Abstract: The main shaft thrust bearing is a critical component of wind turbines, and any failure can lead to significant losses for the unit. In order to achieve early fault warning for the main shaft thrust bearing of wind turbine units and take timely maintenance measures to avoid further expansion of the fault, this paper proposes a stacking fault early warning model based on data acquisition and supervisory control（SCADA） under normal operation of wind turbine units. Firstly, this paper uses the goodness of fit and mean square error of four single models to comprehensively rank the features and obtain datasets with four different combinations of gradient features. Secondly, through the analysis of the predictive performance and correlation of single models, XGBoost, LightGBM, and random forest are ultimately selected as the base learners, and XGBoost as the meta-learner to establish the Stacking ensemble learning prediction model. Experimental results show that the stacking model for predicting the temperature of the main shaft thrust bearing performs the best, with a significant improvement in prediction error compared to the base learners. Finally, the root mean square error（RMSE） of the model temperature prediction is calculated, and the error threshold for the normal state of the main shaft thrust bearing is set based on the exponential weighted moving average（EWMA） method. Experimental results demonstrate that the stacking model established in this paper can issue a fault warning for the main shaft thrust bearing of the wind turbine at least 6 hours in advance.