A neural network system used for failure detection of a liquid rocket engine (LRE) is described and presented in this paper. The system includes two layers: the first processing layer consists of the winner-take-all (WTA) neural network used for LRE fault diagnosis; the second layer consists of the back-propagation (BP) neural network used for displaying the failure value with the output result of the first layer. In the paper，the failure detection simulation of a liquid rocket engine is given to show great advantages for the failure detection system of the neural network.