There is no line B: four rings against unplanned downtime

The problem

There is no line B, and the root cause was visible somewhere.

Every hour of unplanned downtime is hundreds of metric tons not produced, plus the cost of restabilizing the process, plus the incident risk every start-up carries. Which is why the capital pain of this sub-sector is always the same: zero unplanned shutdowns. And now the hopeful part. When an unplanned shutdown is analyzed afterwards, the root cause is almost never a complete surprise: it is a symptom that was visible somewhere — on a round sheet, in a vibration trend, on a utility panel, in a corridor remark — and that did not reach the person who could act, in time. The problem, therefore, is not instrumentation. It is connection.

• A trip on the main air compression train does not stop oxidation: it stops the entire complex, because downstream runs out of feedstock within hours. And a continuous process is not resumed: it is restarted.
• Every hour of unplanned downtime is hundreds of metric tons not produced, plus the cost of restabilizing the process, plus the incident risk every start-up carries.
• When a shutdown is analyzed afterwards, the root cause is almost never a complete surprise: a symptom was visible somewhere — on a round sheet, in a vibration trend, on a utility panel, in a corridor remark — and did not reach the person who could act, in time.
• The problem, therefore, is not instrumentation. It is connection.

How it fits the IRIS system

Four rings that cross-check — it is all of the above, orchestrated, plus SMED AI on the planned shutdown.

Four rings, each with its own case behind it:

No single ring solves the problem: each one exists today in some form and the train still trips. The strength is the correlation — "this pump has had a declining oil level for three weeks, the downstream equipment has risen in vibration and the critical spare is not in the warehouse" — a sentence nobody builds today because its three pieces live in three different places.

• Ring 1 · Connect (cases 1, 2, 4): everything that is paper, an isolated panel or a corridor remark today enters central memory the same day it is observed. They stop being three separate worlds.
• Ring 2 · Connected condition monitoring (case 8): crossing a threshold creates the maintenance notification with its trend and its recommendation, and closing the order returns the outcome to the history.
• Ring 3 · Agentic correlation (cases 3, 6, 9): Agents cross-references the signals from the two previous rings against process data, laboratory analytics, thickness inspection and external supply alerts. What no single signal says on its own — "this pump has had a declining oil level for three weeks, the downstream equipment has risen in vibration and the critical spare is not in the warehouse" — the cross-check does say. Nobody builds that sentence today.
• Ring 4 · JIDOKA AI at start-up (cases 5, 10, 11): no pressurization authorization goes out without a line-up verified point by point and signed, and all the evidence is packaged automatically.
• SMED AI in parallel: it shortens the planned shutdown and the start-up. If shutting down in an orderly way is cheap and fast, the right decision gets taken in time and nobody stretches a machine until it breaks.

See the full IRIS architecture →

Before and after

The four rings, and what each one closes

Ring	Which cases it draws on	What it closes
1 · Connect	Rounds (1) + panels (2) + voice (4)	Paper, isolated panels and corridor remarks enter central memory the same day
2 · Connected condition monitoring	Vibration to CMMS (8)	Threshold to notification to outcome, with nothing keyed by hand
3 · Agentic correlation	Supply alerts (3) + lab sheet (6) + chip (9)	The cross-check no single signal says on its own
4 · JIDOKA AI at start-up	Verification gate (5) + line-up (10) + evidence pack (11)	No pressurization without a verified, signed line-up and its evidence packaged
SMED AI, in parallel	Planned shutdown and start-up	Shutting down in an orderly way becomes cheap, so nobody stretches a machine until it breaks
Humans in command	All twelve	Every alert is a proposal to a person; nothing writes to the DCS

from unplanned shutdowns whose root cause is reconstructed afterwards, to symptoms correlated and acted on days or weeks in advance. From a start-up authorized on a blindly signed list, to one authorized on point-by-point evidence.

Impact estimate

Estimated impact — to validate against your real cost per hour of train downtime.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Every hour of unplanned downtime avoided in a complex of this scale is worth an order of magnitude more than the full cost of the system.
• Which is why this case is justified against the first shutdown avoided, not by accumulated operational savings. Estimated payback 6-12 months, to validate against the trip history.
• It hits all four maintenance indicators at once: safety (ring 4), availability (rings 2 and 3), cost (all of them) and staff efficiency (ring 1).
• From a start-up authorized on a blindly signed list to one authorized on point-by-point evidence.

every hour of unplanned downtime avoided in a complex of this scale is worth an order of magnitude more than the full cost of the system. Which is why this case is not justified by accumulated operational savings: it is justified against the first shutdown avoided. Estimated payback 6-12 months, to validate against the real cost per hour of train downtime and the trip history. It hits all four maintenance indicators at once: safety (ring 4), availability (rings 2 and 3), cost (all of them) and staff efficiency (ring 1). *Estimate to validate*.

And the fair question from the production manager

«Do we have to deploy all twelve cases to get this?» — no, and framing it that way would be the way never to start. The rings are built in layers and each case pays on its own from the first month. With the preservation rounds and the vibration-to-CMMS stitching you already have ring 1 and ring 2 on the critical rotating equipment, which is where most trips on a single train are decided. Where Agents builds the correlation sentence, it builds it from signals that are each anchored — a reading, a threshold, a stock level — where the best models drop below 1.5% error [1], and the sentence is a proposal to the maintenance lead, never an order to the DCS.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.