Abstract: In order to solve the problem of long training time and local minima in BP neural network, a method of neural network initialization based on machine learning is studied in this paper. The experiment results show that the learning efficiency and generalization ability of the initialized neural network are increased and the probability of falling to local minima is decreased.