A neural network method used to identify the operating states of transformers has been proposed and established. It is superior to the traditional transformer protective relays and can correctly identify, within half cycle from the fault inception, the internal faults, magnetizing inrush current state, external faults and switching on internal faults of a no-load transformer. In addition, this method has broad availability and high faulttolerant ability. A lot of simulations have demonstrated its superiority.