Bio-Inspired In-Network Filtering for Wireless Sensor Monitoring Systems Guillermo G. Riva Universidad Tecnol´ ogica Nacional Facultad Regional C´ ordoba, Argentina Email: griva@scdt.frc.utn.edu.ar Jorge M. Finochietto Universidad Nacional de C´ ordoba CONICET, C´ ordoba, Argentina Email: jfinochietto@efn.uncor.edu.ar Guillermo Leguizam´ on Universidad Nacional de San Luis LIDIC, San Luis, Argentina Email: legui@unsl.edu.ar Abstract—In-network filtering schemes can be used for com- puting type-threshold functions in wireless sensor networks. Instead of relaying all data to a sink node, sensor nodes can filter measurements to provide only the set of data required to compute a given function (e.g., maximum, range). In this context, the network can progressively learn where relevant data are available and use this information to compute the function over time by only querying a subset of nodes. Trails between sink and these nodes can be obtained based on bio-inspired strategies, reducing the energy consumption and prolonging the network lifetime. The adaptive behavior of swarm intelligence allows to overcome a lot of obstacles presented in wireless communication networks. In this work, we evaluate the PhINP (Pheromone-based in Network Processing) mechanism, which drives the filtering process based on the integration of metaheuristic and learning algorithms. MAX function computation in one- and multiple-source environment monitoring is used as a case study. We show by simulation that communication cost can be significantly reduced respect to tradi- tional mechanisms, increasing the network lifetime, while keeping a low computational error. Finally, node density requirements for efficient event detection in real applications are analyzed. Keywords—Wireless sensor networks, in-network filtering and computing, pheromone-based swarm intelligence, reactive systems I. I NTRODUCTION A wireless sensor network (WSN) is a set of tiny and inexpensive energy-constrained sensor nodes communicated through a shared and unreliable wireless medium. WSNs can be deployed over an area to monitor physical phenomena such as temperature, pollution, noise levels, etc. In this sense, single- and multi-events monitoring using WSNs has emerged as a novel and efficient solution. WSNs are mainly data-centric networks, and in the form of direct and reversed multicast trees rooted at base station or sink node. From a point of view, a WSN can be viewed as a distributed database in which sensor nodes temporally update their read- ings and to which an user can request for specific information by computing a given function. How to find nodes with relevant information for a given function, to compute and extract the resulting information from the network with minimum communication cost and low error is a challenging problem, especially if no information about the location of these nodes is available a priori. In a WSN, the information can be either proactively (event-driven) or reactively (query-driven) reported by nodes and relayed by intermediate ones to the sink by using multi-hop communication. Hybrid approaches are considered promissory strategies because they take advantage of the best of both. However, sensor nodes have constrained sensing, processing and communication capabilities mainly due to limited energy resources [1]. As data communication is the main source of energy consumption, an efficient resource management is a key factor in the design of WSNs [2]. This issue challenges us to develop distributed routing and processing strategies to maximize the network lifetime. This work focus on reactive (on-demand) schemes for large WSNs, which have less communication overhead than proactive ones. In this sense, the development of energy-efficient distributed algorithms (i.e., query dissemination and data collection) for reactive WSNs is discussed in this paper. One- and multi-source monitoring requires not only an initial discovery stage, but also a continuous search for new relevant data due to the spatial and temporal dynamics of the physical phenomena. Additionally, sensor nodes around the sink run out of energy more quickly than other nodes in the network, producing communication holes around the sink, thereby disconnecting the network. This effect is mainly produced by a many (sources) to one (sink) data flow, where sink’s neighbor nodes forward a lot of packets, becoming a bottleneck for the correct operation and lifetime of the network. WSNs are often deployed for applications in which specific computation functions are performed over collected data. To this end, the sink node disseminates a query message in the network that carries a function evaluation request. Examples include computing AVG and MAX/MIN values of a sensed field. Instead of relaying collected data, nodes can collaborate to compute these functions in a distributed fashion (in-network computation). As a consequence, a smaller amount of data is reported to the sink node; thus, reducing the communication cost and increasing the network lifetime. In this context, the computation of symmetric functions, which are invariants to permutations of their arguments, is of great interest as most statistical functions belong to this class [3]. Among these functions, type-sensitive and type- threshold ones can be considered. The former require to know a minimum fraction of arguments for the function value to be determined. Examples include average, median, histogram and majority. The latter depend only on element-wise maximum (minimum) of the histogram and a threshold vector. Instances of these functions are MAX, MIN, range, bottom-n and top- n. Intuitively, the value of a type-sensitive function cannot be determined if a large enough fraction of the arguments are unknown, whereas the value of a type-threshold function can be determined by selecting only relevant arguments.