Deep causal factorization network: A novel domain generalization method for cross-machine bearing fault diagnosis

Abstract

Domain generalization (DG) has attracted much attention in bearing fault diagnosis since it can generalize the prior diagnostic knowledge to invisible working conditions. However, previous DG-based approaches are easy to introduce personalized bias due to the specificity of the mechanical equipment, which worsens the generalization performance. To solve this problem, a deep causal factorization network (DCFN) is proposed for cross-machine bearing diagnosis without the involvement of target domain data in training. Specifically, by leveraging the structural causal model of bearing fault signal generation, DCFN defines the cross-machine generalized fault representations as causal factors and the domain-related representations as non-causal factors. Afterwards, taking advantage of the causal properties that can be preserved in data distribution shifts, causal task factorization and feature factorization modules are designed to reconstruct causal mechanisms. To separate the two types of underlying factors, causal task factorization aims to maximize the predicted output entropy of the domain classifier using learned causal factors as input, and simultaneously maximizes the entropy of the fault classifier using learned non-causal factors as input. Moreover, causal feature factorization highlights that the ideal causal factors desire cross-domain consistency and inter-dimensional independence, thereby learning causal factors with stable and sufficient distinguishing features. Finally, under broad bearing fault datasets including public, laboratory and simulation datasets, the effectiveness of the proposed DCFN is verified by comparing with various state-of -the-arts diagnosis methods.