Applied Soft Computing 29 (2015) 12–25 Contents lists available at ScienceDirect Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc Heuristic routing with bandwidth and energy constraints in sensor networks S. Kavi Priya a,∗ , T. Revathi b , K. Muneeswaran c , K. Vijayalakshmi d a Mepco Schlenk Engineering College (Autonomous), Sivakasi, India Department of IT, Mepco Schlenk Engineering College (Autonomous), Sivakasi, India Department of CSE, Mepco Schlenk Engineering College (Autonomous), Sivakasi, India d Department of CSE, Ramco Institute of Technology, Rajapalayam, India b c a r t i c l e i n f o Article history: Received 20 August 2011 Received in revised form 15 October 2014 Accepted 15 December 2014 Available online 29 December 2014 Keywords: Sensor networks routing Bandwidth constraint Energy constraint Nearest neighbor tree Distributed algorithm Maximum lifetime a b s t r a c t Most of the routing algorithms devised for sensor networks considered either energy constraints or bandwidth constraints to maximize the network lifetime. In the real scenario, both energy and bandwidth are the scarcest resource for sensor networks. The energy constraints affect only sensor routing, whereas the link bandwidth affects both routing topology and data rate on each link. Therefore, a heuristic technique that combines both energy and bandwidth constraints for better routing in the wireless sensor networks is proposed. The link bandwidth is allocated based on the remaining energy making the routing solution feasible under bandwidth constraints. This scheme uses an energy efficient algorithm called nearest neighbor tree (NNT) for routing. The data gathered from the neighboring nodes are also aggregated based on averaging technique in order to reduce the number of data transmissions. Experimental results show that this technique yields good solutions to increase the sensor network lifetime. The proposed work is also tested for wildfire application. © 2014 Elsevier B.V. All rights reserved. 1. Introduction Wireless sensor network consists of large number of tiny sensor nodes connected via wireless communication channels. These are suitable for lots of applications such as military surveillance, temperature monitoring, wildfire detection, disaster warning, etc. In particular, sensors are deployed to monitor the regions where the human cannot intervene. For instance, sensors deployed for wildfire detection in the forest region continuously monitors the environment to detect the changes in temperature. When the temperature value crosses the threshold value say 40 ◦ C (event detection), sensor routes the data to sink node (typically a base station or a sensor/actuator node or a gateway to larger network with high computing power and energy where information is required) in the remote location through the multi-hop routing algorithms. Therefore, the sink collects the data from all the sensor nodes to derive useful information about the event (for example the geographical map of the wildfire can be plotted) detected. Fig. 1 shows the model of wireless sensor networks used in the proposed work. According to the characteristics of sensor network, the sensor nodes perform sensing, preprocessing, aggregation and transmission of data on its neighboring nodes within the transmission range. Hence, the total data rate increases suddenly in the sensor networks when it detects the event. The sensor data cannot be further forwarded to the neighboring node, if the sensor node runs out of energy or due to network congestion. The sensor network ∗ Corresponding author at: Mepco Schlenk Engineering College (Autonomous), Sivakasi, Tamil Nadu, India. Tel.: +91 9842295563; fax: +91 04562235111. E-mail addresses: urskavi@mepcoeng.ac.in (S. Kavi Priya), trevathi@mepcoeng.ac.in (T. Revathi), kmuni@mepcoeng.ac.in (K. Muneeswaran), vijayasrini9701@gmail.com (K. Vijayalakshmi). http://dx.doi.org/10.1016/j.asoc.2014.12.019 1568-4946/© 2014 Elsevier B.V. All rights reserved. starts to congest when the total link bandwidth between the sensor nodes is smaller than the data rate of the network. Hence the wireless sensor networks are considered as resource scarce, which is manifested in terms of energy, link bandwidth, computing power, etc. In most of the previous works related to sensor networks, the authors tried to increase either energy efficiency through different routing techniques [1–10] or optimize wireless link bandwidth as in [11,12]. The classical routing algorithms like minimum spanning tree [13,14], requires calculation of routing path at every node and results in high computing power to find the optimal path. The use of the distributed algorithm to find the best optimal nearest neighbors for packet forwarding will increase the network’s lifetime. The network lifetime is considered as the time until which the first node in the sensor network drains out of energy. When every sensor node is allowed to forward data only to the nearest next neighboring node with optimal performance factor (energy or bandwidth efficiency) along with data aggregation (that converges number of data received from various sources into few messages), the sensor network’s lifetime will be maximized as discussed in [15–17]. In [18], the authors have devised a routing technique with both energy and link constraints which will have performance degradation since it is executed in a centralized fashion. In some of the recent works [22–24], energy efficiency is attained by increasing the network coverage (resulted in increased hardware cost), standby cluster head (suffered due to central point of failure if cluster node is dead) and efficient location discovery respectively. The researchers also concluded that the distributed routing algorithm may increase the sensor network’s lifetime. The works proposed in [2,5,9,15,17], suggests that using data aggregation in sensor network can utilize bandwidth efficiently. The survey of the papers [25–27] reveals that the performance of the sensor network may also depend on the type of application for which it is used. Therefore, this work proposes a model to tackle bandwidth constraints using link rate allocation and energy constraints using distributed NNT algorithm along with data aggregation considering the issues in the wireless sensor network wildfire application. S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Fig. 1. Model of wireless sensor networks. The rest of the paper is outlined as follows: Section 2 describes some of the research woks in the related area and their significance. Section 3 narrates the various constraints and modules used in the proposed work. Section 4 addresses the overall design of the proposed scheme and the various algorithms used in this work. Section 5 shows the simulation results and performance of the modeled system. Section 6 gives the application details of the proposed work in a model sensor network. Section 7 concludes the paper. 2. Related research works Chang and Tassiulas [1] reduced the sensor network traffic by routing the sensed data only based on the sensor node’s remaining energy. The authors conclude that this type of routing saves ad hoc network lifetime unlike the other algorithms which try to minimize the absolute consumed power. Schurgers and Srivastava [2] derived a practical routing guideline called gradient based routing based on the energy histogram for uniform resource utilization in the network. They also suggested that robust aggregation of packet streams will reduce the sensor node energy consumption by a factor of two or three. The algorithms proposed in [1,2] did not consider the number of messages transmitted between the neighbors that may result in increased communication cost. Bhardwaj et al. [3] proposed a model for bounding the lifetime of the sensor node to formulate energy constraints. The authors assumed that there was sufficient bandwidth and only one node could act as a source in the sensor network. But in some sensor applications like wild fire monitoring, poisonous gas leakage management system, etc. there is only limited bandwidth and many nodes can simultaneously send data at a particular time. Younis et al. [4] modeled a gateway-based routing technique that finds the optimal path in the network. In this scheme, a central network manager between the sensor clusters is responsible for routing. Chamam and Pierre [14] implemented a centralized tabu search heuristic to tackle the exponentially increasing computation time which is addressing the optimal planning of sensor states for cluster based sensor networks. Senthilkumar et al. [23] discussed honeybee technique for re-election of cluster heads to increase the energy efficiency of sensor networks. In this scheme, when a cluster head goes down, immediately the standby node becomes the cluster head. Thus the sensor network works continuously without any delay in order to choose the next cluster. These techniques are less suitable for sensor networks for two reasons: (i) It needs more computing power (ii) When the central manager node dries out of battery power; the entire network is out of control. Krishnamachari et al. [5] used Greedy Incremental Tree (GIT) heuristic scheme for data aggregation. In this scheme, the aggregation tree was built sequentially that will consume more time when compared to distributed schemes. Cui et al. [6] emphasized that the energy efficiency in a multi-hop routing environment must be supported across all layers of the protocol stack. They devised a cross-layer design based on variable-length TDMA schemes where the slot length is optimally assigned according to the routing 13 requirement of each sensor node. But this technique may not be suitable for large scale sensor networks. Ambühl [7] proposed an approximation algorithm that constructed minimum spanning tree based on the Euclidean distance to reduce the battery power consumption. Ok et al. [8] introduced a new metric for energy cost considering sensor node’s remaining energies as well as energy efficiency. This metric gave rise to the distributed energy balanced routing (DEBR) algorithm that increased the life span of sensor network. Misra and Dias Thomasinous [9] suggested a simple, least-time, energy-efficient routing protocol with one-level data aggregation that ensured increased lifetime, reliability and congestion avoidance in sensor networks. Qiu et al. [10] formulated enhanced tree routing (ETR) strategy based on sensor node address assignment schemes. In addition to the parent–child links, ETR also used links to other one-hop neighbors that will lead to a shorter path. It is shown that such a decision can be made with minimum storage and less computing cost. The works in [7–10] considered only the energy efficiency leaving behind the bandwidth efficiency for data forwarding in the sensor network. Hence, the number of network communications is increased whereas the sensor network’s lifetime is reduced. Madan and Lall [11] proposed a linear programming and subgradient distributed algorithm for routing and maximized the sensor network’s lifetime. Chang and Tassiulas [12] formulated a shortest path routing algorithm using sensor link costs. The link costs reflect both the communication energy consumption rates and the residual energy levels between the sensor nodes. This work helps to derive energy constraints that can maximize the sensor node lifetime. But the overhead in executing the algorithm is slightly high and is well suited only for small scale sensor networks. Li et al. [13] proposed a new localized routing algorithm called Incident MST (minimum spanning tree) for broadcasting in wireless adhoc networks. The author concluded that if the total link power was deterministic, then the algorithm might result in high energy efficiency. Cheng et al. [18] discussed the sufficient bandwidth and energy constraints for the sensor nodes to maximize the sensor network’s lifetime. The authors proved that ignoring bandwidth constraints on energy efficient algorithms may lead to infeasible routing solutions. They also aggregated received data in order to minimize the number of wireless transmissions between the sensor nodes. These works confirmed that the link bandwidth is also needed for efficient routing in sensor networks. Luo et al. [15] focused on optimizing both data transmission and aggregation costs by dynamically adjusting the route structure when the sensor nodes joined or leaved the network. This algorithm was slightly deviated from real network energy consumption for complex applications. Gallager et al. [16] designed a distributed algorithm that constructed the minimum-weight spanning tree. The sensor network was implemented as a connected undirected graph with distinct edge weights. The nearest neighbor tree algorithm was proposed based on the above scheme. Khan et al. [17] presented NNT algorithms ex the complete graph model where the maximum transmission range of the nodes were large enough so that any pair of nodes could communicate directly with each other. The authors proposed two NNT algorithms: Random-NNT and Coordinate-NNT based on types of ranking the sensor nodes for single hop wireless network. They also modeled UDGNNT (undirected graph NNT) for multi-hop wireless networks and proved that (1) the tree produced by such a distributed algorithm will be of low cost; (2) the NNT paradigm can be used to design a simple dynamic algorithm and (3) the time, message and work complexities of the NNT algorithms are close to the optimal energy constraint for routing. Konstantinidis and Yang [22] show that better energy efficiency can be achieved in a short period when solved in parallel as a multi objective evolutionary problem based on decomposition. The 14 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 optimization problem has to be devised more problem specific depending on the given application. Molina and Alba [24] proposed a meta-heuristic technique for solving location discovery problem in the sensor network. The authors used a two-stage search procedure that combines minimization of an error norm function with maximization of a maximum likelihood function. This model could be well used to identify the location of the source sensor by which the event location could be detected (for example the location of fire could be detected in wildfire application). Huang et al. [25] discussed Limb prosthesis application using body area sensor networks. The authors exhibited the effectiveness and practicability of the proposed prosthesis training system that used RFID and tags to recognize movement disorders. Huang et al. [26] analyzed underwater sensor network application and suggested routing using forwarding tree trimming mechanism for power efficient communication. The literature study of these papers helped to know about the changes in the behavioral characteristics of the sensor network with respect to the applications for which it is deployed. The above-mentioned sensor routing algorithms that maximized the network lifetime, either assumed sufficient wireless link bandwidth or used centralized routing technique. The key objective of this paper is to provide better network lifetime by combining energy-aware routing heuristics considering bandwidth availability together with data aggregation. First, distributed NNT algorithm is used to find the next neighbor to route the sensor data that makes the network energy efficient [17]. Second, a random bandwidth (based on reserved energy) is allocated for each link between the sensor nodes in order to acquire bandwidth efficiency [18]. The proposed scheme is tested for performance improvement in wireless sensor networks using NS2 simulator. The simulated sensor network implements the data-gathering application called wildfire monitoring system. Each sensor node in the simulation receives temperature data from the neighbor nodes, performs aggregation and sends the resultant data to the next node selected by NNT algorithm. The research simulation shows that the network lifetime is improved on applying such combined heuristics. 3. The proposed algorithm 3.1. Problem definition Wireless sensor network has numerous sensors that are densely deployed randomly in a remote geographic location without human intervention. These sensors are energy-scarce and bandwidth-scarce resources which are used cooperatively to monitor physical or environmental conditions. These sensors are connected to sink node via wireless links. The optimal path between the source node (sensor that generates data) and the sink node is computed using the distributed algorithm [17] that satisfies only energy constraints. This reduces the cost of constructing the minimum spanning tree, decreases the time and work complexity, and reduces the number of message transfers in the proposed work. The proposed model guarantees bandwidth efficiency of the sensor network through application of constraints formulated in [18] at every sensor node to push sufficient amount of data to the network based on available bandwidth. The data received within a time span is aggregated to avoid repetitive data forwarding (that in turn reduces the number of message transfers within the sensor network). If the deviation between the current sensor data and the received data is less, sensor node averages the values and send as a single value. If the value deviates more, then the node forwards the received data without aggregation. Hence, the proposed model reduces the number of message in the network. The modeled wireless sensor network consists of n number of nodes with wireless link bandwidth as B (bits per second). Each sensor node i has initial battery energy Ei (Joules). Each node i generates sensor data at a rate of Ri bits per second (Ri > 0 if node i is a source, Ri = 0 if it is a pure relay node, and Ri < 0 if it is a sink). The nodes consume energy on transmitting, receiving and sensing data, and their energy consumption rates are represented as Pt , Pr and Ps Joules per Bit respectively. Pr and Ps are assumed to be constants, but Pt is handled differently in the two models: in the uniform model, each node transmits at the same power level Pt ; in the non-uniform model, each node can transmit at different power levels. Any node i in the sensor network uses same power to transmit the data throughout its lifetime. But the transmitting power of the sensor node i will be different from the transmitting power of the sensor node j in the network is experimented in the non-uniform model. In simple words, Pti is not equal to Ptj . The energy-bandwidth constrained routing problem is stated as follows: suppose each node i’s rate Ri is known and the transmission rate from node i to node j is unknown, then the sources are preselected using NNT algorithm to find the optimal path. Let T be the total network lifetime. The rate allocation problem is to compute the data rate from node i to node j denoted as Rij on each link (i, j), given each node i’s Ei , Ri and link capacity B, so that the total network lifetime T is maximized and the rate allocation can be accommodated by wireless link capacity and energy reserve. The proposed work is applied in four different scenarios: single disk uniform transmission power model, single disk nonuniform transmission power model, double disk model and data aggregation model. 3.2. Single disk uniform transmission power model All sensor nodes in the network in this model have symmetric links (same transmission power). The transmission power Pti of a node i is same as the transmission power Ptj of the neighboring node say j (i.e. Pti = Ptj ). The effective transmission range is same as the interference range in the single disk model. Only the strong signals are decoded in this scheme. Let Ni , denote the neighboring nodes of i excluding i itself. The packets are routed to sink node through its neighboring nodes that are calculated using distributed NNT algorithm [17]. The proposed model uses TDMA scheme in which number of time slots assigned to the link between a node i and node j is proportional to the data rate from node i to node j at the MAC layer. Hence, for any node i to be a receiver, the TDMA schedule must guarantee (1) when node i is receiving, it cannot be sending, and (2) when node i is receiving from node j, none of its neighbors except j should be sending. The proof for this sufficient condition is given in [18]. If fi denotes the indicator for receiver, then fi = 1 if node i is a receiver or fi = 0 if node j is not a receiver as given below fi =  1, if 0, if  R j∈Ni ji >0 R j∈Ni ji ≤0   (1) The network lifetime T is maximized by minimizing 1/T. To minimize 1/T, the rates are allocated to each node based on the following three constraints. 3.2.1. Flow constraint For any two nodes say i, j in wireless sensor networks, the flow constraint f i,j) is less than or equal to link capacity c(i,j) (i.e. f(i,j) ≤ c(i,j)). Rij satisfies the flow conservation at each node as follows  j∈Ni Rij − Rji = Ri ∀i (2) 15 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 3.2.2. Energy constraint The energy constraint denotes the amount of energy spent at each node and is modeled as Ps Ri +  j∈Ni Pr Rji + Pt Rij = Ei /Ti , (3) ∀i 3.2.3. Bandwidth constraint Since all the nodes transmit sensor data on the same frequency channel, the bandwidth constraint for the network is formulated so that all links in the same collision domain will have bandwidth less than B. If node i acts as a sender, then the sum of all outgoing links should be bound by B, else if node i acts as a receiver, then node i’s sending, receiving, and interfering nodes total bandwidth must be utmost B, else the node i’s bandwidth constraint is automatically satisfied. These constraints are represented using the given equation  j∈Ni Rij + fi ·  j∈Ni  k∈Nj Rjk ≤ B, ∀i (4) The constraints for the variables Rij and fi is given as fi = {0, 1}, Fig. 2. Overall design of the proposed work. (5) ∀i The flow constraint is remains the same as in Eq. (2) described in Section 3.2. Let Ni+ , denotes the neighboring nodes of i that receives from i; and Ni− denotes the neighboring nodes that sends data to node i. Hence, the bandwidth constraint for the node to forward data is redefined as follows aggregation is the process of combining one or more data by taking mean, standard deviation, etc. thereby it is used to reduce the number of data transmissions. Simple averaging of data that is received from all the neighboring nodes in a given time span is performed and the averaged value is sent to the next node. For example, if a sensor node receives three different temperature readings like 30 ◦ C, 20 ◦ C, 40 ◦ C from the neighbor nodes, instead of forwarding all these data, the mean value 30 ◦ C is sent to the next neighboring node in the network. If sensor detects its own readings deviates more than 40 ◦ C compared to the received data from the neighboring nodes, then the sensor simply act as a relay node (Ri = 0). Aggregation reduces the number of message transmissions in the sensor network thereby contributes to increase in the sensor network’s lifetime. To accommodate the neighboring nodes with different data rates, the low-rate flows are combined with high-rate flows to improve the bandwidth efficiency of the sensor network. To aggregate data from different sources at a closer time frame say 5 ms, the flow conservation constraint can be designed as   0 ≤ Rij ≤ B, (6) ∀i , ∀j 3.3. Single disk non-uniform transmission power model The sensor nodes are assumed to have asymmetric links, but still use fixed transmission power. Any node i in the network can use Pti to transmit the data to its neighboring node whereas another node j in the same network can use different power Ptj (such that Pti = / Ptj ) to transmit the data packets to its neighboring nodes. Hence, the energy constraint is modified as follows Ps Ri +  Rij + fi j∈N + i j∈Ni (Pr Rji + Pti Rij ) ≤ Ei /Ti ,  Rjk ≤ B, ∀i ∀i (7) (8) j∈N − k∈N + i j In the case of single disk model which is described in Sections 3.2 and 3.3, the effective transmission range is the same as the interference range. But in reality, the interference range is usually larger than the effective transmission range. Though the nodes located out of radio’s transmission range is not strong to receive the data, sometimes can cause interference at others. So, in the double disk model with uniform transmission power, the bandwidth constraint is fixed considering the neighboring nodes that are in the interference range of node i. Let Ni denote the nodes in the transmission range and NiF be the nodes in the interference range. It is observed that, Ni ⊆ NiF since the interference range is always larger than the transmission range. Hence, the bandwidth constraint is changed to j∈Ni Rij + fi  Rjk ≤ B, ∀i (10) j∈Ni 4. Design 3.4. Double disk model  Rij = max{Ri Rji }, j∈Ni ∀i (9) j∈NiF k∈Nj 3.5. Data aggregation model Section 3.2–3.4 provide energy and bandwidth constraints for the basic routing in sensor network without aggregation. Data The proposed work comprises of four modules: (i) Finding the shortest route from source node to sink node using distributed NNT algorithm; (ii) Dynamic link rate allocation for sending data; (iii) Maximum life time routing (Heuristics I) and (iv) Optimizing the life path of sensor node using bandwidth constraints (Heuristics II) as shown in Fig. 2. Heuristics I and II are repeated for sensor nodes with uniform transmission power, without uniform power transmission, with data aggregation and without data aggregation. 4.1. Module 1: shortest path routing (SPR) The shortest route from the source node (that originates packet based on temperature value) to the sink node is calculated depending upon the constraints at each sensor node formulated using [18] in this module. The packets are forwarded to the next nearest neighbor using the distributed NNT algorithm as in [17]. The sensor nodes in the network are deployed randomly using a topology generator GenSen [19] and are implemented as a graph G. Any sensor node i possess a distinct identifier (denoted as id(i)) from the totally ordered set. Usually, in a sensor network, sink collects information from all nodes. So we designate sink to be the root of 16 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Table 1 Algorithm for Basic NNT Scheme. Algorithm: Basic NNT Scheme Input: All nodes with distinct ids from a totally ordered set. Output: A spanning tree. ===================================================== //Every node i executes the following steps independently Step 1: choose a unique rank rank(i) Step 2: Connect to the nearest node j if and only if rank(i) < rank(j) to add the edge (i,j) to the spanning tree. Step 3: Stop when the edge to sink node is added in the spanning tree. Table 2 Algorithm for Rank selection. Algorithm: Ranking Scheme Input: All nodes with distinct ids from a totally ordered set. Output: Rank of any node i (rank(i)). ===================================================== =============== //The sink node s initiates the algorithm Step 1: Sink chooses a large random number as rank (s) Step 2: Broadcast rank(s) and id(s) to the sink neighboring nodes //When a node i receives the first message from//one of its neighboring node j (may be sink s) 2.1: If (first message) then Calculate rank(i) D [rank(j) − 1, rank(j)] end //When a node i receives the second message from its another neighboring node k (may be sink s) 2.2: If (second message) then store rank(k), id(k) end Step 3: Repeat Step 3 until rank and id of all neighboring nodes is known. the spanning tree with highest rank in the network. Each sensor node has an omnidirectional antenna. The network is modeled as a UDG (unit disk graph), where any sensor node i can communicate directly to the neighboring node j if and only if there is an edge (link) between them. For a given transmission radius TR , there is an edge d(i,j) between two sensor nodes i and j if and only if d(i,j) ≤ TR (i.e. both nodes must lie inside TR ). Each node in the network chooses a unique rank (random number) and connects only to the higher rank node to rule out cycles, and constructs the spanning tree. The edge is established from a node i to the node j if and only if rank(i) > rank(j). Hence, the NNT algorithm constructs the spanning tree with low complexity. Table 1 shows the basic working of NNT algorithm to construct the spanning tree in a single hop network (all nodes lie within TR ). In reality, TR is not sufficient to cover all sensor nodes in the network. Hence, UDG-NNT is applied to support such multi-hop outing scheme. The rank of each node i is chosen as shown in Table 2 that guarantees every node is connected only to the closest node with higher rank in the network. The selection of node i’s rank in the interval [rank(j) − 1, rank(j)] (where j is the neighboring node of i) guarantees each node has atleast one neighboring node with higher rank in the network (i.e. rank(j) > rank(i)). Once the rank is calculated, each node i (except sink node s) selects the nearest node among the neighboring nodes and adds an edge to it in the spanning tree. To route the sensor data to sink node, each node uses three type of messages: request, available and connect to perform the algorithm. Each node begins transaction with broadcasting the request message (with its rank and id value) until it finds a neighboring node with higher rank. If the node that receives the request message has higher rank compared to the sender, then it responds the sender with available message (combination of rank and id value). The sender of request message finds the nearest higher ranked node from the rank information collected due to available messages received. Finally, the sender adds an edge to the spanning tree using the connect message (i.e. message (connect, i, j) is sent to j if j is the nearest higher ranked neighboring node of i). The details are given in Table 3. Table 3 Algorithm for Distributed UDG-NNT. Algorithm: Distributed NNT Input: All sensor nodes with distinct ids from a totally ordered set. Output: A spanning tree. ===================================================== //The algorithm is executed by each sensor node i independently and simultaneously. //Messages are written in the format (message name, sender, [recipient, [other information]) repeat Step 1: Set transmission radius TR Step 2: Broadcast (request, id(i), rank(i)) until (receipt of available message) //receive() is a function to receive data at each node Step 3: While (receive, i, rank(i)) for j = 1 to n //where ‘n’ number of nodes in the sensor network and j = / i, do 3.1: If rank(j) > rank(i) then send (available, i, j) to i end end Step 4: While (available, i, j)) for j = 1 to n//where ‘n’ number of nodes in the sensor network and j = / i, do 3.1: Select min(TR (j)) send (connect, i, j) to j that satisfies min(TR (j)) end end until (j == s)//where s is a sink node 4.2. Module 2: maximum lifetime routing (MaxLife) The lifetime of each sensor node in this module is calculated as in [11]. In [11], the authors proposed a distributed algorithm that maximizes the sensor network’s lifetime without considering bandwidth constraints. The authors formulated equations that optimized the energy efficiency of the network lifetime without taking into account the bandwidth considerations for a sensor node. In this module, Eqs. (4)–(6) are not utilized to provide sufficient conditions on link bandwidth while forwarding sensor data. Ignoring the bandwidth constraints results in infeasible routing solutions since link bandwidth affects the routing topology and data rate on a link. 4.3. Module 3: heuristic-I: scalable rate allocation on shortest paths Heuristic-I works on the shortest paths (in-terms of number of hops) from source to sink node and determines rate on each link using the available bandwidth. The shortest path is calculated based on the distributed algorithm; hence the sensor nodes consume minimal energy as depicted in [17]. The steps in Heuristic-I are described below (1) Calculate the shortest path from the source to sink node using distributed NNT algorithm. (2) Assume source rate is one unit, check bandwidth constraint for each node, and find the most bandwidth contentious node i. Then compute the scale factor ‘a’ as B/LHS, where LHS denotes the required bandwidth of node i’s collision domain. Set f = min{a/2,Ri }; (3) Push out f amount of flow from each source to the sink, then update the remaining input flow Ri′ = Ri − f for each source i. (4) Repeat Steps 5–7 either Ri′ or the network is fully saturated. (5) Find the shortest paths for nodes with Ri′ > 0 based on the current available nodes and links. Nodes that are saturated on bandwidth constraint Eq. (4) and their neighbors are not eligible for relaying. In case of a tie, give higher priority to nodes S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 17 Fig. 3. AVG ratio with uniform transmission power model. with more remaining energy; if there is still a tie, give higher priority to nodes with smaller degree. (6) Decide the scale factor ‘a’ in a similar manner as in Step 2. Let ‘m’ be the amount of flow to the next sensor node calculated as min {a, Ri′ }. If m does not decrease the lifetime, then set f = min {a, Ri′ }; otherwise, set f = min{a/2,Ri }; (7) Push out ‘f’ amount of flow from each source with Ri′ > 0, then update the remaining input flow Ri′ = Ri − f. 4.4. Module 4: heuristic-II: optimizing lifetime with bandwidth constraint This module is used to predict the feasible and infeasible routing solutions from the source to sink node. Heuristic-II works as a linear programming mathematical model based on Eqs. (1) and (4) to produce the best optimal solution. Heuristic-II can be narrated as follows (1) Set fi = 1 for sink and fi = 0 for all other  nodes to make the problem linear and solve; update fi = 1 if Rji > 0; if Eq. (4) is satisfied then return link rates Rij ; otherwise, go to Step (2). (2) Compute the shortest paths from sources to the sink using NNT. Rji > 0 and fi = 0, (3) Set fi = 1 for receiving nodes; solve the LP; if update fi = 1. (4) Repeat Step (3) until there is no update for fi or the linear program becomes infeasible. (5) If the linear problem converges, output Rij for all links (i,j) between nodes i and j. Fig. 4. MAX ratio with non-uniform transmission power model. 18 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Fig. 5. Normalized lifetime with uniform transmission power model.  Rji = 0 and fi = 1, set fi = 0 and Rji = 0 for (6) If it is infeasible: if all j belongs to Ni , solve the linear program again: if it is still infeasible, report infeasible. 5. Simulation and results The simulation is focused to show the relative performance of the shortest path routing using distributed NNT (labeled as SPR), MaxLife routing with energy constraints (MaxLife), Heuristics-I and Heuristics-II. The sensor network is simulated in NS-2 (Network Simulator 2) [21] with 50 nodes. In the simulated sensor network, the sensor nodes are deployed randomly using a standard topology generator which is written in C++ for wireless sensor network called GenSeN [19]. GenSeN is used to perform the deployment of the entire sensor network. The simulated wireless sensor network consists of 50 nodes deployed using random deployment strategy in a 100 × 100 square region in order to predict the performance of the simulation in collision domain. The simulation is carried out for duration of 1000 s and repeated for 15 different deployments. The physical layer selected for wireless communications is IEEE 802.15.4 with TDMA (time division multiple access) scheme at the MAC layer. In this simulation system, the performance of the sensor network is measured in terms of the network congestion and normalized lifetime parameters. The results show that combined heuristics with distributed NNT routing algorithm increases the sensor network lifetime and decreases the computing power Fig. 6. AVG ratio with non-uniform transmission power model. S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 19 Fig. 7. MAX ratio with non-uniform transmission power model. needed since it is executed in a distributed fashion with bandwidth and energy constraints. 5.1. Uniform transmission power model This simulation finds out how the bandwidth constraint can change the routing decision and eventually affect the lifetime of the sensor network. The transmission range is set to 30 m. In the uniform transmission power setup, each link has equal weight say 1. Four source nodes are randomly selected from the network, and increasing source rates are applied on them. Source rate is set to be a percentage of link bandwidth. The proposed schemes Heuristic-I and Heuristic-II are compared with MaxLife and the shortest path routing using NNT. The MaxLife and the NNT algorithm is chosen for comparison because it computes the maximum lifetime without considering bandwidth constraint. When there is enough bandwidth, MaxLife gives the optimal solution. SPR uses the shortest paths from sources to the sink by applying NNT algorithm, with link weight representing the transmission power of the node. The vertical lines in figures indicate after this point, increased data rate cannot be put through. Fig. 3 shows the average ratio (AVG ratio) of the required bandwidth in each collision domain to the offered bandwidth for this model. It is found that when each source node’s data rate Ri is increased above 12% of the given link bandwidth, MaxLife starts to congest since some of the sensor nodes in the collision domain require more bandwidth than the available one. The Fig. 8. Normalized lifetime with non-uniform transmission power. 20 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Fig. 9. AVG ratio with double disk model. SPR starts to congest when the source rate Ri is increased above 15%. Heuristic-I can work without congestion until the load is increased to 18%, and Heuristic-II can support as much as 17%. From the result it is observed that lower the average ratio of required bandwidth; the proposed scheme shows higher bandwidth efficiency. Fig. 4 shows the maximum ratio (MAX ratio) between the required and offered bandwidths. After highest maximum ratio is reached, the model stops working. The maximum throughput of the network is derived at the stop point. The results show that when there is enough bandwidth, MaxLife and Heuristic-II achieve the same optimal solution since they did not consider the bandwidth constraint. However, when bandwidth violation occurs, Heuristic-II can still push through 5% more data than MaxLife, and 2% more data than SPR. Heuristic-I can push through 6% more data than MaxLife and 3% more data than SPR, since they combine energy constraints and used NNT algorithm to find the shortest path. Heuristic-II achieves the best performance on lifetime and second best on throughput; Heuristic-I achieves the best performance on throughput, which is consistent with the observation from Fig. 5 compared to other two algorithms. Networks with more nodes can achieve longer lifetime than networks with 50 nodes because the workload is shared among more nodes. 5.2. Non-uniform transmission power model Transmission range is randomly selected within the range of 25–35 m in this simulation. The performance of the non-uniform Fig. 10. MAX ratio with double disk model. S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 21 Fig. 11. Normalized lifetime in double disk model. transmission power model is shown in Figs. 6 and 7. With asymmetrical edges, the performance is consistent with the Uniform Transmission Power model. But the network lifetime is reduced because the disparity in energy consumption is severe. Thus, Heuristic-II can give better result than the other models as shown in Fig. 8. 5.3. Double disk model The transmission range in this simulation is chosen as 30 m; interference range as 1.7× transmission range and the other data are the same as in Section 5.1. Figs. 9 and 10 show the throughput performance of all the four compared algorithms Fig. 12. Normalized lifetime with data aggregation. 22 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Fig. 13. NNT with wild fire application. in this model. From the figures, it is observed that when the interference range is larger, there is less chance for channel reuse; therefore the network throughput is less. Fig. 11 shows that the network lifetime is increased due to the lower data rate. 5.4. Data aggregation model This simulation computes the network lifetime improvement achieved through the data aggregation method. The solution is compared with the minimum spanning tree generated using NNT Fig. 14. Distance between source and sink. S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 23 Fig. 15. Packets forwarding. algorithm. Assume each node consumes 10% of energy while sending one unit of data. From the observation of Fig. 12, it is observed that there is a dramatic improvement on the network lifetime with data aggregation. Fig. 12 shows that Heuristics-I can push data until source rate is 17% of the link bandwidth. The SPR has stopped working due to congestion when the source rates crosses 14% of the link bandwidth. This indicated a throughput gain of 3% for Heuristics I over SPR. It is observed that the proposed heuristics with data Fig. 16. Result of packets forwarding. 24 S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Fig. 17. Result of proposed heuristic I. aggregation under energy and bandwidth constraints improved the network performance and reduced the congestion rate of the sensor network. 6. Application Wireless sensor networks are widely used in environmental applications like wild forest fire detection. Although wild forest fires occur relatively rarely, they must be detected early in order to prevent severe damages. To minimize needless communication between the sensor nodes for this usage, data aggregation technique is used in this paper. The algorithm is experimentally evaluated for continuous wildfire application as described in [20]. In this application, the sensor nodes comprise tiny temperature measuring devices like thermometer that monitors current temperature continuously. When the input to the sensor crosses a threshold value (normally 100 ◦ C), the sensor node has to send the temperature value to the sink node. For sending the sensor Fig. 18. Result of proposed heuristic II. S. Kavi Priya et al. / Applied Soft Computing 29 (2015) 12–25 Table 4 Energy efficiency of proposed models. Routing Shortest path routing using NNT MaxLife time routing Heuristics I Heuristics II Energy efficiency 16% 13% 19% 18% data, each node utilizes SPR or MaxLife or Heuristic-I or Heuristic-II models. The application details are shown in Fig. 13. The distance between the node and the sink node has to be calculated as in Fig. 14. Figs. 15 and 16 show the result of packet forwarding after applying NNT. Figs. 17 and 18 depict that the proposed model Heuristic-I and Heuristic-II take less energy to route data to the sink node, thereby improving the network performance in terms of lifetime. The energy efficiency of the proposed heuristics compared to other models is given in Table 4. 7. Conclusion In this paper, the network performance is analyzed by applying distributed algorithm for routing with bandwidth constraints to find the next neighbor node. The performance of the existing algorithms is evaluated and compared with the proposed work via simulation. The simulation results show that the sensor network’s lifetime increases by 17% and decreases network congestion due combined heuristics. The routing problem is also solved using linear programming model. Hence, the bandwidth efficiency and the energy efficiency are improved. The proposed technique works well for high traffic rates and also for large homogeneous networks. In future, it is decided to design an algorithm that is energy efficient and bandwidth efficient by applying the fuzzy based sleep scheduling technique (similar to the algorithm described in [28]) in order to improve the sensor network lifetime. Moreover it is also decided to apply complex data aggregation that avoids congestion communication overhead. And also some of the evolutionary algorithms (like fuzzy based optimization as in [29]) are considered to perform routing in the sensor network. Among the questions for future work, the most interesting is to model energy and bandwidth constraints as fuzzy objective function to select the optimal path in the network. References [1] J. Chang, L. Tassiulas, Energy conserving routing in wireless adhoc networks, in: Proc. IEEE INFOCOM 2000, vol. 1, 2000, pp. 22–31. [2] C. Schurgers, M.B. Srivastava, Energy efficient routing in wireless sensor networks, in: Proc. MILCOM 2001, vol. 1, 2001, pp. 357–361. [3] M. Bhardwaj, T. Garnett, A. Chandrakasan, Upper bounds on the lifetime of sensor networks, in: Proc. IEEE ICC’01, vol. 3, 2001, pp. 785–790. [4] M. Younis, M. Youssef, K. Arisha, Energy-aware routing in cluster based sensor networks, in: Proc. IEEE MASCOTS, 2002, pp. 129–136. [5] B. Krishnamachari, D. Estrin, S. Wicker, Modelling Data-centric Routing in Wireless Sensor Networks, Computer Engineering Technical Report CENG 02-14. [6] S. Cui, R. Madan, A. Goldsmith, S. Lall, Joint routing, MAC, and link layer optimization in sensor networks with energy constraints, in: Proc. IEEE ICC’05, vol. 2, May, 2005, pp. 725–729. 25 [7] Christoph Ambühl, An Optimal Bound for the MST Algorithm to Compute Energy Efficient Broadcast Trees in Wireless Networks. Technical Report No. IDSIA-22-04, November 2004. [8] Chang-Soo Ok, Seokcheon Lee, Prasenjit Mitra, Soundar Kumara, Distributed energy balanced routing for wireless sensor networks, Comput. Ind. Eng. 57 (August (1)) (2009) 125–135. [9] Sudip Misra, P. Dias Thomasinous, A simple, least-time, and energy-efficient routing protocol with one-level data aggregation for wireless sensor networks, J. Syst. Softw. 83 (May (5)) (2010) 852–860. [10] Wanzhi Qiu, Efstratios Skafidas, Peng Hao, Enhanced tree routing for wireless sensor networks, Ad Hoc Netw. 7 (May (3)) (2009) 638–650. [11] R. Madan, S. Lall, Distributed algorithms for maximum lifetime routing in wireless sensor networks, in: Proc. IEEE Global Telecommunications Conference (GLOBECOM), 2004, pp. 748–753. [12] Jae-Hwan Chang, Leandros Tassiulas, Maximum lifetime routing in wireless sensor networks, IEEE/ACM Trans. Netw. 12 (August (4)) (2004) 609–619. [13] Xiang-Yang Li, Yu Wang, Peng-Jun Wan, Wen-Zhan Song, Ophir Frieder, Localized Low-Weight Graph and Its Applications in Wireless Ad Hoc Networks, IEEE INFOCOM, 2004. [14] Ali Chamam, Samuel Pierre, On the planning of wireless sensor networks: energy-efficient clustering under the joint routing and coverage constraint, IEEE Trans. Mobile Comput. 8 (August (8)) (2009) 1077–1086. [15] Hong Luo, Yonghe Liu, Sajal K. Das, Distributed algorithm for en route aggregation decision in wireless sensor networks, IEEE Trans. Mobile Comput. 8 (January (1)) (2009) 1–13. [16] R. Gallager, P. Humblet, P. Spira, A distributed algorithm for minimum-weight spanning trees, ACM Trans. Progr. Lang. Syst. 5 (January (1)) (1983) 66–77. [17] Maleq Khan, V.S. Gopal Pandurangan, Anil Kumar, Distributed algorithms for constructing approximate minimum spanning trees in wireless sensor networks, IEEE Trans. Parallel Distrib. Comput. 20 (January (1)) (2009) 124–139. [18] Maggie Cheng, Xuan Gong, Lin Cai, Joint routing and link rate allocation under bandwidth and energy constraints in sensor networks, IEEE Trans. Wirel. Commun. 8 (July (7)) (2009) 3770–3779. [19] Tiago Camilo, Jorge Sá Silva, André Rodrigues, Fernando Boavida, GENSEN: a topology generator for real wireless sensor networks deployment, in: Proceedings of the 5th IFIP WG 10.2 International Conference on Software Technologies for Embedded and Ubiquitous Systems, 2007, pp. 436–445. [20] David M. Doolin, Nicholas Sita, Wireless sensors for wildfire monitoring, in: Smart Structures and Materials 2005: Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems, vol. 5765, March, 2005, pp. 477–484. [21] Network Simulator-2, Retrieved from: http://isi.edu/nsnam/ns/ [22] Andreas Konstantinidis, Kun Yang, Multi-objective energy-efficient dense deployment in wireless sensor networks using a hybrid problem-specific MOEA/D, Appl. Soft Comput. 11 (September (6)) (2011) 4117–4134. [23] J. Senthilkumar, M. Chandrasekaran, Y. Suresh, S. Arumugam, V. Mohanraj, Advertisement timeout driven bee’s mating approach to maintain fair energy level in sensor networks, Appl. Soft Comput. 11 (5) (July 2011) 4029–4035. [24] Guillermo Molina, Enrique Alba, Location discovery in wireless sensor networks using metaheuristics, Appl. Soft Comput. 11 (January (1)) (2011) 1223–1240. [25] Chenn-Jung Huang, Yu-Wu Wang, Tz-Hau Huang, Chin-Fa Lin, Ching-Yu Li, Heng-Ming Chen, Po Chiang Chen, Jia-Jian Liao, Applications of machine learning techniques to a sensor-network-based prosthesis training system, Appl. Soft Comput. 11 (April (3)) (2011) 3229–3237. [26] Chenn-Jung Huang, Yu-Wu Wang, Hsiu-Hui Liao, Chin-Fa Lin, Kai-Wen Hu, TunYu Chang, A power-efficient routing protocol for underwater wireless sensor networks, Appl. Soft Comput. 11 (March (2)) (2011) 2348–2355. [27] Zhi-gang Su, Pei-hong Wang, Jiong Shen, Xiang-jun Yu, Zhen-zhong Lv, Lu Lu, Multi-model strategy based evidential soft sensor model for predicting evaluation of variables with uncertainty, Appl. Soft Comput. 11 (March (2)) (2011) 2595–2610. [28] J.C. Cuevas-Martinez, J. Canada-Bago, J.A. Fernandez-Prieto, M.A. GadeoMartos, Knowledge-based duty cycle estimation in wireless sensor networks: application for sound pressure monitoring, Appl. Soft Comput. 13 (February (2)) (2013) 967–980. [29] Abdel Kareem Jaradat Mohammad, Reza Langari, A hybrid intelligent system for fault detection and sensor fusion, Appl. Soft Comput. 9 (January (1)) (2009) 415–422.