One-class transformer-based feature disentanglement with clustering-driven semisupervised learning for the fault diagnosis of reciprocating machinery

In industrial machinery, fault diagnosis is crucial for preventing significant economic losses and ensuring operational safety. Traditional approaches often rely on extensive datasets containing both healthy and faulty data across various operating conditions, which can be challenging to obtain, especially for high-tech equipment that operates reliably for extended periods. This paper presents a novel decoupled transformer-based domain generalization approach that addresses this limitation by leveraging exclusively healthy operational data, thereby significantly simplifying deployment compared to existing methods that require comprehensive datasets of both healthy and faulty data. Our method employs a clustering-based semi-supervised learning procedure, which not only decouples domain generalization from fault diagnosis training but also efficiently operates without domain information during inference. Notably, our proposal requires only a small amount of labeled data under faulty categories, which can become available after the domain-generalization phase has been completed, further enhancing practical applicability. Experimental results demonstrate promising performance in 46 domain generalization tasks involving a reciprocating compressor with valve-related faults, highlighting the practical industrial value and deployment simplicity of our proposed method.