CNN-BiLSTM-Autoencoder hybrid for prognostics of gearbox Over-Temperature faults in offshore wind turbines

• Novel CNN-BiLSTM-AE model for early gearbox fault detection in offshore wind turbines. • RF and PCC feature selection enhances prediction accuracy for oil temperature monitoring. • Adaptive warning via reconstruction error outperforms fixed-threshold methods. • Validated on six turbines, achieving 930-min average early warning lead time. • Robust cross-turbine generalization supports scalable predictive maintenance. The escalating adoption of offshore wind energy underscores the need for robust fault detection systems, particularly for gearbox failures that account for 30–40 % of turbine downtime, incurring significant economic losses. Traditional threshold-based methods for monitoring gearbox oil temperature suffer from delayed warnings and limited generalization, prompting the development of a novel fault early warning method based on the CNN-BiLSTM-AE model (CBL-AE-FEW). This approach integrates convolutional neural networks (CNN) for spatial feature extraction, bidirectional long short-term memory (BiLSTM) networks for temporal dependency modeling, and autoencoders (AE) for adaptive reconstruction error analysis. Utilizing SCADA data from six offshore wind turbines, feature importance is assessed via Random Forest and Pearson correlation coefficient, selecting key predictors such as hydraulic oil temperature. The method dynamically captures temperature trends, triggering early warnings when reconstruction errors escalate. Validation demonstrates superior performance, with a mean squared error of 0.06513, root mean squared error of 0.25564, and an average lead time of 930 min across turbines, surpassing traditional models like XGBoost and SVM. This study offers a reliable, generalizable solution for predictive maintenance, enhancing the operational stability and economic viability of offshore wind farms.