Large-Kernel Group Convolutional Perceptron Attention Network for Interturn Short Circuit Fault Diagnosis in PMSM

This article proposes a novel network architecture, namely the large-kernel group convolutional perceptron attention network (LCPA-Net), for the purpose of diagnosing interturn short circuit (ITSC) in permanent magnet synchronous motors (PMSMs). The primary focus of this network is to effectively extract informative features from the acquired signals, a crucial aspect when employing data-driven approaches for fault diagnosis. The large-kernel group convolution technique is introduced to partition the input into multiple sub-regions, wherein two pathways, namely convolution and multi-layer perceptron (MLP), are jointly employed to extract local features. And a multi-head self-attention mechanism is implemented to mitigate the reliance of conventional neural networks on sequential ordering. Experimental results demonstrate that the proposed methodology yields a remarkable improvement in fault diagnosis accuracy, achieving an impressive 99.86%. Moreover, the robustness of this approach is validated across various speed conditions.