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

Subscribe >
Skip to content

Neural network processing of ultrasonic test signals - Industrial Member Report 602/1997


TWI Industrial Member Report Summary 602/1997

A M Lank, P Mudge, P Wilkinson* and T J Harris*

This report describes the current status of a project conducted in collaboration with the Neural Applications Group at Brunel University.


Manual inspection is a highly skilled operation and generally requires a great deal of experience to be confident about predicting flaw types from signals appearing on a flaw detector screen. It is not unknown for disagreement, even between experienced operators, when diagnosing signals from weldments. It is clear, therefore, that any means of enabling the ultrasonic test technician to achieve better and more consistent results would be beneficial.

When an operator performing a manual ultrasonic test detects a flaw, he will formulate an opinion about the nature of that flaw by scanning the ultrasonic probe over the surface of the workpiece and observing how the A-scan signals change as the ultrasonic beam impinges on that flaw. The approach used in this project is to feed such A-scan signals into a neural network, so that different populations of A-scan signals can be associated with discrete generic types of flaw, hence providing the basis for classification of flaw type.

The aim of this work was to explore whether application of neural network processing offers the potential to provide the technician carrying out manual ultrasonic testing with an objective interpretation of the nature of the flaw being examined. If this approach proves to be successful, it is intended that TWI collaborates with an NDT company to incorporate the technology into standard flaw detection equipment.

This report details the initial progress made in assessing feasibility of the neural network approach.


  • To develop a flaw characterisation system for manually collected ultrasonic signals capable of discriminating between different types of welder-induced flaw, using a neural network-based classifier.

For more information please email: