Explainable Scheduling in Vehicle-as-a-Conveyor Matrix Manufacturing Systems via Deep Reinforcement Learning

Abstract A Matrix Manufacturing System (MMS) is a highly flexible production system designed to adapt to uncertainties in product demand and shop floor operations, with a focus on maximising production efficiency and adaptability. Recently, the emergence of the Vehicle-as-a-Conveyor (VaaC) concept presents an opportunity to fully leverage the flexibility and parallel processing capabilities of MMS. VaaC is a concept in which a vehicle autonomously navigates among workstations and undergoes various processes during production. To ensure efficient operation of an MMS integrated with VaaC (VaaC-MMS), it is crucial to develop an optimisation methodology. This paper proposes a methodology for explainable optimisation via deep reinforcement learning to enhance dynamic scheduling and resource utilisation of the VaaC-MMS. The learning processes and learned policies are interpreted using frequency-map analysis, action-occlusion sensitivity analysis, and SHAP-based feature attribution. The proposed methodology applies Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Asynchronous Advantage Actor-Critic (A3C) algorithms. To validate the proposed methodology, a case study was conducted focusing on the trim part assembly process in the automotive industry. This paper contributes to the realisation of VaaC-MMS and provides a valuable reference for envisioning the factory of the future in the automotive industry.