Aiming at the uncertainty and fuzziness of the relationship between fault symptom and fault type in the process of traditional transformer fault diagnosis

we construct a multi-level fault diagnosis model of transformer based on rough set knowledge and multi-kernel learning theory. In this model

five kinds of characteristic gases and 16 kinds of gas ratios are taken as the initial characteristic parameters based on the dissolved gas analysis（DGA） diagnosis standard

and the minimum fault characteristic information set with high importance on the diagnosed fault type is obtained by the attribute of the neighborhood rough set. Based on the deep excavation of the fault information contained in the DGA

a multi-level fault diagnosis model is established

and the binary-class support vector machine is used as a classifier to diagnose the fault with the minimum fault characteristic information set. In addition

in order to avoid the erroneous situation caused by the long tail distribution of the dissolved gas in the oil

the input characteristics are treated by an arctangent transformation. At the same time

all support vector machines use multi-kernel learning to solve the defects of data sensitivity and low robustness in the single kernel support vector machine. The example analysis shows that the accuracy of each diagnostic level under the characteristics can reach more than 88%

and the minimum running time can be 0.337 5 s

which has the obvious advantages of improving the classification accuracy and reducing the running time and the algorithm structure in comparison with traditional characteristics. Secondly

compared with the traditional fault diagnosis method

the multi-level diagnosis can not only extract the fault characteristic information deeper

reduce the complexity of redundant characteristics

but also effectively improve the average accuracy more than 3%

which have higher accuracy and reliability.