Physics-guided SCADA-based transformer model for robust early detection of wind turbine gearbox faults

Abstract

Among the various components of a wind turbine, the gearbox incurs the highest maintenance costs. Therefore, early fault prediction of wind turbine gearboxes has become an area of growing interest, with the aim of enabling predictive maintenance and reducing maintenance costs. This study has developed a Transformer-based deep learning model for early gearbox fault prediction using only SCADA data collected under healthy operating conditions. A reliable normal behaviour model is established through a systematic selection of SCADA variables that are correlated with the gear bearing temperature, combined with rigorous data cleaning and normalisation procedures. In selecting SCADA variables for training the normal behaviour model, the first principle of thermodynamics is applied to the bearing system to identify key variables that most strongly influence bearing temperature. In contrast, many existing studies rely primarily on Pearson correlation to select variables based on statistical relationships with the target signal. The proposed physics-based approach provides a more robust and physically interpretable representation of the relationships between input variables and the target output. The model is validated using data from a real wind turbine and demonstrates its ability to identify gearbox bearing temperature anomalies prior to fault occurrence, thereby providing valuable lead time for maintenance interventions before faults escalate into major gearbox failures. Furthermore, a comparative analysis with an LSTM-based model shows that while the Transformer does not consistently provide earlier fault alarms across all events, it demonstrates improved robustness in distinguishing genuine faults from non-critical high-temperature episodes, thereby reducing false alarms.