Study On Ultrasonic Testing Based On Video Positioning Method And Defect Recognition

inspection of welds structures are accomplished by the A-Scan ultrasonic detector. As we all known, this method requires the measurement and recording of the position of ultrasonic probe manually once a defect is found. Furthermore, the testing data is hard to store and the results differs from each operator. To solve such practical engineering problems, an ultrasonic testing system based on the video positioning is developed in this research. This system could show the projection imaging of weld defects in three-dimensional. At the same time, the system has many advantages such as simple structure, convenient application, little cost.This thesis introduces the mechanism of video camera positioning for ultrasonic probe, the corresponding relationship between probe position and echo signal in manual scanning and image distortion correction of video camera. The reliability of the system is verified.A lot of weld samples containing four kinds of defects: porosity, slag, cracks, and lack of penetration are detected by the constructed system and the characteristics and rules of different flaw signal are analyzed in time domain, frequency domain and wavelet analysis. Defect features corresponding to different categories are formed by the result of signal analysis form and three-dimensional projection imaging pictures. So the original feature set of different defect categories is constituted. Divisibility measure of distance is adopted as the evaluation function of feature extraction, and three feature selection strategies: the Sequential Forward Search, the Sequential Backward Search, the Plus l Take Away r algorithms are used to choose features in original feature set. By comparison the result, the Plus l Take Away r algorithm is better than the other two feature selection algorithms. The original feature set of 32 dimensions is reduced to 5 dimensions.Finally, the BP network and support vector machine (SVM) are both used to classify and identify flaw styles. By comparing the recognition results under the conditions of same training samples and testing samples in the feature set before and after feature selection, it is concluded that the defect classification results by feature selection is better than not use feature selection; SVM classification is better than the BP network.