Joint scheduling of distributed mixed-line high-speed train bogie maintenance: a multi-agent hierarchical deep reinforcement learning approach

Abstract

Distributed mixed-line reassembly has emerged as a vital paradigm for high-speed train bogie maintenance. However, spatio-temporal heterogeneity complicates the joint scheduling of line assignments, operations and heavy component transportation. Because reassembly and transport constraints are tightly coupled, traditional methods struggle to balance turnaround efficiency with resource utilization. To overcome this, a multi-agent hierarchical deep reinforcement learning framework is proposed for joint scheduling. To balance makespan and critical equipment workload dynamically, the decision-making process is decomposed spatio-temporally. Spatially, an upper layer leverages a Graph Attention Network (GAT) to model the physical depot layout and dynamic transportation constraints. Temporally, a lower layer employs a transformer to capture long-range sequential dependencies among complex maintenance operations. Additionally, a Pareto-front-based dynamic weighting mechanism achieves adaptive trade-offs between conflicting objectives. Extensive experiments demonstrate that the proposed method achieves superior joint scheduling performance and exhibits strong zero-shot generalization across varying problem scales.