Combining Graph Neural Networks and Deep Reinforcement Learning for Aircraft Maintenance Stand Scheduling

This study delves into the complex aircraft maintenance stand scheduling problem with the primary aim of enhancing safety during maintenance operations. We first establish a Integrated heterogeneous graph model that effectively integrates both maintenance stand and maintenance item nodes, offering high information density and adaptability to dynamic conditions. Building on this foundational model, we develop a maintenance stand flexible scheduling deep network that leverages the power of graph neural networks. This innovative network performs targeted embeddings for both maintenance stands and maintenance items, enabling precise feature extraction from the heterogeneous graph and facilitating better representation of their intricate relationships.Decision-making within this framework is achieved through advanced deep reinforcement learning techniques, specifically designed to focus on minimizing aircraft movements while ensuring the safest and most efficient scheduling solutions. Experimental results on real-world data show that our method outperforms Priority Dispatching Rules (PDRs) in all problem scales, with at least a 0.62% lower Gap value. Notably, in the 20x10 instance, it surpasses OR-Tools with a 0.47% Gap. Additionally, it exhibits rapid convergence speed and excellent generalization performance.