Fault Diagnosis for Inverter-Fed Motor Drives Using One Dimensional Complex-Valued Convolutional Neural Network

This paper proposed a new one dimensional complex-valued convolution neural network (1D CVCNN) model to diagnose power switch open-circuit fault of three-phase inverter-fed PMSM system. The 1D CV convolution operation was defined and the CV rectified linear unit (ReLU) activation function was chosen. A CV backpropagation algorithm is also proposed for 1D CVCNN training. The 1D CVCNN framework model is built, where 1D inputs and all the weights between the layers are complex numbers. The Clarke transformation is used to process the three-phase current of the inverter to obtain a complex-valued signal. The non-overlapping sliding window sampling method is used to obtain CV data set. The fault classification accuracy of the 1D CVCNN has been verified by experiments, and the experimental results show that the 1D CVCNN has better feature extraction ability than any other conventional deep learning method and better robustness to noise.