The refinery in question runs two high-pressure steam header networks feeding 9 process areas. Steam is the workhorse of the site — it drives reboilers, strips columns, traces miles of pipework — and at 38 bar / 380 °C, every kilogram of it carries a real number on the invoice.
01 · The brief
The reliability lead handed us a stack of trap-survey PDFs going back to 2019 — annual ultrasonic walkdowns, three contractors, four reports, none of them in agreement. The ask was simple: tell us which traps are leaking today, and rank them by what it costs us in steam. Without a shutdown, without IT involvement, and before the next audit.
The unsaid part of the brief was harder. Trap surveys produce binary labels — passing, failed-open, failed-closed — but the cost of a failed-open trap varies by two orders of magnitude depending on orifice size, upstream pressure, and how much of the year it’s been failed. A list of red dots isn’t actionable. A ranked steam-loss ledger is.
“A trap survey gives you a list. A continuous baseline gives you a ledger.”
02 · Why this header
We started on the HP header in process area 4 for three reasons: the longest piping run, the densest population of traps, and a 2024 survey flagging “elevated bypass loss” without quantifying it. There are 41 traps on this header. Walking it takes 40 minutes at a brisk pace. Surveying it properly takes a week. Continuously monitoring it should, in theory, take an afternoon.
03 · Instrumenting in 5 days
Sixty-one clamp-on sensors — temperature pairs across each trap, four ambient references on the deck, and two flow-correlation nodes on the condensate return — went up in five working days with two engineers and a scissor lift. Nothing was cut, threaded, or shut. The radio gateway sits in the rack room and uses the site’s existing LoRaWAN coverage; the data path never touches the OT network.
We deliberately oversized the instrumentation. Survey-grade work would have used half the sensors. The extra channels exist so that fouling models can be back-fit against ambient and load — the rest of the year of analytics flows out of that redundancy, not out of the trap data itself.
04 · What we found
Eleven traps were leaking on the day we energised. By week four — once the baseline was clean enough to trust — that number climbed to thirteen, then settled. Three traps (T-02, T-04, T-07 in Fig. 01) accounted for 62% of the lost steam. One of them, T-04, had been recorded as passing in the 2024 walkdown.
The interesting failures weren’t the loud ones. Two traps showed the signature of partial bypass: not enough to register on a ultrasonic sweep, enough to quietly cost €18k/year each. Those are the kind of losses that don’t show up until the year-end energy reconciliation, and even then they get absorbed into “process drift.”
| TRAP | MODE | STEAM LOSS | ANNUALISED € | STATUS |
|---|---|---|---|---|
| T-02 | Failed open | 1.42 t/h | €124,000 | Replaced wk-7 |
| T-04 | Failed open | 0.71 t/h | €062,400 | Replaced wk-7 |
| T-07 | Failed open | 0.48 t/h | €041,800 | Replaced wk-9 |
| T-11 | Partial bypass | 0.21 t/h | €018,600 | Rebuilt wk-9 |
| T-23 | Partial bypass | 0.19 t/h | €017,200 | Rebuilt wk-9 |
| T-30 | Plugged | — | — | Cleaned wk-11 |
05 · The €310k itemised
We don’t publish savings numbers we can’t itemise to a meter and a tariff. The €310k here is the sum of measured steam losses, valued at the site’s published 2026 steam tariff of €37/t, with downtime and labour subtracted. Three traps did most of the work; the rest is a long tail of partial-bypass and plugged condensate returns.
06 · What changed operationally
The reliability team didn’t really need a dashboard — they needed a queue. Each Monday morning the system emails the planner the three traps most likely to deliver the largest recovery this week, ranked by a simple €/hour of millwright time heuristic. Eight months in, the queue has never been empty, and the steam reconciliation on this header is the tightest it has been since the network was commissioned in 2008.
The pilot is now running on a second header, and the methodology is being extended to heat exchangers in process area 7. We’ll write that one up when it’s no longer interesting, which is the only honest standard for a case study.