Enabling knowledge representation and management for Smart manufacturing using KGs and LLMs

Wan, Yuwei 2024. Enabling knowledge representation and management for Smart manufacturing using KGs and LLMs. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (4MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (1MB)

Abstract

Smart manufacturing (SM) represents a transformative shift and offers a strategic way to achieve digitalisation, automation and intelligence. Despite these benefits, realising the full potential of SM poses several challenges. The implementation of advanced technologies under SM has generated and accumulated vast amounts of data. Embedded within these datasets is a wealth of knowledge that needs to be discovered, properly represented, and effectively managed. However, several challenges impede this process, such as integrating heterogeneous and multi-source data, the lack of both human-understandable and machine-queryable formats, the absence of a unified knowledge representation model, and hard-to-represent rich-semantic information etc. As the technologies advance, knowledge graphs (KGs) and large language models (LLMs) have demonstrated remarkable potential in representing and managing domain knowledge for SM. KGs and LLMs naturally compensate for each other. To be specific, KGs provide the semantic foundation and confidence, while LLMs contribute adaptability and usability. KGs have some advantages, such as structural and accurate knowledge, interpretability and transparency, and an evolving domain knowledge base. Also, KGs have their own limitations, such as incompleteness, lacking language understanding, and hard-to-handle unseen facts. These limitations can be compensated by LLMs’ advantages, such as broad general knowledge, powerful language processing, and generalisability. On the other hand, KGs’ advantages can make up for LLMs’ inherent limitations, such as implicit knowledge representation, hallucinations, black-box nature, and limited up-to-date or domain-specific knowledge. In this case, by combining complementary strengths of KGs and LLMs, this thesis investigates a promising framework for knowledge representation and management in SM can be achieved. Firstly, a constructing and completing manufacturing KGs approach is presented. This involves designing ontologies to serve as the backbone of KGs and introducing a iv Abstract ‘know-how’ model that unifies node and link prediction for KG completion. Incorporating numeric data and correlations improves the efficiency and scalability of the ontology design process by reducing human cognitive load. A ‘know-how’ model is further proposed to address the key issues of information sharing and data collaboration in the processes of KG completion. To validate the effectiveness of the proposed KG construction and completion techniques, a real-world case study is conducted on cold rolling processes in the steel industry. Secondly, the thesis introduces a methodology termed text-to-Cypher with semantic schema (T2CSS) to facilitate the transformation of user inquiries with natural language into structured query statements for graph databases. By integrating domain-specific semantic schemas with LLMs, T2CSS enables precise and accurate query generation, enhancing the accessibility and utility of KGs. Experiments conducted using data from a cold rolling plant show that LLMs guided by T2CSS outperform baseline models in generating logically coherent and executable query statements. Finally, a hybrid KG-Vector retrieval-augmented generation (RAG) approach is proposed to leverage the deep reasoning capabilities of LLMs and the structured domain knowledge represented in manufacturing KGs. By integrating KGs with vector-based retrieval methods and introducing semantic alignment mechanisms, the hybrid RAG framework addresses the limitations of existing RAG implementations in handling domain-specific queries. A case study in design for additive manufacturing (DfAM) illustrates the improved accuracy and relevance of generated responses in domain question-answering (Q&A) tasks. In conclusion, this thesis contributes to advancing knowledge representation and management in SM by developing methods that integrate KGs and LLMs. The proposed approaches enhance the accuracy, relevance, and contextual understanding in decision-making processes and provide timely and relevant solutions to critical challenges faced by SM nowadays.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	knowledge representation; knowledge management; smart manufacturing; knowledge graph; large language models
Date of First Compliant Deposit:	14 May 2025
Last Modified:	14 May 2025 15:35
URI:	https://orca.cardiff.ac.uk/id/eprint/178287

Actions (repository staff only)

Edit Item