As the key component of high-speed train bogie, the fault characteristics of gearbox are mainly reflected in its vibration signal. The vibration signals collected in the process of gearbox fault diagnosis are usually complex and changeable, and have strong randomness and contingency. A gearbox fault diagnosis method based on multi feature extraction, principal component analysis (PCA) and adaptive genetic algorithm is proposed to optimize the back propagation neural network analysis. The original vibration data in the gearbox fault diagnosis experiment published by Jiangsu qianpeng Diagnostic Engineering Co., Ltd. is extracted by multi eigenvalues. The feature set is reduced by PCA. The selected principal components are diagnosed and analyzed by AGA-BP neural network. The final diagnosis result is that the root mean square error (MSE) of AGA-BP neural network is 0.0116, and the recognition rate of gearbox fault is 100%.