Abstract: During Earth-Moon transfer, a spacecraft's orbit insertion errors propagate and amplify rapidly in a strong nonlinear dynamical environment, critically impacting mission success. Traditional mid-course correction methods, often reliant on the "small deviation" assumption and ground-based support, struggle to meet the demands for autonomous, real-time onboard execution. This study develops a lightweight neural network-based method for the rapid estimation of mid-course correction impulses, designed for limited onboard computational resources. The approach begins with high-fidelity shooting simulations and theoretical analysis, which reveal a near-linear relationship between initial velocity errors and the required correction impulses. Subsequently, a lightweight fully-connected network is constructed to establish a direct, end-to-end mapping from nominal orbital parameters and correction time to the corresponding impulse sensitivity coefficients. Validation results demonstrate that the relative error in predicting these coefficients remains below 3%. The proposed method reduces dependencies on traditional assumptions and ground support, offering a viable pathway for autonomous mid-course correction under stringent onboard resource constraints.

Key words: human lunar landing mission, trans-lunar trajectory, mid-correction, deviation propagation, DNN