Dynamic aircraft maintenance stand scheduling via graph attention network and deep reinforcement learning

Abstract

Dynamic aircraft maintenance stand scheduling (DAMSS) aims to optimize aircraft stand allocation by jointly minimizing unnecessary aircraft movements, balancing workload distribution across stands, and maximizing stand utilization, thereby enhancing both operational efficiency and safety during maintenance operations. The traditional methods often struggle to adapt to the dynamic and complex nature of maintenance tasks. This paper presents a novel approach that integrates a heterogeneous interactive graph neural network (HIGN) for dynamic task completion and stand embedding generation. The method uses a three-layer interaction structure: a bottom layer based on a graph attention network (GAT) for neighborhood interactions, a middle layer leveraging transformer structures to model long-range dependencies, and a global aggregation layer using average pooling. These embeddings are then fed into a dual-policy proximal policy optimization (DP-PPO) framework, where two actors share a critic to jointly optimize the scheduling policy. The reward function is designed to minimize aircraft movements, improve stand utilization, and balance workload by encouraging consecutive operation assignments to the same stand. The proposed method achieves superior overall performance compared to PDR, metaheuristic, and RL baselines on realistic simulation datasets.