Abstract
In complex mechanical systems, data-driven fault diagnosis faces significant challenges when encountering novel compound faults absent from the training set, a problem framed as Generalized Zero-Shot Learning (GZSL). Existing generative approaches often fail to capture the intrinsic physical coupling of fault signatures, leading to poor generalization due to the domain shift between synthetic and real data. To address this, we propose a novel framework termed Physics-Guided Causal Diffusion with Confidence-Rectified Ensemble (PGCD-CRE). First, we formulate a causal diffusion model that integrates multi-hot semantic conditioning with an adaptive physics envelope loss. This mechanism explicitly models transfer path attenuation and enforces the nonlinear superposition of fault impulses, enabling the high-fidelity synthesis of complex compound signals using only seen single-fault data. Second, to break the restrictive assumption of knowing the exact number of unseen fault types in advance and to mitigate the spectral confusion inherent in coupled faults, we design a specialist ensemble diagnosis strategy augmented by a density-based topological discovery (HDBSCAN) and a threshold-guided cross-confidence rectification mechanism. This module dynamically isolates environmental noise and aligns real unseen samples with their correct semantic categories without relying on predefined cluster counts. Extensive experiments on the XJTU and HUST datasets under complex industrial conditions demonstrate that PGCD-CRE significantly outperforms state-of-the-art methods, achieving robust accuracy in identifying unseen multi-component faults by bridging the semantic gap through physical causality and open-world discovery.

This publication has 12 references indexed in Scilit: