Multiscale Kernel-Based Residual CNN for Estimation of Inter-Turn Short Circuit Fault in PMSM

Abstract

The diagnosis of an inter-turn short circuit (ITSC) fault at its early stage is very important in permanent magnet synchronous motors as these faults can lead to disastrous results. In this paper, a multiscale kernel-based residual convolutional neural network (CNN) algorithm is proposed for the diagnosis of ITSC faults. The contributions are majorly located on two sides. Firstly, a residual learning connection is embedded into a dilated CNN to overcome the defects of the conventional convolution and the degradation problem of a deep network. Secondly, a multiscale kernel algorithm is added to a residual dilated CNN architecture to extract high-dimension features from the collected current signals under complex operating conditions and electromagnetic interference. A motor fault experiment with both constant operating conditions and dynamics was conducted by setting the fault severity of the ITSC fault to 17 levels. Comparison with five other algorithms demonstrated the effectiveness of the proposed algorithm.