An augmented variable neighborhood search for mixed-model two-sided assembly line balancing considering PM scenarios

Abstract

In a real mixed-model two-sided assembly line, preventive maintenance (PM) activities cause capacity waste at available stations and production halts. To mitigate these issues, multiple task assignment schemes with high interchangeability are required, each tailored to one specific scenario. However, the resulting mixed-model two-sided assembly line balancing problem considering PM scenarios (MTALBP-PM) has not been studied. Therefore, a mixed-integer linear programming model is formulated to minimize total cycle time and task adjustment simultaneously. Meanwhile, driven by knowledge and learning, an augmented variable neighborhood search (AVNS) is designed. Concretely, with the guidance of problem-specific knowledge, a decoding mechanism and three objective-oriented neighborhood structures are designed to achieve solutions with better objectives. Using unsupervised learning, an initialization heuristic is mined from tacit information to obtain high-quality initial solutions. With historical search information, a self-adaptive strategy based on Q-learning is proposed to recommend the best-fit neighborhood structure for higher efficiency. Besides, an auto-tuning restart operator based on multi-domain knowledge is employed to escape local optima. Experimental results show that the espoused policy is effective, and AVNS outperforms eight other state-of-the-art meta-heuristics in deriving well-converged and -distributed Pareto fronts. In a statistical sense, the average GD, IGD, and HVR of AVNS reach the best values among all tested meta-heuristics based on 40 benchmark cases, which are 0.4599, 0.8021, and 0.8943, respectively.