Fault Diagnosis of Electric Motors by a Novel Convolutional-based Neural Network and STFT

This paper proposes a novel method for fault detection in electric motors based on Short-Time Fourier Transform (STFT) and a regulated convolutional-based neural network. In this method, STFT is applied over the raw signal measured from the motor to generate a 2D matrix as an input to a classification model. The classification model constitutes a regulated network combining Convolutional Long Short Term Memory (ConvLSTM) andConvolutional Neural Network (CNN) which is developed to be fit for the low-size 2D STFT matrices. The proposed method is evaluated over, a Permanent Magnet Synchronous Motor (PMSM) with healthy and three levels of Inter-Turn Short Circuit (ITSC) fault conditions. The model is also tested under the influence of measurement noise. The model has shown a good performance for all these conditions.