Cyclostationarity-enhanced multi-source domain generalization with mixture-of-experts for rolling bearing fault diagnosis

Cross-domain fault diagnosis for rolling bearings under unseen working conditions is a challenging yet essential task, as the distribution bias significantly degrades the performance of data-driven methods. Causality-inspired domain generalization aims to address this challenge, with existing methods primarily focusing on alignment- or gradient-based operations from an entire signal perspective, which, however, overlooks the risk that biased alignments may induce spurious correlations and misguide learning. To this end, we reformulate the problem from a more physical and fine-grained perspective by treating fault-irrelevant frequency bands as confounders and aiming at localizing causal frequency bands encoded with robust and interpretable fault-related features to explicitly extract causal features. We propose a frequency band-aware method with a cyclostationarity-enhanced representation. Specifically, we first introduce a representation based on spectral correlation density (SCD) for wavelet packet-decomposed frequency bands. Then, treating cycle frequencies as channels, an adaptive feature extractor is designed based on a mixture-of-experts (MoE) block with a multi-view router, which integrates the views of spectrum, frequency band, and entire sample, to adaptively extract features across samples. In addition, prior knowledge guidance is introduced to enhance robust features. With cycle frequency-level features of frequency bands, a frequency band-aware attention module based on a tokenized Transformer, enhanced with an entropy-based sparsity regularization, is designed to model inter-band dependencies and localize fault-related frequency bands for diagnosis. Experiments are conducted on Case Western Reserve University (CWRU), Paderborn University (PU), and Harbin Institute of Technology (HIT) bearing datasets, and the proposed method shows effectiveness and interpretability across transfer tasks with different spans.