Early fault warning for wind turbine main bearings using XGBoost and parameter-based transfer learning with SCADA data

Main bearings in wind turbines, operating under low-speed heavy-load conditions, are prone to incipient faults that cause extended downtime and high maintenance costs, especially in remote offshore environments. Conventional SCADA-based diagnostic methods often lack cross-turbine transferability and struggle with data scarcity in new or data-limited units. This study proposes a cost-effective early fault warning framework combining Extreme Gradient Boosting (XGBoost) for accurate main bearing temperature prediction with parameter-based transfer learning. For cross-farm transfer, the model was retrained using five universal SCADA features shared by both source and target datasets (wind speed, cabin temperature, ambient temperature, gearbox oil temperature, main bearing temperature), the model is trained on data-rich 6.8 MW offshore units (Guangdong Bodhi Sea) and transferred to 8.3 MW units in the same farm and 2 MW onshore units (Yunnan Daheishan), requiring only ∼3 days of target-domain fine-tuning. Results show competitive performance (R 2 > 0.95, RMSE ≈ 0.432°C) over LSTM, GRU, LightGBM, and recent transfer learning methods; note that the R 2 metric is partly driven by the autocorrelation of the temperature signal and should not be taken as a standalone measure of model skill. Real-world validation on a single fault event (rolling element fracture, Unit 15# at Yunnan Daheishan) detected the anomaly approximately 40 days before the fault was first discovered during a scheduled periodic inspection; the true fault onset is unknown, so this figure represents a lower bound on advance warning. The one-sided adaptive residual threshold ensures high sensitivity and low false alarms. The approach offers a scalable and transferable solution for predictive maintenance in multi-model wind farms.