For accurate estimation of the main types of transformer faults with relatively fewer fault information samples, this paper presents an approach to transformer fault diagnosis based on the probability output of the least squares support vector machine(LSSVM)and DS evidence theory according to the ideas of intelligence complementarity and information fusion. This diagnosis method has the following features: it integrates multiple feature information of the operating state of the power transformer, outputs the probabilities of various transformer faults, and provides more available information for the maintenance and repair of the power transformer. This gives full play to the strong generalization ability of the LSSVM in the case of small samples. In case studies, the diagnosis accuracy of the proposed method reached 91. 1%, which was higher than that of the three-ratio method(with an accuracy of 75. 6%)and that of the LSSVM method(with an accuracy of 82. 2%). The proposed method effectively reduces the risk of misdiagnosis of transformer faults.