Intelligent fault diagnosis of bearings under small samples: A mechanism-data fusion approach

In recent years, deep learning has been extensively applied to bearing fault diagnosis with remarkable achievements. However, in real industrial scenarios, the primary challenge in developing an effective intelligent diagnosis model is the scarcity of fault samples required for training due to shutting down operations during failure behavior. Simulation data-driven and data augmentation-based small sample diagnosis methods have their own limitations, including insufficient diagnostic performance and low data quality. In view of these, a novel mechanism-data fusion diagnosis scheme with bearing dynamic model and multi-agent diverse generative adversarial network (MAD-GAN) is proposed in this study. Specifically, the bearing dynamic model addresses the scarcity of failure samples by simulating vibration signals across various operating conditions. Besides, A simulation-real transformation model based on MAD-GAN is developed to achieve the conversion between simulated domain and real domain, which shares the diagnostic knowledge between two domains. Finally, the domain adversarial based comprehensive generalization network is improved by maximum mean discrepancy and metric learning to further generalize diagnosis knowledge from simulation domain to real domain. Two bearing experimental datasets are applicated and case studies are conducted under small sample, validating the effectiveness and superiority of the proposed method. Experimental results show the potential application of this method in real industrial scenarios.