Rolling bearing fault diagnosis under varying operating conditions using a convolutional-Transformer multi-alignment approach

In this study, a novel deep transfer learning method, termed the Universal Domain Alignment and Multi-level Alignment Network (UDAM-Net), is established to address the significant feature distribution discrepancy between the source and target domains caused by the difficulty of feature extraction for rolling bearings under complex operating conditions. In the feature extraction stage, a one-dimensional convolutional neural network is integrated with a Transformer-based multi-head attention mechanism to effectively enhance the feature representation capability of raw signals. Additionally, the complementary advantages of Joint Maximum Mean Discrepancy (JMMD), Multi-kernel Maximum Mean Discrepancy (MK-MMD), and a domain discriminator are explored by designing a collaborative mechanism combining distribution metrics and adversarial learning in conjunction with transfer learning. Furthermore, an uncertainty-based weighting strategy is introduced to adaptively adjust the loss function and dynamically balance the contributions of different alignment modules. Finally, the Adam optimizer is employed to accelerate model convergence. Experimental results on the CWRU public dataset and a laboratory-built test rig dataset demonstrate that UDAM-Net achieves classification accuracies of 96.67 % and 96.17 %, respectively. This study enlightens fault diagnosis based on feature extraction and cross-domain feature alignment.