Efficient Signal Processing and Anomaly Detection in Wireless Sensor Networks

Abstract

Abstract. In this paper the node-level decision unit of a self-learning anomaly detection mechanism for office monitoring with wireless sensor nodes is presented. The node-level decision unit is based on Adaptive Resonance Theory (ART), which is a simple kind of neural networks. The Fuzzy ART neural network used in this work is an ART neural network that accepts analog inputs. A Fuzzy ART neural network represents an adaptive memory that can store a predefined number of prototypes. Any observed input is compared and classified in respect to a maximum number of M online learned prototypes. Considering M prototypes and an input vector size of N , the algorithmic complexity, both in time and memory, is in the order of O(MN). The presented Fuzzy ART neural network is used to process, classify and compress time series of event observations on sensor node level. The mechanism is lightweight and efficient. Based on simple computations, each node is able to report locally suspicious behavior. A system-wide decision is subsequently performed at a base station.

Wireless Sensor Network Anomalies: Diagnosis and Detection Strategies

“…Octopus [15] is a sensor network monitoring and visualization tool that provides live topology and link state information on the nodes and enables users to issue simple preset commands to the nodes. [15] topology/state tool Network centralized visualization Momento [5] variance-based modules Network distributed target FP rate Sympathy [6] metrics collection tool Network centralized specify epoch Wang et al [16] Bayesian algorithm Data centralized n/a Rajasegarar et al [17] cluster-based algorithm Data hybrid n/a Walchli and Braud [7] prototypes algorithm Data hybrid n/a Ramanathan et al [9] rules-based tool Data/Node centralized n/a Chen et al [10] majority voting algorithm Data/Node distributed unattended Krishnamachari Bayesian algorithm Data distributed n/a and Iyengar [8] Chang et al [18] Recurrent NN algorithm Data centralized n/a Obst [19,20] Recurrent NN algorithm Data distributed n/a A separate class of approaches distributes the anomaly detection logic at the nodes. In contrast with tools, these approaches are algorithmic in nature.…”

“…Hybrid anomaly detection strategies [5,7,17] try to combine the best of both worlds: the availability of network state information at the back-end of the centralized approach and the responsiveness and low communication overhead of the distributed approach. A hybrid approach uses a centralized strategy for anomalies whose detection metric can be readily examined at the database, such as connectivity anomalies, broadcast storms, and node resets, and complex data anomalies.…”

“…However, decisions on whether anomalies have occurred are left for the central sink node, where the results of computations from all nodes converge. A similar computation model is proposed in [7], where sensor nodes classify incoming signals according to learned prototypes, prior to sending the classification to a fusion center for system-wide access patterns.…”

“…in space [4]) are more likely than conventional networks to experience anomalies related to connectivity or hardware failures. Recent work also focuses on devising detection strategies that target network level [5,6], data level [7,8], or node and data level [9,10] anomalies.…”

Improved Data Discrimination in Wireless Sensor Networks

“…Markus Wälchli and Torsten Braun [6] propose a sensor node level unsupervised anomaly detection mechanism, based on the Fuzzy Adaptive Resonance Theory (ART) neural network. The mechanism can be used for office monitoring and is able to distinguish abnormal office access from normal access.…”