Multi-source variational decoupling network with adaptive source activation for zero-shot compound fault diagnosis

Abstract Modern mechanical systems are structurally complex, and the number of potential compound fault patterns can be exponentially large. Acquiring training samples for all possible fault combinations in advance is impractical, which severely limits the generalization capability of intelligent diagnostic models to previously unseen compound faults. Unsupervised decoupling methods provide a potential solution by decomposing compound fault signals into single-fault source components. However, a major limitation of existing approaches lies in inaccurate identification of the number of fault sources, which can easily lead to false alarms or missed detections. In addition, the decoupled source components often lack explicit fault semantics, resulting in misalignment with subsequent diagnostic objectives. To address these challenges, a multi-source variational decoupling network is proposed for zero-shot compound fault diagnosis. Trained solely with single-fault samples, the proposed method employs variational inference to adaptively identify the number of fault sources present in previously unseen compound signals and decomposes them into semantically explicit source components, followed by fault type recognition for each source. Experimental results on two compound fault diagnosis tasks demonstrate that the proposed method effectively handles zero-shot compound faults involving two or more fault types in rotating machinery, achieving higher diagnostic accuracy and improved interpretability compared with existing methods.