Efficient Sensor Fault Diagnosis in Wireless Sensor Networks.

Abstract

Nodes in wireless sensor networks (WSN) are prone to faults due to their inexpensive components and due to the harsh environments in which they are deployed. Therefore, automated fault diagnosis algorithms are necessary to ensure network functionality and measurement quality. Because wireless sensor networks have limited energy resources and consist of a large number of sensors, there is a need for fast and power efficient sensor fault diagnosis algorithms. This thesis proposes two frameworks of efficient sensor fault diagnosis. The first is a distributed model-based fault diagnosis framework for embedment in the WSN nodes. Fault specific algorithms are designed under this framework for detecting and identifying spike and non-linearity faults without the use of reference sensors. These algorithms fill the gap between existing centralized model-based and distributed model-free frameworks. In addition, they have the benefit of being scalable, power efficient and highly accurate. In the second framework, group testing-based fault diagnosis algorithms are proposed for situations where the number of faulty sensors is much smaller than the number of sensors in the network. These group testing algorithms evaluate sensors on a collective basis instead of individual basis. This study designs a Kalman-filtering based method for evaluating a group of sensors to determine if faulty sensors exist in the group. This method, together with the combinatorial group testing technique, is able to detect faulty sensors in O(d^2log(N)) tests, where d is the number of faulty sensors and N is the size of the network. This study also develops a Bayesian adaptive group testing algorithm in which test pools are designed based on previous group test results. This method further reduces the required number of tests and is suitable for noisy group test systems. Algorithms of both frameworks are evaluated by simulated and real sensor data with faults present. Results show that the distributed algorithms are able to achieve a detection rate of 85% or higher while keeping the false alarm rate low (~1%) under typical faulty signals. The group testing algorithms are able to reduce the required number of tests significantly while achieving similar accuracy as the traditional fault detection methods.