Abstract: Aiming at the complex working conditions of wind turbine gearbox, a new early fault warning method was proposed based on the Gabor rearrangement logarithmic time-frequency ridges manifold. Firstly, the ridges of Gabor rearrangement logarithmic time-frequency spectrum were extracted and the high dimensional early fault feature vector was built. Then, LTSA (local tangent space alignment) manifold learning method was studied and improved to achieve the reduction of high dimensional feature vector. Finally, the K-nearest neighbor classifier was applied to complete the early fault identification and warning of variable conditional wind turbine gear box. Many experiments were carried out to get verifying data from different condition, including variable speed, load working conditions of planetary gearbox and wind turbine operation filed. The results show that the proposed method can improve the early fault warning accuracy of wind turbine gearbox that works under complex non-stationary conditions, and can provide a reliable basis for predictive maintenance.