A Framework for Deep Reinforcement Learning-Based Assembly Sequence Planning of Aircraft Tube

Assembly sequence planning is a well-known NP-hard problem that must satisfy complex precedence constraints while minimizing time, cost and error rates. To address these challenges, we propose a three-layer deep reinforcement learning framework for aircraft tube assembly sequence planning. Under this framework, first, a unified data standardization scheme processes both CAD-extracted and topology-generated tube assembly information into fixed-size assembly state, precedence and direction matrices; second, the decision layer employs a Double Deep Q-Network (Double DQN) agent enhanced by prioritized experience replay and the sigmoid learning rate scheduler to generate optimal sequences under multiple constraints; third, the application layer converts the learned sequences into step-by-step instructions and interactive visual simulations. Finally, a case study on a forty-tube assembly is conducted to validate the effectiveness of the entire framework.