The relationships between fault symptoms and fault reason of oil-immersed electrical power transformer are very complicated. One single fault diagnostic method cannot always achieve the true result due to its particular information used for diagnostic and specific approaches used. Therefore， a new algorithm，which synthesizes several conventional transformer fault diagnosis methods， is put forward to eliminate the influence of a single method's preference on the final diagnosis result. Fuzzy C-means clustering（FCM） algorithm， BP neural network， rough set theory diagnosis algorithm and IEC 60599 recommended three-ratio method are adopted as monomial methods to diagnose fault independently in the new diagnosis model. The optimal weights of every single method are calculated when the total diagnosis error squared is minimal. Then the optimal voting combination model is set up and the fault type with the largest fault probability is obtained. The results of diagnosis instances show that the new method can not only decreases misjudgment risk and get much higher accuracy rate， but also enhance diagnostic robustness compared with single conventional method.