Large language models empower the reliability of disassembly in remanufacturing

Disassembly is a crucial stage in the product lifecycle, significantly contributing to remanufacturing, waste minimization, and resource recovery by facilitating the reuse and recycling of components and materials. However, engineers often face a plethora of choices during the disassembly process, each linked to uncertain outcomes, leading to substantial challenges. To address this, we introduce an assembly sequence evaluation model offering a quantitative analysis for diverse conditional decisions. Initially, an event graph is devised to outline the hierarchical structure, effectively shrinking the sequence space through the utilization of engineering semantics. Subsequently, a reinforcement learning model is established, with the reward function defined by large language models that use tailored prompts for their sequences. Deep Q-learning is then applied to train the reinforcement learning model, incorporating a selected set of ground truth sequences to highlight the correct one. Finally, a case study on a decelerator disassembly is presented to illustrate the effectiveness of the proposed method.