A Review of Knowledge Base Construction Strategies for LLM-based Intelligent Decision Support System in Material Selection and Process Planning

초록

Material selection and process planning are core decision-making tasks in manufacturing, directly impacting cost, quality, productivity, and sustainability. However, diverse design alternatives, data scarcity, and qualitative requirements limit the effectiveness of existing data-driven approaches and expert systems, creating challenges in generalization and knowledge acquisition. Large language models (LLMs) have emerged as promising tools due to their natural language interaction and complex query-handling capabilities, yet hallucination and outdated knowledge pose significant risks in high-reliability manufacturing environments. To address these issues, this study proposes retrieval-augmented generation (RAG), which integrates reliable external knowledge, as the core architecture for an LLM-based intelligent decision support system (IDSS). The objective is to identify appropriate knowledge base (KB) structures and practical construction strategies for material selection and process design. This study reviews decision-making methodologies, research on material selection and computer-aided process planning (CAPP), and prior KB construction approaches, reframing them through a RAG-centric perspective. The analysis reveals that CAPP involves heterogeneous tasks across feature, operation, and system levels, requiring integrated use of rule-based reasoning, optimization, knowledge inference, and learning-based algorithms. Although knowledge graphs (KGs) effectively structure complex manufacturing knowledge, manual construction remains costly and time-consuming. Recent LLM-based techniques that automatically extract and refine knowledge from unstructured data have emerged as promising solutions to overcome these limitations and improve KG scalability. The combination of LLMs and KGs can alleviate the generalization constraints of traditional data models and the knowledge acquisition bottlenecks of expert systems, forming a foundation for next-generation intelligent manufacturing systems. This study outlines key future research directions: developing an upper-level integrated schema applicable across diverse manufacturing domains, establishing bottom-up methodologies for constructing reliable domain-specific KGs, and creating quantitative evaluation and management frameworks for the LLM–KG–RAG pipeline.

키워드