Deep reinforcement learning-based approach for dynamic disassembly scheduling of end-of-life products with stimuli-activated self-disassembly

Remanufacturing is one of the most critical strategies for end-of-life product management to promote a circular economy; however, it has been seen very limited implementation due to the labor-intensive and time-consuming disassembly processes for component retrieval. The newly emerged 4D printing technology enables the fabrication of stimuli-responsive reconfigurable structures, outlining new ways to achieve non-destructive and simultaneous self-disassembly of components with different geometry. However, large uncertainties and increased process dynamics have also emerged directly pertaining to the real-time scheduling in disassembly lines with self-disassembly workstations, which the existing scheduling methods are not equipped to handle. In this study, a constrained multi-agent deep reinforcement learning approach is proposed to maximize the disassembly profit by dynamically changing the batch mixing ratios of different-sized components in self-disassembly workstations and adapting real-time scheduling to stochastic product quality, changes in operational sequences, and self-disassembly failures. The proposed approach is validated on a disassembly line for hand pulse detectors that contain heat-activated self-disassembly components. Numerical results show that the proposed achieves stable convergence under uncertainties, and the implementation of a dynamic batch mixing scheme in self-disassembly operations yields a substantial improvement in disassembly profit over the scheduling period. In addition, sensitivity analyses are conducted to evaluate the impacts of system uncertainties on the profitability of the disassembly line.