a b s t r a c tThis paper presents artificial neural networks (ANN) and wavelet analysis as methods that can assist high resolution of multiple defects in close proximity in components. Without careful attention to analysis, multiple defects can be mis-interpreted as single defects and with the possibility of significantly underestimated sizes. The analysis in this work focussed on A-scan type ultrasonic signal. Amplitudes corresponding to the sizes of two defects as well as the phase shift parameter representing the distance between them were determined. The results obtained demonstrate very good correlation for sizes and distances respectively even in cases involving noisy signal data.

A method is presented to demonstrate the use of artificial neural networks (ANNs) in providing additional information regarding defects or flaws when used in conjunction with the ultrasonic A-scan method. ANNs were employed both as pattern classifiers and as function approximators to maximise the amount of data available from the temporal A-scan signal. A steel bar was modelled in 2D using ABAQUS finite element analysis (FEA) software. A single defect was introduced to the bar, modelled as a void, and parametric studies conducted to record data with the defect at various locations. An ultrasonic Lamb wave was introduced at the top of the bar. The longitudinal wave propagated along the length of the bar and was partially reflected by the defect. Multiple cases were simulated, modelling voids between 1mm and 6mm in width in various locations. Mean displacement of all the nodes at the top of the bar was recorded throughout the simulation, and features extracted from this waveform to create the data set for the ANNs. The ANNs were trained with a percentage of the data collected, selected at random, and assessed with the remaining data. The target data for the ANNs were the depth and size of the defect. The case of two separate defects was also investigated. The procedure was carried out in the same manner as for one defect, but in this case the target data for the ANNs were the depth of the first defect and the distance between the defects. A separate ANN was employed as a pattern classifier, to determine if the reflected A-scan signal represented one or two defects. The final system was tested using previously unseen data, and provided very good results both in determining the number of defects and the size and location of the defects, even with data to which noise had been added.