A conjugate gradient back-propagation (CGBP) algorithm of a multilayered neural network is presented with the theory and methods for nonlinear optimization. Since this learning algorithm needn't compute the second derivative，it is worked out easily. Furthermore，the CGBP belongs to the superlinear convergence algorithm so that the performance of the multilayered neural network is improved in comparison with the general back-propagation algorithm. The identification techniques have been widely applied to the field of the liquid rocket engine (LRE). Because the engine model must have been known before the system identified with traditional mathematical method，the identification method of parameter is extremely restrained in the engine. Based on the neural network conjugate gradient algorithm，the experimental data of a variable thrust rocket engine is presented to the neural network，and a satisfied identification model is obtained in this paper.