Fault Diagnosis of Power Transformers is a kind of pattern recognition. The paper modifies the traditional ART-2 network based on this idea: different distances for samples to center of cluster produce different influence on excursion of center of cluster and the lack of pattern drifting is improved. First the input vector of sample is extended, then it is diagnosed by the modified ART2-network .The results are compared with those obtained by the IEC three-ratio and BPNN method. Comparison result verify that the algorithm proposed is effective and feasibile.