This paper describes a novel application of ultrasonic Lamb waves combined with network methods for data analysis for a non-destructive evaluation of the adhesive fillet size in the cusp of an aluminium T-peel joint. The ultrasonic signals were transmitted and received using a purpose-built cross correlator system, which generated filtered pseudo-random binary sequences tailored to excite s0 and a1 wave modes in the specimens. Signals received after propagation across the joint area were analysed in the modulus frequency domain by means of a standard fast Fourier transform. Simple statistical measures (such as mean and standard deviation) applied to peaks in the frequency spectra did not provide a robust basis for automatic discrimination between classes of bond fillet size, with 'recognition' success being of the order of chance alone. The most basic form of artificial neural network, the linear network, was then trained to recognize bond fillet radius as belonging to one of three categories of size. When presented with regions of bond fillet that were not included in its training data, it was able to 'recognize' fillet sizes with a success rate of 95%. The sensitivity of the method to experimental arrangements was examined by comparing the results obtained with well-collimated water-coupled transducers with those obtained by using mode-converting contact probes, which exhibited greater angular dispersion in the excited waves. Comparisons were made between different transducer excitations designed to excite s0 alone, a1 alone, s0 and a1 together and also a broadband excitation. Various training protocols for the network were also compared as were the results of output thresholding to minimize the number of wrong decisions made by the network. Overall we find that the best fillet size recognition performance was obtained with the broadband excitation applied to mode-converting wedge transducers set on either side of the bond.
Read full abstract