OAIFAN: A Noise-Robust Discriminative Feature Unification Framework for Cross-Speed Fault Transfer Diagnosis
Yuhan Huang, Xiaoxi Hu, Huan Wang, Yiming He, Jingming Cao
Abstract
Despite impressive advances in multisource domain generalization for cross-speed fault diagnosis, two critical challenges remain unresolved. First, existing methods neglect inherent discrepancies among source domains when extracting generalizable knowledge. This oversight significantly hinders effective model training and generalization. Second, industrial signals inevitably contain noise interference. Current denoising techniques overly rely on classification objectives, preserving only classification-relevant features rather than accurately distinguishing noise from genuine fault signals. To address these challenges, we propose the Order Analysis-Informed Fault-Aware Network (OAIFAN). For the first challenge, the framework employs an equal-angle resampling strategy. This approach maps multisource signals into a unified angular domain, effectively reducing cross-speed distribution discrepancies. For the second challenge, OAIFAN integrates two physics-guided modules. The Dynamic Reweighted Thresholding Denoising Module (DRTDM) combines adaptive thresholding with positional attention mechanisms. This design achieves precise noise removal while retaining fault-specific temporal features. The Adaptive Feature Band Enhancement Module (AFBEM) leverages fault-sensitive indicators, including Normalized Energy Ratio (NER) and kurtosis, to selectively enhance fault-relevant frequency bands. This physics-informed architecture ensures explicit noise identification and optimal fault feature extraction. Furthermore, a joint distribution alignment loss is incorporated to refine category-wise feature alignment, enhancing domain generalization. Experimental results on two test rigs demonstrate that OAIFAN outperforms state-of-the-art methods in cross-speed fault diagnosis, effectively addressing domain shifts and noise interference.