Application of dynamic neural networks with exogenous input to industrial conditional monitoring

Intelligent assessment of information gathered from industrial-grade data loggers for preemptive maintenance is one of the foremost areas of research in conditional monitoring. Due to the general operating environment, there exists a non-linear relationship between the input and output data gathered from these sensors. Moreover, the transmission of data from such dynamic environments is generally marred by a large SNR with substantial level of “false-noise” belonging to the normal movement pattern of the mechanical parts. Within this context, the goal of this paper is to explore, evaluate and develop an optimal, dynamic neural network to improve the fault prediction accuracy of condition monitoring systems. The training data for this research was obtained from a vibration and a thermal sensor connected mounted over a polyphase induction motor. The objective was to identify any anomalies in the motor's fan-based cooling system. Moreover, the model presented a comparative analysis of a dynamic neural network (DNN) model against a non-linear autoregressive neural system (NARX) with exogenous input. The validation outcome presented a close regressive relationship of 0.9734 between observed and targeted outcomes over a 7-second delay with a NARX model giving a 4.56% and 5.23% classification accuracy. The best model system was evaluated against unseen anomaly data and demonstrated high prediction accuracy.