A fault diagnosis method was presented for motor rotor broken bar fault based on wavelet package analysis (WPS) and Elman neural network. The sideband frequency current, which reflects the broken bar fault, was analyzed with the technology of wavelet package decomposition. The frequency segment power under operating states was abstracted as fault characteristic vectors, which weaken the influences of variable load and noise. Elman neural network was adopted to identify the broken bar fault. To improve its performance, Elman neural network was modified by adding a self-feedback gain factor in the context nodes. Energy of various frequency bands acting as the fault characteristic vector was input into the modified Elman neural network to realize the mapping between the feature vector and the fault mode. Experiment results show that the fault characteristic vectors abstracted by WPA are evident. The diagnosis system based on wavelet package and Elman neural network could identify motor rotor broken bar fault efficiently and accurately.