BRENNEN

Industries / Industrial equipment and services / Industrial services / Welding and fabrication

Weld inspection

Weld inspection is checking finished welds, by eye and with magnetic powder, X-rays and ultrasound, against the limits a code sets for flaws.

6

Milestones achieved, from 1920s to 2009.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1920s

    Iron powder shows cracks

    In the early 1920s William Hoke noticed that metal grindings held on a magnetic chuck gathered in patterns over cracks in the parts. A crack in magnetized steel bends the magnetic field out of the surface, and fine iron powder piles up there. By the early 1930s, magnetic particle inspection was replacing older oil-and-chalk tests on railroad boilers, wheels and axles, and it is used today to check structural welds on bridges and storage tanks.

    nde-ed.org

    Credit

    William Hoke, as told by Iowa State University's NDE-Ed resource

  2. 1931

    X-rays let vessels be welded

    A 200,000-volt X-ray tube of 1922 made pictures of thick steel parts in a reasonable time, and in 1931 General Electric built 1,000,000-volt X-ray generators. The same year the American Society of Mechanical Engineers (ASME) permitted X-ray approval of fusion-welded pressure vessels, which opened industry to radiography of welds.

    nde-ed.org

    Credit

    ASME; General Electric

  3. 1962

    King's Bridge breaks

    In July 1962, about fifteen months after it opened, one span of King Street Bridge in Melbourne collapsed under a loaded semi-trailer. A royal commission examined the bridge's specifications, design, steel quality and welding practices. It found brittle fractures where cover plates, extra steel plates added to strengthen the girders, had been welded onto their tension flanges, the parts stretched by the load, and it blamed an "inherent defect" in the bridge, not any outside force.

    LessonThe commission traced the collapse to defects built into the welded girders, not to the truck crossing them.

    mmhn.org.au

    Credit

    a failure; investigated by the Royal Commission into the Failure of Kings Bridge (1962–63), as summarized by the Melbourne Maritime Heritage Network

  4. 1976

    Inspectors get certified

    The American Welding Society (AWS) published its standard for certifying welding inspectors, QC1, in 1975 and held the first examinations in the spring and fall of 1976. By its 2007 edition, AWS counted more than 55,000 people certified under it. In 1976 The Welding Institute (TWI) in Britain also began certifying welding inspectors through its scheme CSWIP.

    s3.amazonaws.comtwi-global.com

    Credit

    AWS; TWI and CSWIP

  5. 1996

    Ultrasound may replace X-rays

    Ultrasonic testing sends short pulses of high-frequency sound into a weld and times the echoes; a hidden flaw sends back an early echo. Iowa State University's NDE-Ed resource credits Floyd Firestone (1940) and Simons (1945) with this pulse-echo method. In 1996 ASME's Boiler and Pressure Vessel Code Committee approved Code Case 2235, which allows welds in Section VIII pressure vessels to be examined by ultrasound instead of radiography. Later revisions lowered the thinnest wall it covers from 4 inches to 0.5 inch (12.7 mm) and added flaw acceptance limits worked out with fracture mechanics, the study of when a crack of a given size will grow.

    api.crossref.orgnde-ed.org

    Credit

    ASME Boiler and Pressure Vessel Code Committee; the technical basis was published by Mahendra Rana, Owen Hedden, Dave Cowfer and Roger Boyce

  6. 2009

    Phased arrays, blind-tested

    A phased array probe holds many small ultrasonic elements fired in sequence to steer and focus the beam, and it records the scan. In a joint industry project with nine sponsors, TWI had 10 phased array operators, from TWI, equipment makers and inspection firms, scan eight steel test blocks holding 40 hidden flaws without knowing where they were, giving 400 size measurements. TWI then checked the true sizes with another ultrasonic method and by cutting the blocks where needed. Leaving out two phased array operators and one manual operator who did far worse than the rest, flaws were found 96.6% of the time with phased arrays and 93.8% with conventional manual ultrasound.

    LessonTWI found that sizing errors over 6 mm came from how operators read the data, not from the method, and that operators with more than five years' experience had about 40% smaller random errors on the thinnest plate.

    twi-global.com

    Credit

    Charles Schneider of TWI and Colin Bird of Doosan Babcock

Watch for

  1. Ultrasound watches welds form

    Lack of sidewall fusion is a flaw buried inside a multi-pass weld, where the new weld metal fails to melt into the side of the groove, and it is usually found only after welding ends. Researchers at the University of Strathclyde placed a phased array probe ahead of the torch and recorded images while welding. They report that the echo from the groove wall drops from 60% to 0% just before the wall fuses and returns once the metal solidifies, a sign of whether fusion took place, and they conclude that imaging during welding can predict and confirm fusion quality. Their tests used lab welds and artificial flaws.

    Achieved whenthe method flags real lack-of-fusion flaws in production welds, confirmed by cutting or a second inspection method, in a trial reported by someone other than the developers.

    strathprints.strath.ac.uk

    Claimed by

    Angelos Dimakos and colleagues, University of Strathclyde, in Welding in the World, online November 2025

  2. A cobot hears hidden flaws

    Mobina Mobaraki, Mahyar Asadi, Klaske van Heusden and Guy Dumont trained a model on images and sound recorded from an industrial cobot doing MIG welding of fillet joints, the corner welds that join two plates at an angle. They report that it detects porosity (gas bubbles), lack of penetration and fusion, undercut and cold lap with an F1 score, a combined measure of misses and false alarms, of up to 0.99. All the data came from one robot.

    Achieved whenthe model, or one like it, keeps its score on welds from a different robot cell, with flaws confirmed by X-ray or by cutting, in a test run by an outside group.

    arxiv.org

    Claimed by

    the authors, in an arXiv preprint, September 2026

Open

  1. 01

    Ultrasound scans with cut truth

    An open set of raw phased array ultrasonic scans of welded steel plates with flaws made on purpose, such as lack of fusion, slag and porosity, in which every flaw is afterward cut through, etched and photographed so that its true height and position are known. A scoring script lets any person or program size the flaws from the scans and see its error.

    The brief
    Credit · unclaimed
  2. 02

    One model, two welding cells

    A test of whether a model that spots weld flaws from sound, video and welding current, trained on public recordings from one robot cell, still works on a different cell. Record 300 welds on a second robot cell, label each by inspection, score models trained on the public sets, and publish how much the score drops and how many new labeled welds it takes to recover.

    The brief
    Credit · unclaimed

Open 01 · the brief

Ultrasound scans with cut truth

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open set of raw phased array ultrasonic scans of welded steel plates with flaws made on purpose, such as lack of fusion, slag and porosity, in which every flaw is afterward cut through, etched and photographed so that its true height and position are known. A scoring script lets any person or program size the flaws from the scans and see its error.

Why it is still open

TWI's blind trial found that the largest sizing errors came from operators misreading the data, but it published statistics, not the scans. The Submerged Arc Welding Open Repository, collected by the research center AIMEN for the PENELOPE project (Zenodo, March 2025), gives phased array files, X-ray and ultrasonic reports and defect locations found by testing, not by cutting the welds open. No public set that pairs raw phased array scans with cut sections was found in November 2026.

twi-global.comzenodo.org

Built with what exists

Test plates welded with planned flaws by a welding school; a rented portable phased array instrument that records full scans and can do total focusing (a method that builds a sharp image from every element pair), such as an Evident OmniScan X3 or a Sonatest Veo3, run by a qualified Level 2 operator; a band saw, polishing and etching done at a materials lab; Python with NumPy for the scoring script.

  1. First step

    Ask a welding school to make three 25 mm plates with planned lack-of-fusion flaws, and a local inspection company to scan them before they are cut.

  2. Done looks like

    20 plates holding 60 flaws, with all raw scans and at least 300 section photos released under an open license, and the sizing errors of at least two human operators and one automated method published against the cut sections.

Open 02 · the brief

One model, two welding cells

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test of whether a model that spots weld flaws from sound, video and welding current, trained on public recordings from one robot cell, still works on a different cell. Record 300 welds on a second robot cell, label each by inspection, score models trained on the public sets, and publish how much the score drops and how many new labeled welds it takes to recover.

Why it is still open

Two public sets of welding recordings exist, Intel Labs' Robotic Welding Multimodal Dataset on Hugging Face (2024, access by request) and the therness set (February 2026), which holds 2,330 labeled runs with sensor files, audio and video, but each scores models on data from its own cell. The closest transfer work moved a model between TIG and laser welding, two different processes (Li and colleagues, arXiv, June 2026), and the 2026 cobot study above tested on its own robot's data.

huggingface.cohuggingface.coarxiv.orgarxiv.org

Built with what exists

The two public datasets; a robot MIG welding cell at a school or shop, run by its staff; a microphone, a camera behind a welding filter and a current and voltage logger; PyTorch; labels from a certified welding inspector's visual check, with a sample of welds cut and etched.

  1. First step

    Request access to Intel's dataset, download therness, train a baseline model, and record 20 welds on a school's robot cell.

  2. Done looks like

    300 labeled welds from the second cell released, and for at least two models trained on the public sets, the F1 score on the new cell before and after fine-tuning on 30 of its welds, published with the code.