It's important to improve the performance of the weapon by compensating for errors of inertial sensors. Identification of error model is the key of compensating for errors. A multi-layer local recurrent neural networks is adopted to model the errors of inertial sensors. The framework of networks and adaptable dynamic gradient arithmetic are presented in detail. The result of simulation example shows that multi-layer local recurrent neural networks has some advantage for modeling errors of inertial sensors' output: rapid convergence, good performance of tracking and stabilization.