Early fault detection for power plant fans based on dynamic neural network

Abstract Early fault detection is increasingly important for the reliable and secure operation of power plant fans. In this paper, we propose an early fault detection strategy for power plant fans based on dynamic neural network. First, the nonlinear autoregressive (NARX) with exogenous inputs network is utilized to predict the behavior of fans by using normal operating data, and the discrete particle swarm optimization (DPSO) is utilized to optimize the hyper-parameters of NARX network to enhance prediction accuracy. Then, the fault alarm strategy is proposed by the prediction deviations and generalized extreme value (GEV) theory. Finally, the proposed method is applied to a forced draft (FD) fan in a coal-fired power plant. Experiment results show that the proposed model has high prediction accuracy on normal operations and produces large prediction errors when a failure occurs. Furthermore, it can detect early faults accurately and timely before the failure occurs.