TimeXer-based main bearing temperature prediction and Bayesian fault early warning for offshore wind turbines via transfer learning

Main bearing failures are a primary driver of offshore wind turbine unplanned downtime, yet data-driven monitoring methods lack cross-turbine generalization and probabilistic fault detection. This study proposes an integrated framework combining the TimeXer model, transfer learning, and a two-stage Bayesian fault early warning system for main bearing temperature prediction. The TimeXer model applies patch-level tokenization to endogenous variables and cross-variable attention to exogenous SCADA inputs, capturing temporal dependencies and inter-variable coupling. Fault early warning sequentially applies a three-sigma threshold, Bayesian posterior probability estimation, and a three-step persistence criterion to suppress false alarms. A two-phase transfer learning protocol adapts the source-domain model to target turbines using limited data. Validated on 41,906 ten-minute SCADA samples from Yangjiang, China, the model achieves R 2 = 0.9898, RMSE = 0.3172 °C, and MAE = 0.1636 °C on the 58# 8.3 MW source turbine, surpassing Transformer, LSTM, and XGBoost baselines by up to 42.1% in RMSE. Transfer to the 24# 6.8 MW and 33# 8.3 MW target turbines yields R 2 = 0.9851 and 0.9840, with RMSE below the source baseline. The early warning system attains a step-level false alarm rate of 2.1% with a 94.4% detection rate, and component-wise ablation, prior sensitivity analysis and residual normality testing are reported to substantiate each design choice. Normalized error analysis further shows that the lower target-domain RMSE reflects the reduced temperature variability of the target machines rather than an accuracy gain. The framework provides a scalable pathway for offshore wind turbine predictive maintenance.