PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Graph Unification for Industrial Digital Threads: Bridging 11 Heterogeneous Manufacturing Systems Through Ontology-Driven Knowledge Graphs

Grama Chethan

arXiv:2608.24918v1cs.AI

TL;DR

Manufacturing systems operate under independent governance, motivating a coherent ontology for industrial digital threads. The paper introduces an ontology-driven discovery framework whose strongest result is that blocking cross-system joins reduces recall from 1.00 to 0.31.

  • Problem

    Independent governance across industrial systems creates overlapping but inconsistent data environments that motivate unified digital-thread analysis.

  • Method

    The paper combines a purpose-built ontology spanning 10+ industrial standards with an automated discovery engine containing nine pluggable statistical strategy categories.

  • Results

    Blocking 24 cross-system tools reduces recall from 1.00 to 0.31 and F1 from 1.00 to 0.48; verification against a 65-signal synthetic manifest yields F1 = 1.00.

  • Takeaways & Limitations

    The framework exposes the graph as a 287-tool MCP semantic layer for discovery across system boundaries.

  • Takeaways & Limitations

    The validation uses a synthetic, author-constructed manifest, while simulators omit message loss and legacy-system quirks; real-data F1 is expected to be lower.

Abstract

from arXiv · show

Modern manufacturing enterprises operate heterogeneous systems -- ERP, MES, PLM, SCADA, QMS, SCM -- each with its own data model and API. The resulting silos prevent holistic analysis, delay root-cause investigation, and obstruct Industry 4.0 traceability. Point-to-point integration scales as O(n^2) and accumulates brittle dependencies. This paper presents an open framework for semantic graph unification of industrial digital threads. An ontology-driven RDF knowledge graph unifies data from 11 simulated sources across nine domains through a five-stage ETL pipeline with automated entity resolution spanning 97 owl:sameAs identity links. The ontology encompasses 78 RDFS classes, 108 object properties, and 243 data properties, drawing on ISA-95, OPC UA, eClass, the Asset Administration Shell, RAMI 4.0, and additional standards. An automated discovery engine applies nine strategy categories -- cross-station correlation, alarm coverage, ECN impact, CUSUM/EWMA drift detection -- to surface insights spanning system boundaries. The primary empirical result: blocking 24 cross-system tools reduces recall from 1.00 to 0.31 (F1 from 1.00 to 0.48), showing that 69% of discoverable signals require cross-system graph joins. Leave-one-out ablation confirms six of nine strategies contribute unique signals. Verification against a 65-signal manifest (16 positive, 49 null) yields F1 = 1.00 (95% Clopper-Pearson CI [0.79, 1.00]); as the manifest was author-constructed, this constitutes verification not independent validation. The graph is exposed to LLM agents via 287 Model Context Protocol tools as a SPARQL-native semantic layer. Five industry templates (aerospace, CPG, pharma, medical devices, turbine blades) demonstrate schema stability across manufacturing verticals.

1. INTRODUCTION

Manufacturing data remains fragmented across heterogeneous systems, while existing integration approaches leave gaps in semantic alignment and cross-system querying. The paper proposes an ontology-driven RDF framework, ETL pipeline, discovery engine, and MCP layer, with experiments showing that cross-system joins are central to discoverable signals.

  • Motivation: Manufacturing systems use independent data models and APIs, leaving information about quality, yield, cost, and compliance trapped in silos.Bridging these silos requires laborious manual export-and-reconciliation cycles.
  • Motivation: The digital-thread vision links design, manufacturing, quality, and field service, but requires ingestion, entity resolution, semantic representation, and insight access.The paper frames traceability queries as a practical use case for this linkage.
  • Research gap: Existing ESB, data-lake, canonical-model, federated-SPARQL, and iPaaS approaches do not simultaneously provide semantic richness, automated entity resolution, and AI-ready access.Point-to-point adapter mappings scale as O(n^2), while most industrial systems expose proprietary REST or OPC UA APIs rather than SPARQL endpoints.
  • Contributions: The discovery engine applies nine statistical strategy categories and exposes 287 SPARQL tools to LLM agents through an MCP semantic layer.The framework is demonstrated across five industry-specific templates.
  • Empirical findings: 69% of discoverable signals require cross-system graph joins, as blocking 24 tools reduces recall from 1.00 to 0.31 and F1 from 1.00 to 0.48.Internal verification against a 65-signal manifest reports F1 = 1.00 with 95% Clopper-Pearson CI [0.79, 1.00].
  • Contributions: The framework combines an ontology spanning 10+ industrial standards with 78 RDFS classes, a five-stage ETL pipeline, and 97 owl:sameAs identity links.The pipeline ingests 11 simulated systems and validates the resulting graph before persistence.

2. RELATED WORK

Industrial standards provide rich but fragmented vocabularies, and prior approaches address only parts of cross-system semantic integration. The paper positions its framework as a broader open implementation combining ontology alignment, entity resolution, multi-strategy discovery, and LLM access.

  • Knowledge graph construction: Industrial ontology mappings can represent ISA-95 semantics yet still fail to align identically named but differently modeled ERP and PLM entities.Distinguishing an ERP material from a PLM part requires workflow-aware domain semantics beyond string similarity.
  • Industrial standards: OPC UA supports device-level interoperability but is designed for real-time communication rather than persistent semantic integration with enterprise systems.The framework extends beyond OPC UA by unifying SAP ERP, Teamcenter PLM, and supply-chain systems.
  • Industrial standards: Compared with OPC 30530, the framework adds explicit owl:sameAs resolution and SPARQL-based multi-hop traversal across non-OPC-UA systems.OPC 30530 is described as complementary through potential use as a source adapter.
  • Industrial standards: AAS provides a digital-twin serialization and exchange structure but does not specify ontological axioms for cross-system resolution or integration beyond the shell boundary.The framework embeds AAS-derived triples in an 11-source graph contextualized with ERP, MES, and supply-chain data.
  • Industrial standards: The collective standards gap is an implementation-ready ontology that expresses cross-standard alignment, supports automated entity resolution, and offers a consistent interface for analysts and AI agents.RAMI 4.0 provides an architectural reference while leaving cross-standard semantic details unspecified.
  • Framework positioning: The framework spans 11 simulated systems across the ISA-95 hierarchy and combines RDF ontology grounding with multi-hop tools over heterogeneous schemas and identity spaces.Its MCP layer is designed as a SPARQL-native semantic layer for manufacturing knowledge graphs.

3. SYSTEM ARCHITECTURE

GraphUnifier uses a four-layer stack to connect eleven industrial simulators, a five-stage ETL pipeline, a partitioned RDF store, and downstream applications. Its many-to-many simulator-to-domain structure supports cross-system SPARQL queries over unified manufacturing data.

  • Layers 2–3: The five-stage ETL pipeline connects the source services to an rdflib quad store partitioned into four named graphs.The graphs separate ontology, eCl@ss classification, plant instance data, and rolling live telemetry.
  • Layer 4: Layer 4 exposes the unified graph to downstream applications and AI agents through cross-graph SPARQL 1.1 GRAPH queries.These queries can join named graphs in a single round trip, such as correlating a live alarm with equipment hierarchy and OPC UA tag type.
  • Coordination and storage: The coordinator owns the endpoint registry, orchestrates ETL runs, and maintains cumulative triple, link, and timing statistics.Thread-safe graph access and locking prevent concurrent ETL cycles from corrupting in-progress queries.
  • Layer 1: Eleven independent simulator services generate synthetic data across ISA-95, OPC UA, SAP, Teamcenter, Insights Hub, SCM, and related industrial domains.The simulators collectively represent seven industrial domains, with several domains served by multiple systems.
  • Cross-system integration: The many-to-many relationship between simulators and industrial domains motivates owl:sameAs links that connect semantically equivalent entities across source systems.Production Operations spans ISA-95, OED, and SAP, while Automation and Control spans OPC UA, TIA Portal, and ISA-18.2.

4. ONTOLOGY DESIGN

The ontology maps existing industrial standards into one coherent namespace using competency questions and minimal alignment axioms. It supports cross-standard SPARQL discovery while retaining explicit boundaries around causal inference, real-time aggregation, and OWL expressivity.

  • Ontology scope: 78 RDFS classes, 108 object properties, and 243 data properties form the ontology’s cross-standard semantic foundation.The class map spans nine industrial standards, with coloured regions grouping classes by standard and dashed arrows marking inter-standard relationships.
  • Design methodology: Six representative competency questions guided schema development, with required query patterns determining vocabulary, properties, and cross-system link categories.The ontology was refined over five cycles by adding CQ-required classes and removing speculative ones.
  • Scope boundaries: The ontology does not model continuous real-time OEE aggregation or causal root-cause inference beyond its correlational scope.The paper identifies these negative competency questions as boundaries that inform future extension priorities.
  • Reasoning profile: Using RDFS rather than OWL DL supports the SPARQL-based integration use case but leaves constraints such as cardinality and class disjointness to procedural checks.The choice reflects rdflib’s reasoning support, pipeline-latency considerations, and the downstream tools’ use of SPARQL rather than DL subsumption queries.
  • Entity alignment: 97 owl:sameAs triples connect entities across seven boundary categories, using deterministic identifier-matching rules in the simulated sources.The links include SAP–OED orders, SAP–Teamcenter materials, OPC UA–ISA-95 stations, SAP–SCM purchase orders, and related alignments.
  • Bridging philosophy: Minimal alignment axioms map existing standards to one another while preserving each standard’s semantic commitments and enabling cross-standard inference.Examples connect ISA-95 equipment hierarchies with OPC UA namespaces and link Teamcenter design objects to ISA-95 execution through SAP and OED objects.

5. PIPELINE AND ENTITY RESOLUTION

The five-stage pipeline extracts, normalizes, enriches, links, and validates data from eleven simulators. Entity resolution creates bidirectional owl:sameAs links, while quality checks expose two partial failures associated with timestamp skew.

  • Pipeline overview: 5.2 s is the total wall-clock latency for one complete ETL cycle across all eleven simulators.Figure 3 reports each stage’s operation, timing, and approximate triple volume at stage boundaries.
  • Entity-resolution matrix: The entity-resolution matrix covers seven source systems and eighteen cross-reference categories, with colour indicating link density.Non-trivial links are bidirectional via owl:sameAs with symmetry enforcement.
  • Extract and Normalize: The Extract stage issues parallel requests with timeouts and retries, while Normalize converts source JSON into ontology-typed rdflib triples.A standard cycle produces approximately 8,000 triples after normalization.
  • Enrich and Link: The Classify/Enrich stage uses IRDI dictionary lookups, and the Link stage resolves eighteen cross-reference categories into bidirectional owl:sameAs triples.Enrichment adds approximately 200 triples, while linking adds approximately 300 owl:sameAs triples per cycle.
  • Validation: Ten Golden Triangle SPARQL ASK rules and more than twenty data-quality checks validate integrity across design, execution, feedback, and supply-chain flows.The checks cover completeness, consistency, validity, timeliness, and uniqueness.
  • Validation results: 18/20 data-quality checks pass consistently (90%), with two partial failures caused by timestamp skew at high time-dilation factors.After validation, the plant graph is atomically replaced under a lock.

6. AUTOMATED STATISTICAL DISCOVERY

The framework adds a continuously operating discovery layer to the unified graph, combining statistical testing, evidence tracing, recommendations, and human oversight. Nine strategies analyze cross-system patterns while conservative multiple-testing controls and actionability scoring prioritize findings.

  • Discovery architecture: Nine pluggable strategies run through DiscoveryRunner to query the unified graph, quantify findings, and produce evidence-linked recommendations.Strategies execute independently, with failed or timed-out strategies logged without aborting the discovery cycle.
  • Discovery architecture: Each strategy separates candidate discovery, evidence-chain construction, statistical quantification, and recommendation generation.This separation allows new strategies to be added without modifying the orchestrator.
  • Statistical methods: Six hypothesis tests support categorical, continuous, multi-group, small-cell, and cohort-rate analyses, including chi-square, Welch's t-test, Pearson correlation, ANOVA, Fisher's exact test, and rate ratios.Test selection follows the data distribution and comparison structure.
  • Statistical safeguards: Two-level Bonferroni correction controls multiple testing within strategies and across strategies, reducing family-wise error at the cost of some power.The choice is motivated by safety-critical recommendations; BH-FDR remains configurable for deployments prioritizing discovery breadth.
  • Insight prioritization: The composite insight score combines statistical evidence, estimated financial impact, and actionability, with effective weights of approximately 40%, 35%, and 25%.The design emphasizes financial and operational consequences alongside inferential measures.
  • Strategy coverage: Nine strategies span safety, quality, and operational categories while engaging between two and four source systems each.The strategy set targets commercially significant manufacturing intelligence phenomena.

7. MULTI-INDUSTRY GENERALIZATION

The framework addresses single-domain specialization by separating a shared ontology core from industry-specific configuration and terminology. Across five manufacturing verticals, the schema and tools remain stable while data values, mappings, and prompts vary.

  • Configuration stability: The 78-class ontology core is domain-agnostic, while a configuration layer supplies seed data, regulatory mappings, and industry terminology.This separation is designed to avoid architectural changes when extending to new industries.
  • LLM-assisted tooling: A tool learned against the aerospace graph remains valid against the pharma graph because shared ontology terms preserve the underlying tool interface.Industry-specific terminology is supplied through customized system prompts rather than separate implementations.
  • Configuration stability: Five industry configurations use the same ontology schema, MCP tools, and discovery pipeline while differing in plant, catalog, supplier, and regulatory data.The configurations cover aerospace, CPG, pharma, medical devices, and turbine blades.
  • Cross-industry discovery: Cross-industry discovery groups quality patterns by plant type to distinguish effects shared across verticals from findings specific to one industry.The strategy can examine whether alarm-coverage deficits are concentrated in high-complexity assembly stations.
  • Regulatory awareness: Regulatory annotations attach applicable standards and clauses to station instances so safety strategies can assign higher actionability to findings under mandatory requirements.The annotation supports risk-aware prioritization rather than operational compliance.
  • Regulatory scope: The framework does not implement manufacturing compliance machinery such as electronic signatures, validated state control, audit-trail integrity, or CAPA workflows.Compliance-focused extensions would require integration with systems such as EBR or QMS.

8. EVALUATION

Evaluation on the aerospace configuration combines blind discovery testing, pipeline performance, data quality, tool coverage, and cross-system use cases. Results show perfect manifest classification, broad cross-system dependence for discoverable signals, and measurable schema and pipeline trade-offs.

  • Discovery accuracy: Precision, Recall, F1, and Specificity each equal 1.000 on the 65-signal manifest, with FDR = 0.000 across 16 positive and 49 null signals.The confusion matrix is TP=16, FP=0, FN=0, TN=49, and the reported Recall, Precision, and F1 thresholds are all satisfied.
  • Discovery accuracy: The 95% Clopper-Pearson interval for the manifest result is [0.79, 1.00], and the author-constructed manifest does not provide independent validation.The paper notes possible confirmation bias in manifest construction and uncertain transfer to out-of-distribution manufacturing data.
  • Pipeline performance: An eleven-source unification cycle averages 5.2 seconds, placing the framework in the near-real-time batch analytics tier.The stated use cases include shift-level dashboards, daily quality reviews, and ad-hoc investigative queries.
  • Pipeline performance: The dominant performance cost is graph loading, where rdflib SHACL validation processes the full approximately 27,800-triple graph.The SHACL validation stage alone is reported at 4.89 seconds.
  • Tool coverage: Tool coverage rises from 89% to 96% after adding 23 LLM-generated tools, all human-reviewed before inclusion.The extension mechanism adds learned SPARQL queries to the existing authoring coverage.
  • Cross-system use cases: The five validated use cases connect engineering changes, materials, alarms, thermal propagation, and design intent across the unified graph.The paper states that no single source system contains enough information to reconstruct this end-to-end chain.
  • Cross-system dependence: Blocking 24 cross-system tools reduces recall from 1.000 to 0.312, a 69% absolute drop, while precision remains 1.000.Eleven of 16 positive signals require cross-system graph joins; the remaining recoverable signals are process-drift or alarm-coverage findings.
  • Ablation: Leave-one-out ablation finds that six of nine strategies contribute unique signals, led by characteristic_lifecycle and supplier_quality with ΔTP = 5 and ΔF1 = −0.185 each.Process_drift and cross_station follow with ΔTP = 4, ΔF1 = −0.143 and ΔTP = 3, ΔF1 = −0.103, respectively.

9. DISCUSSION AND LIMITATIONS

The cross-system ablation identifies graph unification as central to signal discovery, while verification and generalization results are bounded by synthetic data, scalability, and deployment constraints.

  • Empirical findings: 69% of discoverable signals require cross-system graph joins, as blocking 24 cross-system MCP tools reduces recall from 1.00 to 0.31.The corresponding F1 decreases from 1.00 to 0.48.
  • Empirical findings: The unified graph architecture, rather than only the statistical tests, is responsible for the majority of discovered signals.
  • Evaluation scope: Internal verification reaches F1 = 1.00 on a 65-signal synthetic manifest, with a 95% confidence interval of [0.79, 1.00].The manifest contains 16 positive signals and 49 null signals.
  • Scalability: The in-memory implementation performs adequately at approximately 27,800 triples but is not expected to scale to live production volumes without architectural changes.At 10× scale, complex-query latency is expected to become multi-second; at 100× scale, memory pressure and O(n²) joins are likely infeasible.
  • Deployment boundaries: Production deployment remains bounded by simulated sources, missing tool-level authentication and authorization, and regulatory metadata that does not enforce compliance.Real deployments would require adapter changes, coordinated identity controls, and regulated workflows such as signatures, audit integrity, and CAPA automation.
  • Evaluation scope: The verification is weaker than independent validation because the development team designed the discovery strategies, synthetic data, and evaluation manifest.Independent replication on production datasets with third-party expert manifests remains necessary.

10. CONCLUSION AND FUTURE WORK

The paper concludes that semantic unification, automated discovery, and LLM-native access provide a framework for manufacturing digital threads, while identifying production-scale and causal extensions as future work.

  • Conclusion: A 78-class ontology and five-stage ETL pipeline unify 11 simulated source entry points across nine data domains.The pipeline uses owl:sameAs entity resolution and reports 90% data-quality compliance in 5.2 seconds.
  • Conclusion: Blocking cross-system joins reduces recall from 1.00 to 0.31 and F1 from 1.00 to 0.48, while six of nine strategies contribute unique findings.These results support the paper’s claim that cross-system graph unification is a prerequisite for most discoverable manufacturing signals.
  • Conclusion: The nine-strategy discovery engine is verified against a 65-signal synthetic manifest with F1 = 1.00 and a 95% Clopper-Pearson CI of [0.79, 1.00].The framework also exposes a 287-tool MCP semantic layer for LLM agents.
  • Conclusion: Multi-industry generalization is demonstrated through templates spanning aerospace, CPG, pharmaceutical, medical-device, and turbine-blade manufacturing.The turbine-blade template is described as a structurally related extension.
  • Future work: Future work prioritizes native triple-store migration, event-driven ingestion, expanded industry coverage, causal inference, and federated learning across factories.The proposed streaming architecture would replace polling-based simulation and support millisecond-latency graph updates aligned with physical process events.

CREDIT AUTHOR CONTRIBUTION STATEMENT

The author contribution statement assigns responsibility for conceptualization, methodology, software, validation, analysis, investigation, data curation, and writing.

  • Contributions: Grama Chethan contributed conceptualization, methodology, software, validation, formal analysis, investigation, and data curation.
  • Contributions: The author also contributed to the original draft, review and editing, and visualization.
Loading 2608.24918v1…