The invention discloses a fault early warning method of a gearbox of a wind turbine generator. The early warning method mainly includes the following steps of data acquisition, wherein a terminal acquires the historical data of wind turbine generator power, environment temperature and wind speed through a hard contact temperature sensor, a soft contact temperature sensor and a temperature acquisition card arranged in the gearbox or calls the historical data of wind turbine generator power, environment temperature and wind speed from wind power plant recorded data, and the time span of the datacovers the whole fault evolution section; data processing, wherein input data is normalized, the integrity of the input data is verified, abnormal data and incomplete data are deleted from the inputdata, a BP neural network including an AIS-SA hybrid network prediction algorithm is established, and then the network size and various initial connection weights and thresholds are determined; earlywarning calculation, wherein the acquired temperature data is predicted through the optimized network, the wind speed is subjected to approximate substitution in combination with numerical weather prediction, a temperature prediction conversion curve in a certain future time is obtained, the residual of predicted temperature and actual temperature is calculated, and with several recording points as an early warning section, the residual mean and standard deviation of the two temperatures in the section are worked out.