Dynamic Balancing of U-Shaped Robotic Disassembly Lines Using an Effective Deep Reinforcement Learning Approach

Abstract

Disassembly line balancing (DLB) is used for efficient task planning of large-scale end-of-life products, which is a key issue to realize resource recycling and reuse. Robot disassembly and U-shaped station layout can effectively improve disassembly efficiency. To accurately characterize the problem, a mixed-integer linear programming model of U-shaped robotic DLB is proposed. The aim is to minimize the cycle time to shorten the offline time of the product. Since there are many dynamic disturbances in the actual disassembly line, and traditional optimization methods are suitable for dealing with static problems, this article develops a deep reinforcement learning approach based on problem characteristics, namely deep Q network (DQN), to achieve a dynamic balancing of disassembly lines. Eight state features and ten heuristic action rules are designed in the proposed DQN to describe the disassembly environment completely. The effectiveness and superiority of the proposed DQN are verified by numerical experiments. In the case of a laptop disassembly line, not only the cycle time of the robots is reduced, but also intelligent decision-making and dynamic planning of disassembly tasks are realized.