Abstract
Disassembly sequence planning (DSP) is critical for increasing efficiency and reducing environmental hazards of end-of-life (EOL) products. As an NP-hard combinatorial optimization problem, DSP becomes increasingly difficult to solve as product complexity grows. This paper proposes a Modified Grey Wolf Optimizer (MGWO) to address these challenges. By introducing a Feasible Solution Generator (FSG), a Neighborhood Search Operator (NSO), and a Guided Search Operator (GSO), the MGWO ensures solution feasibility and improves search efficiency. Validated against engineering benchmarks and real-world industrial cases, the MGWO demonstrates superior performance compared to traditional meta-heuristic algorithms.
Introduction: The Challenge of End-of-Life Products
In modern manufacturing, the efficient disassembly of products at their end-of-life (EOL) is essential for recycling, remanufacturing, and waste reduction. However, determining the optimal order in which to remove components—Disassembly Sequence Planning (DSP)—is a complex "puzzle" where the number of possible solutions grows exponentially with the number of parts.
The goal of DSP is to minimize disassembly time, tool changes, and material mixing while maintaining structural stability. Because this is an NP-hard problem, traditional exact algorithms struggle with complex assemblies. While various "intelligence algorithms" have been applied, they often require extensive parameter tuning and can struggle with the strict physical constraints of mechanical assemblies.
Methodology: The Modified Grey Wolf Optimizer (MGWO)
The researchers developed the MGWO by adapting the standard Grey Wolf Optimizer, which mimics the social hierarchy and hunting patterns of wolves, for the discrete constraints of mechanical disassembly.
1. Ensuring Physical Feasibility
A primary challenge in DSP is "precedence constraints"—certain parts cannot be removed until others are out of the way. To handle this, the authors implemented a Feasible Solution Generator (FSG). This operator utilizes a decimal interference matrix to represent spatial constraints between components.
The FSG algorithm uses this matrix to identify which components can be removed at any given step without interference, ensuring that every sequence generated is physically executable.
2. Enhancing Search Capabilities
Beyond feasibility, the MGWO introduces two additional operators to optimize the sequence:
- Neighborhood Search Operator (NSO): This allows the top-performing "leader wolves" ($\alpha, \beta$, and a newly defined $\delta$) to self-evolve by exploring local variations in their sequences.
- Guided Search Operator (GSO): This manages the global search by using self-adaptive factors to determine how the rest of the "pack" follows the leaders, balancing exploration of new sequences with exploitation of known good ones.
The algorithm evaluates sequences using a multi-indicator fitness function that accounts for disassembly direction changes, tool changes, material grouping, and structural stability.
Evaluation and Results
The MGWO was tested against three benchmarks: a 11-component Vise, a 17-component Ball Valve, and a real-world 17-component Azimuth Thruster Propeller.
The results were compared against five established algorithms: Genetic Algorithm (GA), Ant Colony Optimization (ACO), Simplified Swarm Optimization (SSO), Teaching-Learning-Based Optimization (TLBO), and Improved Harmony Search (IHS).
- Reliability: In the Vise case, MGWO found the optimal solution with a 63.33% success rate, significantly higher than GA (13.33%) and ACO (16.67%).
- Convergence Speed: MGWO reached optimal fitness values in fewer iterations across all test cases.
- Industrial Application: For the industrial propeller, MGWO achieved the best mean fitness (5.5669) and the highest probability of hitting the absolute optimum (76.67%), proving its robustness for high-value engineering recovery.
Conclusion and Future Directions
The MGWO offers a powerful, self-adaptive solution for the DSP problem, outperforming traditional methods in both reliability and efficiency. However, the authors note that the current model primarily focuses on spatial interference. Future work will aim to integrate more nuanced real-world constraints, such as varying labor costs and specific environmental hazards, to further enhance the practical utility of the optimizer in industrial settings.
Summary updated