With the development of the wind turbine, the icing problem of wind turbine blades under low temperature and humidity has been paid more and more attention. When the blades freeze, it causes the power loss of the wind turbine, causing the blades to break up in severe time, resulting in huge power loss and maintenance costs. At present, when the wind turbine issues an alarm about icing, the ice conditions are usually very serious. So it is necessary to achieve the early prediction of icing about wind turbine blades. In this paper, data driven method is used to predict the blade icing of wind turbine, the blade states are identified by initial data preprocessing, feature extraction, model construction and training. Logical Regression model and XGBoost model are constructed respectively. Then, the accuracy and prediction time of the above two models for blade icing prediction are analyzed and compared. Finally, the model suitable for online detection of blade icing is obtained. Through calculation and analysis, the XGBoost model is better than the Logistic Regression model in predicting blade icing, and the prediction time is only 0.449 seconds, which can fully achieve the purpose of online detection.