Based on rough set and fuzzy set and Bayesian optimal classifier, a novel transformer fault diagnosis and maintenance method is proposed in the paper. The method firstly applies fuzzy subjection degree function of the observed information to establish posterior probability of original assumption in Bayesian optimal classifier, the classified results based on each fault information then are calculated, the best diagnosis result is acquired after all these results are weighted average. Then based on rough model of Bayesian risk decision, the diagnosis results of all faults information are identified to constitute possible maintenance strategies. Actual application shows that the proposed method can deal with the "bottle neck" of fuzzy knowledge acquisition in Bayesian optimal classifier and possesses stronger self-learning abilities, and is an effective transformer fault diagnosis and maintenance method..