The stability loop of the Satcom On-the-Move (SOTM) servo system is strongly nonlinear due to the influence of backlash friction and other factors.Although the standard Back-Propagation (BP) algorithm has good identification effect for nonlinear systems, it has some shortcomings such as slow network convergence, process oscillation and poor generalization ability.To make up for the shortcomings, an improved adaptive BP algorithm based on the cumulative error function gradient and the double learning step is proposed to accelerate convergence and reduce oscillation.The global error threshold is also set to control training times in order to further enhance its generalization ability.Experiments are designed on the SOTM platform, and it is verified that the improved algorithm has obvious improvement in convergence, identification accuracy, the generalization ability and so on.A more accurate BP network model with nonlinear stability loop can be obtained.