Incremental learning-based dynamic multiobjective optimization algorithm for disassembly line balancing

Disassembly line balancing is a key operational problem in remanufacturing and circular manufacturing systems. In practical disassembly lines, task processing times are not stationary because component corrosion, wear and deformation, together with robot performance degradation, continuously change operation durations and resource availability. These physical-state variations make the Pareto-optimal solution set and Pareto front migrate across environments. Existing dynamic multiobjective evolutionary algorithms commonly reuse raw historical populations, rebuild archives or reconstruct learning models after environmental changes, which may cause redundant knowledge accumulation, negative transfer, high repeated learning cost and catastrophic forgetting. To address these issues, this paper proposes an incremental learning-based dynamic multiobjective optimization algorithm, termed IL-DMOEA, for physical-state-driven dynamic disassembly line balancing. The algorithm integrates physical-state modeling, density-based historical knowledge clustering, environment-similarity matching, prototype memory construction, knowledge incremental update, memory retention, prototype-driven initialization and evolutionary search into one unified method chain. Historical nondominated solutions are first organized into transferable cluster-level knowledge, and this knowledge is then compressed into compact prototype memory units that can be locally updated when new environments arrive. Computational experiments on 12 dynamic disassembly instances under high-, medium- and low-similarity environments show that IL-DMOEA obtains competitive or superior mean inverted generational distance, mean hypervolume and average running time results compared with representative dynamic multiobjective and disassembly line balancing algorithms. The results indicate that prototype-memory-based incremental learning can improve both solution quality and response efficiency in long-term dynamic disassembly optimization.