Facility layout optimisation of an intelligent manufacturing unit based on queueing network and deep reinforcement learning
针对需求不确定下制造单元设施布局问题，提出一种结合Transformer自注意力机制和指针网络的深度强化学习方法，利用排队网络快速评估性能，在大型实例中平均改进6.89%，并具备跨场景泛化能力。
Facility layout critically affects manufacturing efficiency and cost. Fixed layouts struggle to adapt to uncertain demand, constraining system performance. In facility layout problems (FLPs) under demand uncertainty, traditional metaheuristics suffer from limited responsiveness to production changes. Therefore, we propose a deep reinforcement learning (DRL)-based method capable of handling high-dimensional inputs and exhibiting cross-scenario generalisation. Specifically, we model the FLP as a Markov Decision Process (MDP) and leverage Transformer-based self-attention mechanism combined with Pointer Network to enable adaptive decision-making. Considering inherent stochasticity in manufacturing systems, the system is modelled as an open queueing network with finite buffers using the Generalised Expansion Method (GEM). Compared with analytical or simulation models, the queueing network model enables rapid estimation of real-world production performance, providing efficient reward feedback for DRL training. Numerical experiments validate that the proposed method achieves the best performance among baselines, with an average improvement of 6.89% in large-scale instance. Generalisation tests show that integration with the 2-opt heuristic enables swift generation of high-quality layouts in unseen scenarios. A real-world case study further validates the framework's applicability, offering a practical solution for dynamic layout reconfiguration in intelligent manufacturing, and significantly enhancing production flexibility and resource utilisation under demand uncertainty.