Combining fuzzy logic and neural networks in classification of weld defects using ultrasonic time-of-flight diffraction

Ultrasonic TOFD is good at detecting most internal defects, results are available immediately, testing can be performed ‘on site’ – equipment is fully portable, running costs and safety hazards are low. However, TOFD requires experienced operator skills for the interpretation and detection of defects. These days, with advanced technology, ultrasonic testing can be made more accurate, reliable and feasible in most inspection systems(2-7). Although the data acquisition configuration lends itself conveniently to automation, and methods such as robotic scanning and computer-conditioned data acquisition are routinely used, the crucial processes of data processing and interpretation are still performed off-line manually depending heavily on the skills, experience, alertness and consistency of the trained operator. Results typically suffer from inconsistency and errors, particularly when dealing with large volumes of data. In the light of industrial pressure, the recent trend has been to automate the TOFD data interpretation process in software by adding an element of robustness, accuracy and consistency. This can be achieved by advanced image processing and artificial intelligence techniques to discriminate between subtle variations in visual properties of the data, reducing the overall interpretation time, effort and cost. The TOFD image provides important characteristics and patterns in recognition of defect types. However, lack of unique visual defect signature interpretation of volumetric flaws is not trivial. Further, TOFD standards and acceptance criteria introduce their own complexities. The prescribed(1) defect classes are planar flaws, volumetric flaws, thread like flaws and point flaws. In this research, statistical image processing methods and algorithms have given excellent results. In the classification stage three different classifiers are used to demonstrate performance of different methods: neural classifier, fuzzy classifier and neural-fuzzy classifier. The TOFD automatic inspection system can be semi or fully automated(3-4)(8-9) for the flaw detection in the metal structures.