Learning Category-Invariant Disentangled Features for Domain Generalization in Machine Fault Diagnosis
Zuoyi Chen, Jiandi Wu, Zhongwei Deng, Hong-Zhong Huang
Abstract
Existing domain generalization (DG) methods typically align multidomain data into a shared feature space to extract domain-invariant representations for machine fault diagnosis. Departing from this alignment paradigm, this article proposes a novel learning category-invariant features (LCIFs) framework. Motivated by the concept of genetic markers, the LCIF explicitly decomposes machine health states into domain attributes (reflecting operating conditions) and category attributes (reflecting fault types), and disentangles category-invariant features from multidomain observations. To this end, a patch-level feature adaptive aggregation is built in the feature extraction module to suppress noise attributes, and learnable condition masks are constructed to capture multigranularity local and global features. Parallel attribute disentanglement branches are then established to extract category and domain features separately. For the category attribute branch, an Earth Mover’s Distance based cross-attention mechanism is designed to strengthen category-specific channels and suppress domain interference. Experiments on public benchmarks and a self-collected dataset, covering single and multiple unseen domain diagnostic tasks, show that the LCIF substantially outperforms several state-of-the-art DG methods on unseen operating conditions.