MWCDGN: a spectral-kurtosis guided multi-wavelet causal disentanglement generalization network for gearbox cross-domain fault diagnosis

Abstract

Domain generalization-based fault diagnosis methods have gained widespread attention for addressing cross-domain challenges under unseen operating conditions. However, current statistical methods primarily focus on capturing the relationships between vibration signals and labels, which often leads to entangled features and limited generalization performance. To this end, we propose a spectral-kurtosis guided multi-wavelet causal disentanglement generalization network (MWCDGN) for gearbox cross-domain fault diagnosis. First, a physics-enhanced encoder is developed based on the spectral-kurtosis guided multi-wavelet convolutional module to extract multi-scale fault-sensitive features with explicit physical significance. By integrating multi-wavelet convolution and spectral-kurtosis-guided adaptive fusion, the proposed encoder can effectively emphasize discriminative impulsive and resonance-related characteristics. Building upon this, a structural causal model (SCM) is integrated within a spatiotemporal fusion framework to disentangle the extracted features into fault-related causal representations and domain-specific non-causal representations. An SCM-inspired spatiotemporal framework then learns fault-related and domain-related representations, with geometric aggregation and redundancy reduction promoting functional specialization and reducing cross-branch correlation. Finally, dual counterfactual causal intervention mechanism applies directional latent interventions to reduce domain-dependent interference and enhance the robustness of fault-related representations under domain shifts. Diagnostic case studies on three datasets under various working conditions demonstrate the effectiveness and superiority of MWCDGN.