Your contractor's numbers set your due dates. Make them earn it.

Why a steel mill's readings come from outside

An integrated mill runs with a small integrity group and a large contracted inspection spend, and the reason is structural rather than budgetary. A blast furnace campaign fixes when anything can be opened. A coke battery essentially never stops. The mill's own crews are consumed by mechanical maintenance during the same windows. So when a stove goes off-blast or the BOF shop takes its shut, the mill mobilises three or four NDT service companies for ten days, and those crews generate most of the thickness data that will drive due dates for the next several years.

The consequence is that the mill's integrity record is assembled from submissions it did not control. Different employers, different written practices under SNT-TC-1A, different instruments, different couplants, different calibration blocks, and different readings of the same CML list. Each contractor is entirely competent within their own convention. The mill inherits the mixture, and the mixture is where the errors live — not in any single crew's work.

This is not a contractor problem to be solved by hiring better contractors. It is a data problem that has to be solved at ingestion by the receiving system, because the mill is the only party that sees every submission and the only party that owns the trend across them. A vendor pitching you an interval engine should be able to explain what it rejects, not merely what it calculates.

Single-echo, echo-to-echo, and wall that appears from nowhere

Almost every external line in a steel mill is coated, painted or heavily scaled. A conventional single-element thickness measurement that times from the interface echo to the first backwall includes the coating in the reading. An echo-to-echo measurement between successive backwall echoes excludes it. The difference is the coating thickness — commonly 0.015 to 0.025 inch on mill piping, and considerably more on a line that has been painted every campaign for thirty years.

Now mix conventions across visits. Consider a 60-inch blast furnace gas main, nominal 0.375 inch, minimum structural thickness 0.200 inch, thinning at roughly 0.012 inch per year. Contractor A reads echo-to-echo in 2024 and reports 0.260. Contractor B reads single-echo in 2026 through 0.020 inch of paint and reports 0.256. The apparent loss is 0.004 inch over two years — a rate of 0.002, a remaining life of twenty-eight years, and a next date sitting comfortably at the code ceiling. The true 2026 metal thickness is 0.236, the true remaining life is three years, and the correct interval is eighteen months on a line carrying carbon monoxide through occupied plant.

Nothing in that sequence looks like an error. Both readings are correct measurements of different things, taken competently. The only defence is that every reading carries its technique as a mandatory field, and that the engine refuses to trend across a technique change without an explicit reconciliation. That single rule catches more bad due dates in a steel mill than any amount of risk-based sophistication layered on top of contaminated data.

The temperature correction nobody applied

Ultrasonic thickness gauges compute wall from time of flight using a velocity constant, and that constant is set at ambient temperature on a calibration block. Sound velocity in carbon steel falls as temperature rises, so a gauge calibrated cold and used on hot steel reports a thickness greater than the truth. The accepted field rule is to reduce the reading by approximately one percent for every 100 degrees Fahrenheit above the calibration temperature.

A steel mill is full of surfaces where this matters, and the refinery rules of thumb do not obviously map onto them. Hot blast stove shells and the downcomer sit in the 250 to 450 degree band. The bustle pipe and hot blast main, refractory lined internally, run warm on the outside. Reheat furnace recuperator ducting, waste heat boiler casings and coke oven gas mains downstream of the primary cooler are all warm enough to matter. On a 0.500 inch stove downcomer at 400 degrees the uncorrected reading is about 0.015 inch high — in a service consuming a few thousandths a year, that is several years of loss erased by an arithmetic omission.

The correction itself is easy. Knowing whether it was applied is not. If the submission does not record surface temperature at the time of measurement and whether a correction was made, the mill cannot distinguish a corrected reading from an uncorrected one, and cannot legitimately compare this campaign's numbers with the last one. Surface temperature belongs in the required intake fields alongside probe frequency, couplant and calibration block, and it costs the technician nothing to record.

A negative corrosion rate is a decision, not an error

The most frequent pathology in third-party thickness data is a reading thicker than the one before it. The causes are ordinary: a different technique, a different point within the same CML, a repad or weld overlay, a replaced spool nobody logged, an operator resting the probe on a scale nodule. The effect inside a spreadsheet is uniform and dangerous. Either the rate goes negative and the sheet throws an error somebody suppresses, or someone clamps it at zero and remaining life becomes infinite.

Both codes anticipate exactly this situation. API 510 and API 570 allow the owner-user to establish a corrosion rate from equipment in the same or similar service when measured data is unavailable or unreliable, and to work from an estimated rate until enough trustworthy data exists to compute one. That is the correct handling: substitute a defensible rate rather than accept an impossible one, and never let an impossible one propagate silently into a date.

What separates a real system from a spreadsheet with a better interface is that the substitution is a recorded act. The reading is quarantined. A reason is selected from a controlled list. A substitute rate is applied with its source stated — circuit default, similar service, engineering estimate. A named person approves it. Six months later, when someone asks why this line is on a two-year cycle while the identical line beside it is on five, the file answers the question without anyone reconstructing it from memory. That is what reproducible means in practice.

CML identity is the asset, not a label

A condition monitoring location is a physical spot with a history. A tag is a string that points at it. Contractors routinely rewrite the string: BFG-101-E1 comes back as BFG101 EL1, or a crew that could not safely reach a point at elevation invents BFG-101-E1A for a substitute location three feet away and reports it in the same column as the original. Two visits later the mill holds two histories of half the length, no rate on either, and no indication that anything went wrong.

The mill's own actions do identical damage. A spool is replaced during a shut and the CML now sits on new steel at full wall. A repad is welded over a wasted area. A fitting is cut out entirely and the CML no longer physically exists. None of these are inspection findings; they are maintenance events, and in most mills they are recorded in the maintenance system and never reach the inspection file at all.

The design answer is that CMLs are created by the mill and never by a submission. A contractor's file may only report against locations that already exist in the register; anything else is rejected at intake with a request to add the location formally. And the engine has to accept asset change events — replaced, repadded, retired, re-rated — that close one series and open another with a stated starting thickness, so that a discontinuity in the trend is explained rather than smoothed into a corrosion rate that never happened.

Where thinning math is the wrong model entirely

Two of a steel mill's highest-consequence damage mechanisms consume no measurable wall. Coke oven gas is wet, sour and ammoniacal, and carbon steel in that service is exposed to hydrogen-induced cracking, stress-oriented HIC and sulfide stress cracking, all described in API 571. A thickness history running through a decade of wet H2S service can be perfectly flat and entirely reassuring while the steel develops a network of internal blisters and step-wise cracks that no thickness gauge is looking for.

Blast furnace stove shells fail by a different crack mechanism again. Condensed moisture carrying nitrogen compounds on the cold face of a stove shell, combined with high residual stress at the combustion chamber welds, produces nitrate stress corrosion cracking — a well-documented stove problem that shows up at shell seams and manhole reinforcements and that no corrosion-rate table anywhere will predict. Stove shell integrity programmes are built on crack detection: magnetic particle over ground welds, ACFM where grinding is impractical, and shell movement and temperature monitoring between openings.

For both, the interval basis is inspection coverage and effectiveness, not remaining wall. The engine has to be able to say that this circuit's date is set by the last magnetic particle coverage of the weld network, the percentage of that network actually examined, and the indications carried forward with their last sizing — and to hold that alongside a thinning-driven date on the same asset, taking whichever comes sooner. A product that forces every asset into a corrosion-rate field will hand you a ten-year date on a cracking service without blinking.

The intake gate

Everything above resolves to a single design decision: does data enter the system through a gate, or through a paste. A gate is a defined submission format with mandatory fields — CML drawn from the register, reading, technique, probe frequency, couplant, surface temperature and whether a correction was applied, instrument serial number, calibration expiry, technician identity with certification scope and expiry, date and time — validated on arrival and either accepted, quarantined or rejected with a reason returned to the contractor while they can still act on it.

The rules that earn their keep are few. Reject any CML not present in the register. Quarantine any reading that increases beyond measurement uncertainty relative to the prior reading. Quarantine any reading whose technique differs from the technique established for that CML. Reject any reading taken against an expired certification or a lapsed instrument calibration. And escalate immediately, before the crew demobilises, any reading at or below the required minimum thickness.

That last rule is worth more than all the others combined. In a ten-day shut, a reading that arrives in a PDF three weeks later and turns out to sit below minimum forces a decision with no measurement option left: run, repair on the fly, or take an unplanned outage in a plant where the furnace does not wait for anybody. A same-shift flag turns that whole scenario into a re-shoot by a technician who is still standing on the scaffold.

How to evaluate the product

Bring your worst submission, not your best. Every mill has one campaign's data that never reconciled and got quietly set aside. Hand that file to the vendor and ask them to load it and tell you what they reject and why. A product that reports no rejections is not clean; it is credulous, and it will convert your contractors' inconsistencies into confident due dates with a completely straight face.

Then test the awkward states directly, one at a time. What date does it compute for a CML with a single reading? With a negative rate? With a technique change halfway through the history? On a circuit whose minimum thickness has never been derived? Where a spool was replaced last year and logged only in the maintenance system? Ask to see the full calculation chain for one asset rendered on a single page, and hand that page to your Level III. If they cannot verify it unaided, the mill has bought software, not reproducibility.

Atlantis builds the interval engine on Odoo, so the contractor purchase order, the submission, the technician's certification record, the instrument calibration register and the equipment file are all the same database — which is why the intake gate can check credentials and calibration validity without anybody looking anything up in a folder. Affordable, accessible, fully customizable; validation rules are configured to your own written practice rather than fixed by ours. To run your worst dataset through it, contact info@atlantisndt.com.

Request a consultation