CEMS stack emissions with AI — an analyzer in drift is not an incident, it is the quarter's reporting invalidated.

The problem

The CEMS works fine — until the lab finds out it had been lying for weeks.

The environmental manager of a refinery or a cogeneration plant does not need anyone to explain what a CEMS is. They live in it. What keeps them awake is something else:

1. The analyzer looks operational on the SCADA because it still delivers a value — but that value has been running with a zero drift for three weeks that nobody looked at, and QAL3 is out.
2. The QAL2 laboratory comes every few years, not every month. Between campaigns, the only safety net is continuous statistical control — and that usually lives in a spreadsheet from which the traceability of the whole period hangs.
3. BREF, CORINAIR and E-PRTR reporting gets rebuilt by hand at year-end: hourly averages, daily averages, validity flags, gap substitution. A week of work, and a risk of error that is not the analyzer's responsibility but that of whoever compiled it.
4. When the inspection arrives, what is requested is not just the data — it is the data lineage: which calibration was in force, what incident occurred, what human decision was taken, where the signature is.

The classic system (analyzer + DCS + spreadsheet + the manager's eye) works most months. But the cost is not in the majority: it is in the month it fails, because any invalidated hour drags the whole report with it, and by then it is a problem that leaves the department.

How it fits the IRIS system

iLEAN does not replace the CEMS — it seals the cracks between the analyzer, the DCS and the reporting.

A CEMS does not fail for lack of hardware: the shelter is there, the analyzer is there, the QAL2 laboratory comes when it should. It fails in what happens in between — the undetected drift, the incident nobody logged, the report assembled at the end with patched data. iLEAN acts as the putty between the pieces you already have: it reads from the analyzer and the DCS without asking them to change, validates continuously, anticipates failure and prepares the regulatory dossier ready to sign.

Edge reads the analyzer data the way the manager would read it walking up to the screen. Brain computes QAL3 continuously and cross-references it with process conditions. Writer generates the BREF/CORINAIR report. The person signs — always.

The three iLEAN pieces applied to CEMS emissions control:

• iLEAN Edge — a terminal that captures analyzer data by whichever route exists (OPC/Modbus bus, 4-20 mA, OCR of the local screen if the shelter is old and isolated). It also reads auxiliary measurements: flow, temperature, pressure, reference oxygen. It works without a network: if the plant loses connectivity, Edge keeps recording locally, because the regulator does not accept “there was no WiFi” as a justification for a gap in the report.
• iLEAN Brain — continuously computes QAL3 statistical control (CUSUM, Shewhart, zero and span drift), cross-references it with real process conditions (furnace load, fuel type, operating mode) and raises the alert while the drift is still recoverable, not once it has already invalidated the period. It learns the normal pattern of your stacks and separates what is a process phenomenon from what is an instrument failure.
• iLEAN Writer — prepares the regulatory deliverables (monthly and annual report per stack, E-PRTR, CORINAIR table with emission factors), with the data lineage attached: the QAL2 calibration in force, QAL3 incidents, human validations, justified gap substitutions. The manager reviews and signs; they do not rebuild anything by hand.

See the full IRIS architecture →

Before and after

Classic CEMS management vs. a continuously validated CEMS with iLEAN

Aspect	Classic CEMS + spreadsheet + eyeballing	With iLEAN Edge + Brain + Writer
Continuous QAL3	Monthly chart reviewed by hand by the manager	Computed on every reading, alert before going out of tolerance
Drift detection	Found at the next calibration or audit	Anticipated from a pattern cross-referenced with the real process
Integration with the analyzer	Whatever the integrator wired to the DCS	Modern bus, 4-20 mA or screen OCR — whatever there is
Gap substitution	Manual rule applied at closing	Configured rule, applied on the spot, justified in the dossier
BREF / CORINAIR / E-PRTR report	Annual reconstruction, several person-days	Permanent draft, validated and signed at closing
Regulatory inspection	“We'll have to look for it” — several weeks	Dossier with full lineage in a couple of clicks

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the data of your stacks and your real incident frequency. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Refinery or petrochemical plant with several CEMS stacks (boilers, furnaces, flare) measuring NOx, SOx, CO, CO2, particulates and reference oxygen.
• Pilot on one stack: integration with the existing analyzer, continuous QAL3, predictive drift alert, monthly dossier with lineage. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the current cost of rebuilding regulatory reporting, the frequency of documented QAL3 incidents and the avoided cost of a single invalidated hour multiplied by your emission factor.
• Expected reduction in compliance effort ≥ 30% (to be validated against your environmental team's baseline). The hard lever is not saving person-hours, it is avoiding the audit that reopens a quarter.

And the environmental manager's reasonable doubt

“What if the AI gets it wrong and validates an hour that was not valid?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply recontextualizes a piece of data from one system into another (reading the analyzer, comparing it with the QAL2 calibration in force, applying the QAL3 rule), the best models brought error below 1.5% ^[1]. And even so, nothing regulatory is closed alone: iLEAN proposes the flag and the dossier, the person signs. The three safety rings exist precisely for this — the CEMS OT network stays isolated, the AI lives in the outer ring, and only what has been validated enters the inner ring.

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