AI Tank Inspection and Corrosion: The Data It Needs First

Can AI prioritise tank inspections and flag corrosion from your data?

AI can rank a tank terminal's inspections and flag corrosion once drone imagery, ultrasonic thickness readings and past inspection reports sit in one dataset. PGS 29 allows terms up to 20 years under RBI, and the data decides the term.

Yes, once the data sits in one place: drone and crawler imagery, ultrasonic thickness readings per plate, the inspection reports of the last two terms, the tank register and the CMMS work orders. With that set a model ranks which tanks to open first and marks the plates that lost the most steel. It does not replace the inspection; it decides where the inspection goes. This guide says what the data must contain, how it fits PGS 29, EEMUA 159 and API 653, and what it costs.

By Tom Joseph · Last updated: 14 October 2026

What an AI integrity layer does that the inspection calendar cannot

Ranks the tank list. Every tank gets a score from its last UT readings, the corrosion rate since the previous internal inspection, the product service (product class and climate code as EEMUA 159 table B1 uses them), age, lining, cathodic protection and the settlement measurements. The integrity engineer sees which five tanks to open first, and why.

Reads the imagery. A drone or a crawler brings back hundreds of frames per tank. A vision model marks coating breakdown, pitting, weld-seam corrosion and floor-plate indications, so the inspector reviews the marked frames instead of all of them. Vopak contracted Terra Inspectioneering in 2020 for drone-based confined-space inspection, including ultrasonic thickness testing and deformation analysis, across 11 sites in Belgium and the Netherlands under EEMUA 159 and API 653 (Tank Storage Magazine).

Turns readings into a corrosion rate per plate. UT grids from successive inspections are aligned to one plate map, so the model reports metal loss per shell course, annular plate and floor plate and the remaining life against the EEMUA 159 or API 653 minimum, not one average per tank.

Feeds the CMMS. The ranked list and the flagged indications become inspection tasks and work orders in SAP PM, IBM Maximo or Ultimo, with the frames and readings attached, so the planner, the integrity engineer and the NL-CBIT look at the same evidence.

Which data source yields which answer

Data source	What it yields	Where it usually lives
Drone and crawler imagery (shell, roof, floor, annular plate)	Marked indications: coating breakdown, pitting, seam corrosion, deformation	The inspection contractor's delivery, a file share, an inspection platform
Ultrasonic thickness readings (UT grid per plate)	Corrosion rate per plate, remaining life to the minimum thickness	Inspection reports as PDF, contractor spreadsheets
Historical inspection reports (NL-CBIT, EEMUA 159, API 653)	The trend over two or three terms, repair history, earlier findings	Document management, often PDF only
Tank register and tank data sheet	Design standard (NEN-EN 14015 or API 650), as-built plate thickness, foundation, lining, cathodic protection	Engineering archive, CMMS equipment master
CMMS work orders (SAP PM, IBM Maximo, Ultimo)	Repairs, coating jobs, leak reports, cost per tank	The CMMS
Product service history	Product class and climate code for the EEMUA 159 term, aggressive or inert service	Terminal management system, tank register
Settlement and floor-profile measurements	Differential settlement, floor deformation, edge settlement	Survey reports, inspection reports

The data and systems it needs

Start from the tank register, not from the model. Per tank it must state the design standard (NEN-EN 14015 for tanks built in Europe since 2005, API 650 for most older ones), the as-built thickness per shell course and per floor plate, the lining, the cathodic protection, the foundation type and the product service. PGS 29 ties the calculation method to that design standard: EEMUA 159 methodology by default, API 653 for tanks designed to API 650 (vs 3.7.9). A model that does not know which standard a tank was built to cannot compute remaining life.

Then three joins. The UT grids of at least two inspections must land on the same plate map, so a reading in 2016 and a reading in 2026 describe the same square metre of steel. The imagery must carry position: which shell course, which plate, which weld. And the inspection reports must be read out of PDF into fields: finding, location, severity, repair, next term. Most of the audit is this extraction, because NL-CBIT reports, contractor spreadsheets and drone deliveries arrive in three formats.

The systems are known. Maintenance runs in a CMMS: Stolthaven standardised maintenance across ten terminals on Ultimo after moving to one system in 2021 and now runs AI functions inside it (IT Brief, March 2026); SAP PM and IBM Maximo are the other two you meet on chemical sites. Inspection data is moving into dedicated platforms: Vopak signed in October 2025 for worldwide deployment of Falcker's tank inspection platform, a site explorer, condition monitoring and a life cycle planner (Tank Storage Magazine), and Evos Hamburg launched DigiTank in January 2025, a three-year IHATEC II research programme combining digital twins, AI and robotics for tank storage (Hydrocarbon Engineering). The AI layer reads from those systems and writes back into them; it is not a fourth place to look.

The regime decides the governance. A Seveso III site (in the Netherlands under the Besluit activiteiten leefomgeving since 1 January 2024, formerly Brzo 2015) must have its inspection terms procedurally secured, PGS 29 requires ultrasonic plate-thickness measurement of wall and floor at every internal inspection, and an RBI term needs an RBI methodology approved by an NL-CBIT (vs 3.7.6). The model's ranking is evidence in that review. The term is still set by the inspection body.

What AI tank inspection does not fix

A camera measures surface condition, not metal loss. Imagery tells you where to measure; only the UT reading says how much steel is left. A model trained on imagery alone misses under-coating corrosion and soil-side floor corrosion entirely, which is where the leaks are. Treat the vision model as a triage of frames, never as a thickness measurement.

Two inspections ten years apart with a different UT grid spacing cannot be aligned to a plate, and the corrosion rate is then a guess dressed as a number. The audit checks alignment on five tanks before anyone talks about a model. The same goes for a tank register with the lining, the cathodic protection or the last floor repair missing: those fields move the term under EEMUA 159 table B1 and the model cannot infer them.

It does not change the term and it does not inspect. The out-of-service inspection, with its draining, cleaning, gas-freeing and entry, stays; the point is to open the right tank in the right year and to hold the long term with evidence. A robotics vendor puts the average saving per deferred internal inspection at 400,000 dollars (Square Robot, a vendor figure); treat that as a claim to test on your own cost per outage, not as a result. And Crux Digits builds the data layer and the models alongside your ICT partner and your inspection contractor; the drones, the crawlers and the NL-CBIT remain theirs.

A terminal with 60 tanks: assumptions stated, cost stated

Take a terminal with 60 atmospheric tanks under PGS 29, 45 on a time-based ten-year term and 15 on an RBI term, about six internal inspections a year, UT grids from the last two terms as contractor spreadsheets, inspection reports as PDF, drone imagery of shell and roof for 20 tanks from the last three years, and maintenance in SAP PM. These are assumptions for the illustration, not a client's numbers. The question the integrity manager asks is which six tanks to open next year and whether three of the RBI tanks can hold their term with evidence.

Three fixed steps. The €2,500 AI audit (one to two weeks) inventories the data per tank, tests UT-grid alignment on five tanks and reads one term of reports into fields; you keep the per-tank data inventory and the readiness list, whatever you decide next. The €20,000 Production-ready MVP (four to six weeks) builds the ranking on 20 tanks' back data and the frame triage on the drone imagery, scored against what the inspectors actually found in the last term; you keep the code, the models and the aligned dataset. Production from €50,000 (eight to twelve weeks) connects it to SAP PM and to the inspection contractor's delivery, with the integrity engineer approving every term proposal. All excl. VAT, code and models handed over, and the audit is credited in full if you continue within 60 days. See computer vision, machine learning, data engineering, AI for energy and tank storage and AI implementation in the Netherlands.

Frequently asked questions

Can AI detect corrosion on a storage tank from drone images?

It can mark coating breakdown, pitting, seam corrosion and deformation in the frames, so the inspector reviews the marked frames rather than every frame. It cannot measure metal loss; that is the ultrasonic thickness reading, and a model that joins the two per plate is what makes the ranking defensible.

Does it change our EEMUA 159 or API 653 inspection intervals?

Not by itself. Under PGS 29 the time-based term follows EEMUA 159 table B1 and an RBI term needs a methodology approved by an NL-CBIT, with a maximum of 20 years, 25 with approval. The model's corrosion rates per plate are evidence in that review; the inspection body sets the term.

What data do we need before anything can be built?

The tank register with design standard, as-built thickness, lining and cathodic protection; UT grids from at least two inspections that can be aligned to one plate map; the inspection reports of those terms; the imagery with position; and the CMMS work orders. The audit tells you per tank what is present and what is missing.

Does it work with SAP PM, IBM Maximo or Ultimo?

Yes. The ranked list and the flagged indications become inspection tasks and work orders in the CMMS you run, with the evidence attached, and the model reads repair history back from it. The integration is scoped in the audit, together with your ICT partner.

What does AI tank inspection cost?

At Crux Digits: a €2,500 audit on the data of your tanks, a €20,000 Production-ready MVP on 20 tanks' back data, production from €50,000 integrated with the CMMS and the inspection contractor, all excl. VAT, with code and models handed over. The audit is credited in full if you continue within 60 days.