A method for rolling bearing fault diagnosis based on domain-adaptive learning

Abstract Bearings are a vital component of rotating machinery, and their operating condition directly affects the stable operation of the equipment. However, due to changes in operating conditions, the diagnostic accuracy of traditional models fluctuates, and they are unable to adapt quickly to new operating conditions. Consequently, a diagnostic method has been devised for variable speeds and variable loads, with the aim of improving efficiency whilst maintaining high diagnostic accuracy. Specifically, the maximum mean difference is used to reduce the feature distance between the two operating conditions, whilst domain-adversarial and featureextraction modules are constructed to extract domain-invariant features. This method of aligning first and then extracting reduces the number of ineffective adversaries in adversarial training, thereby improving diagnostic speed. On the other hand, domain adversarial learning improves diagnostic accuracy under varying loads by extracting generalized features. The experimental results show that diagnostic accuracy exceeded 98% for both the CWRU dataset and simulated data. Ablation experiments have demonstrated the superiority of feature alignment followed by domain-invariant feature extraction. Using this platform and the data scale and partitioning methods employed in this study, a single training run takes approximately 11 seconds. This study achieved optimal diagnostic accuracy after just 30 training rounds; meanwhile, the model's total parameter size is 13.12 MB, occupying approximately 0.4 GB of video memory, making it suitable for deployment on small platforms. Consequently, this study is better suited to bearing fault diagnosis.