Abstract

A new Cluster-based methodology for real-time Novelty Detection and Isolation (NDI) in sensor networks, is presented. The proposed algorithm enables uniform clustering across time-frames to indicate the presence of a “healthy” network. In the event of novelty, the associated sensor is seen to be clustered in a non-uniform manner with respect other sensors in the network, thereby facilitating fault isolation. Moreover, a statistical approach is proposed to determine a noise tolerance level for reducing false alarms. Performance of the proposed algorithm is examined using datasets obtained from a number of industrial case studies, and the significance for fault detection for such systems is demonstrated. Specifically, it is shown that through a correct selection of the noise tolerance level, an emerging failure is successfully isolated in presence of other abrupt changes that visually might be perceived as indication of a failure.