Classification of Defects by the SVM Method and the Principal Component Analysis (PCA)

Analyses carried out on examples of detected defects echoes showed clearly that one can describe these detected forms according to a whole of characteristic parameters in order to be able to make discrimination between a planar defect and a volumic defect. This work answers to a problem of ultrasonics NDT like Identification of the defects. The problems as well as the objective of this realized work are divided in three parts: Extractions of the parameters of wavelets from the ultrasonic echo of the detected defect the second part is devoted to principal components analysis (PCA) for optimization of the attributes vector. And finally to establish the algorithm of classification (SVM, Support Vector Machine) which allows discrimination between a plane defect and a volumic defect. We have completed this work by a conclusion where we draw up a summary of the completed works, as well as the robustness of the various algorithms proposed in this study. Keyword: NDT, PCA, SVM, Ultrasonic, Wavelet. Introduction The Non Destructive Testing (NDT) has to allow obtaining the highest possible detection probability, the most exact size and the exact orientation of dangerous defects that the specimen to test can contain. The identification or the knowledge of detected defects nature is very difficult in Ultrasonic technique. This stage of the inspection is based on the experience of the expert controller. This one proceeds by changing the angle of ultrasonic beam and a lot of other tricks in order to find out a diagnosis on the defect nature: planar or volumic. This verdict is very important since norms and standard accept some volumic defects but refuse others planar defects. Therefore the acceptance or the refusal of defect could stop the functioning of an industrial installation. Analyses carried out on examples of detected defects echoes showed clearly that one can describe these detected forms according to a whole of characteristic parameters in order to be able to make discrimination between a planar defect and a volumic defect. This work answers to a problem of ultrasonics NDT like Identification of the defects. In reference [1], we applied artificial neurons networks (ANN) as a classification method. In this study, we use Support Vector Machine (SVM). This paper will be divided into three parts: In the first part, we describe the extraction from the detected defect signal, the wavelet parameters by the Discrete Wavelet Transform (DWT). With these parameters, we build an attribute vector and then, we optimize this one by Principal Components Analysis (P.C.A) method, this stage will be seen in second part. In the third part, we carry out defects classification by SVM method into: planar or volumic defects. Results of this paper will be compared to those obtained in [1]. ECNDT 2006 We.2.4.4