Catching the early warning signs

Ammonia plants almost always show warning signs before they fail. The problem is that the signs are subtle and easy to miss: the reformer, the synthesis loop, the big rotating machines on refrigeration and CO2 compression all degrade in ways that show up first as a change in the relationship between dozens of tags, long before any one of them crosses an alarm limit. By the time an alarm fires, the cushion to act has often gone. A modern plant streams thousands of tags into its historian every second, yet most monitoring still waits for a fixed threshold to be crossed. The data is all there; the early warning is what is missing.

That is the gap a major nitrogen producer set out to close, beginning with one of its ammonia plants as the first site in a planned, fleet-wide rollout. It is one of the world’s largest nitrogen producers, turning out more than 5 million tonnes of urea and nearly 2 million tonnes of ammonia per year. ControlRooms deployed its AI troubleshooting agent on top of the plant’s existing data infrastructure, with no change to control logic or hardware. Within weeks it was surfacing the early signs the plant’s alarms were never able to catch.

Two models, one job: catch it before it becomes a problem

Harmony Models lay the foundation

For each area of the plant, ControlRooms builds a single Harmony Model, a multivariate model that learns how all the relevant tags in that area normally behave in concert. These models are not trained on a library of known failure patterns. They learn what normal looks like and flag departures from it, allowing them to catch problems the plant never thought to watch for: a group of variables drifting out of their usual harmony while all individual readings still look fine. A handful of them covers an entire plant within weeks. Catching the unknown unknowns, the problems no one thought to watch for, is exactly what Harmony Models are built for.

Predictive models add the scaffolding

Alongside that broad coverage, smaller asset-level models tackle known failures and pivotal performance-improvement areas: a bearing with a failure history, an analyser prone to drift, a control valve that wears, a heat exchanger that fouls, a pump losing suction, a compressor stage running hot. A Harmony Model watches a whole area at once; a predictive model zeroes in on one asset and the specific way it tends to fail.

Catches from the first six months

A blind spot on a critical bearing

A Harmony Model flagged a subtle anomaly on a combustion air fan journal bearing in the reforming and heat recovery area. The cause turned out to be the bearing’s temperature sensor, which had failed. A dead sensor on a critical bearing is dangerous precisely because it is quiet: it leaves the bearing’s real temperature invisible and can mask an overheating event entirely. Because the alert put it in front of the team, they took the machine down to fix it rather than keep running blind. A senior process engineer at the site, who had watched the signature build over the preceding days, was direct about the value: “These alerts are very useful. As soon as something happens we get informed. Never had this functionality in the past. If you guys are catching that signature, that’s a great sign”.

A compressor tracked toward its turnaround

A predictive model on a large ammonia refrigeration compressor flagged slowly rising journal-bearing vibration. Rather than a single alarm, this became a months-long watch: reviewing it week by week, the team correlated the vibration with falling lube-oil temperature and the anti-surge valve position at reduced rates. The bearing was tracked intelligently with a predictive model right up to the planned turnaround, instead of failing between turnaround cycles. This is the kind of slow, interconnected degradation that no single alarm could ever explain.

A worn valve no one could see

A Harmony Model on the synthesis area picked up a developing oscillation across the ammonia flash-drum and cold-exchanger levels. The cause was a level control valve whose plug and seat had worn, quietly forcing operators onto a bypass to hold level. Nothing had alarmed. The same engineer put plainly why whole-system monitoring earns its keep: “I would not have known about this defect if I didn’t have the visibility via ControlRooms.ai”.

A signature that pointed at fatigue

The model surfaced temperature fluctuations on the wall of the ammonia synthesis reactor that tracked pressure cycling in the synthesis loop. That is not a nuisance reading: pressure cycling drives cyclic fatigue, and a sister facility had previously suffered a through-wall crack on the reactor’s outlet cone. Catching it at the temperature-fluctuation stage gives the engineering team the data to investigate the fatigue risk early, instead of finding the damage at an inspection.

Twelve hours of warning on a reformer tube leak

A focused oxygen model on the primary reformer, run against data from a past tube-leak event, flagged a clear deviation roughly twelve hours before the unit came down, with measured oxygen falling below the model’s prediction in step with the onset of the leak. That is a whole shift of warning on one of the most expensive failures an ammonia plant can have.

Cleaner data and sharper operations

Some of the earliest flags were not process failures at all. The models surfaced bad tags, incorrect tag metadata and poorly set alarm limits. These are the subtler operational-excellence gains that come with real visibility into the process: cleaner data, better-tuned alarms and a more trustworthy picture of the plant, whether or not any given flag turns out to be a failure. One of the site’s engineers put it directly: “Not only is ControlRooms.ai detecting anomalies in the measured values but also pointing out errors from when the historian was set up”.

Beyond the data itself, the catches kept coming across the plant: a CO2 compressor signal loss the engineer learned about from the alert before anywhere else, a failed desuperheater starved of boiler feedwater, a reformer running with burner cocks closed. Different equipment, same pattern: the system saw it first.

Where the value compounds

Early detection only pays off if someone acts on it, and that is what this deployment was built around. Alerts were routed into the team’s existing collaboration channel, not a new dashboard to remember to open. Each one arrived as a short, readable message naming the area and tags, with a trend and a one-click link straight into a view of the anomaly inside the app.

An engineer could open the alert, pull related tags into the same view, and share a link that put anyone else into the exact view they were looking at, instead of trading screenshots and stale reports. Investigations that used to take days closed out in a single shift, thanks to those streamlined troubleshooting features. The channel became a living record of what the plant was doing and why, with each engineer and operator’s explanation attached to the alert that prompted the investigation. Even alerts that proved benign or already known were not wasted: each added context or prompted a slight model tuning, and a standing weekly review kept improving the system. The result was a plant that ran better, with teams troubleshooting together instead of in isolation, fewer surprises, and more work shifting from reactive firefighting to preventive intervention.

The impact

The operator’s own before-and-after captures both the depth and the breadth of the value the system delivers. Earlier, an undetected temperature excursion driven by a control-valve failure had taken a unit down for the better part of two days, with several thousand tons of lost production. Later in the same year, when a positioner began to fail on a cold exchanger, AI anomaly detection flagged the developing oscillation early enough for a planned, targeted intervention. The unit kept running and the event closed out with no downtime at all.

That contrast, undetected and costly versus seen early and contained, is now something the operator points to in its own performance reporting, and it factors AI-based anomaly detection into its multi-year value-creation targets. None of it required new instruments or changes to the control system. It came from catching developing problems hours, sometimes days, earlier than the alarms could. Across the first six months, catches like these added up to thousands of tons of saved production.

From an initial deployment to a fleet

This was the first site in a planned, fleet-wide rollout, and the early wins here are what carried it forward. Six months of incremental catches built the case: real coverage across the plant, a team that reviewed proactive AI alerts, and a range of findings from a worn valve to a fan bearing to a reformer tube leak. With that foundation the producer rolled the system out to four more production sites, where it is live today.

The takeaway applies to any continuous process. The plant did not lack data and it did not lack alarms. What it gained was the few hours of warning between a pattern starting to drift and a limit being crossed, delivered to the people who could act, via an intuitive interface. On an ammonia complex, those few hours are the difference between proactive intervention and a multi-day shutdown.