Moving decades of steel mill thickness history into a system that can use it

What is actually in forty years of thickness books

Before anyone designs a migration, someone has to open the boxes. The record in an integrated steel plant is not one legacy system, it is five or six strata laid down by different eras of practice. The oldest layer is handwritten: thickness books in mils, keyed to sketches drawn by an inspector who knew the plant intimately and wrote locations accordingly — a reference to a walkway, a valve, a building that was demolished in 1994. That layer is often the most careful work in the whole archive and the hardest to use.

Above it sits a database era. Most plants had something: a mainframe module, a departmental Access application, occasionally a commercial package long since unsupported. These extracts look attractive because they are already structured, which is exactly why they are dangerous. They carry tag, date and thickness, and almost never carry instrument, technician, surface condition, grid definition or measurement basis. Structured data of unknown provenance is harder to treat sceptically than a handwritten book, because it arrives looking authoritative.

The recent layer is contractor PDFs, typically fifteen to twenty years of them, often with the raw instrument files long gone. These are usable but pre-rounded, and the rounding was applied by somebody else's report template. Cataloguing what exists, in what precision, with what supporting metadata, is the first deliverable of a migration. Skipping it is what produces a project that appears to succeed and quietly loses the ability to justify any number in it.

Nominal wall is not a baseline

The most consequential silent defect in legacy steel plant data is the substitution of a catalogue nominal for a measured baseline. A line is recorded as twenty-inch standard weight, someone looks up the schedule thickness, and that figure becomes the starting point for every corrosion calculation performed on it for the next thirty years. It is not what left the mill, and mill tolerance runs in the direction that matters.

In practice the error surfaces in two forms. The benign form is a modest understatement of loss because actual as-installed wall exceeded nominal, which produces a conservative rate. The harmful form is a transcription error in the schedule. Record a line one schedule heavier than it is and you introduce roughly a tenth of an inch of apparent wall loss on a large-diameter pipe, which on a slow-corroding gas main is more than a decade of phantom corrosion. That line will be flagged, investigated, possibly replaced, on the strength of an arithmetic error made during data entry.

Migration is the one opportunity to fix this at scale, and the fix is unglamorous. Where a measured baseline exists in the earliest inspection record, use it. Where it does not, load nominal but flag it as assumed, and mark the resulting long-term rate as indicative rather than computed. A rate labelled indicative is still useful. A rate presented as computed when its baseline was guessed is a liability.

Long-term and short-term rates die on import

API 570 draws a deliberate distinction between long-term corrosion rate, computed from an initial thickness against the current reading, and short-term rate, computed from a recent previous reading. The two exist because they answer different questions: the long-term rate describes the service, the short-term rate detects a change in it. Comparing them is how an inspector notices that something has shifted in the last two years, and losing either one halves the diagnostic value of the dataset.

A migration that carries only the most recent reading per point — which is what a fast import scoped as "get current thickness into the new system" produces — eliminates the long-term rate for the entire plant. It cannot be recomputed later, because the source books are usually boxed and disposed of once the new system is signed off. The plant then spends five to eight years rebuilding the ability to compute a number it already had, on lines that are inspected annually.

There is a second, subtler version of the same loss. Where a point has a long series with gaps, some import routines interpolate to produce an even sequence. The interpolated values are then indistinguishable from measured ones, and any subsequent statistical treatment of the trend is contaminated. Load the gaps as gaps. A sparse honest series supports a defensible rate; a dense synthetic one does not.

Replacements break the series, and nobody told the inspection department

Steel plants replace piping. Blast furnace gas mains, coke oven gas lines, pickle line acid headers and cooling water lines all get sections renewed, sometimes on an outage, sometimes on a leak, and the record of that renewal lives in a maintenance work order. The inspection department is not always copied. The result is that the CML register carries a point which continues to accumulate readings across a piece of metal that was replaced in 1998.

The data signature is a step upward in thickness followed by a very low apparent corrosion rate. Neither is questioned reliably. The step is often written off as a bad reading, and the low subsequent rate is a comfortable number that supports interval extension, so it survives review. Years later a line that was assumed good on the strength of a nine-year rate turns out to have twenty-eight years of accumulated service on the sections that were never replaced.

Reconciling inspection history against maintenance work orders is therefore a required step in migration, not an optional enhancement. Every replacement found becomes a break: the old point closes at the replacement date with its history intact, a new point opens with a new baseline, and both remain visible under the same location. This is also the strongest single argument for holding inspection and maintenance in one system afterwards, so the next replacement is visible to the inspection record automatically rather than through an archaeology exercise a decade later.

Coke by-products and gas mains: the points you keep versus the points you need

Migration is also the moment plants discover why their historical point locations are where they are. A large share of legacy CMLs in a steel plant were positioned for access rather than for corrosion likelihood — near a platform, at a convenient elevation, where a ladder reached. That was rational when every reading required a scaffold, and it means the historical population systematically under-samples the locations that actually govern.

In coke oven gas and by-product service the governing locations follow the condensate. Gas leaving the ovens carries hydrogen sulfide, ammonia, hydrogen cyanide and tar, and the aqueous phase that drops out through primary coolers and downstream recovery creates the wet H2S conditions that API 571 associates with hydrogen blistering, hydrogen-induced cracking and stress-oriented hydrogen-induced cracking. Those are examination findings at low points, dead legs and weldments, not thickness readings at accessible elevations. Blast furnace gas mains have their own geography of water drop-out and dust accumulation that governs where thinning concentrates.

The disciplined approach during migration is additive: keep every historical point and its identity intact, because its value is precisely its length of record, and add the points the mechanism argues for as new CMLs with new baselines. What must not happen is renumbering. A migration that tidies up the numbering scheme severs the link between the new register and every paper report, work order and contractor file that references the old numbers, which is a cost paid repeatedly for years afterwards.

Provenance: the reading has to be defensible

The test of a migrated dataset is not whether it renders a chart. It is whether, when an insurer's engineer or a jurisdictional inspector picks a number and asks where it came from, you can answer in under a minute. In a steel plant that question tends to arrive attached to a pressure vessel under state or provincial registration, where National Board Inspection Code NB-23 routes govern repairs and alterations, or attached to a by-product plant unit where inventories of listed substances bring OSHA 29 CFR 1910.119 into play.

Providing that answer requires the migration to carry provenance rather than values alone. Every migrated reading should reference its source — the book and page, the report number, the extract file — and ideally the scanned image of the source document. Scanning is cheap and is only cheap while the boxes are still on site and someone still knows what they contain. Plants that skip it during migration almost never go back and do it.

Provenance also protects the new system's credibility internally. When a migrated series looks wrong, the ability to open the original page and see the inspector's own note is what converts an argument into a five-minute check. Without it, every anomaly becomes a debate about whether the migration was done correctly, and confidence in the whole dataset erodes on the strength of a handful of unresolvable rows.

Reconciliation, parallel running and freezing the old system

A migration should produce a reconciliation report before it produces a live system: how many points existed in each source, how many were loaded, how many were merged as duplicates, how many were rejected and why, and how many long-term rates could be computed against a measured baseline versus an assumed one. That last figure is the honest measure of what you have. A plant that migrates eleven thousand readings and can compute a genuine long-term rate on forty percent of its points knows something useful; a plant that migrated everything and never counted does not.

Run in parallel for at least one full inspection cycle. On an annually inspected plant that is a year, which project schedules dislike, but the alternative is discovering a systematic import defect after the source is archived. Parallel running does not mean duplicating all work — it means keeping the legacy system authoritative for a defined set of points and comparing outputs at the end of the cycle.

Then freeze rather than decommission. Set the old system read-only and keep it reachable. The cost is negligible and the scenario it protects against — needing to substantiate a migrated number years later, to an insurer, a regulator or your own board after an incident — is exactly the scenario in which having the source available is worth more than the entire project budget.

Evaluating a migration, not a product

When this is the job, you are not buying software features, you are buying a migration outcome, and the evaluation should reflect that. Give a vendor a genuinely messy sample: one paper thickness book, one database extract, one folder of contractor PDFs, all covering the same circuit. Ask them to produce the reconciled result and the reconciliation report. What comes back tells you more than any demonstration, because handling that sample well requires understanding the failure modes rather than having an import wizard.

Check specifically for the four capabilities that decide whether the migrated data stays usable: a baseline that can be marked measured or assumed, a series that can be broken at a replacement date with both halves retained, native precision preserved from the source, and provenance carried per reading down to a scanned page. A system missing any one of these will flatten your history into something that looks tidy and cannot be defended.

Atlantis builds the CML and TML registry on Odoo, which means the migrated inspection history sits in the same system as the maintenance work orders that record replacements, the equipment register, the calibration record of the instrument used and the certification status of the technician who took the reading — which is precisely the cross-referencing that migration exposes as missing. The approach is affordable, accessible and fully customisable to the numbering scheme your plant already uses, because renumbering is the one thing a migration must not do. To discuss a reconciliation of your own records, request a consultation or a scoped quote at info@atlantisndt.com.

Request a consultation