Welding Defects Modeling And Recognition Algorithm Reasearching Based On Support Vector Machine

With the development of machine vision and pattern recognition technology, theintelligent method for defects recognition has become a hot research topic in the field ofNondestructive Testing. However, the traditional intelligent recognition algorithms requireplentiful number of training samples, and it leads to lots of operation time with theincreasement of samples’ dimension and category. According to this, a method demands smallamount of training samples is proposed in this research to classify different welding defects ofpetroleum welded pipe, it not only distinguish the welding defects from image noise, but cantell the specific category. Moreover, it will help to improve the inspection efficiency andcontrol the quality of welded pipe once applied in the production site.In this research, the X-ray images of submerged-arc welded pipe are taken as the researchobject. To identify the welding defects, such as: air hole and crack, which may exist in thewelding seam, a process includes: defects segmentation, feature parameters extraction andobject classification is studied. Firstly, the average filter, Sine enhancement, OSTUsegmentation and sobel method are used to detect the welding seam boundary in the entireimage, and then the Hough transformation is adopted to demarcate the weld area bycalculating the expression of boundary line. Secondly, the effect of OSTU and theDensity-Based Spatial Clustering method in defects segmentation are studied, and the latter ischosen to segment the welding defects and image noise in the determinate weld area.Furthermore, the feature vector is built up by calculating6classes of shape feature of thesegmented defects to realize the defects modeling, and the Principal Component Analysismethod is applied to preprocess this vector. Finally, the recognition between the image noiseand welding defects is studied based on the Support Vector Machine (SVM) theory, of whichthe preprocessed vector is taken as the input parameter.What’s more, the PSO-SVM (Particle Swarm Optimized SVM), GA-SVM (GeneticAlgorithm based SVM) and LS-SVM (Least Square Support Vector Machine) are studied aimat reducing the operation time of classical SVM in the process of defects recognition. Theexperimental result shows that the LS-SVM method achieves a better performance on welding defects recognition, with an average accuracy of92.8287%, and its operation time is0.394559s, compared with the classical SVM method, its accuracy increased by1.6%,operation time reduced by72%, which can be proved as a rapid and accurate welding defectsrecognition method.