Skip to content
Beaconby Winda
FactInspection, Testing and Monitoring · Robotics and drones

How is AI used in industrial inspection?

·7 sources

Short answer

Artificial intelligence (AI) in inspection means machine learning models that find and classify defects in images, radiographs and ultrasonic signals. Computer vision is the image-based branch. Published work uses these models mainly as an aid to human inspectors, and their reliability has to be shown with NDT measures such as probability of detection.

By , Editor·Researched and drafted by ·Checked by , Editor·Last checked ·Next review

Why does it matter?

A high accuracy score is not proof that an AI tool finds flaws reliably. A 2022 study of deep learning on weld radiographs notes that industrial acceptance needs performance shown with standard NDT measures such as probability of detection (POD), and that these were not commonly used in earlier studies 2. A 2023 review from Pacific Northwest National Laboratory (PNNL) on machine learning for ultrasonic weld inspection called for ways to quantify and validate performance, including common data sets 1. The US National Institute of Standards and Technology (NIST) adds that accuracy figures should include false positive and false negative rates and be tested on data that represent real conditions of use 4.

What does AI do in inspection?

Machine learning is applied at several steps of non-destructive testing: cleaning up raw data, detecting flaws, characterizing them and measuring properties 3. Most published work trains a model on labelled examples, then lets it mark likely flaws for review.

Inspection dataWhat the model is trained to doPublished example
Photos and drone images of structuresFind and classify corroded areas (computer vision)Deep learning models on drone images of a telecommunication tower classified corrosion more accurately than earlier approaches 5
Digital radiographs of weldsDetect, outline and rate flawsAn aerospace weld study reached a flaw size of 0.6 mm detected with 90% probability at 95% confidence (a90/95) and was trialed as an inspector's aid in production 2
Ultrasonic signals and imagesRemove noise, detect and classify weld flaws, estimate sizeReviews cover nuclear plant weld inspections and ultrasonic testing generally 1,3
Thermal images of insulated pipingFlag likely wet insulationA 2022 review describes patented systems combining infrared images and machine learning for corrosion under insulation 6

How is an AI inspection model trained and tested?

Data. Real flaws are rare, so the shared problem across studies is too little training data 3. Researchers add data by augmentation (altered copies of real samples), by transfer learning from related tasks, and by simulation 3. In the radiography study, adding virtual flaws improved detection, most of all for small data sets and rare flaw types; with less data, the detectable flaw size grew and results varied more between training runs 2.

Testing. The radiography study measured POD and a false call rate per unit length of weld, the same terms used for human inspection 2. A University of Bristol review proposes that models used for detection report POD, false alarm probability or similar curves 3. NIST asks for test sets that represent the expected conditions, with the test method documented 4.

Scope. The Bristol review holds that a model should be trained for a defined domain: the inspection method, the structures and materials, and the expected environment, because one model for all conditions is not realistic 3. NIST warns that systems which generalize poorly beyond their training data raise risk, and that training data can become outdated relative to where the system is used 4.

Who makes the decision?

The Bristol review proposes automation levels running from classical NDT, where the operator does everything, through assistance and human-machine collaboration, to full automation without human input; it aligns these with automation levels published by the European Union Aviation Safety Agency 3. It describes the technology as still at an early stage for safety-critical industrial use 3. NIST notes that human roles in overseeing AI need to be defined, from fully manual to fully autonomous 4.

What are the limits?

  • Images show surfaces only. HSE notes that visual methods detect surface damage, and warns that with remote video systems, image processing must not hide defects 7.
  • Domain shift. A model tested on one weld type, material or set-up has no demonstrated performance on another 3,4.
  • Rare damage. Flaw types that are scarce in the training data are the hardest to learn 2.
  • Thermal images depend on conditions. The 2022 CUI review notes that infrared results are affected by temperature, wind speed, insulation material and pipe diameter 6.

Where integrity problems show up

What it finds wellWhat it can miss or get wrongIn the Library
Corrosion visible on structures in drone images 5Damage under insulation or inside equipment, which images cannot show 7What is atmospheric corrosion?
Small flaws in weld radiographs; a90/95 of 0.6 mm in one study 2Flaw types rare in the training data; performance drops as data shrink 2Radiographic testing
Repetitive review of large ultrasonic data sets 1,3Data from materials, geometries or set-ups outside the trained domain 3,4Ultrasonic testing
Consistent flagging of indications for an inspector 2Misses and false calls that go unmeasured when POD and false call rates are not reported 2,4Probability of detection (POD)
Likely wet insulation from thermal images 6Readings shifted by temperature, wind and insulation type 6What is corrosion under insulation (CUI)?

Sources: PNNL review 1; Tyystjärvi et al., 2022 2; Cantero-Chinchilla et al., 2021 3; NIST AI 100-1, 2023 4; Forkan et al. 5; Cao et al., 2022 6; HSE COMAH technical measures document 7.

Common questions

Is computer vision inspection the same as AI inspection?
Computer vision is the part that works on images, such as drone photographs. AI in inspection also covers radiographs and ultrasonic signals.
Can AI replace a certified inspector?
Not in the work reviewed here. The radiography model was deployed as an aid to inspectors, and the Bristol review places full automation at the top of its scale, beyond current safety-critical practice.
How is this different from predicting corrosion with machine learning?
Inspection AI interprets measurements and images that already exist; prediction models estimate future corrosion rates from process and inspection history. See [How is machine learning used to predict corrosion rates?](/library/machine-learning-corrosion-prediction/).

Beacon explains these concepts in its own words for education. It does not reproduce or replace the standards cited, and Winda is not affiliated with or endorsed by the organizations that publish them. Engineering decisions need the current edition of the standard and a qualified engineer.

Sources

  1. 1.Machine Learning for Ultrasonic Nondestructive Examination of Welding Defects: A Systematic Review (H. Sun, P. Ramuhalli, R. E. Jacob), Ultrasonics 127, PNNL-SA-169657 · Pacific Northwest National Laboratory (PNNL), US Department of Energy (publication page)
  2. 2.Automated defect detection in digital radiography of aerospace welds using deep learning (T. Tyystjärvi, I. Virkkunen, P. Fridolf, A. Rosell, Z. Barsoum), Welding in the World 66:643-671 · Springer (open access, CC BY), via Aalto University repository ·
  3. 3.Deep learning in automated ultrasonic NDE: developments, axioms and opportunities (S. Cantero-Chinchilla, P. D. Wilcox, A. J. Croxford), arXiv:2112.06650 · University of Bristol, via arXiv (preprint) ·
  4. 4.Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 · US National Institute of Standards and Technology (NIST)
  5. 5.CorrDetector: A Framework for Structural Corrosion Detection from Drone Images using Ensemble Deep Learning (A. R. M. Forkan et al.), arXiv:2102.04686; published in Expert Systems with Applications · arXiv (preprint record with journal DOI 10.1016/j.eswa.2021.116461)
  6. 6.A Review of Corrosion under Insulation: A Critical Issue in the Oil and Gas Industry (Q. Cao et al.), Metals 12(4):561 · MDPI (open access, CC BY 4.0)
  7. 7.Inspection/Non Destructive Testing (COMAH technical measures document) · UK Health and Safety Executive (HSE)