In order to improve the accuracy of fault diagnosis of power transformer and the inherent two classification attributes of a single relevance vector machine kernel function and the problem of weak robustness to prediction classification, a new model for fault diagnosis of transformer based on fuzzy C means clustering and improved relevance vector machine is proposed. First, the sample is clustered by fuzzy C means, and then the sample is divided by complete binary tree structure of relevance vector machine. The kernel function of relevance vector machine is the mixed kernel function constructed by combining Gauss kernel function and Polynomial kernel function, and the parameters of the mixed kernel function are optimized by using the Double subgroups FOA with the characteristicsof Levy flight. The experimental results show that, compared with single kernel function and particle swarm optimization for optimizing parameters of mixed kernel function, the proposed method has high accuracy, good stability and fast classification speed, and meets real-time online fault diagnosis.