An Optimal Disassembly Sequence Planning for Complex Products using Enhanced Deep Reinforcement Learning Framework

Disassembly Sequence Planning (DSP) is a crucial problem in the field of repair and maintenance. There is a pressing need for an efficient technique to solve the Complete Disassembly Sequence Planning (CDSP) problem for large, highly complex products without compromising time and computational resources. Since exact methods fail to handle complex products, and meta-heuristic approaches often do not produce optimal results, its solution requires a deep reinforcement learning approach. This work proposes a novel Enhanced Deep Reinforcement Learning (EDRL) approach in which the Actor-Critic Networks with an attention mechanism is employed in both networks to assign weightage to important actions, generating optimal sequences based on these actions. To further improve the model and accurately predict the loss value, a hybrid loss function is developed by combining the Categorical Cross-entropy and Log-Cosh loss functions. This work considers DSP attributes like stability, liaison, geometric feasibility, and precedence for experiments on various products. The proposed EDRL-CDSP method outperforms existing techniques in terms of optimality while requiring less time to generate the solution. The optimized disassembly sequence, when followed, consumes less time for the total disassembly of all the parts of the product. These findings suggest that the proposed EDRL-CDSP approach can effectively address the challenges in DSP and provide practical benefits for industrial applications.