A sequential fusion modelling approach for hypersonic vehicles based on reinforcement learning

During the integrated design phase of hypersonic vehicles, performance analysis necessitates substantial high-fidelity model data to enhance design capabilities. However, acquiring such data proves prohibitively costly. A critical technical challenge lies in reducing the volume of high-fidelity data while maintaining reasonable accuracy metrics, thereby supporting iterative design of the vehicle’s guidance and control systems. To address this, this paper proposes a reinforcement learning-based sequential fusion modelling approach for hypersonic vehicles, aiming to resolve the conflict between model accuracy and modeling costs during the iterative design process. Firstly, a Co-RBF multi-fidelity surrogate model architecture is established by employing a biased RBF kernel to design a roughness penalty criterion. Subsequently, an active exploration strategy based on response surface extrema points is proposed to sample the high-fidelity model data space of the hypersonic vehicle. Furthermore, based on reinforcement learning strategies, we design an adaptive sequential iterative method for model data, which achieves efficient fusion of multi-source model data for hypersonic vehicles while reducing computational costs. Finally, we demonstrate the effectiveness of the proposed methodology through multiple numerical and simulation examples.