ITSC Fault Diagnosis for Five Phase Permanent Magnet Motors by Attention Mechanisms and Multiscale Convolutional Residual Network
Qian Chen, Xushu Dai, Xiangjin Song, Guohai Liu
Abstract
This article proposes a multiscale convolutional residual neural network algorithm with attention mechanisms for early interturn short circuit (ITSC) fault diagnosis of five-phase permanent magnet synchronous motors (FPPMSMs). First, a multiscale convolutional neural network with channel attention is used to highlight the fault features. Second, a spatial attention residual module is developed to improve feature learning performance, which can alleviate the problem of disappearing gradients and further enhance the fault features in the feature map. Finally, a self-attention structure is adopted to reduce reliance on manually set parameters, capture the internal correlation of features, and improve the interpretability of the network model. Experiments on ITSC faults of FPPMSM are carried out. The experimental results and the comparison with the other four methods highlight the superiority of the proposed method.