In order to solve the problem of data mining in diagnosing transformer fault, that data of dissolved gas-in-oil is divided without considering the randomness and fuzziness, Cloud model is applied. The efficiency of mining association rules is also improved through Cloud model; For the assumption in Naive Bayes Classifier is not conformed to the actual situation, An association rule forest and a method of the joint probability calculated is applied to improve Naive Bayes Classifier. The new Bayes Classifier is proved to be practical in the diagnosis of transformer by comparing with other classifier and testing example.