An Intelligent Multi-Fault Diagnosis Method for Asynchronous Motors Based on Depth-Wise Convolutions Enhanced Transformer

This paper proposes an intelligent diagnostic network, termed Depth-Wise Convolutions Enhanced Transformer (DWCET-Net), to address the challenges of concurrent multi-fault identification in Asynchronous Motors operating within complex industrial environments, where conventional diagnostic approaches remain constrained by manual feature extraction dependencies and suboptimal generalization capabilities. The proposed methodology develops a multi-channel image fusion technique based on Gramian Angular Field with Guided Filtering (GAGF), which transforms time-frequency characteristics of multi-phase voltage signals into noise-robust 2D image representations. By effectively integrating the local feature extraction capability of Depth-Wise Convolutions with the global temporal modeling strength of Transformer network, this approach overcomes the technical limitations of conventional methods in feature engineering construction and concurrent multi-fault scenarios. To validate model performance, a comprehensive experimental dataset encompassing six categories of typical faults was established, including normal condition, inter-turn short circuit, eccentricity, inner/outer bearing ring failure and broken bar failure. Experimental results demonstrate that compared with baseline models such as AlexNet, ResNet, and standard Transformer network, the DWCET-Net achieves significant improvements in key metrics including mean Average Precision (mAP), mean Average Recall (mAR), and Macro-F1 score, exhibiting superior classification performance particularly in concurrent multi-fault identification scenarios. This advancement provides a robust framework for intelligent fault diagnosis in complex electromechanical systems.