Hydraulic system is a typical highly nonlinear system, and the failure modes and fault mechanisms are complicated and varied because of the interference among the various loops of the system. The power transmission in the system is closed, the parameters are not measurable, and the fault information is difficult to extract, which make it difficult to diagnose the fault of the hydraulic system. Especially in some practical applications which lack system model, expert knowledge and prior probability, the traditional fault tree-Bayesian network diagnosis method can not be effectively applied. In view of this application scenario, Bayesian network learning based on simulation data mining was proposed, and BNFinder software was used to deal with the simulation data. The structure of the Bayesian network is optimized and the efficiency of fault diagnosis is improved.