Leak detection with thermal vision and AI — the leak you can already smell arrived too late.

The problem

Classic LDAR finds what already escaped weeks ago.

The environmental / HSE lead of a chemical plant has lived with the same dilemma ever since LDAR rules have existed:

1. The manual survey is a snapshot. A technician with an OGI camera or a handheld sniffer covers thousands of components in a quarterly or biannual campaign. The leak that appears on the Monday after the survey has been emitting for three months by the time it is detected.
2. The smell arrives late and depends on the wind. By the time an operator “smells something”, the plume has been emitting for hours — sometimes days. Smell is an indicator, not a sensor.
3. Earlier continuous OGI stumbled on false positives. The first fixed thermal camera deployments without AI triggered on water vapor, on sunlight reflecting off hot flanges, on vehicle exhaust plumes. They ended up muted or ignored.
4. The regulator's file is paper. An EPA or competent-authority audit means rebuilding traceability, calibrations, events and repairs by hand — weeks of work per campaign.

The classic system works most of the time. But the serious leak is the one that escapes between surveys. This is not a people problem — it is a sampling frequency problem, and the answer is not hiring more technicians: it is having the plant see itself continuously.

How it fits the IRIS system

iLEAN does not add one more standard — it seals the cracks between the camera, the regulator and the maintenance team.

The plant already has OGI cameras (sometimes), a documented LDAR program and an EAM/CMMS for maintenance orders. Each of them does its own job well. The problem lives in the joints — between what the camera sees and what the technician interprets, between what is detected and the repair order, between the repair and the file for the regulator. iLEAN acts as the putty that fills those gaps.

The thermal camera sees the plume. The AI model filters noise. The agent opens the order, follows the repair and signs off the file. The person validates; the system does not close it alone.

The three iLEAN pieces applied to leak detection in a chemical plant:

• Edge + thermal Vision — machine-vision terminals (CNN) on a fixed thermal camera watching the pipe rack, the critical units or the tank farm. The model learns to tell a gas plume from water vapor, from sun glare, from people moving through the frame. It raises an alert only when confidence is high. It works without a network: the basic detection loop keeps running on cabinet power.
• Connect — captures the operator's report that they “smell something” over the radio or WhatsApp, correlates it with the thermal image at that moment, and cross-references it with wind direction and line pressure. It also captures the reports from the last manual OGI inspection, the CoAs of the gaskets installed and the maintenance technician's work sheets.
• Agent — when there is a high-confidence detection, it opens an order in the EAM/CMMS, assigns the technician, follows the repair, schedules the post-repair verification with the same camera and keeps the LDAR dossier signed and ready for the auditor. It coordinates with sustainability to report the corrected leak in footprint KPIs. The person signs every critical step.

See the full IRIS architecture →

Before and after

Survey-based LDAR vs. continuous LDAR with iLEAN

Aspect	Manual survey LDAR	With iLEAN Vision + Edge + Agent
Inspection frequency	Quarterly or biannual	Continuous, 24/7, per component under camera
Average time leak → detection	Weeks to months between surveys	Hours, sometimes minutes
False positives on fixed OGI	Previous fixed camera muted in practice	AI model filters water vapor, sun, movement
Repair order	Manual, paper, days of delay	Agent opens the order in the EAM on the spot
Post-repair verification	At the next survey	Same camera, same shift
File for the auditor	Rebuild a spreadsheet, weeks	Live LDAR dossier, signed per component

Impact estimate

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

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

• Mid-sized chemical plant or refinery with a formal quarterly LDAR, an OGI program with a handheld camera, 10-30 critical units and an EAM/CMMS in place.
• Thermal Vision + Edge pilot on 2-3 prioritized racks or units, integrated with the EAM and with LDAR reporting. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the cost of a manual LDAR campaign, documented regulatory incidents and the internal valuation of penalty risk.
• The hard lever combines: product recovered because a leak was fixed in hours instead of months, a measurable reduction in reportable footprint, avoiding an unscheduled inspection by the regulator, and fewer hours spent on manual LDAR campaigns.

And the HSE lead's reasonable doubt

“What if the model invents a leak that does not exist and shuts down the unit?” — hallucination is a problem of free generation, not of anchored tasks. Leak detection is exactly an anchored task: the image is the source, the model segments and classifies. In this kind of task, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN alerts and proposes; the person validates and signs the shutdown or the repair. The three safety rings exist precisely for this.

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