Eddy Current Testing is one of the important techniques in Nondestructive Testing. Automated characterisation of eddy current signals (ECS), either in the form of Lissajous patterns (figure-of-eight) or individual voltage vs. time signals is an area of growing interest. This is particularly relevant in environments where the signal-to-noise ratio (SNR) of ECS are very poor. Intelligent, timely and precise interpretation of resulting data, is the key for improving the efficiency of NDT&E. A comprehensive study has been undertaken by the authors for the characterisation of ECS having poor SNR, using three types of artificial neural networks (ANNs). Among the many methodologies to automate interpretation and classification of noisy NDT&E data correctly, the choice of artificial neural networks (ANNs), seems appropriate, in view of the fact that ANNs capture the underlying statistical regularities (or lack of it) of the data, which are otherwise not possible by analytical means. The types of ANNs used in this study are
- the error-back propagation model,
- the binary Hopfield model and
- the Kohonen's self-organising maps model.
These three types of ANNs have been used with the express purpose of classifying ECS where - a number of examples are available,
- the signals are in the form of Lissajous patterns and
- where ECS are available but their classification is not known a priori.
Eddy current signals, acquired from different types of defects (both artificial and natural) such as holes, notches, cracks and pits on stainless steel Type 304 sheets were used in this study. The paper discusses the issues of implementing the various neural network models, the classification results obtained and the utility of the respective models in field applications.
NDTnet
