Fault Diagnosis Algorithm for Power Module Based on A Hybrid Attention Network

Fault diagnosis has a significant contribution to the reliability of the power module, the stability and accuracy of the fault diagnosis algorithm based on single-channel data have major limitations, and existing fault diagnosis algorithms based on multi-channel data have limitations such as insufficient capability of extracting features and low utilization rate of inter-channel correlation information. Therefore, in this paper, a multi-channel fault diagnosis algorithm for power modules based on a hybrid attention network is proposed, aiming at solving the problems of limited fault characterization capability of single-channel data and insufficient feature extraction of existing multi-channel input models. The hybrid attention module of this network adopts a mechanism that combines time attention and channel attention to extract the features of multi-channel data from both channel and time series dimensions of the input data in a holistic manner. The experiments have demonstrated that the hybrid attention network proposed in this paper achieves an accuracy of 99.40% in the multi-channel fault diagnosis test set, which is highly state-of-the-art and practicable, and can provide new ideas and methods for the research in fault diagnosis.