Rule generation is an important technology in data mining. In this paper, based on rough set theory, using the concept of generalized decision attribute function or uncertain class, an inconsistent decision system is transformed to one that is consistent as an initial preprocessing step. On the basis of this, by means of discernibility matrix and decision function, a rule generation algorithm is presented. The algorithm can generate all decision rules directly, instead of computing all the reducts of decision system． Furthermore the obtained rule set using the algorithm keeps all useful information，which is different from that using some other algorithm ．In the end,an example illustrates that the algorithm is reasonable and effective．