Abstract: Support vector machine (SVM) is an intelligent method which can create diagnostic models automatically by using off/on-line data sets, but it needs to be extended to a multi-class classifier for multi-fault diagnosis. Decision directed acyclic graph (DDAG) is an extending strategy with outstanding performance. However, its decision largely depends on the sequence of nodes which is arbitrarily selected. This affects the accuracy of diagnosis. In this paper, we proposed a method to refine the sequence of nodes according to the distribution of fault data sets, so as to improve the accuracy of SVM-based diagnosis. Multi-class SVM extended by our method has been employed as a transformer fault diagnosis, and satisfactory results have been obtained.