Reports as Hubs, Findings as Authorities: Investigating Graph-Centrality Methods for Discovery in Heterogeneous NDE Inspection Corpora

Three decades of non-destructive examination (NDE) have produced large inspection corpora of reports, findings, components, weld histories, standards, procedures, and time-linked asset information whose original purpose was compliance, not discovery. These records anchor inspection planning, review, and decision-making, yet most corpora remain document-centric rather than structurally represented, making it difficult to identify which records are authoritative, governing, or influential within a broader inspection history. The field lacks a principled way to ask, of its own archives, which findings govern an asset's history, which procedures anchor a body of work, and which earlier inspections set the framework that later inspections operate within. NDE inspection corpora can be represented as heterogeneous knowledge graphs. This paper investigates whether such a representation, combined with graph-centrality methods developed originally for hyperlink and citation analysis, supports authority analysis in NDE corpora. We model documents, findings, components, defects, methods, and standards as distinct Node classes connected by typed Edges, and within this representation we examine degree centrality as a baseline, PageRank and personalized PageRank for recursive authority estimation, HITS for hub–authority decomposition, and betweenness centrality as an indicator of structurally critical records. We pay particular attention to NDE corpora's sparse, multi-component, temporally layered, and naturally bipartite structure — in which reports function as hubs while findings, components, and standards function as authorities Accordingly, we evaluate not only ranking behavior but also descriptive graph properties such as degree distribution and connectivity. The investigation contributes (1) a typed-graph schema for NDE inspection corpora, (2) measurements of descriptive graph properties — degree distribution, connectivity, and component structure — that condition the behavior of standard centrality measures, and (3) a comparative analysis identifying where these methods fit, where they fail, and where domain-specific adaptation is warranted. Results that contradict the working hypothesis will be reported. By outlining these methodological constraints, this paper sets the groundwork for advanced, graph-based inspection discovery in industrial datasets.