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

The ultrasonic Time-of-Flight Diffraction (TOFD) technique is gaining rapid prominence in non-destructive testing due its high accuracy in detection, positioning and sizing of weld flaws in steel structures. Until lately, TOFD was used reliably only in fast-track inspections due its portability in automatic scanning and data acquisition. However, data processing and interpretation of TOFD data requires expert knowledge and accuracy largely depends on the operator experience. Hence, results suffer errors as interpretation is often carried out offline, especially when dealing with large volumes of data. A fully comprehensive automatic detection and interpretation can be achieved using advanced image and signal processing and artificial intelligence techniques, thus reducing time, cost and errors in the detection, interpretation and classification of flaws in steel structures.

This paper presents current research using advanced methods for automatic interpretation and classification of weld defects in TOFD data. In the classification stage three different classification techniques are employed and compared: an artificial neural network-based classifier, a fuzzy logic-based classifier and a hybrid neural-fuzzy classifier. A neural classifier can learn from data, but the output does not lend itself naturally to interpretation. A fuzzy classifier on the other hand consists of interpretable linguistic rules, but they cannot learn. A neural-fuzzy classifier is based on a three-layer feed-forward neural network and combines the merits of both neural and fuzzy classifiers while overcoming their drawbacks and limitations. The developed neural-fuzzy classifier exhibits high levels of accuracy, consistency and reliability, with acceptably low computational time and is a promising new development in the field of semi-automatic weld inspection.