In order to improve the diagnosis accuracy of wind turbine bearing fault signals, a new fault diagnosis method for wind turbine bearing was proposed to solve the problem of misidentification of unknown fault signals. Firstly, the vibration signals of wind turbine bearings were processed by empirical wavelet transform, and 15 time-frequency domain features were extracted from the decomposed inherent mode components to form a feature vector set. Then, the feature classification ability was evaluated by Gini index, and the optimal feature set was constructed. Finally, a hierarchical hybrid classifier combining single-class support vector machine and extreme learning machine was used for fault diagnosis. Compared with ELM and SVM classifier, the new method can identify the wind motor bearing fault signals with unknown fault types well.