The multi-classsified probability output method is capable of yielding satisfactory classification results and class probabilities for general transformer fault diagnosis. Based on this method

a new transformer fault diagnosing method is proposed which gives fault probabilities more comprehensive than those of the conventional methods by producing binary outputs. An additional sigmoid function is trained to transfer the traditional support vector machine (SVM) binary outputs 0 and 1 to probabilities of the incipient faults. With the results of binary classifiers

multi-classified probabilities are determined by convex quadratic programming. By providing comprehensive fault probability information

not only can the proposed fault diagnosing method achieve high classification accuracy but also it will give rational results possibly misdiagnosed by other methods. For samples misclassified by traditional methods such as SVM and three-ratio method

the proposed method is able to correctly classify the fault type with the greatest probability

and remind the decision-makers of the real fault type hidden in two or three likely types. This work is supported by National Natural Science Foundation of China (No. 51177143)

National High Technology Research and Development Program of China (863 Program) (No. 2011AA05A120) and Zhejiang Provincial Natural Science Foundation of China (No. LZ12E07002).