Automatic Defect Classification in Time-Of-Flight-Diffraction Data Using Neural Networks

摘要

Ultrasonic Time-Of-Flight Diffraction (TOFD) is a fairly recent innovation in NDT that has proved highly effective for the inspection of welds in steel structures, providing highly accurate detection, characterisation, positioning and sizing of defects with a high probability of detection. This has enabled it to gradually replace other more conventional ultrasonic testing techniques. Currently most TOFD data interpretation is done off-line manually by a trained operator and using advanced interactive processing tools in software. This processing is highly dependent on operator skill, experience, alertness and consistency and is a cumbersome, tedious and time-consuming process. Results typically suffer from inconsistency and slight inaccuracies as a result of natural human error when dealing with such large volumes of data that are commonly generated in TOFD investigations. The recent trend in the related disciplines of remote sensing and medical imaging is to automate the data processing and interpretation process as far as possible, relieving the expert to some extent of unnecessary or repetitive tasks. In light of industrial pressure, it is anticipated that TOFD interpretation could benefit from such automation, potentially improving the interpretation procedures by adding an element of robustness, accuracy and consistency. This can be achieved by utilising computational tools that are better suited to discriminating between subtle variations in visual and spectral properties of the data. Furthermore, this would result in a saving in time, effort and cost. Although each defect category has unique characteristics and patterns but there are some similarities between these categories which make the discrimination between these categories not an easy task. This paper presents a novel system for rapid and consistent automatic classification of detected defects in TOFD data as essential stage of a comprehensive unsupervised TOFD inspection and interpretation aid. This system is based on extracting some visual and spectral distinguishable features from A-scan and D-scan segments which represent different defect classes in TOFD data to produce the discrimination bases for the neural classifier to distinguish between these classes. This classifier was applied to a variety of TOFD data sets gathered from a variety of steel plates and tubular pipe lines with different thickness and different defect classes have been successfully discriminated with a consistency greater than that of the expert operator, but in a fraction of the time. The results achieved are rapid, with satisfactory levels of accuracy and reliability to form a robust automatic classification of detected defects. It is hoped this will form the basis for a new paradigm in ultrasonics for fully-automatic batch processing and interpretation.