Abstract: The fault features based on dissolved gas analysis (DGA) is the important prior information for transformer fault diagnosis, and the quality of features directly affect the diagnostic effect. At present, there are many fault features based on DGA gas, whereas, the types are relatively single and the diagnostic effect is limited. To realize transformer fault diagnosis with higher accuracy, the cloud feature method is proposed to enrich the existing ratio feature set. To adapt the cloud transformation of high-dimensional cloud feature, the neural network of self-organized cloud concept extraction (SOCCE) is proposed to extract cloud concepts, so as to deeply mine the associated information between multi DGA gases and improve the diagnostic capability of intelligent algorithm. Finally, the optimal DGA hybrid feature set is selected through the feature optimization strategy in which ranking is followed by optimizing. Through the comparative diagnosis based on the IEC TC10 database, it can be seen that the optimal hybrid new features can achieve the fault diagnosis accuracy of 92.4%, which is an improvement of 13.2%~30.8% compared with the diagnosis accuracy of the traditional features. In addition, the new features show the strong ability in field application and multi diagnostic models.