Jump to main content Jump to navigation Jump to search Jump to footer
Jump to main content
Distributed and Self-organizing Systems
From Silos to Semantics: Building and Evaluating a Validated Industrial Knowledge Graph Pipeline
From Silos to Semantics: Building and Evaluating a Validated Industrial Knowledge Graph Pipeline | Distributed and Self-organizing Systems
 

PUBLICATION

From Silos to Semantics: Building and Evaluating a Validated Industrial Knowledge Graph Pipeline

Type

Conference Paper

Year

2026

Authors

Research Area

Web Engineering

Event

4th International Workshop on Semantic Industrial Information Modelling (SemIIM 2026)

Abstract

Industrial process data are typically distributed across heterogeneous tabular sources, making semantic integration, validation, and reuse difficult. Knowledge graphs provide a principled foundation for interoperable industrial data, yet existing work rarely evaluates whether the semantic validation mechanisms protecting these graphs are themselves effective. As ontology engineering and constraint generation increasingly incorporate large language models (LLMs), understanding the trustworthiness of validation pipelines becomes all the more important. This paper presents an end-to-end semantic engineering workflow for transforming heterogeneous process tables into a semantically validated knowledge graph and quantitatively evaluating the validation layer that safeguards it. The workflow combines CMPO, an openly released OWL ontology for chemical mechanical planarization (CMP) built on the W3C SOSA/SSN observation pattern, with ModPipe, a six-stage pipeline that performs data ingestion and profiling, ontology alignment, value normalization, RDF generation, validation, and persistence. Beyond constructing the knowledge graph, we evaluate the validation process itself. Using mutationbased experiments, competency-question evaluation, an audit of a previous deployment, a comparison with an external SHACL baseline, and an assessment of LLM-generated validation rules, we measure which classes of semantic errors are detected, which remain undetected, and why. The results show that node-local constraints reliably detect structural and completeness errors, while graph-level rules are required for vocabulary, identity, and domain-specific semantic constraints. The study also reveals that validation suites can silently become ineffective through model evolution or automatically generated constraints, despite the supporting tools reporting successful execution. We argue that trustworthy industrial knowledge graphs require not only semantic modelling and validation but also systematic evaluation of the validation mechanisms themselves. By releasing the ontology, construction pipeline, validation framework, and reproducible evaluation artifacts, this work provides a reusable methodology for trustworthy semantic industrial information modelling.

Reference

Walpola, Maheshika Hansamalee; Samuel, Sheeba; Gaedke, Martin: From Silos to Semantics: Building and Evaluating a Validated Industrial Knowledge Graph Pipeline. 2026.

Cite this

BibLaTeX