A Multiscale Feature Residual Network With Staged Optimization for Unsupervised Cross-Domain Fault Diagnosis

In the fault diagnosis of industrial equipment, the traditional supervised learning method is difficult to effectively deal with the unlabeled data in the target domain because of the significant differences of working modes. To address this issue, a multi-scale feature residual network with staged optimization (MFSResnet) is proposed in this paper. First, an enhanced dilated convolution module and two novel residual blocks are designed to improve feature representation ability. Building on these components, a multi-scale feature residual network is then proposed, integrating the enhanced dilated convolution module, the residual blocks, and an SE attention module for effective feature extraction. Additionally, a staged optimization method is employed throughout the training process to optimize multiple loss functions. Finally, experiments analysis conducted on two bearing datasets, CWRU and PU, demonstrate the efficacy of the proposed approach in addressing the challenge of unlabeled data in the target domain, and its superiority over existing methods.