An Enhanced Partial Transfer Fault Diagnosis Network Aided by Dual-Force Boundary Refinement and Inverse-Forward Adaptive Alignment

Partial domain adaptation (PDA) is better suited to address the practical challenges of transfer fault diagnosis in rotating machinery. However, existing PDA methods often suffer from two key issues: the propagation of uncertainty caused by high prediction scores for incorrect source domain classes and instability in label distribution alignment due to relying on only low-weight filtering by the classifier. To overcome these issues, this article proposes an enhanced partial transfer fault diagnosis network with dual-force boundary refinement (DFBR) and inverse-forward adaptive alignment (IFAA). First, the 1-D vibration signals are converted into 2-D representations using the Gram angle field (GAF). ResNet50, embedded with coordinate attention (CA), is employed as a feature extractor to capture high-dimensional, fine-grained features for knowledge transfer. Second, the developed loss fusion enhancement strategy (LFES) increases classifier discrimination and stabilizes label distribution alignment. DFBR refines decision boundaries by promoting high-confidence predictions for correct classes while penalizing incorrect ones. IFAA steadily balances cross-domain label distribution by combining forward filtering with inverse progressive supplementation, extending the outlier samples filtering strategy by gradually enriching the target domain with source domain features. Finally, the superiority of the proposed method is validated through extensive comparisons with state-of-the-art models on PDA mechanical fault diagnosis tasks.