A multi-stage domain generalisation method via exploiting causal disentanglement for machinery fault diagnosis

Traditional intelligent diagnosis models overly rely on the independent and identically distributed assumption, which makes it difficult to generalise to new operating conditions. In addition, it is difficult to collect fault sample data in practical engineering applications. To address the above problems, a multi-stage domain generalisation method via exploiting causal disentanglement is proposed for machinery fault diagnosis. The proposed framework filters out noise interference while preserving critical features through Fast Fourier Transform preprocessing combined with a multi-scale feature adaptive shrinkage ConvNeXt. Specifically, according to the causal analysis hypothesis, signal features are divided into domain-invariant and domain-specific features. As a result, this technique can effectively eliminate interference factors that are not directly causally related to fault categories, while retaining domain-invariant feature classifiers to achieve generalised diagnosis of unknown data. Meanwhile, the generalisation performance of the domain-invariant classifier is further enhanced by a multi-source domain mixup-based augmentation method. Finally, experimental results illustrate the superiority and generalisation ability of the proposed method compared with several popular fault diagnosis methods.