System design for physical AI: steam turbine condition monitoring that can never write to the DCS

Posted on Oct 25 • Originally published at codeatoms.ai

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/steam-turbine-condition-monitoring-pakistan/. DOI 10.5281/zenodo.23365092. The operator is described by class, never by name.

A combined cycle plant in Pakistan is rehabilitating a steam turbine and installing a new 280 MW, 20 kV hydrogen-cooled generator. The protection layer already trips the unit when bearing vibration reaches Zone C. The question the design answers is the one before the trip: can the shift operator see a developing excursion, with a prescribed action, while the bearing housing is still drifting up from its nominal 2.3 mm/s RMS?

The DCS is read through a hardware one-way egress device into a DMZ OPC UA collector. The platform gets process, alarm and vibration context and has no path back into the control network, so the existing trip philosophy is untouched. The design chose this over a direct platform connection into the control network because the NEPRA Security of IT and OT Regulations 2022 and the sector's own pitfalls rule out a live bidirectional path for an analytics layer. A signed zone and conduit diagram is a phase one deliverable, not an afterthought.

Eight named systems feed the platform: the DCS, the 3500 vibration monitoring rack, the generator condition monitoring systems, the document management system, the programme schedule, the plant CMMS or EAM, SCADA and outage management. Four adapter families (integration, file and document, operational systems, telemetry) turn each source's protocol and naming into instances of one ten-object model, so nothing downstream knows that a vibration channel arrived as an OPC UA node and a test report arrived as a document. The plant CMMS is marked domain typical: it carries no commitment until the operator confirms it.

The condition model holds the steam turbine, the new generator, the vibration rack, the condition anomaly alert, the condition work order, the instrument calibration record, the commissioning and performance test record, the shift operator, the warranty engineer and an availability measure. The links are typed, so one traversal answers "every vibration anomaly since the rehabilitation, the calibration state of each channel when it fired, and the work order and warranty engineer now attached". In a document store that question is application code that joins loose records by hand.

MOMENT-1 (MIT) runs zero-shot anomaly detection over vibration, hydrogen, winding and process telemetry, because the rehabilitated unit's recovered history may be too sparse to baseline every channel. IBM Granite 4.2 8B (Apache-2.0) drafts diagnostic notes and test reports for engineer review; it was chosen over a 27B model as the smaller, licence-clean fit with real KV headroom on one card. Both run on one 48 GB PCIe inference class card on ruggedised edge servers inside the plant perimeter. Pakistan has no hyperscaler region, so no tier assumes a cloud, and model updates arrive by mirror-archive delivery on the site's own import cycle.

Appendix A prices the site envelope (the host share for one L40S-class card, a PTP grandmaster time card, two industrial switches and a UPS) at 21,600 to 23,400 US dollars over three years, about the same as the deepest three-year AWS commitment for one L40S card in the UAE region, which could not keep the read one-way in any case. The data diode and enclosure are priced by quotation and left out.

Alert flooding trains operators to ignore alerts, so the design carries a gate of 2 false alerts per week, curated unit-checked series, and channel health separated from process alarms. The platform advises; the protection layer trips.

Designed on CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. Paper, ontology and register: https://codeatoms.ai/steam-turbine-condition-monitoring-pakistan/

One of the CodeNinja Atoms sovereign AI reference architectures: every published design with its sector, country, compute and three-year cost.