In SV10PB1-30B hydraulic control check valve, the sensor was used to collect the vibration signals of 10 valve cores in 3 different leakage modes. A deep convolution model and pattern recognition test were performed with different measuring points(upper surface and seat of check valve) and different signal feature extraction methods(original signal, eigenvalue, feature map). The results showed that the fault type could be effectively identified, and the recognition accuracy rate on the verification set was as high as 88.293% with eigenvalues of the axial impact signal and the deep convolutional neural network. The accuracy was 7.79 times based on the feature map and 1.16 times based on the original time domain impact signal. The optimal number of training steps was 100. The model showed optimal classification effects on normal spool and different damaged spool.