A prompt-guided domain generalization network with post-hoc interpretability for mechanical fault diagnosis under unknown operating conditions

In mechanical fault diagnosis, substantial discrepancies in operating conditions often exist between the source and target domains, which severely constrain the performance of domain generalization (DG) methods. Moreover, DG methods generally lack post-hoc interpretability, making it difficult to explicitly identify the physical causes underlying the diagnostic results. Therefore, this paper first proposes a prompt-guided DG network (PDGN) for mechanical fault diagnosis under unknown operating conditions. PDGN integrates multi-dimensional knowledge prompts to explicitly guide the decoupling of multi-source features, enhancing the extraction of domain-invariant features and thereby improving generalization to unknown operating conditions. Moreover, this paper proposes a method named physical feature sensitivity analysis and GPT semantic reasoning (PFSA-GPT) to provide post-hoc explanations for the diagnostic results of PDGN. PFSA-GPT perturbs physical features and quantifies their contributions to diagnostic performance, and further leverages the reasoning capability of GPT to explicitly identify the physical causes underlying the diagnostic results. The experimental results demonstrate that, across three datasets, the proposed method exhibits strong generalization capabilities and high interpretability in mechanical fault diagnosis under unknown operating conditions, thus offering an efficient and transparent diagnostic tool for engineering applications.