Abstract: Feature subset Selection is of prime important for effective fault diagnosis. But the classification boundary of real fault diagnosis data sets is often ambiguous, and the relationships between faults and symptoms are always uncertain. Rough set theory is a novel mathematical tool dealing vagueness and uncertainty. This paper introduces rough set theory and proposes a method for fault diagnosis feature subset selection. By two fault diagnosis examples, this paper validates the method. The results show that this method can efficiently extract the main fault features while the fault classification result is invariable. The research in this paper supplies a basis for further study of applying rough set theory in fault diagnosis.