Deep reinforcement learning-based energy-aware disassembly planning for end-of-life products with stimuli-activated self-disassembly

Remanufacturing stands as a cornerstone strategy for end-of-life (EOL) product management, playing a vital role in fostering a circular economy. Despite its significance, the widespread implementation remains difficult, mainly due to challenges such as labor-intensive operations, diminished quality, and time-consuming processes involved in component disassembly. A potential solution emerges in stimuli-activated self-disassembly, offering a non-destructive pathway that encourages seamless human–machine collaboration. This innovative approach facilitates the simultaneous disassembly of multiple components, reducing damage, labor costs, and energy consumption. Notably, limited studies have addressed real-time disassembly planning (DP), especially within self-disassembling workstations. Our research aims to maximize disassembly profit and energy recovery by optimizing disassembly sequences, EOL options, and a hybrid scheme that combines manual and self-disassembly operations. We propose an advanced deep reinforcement learning (DRL) algorithm that incorporates an innovative loss function, a revised training scheme, and parameter embedding to generate the Pareto frontier. Additionally, we propose a compact product representation that captures dynamics and uncertainties, such as product type variations, missing components, potential disassembly failure, and stochastic product quality. The effectiveness of our approach is demonstrated through a case study involving a TV disassembly line, benchmarked against six baselines. Furthermore, a sensitivity analysis is conducted to elucidate the impact of labor expenses and hybrid disassembly schemes on the ultimate profit recovery.