Adaptive Time-Leaping Gradient Aligned Based Unsupervised Domain Adaptation for Bearing Fault Diagnosis Under Imbalanced Samples

Unsupervised domain adaptation shows promise in fault diagnosis by aligning cross-domain feature distributions. However, class imbalance in the target domain degrades the discriminative feature structure and hinders domain shift reduction. To address the challenge, a novel Time-Leaping Gradient Aligned Dynamics Gramian Angle Feature Pyramid Network (TGA-DGAFPN) is proposed. Firstly, a dynamic time leaping sampling mechanism that adaptively predicts reconstructed time points, replacing fixed intervals in Gramian Angular Difference Fields. Secondly, a gradient alignment loss is designed based on time-leaping parameters and analyzed for its correlation with domain shift. This loss aligns the gradient directions between source and target features with respect to time leaping offsets, thereby promoting the dynamic time-leaping sampling mechanism to reduce domain discrepancies. These innovative modules are incorporated into a Feature Pyramid Network framework to integrate multi-scale feature information, thereby capturing more comprehensive and nuanced fault patterns. Extensive evaluations across nine tasks under three imbalanced scenarios demonstrate the superior performance of TGA-DGAFPN in fault diagnosis. In a slightly imbalanced scenario, it achieves 93.04% accuracy, 89.13% precision, and a 91.85% F1-score. Based on various scenarios, the performance of TGA-DGAFPN is superior to the current state-of-the-art models.