The fault of a transformer cannot be accurately diagnosed when the fault information is small. An improved artificial bee colony algorithm is proposed to optimize the fault diagnosis method of the support vector machine. First, Principal Component Analysis (PCA) is used to extract the features of the input variables. This reduces the dimension of the feature vector and avoids the overlap of the variable information. Secondly, through two-dimensional uniform based population initialization and an Euclidean distance-based food source update, this paper improves the traditional Artificial Bee Colony (ABC) algorithm, and then tests the performance of the Improved Bee Colony Algorithm (IABC) and ABC and Particle Swarm Optimization (PSO). Search rate and convergence are improved significantly. By using IABC optimization Support Vector Machine (SVM) parameters, the new eigenvalues extracted by PCA are input into IABC-SVM, GA-SVM, PSO-SVM models and the diagnostic results are compared. Finally, the method has high diagnostic accuracy, uses a simple model, and has strong generalization ability. This work is supported by National Natural Science Foundation of China (No. 51974151) and Foundation of Key Laboratory of Liaoning Education Department (No. LJZS003)