Making Thickness Data Roll Up Across Four Fabrication Shops

The two rates, and why one of them is not optional

Every thickness reading a shop takes belongs to two calculations at once. The short-term rate divides the loss between the previous inspection and the current one by the interval between them. The long-term rate divides the loss from the as-built baseline by the total service age. API 570 and API 510 both expect the owner-user to have both numbers and to use the one that best represents the damage the component is actually experiencing. A defensible system computes both, presents both, and lets the higher one drive remaining life unless an engineer records a reason to override it.

The two rates diverge because they answer different questions. The long-term rate averages over the whole life, so it dilutes a recent process upset across twenty years of clean service. The short-term rate is sensitive to exactly that upset — and equally sensitive to measurement noise. A gauge with 0.001 in resolution read across a two-year interval carries a noise floor near 0.0005 in/yr before any real metal has moved. On a component genuinely losing 0.002 in/yr, that is a twenty-five percent error band contributed by the instrument alone, which is why a single short-term rate on a short interval should never drive a decision by itself.

Remaining life then follows as t_actual minus t_required, divided by the governing corrosion rate. That subtraction is where the shop's own work enters the calculation, because t_required comes from the design the shop built to and t_actual depends on the wall the shop actually delivered. Neither of those is the nominal number printed on the drawing, and the gap between the drawing and the metal is where most multi-site roll-ups fail.

Why a fabrication shop's baseline is the hardest number in the calculation

The single most consequential field in a fabrication shop's corrosion record is t_initial, and it is the field most often filled in wrong. Seamless pipe to ASTM A106 or A53 carries a 12.5 percent minus tolerance on wall thickness. A nominal 0.280 in Schedule 40 NPS 6 pipe can leave the mill at 0.245 in and pass every acceptance check there is. If the record says the baseline was 0.280 in and the first in-service reading five years later is 0.248 in, the system reports roughly 0.0064 in/yr of corrosion. The true loss might be nothing at all.

Plate is more forgiving but not free. ASTM A20 governs undertolerance on pressure-vessel plate, and ASME Section VIII, Division 1 permits a small mill undertolerance to be neglected in design. That is a design allowance, not a measurement allowance, and the distinction is routinely lost. A rule that says you may ignore 0.01 in when sizing a wall does not mean your corrosion-rate baseline may ignore it. That same 0.01 in, treated as metal loss across a four-year interval, manufactures 0.0025 in/yr of corrosion that never happened.

The remedy is not complicated, only unpopular: measure the as-built thickness at each CML before the item ships, and record which convention the baseline came from. A system that stores one unlabelled baseline number cannot be audited and cannot be handed to a client's integrity programme without a caveat. A system that stores baseline value, baseline basis, baseline date and the source document is portable, and portability is what a multi-shop fabricator is actually buying.

Four shops, four conventions, one impossible roll-up

The standardisation problem is rarely that one location is wrong. It is that each location is internally consistent and mutually incompatible. One shop numbers CMLs sequentially per spool. Another numbers them per line, restarting at each isometric. A third uses the client's tag plus a clock position. When all three feed a common register, the same physical location appears under three identities, and the corrosion rate calculated for each is built from a fragment of the history that exists.

The second divergence is orientation and statistic. A CML on horizontal piping is normally read at four quadrants — 12, 3, 6 and 9 o'clock. If one shop stores all four readings and another stores only the lowest of the four, the two datasets are not comparable. The 6 o'clock reading on a line carrying entrained water will corrode several times faster than the 12 o'clock reading, so averaging a four-quadrant set against a min-of-four set simultaneously understates the aggressive location and overstates the benign one, inside the same roll-up.

Units are the least interesting and most persistent failure mode. A shop working in millimetres per year and a shop working in mils per year will both produce plausible-looking reports, and 0.1 mm/yr sitting next to 0.1 mpy is a factor-of-four-plus error that no reviewer catches by eye. Storing every rate in a single canonical unit and rendering per user preference removes an entire class of error that training never fully removes.

Negative rates, hot readings and the disposition nobody records

Ask a group of shop QC leads how they handle a thickness reading that comes back thicker than the last one and you will get four answers: record it, discard it, zero the rate, or re-read until it agrees. All four are common practice and only one is correct in any given circumstance, so the system has to know which was applied. A genuinely thicker reading usually means the probe was not on the same spot, couplant film was measured along with the wall, or the earlier reading was taken at temperature without velocity correction.

Temperature is the routine culprit and it is systematically misunderstood. Longitudinal velocity in carbon steel falls roughly one percent per 55 °C (100 °F) rise, so a gauge calibrated on a room-temperature block and used on a line at 260 °C (500 °F) reads approximately four to five percent thick — about 0.013 in on a 0.300 in wall. Against a cold baseline that presents as an apparent thickness gain, and the operator who resolves it by re-reading until the number drops has just discarded the only honest measurement in the set.

Silently clamping negative rates to zero is the worst available option, because it conceals the calibration problem that generated them. Keeping the reading, flagging it, and requiring a disposition — re-read, reject with stated cause, or accept as measurement scatter — converts a nuisance into a calibration audit trail. Across four shops the pattern in those dispositions identifies the location with a probe or procedure problem long before a client audit does.

Corrosion that happens in your own yard

Fabricators do not usually think of themselves as owning corroding assets, but a spool that has been hydrotested and then stood in a coastal yard for eleven months is corroding, and the client will find it. Test water left in a low point supports microbially influenced corrosion under sludge, and the pitting it produces is local, deep, and effectively invisible to a four-quadrant thickness survey designed to detect general wall loss. The general-corrosion rate reports a healthy wall right up to the point the spool leaks in service.

This is where a corrosion-rate module earns its place inside a shop rather than a plant. The record that matters is not a rate at all — it is a preservation clock: hydrotest date, drain and dry-out verification, nitrogen blanket or desiccant status, end-cap integrity, and days in yard. Most specifications cap chlorides in test water for austenitic stainless steel at around 50 ppm precisely because the mechanism there is not general corrosion; it is chloride stress corrosion cracking, which leaves no thickness signature until it leaks.

A shop that tracks preservation state alongside thickness has an answer when a client asks why a two-year-old spool shows pitting. A shop that tracks only thickness has a corrosion rate that says the wall is fine and a leak that says otherwise. The two are not in conflict — the general-corrosion rate was simply never the right instrument for that damage mechanism, and the system should make that visible rather than reporting a comfortable number.

Client-owned equipment in for repair: whose corrosion rate is it?

When a client vessel arrives for an R-stamp repair under NBIC NB-23, the shop takes the thickness survey but the owner-user owns the integrity decision. The shop's data therefore has to land in a form the owner's API 510 programme can consume without re-keying: CML identity that maps to their register, baseline basis stated explicitly, readings tagged with instrument, probe, calibration block and technician certification level per ASNT SNT-TC-1A, and the date on which the repair changed the metal.

That last item is the one most shops omit and every owner needs. A patch plate installed in the shop resets t_initial at those locations and only those locations. If the owner's system later computes a long-term rate from an original 2009 baseline to a 2026 reading on 2019 metal, it produces a corrosion rate for a plate that was absent for ten of the seventeen years. The shop is the only party that knows exactly which CMLs were affected, and the handoff is the only moment when capturing that knowledge is cheap.

This is also the commercial argument for standardising, and it is stronger than the internal one. A fabricator whose four shops hand over an identical structured record — same fields, same units, same baseline convention, same certification metadata — is materially easier to work with than one whose data quality depends on which location won the job. Owners notice that, and it shows up in repeat work rather than in a marketing claim.

How to evaluate the module

Evaluating a corrosion-rate module for a multi-site fabricator is mostly a matter of asking what the software refuses to do. Will it accept a thickness reading with no baseline basis? Will it accept a CML with no parent component and no material specification? Will it compute a long-term rate straight through a recorded repair date? A system that accepts everything lets each shop keep its habits and produces a roll-up that averages incompatible data into a number nobody can defend in front of a client.

Ask to see the calculation exposed rather than the result. For any remaining-life figure on screen, the system should show both rates, the readings each was built from, the interval in days, the baseline record with its basis, and the identity of whoever accepted a flagged reading. If any of that has to be reconstructed on demand or dug out of an export, the audit trail does not exist — there is only a number and a recollection.

Then put the migration question directly, because a fabricator with four shops has four spreadsheets. The right answer is not a bulk import. It is a mapping exercise per shop covering CML identity, unit, baseline basis and reading statistic, with the unmapped remainder quarantined and visible rather than silently defaulted. Data imported against a guessed baseline is worse than data you know is missing, because six years from now it will look exactly like data you measured.

Standardising without stopping production

The practical sequence is baseline convention first, CML identity second, reading statistic third, units last. Units convert retroactively; baselines do not. Choose one convention across every site, write it into the shop QC procedure so it survives staff turnover, and mark each historic record with the convention it was actually taken under rather than rewriting the past to match the new rule.

Expect the first honest roll-up to look worse than the spreadsheets did, and brief management before it happens. When four shops stop averaging min-of-four against mean-of-four, the aggressive locations surface and the reported corrosion rates go up. A leadership team that reads that as a software defect abandons the exercise in month three. A leadership team told in advance reads it as the first comparable number the group has ever had.

Run the new system in parallel with the spreadsheets for one full inspection cycle at a single shop, not all four. The test is not whether the numbers match — they will not — but whether every difference can be explained. Differences you can explain are evidence the migration worked. Differences you cannot are unmapped data, and finding them at one site costs a fraction of finding them at four.

Request a consultation