Knowledge Structure and Evolution of Industrial Heritage Digitalization: A Comparative Bibliometric Analysis Using CiteSpace and AI-Assisted Methods

Industrial heritage digitalization increasingly connects documentation, conservation, adaptive reuse, regeneration, and heritage management, yet its interdisciplinary terminology complicates systematic characterization of the field. This study examines the knowledge structure and evolution of industrial heritage digitalization by applying CiteSpace-based and artificial intelligence (AI)-assisted bibliometric analyses to the same literature samples. A total of 431 publications from the China National Knowledge Infrastructure (CNKI; n = 260) and Web of Science Core Collection (WoSCC; n = 171) were analyzed separately. The two approaches examined corresponding dimensions of keyword prominence and relationships, thematic structure, and temporal evolution, with their outputs compared in terms of substantive correspondence rather than numerical equivalence. The findings indicate that industrial heritage digitalization has expanded from documentation and geometric representation toward the wider use of digital information in conservation, adaptive reuse, regeneration, and management. The two approaches showed broad correspondence in keyword prominence, major network nodes, major thematic domains, and overall temporal evolution, whereas specific network relationships, thematic organization, and temporal signals were only partially consistent. Such correspondence indicates convergence across analytical procedures rather than validation against an external ground truth. Differences between the two representations were associated with terminology processing, network construction, clustering, and temporal indicators rather than semantic processing alone. Overall, the findings indicate increasing integration of digital information across heritage functions and spatial scales and emphasize the importance of information continuity across conservation and adaptive-reuse processes. Methodologically, researcher-supervised large language model (LLM)-assisted semantic processing can complement established science-mapping approaches when combined with transparent and reproducible analytical procedures.