Why the Same UT Grid Gives Four Different Rates Across a Fleet

Four sites, four answers, and one fleet number nobody trusts

The search that lands on a page like this almost never starts as a technical dispute about corrosion. It starts in a quarterly fleet reliability review, where the same component family — feedwater elbows downstream of a control valve, extraction steam lines, heater drain piping — carries a wear rate at one station that is half the rate at another. The engineering manager cannot say whether that gap is a real difference in chemistry and hydraulics or a difference in how two procedures divide a number. Until that is settled, the fleet roll-up is a slide nobody defends, and every station quietly keeps its own workbook running alongside the corporate system.

The uncomfortable finding, most of the time, is that the difference is arithmetic. Two stations measured the same wall loss and divided it by different denominators, subtracted it from different baselines, or reduced a twelve-point grid to a single governing number by different rules. None of those choices is wrong on its own. Each is written into a site procedure, each has been through review, and each produces a different rate from the same instrument output.

This is why the instinct to normalise everything into one house convention backfires. Recomputing historical rates under a new rule destroys the traceability that made the original numbers acceptable in the first place, and it hands the next audit a set of results that no longer match the reports issued at the time. The correct move is to store the convention alongside the result and compute the fleet view as a second, clearly labelled number.

Flow-accelerated corrosion is not a corrosion rate in the ordinary sense

Most corrosion rate modules were built for a hydrocarbon plant, where a circuit corrodes because of what is in the fluid, roughly in proportion to how long the fluid has been in contact with it. FAC is a mass transfer problem. The protective magnetite layer on carbon steel dissolves into flowing water at a rate governed by temperature, pH, dissolved oxygen, local velocity and geometry, and the alloy content of the base metal. It peaks in single-phase water near 300 degrees Fahrenheit and falls away on both sides of that window. Even a tenth of a percent of chromium in the steel reduces it measurably, and low-alloy chromium-molybdenum material is effectively immune.

Two consequences follow directly into the calculation. First, wear accrues only while the plant is operating at susceptible conditions, so time at shutdown genuinely contributes nothing and any denominator that includes it is wrong. Second, the rate is a property of the operating state rather than of the component alone. A feedwater chemistry change, an oxygen injection programme, a power uprate, or a single spool replaced in a chromium-bearing alloy is a step change in the rate, not a point on a trend.

That is why the industry writes both a short-term and a long-term rate and takes the more conservative of the two. This is not an abundance of caution. The long-term rate carries the memory of a chemistry regime that may no longer exist, and the short-term rate carries the current regime along with most of the measurement noise. Neither is sufficient alone, and a module that stores only one of them has already lost the argument.

The time base is the first thing that breaks

Take an eighteen-month operating cycle, roughly 547 days. A unit that spent 55 of those days in a refuelling outage and a short forced shutdown operated about 1.35 equivalent years. Dividing measured wall loss by 1.5 calendar years instead of 1.35 gives a rate about 10 percent lower. A sister unit that spent 90 days in a steam generator replacement outage inside the same nominal cycle operated 1.25 equivalent years, and its calendar-basis rate lands about 16 percent low. Neither station has done anything unusual. The fleet comparison is now carrying a sixteen point spread that has nothing to do with corrosion.

It gets worse across a mixed fleet, because a rate expressed per year is not even comparable when one station runs an eighteen-month cycle and another runs twenty-four. The per-interval loss is being annualised against different amounts of exposure, and the annualisation itself introduces the error. The only stable denominator is hours at susceptible conditions, taken from the plant process computer rather than inferred from outage start and end dates.

The practical requirement for the module is narrow and specific: it must accept an operating-hours value per interval as an input field, not derive one from two calendar dates. Effective full power hours is the usual source, adjusted where a susceptibility screen defines a narrower window. If the system cannot store that number, the fleet view cannot be fixed by configuration, only by a spreadsheet outside the system — which is where the station already is.

Short-term, long-term, and where the noise actually lives

The short-term rate divides the loss between the two most recent inspections by the exposure between them. The long-term rate divides the loss from baseline to the current reading by the total exposure since baseline. Remaining life is projected on whichever is higher. The mechanics are simple; the interesting part is what the two rates are made of.

Consider a hand-scanned grid where point-to-point repositioning between outages is realistically 0.005 to 0.010 inch, even though the instrument displays to 0.001 inch. Over one eighteen-month interval, 0.010 inch of repositioning error becomes roughly 0.0067 inch per year of pure noise in the short-term rate. Over a twenty-year baseline, the same 0.010 inch contributes about 0.0005 inch per year. The short-term rate is where almost all of the uncertainty sits, and the long-term rate is comparatively quiet. Taking the higher of the two therefore converts measurement noise into conservatism on a routine basis.

That is safe for remaining life and expensive for outage scope, because a noise-driven short-term rate can push a component onto a repair list it does not belong on. The resolution is not to abandon the governing rule but to carry an uncertainty band with each rate. When the short-term rate exceeds the long-term rate by less than the noise band, the correct response is re-inspection or a wider grid, not a repair package. A module that stores a rate as a bare number cannot make that distinction, and the engineer ends up re-deriving it by hand every outage.

The baseline argument, and why measured-versus-nominal is a screening tool

Seamless pipe is permitted a substantial under-tolerance on wall thickness. A component delivered near the low end of that tolerance will read below schedule nominal on the day it is installed. Compute a long-term rate as nominal minus current, divided by years in service, and the result includes manufacturing tolerance as if it were corrosion. On a thin-walled small-bore line this can be the majority of the reported loss.

The defensible baseline is the first grid inspection at that location, treated as an as-built measured value, which makes the long-term rate a measured-versus-measured calculation. That baseline is itself a measurement with its own error, and for older units it may not exist at all — in which case measured-versus-nominal is retained as a screening number, clearly labelled as such, and never used as the basis for a fitness-for-service evaluation under Code Case N-597-2 without qualification.

Replacements complicate this further. A repair or replacement performed under ASME Section XI IWA-4000 creates a new component at an old location, often in a different material. If the spool goes in as a chromium-bearing alloy, the FAC susceptibility of that location has changed fundamentally and the prior history is not merely a different baseline, it is a different problem. The module needs a first-class reset event that starts a new baseline, records the new material, and keeps the prior history visible without letting it into the rate.

Grid statistics: what the band actually reports

A FAC inspection does not produce a thickness. It produces a set of readings on a defined grid within a defined band, and a rule for turning that set into one governing value. The common rules are the single lowest reading, the mean of the lowest few after outlier rejection, or a statistical treatment that reports a lower bound at a stated confidence. Each answers a slightly different question, and each has a different sensitivity to a single bad reading.

The single-lowest rule is the most conservative and the noisiest, because the whole band rate is then driven by one point that may have been taken slightly off the grid location, on a scan mark, or over a weld crown. Averaging the lowest few suppresses that but risks smearing a genuine local thin spot into its neighbours. Neither is wrong. What matters is that the fleet knows which rule produced a given number, and that the underlying point data is preserved so a different rule can be applied later without a re-inspection.

This is the most common data loss in a legacy corrosion database: the reduced number is stored and the grid is not. Once the individual points are gone, the station cannot re-analyse an old inspection under a revised procedure, cannot re-examine a suspicious rate, and cannot answer the question of whether an apparent step change was a real event or one poor reading. Storing the full grid, its coordinate origin and its band definition costs almost nothing and is the difference between a database and a report archive.

Negative wear, and what it means when a fleet handles it two ways

Grid points routinely report more metal at the second inspection than at the first. Steel does not grow. The reading pair reflects a probe placed a few millimetres away from where it was placed last outage, a different couplant condition, a different technician, or a scan surface that has been cleaned since. In a well-run programme, negative results appear on a predictable fraction of points and are a useful indicator that the grid is being reproduced within tolerance.

What the programme does with them matters more than most people expect. Clamping each negative point to zero before averaging the band raises the band mean, and does so more at low-wear locations than high-wear ones, which is precisely where a fleet comparison is most sensitive. Retaining the sign and averaging gives an unbiased band estimate and lets negative band results stand as evidence that wear is below the measurement floor.

Across a fleet where one station clamps and another does not, the same physical component produces different rates, and the difference is largest exactly where you are trying to distinguish a good performer from an average one. The module must therefore never clamp silently. It should preserve raw values, apply the site's declared rule as a labelled transformation, and be able to report both the treated and untreated result on demand.

What the Appendix B record has to carry

Where results feed the safety-related programme, 10 CFR 50 Appendix B applies to the calculation and not only to the inspection. Criterion V requires the work to follow a documented procedure. Criterion XVI ties an exceedance to the corrective action programme. Criterion XVII governs the records, and 10 CFR 50.9 makes completeness and accuracy of information a licensing obligation rather than a housekeeping preference.

In data terms this translates into one demanding requirement: a rate computed in 2019 must still be reproducible in 2031 from stored inputs, even though the calculation engine has been revised twice in between. That means versioning the engine, storing the version with the result, and never recomputing history in place. A system that silently re-derives all historical rates when a rule changes has quietly invalidated every condition report those rates generated.

The corrective action interaction is where this bites first. A rate crossing a threshold generates a condition report, and that report is a record of the value at the moment of the decision. If a subsequent reading changes the rate and the system overwrites it, the condition report now references a number that no longer exists anywhere in the database. The rate that triggered the action has to be frozen against the action, with the current rate shown beside it.

Evaluating a corrosion rate module for fleet use

Ask for a demonstration, not a feature list, and bring two real datasets from two stations that currently disagree. The single most informative test is to load both and ask the system to explain the delta by named convention: how much of the gap is time base, how much is baseline source, how much is grid reduction, how much is negative-wear handling. A system that can decompose the difference is one that can eventually close it. A system that just produces a third number has added a fourth opinion.

Then work through the storage questions, because they are irreversible. Does it accept operating hours per interval as an input rather than deriving them from dates? Does it store every grid point, or only the reduced band value? Does it hold the procedure and revision with each result? Does it support a dated reset event that starts a new baseline while preserving prior history? Does it record the material at the location, so that an alloy replacement is visible in the data rather than only in a work order? Does it freeze a rate against the condition report it triggered?

Finally, test the negative case deliberately. Feed it a grid with negative wear on three points and see what it does without being told. Feed it a location with a replacement recorded and confirm the long-term rate restarts. Feed it a measured-versus-nominal case where the component was delivered thin and see whether the result is labelled as a screening value. What a module does with awkward data, unprompted, tells you more about whether a fleet can standardise on it than any amount of configuration documentation.

Request a consultation