This paper describes a machine learning framework for gearbox health assessment in wind turbines, built around gearbox oil pump pressure rather than the thermal signals that dominate the existing literature. A Random Forest Normal Behavior Model (NBM) is used based on a healthy operational baseline to predict expected lubrication pressure, and the residual between predicted and actual values is smoothed over a 30-day rolling window to filter out operational variability and expose structural drift. The model was chosen after benchmarking against XGBoost and Support Vector Regression.

Keywords

Wind turbine, gearbox condition monitoring, SCADA, normal behavior modeling, Random Forest.