A novel sliding mixing graph contrastive domain adaptation method for fault diagnosis under time-varying speeds

Deep learning-based bearing fault diagnosis methods at time-varying speeds have been extensively investigated, achieving considerable outcomes. However, the non-stationary characteristic of the fault data distribution over time at varying speeds poses a significant challenge in adapting to the dynamic distribution for fault identification in industrial applications. To overcome these difficulties, a novel sliding mixing graph contrastive domain adaptation (Mix-GraphCDA) method is proposed in this paper. Initially, acquired raw signals are transformed into frequency domain signals, followed by the construction of a spectral correlation weighted graph. Subsequently, an improved graph feature learning model is developed. Instead of limiting itself to node feature induction, this model explicitly considers edge weights and the nodes' own features in the aggregation process. It facilitates the capture of node feature neighborhood information as well as latent distributions. By employing the contrastive learning method, maximize the similarity between each domain and its mixed view to eliminate distribution differences and alleviate feature coupling caused by time-varying speeds. Finally, the proposed diagnosis method was verified comprehensively with two laboratory and one publicly available dataset. The experimental results indicate the high diagnostic accuracy and robustness of our diagnostic method across different and conditions.