A multi-agent control architecture for flexible manufacturing system

Abstract

As Flexible Manufacturing System (FMS) transitions from human-oriented management to automation processes, the modeling strategy of FMS evolves. The multi-agent system (MAS) concept is widely used in modern factories and is expected to reduce computational complexity and conduct fault-tolerant control. This dissertation begins by reviewing modeling strategies from past decades, focusing on flexibilities, deadlock avoidance, and optimization methods. The review also highlights recent developments including Industry 4.0, multi-agent systems, and multi-agent reinforcement learning. Subsequently, a multi-agent-based FMS structure is introduced. Automata-based deadlock avoidance strategy is integrated into a top-down system structure, making it highly decoupled. A straightforward action-prediction bidding algorithm is developed to minimize information share, alleviating the communication bandwidth burden. A novel dynamically generated agent neighborhood structure is designed to eliminate redundant information between agents, reducing computational time in highly customized systems. Various simulation environments are discussed before implementing the proposed system with Python. Following the implementation, tests are conducted to verify the functionality of the proposed framework. The performance of the proposed neighborhood structure is also compared in detail. Finally, conclusions are drawn in three folds, limitations are listed and recommendations for future work across multiple aspects are discussed.