The traditional BP algorithm has the inherent shortcomings such as low convergent speed and local minimum. So this paper proposes a simulated annealing algorithm with globalsearch property to optimize the BP neural network for avoiding the local minimum and improving its stability. In order to improve the convergent speed of neural network, the Powell algorithm is proposed to optimize the simulated annealing algorithm. Finally, the simulated results of gearbox fault diagnosis are presented to show that the improved BP neural network has higher convergent speed and precision than those of the traditional BP neural network.