Deformation-aware positioning optimization in aircraft assembly using surrogate model-assisted deep reinforcement learning

Abstract

Assembly positioning processes play a crucial role in determining the final manufacturing precision of aircraft components. Traditional methods typically treat components as rigid bodies, focusing on adjusting their position and orientation while overlooking the complexities associated with deformable structures. This paper proposes an innovative methodology to optimize the positioning process of aircraft components by incorporating deformation considerations. A two-stage surrogate model, enhanced by machine learning techniques, is introduced to approximate the deformation of structures under various locator configurations. Deep Reinforcement Learning (DRL) is subsequently applied to leverage the surrogate model-based simulation environment. The high-dimensional stress field, compressed by the surrogate model, is used as the state input for the DRL agent, significantly reducing training complexity and enhancing stability. The agent's action corresponds to adjusting the locator's end effector position, while the reward function is designed to minimize the deformation indicator. Upon training, the resulting policy demonstrates strong generalization on the test dataset, achieving a median structural deformation reduction of 99.3 %, with 95 % of the test samples showing a reduction of over 92 %. This approach not only improves the precision but also increases the productivity of aircraft assembly, establishing a new benchmark for intelligent assembly systems that involve deformable components.