Management of Change (MoC) in petrochemical operations requires engineers to trace cross-document dependencies manually across many interdependent document types per process unit. These dependencies are latent: shared equipment tags, instrument identifiers, and operating parameters are distributed across Word reports and Excel spreadsheets with no machine-readable links. To our knowledge, no existing system infers cross-document dependencies at the section level by reconstructing entity relationships from document text across heterogeneous engineering formats under cold-start conditions; commercial systems organise documents by metadata (and in some cases link documents at the identifier level) but do not perform content-level dependency inference, and NLP-based traceability research targets software requirements rather than multi-format engineering documentation.

This thesis constructs a knowledge graph from 9 document types under cold-start conditions, with no pre-existing structured metadata or asset registries (though the pipeline relies on domain conventions such as ISA-5.1 tag formats). Custom parsers and deterministic entity extractors preserve document-structural provenance; entity resolution links variant mentions to canonical representations; dependency classification and graph traversal identify impacted documents and sections. A scope expansion step using cause-and-effect matrix (CEM) relationships discovers instruments attached to scope equipment, enabling the system to reach instrument-related sections that the scope narrative does not explicitly name. A two-tier addition gap strategy handles entities introduced by the change that have no prior graph representation: same-prefix peer lookup within the process unit, followed by entity-class profiles computed across units. A topology-based review ordering derived from entity and parameter density produces consistent document priority across all six units (mean pairwise Kendall’s τ = 0.795) and matches the review workflow confirmed by Shell process engineers.

Evaluation covers 12 MoC cases from Shell spanning 6 process units, with 467 ground-truth change locations derived from before-and-after document diffs (capturing what engineers actually changed, which may undercount true dependencies). The system achieves document-level recall (macro-averaged) of 0.904, precision of 0.706, and F1 of 0.776. In 9 of 12 cases, all modified documents were correctly identified; of the three cases with a missed document, two involve enclosure-calculation documents connected through indirect chemistry-mediated dependencies, and one is a pure-addition case that missed a P&ID drawing. Addition-gap mitigation raises document recall from 0.756 to 0.904. Region-level localisation reaches recall of 0.625 (F1 0.507). Baseline comparisons against five non-graph approaches confirm that the KG pipeline outperforms the best non-graph baseline (dense embedding, F1 0.612) by 16.4 F1 percentage points; no baseline simultaneously matches the KG’s recall and precision. Results demonstrate robustness across 6 process units of a single operator, not cross-client generalisability. A structured evaluation session with Shell process engineers indicated that the impact scope predictions align with engineering expectations and that cross-document dependency visibility is the system’s strongest practical contribution. ... Read more