EAAI Journal 2026 Journal Article
Domain adaptation fault diagnosis method based on discriminative feature enhancement)
- Chenhui Qian
- Zhaojun Yang
- Jialong He
- Chi Ma
- Chenchen Wu
- Shaoyang Liu
Many unsupervised domain adaptation methods have been proposed to address the cross-domain fault diagnosis problem. However, when applied to complex mechanical equipment, these methods often fail to achieve satisfactory results, primarily due to the strong local dependency of their fault features. Local dependency makes the relationships between features more complex, leading to poor performance in handling feature differences across domains, which in turn affects the model's generalization ability and diagnostic accuracy. To address this, a domain adaptation fault diagnosis method based on discriminative feature enhancement (DADFE) is proposed. First, a contrastive mixed attention mechanism (CMA) is proposed, which calculates attention for fault features through grouped attention and incorporates contrastive learning to create a contrastive normalization layer. This improvement enhances the uniformity of the feature space and resolves the issue of feature collapse caused by insufficient constraints in the attention mechanism. Next, a multi-angle Taylor metric layer (MATM) is designed to assess feature diversity from various perspectives. The Taylor series expansion is utilized to perform nonlinear expansion on high-dimensional feature clusters, further enhancing the separability of the discriminative space. The experimental results on bearing and reducer datasets fully validate its effectiveness and advantages.