An unsupervised transfer learning gear fault diagnosis method based on parameter-optimized VMD and residual attention networks

Abstract

Traditional gear intelligent fault diagnosis methods require a large amount of labeled training data. It is challenging to train a high-precision fault diagnosis model due to the issue of insufficient fault data. Transfer learning can reduce the requirement for sufficient labeled data. When the data from the source and target domains differ significantly, the accuracy of the current transfer learning-based fault diagnosis techniques is poor. To deal with this problem, a deep transfer learning gear fault diagnosis method is presented. Firstly, a variational mode decomposition (VMD) and gramian angular field (GAF)-based data preprocessing technique is suggested to denoise the signal and convert the one-dimensional signal into two-dimensional images. Next, this paper proposes an improved residual attention convolutional neural network (IRACNN) to extract the signal's multi-scale spatial features, thereby improving the network's capability to extract gear fault features. Finally, this paper suggests a staged transfer training strategy to align the class-level feature distribution. This paper sets up a gear fault test platform in the laboratory to verify the suggested method and demonstrate its superiority.