Subscribe to our newsletter to receive the latest news and events from TWI:

Subscribe >
Skip to content

Automatic Defect Classification Using Machine Learning and Computer Vision Techniques for Ultrasonically Acquired Data

Project Code: 32903
Start date and planned duration: February 2019, 36 months


  • Review current Artificial Intelligence (AI) techniques and their fields of applications to assess their suitability for NDT, including: currently implemented medical imaging solutions based on neural-network and deep-learning; binary decision trees; clustering algorithms; and probability distribution.
  • Develop AI algorithms for use with ultrasonic imagery.
  • Develop flexible modelling and simulation capability, for the generation of input parameters for use with the AI model-assisted Automated Defect Recognition (ADR) solution.
  • Demonstrate the benefits and drawbacks of the developed methods through a parametric study based on industry provided samples.


Project Outline

Full Matrix Capture (FMC) and Virtual Source Aperture (VSA) ultrasonic testing exploit the Total Focussing Method (TFM) allowing for fully-focussed imagery throughout the inspection volume by synthesising ultrasonic transmission and reception to a given pixel within an image. Unlike conventional and phased array imaging methods, a column of data within a TFM image is not directly linked to a single underlying A-Scan, but rather a contribution from many, making the technique more image-based than previous ultrasonic inspection methods. Approaching ADR as an image-based problem lends itself more readily to existing AI solutions, allowing developed knowledge to be applied to other NDT technologies.

Computer Vision (CV) is a method that processes images to extract features. When combined with AI this allows fully automated defect detection and classification. AI is an umbrella term covering a number of underlying techniques. More commonly neural-network based solutions are used, comprising a series of probability and weighting functions automatically implemented through the use of training data. Deep learning networks extend this by taking a much larger set of input parameters (typically every pixel in an image).

While neural-network based solutions have proved effective, they require a large set of training data, typically in excess of 1,000 training samples per classification type. This project will develop a technique termed ‘model-assisted ADR’, where modelled data in combination with experimental data are used to simulate training data for the neural network. This will have a big impact in the adoption of neural-network based systems for NDT. However, as the neural-network is essentially a ‘black-box’, once trained there is no means to determine how the classification was achieved, and so no audit trail for false-positive results. The project will therefore also explore a range of AI techniques including binary decision trees and clustering algorithms (both of which provide a full audit trail for classification).

The scope of work for this project will be limited to FMC ultrasonic inspection for the well-defined girth-weld inspection scenario only. This is to ensure parameter inputs are minimised while algorithms are developed. It is anticipated that the outcomes will be applicable to a wide range of materials and NDT techniques.

Industry Sectors

Benefits to Industry

There is a lack of commercially available equipment exploiting AI to automate NDT processes. Such equipment can be used to make quality assurance more consistent at lower cost than manual techniques, and will allow the full potential of advanced techniques, such as FMC and VSA, to be realised. The project will investigate and demonstrate techniques and technology to underpin the development of equipment for industrial applications.


 Research Board Portal

For more information please email: