Design and optimisation of a distributive model-based sensor fault detection method for automated in-network execution in a wireless sensor network

Abstract

In this paper, a distributed model-based sensor fault detection method is presented for detecting and identifying spike faults without the requirement of the existence of reference sensors. This method partitions the sensor network into sensor pairs and carries out fault diagnosis within these sensor pairs based on autoregressive with exogenous input time series analysis. The performance of the proposed method is evaluated by implementing the algorithm in a 16-node wireless sensor network deployed to monitor the traffic induced accelerations of the Grove Street Bridge (Ypsilanti, Michigan). Spike faults are generated on-site and superimposed on the acceleration measurement before being acquired by some of the monitoring system wireless sensors. In addition to accuracy evaluation, this study focuses on the relationship between the detection accuracy and three different network partition methods. Based on this relationship, a communication energy saving partition method is presented. The proposed algorithm achieved a detection rate of over 85% yet reduced communication energy by more than 54% when compared to a centralised fault detection method implemented in the monitoring system base station.