of anchors is reached, a new convergecast tree is built. Therefore more energy is spent on this process, more frequently.

In Fig. 6, the average sink speed is 5 m/s, the average event speed is 9 m/s, and the number of nodes is n = 3125. Results are measured for small, medium, and large events, with one event in the network. More energy is spent by the network on data reporting for larger events since more clusters are formed and report data to the sink.

Fig. 6
figure 6

Average residual energy of the network with different size of the event. a Small event. b Medium event. c Large event

Full size image

In Fig. 7, 15 min of simulation time (or 25% of the simulation time) are used to compute the average number of active clusters formed by each algorithm when we vary the event speed, the event size, and the sink speed. The network has n = 3125 nodes and A.L = 1100 m. The number of clusters vary with the event size. Larger events result in more clusters. A higher event speed results in more clusters. The speed of the event affects the number of clusters because the area of coverage by the event varies, as mentioned previously. The speed of the sink does not affect the number of clusters formed in the network because these values are not dependent. In all measurements, ASC(CF&DC) and ARER(CF&DC) have a smaller number of clusters than ASC(CF) due to the dynamic clustering technique.

Fig. 7
figure 7

Average number of clusters in the network. a Different event speeds. b Different event sizes. c Different sink speeds

Full size image

For medium events with event speed of 9 m/s, an average of 35.42% of the nodes joining a cluster has changed their cluster at least once. The average number of cluster changes is 1.83, and the maximum number of times a node changes its cluster is 2 (e.g., 3 different cluster memberships during the experiment, as the event moves).

Figure 8 shows the percentage of composite events processed successfully by the sink, when n = 3125 nodes and A.L = 1100 m. The two experiments vary the average speed of the event and the average speed of the sink, respectively.

Fig. 8
figure 8

Percentage of composite events successfully processed at the sink. a Different speeds of the event. b Different speeds of the sink

Full size image

We can observe that ARER(CF&DC) improves the percentage of composite events processed successfully by the sink compared to the ASC(CF&DC) protocol. ASC(CF) and ARER(CF&DC) yield similar results, which are better than ASC(CF&DC). ASC(CF&DC) has the smallest percentage because, when nodes join new clusters (e.g., change clusters), some data packets are lost when the nodes send the acknowledgment to the new cluster head. In the ARER(CF&DC) protocol, no acknowledgment messages are sent when nodes join a cluster.

Both higher sink speed and higher event speed negatively impact the percentage of composite events processed successfully by the sink. A higher sink speed triggers more frequent reconstruction of the convergecast tree, resulting in packets being dropped. A higher event speed leads to new clusters forming more frequently. Packets may be dropped during the cluster formation when messages for joining the cluster and acknowledgment are exchanged, and also some possible contentions and collisions when event-reporting messages are sent to the sink.

From both experiments in Fig. 8, we can see that the percentage of composite events processed successfully at the sink is slightly higher for larger events. This is because more sensors detect the event, thus the redundancy in event reporting helps alleviate the impact of packet dropping.

9 Conclusions

This paper presents ARER, an improvement of the anchor-based routing protocol with constrained flooding and dynamic clustering from [1]. A new clustering algorithm and a new rule for dynamic clustering are proposed. ARER has a better performance in terms of energy consumption and composite events detected successfully by the sink.