A Distributed Fault Detection System Based on IWSN for Machine Condition Monitoring

This paper introduces a novel framework for industrial wireless sensor networks (IWSNs) used for machine condition monitoring (MCM). Our approach enables the use of state-of-the-art computationally intensive classifiers in computationally weak sensor network nodes. The key idea is to split data acquisition, classifier building and training, and the operation phase, between different units. Computationally demanding processing is carried out in the central unit, while other tasks are distributed to the sensor nodes using over-the-air programming. The system is autonomously trained on the healthy state of a machine and then monitors a change in behavior which indicates a faulty state. Thanks to one-class classification, there is no need to introduce the faulty state of the machine in the training phase. We extend the diagnostic capability of the system using dynamic changes in the data acquisition and classification parts of the program in the sensor nodes. This enables the system to react to ambiguous machine states by temporarily changing the diagnostic focus. Compressing the information in the individual sensor nodes provided by in-node classification allows us to transmit only the classification result, instead of full signal waveforms. This enables the MCM system to be deployed with a large number of nodes, even with high sampling rates. The proposed concept was evaluated in IRIS IWSN by means of a rotary machine simulator.