Detecting Faulty Wireless Sensor Nodes through Stochastic Classification

In many distributed systems, the possibility to adapt the behavior of the involved resources in response to unforeseen failures is an important requirement in order to significantly reduce the costs of management. Autonomous detection of faulty entities, however, is often a challenging task, especially when no direct human intervention is possible, as is the case for many scenarios involving Wireless Sensor Networks (WSNs), which usually operate in inaccessible and hostile environments. This paper presents an unsupervised approach for identifying faulty sensor nodes within a WSN. The proposed algorithm uses a probabilistic approach based on Markov Random Fields, requiring exclusively an analysis of the sensor readings, thus avoiding additional control overhead. In particular, abnormal behavior of a sensor node will be inferred by analyzing the spatiotemporal correlation of its data with respect to its neighborhood. The algorithm is tested on a public dataset, over which different classes of faults were artificially superimposed.