Efficient feature extraction for an automatic ultrasound testing decision support system

The ultrasound testing is a nondestructive evaluation method widely applied in the industry for detection of flaws in different materials. Its efficiency strongly relies on the operator knowledge, as he is responsible for interpreting the acquired signals and detecting the existence of defects. In view of this, some works have been developed to design automatic decision support systems for nondestructive ultrasound testing. Their aim is to provide valuable information to assist the operator in the decision making process. Most of these automatic systems comprise different signal processing steps (i.e. feature extraction and classification) in order to reach a condition indication for the evaluated material. This work presents a study on the relevance of distinct feature extraction techniques for a steel welded joint ultrasound testing decision support system. The Fourier, cosine and wavelet transforms were applied to estimate relevant attributes used to feed neural network based classifiers. Three different welding defects were considered (lack of fusion, porosity and slag inclusion). The discrimination results obtained with the different preprocessing techniques were presented and compared.