A power transformer is an important fault reason for a power system, some uncertain factors such as randomness and funniness exist between fault phenomenon and fault mechanism. There are no definite corresponding relations between fault characteristic quantity and fault omens, which make fault diagnosis more difficult. The transformer fault of artificial neutral network diagnosis is more and more paid attention to, but when the normal artificial neutral network calculations to the data are trained, most of them have slow convergent speed, even no convergent. The paper proposed to improve artificial neutral network calculation, thus increase training speed and diagnostic reliability. It states that artificial intelligence system is very useful tool for transformer early hidden faults achieves the possibilities and accuracy of primary diagnosis.