To address the critical challenges of real-time responsiveness and autonomous evolution in highly dynamic flexible production lines, this paper proposes a novel dynamic modeling framework based on a Large Language Model-driven Multi-Agent System (LLM-MAS). Departing from conventional digital twin (DT) paradigms constrained by static configurations, our approach facilitates seamless synchronization and adaptive evolution by decomposing the DT into three interlinked dimensions: knowledge, behavior, and geometry. Leveraging the semantic reasoning of LLMs, an evolving knowledge graph is first synthesized from heterogeneous multi-source data to empower agents with deep situational awareness. Building upon this, a task-driven coordination mechanism dynamically generates and reconstructs behavior models, enabling robust, goal-oriented logic amidst environmental disturbances. This behavioral intelligence further guides the spatio-temporal reconfiguration of geometric models, effectively translating high-level semantic decisions into physical structural and parameter adaptations. Experimental results from a parts defect monitoring and sorting line demonstrate that, compared to conventional methods, the proposed approach reduces changeover and recovery times by 47% and significantly reduces the false positive rate, proving that the LLM-MAS framework provides a robust foundation for the autonomous operation of next-generation smart factories.
Gao et al. (Sun,) studied this question.