Abstract: The coastal water-rich sand layer has the characteristics of high porosity, strong permeability and low strength, which makes it difficult to control the axis of shield tunneling. The traditional manual control has the problems of response lag and insufficient accuracy. In order to realize intelligent axis control, a lightweight intelligent prediction method is proposed, and a hybrid teacher model integrating Long Short-Term Memory ( LSTM ), Transformer and Random Forest ( RF ) is constructed to capture the multi-dimensional time-series nonlinear characteristics of shield-stratum dynamic interaction. Furthermore, the predictive ability is transferred to the lightweight student model through knowledge distillation, and FLAC 3D is embedded to realize the dynamic optimization of construction parameters. Based on the measured data of the water-rich sand section of the Jingu Haihe Tunnel, the model is verified. The results show that the mixed teacher model predicts R² > 0.93 ; the lightweight student model improves the computational efficiency by 30 times while maintaining the accuracy, which meets the real-time control requirements of the airborne system. In the interactive simulation with FLAC 3D, the surface subsidence is reduced by 45.6 %, and the axis offset is reduced by 50.5 %.