As the core component of wind turbines, gearboxes frequently fail. It is significant to study the fault diagnosis methods of the wind turbine gearboxes. Considering that the Knearest neighbors (KNN) diagnosis method was insensitive to noise and the accuracy of fault diagnosis was low, a fault diagnosis method based on wavelet packet and improved kernel Knearest neighbors algorithm was proposed. This method used wavelet packet analysis technology to extract the fault features, and eliminated the noise by mutual nearest neighbor criterion. Then, an improved Knearest neighbors classification decision rule based on kernel method was established. Experiments showed that this method could effectively improve fault diagnosis accuracy and robustness, and provide new ideas for the research of intelligent diagnosis technology.