Internal leakage which may be caused by seal damage is concerned for leading to unstable operation of the hydraulic systems. To study this issue, a convolution neural network based method to detect the internal leakage fault is proposed. First through simulation, internal pressure signal of a hydraulic cylinder was obtained under the working conditions of four kinds of leakages as no, small, medium and large, after studying and training, then the neural network was used to detect the different internal leakage degrees automatically under uncertain working conditions by input of the signal. Comparing the traditional method based on modelling, this method proposed overcome the drawback of the difficulty in modelling of the nonlinear hydraulic system, sampling of pressure signals was needed only, and it was feasible and simple, with very promising reliability. At last, the traditional back propagation (BP) neural network is compared, which shows the excellence of the convolution neural network.