A Virtual Data-Driven Fault Diagnosis Method Based on Adaptive Regularisation Consistency

The success of traditional fault diagnosis methods depends on obtaining a large number of training samples in the form of full faults, which may not be available in practical applications. Recent studies have shown the feasibility of constructing virtual representations reacting to fault situations by building physical models when fault samples are insufficient. In this paper, a virtual data-driven domain-based adaptive fault diagnosis method based on adaptive regularisation consistency, ARC, is proposed to ensure the consistency between the virtual and real domains. Firstly, the bearing dynamic model is constructed to obtain the simulation data, and the simulation data and the real data are input into the trained original model and the target model at the same time. The data distributions of the simulated and real samples are adaptively brought closer by adding weights to the maximum mean discrepancy (MMD); Kullback-Leibler Divergence (KLD) is used to make the feature representations extracted from the target model similar to each other and between the original model and the feature representations extracted from the target model, and regularization is used to fine-tune the feature extraction of the model. An entropy-based adaptive pseudo-labelling selection method is proposed to filter low-quality samples and prevent negative transfer. The diagnostic results of the case study show that the proposed method is able to utilize the diagnostic knowledge from the simulation data to achieve fault diagnosis of mechanical equipment and outperforms the commonly used unsupervised cross-domain fault diagnosis methods.