Due to the failure of existing wind turbine blade icing detection ensemble methods in effectively utilizing the strengths of different individual classifiers, a blade icing detection model based on feature extraction and optimal weighted ensemble learning was proposed. Firstly, the features associated with icing were extracted using a stacked denoising auto encoder. After evaluating the performance of various individual classifiers and comparing their effectiveness in binary classification applications, the random forest, extreme gradient boosting tree, light gradient boosting machine, and K-nearest neighbor algorithms were selected as individual learners. The algorithms were then optimized for hyper parameters using the Bayesian optimization algorithm. Then, an optimal weighted ensemble strategy, based on sequential quadratic programming, was proposed to identify the state of the blade. Finally, the historical data of the No. 15 wind turbine and No. 21 wind turbine were simulated. The experimental results show that the proposed detection model has improved numerous indicators compared to the individual models and other ensemble models. The accuracy has reached 99.2%, indicating its effectiveness in detecting icing.