Pipe wall thinning assessment is essential for ensuring the long-term operational safety of nuclear power plants, which requires accurate acquisition of the actual chemical composition of piping materials. However, the Quality Verification Documents (...
Pipe wall thinning assessment is essential for ensuring the long-term operational safety of nuclear power plants, which requires accurate acquisition of the actual chemical composition of piping materials. However, the Quality Verification Documents (QVD) that contain this information exist primarily as unstructured scanned images, and their visual layout and structural patterns are highly inconsistent, making systematic database construction fundamentally difficult.
To address these challenges, this study proposes an intelligent system that automatically extracts and links piping chemical composition data by integrating deep learning-based document analysis with the stepwise reasoning capabilities of a Multi-Modal Large Language Model (MM-LLM). Although conventional approaches that combine object detection and Optical Character Recognition (OCR) are effective for structured documents, they exhibit inherent limitations in reconstructing complex logical relationships within unstructured documents. To overcome this, we designed a hierarchical inference methodology for the MM-LLM based on a Divide-and-Conquer strategy and Conditional Prompting.
Instead of interpreting a full document at once, the proposed system first analyzes small, high-confidence regions containing chemical composition data, then progressively expands its reasoning scope to larger table regions and full-page images containing identifying information such as Heat No. At each stage, tailored conditional prompts guide the MM-LLM to infer precise semantic relationships between identifying information and corresponding chemical composition values by leveraging visual cues. Additionally, recognizing the practical difficulty of complete automation for industrial documents, a user correction interface was integrated to establish a realistic human-in-the-loop workflow that ensures data integrity.
Experiments conducted using actual QVD from nuclear power plants demonstrated that the proposed divide-and-conquer-based reasoning strategy significantly improved the stability of chemical composition table recognition and the accuracy of data linkage compared with conventional single-image processing. The system effectively restored the underlying logical structure even in irregular document layouts, substantially reducing manual workload and improving operational efficiency.
Overall, this study presents an intelligent system that maximizes the inferential capabilities of MM-LLM for unstructured industrial documents. The proposed approach provides a technological foundation for automatically constructing key data required for pipe wall thinning prediction and is expected to contribute to digital transformation and the advancement of predictive maintenance strategies in the nuclear sector.