Neural Network Based Optical Sensors for Metal Welds

The problem addressed was to develop automated image analysis of metal welds to replace unreliable and costly 100% visual inspection. The objective was to evaluate neural networks combined with features from gray level surfaces. Photomicrographs of the welded seam in Zircaloy-4 channels used in nuclear fuel rod assemblies were examined. The methodology used was to select a set of morphological feature coefficients from Bessel Fourier morphological feature coefficients extracted from digitized photomicrographs. The selected morphological feature coefficients were then used to train a neural network. The approach was verified initially by successfully training a neural network using the known, secondary chemistry of the alloy welds. The results were good. The Bessel Fourier coefficients captured sufficient information from the metallographic images to distinguish weld quality. This approach has the combined advantages of compressing the images through Bessel Fourier coefficients and give the ability to rapidly process large amounts of data using neural networks.