In view of the high complexity and difficulty in identification of hydraulic signals, a method based on deep belief network was proposed for the diagnosis of leakage state in hydraulic pump. The wavelet transform and HHT were used to extract features form the pressure signals and flow signals. The stacked RBM network was used to optimize the raw feature set, and the advanced fusion features were extracted. Finally, BP was used for prediction. The experimental results show that DBN can be used to effectively extract the intrinsic characteristics of the original feature set, so that the hydraulic signals is better expressed; the DBN’s identification accuracy reaches 98.77%;compared with SSAE and H-ELM classifiers, DBN has better identification ability and stability for leakage state in hydraulic pump.