Neural-fuzzy classification of weld defects using ultrasonic time-of-flight diffraction, BINDT Mini Conference on Advanced Ultrasonics

TLDR: In this paper, a Neural-Fuzzy classifier is developed to overcome the dimensionality problems associated with Fuzzy logic and to provide the training and learning ability of the neural network s which makes this classifier more powerful and reliable in clustering of each defect class.

Abstract: Ultrasonic Time-Of-Flight Diffraction (TOFD) is rapidly gaining prominence as a reliable non-destructive testing technique for weld inspection in steel structures, providing highly accurate positioning and sizing of flaws with a high probability of detection. Use of TOFD as a rapid non-destructive i nspection tool for steel plates, pipelines and vessels has grown tremendously during recent years, bringing into light the challenge of developing a fast and reliable dat a processing and interpretation platform. This paper presents several innovative procedures d eveloped and implemented to great success for the automation of the classification of weld flaws in TOFD data according to the adopted defect classification standard as an essential stage of a comprehensive TOFD inspection and interpretation aid. A Neural-Fuzzy classifier is developed to overcome the dimensionality problems associated with Fuzzy logic and to provide the training and learning ability of the neural network s which makes this classifier more powerful and reliable in clustering of each defect class. A number of advanced image processing tools have been developed to characteris e TOFD images and extract distinguishable features to be used in defect class ification. Several features have been investigated and selected, which prove to produce a good discrimination between different defect classes. Neural-Fuzzy classificati on is a powerful tool and considered as the key technology for representing knowledge of th e human experience and for constructing adaptive systems. Combining the technologies of Neural-Fuzzy classification with advanced visual scan features e nables the differentiation between different defect classes in a fully automatic and u n-supervised manner.