Neural Network Reentry Guidance for Reusable Launch Vehicle Based on Sample Dimensionality Reduction

Abstract

In this paper, a neural network guidance based on sample dimension reduction is proposed for a reusable launch vehicle. Different from previous works, the training samples of neural networks use Pearson correlation matrix analysis for dimension reduction, which significantly reduces the data volume of training samples and improves the efficiencies of sample construction, data storage, and neural network training. First, the analytical samples are generated by integrating dimensionless energy longitudinal motion equations. Then, the input parameters are processed by Pearson correlation matrix analysis, significantly reducing the data volume of training samples while ensuring the training effectiveness. Furthermore, the neural network is utilized to learn the mapping relationship between reduced-dimension flight states and predictive range-to-go, eliminating numerical integration calculations and improving the real-time performance of guidance. Finally, an extended Kalman filter (EKF) is implemented online for aerodynamic parameter identification, which dramatically enhances the adaptability to internal uncertainties and external disturbances. Compared with the traditional numerical guidance method, simulation results demonstrate that the proposed guidance exhibits superior real-time performance and faster computation speed while ensuring high accuracy and strong robustness to perturbations.