Efficient feature selection for neural network based detection of flaws in steel welded joints using ultrasound testing.

This work studies methods for efficient extraction and selection of features in the context of a decision support system based on neural networks. The data comes from ultrasonic testing of steel welded joints, in which are found three types of flaws. The discrete Fourier, wavelet and cosine transforms are applied for feature extraction. Statistical techniques such as principal component analysis and the Wilcoxon-Mann-Whitney test are used for optimal feature selection. Two different artificial neural network architectures are used for automatic classification. Through the proposed approach, it is achieved a high discrimination efficiency by using only 20 features to feed the classifier, instead of the original 2500 A-scan sample points.