STFT–LoRA–DANN: a lightweight adversarial transfer learning framework for cross-load bearing fault diagnosis

Abstract Due to the significant differences in energy distribution, noise characteristics, and feature patterns of vibration signals under varying load conditions, cross-load fault diagnosis has become a central challenge in rolling bearing health monitoring. To address the weak cross-domain generalization and high adaptation cost of traditional deep models, this study proposes an integrated lightweight transfer learning framework termed STFT–LoRA–DANN. The method first constructs stable time–frequency representations via STFT to enhance feature consistency across different loads. Subsequently, only a small number of low-rank adaptation parameters are inserted into the pretrained SeResNeXt50 backbone, enabling controllable feature-space adaptation with substantially fewer trainable parameters and reduced optimization cost. Finally, adversarial domain alignment based on DANN is employed to explicitly reduce the feature distribution discrepancy between source and target domains. Across twelve cross-load transfer tasks on the CWRU dataset, the proposed framework achieves near-saturated and competitive performance, reaching an average accuracy of 99.4% and outperforming classical transfer baselines such as MMD, MK-MMD, and DANN in most transfer directions. Confusion matrix and training-dynamics analyses further demonstrate that the framework maintains stable convergence and good generalization behavior with limited parameter updates, suggesting potential practical value for cross-load diagnosis scenarios with data scarcity and varying operating conditions. The results indicate that the combination of LoRA and adversarial domain adaptation provides a scalable parameter-efficient paradigm for low-cost intelligent mechanical fault diagnosis.