Study On Ultrasonic Testing Of Defects Based On Wavelet Transform And Support Vector Machine

After studying the development of ultrasonic nondestructive testing and its application, an experimental system using ultrasonic echo to detect the defects was established and experiments were conducted to detect defects perpendicular to the surface of steel block. This paper also does ultrasonic testing signal analysis and studies ways to determine the defect size. The main works are the followings:To achieve the classification of ultrasonic flaw signals, an experiment system with OmniScan MX UT portable ultrasonic detector, ultrasonic sensor, the test piece with defect, the probe scanning control system and test artifacts was set up; The experiment system uses ultrasonic echo to detect the inside flaw of the metal test piece and ultrasound incident angle probe method to do signal acquisition; Through simple analysis of data, incorrect data were deleted.Based on in-depth study of wavelet transform theory and wavelet de-noising based on wavelet transform, this paper determines the wavelet, de-noise the ultrasonic testing signal; Then this paper does multi-resolution analysis to the signals, extracts gain and energy proportion as features of flaw size.Based on in-depth study of the theory of SVM, the paper chose an appropriate kernel function and way to select the parameter of kernel function that are suitable for ultrasonic testing signal classification and method of defect size classification.After in-depth study of model of ultrasonic impulse detection signal, the paper proposes LM algorithm of parameter estimation of ultrasonic signal, applies it to experimental signals and gains good performance. The paper extracts the result of parameter estimation as features of flaw size.Choose the gain and energy characteristics based on wavelet transform and the ultrasound signal parameter estimation results calculated by LM algorithm as input vector respectively; use the above mentioned input vector to train support vector machine and obtain classification models, then use the resulting models to classify the size of the ultrasonic signal of different flaw size, and use BP neural network to test the feasibility and effect of the proposed methodTo achieve better classification performance, the paper proposed the combination of two kinds of features that is making gain and energy as features of flaw sizes that are smaller than 1.Omm and result of parameter estimation as features of flaw sizes that are bigger than 1.0mm.