PaperlayerView exact revision on arXiv

Source-linked AI summary

Composable Trust Infrastructure for Manufacturing Knowledge Graphs: Cross-System Provenance, Temporal Reasoning, and Decision Traceability

Grama Chethan

arXiv:2608.21418v1cs.AIcs.SE

TL;DR

Manufacturing knowledge graphs lack a unified way to establish data validity, provenance, decision-time state, and action traceability across heterogeneous systems. The paper integrates four trust capabilities into a composable RDF architecture, concluding that each is load-bearing for full-chain auditability while evaluating a structurally realistic but proof-of-concept testbed.

  • Problem

    Heterogeneous manufacturing knowledge graphs do not jointly establish whether data is valid, valid at decision time, sourced correctly, and acted upon traceably.

  • Method

    The paper integrates SHACL validation, PROV-O provenance, bi-temporal versioning, and graph-native decision objects through shared identifiers in an RDF architecture.

  • Results

    Full-chain auditability is an emergent property that no proper subset of the four capabilities can deliver.

  • Takeaways & Limitations

    The results support composing the four trust capabilities rather than treating them as independent metadata or workflow functions.

  • Takeaways & Limitations

    The proof-of-concept testbed contains 6,582 triples across 329 plant entities and requires migration to a scalable triple store for larger deployments.

Abstract

from arXiv · show

Manufacturing knowledge graphs that integrate data from heterogeneous industrial systems face a trust deficit: consumers cannot determine whether queried data is valid, whether it was valid when a decision was made, where it originated, or how it was acted upon. We argue that four trust capabilities -- SHACL validation, PROV-O provenance, domain-aware bi-temporal versioning, and graph-native decision objects -- compose through shared correlation identifiers to produce emergent trust properties that no single capability delivers alone. We present a composable trust infrastructure that integrates these four capabilities into a unified RDF architecture. Capabilities compose through shared entity URIs, ingestion activity identifiers, and temporal correlation keys, enabling compound queries spanning all four dimensions. An experimental ablation confirms that removing any single capability causes exactly three of six composition queries to fail, demonstrating that all four are equally load-bearing. Analysis of higher-order compositions reveals four emergent three-way properties and one irreducible four-way property (full-chain auditability, 31ms execution). The infrastructure is validated on a testbed integrating eleven industrial sources -- OPC UA, TIA Portal, eClass, AAS, ISA-95, ISA-18.2, SAP S/4HANA, Teamcenter, Opcenter EX, Insights Hub, and SCM -- under an 89-class ISA-95-aligned ontology. The unified graph contains 8,743 triples across five named graphs, stitched by 81 owl:sameAs identity edges. Evaluation uses simulated but structurally realistic data from purpose-built emulators; data structures and cross-system linkage patterns are representative of real industrial installations.

1. Introduction

Manufacturing knowledge graphs must combine validation, provenance, temporal reasoning, and decision traceability to answer cross-system trust questions. The paper presents an RDF infrastructure in which shared identifiers compose these capabilities into compound queries and emergent trust properties.

  • Motivation: Heterogeneous manufacturing systems leave consumers unable to establish data validity, decision-time state, origin, and downstream action together.The motivating scenario spans provenance, SHACL validation, temporality, and decision traceability across system boundaries.
  • Motivation: No single capability answers all four trust questions: provenance lacks structural validation, SHACL lacks version alignment, temporal reasoning lacks decision context, and decision objects lack evidence provenance.The paper frames the trust deficit as a compositional gap rather than a deficiency in one isolated technique.
  • Contribution: Four capabilities compose through shared entity URIs, ingestion activity identifiers, and temporal correlation keys to produce trust properties unattainable by subsets.The shared identifiers connect validation reports, provenance activities, temporal version chains, and decision evidence links.
  • Testbed: The testbed integrates eleven industrial source systems into one RDF graph using owl:sameAs identity edges and an ISA-95-aligned ontology.The sources include PLM, ERP, MES, IoT, alarm, equipment, classification, digital-twin, and supply-chain systems.
  • Contribution: The infrastructure enables provenance-scoped validation, temporally qualified decisions, and decision-outcome attribution through compound queries crossing the trust dimensions.These query classes ask which source data failed validation, what validated evidence existed at decision time, and whether triggered actions improved relevant metrics.
  • Engineering contributions: The engineering contributions include graph-native Decision, ActionExecution, and ActionOutcome objects, hybrid PROV-O migration, domain-aware temporal derivation, and incremental SHACL validation.The validator reduces validation from O(n) full-graph traversal to O(k) ingestion-batch processing and adds an 11-rule cross-system consistency check.

2. Background and Related Work

The paper builds on manufacturing knowledge graphs, ISA-95, quality assessment, provenance, temporal models, and decision traceability. Its distinct contribution is integrating eleven sources and treating trust dimensions as first-class graph entities.

  • Manufacturing knowledge graphs: Prior manufacturing knowledge-graph work established cross-system linkage, analytics, and ISA-95-based semantic backbones, but often relied on proprietary connectors or stopped before runtime integration.The cited literature covers industrial analytics, quality prediction, semiconductor manufacturing, and ontology-driven integration.
  • Manufacturing knowledge graphs: The paper integrates eleven industrial sources into a single RDF graph spanning engineering, production, alarms, equipment, classification, digital twins, and supply-chain data.The integration uses an ISA-95 foundation extended across the source systems.
  • Composable trust infrastructure: The architecture overlays SHACL validation, PROV-O provenance, bi-temporal versioning, and decision objects as first-class graph citizens rather than external metadata stores.This positions data quality, lineage, temporality, and decision traceability within the same graph.
  • Quality assessment: The work extends established OWL-to-SHACL practice by integrating generated shapes into incremental manufacturing-pipeline validation and adding 11 cross-system data-flow rules.Validation targets newly added entity subgraphs against the full shape graph and includes domain-specific identity consistency.
  • Temporal data models: Its temporal contribution is domain-aware valid-time derivation, mapping entity types to meaningful timestamp fields for point-in-time queries without application-level temporal logic.Examples include WorkOrder actualStart/actualEnd, Alarm activatedAt/clearedAt, and MaterialLot producedAt/consumedAt.
  • Decision traceability: Graph-native decision objects link decisions to evidence, affected entities, actions, and outcomes, enabling SPARQL joins between decision trails and operational results.The model introduces Decision, ActionExecution, and ActionOutcome as RDF classes.

3. Competitive Landscape

The competitive landscape compares trust capabilities, standards, query depth, scalability, security, and support across manufacturing integration platforms. The paper positions its prototype as open and composable, while acknowledging commercial advantages in production readiness.

  • Comparison framework: The paper evaluates manufacturing integration systems across composable trust, provenance, temporal data, validation, decision traceability, open standards, and cross-system query depth.The comparison is presented as a capabilities-based landscape rather than a single performance benchmark.
  • Limitations: Commercial platforms provide production-grade scalability, security, high availability, failover, and support that the in-memory RDF prototype lacks.The paper specifically identifies dedicated triplestores as necessary for enterprise deployment.
  • Differentiators: The proposed infrastructure uses W3C standards including SHACL, PROV-O, OWL, and SPARQL, allowing components to be replaced or extended without vendor lock-in.Open-standards composability is presented as a primary differentiator.
  • Differentiators: The widest SPARQL query traverses seven systems from an engineering change notice through production, execution, station, sensor, alarm, and HMI data.The stated path runs through Teamcenter, SAP, OED, ISA-95, OPC UA, ISA-18.2, and TIA Portal.
  • Differentiators: Decision traceability is represented as graph-native entities embedded with the manufacturing data they reference, enabling provenance-aware queries from sensor anomalies through decisions to quality outcomes.Existing platforms are described as externalizing comparable functions to workflow engines or audit logs.
  • Positioning: The paper characterizes its approach as complementary to Opcenter Intelligence: the prototype supplies research-grade trust infrastructure, while Opcenter supplies production-grade operational capabilities.The comparison identifies differences in SPARQL querying, PROV-O chains, bi-temporal queries, and graph-linked decisions.

4. Method

The method integrates heterogeneous industrial data through a five-stage RDF pipeline with ontology-derived validation, provenance tracking, temporal state, and decision objects. Cross-system identity links and shared graph structures support validation and audit queries across industrial boundaries.

  • Architecture Overview: The five-stage pipeline extracts data from eleven simulators, normalizes JSON into RDF, links identities, validates the graph, and persists temporal and decision state.Bi-temporal state and decision chains are persisted to PostgreSQL in parallel with graph loading.
  • Cross-System Identity: The Linker materializes 81 owl:sameAs edges across seven system boundaries, enabling SPARQL traversal without prior schema alignment.These links connect entities such as stations, work units, orders, materials, purchase orders, equipment, and IoT assets.
  • Validation: SHACL shapes are auto-generated from the ontology, with domain-range and class-specific constraints enforcing labels, provenance, required properties, and referential links.The transformation from OWL to SHACL is treated as established practice; the implementation contribution is incremental validation integrated into the pipeline.
  • Validation: Incremental validation reduces validation time from O(|G|) to O(|ΔG|) by validating subgraphs containing newly added entity URIs against the full shape graph.Full-graph validation remains in the Loader stage because isolated subgraphs can miss cross-entity violations when related entities have not yet been ingested.
  • Provenance: Entity-level PROV-O records link ingested entities to ingestion activities and pipeline-created entities to generation activities, avoiding triple-level provenance overhead.The provenance graph remains bounded at O(|E| + |R|), rather than potentially unbounded O(|E| × |R|).

5. Composable Trust Infrastructure

The infrastructure defines trust capabilities that compose through shared identifiers, enabling compound queries unavailable to individual capabilities or proper subsets. Ablation and higher-order analysis show symmetric load-bearing behavior and an irreducible four-way audit query.

  • Composition Model: The paper’s central contribution is composing provenance, validation, temporal versioning, and decision recording so that the combined system yields trust properties unavailable to proper subsets.Composition is defined through shared correlation identifiers whose joins produce assertions not derivable from either capability alone.
  • Capability Definitions: The four capabilities provide source attribution, structural validation, point-in-time reconstruction, and evidence-linked human or automated decision records.Each capability contributes queryable assertions that can be joined through shared entity, activity, and temporal identifiers.
  • Ablation Analysis: Removing any single capability causes exactly three of six composition queries to fail, while no query produces a degraded result.Every query either fully succeeds or fully fails, and no single capability is sufficient for any composition query.
  • Higher-Order Compositions: All four three-way capability subsets yield emergent queries, because temporal scoping changes current conformance or provenance into decision-time evidence.The analysis identifies four distinct three-way subsets and shows that their queries are not reducible to pairwise combinations.
  • Higher-Order Compositions: The four-way full-chain auditability query combines provenance, SHACL conformance, decision-time graph state, and decision rationale in one result set.It cannot be decomposed into sequential pairwise or three-way queries because temporal constraints must jointly scope provenance and validation for the same evidence entity.
  • Conclusion: The infrastructure exhibits 6 pairwise, 4 three-way, and 1 four-way emergent composition properties, enabled by shared entity, activity, and temporal identifiers.The full-chain auditability property is presented as the paper’s central claim, with no additional coordination mechanism required.

6. Cross-System Identity Resolution

The unified graph bridges eleven source systems through 81 owl:sameAs identity edges across seven system boundaries. Identity resolution uses curated deterministic rules that suit the proof-of-concept testbed but require more scalable approaches for enterprise deployment.

  • Identity graph: 81 owl:sameAs edges bridge 11 source systems across seven system boundaries.The edges link cross-system entities and aliases, including OPC-UA with Insights Hub, SAP with OED, Teamcenter, and SCM, plus intra-system variants.
  • Identity graph: Identity links cover sensor tags, production and work orders, material masters, purchase orders, tag aliases, and SAP material-code variants.The reported categories include 19 OPC-UA–Insights Hub edges, six edges for each of three SAP-related boundaries, 10 intra-OPC-UA edges, and 34 intra-SAP edges.
  • Resolution strategy: The current approach avoids, rather than resolves, the open-world inference complexity associated with owl:sameAs assertions.The paper notes that curated rules are suitable for the testbed, while OWL identity semantics can produce unintended inferences.
  • Resolution strategy: Deterministic, rule-based matching is appropriate for the proof-of-concept testbed but does not address enterprise-scale identity resolution.The paper identifies probabilistic matching, confidence scores, skos:closeMatch for non-exact identities, and human-in-the-loop validation as needed for larger deployments.

7. Evaluation

The evaluation uses a proof-of-concept testbed with eleven simulators generating structurally realistic data conforming to relevant industrial schemas. It evaluates architectural soundness and capability composition rather than production deployment.

  • Data generation: Simulator outputs conform to relevant industrial schemas and include referential integrity, realistic value ranges, and cross-system structure.The stated standards include ISA-95, ISA-18.2, ISA-88, OPC-UA, eCl@ss, AAS, SAP S/4HANA, Teamcenter, and SCM models.
  • Data generation: The testbed spans control engineering, plant hierarchy, taxonomy, digital twins, alarms, work execution, IoT, ERP, PLM, and supply-chain data.Each simulator produces domain-specific entities such as production orders, asset hierarchies, item revisions, inspections, and warehouse staging lots.
  • Scope: The evaluation validates architectural soundness and capability composition, not a production deployment.The paper explicitly distinguishes the proof-of-concept testbed from production evaluation.

7.2 Quantitative Results

The evaluation reports cross-system graph coverage and executable composition queries for validation, provenance, temporal reasoning, and decision traceability. Results show that shared identifiers support audits of evidence validity, origin, historical state, and decision context, while trust overhead remains small relative to pipeline execution.

  • Testbed metrics: 11 source systems and 329 plant entities quantify the proof-of-concept graph’s integration breadth.Insights Hub contributes 152 entities, followed by ISA-95 with 48, OED with 30, and OPC-UA with 27.
  • Composition validation: Six composition queries are verified executable against the live knowledge graph.The queries cover SHACL with PROV-O, decision context preservation, provenance-aware historical reconstruction, decision provenance, conformance timelines, and decision integrity.
  • Composition validation: A shared focus-node URI joins SHACL validation with PROV-O provenance, exposing invalid evidence together with its source activity and agent.Neither SHACL nor PROV-O alone provides both validity and source attribution.
  • Composition validation: Bi-temporal joins reconstruct evidence state at decision time rather than returning only the current state.Without bi-temporal versioning, later evidence changes could make retrospective auditing misleading.
  • Composition validation: Decision–PROV-O composition verifies whether decision evidence originated from a known, authorized source system.The example checks whether an NCR disposition used data from the OED simulator rather than an untrusted external source.
  • Performance: ~500ms of total trust overhead is negligible relative to the ~250-second full pipeline cycle.The reported performance context includes point-in-time queries remaining sub-5ms for single entities and incremental SHACL validation reducing cost by an order of magnitude versus full-graph validation.
  • Performance: 50 pipeline runs at a 5% change rate produce ~1,139 rows rather than 16,450 through change detection.The benchmark attributes the reduction to avoiding storage of unchanged entities across runs.
  • Integration: Trust triples preserve backward compatibility: 270+ tools execute without modification and existing SPARQL queries require no changes.Trust metadata resides in separate named graphs, supporting the reported hybrid provenance migration pattern.

8. Limitations and Future Work

The work is limited by simulated, small-scale data, manually curated identity links, an in-memory graph store, absent production connectors and governance controls, and no user study. Future work targets production validation, scalable storage, probabilistic identity resolution, usability, and security hardening.

  • Limitations: The proof-of-concept contains 6,582 triples across 329 plant entities, three to four orders of magnitude below industrial reality.Reported industrial examples include 10M+ manufacturing events per day, 50M+ Teamcenter item revisions, and 100K+ SAP material masters.
  • Limitations: All 11 source systems are simulators whose data does not capture production noise, inconsistency, missing values, or evolving schemas.The simulated data conforms to correct schemas and uses realistic value ranges, but production representativeness is not established.
  • Limitations: Production integration requires connector engineering for Teamcenter, SAP, and OPC-UA that is not addressed here.The paper specifically cites ITK/SOA, RFC/BAPI, and OPC-UA subscription and session-management requirements.
  • Limitations: The 81 deterministic owl:sameAs links are manually curated and do not establish a production-scale identity-resolution solution.Larger deployments would require probabilistic resolution, confidence scoring, candidate blocking, and human review for high-stakes assertions.
  • Limitations: The rdflib in-memory graph store is adequate at testbed scale but is not viable for production-scale data volumes.Migration to a scalable triple store would affect query optimization, transaction management, cluster configuration, and potentially SPARQL queries.
  • Limitations: Security, access control, encryption, audit logging, and named-graph governance remain engineering work beyond the proof-of-concept.The paper identifies ITAR, GDPR, role-based access control, and SOX/FDA-compliant requirements as relevant governance concerns.
  • Limitations: Decision objects have not been evaluated with manufacturing engineers for usability, cognitive load, or workflow integration.The paper states that decision traceability’s value depends on practitioner adoption and user-experience factors not addressed in this work.
  • Future work: Future work includes a production pilot, scalable-store benchmarks at 10^6–10^8 triples, probabilistic matching, user studies, and security hardening.The proposed axes include real Teamcenter, SAP, and OPC-UA sources; confidence-scored identity links; human validation; and named-graph-level controls.

9. Security and Governance Considerations

The section identifies security, privacy, access-control, sovereignty, and audit requirements for manufacturing knowledge graphs, while noting that these mechanisms remain largely architectural or incomplete in the current proof of concept.

  • ITAR and Export Control: Provenance chains may expose supplier relationships, material specifications, and process parameters subject to ITAR controls.The provenance graph itself may constitute controlled technical data when it reveals defense manufacturing workflows.
  • GDPR and Personal Data: GDPR erasure rights can conflict with 10–15-year regulatory retention requirements for manufacturing decision records.The section proposes pseudonymization with role-based tokens and a separately protected mapping table.
  • Access Control Architecture: Named graphs support tiered role-based access across ontology, plant, provenance, and runtime graphs, but the current implementation does not enforce this restriction.The proposed tiers align with manufacturing role hierarchies and need-to-know access.
  • Data Sovereignty: Jurisdictional graph partitioning supports data residency, while federated queries introduce latency and complexity not addressed here.The concern includes potentially conflicting European GDPR and United States ITAR requirements.
  • Audit Logging: Regulatory auditability requires mutation logs with timestamps, actors, and before/after state, plus tamper-evident properties beyond bi-temporal versioning.The authors describe these requirements as substantial engineering work beyond the current proof of concept.

10. Revised Claims of Novelty

The paper narrows its novelty claims after peer review, retaining the composition of established trust capabilities as the central contribution and withdrawing claims with substantial prior art.

  • Architecturally novel: Four retained claims center on SHACL validation, PROV-O provenance, bi-temporal versioning, and decision traceability composing through shared correlation identifiers.The revised positioning distinguishes the composition from the individual capabilities, which already exist in isolation or established standards.
  • Architecturally novel: The retained composition is formally defined and evaluated through six pairwise composition queries and ablation analysis.The paper positions systematic composition within a manufacturing knowledge graph as the demonstrated novelty.
  • Architecturally novel: The paper states that no proper subset delivers full-chain auditability.This is presented as the emergent result of composing the trust capabilities rather than as a property of any individual mechanism.
  • Dropped claims: The paper withdraws automated SHACL generation from OWL structure because substantial prior art exists, including SHACLGEN and TopBraid.The withdrawal reflects explicit prior-art positioning rather than a claim that the technique is infeasible.
  • Dropped claims: The paper also withdraws tombstone-based change detection because it applies the established Slowly Changing Dimension Type 2 technique to RDF graphs.The authors retain novelty for the application context while acknowledging that the mechanism itself is not novel.

11. Conclusion

The paper concludes that composing four trust capabilities through shared identifiers yields emergent trust properties, while its proof-of-concept validation remains bounded by simulated, small-scale, and incomplete infrastructure.

  • Contribution: The infrastructure integrates SHACL validation, PROV-O provenance, bi-temporal versioning, and graph-native decision traceability through shared correlation identifiers.The conclusion frames their composition, rather than any individual capability, as the central contribution.
  • Contribution: Six pairwise combinations produce emergent trust properties, culminating in full-chain auditability that no proper subset can deliver.The conclusion presents the full capability set as necessary for this final property.
  • Validation: The proof-of-concept testbed uses 11 source system simulators, an 89-class ontology, 81 identity edges, and five named graphs.The reported testbed spans ISA-95, ISA-88, ISA-18.2, eCl@ss, and AAS standards.
  • Validation: All 270+ pre-built analytical tools continue functioning without modification after the trust infrastructure is added.This result is reported as part of the proof-of-concept validation.
  • Limitations: The evaluation remains limited because its scale is three to four orders of magnitude below industrial requirements and all data sources are simulated.Additional boundaries include deterministic identity resolution, unimplemented security and governance mechanisms, and no user study.
Loading 2608.21418v1…