Aiming at the problems of complex process in traditional fault feature extraction, single diagnosis and poor accuracy, a bearing fault identification scheme based on multithreshold wavelet packet and deep belief network (DBN) was proposed. The optimal wavelet basis function and softhard threshold combination method were used to perform threelayer decomposition and noise reduction on the original vibration signal, and eight signal components from low frequency to high frequency band were obtained. The combined reconstruction of 8 frequency bands was used as the input sample of the neural network. Based on the advantages of DBN in data processing, the bearing fault identification model of DBNBP neural network was established and various parameters of the model were determined. The effects of different sample inputs on the identification rate of the model was explored and compared with the traditional neural network model. The results show that the trained DBNBP bearing fault identification model can accurately predict the bearing fault signal from the original data and wavelet packet decomposition signals. Feature learning and classification, combined with recognition rate and diagnostic time considerations, have better diagnostic efficiency through wavelet packet decomposition signal input.