Detection and Strong Classification of Natural Weld Defects by Magneto-Optical Imaging Under Rotating Magnetic Field Excitation

The accurate detection and classification of invisible weld defects is very important to ensure the quality of welding products. A magneto-optical (MO) imaging detection system excited by a rotating magnetic field is proposed for feature extraction and detection classification of invisible natural weld defects. A finite element analysis (FEA) model of multidirectional natural weld cracks is developed to study the distribution of rotating magnetic fields at different transient times, and the distribution of leakage magnetic field for four types of invisible defects is analyzed. The correctness of the finite element model is verified by MO imaging test of multidirectional natural weld defects. The grayscale values and texture features of MO images can reflect the leakage magnetic field characteristics of weld defects. The MO imaging detection system excited by a rotating magnetic field is used to obtain MO images of defects such as non-penetration, pit, subsurface crack, surface crack, and no defect for weld defect diagnosis. The principal component analysis (PCA) method and the Tamura method are used to extract the grayscale values and texture features of natural weld defect's MO images, respectively. Using backpropagation (BP) neural network as the weak classifier, a strong classification model of BP-Adaboost weld defects is established by the Adaboost algorithm. Experimental results show that the recognition accuracy of the PCA + Tamura- BP-Adaboos model is higher than that of the PCA + Tamura-BP model, and its overall recognition rate reaches 97.8%, and the detection and classification of multidirectional natural weld defects can be realized.