A machine learning approach for identifying and classifying faults in wireless sensor networks

Wireless Sensor Network (WSN) deployment experiences show that collected data is prone to be faulty. Faults are due to internal and external influences, such as calibration, low battery, environmental interference and sensor aging. However, only few solutions exist to deal with faulty sensory data in WSN. We develop a statistical approach to detect and identify faults in a WSN. In particular, we focus on the identification and classification of data and system fault types as it is essential to perform accurate recovery actions. Our method uses Hidden Markov Models (HMMs) to capture the fault-free dynamics of an environment and dynamics of faulty data. It then performs a structural analysis of these HMMs to determine the type of data and system faults affecting sensor measurements. The approach is validated using real data obtained from over one month of samples from motes deployed in an actual living lab.