The back propagation（BP）neural network optimized by modified genetic algorithm(MGA-BP) is proposed to improve accuracy of aero-engine bearing fault diagnosis. The traditional genetic algorithm optimized by using the fixed individual selection probability, trigonometric function and Gauss mutation operation to solve defect problem of genetic algorithm. The BP neural network weight and threshold is optimized by MGA-BP. Four cases of rolling bearing normal, inner ring fault, out ring fault and ball fault are diagnosed by using the optimized BP neural network. The influence of network output mode and diagnostic sample proportion on the accuracy of diagnosis is fully considered. In order to verify the effectiveness of MGA-BP in the bearing fault diagnosis, the BP neural network is optimized by other improve genetic algorithm as a contrast algorithm. The comprehensive comparison results show that MGA-BP can better adapt to different output modes and different sample proportions than other algorithms in this paper. And its noise immunity, diagnosis accuracy, convergence speed and error convergence value are all better than other improved genetic algorithm. In consequence, MGA-BP is more suitable for bearing fault diagnosis.