BRENNEN

Industries / Industrial equipment and services / Industrial services / Non-destructive testing

Visual and drone inspection

Visual inspection is looking at a surface for cracks, corrosion and damage, by eye or through cameras, borescopes, drones and crawling robots; this domain also covers cameras that see heat and lasers that show how a surface moves.

6

Milestones achieved, from 1965 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1965

    A camera that sees heat

    Thermography finds faults by the heat they give off or hold in, such as an electrical joint that runs hot or a gap in insulation. In 1965 the Swedish company AGA built its model 650, which a history by FLIR, AGA's successor, calls "the first camera for predictive maintenance". It was a large unit that used an oscilloscope for a screen, and utilities, among the first users, mounted it on vans to check substations. The AGA 750 of 1973 was the first to run on batteries.

    irinfo.org

    Credit

    AGA (later Agema, now part of FLIR), as told by FLIR's Ed Kochanek

  2. 2005

    Lasers check shuttle foam

    After the Space Shuttle Columbia was lost in 2003, NASA put the five most promising methods for finding voids in the foam insulation on the shuttle's external tank through a blind test, and chose two that see below the surface, backscatter X-rays and terahertz imaging. After foam came off during Discovery's flight in July 2005, NASA added laser shearography. It films a laser speckle pattern on the foam before and after sound gently shakes the surface, and shows where the surface moves differently. NASA engineers reported that it found foam crushed about 0.1 inch (2.5 mm) deep even after the foam had sprung back to its original shape.

    LessonIn a 9-month shearography inspection of 3,100 square feet of damaged foam on one tank, none of the 21 hidden indications turned out to be real damage, because the team made its calls cautious where it had no experience to go on.

    ntrs.nasa.govntrs.nasa.gov

    Credit

    NASA engineers Joel Richter and James L. Walker; later shearography work by Warren Ussery and colleagues

  3. 2017

    Neural networks find cracks

    A convolutional neural network is an AI model that learns visual patterns from labeled examples. Y. Cha, Wooram Choi and O. Büyüköztürk trained one on 40,000 photos of concrete, each 256 by 256 pixels, to tell cracked from sound surfaces, and report about 98% accuracy. On 55 large photos of a structure it had not seen, taken in strong light and shadow, it found cracks better than two standard edge-finding methods. Semantic Scholar counts more than 3,000 papers citing the work.

    api.semanticscholar.org

    Credit

    Cha, Choi and Büyüköztürk, in Computer-Aided Civil and Infrastructure Engineering, May 2017

  4. 2018

    Crack opening beats length

    Remote visual testing uses cameras to look where people cannot go, such as the inside of a reactor vessel under water. For the US Nuclear Regulatory Commission (NRC), Pacific Northwest National Laboratory (PNNL), working with the Electric Power Research Institute (EPRI), ran round-robin tests in which teams from commercial inspection vendors scanned specimens without knowing where the cracks were. The width of a crack's opening mattered more than its length. Detection was reliable (80% probability at 95% confidence) for openings above about 25 micrometers (0.001 inch), and cracks near scratches, weld ripples or weld edges were harder to see. Before each test the teams checked their cameras with the usual card of lowercase letters no more than 1.1 mm (0.044 inch) high.

    LessonCrack opening, lighting and surface texture decide what a camera finds, and PNNL called for resolution targets etched onto specimens with realistic surfaces.

    osti.gov

    Credit

    P. Ramuhalli and colleagues at PNNL, for the NRC, with EPRI

  5. 2024

    Tanks surveyed, nobody inside

    A floating production, storage and offloading vessel (FPSO) stores oil at an offshore field, and its cargo tanks must be surveyed for its classification society, the body whose rules keep a vessel insured and working. Inspectors normally climb inside. In July 2024 Geo Oceans, with Shell and its sister companies Sonomatic and AUAV, finished the full Special Survey of four cargo oil tanks on an FPSO in the Gulf of Mexico using only remote robotic methods for the thickness readings and visual checks, and the American Bureau of Shipping (ABS) signed it off. The company calls it the first such survey with no one entering the tanks.

    jpt.spe.org

    Credit

    Geo Oceans, Sonomatic, AUAV and Shell; accepted by ABS

  6. 2024

    Rules for inspection drones

    ASME's standard MUS-1-2024 sets out how unmanned aircraft may be used in industrial inspections: planning and hazard checks, performance requirements for cameras, lighting and sensors, the roles of the pilot and the qualified inspector, and the records to keep. It does not say which flaws are acceptable. The 2025 edition of Section V of ASME's boiler and pressure vessel code lets a remote camera replace the eye only if it is shown to resolve detail and light the surface at least as well, and names resolution charts such as the USAF 1951 target and ISO 12233.

    nationalboard.org

    Credit

    ASME, as presented by Paul Coco of Hartford Steam Boiler to the National Board of Boiler and Pressure Vessel Inspectors in 2026

Watch for

  1. Docked drones on oil platforms

    A "drone in a box" launches from a dock, flies its inspection, lands and recharges with no pilot on site. Shell says it ran such trials in April 2026 from its Mars floating production platform in the US Gulf, and that it is the first operator in the Gulf with Federal Aviation Administration (FAA) approval to authorize its own drone flights beyond the pilot's line of sight from any of its Gulf facilities. It wants the drones to replace helicopter flybys for visual safety checks after hurricanes, so that platforms can restart sooner.

    Achieved whena docked drone works from an offshore platform and Shell or a regulator publishes how many post-storm inspections it flew and how restart times compared with helicopter checks.

    jpt.spe.org

    Claimed by

    Shell, in the Journal of Petroleum Technology, September 2026

  2. Detection curves for AI

    Andrii Kompanets and four colleagues propose judging AI crack finders the way inspection methods are judged, by the probability of detecting a crack of a given size, with image resolution taken into account. Applied to the "Cracks in Steel Bridges" set of annotated photos, they report that their computer vision method is robust and adds value when used alongside conventional visual inspection.

    Achieved whena bridge owner or inspection firm runs the method and its own inspectors on the same structures and publishes both detection curves.

    arxiv.org

    Claimed by

    Kompanets, Sherry, Duits, Leonetti and Snijder, in an arXiv preprint, August 2026

Open

  1. 01

    Crack-width targets for cameras

    Metal test plates with lines and slots of known width, from about 15 to 150 micrometers (µm), cut into surfaces that look like real welds and machined steel, and a public table that gives, for each inspection camera tested (phone cameras, borescopes, crawler cameras and drone payloads), the narrowest line it shows at each distance and light level. Each camera is also checked with the usual ASME letter card and the USAF 1951 chart, so the table shows whether passing those checks predicts seeing a crack-sized line.

    The brief
    Credit · unclaimed
  2. 02

    Scoring an AI second reader

    A blind test on recorded remote visual video of metal specimens with real cracks of measured opening and length. Certified inspectors, an AI crack finder, and inspectors helped by the AI each call every crack site. Each group gets a probability of detection curve and its a90/95, the crack size found 90% of the time at 95% confidence, computed as MIL-HDBK-1823A sets out, along with false calls per hour of video.

    The brief
    Credit · unclaimed

Open 01 · the brief

Crack-width targets for cameras

Credit · unclaimed

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

What to build

Metal test plates with lines and slots of known width, from about 15 to 150 micrometers (µm), cut into surfaces that look like real welds and machined steel, and a public table that gives, for each inspection camera tested (phone cameras, borescopes, crawler cameras and drone payloads), the narrowest line it shows at each distance and light level. Each camera is also checked with the usual ASME letter card and the USAF 1951 chart, so the table shows whether passing those checks predicts seeing a crack-sized line.

Why it is still open

PNNL's 2018 study for the NRC found that crack opening, not length, decided detection, with reliable detection above about 25 µm, and it called for resolution targets etched onto specimens with realistic surfaces. An earlier PNNL study (Cumblidge and colleagues, 2008) ran a limited laboratory test of one commercial camera system on cracks of various widths. ASME's code accepts letter cards and line charts, and ASME MUS-1 sets performance requirements for drone cameras without acceptance criteria. Searches in November 2026 found no public table that ties inspection cameras to the narrowest crack-like line they can show.

osti.govosti.govnationalboard.org

Built with what exists

Lines and slots cut into stainless steel coupons by a laser micromachining service and measured under a calibrated microscope; a USAF 1951 chart; a light meter; a motorized rail to set the distance; the cameras under test, borrowed or rented; Python with OpenCV to measure line contrast; two certified visual inspectors (Level 2 under ISO 9712 or SNT-TC-1A) who call each image without knowing the line width.

  1. First step

    Have a micromachining service cut lines from 15 to 150 µm wide into two stainless coupons, measure each line under a microscope, and photograph them with three phones at 100, 300 and 600 mm.

  2. Done looks like

    At least 20 camera systems tested at five distances and three light levels; the table and every image released under an open license; and the share of camera setups that passed the letter card but missed lines 50 µm wide or wider, published.

Open 02 · the brief

Scoring an AI second reader

Credit · unclaimed

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

What to build

A blind test on recorded remote visual video of metal specimens with real cracks of measured opening and length. Certified inspectors, an AI crack finder, and inspectors helped by the AI each call every crack site. Each group gets a probability of detection curve and its a90/95, the crack size found 90% of the time at 95% confidence, computed as MIL-HDBK-1823A sets out, along with false calls per hour of video.

Why it is still open

PNNL's 2018 round robin measured human teams only, and its specimens and video are not public. Kompanets and colleagues (August 2026) built detection curves for AI on photos of steel bridges, not on inspection video, and Virkkunen and colleagues (2019) compared an AI with human inspectors on ultrasonic data. A 2025 study compared manual crack inspection with inspection helped by a mobile robot running AI crack detection, and reported accuracy and workload, not detection curves. Searches in November 2026 found no such test on remote visual video.

osti.govarxiv.orgarxiv.orgarxiv.org

Built with what exists

Specimens with real fatigue cracks grown by a supplier such as Trueflaw, which grows real cracks on real components with its thermal fatigue process, each crack's opening measured under a microscope; a borescope or underwater camera on a motorized scanner; an AI crack finder trained with Ultralytics YOLO11 or Meta's Segment Anything Model 2; the mh1823 POD software, which runs in the free R statistics system.

  1. First step

    Borrow or buy ten cracked specimens, measure each crack's opening, and record five minutes of borescope video of each to see whether the openings span the 25 µm range where PNNL saw detection fall off.

  2. Done looks like

    At least 60 cracked sites and 60 clean sites recorded under two light settings and called by at least eight certified inspectors; a90/95 and false calls per hour published for inspectors alone, the AI alone and the two together; video and crack measurements released.