Joint Optimization of Time-Aware Condition-Based Maintenance and Repair Resource Management for Gantry Crane Clusters Based on Improved MADDPG

Abstract

Conventional maintenance strategies for port cranes often lack intelligence, flexibility, and global optimization, with insufficient consideration of time awareness. To optimize condition-based maintenance and resource management for crane clusters, this study decouples maintenance decisions for individual cranes from the overall cluster resource management. We formulate a decision-making model, incorporating uncertainties in procurement lead times, costs, equipment downtime, and spare parts shortages. To improve the model-solving process, we present the evolutionary multi-head attention critic with adaptive strategy–multi-agent deep deterministic policy gradient (EMACAS-MADDPG) algorithm, an enhanced version of the multi-agent deep deterministic policy gradient (MADDPG) algorithm. This algorithm initially evolves policy network parameters through a genetic algorithm and subsequently refines them using experience buffer data. Furthermore, a multi-head self-attention mechanism is embedded into the critic network, and an adaptive exploration strategy is utilized during action execution. The implementation of the EMACAS-MADDPG algorithm in the joint optimization model significantly reduces the average maintenance cost by 22.37% compared to the original MADDPG and by 51.73% compared to the Independent Proximal Policy Optimization (IPPO) algorithm.

Keywords

• Gantry crane condition-based maintenance
• repair resource management
• time-aware
• multi-agent reinforcement learning