Abstract: In the light of the unbalanced data,adaptive neighboring hybrid resampling method is used to process small types of data in the original data,increase the effective examples of small types of data.A similar function-based undersampling algorithm processing is designed to reduce the repetitive data in large types of data.The dimensionality of the data is reduced without changing the high information of the data,and finally the characteristic data is imported into the support vector machine（SVM）and is optimized with PSO.The experimental results show that the prediction performance of the feature extraction in the model reaches 79.21%,and the improvement is 22.93%and 48.83%in the Extreme Learning Machine （ELM） and Random Forest （RF） algorithms,both the significant improvement provides a new idea for the diagnosis of wind turbine blade icing faults.