Simulation Data-Driven Semi-Supervised Fault Diagnosis Method for Rolling Bearings

In addressing the challenges of acquiring substantial and effective rolling bearing fault data in industrial environments, as well as the limited availability of labeled samples which hinders the generalization performance of intelligent diagnostic models, this paper proposes a semi-supervised fault diagnosis method driven by simulation data. Firstly, a fault-bearing simulation model under non-stationary conditions is established, generating comprehensive fault data by simulating various fault forms based on specific parameters set according to bearing type and operating conditions. Secondly, the proposed approach integrates a deep multi-scale residual network with traditional semi-supervised methods, effectively utilizing a large number of unlabeled samples along with a limited number of labeled samples for fault classification. Finally, by comparing the time-domain and envelope spectrum plots of the simulated data with real data, it is observed that the vibration waveforms and fault characteristic frequencies are highly similar. The proposed method demonstrates a significant improvement in diagnostic accuracy compared to traditional semi-supervised methods in various experiments. The results indicate the feasibility of using simulation-generated data to compensate for the lack of real data and highlight the superiority of the proposed semi-supervised method in fault diagnosis with scarce labeled samples.