Abstract: Based on the fault features of Nuclear Power Plant,the ES（expert system） of fault diagnosis has been programmed.The knowledge in the ES adopts the production systems,which can express the certain and uncertain knowledge.For certain knowledge,the simple reasoning mechanism of prepositional logic is adopted.For the uncertain knowledge,CF（certain factor） is used to express the uncertain,thus to set up the reasoning mechanism.In order to solve the "bottleneck" problem for knowledge acquisition,rough set theory is incorporated into the fault diagnose system and the reduction algorithm based on the discernibility matrix is improved.In the improved algorithm,the measure of attribute importance first calculate the attribute which have the same value in the same decision-sort,then calculate the degrees of attribute in the discernibility ma-trix.Several different faults have been diagnosed on some emulator with this expert system.