KG-CoPlan: A knowledge-grounded and H-FSM driven large-small model collaborative framework for generative manufacturing process planning in industry 5.0

The emergence of Industry 5.0 necessitates a human-centric and resilient manufacturing ecosystem, imposing higher standards on the adaptability of Manufacturing Process Planning (MPP). However, traditional methods struggle with flexible production, while general-purpose Large Language Models (LLMs) are prone to factual hallucinations due to insufficient domain knowledge. To address this, this paper proposes KG-CoPlan, a knowledge-grounded and Hierarchical Finite State Machine (H-FSM) driven large-small model collaborative framework. Specifically, the framework constructs the domain-specific Multimodal Process Knowledge Graph (MPKG) as a digital foundation and integrates Gradient Boosting Regressor (GBR)-based small models for deterministic performance prediction, strictly constraining LLM reasoning within engineering feasibility. Furthermore, H-FSM acts as the orchestration core to regulate collaborative decision-making, ensuring structural rigor. Additionally, an automated traceback mechanism is devised to drive closed-loop system corrections by dynamically parsing verification feedback. Experiments on a dataset of 67 complex components demonstrate that KG-CoPlan significantly enhances plan validity. The results achieve a 98.12% Micro-Step Resource Validity (MSRV) score and limit the cost-time deviation to within 5% of expert baselines. By establishing a self-correcting generative framework, this work presents a robust technical paradigm for human-machine collaborative manufacturing in the context of Industry 5.0.