AI in non-destructive testing: what AI adds to NDT

Non-destructive testing produces images and signals: digital radiographs, ultrasonic scans, thermal sequences, eddy-current traces. Artificial intelligence replaces none of these methods and signs no verdict. It adds a layer of computer-aided analysis wherever a digital record exists: flagging and pre-classifying indications, offering a consistent second reading, structuring results into a searchable history, and tracking how an indication evolves from one campaign to the next. This article extends the panorama of non-destructive testing methods: it does not re-explain the techniques, it sets out what AI concretely adds to them, and where it stops.

Detecting and classifying defects on images, method by method

A vision model works only on a digital record. That splits NDT methods into two straight away: those that produce a usable image or signal, and those that stay manual. It is on the first group that AI has most to offer, and its contribution differs by method.

• Digital radiography (RT). On weld radiographs, convolutional neural networks flag porosity, lack of fusion and inclusions, and propose a pre-classification of the indication type. The precondition is a digital acquisition, never a silver-based film, and a set of images labelled by qualified readers.
• Phased array and TOFD ultrasonics (UT). Models help segment the scans, tell a geometric echo from a genuine reflector and assist the sizing of an indication. They call for position-encoded, properly calibrated acquisitions.
• Infrared thermography. On a thermal sequence, AI brings out disbonds and delaminations by subtracting the background and following the cooling dynamics, where the eye sees only a gradient.
• Array eddy current. On heat-exchanger tube bundles, pattern recognition on the impedance signals flags the suspect tubes to examine first.

The common thread is a digital acquisition that is repeatable and traceable. On a penetrant, magnetic particle or visual examination carried out with no image capture, a model has almost nothing to work on, even if photo-assisted visual inspection is starting to change that.

Assisted interpretation and second reading

Beyond flagging, AI acts as a constant second pair of eyes. Human reading drifts with fatigue, volume and the end of a shift; a model applies the same criterion to the first image as to the nine-hundredth. Two arrangements coexist. In the first, the model pre-screens and the inspector confirms each retained indication. In the second, the inspector reads first and the model passes back over the work to catch a miss, on the double-reading principle used in other imaging fields.

The real gain is consistency and throughput on large campaigns, for instance thousands of weld radiographs to examine, not the replacement of judgement. A known risk comes with this comfort: automation bias, the tendency to trust the assistant too readily. That is why the inspector keeps control of every decision, including the decision to set aside a model's proposal without having to justify it to the tool.

Digitising and structuring results into a usable history

This is the area where AI adds value without touching a single image. The normal output of an NDT test is a document: a test certificate, a penetrant report, a thickness map, a radiograph, all filed as PDFs, held by different contractors, year after year. The information exists, but it stays impossible to find at the moment you need it.

Document-reading models, combining optical character recognition and extraction, read these reports and pull out structured fields: equipment tag, method used, date, indication type, location, size, accept-or-reject verdict. The pile of PDFs becomes a searchable base linked to each item of equipment. This foundation is exactly what thickness monitoring locations and a risk-based inspection plan need, since both are worth only as much as the history they can draw on.

Tracking indications over time

A single test is a snapshot. Value builds through comparison: is this indication new, stable or growing? AI helps to align repeated acquisitions on the same weld, the same measurement location or the same tube, from one year to the next, then to surface the deltas: a porosity flagged again in the same spot, a rate of metal loss accelerating from one campaign to the next.

This tracking feeds directly into remaining-life estimation and inspection intervals. But it rests on one condition: structured, comparable results, because scattered PDFs cannot be compared. Documentary structuring is therefore the precondition for trending. This is where a documentary intelligence tool for inspection reports such as Integrity Loop fits: it reads and structures existing NDT reports in PDF to rebuild the history of an item of equipment. It does not carry out the defect analysis, it is not an NDT interpretation engine, and it does not perform predictive maintenance: it simply unlocks the data trapped in the reports so it can be tracked over time.

The limits, and the principle that does not move

None of these aids is neutral. Four limits recur, and ignoring them costs more than doing without.

• False calls and missed defects. A model sometimes takes a geometry for a defect, or lets an atypical anomaly through. The false-call rate and the probability of detection are measured, not assumed.
• Data hunger. A model needs representative examples, labelled and validated. Yet real defects are rare: training sets are often small and imbalanced, and a model trained on one material, geometry and procedure does not transfer as-is to another.
• Out-of-distribution use. A new welding procedure, a different probe, another material, and the model is working outside the envelope in which it was validated, without always flagging it.
• Traceability. A discipline governed by codes calls for an explainable, auditable decision; an opaque model output is a problem, not a convenience.

Above all this sits a simple principle. AI is a tool the qualified inspector uses, in the same way as a flaw detector. It is the inspector certified to Level 2 or 3 under ISO 9712 who interprets the indications and pronounces acceptance or rejection; inspection codes such as API 510, API 570 and API 653 assign that responsibility to qualified people, never to an algorithm.

A model is validated, moreover, like any NDT procedure: on representative samples carrying known defects, measuring the probability of detection and the false-call rate, documenting the result, and revalidating it as soon as the process changes. Treated as part of the inspection procedure rather than as a black box, it earns its place. This logic, where the tool informs and the human decides, runs through our whole view of industrial AI.

Recap: where AI helps, and on what condition

The table below sums up, method by method, what AI adds, the data it calls for and the limit that remains.

Table scrolls horizontally on small screens.