Neural Network-Based Descent Control for Landers with Sloshing and Mass Variation: A Cascade and Adaptive PID Strategy

Abstract

Autonomous control of lunar landers is essential for successful space missions, where precision and efficiency are crucial. This study presents a novel control strategy that leverages proportional, integral, and derivative (PID) controllers to manage the altitude, attitude, and position of a lunar lander, considering time-varying mass and sloshing behavior. Additionally, neural network models are developed, to approximate the lander’s mass properties as they change during descent. The challenge lies in the significant mass variations due to fuel, oxidizer, and pressurant consumption, which affect the lander’s inertia and sloshing behavior and complicate control efforts. We have developed a control-oriented model incorporating these mass dynamics, employing multiple PID controllers to linearize the system and enhance control precision. Altitude is maintained by one PID controller, while two others adjust the thrust vector control (TVC) gimbal angles to manage pitch and roll, with a fourth controller governing yaw via a reaction control system (RCS). A cascade PD controller further manages position by feeding commands to the attitude controllers, ensuring the lander reaches its target location. The lander’s TVC mechanism, equipped with a spherical gimbal, provides thrust in the desired direction, with control angles α and β regulated by the PID controllers. To improve the model’s accuracy, we have introduced time delays caused by fluid dynamics and actuator response, modeled via computational fluid dynamics (CFD). Fluid sloshing effects are also simulated as external forces acting on the lander. The neural networks are trained using data derived from computer-aided design (CAD) simulations of the lander vehicle, specifically the inertia tensor and the center of mass (COM) based on the varying mass levels in the tanks. The trained neural networks (NNs) can then use lander tank levels and orientation to inform and accurately predict the lander’s COM and inertia tensor in real time during the mission. The implications of this research are significant for future lunar missions, offering enhanced safety and efficiency in vehicle descent and landing operations. Our approach allows for real-time estimation of the lander’s state and for precise execution of maneuvers, verified through complex numerical simulations of the descent, hover, and landing phases.