In response to the challenge of early warning for abnormal oil sump temperatures in wind turbine gearboxes, a fault warning method based on supervisory control and data acquisition (SCADA) data is proposed to enhance the operational reliability of the turbines. Firstly, by integrating wind speed-power distribution characteristics, an outlier detection approach utilizing the interquartile range and longitudinal filtering based on data dispersion is employed to eliminate anomalous power points. Subsequently, key input features influencing oil sump temperature are identified using a random forest algorithm, leading to the development of a temperature prediction model based on categorical boosting (CatBoost). The hyperparameters of this model are optimized using tree-structured parzen estimator (TPE). Finally, dynamic warning thresholds are established through statistical process control based on residual distributions. In a practical case study from a specific wind farm, this model issued effective warnings approximately 5 hours prior to gearbox failure; notably, the time points at which residuals exceeded control limits closely aligned with the progression of faults.The proposed method demonstrates significant efficacy in identifying abnormal conditions related to oil sump temperatures and possesses strong early warning capabilities along with substantial engineering application value.