Dynamic-static knowledge graph enhanced LLM reasoning for intelligent process planning in steel manufacturing

Chen, Yuxuan, Rao, Jiangping, Zeng, Yonglong, Zha, Meng, Yan, Wei, Jiang, Zhigang and Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940 2026. Dynamic-static knowledge graph enhanced LLM reasoning for intelligent process planning in steel manufacturing. Journal of Manufacturing Systems 86 , pp. 203-217. 10.1016/j.jmsy.2026.03.007

PDF - Accepted Post-Print Version Available under License Creative Commons Attribution. Download (0B)

Abstract

As a crucial step in steel manufacturing, process planning significantly improves production efficiency and product quality. With the development of knowledge graph (KG) technologies, significant progress has been made in using historical knowledge for intelligent process planning. However, existing KG-based methods mostly rely on static knowledge reuse, which fails to accommodate real-time fluctuations of smelting conditions in steel manufacturing. Moreover, the generation of process planning strategies remains heavily dependent on human expertise, making it time-consuming and labor-intensive. To address these challenges, this paper proposes a dynamic-static KG-enhanced Large Language Model (LLM)-based intelligent process planning method for steel manufacturing. Firstly, a static smelting process KG (SSPKG) is introduced to organize historical smelting knowledge using a convolutional block attention mechanism-enhanced partitioned fusion Kolmogorov-Arnold network (CBAM-PFKAN). Then, a dynamic smelting condition KG (DSCKG) is established to convert the real-time smelting conditions into structured knowledge triples based on predefined rules. Secondly, a plan-and-solve (PS) prompting strategy is designed to guide LLMs in decomposing complex process optimization queries, retrieving both static and dynamic knowledge from the KGs, and generating process schemes step-by-step. Finally, the proposed method is validated through an industry case study. Compared to baseline methods, the proposed approach achieves superior performance in process knowledge extraction, improving the accuracy of generated process strategies by 11.13% on DeepSeek and 8.63% on ChatGPT-3.5, respectively.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Additional Information:	RRS policy applied
Publisher:	Elsevier
ISSN:	0278-6125
Date of First Compliant Deposit:	16 March 2026
Date of Acceptance:	10 March 2026
Last Modified:	16 Mar 2026 14:45
URI:	https://orca.cardiff.ac.uk/id/eprint/185790

Actions (repository staff only)

Edit Item