Proposed in this paper is a two-level fault diagnosis procedure of hoist. In the first-level diagnosis, fault samples of hoist are collected and designed based on the experience and knowledge of experts and on a fault tree of the brake system, and seven kinds of failure modes of the brake system are automatically classified with SOM network. In the second-level diagnosis, a fault tree of sub-system-hydraulic station is built and the corresponding fault samples are collected and designed. Afterwards, the cause and degree of the hydraulic station fault are determined based on the diagnosis with BP network, BP network state classifier and Elman network. Test results of the hydraulic station fault show that （ 1 ） the structures, the intelligent algorithm trainlm, the inputs and outputs of the three above-mentioned networks all meet the requirements of fault diagnosis and prediction ; （2） Elman network is of the most stable diagnosis performance slightly affected by the number of hidden layer neurons; and （3） BP network state classifier is of the highest test precision while Elman network is of the lowest one.