Neural networks for the classification of nondestructive evaluation signals

The paper proposes the use of massively parallel learning networks for interpreting signals from electromagnetic transducers used in nondestructive evaluation (NDE) problems. Nondestructive testing techniques are used in a variety of industries for evaluating the structural integrity of critical components in a noninvasive manner. A major aspect of research in nondestructive testing is related to the inverse problem and is commonly referred to as defect characterisation. This involves the classification of the eddy current signals in terms of the shape and size of the underlying defects in the test object. A key contribution of the proposed network is the ability to obtain a rotation and translation-invariant internal representation of the signal. Results showing the merits of this approach as well as a comparison with traditional techniques for the classification of signals are presented.