Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders

URL SCAN: Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This paper is building a memory prosthesis for industrial plants. Its real product is not cleaner maintenance text; it is the conversion of technician experience into verified, queryable maintenance capital. Rule-first extraction, provenance, human verification, ontology governance, and MCP delivery create a pipeline that moves diagnostic competence from individual technicians into an organizational graph accessible to agents.

The system correctly identifies a real failure in SAP PM data: abbreviated, contradictory, incomplete work orders leave valuable failure knowledge trapped in human memory. But its deeper function is obsolescence infrastructure. Once recurring diagnosis is captured, verified, cited, and served to agents, the technician who carried that knowledge becomes less economically scarce.

The Core Fallacy

The paper treats preservation of knowledge as preservation of human productive necessity. Under the Discontinuity Thesis, those are opposites. A graph that remembers the expert is also a mechanism for replacing the expert.

The reported evidence is structurally narrow. A synthetic 12-month export of 400 work orders and 28 diagnostic questions cannot establish generalization across real plants, equipment classes, languages, sensor contexts, documentation cultures, or rare catastrophic failures. The claimed 100% root-cause precision applies only at 62% coverage; the remaining cases are outside that guarantee. The abstract provides no evidence that ontology accuracy produces lower downtime, safer interventions, or sustained economic value in live operations.

Hidden Assumptions

• Synthetic work orders adequately represent the disorder of production environments.
• The labels used to measure component, failure-mode, and root-cause accuracy are themselves reliable ground truth.
• Technicians will consistently verify facts, answer knowledge-gap questions, and approve ontology changes.
• Human verification remains cheaper than improving the extraction and agent layers.
• Textual failure knowledge is sufficient without embodied inspection, tacit context, sensor history, or physical judgment.
• Agents will respect evidence citations and uncertainty rather than convert partial coverage into false confidence.
• Knowledge captured in the ontology will remain current as equipment, terminology, and procedures change.
• Better diagnosis will preserve technician employment rather than reduce the number of technicians required per plant.

Social Function

Primary classification: transition management. Secondary classifications: elite self-exoneration, prestige signaling, and partial truth.

The paper accurately exposes unusable maintenance records and proposes meaningful controls against hallucinated knowledge: provenance, confidence, verification states, and disagreement queues. That is the partial truth. Its ideological function is to let owners describe labor displacement as “knowledge retention” and “agent access.” The technician is recast from knowledge holder into validation infrastructure, then becomes removable when validation itself is automated.

The Verdict

Technically useful, evidentially immature, and strategically hostile to routine diagnostic labor. Plant Memory does not defend human productive participation. It packages maintenance experience into AI-readable capital and shifts leverage toward whoever controls the plant’s data, agents, and physical assets. Human work survives longest in inspection, repair, safety accountability, and exception handling—the maintenance flank of the New Power Trinity. Plant Memory is not a barrier to obsolescence. It is an early component of the machinery that produces it.