AI-driven integrated online layout planning and scheduling in flexible assembly systems

Manufacturing systems are undergoing systematic changes, balancing the trade-off between customer needs and economic and ecological pressures. Assembly systems, in particular, must become more flexible to adapt to resulting multiple product generations and unpredictable fluctuations in materials and demand. Therefore, flexible assembly systems employ individual job routes and layout reconfigurations during operation through movable multi purpose resources and flexible transportation systems. Simulation studies indicate that such systems enhance scalability and responsiveness to internal and external disturbances, significantly improving employee utilization, reducing labor costs, and optimizing makespan and throughput under dynamic conditions. However, effective control of these flexible systems remains an open research question. Critical tasks include planning individual job routes and dynamic layout changes for mobile stations. Traditional methods, which treat these planning problems offline and separately, often lead to inefficiencies. Additionally, these control tasks demand exponentially increasing computational capacities as problem sizes grow. Emerging machine learning techniques, especially reinforcement learning, can address these challenges. Moreover, multi-agent reinforcement learning can enable multiple agents to learn and adapt in dynamic environments, offering a potential solution for integrating both optimization problems. Consequently, this dissertation introduces a framework for integrated online layout planning and scheduling in flexible assembly systems with mobile resources, emphasizing a multi-agent reinforcement learning algorithm. The framework comprises digital twin modeling, training and hyperparameter optimization, ramp-up, and operation. Several software modules interact within these phases: A GUI and digital twin model facilitate communication with users and subsequent systems. A discrete event simulation and an automatic scenario generator enable the training and evaluation of different scenarios. The algorithm comprises a graph neural network encoder and two decision agents, allowing for various-sized system state descriptions. The thesis details training these modules with real-world data and successful verification and validation using established methods. Sensitivity analyses confirmed plausible algorithm behavior. However, performance improvements varied depending on system characteristics compared to a random agent and a mixed-integer mathematical model.