A new approach for selecting threshold of data discretization is brought forward, aimed at solving difficult to establish state recognition decision table in less fault samples case. Boundary values of parameter variation interval in normal working condition were used as discrete threshold. Decision table of fault diagnosis was built, and was used to solve the problem of fault diagnosis of car dumper hydraulic system. Attributes of the built decision table were reduced based on rough set theory. The simplified rules of system fault diagnosis were obtained. These rules were stored in knowledge base of fault diagnosis expert system for car dumper hydraulic system. On-site diagnosis result of expert system proves the fault diagnosis rules extracted from application of this method have high robustness.