Based on the background of rolling bearing fault diagnosis,taking the JZQ250 type transfer-box as test object,the shuffled frog leaping algorithm（ SFLA） is combined with back propagation（ BP） neural network,by using the efficient computing performance and the excellent ability of global optimization of shuffled frog leaping algorithm,the network structure of BP neural network is optimized. Through comparison,it is found that the BP neural network model optimized by shuffled frog leaping algorithm can avoid making it fall into local optimum,reduce the training time and improve the training accuracy during the training of the network,and have several advantages,such as relatively higher convergence rate and ability to accurately diagnose.Through a series of training and testing,the results show that this approach can improve the accuracy and reliability of fault diagnosis.