Wavelet packet is good at de-noising and analyzing high frequency signal. Moreover the probability neural network can be well used to classify. The wavelet packets were decomposed and used to reconstruct the failure signal of hydraulic pump characteristics, and the node energy in each frequency bang at third layer was extracted and used to group as feature vectors. A probability neural network of the feature vectors was modeled and input as vectors to recognize the failure model of hydraulic pump. Labview and MATLAB were used in integration to program a recognition software system to do failure recognition of hydraulic pump. Experimental results show that the method is good at model recognition of hydraulic pump, and has achieved good effects.