Transfer learning-based fault diagnosis of wind turbine high-speed shaft bearings using CWRU pretraining and public run-to-failure data

Fault diagnosis of wind turbine high-speed shaft bearings (HSSBs) is hindered by the distribution gap between laboratory data and field signals, and by the lack of labeled run-to-failure recordings. This study proposes a three-stage transfer learning framework to address both problems. Stage 1 pretrains a lightweight convolutional backbone with a frequency-domain attention module (FAM) on the CWRU bearing dataset. Stage 2 applies a hybrid alignment strategy combining maximum mean discrepancy and adversarial training to reduce marginal and class-conditional domain divergence. Stage 3 fine-tunes the adapted model on run-to-failure data using health-indicator-driven dynamic sample weighting for degradation-stage recognition and early warning. Experiments on PRONOSTIA, XJTU-SY, and a wind-turbine field dataset collected from an operational 1.5 MW turbine show that the proposed method outperforms six baselines in cross-domain accuracy and early warning time. Ablation results confirm that FAM, hybrid alignment, and dynamic weighting each contribute independently, and their combination is super-additive.