Abstract: Aiming at the problem that the roll angle of a trajectory correction projectile is in a random state after launch, and the Kalman filter is difficult to converge when the misalignment angle of the strapdown inertial navigation system (SINS) is too large, an improved method for rapid estimation of the initial roll angle in the GPS denied environment is proposed based on the neural network. A small number of radio beacons are set nearby the gun muzzle, and a backpropagation (BP) neural network is established to fit the nonlinear mapping model between the initial roll angle and the observations. Regarding the problem of weak attitude observability assisted by beacon, strapdown inertial navigation measurement parameters are introduced as input neurons to improve the estimation accuracy. The principal component analysis method is used for feature extraction to simplify the network structure. The simulation results show that, compared with the alignment method based on nonlinear Kalman filter, the proposed algorithm can achieve rapid coarse alignment at any roll angle. Simulation is also carried out on the scenarios where the firing angle error and initial pitch angle error are not in the training range and there are layout errors. The accuracy is higher and the robustness is better compared with the unoptimized BP network.

Key words: Strapdown inertial navigation, Initial alignment, BP neural network, Radio beacon