LLM-driven human-AI collaborative decision support system for complex industrial processes: A case study in metallurgy.

Abstract

Multi-source heterogeneous data in complex industrial processes often form data silos, hindering data integration and reducing human decision-makers' cognitive efficiency. We present a human-AI collaborative decision support system for complex industrial processes, driven by Large Language Models (LLMs). The system employs a meta-cognitive reasoning framework to decompose tasks and orchestrate tools. It integrates feedback-based knowledge retrieval, which maintains an evolving domain knowledge base, and a memory-augmented SQL constructor. The latter enables natural language access to time-series and relational data without model fine-tuning. These components form a unified architecture that coordinates technical documents, process data, and operator feedback under real-time and safety constraints. The system provides interpretable high-level recommendations while keeping human decision-makers in control. We deploy the system in a ladle preheating process in a steelmaking plant and construct a 340-query expert-annotated evaluation corpus. With DeepSeek-V3 as the driving LLM, the system achieves 99.1% answer correctness on this corpus and provides reliable and efficient decision support under practical inference-time constraints.