Ontology-Constrained Causal Discovery for Root-Cause Reasoning in Cyber- Physical Production Systems: The Manufacturing Domain Causal Ontology (MDCO)

Root-cause isolation in cyber-physical production systems is constrained by a structural asymmetry: data-drivenanalytics recover association rather than causal structure and degrade under non-stationarity, while semantic assetmodels encode engineering knowledge but lack the directional and interventional semantics causal inferencerequires. We argue that the productive relationship between the two is constraint rather than fusion, and introducethe Manufacturing Domain Causal Ontology (MDCO), a Basic Formal Ontology-conformant ontology reusingSSN/SOSA and Industrial Ontology Foundry terms and extending them with reified causal links carrying direction,propagation delay, mechanism and provenance. The central contribution is a formal compilation Φ from MDCOaxioms — stratified into hard, verified and soft knowledge by the warrant each carries — into the background-knowledge interface of causal discovery: forbidden edges, required edges, tier orderings and soft priors. We state theconditions under which compiled constraints preserve the true graph, and gate compilation with SHACL validation.Across paired experiments on random and layered manufacturing-style DAGs, constrained discovery improvedstructural recovery over the identical unconstrained estimator in every configuration — with the largest gain on thedensest manufacturing-structured topology — while reducing conditional-independence testing by roughly 40–49%at scales up to 500 variables; the advantage replicated under a score-based estimator and scaled monotonically withknowledge coverage. Corruption experiments establish the boundary condition: the benefit reverses onceapproximately 5–10% of true edges are falsely constrained, with hard exclusions degrading fastest. Industrialdiagnostic validation is reserved for prospective work under a pre-registered protocol specified herein.