The Research Of Pipe Weld Defects Identification Based On Machine Learning

In the manufacturing process of SAWH(Helical Submerged Arc Welding)pipe,various welding defects are often produced in the welding seam due to the restriction of production process and production environment.And the existing artificial X-ray image weld defect detection and identification method is inefficient and inaccurate.In this paper,the detection and identification methods of weld defects are studied based on the X-ray images of the SAWH pipe,aiming to propose a weld defect identification algorithm which can realize better identification index,so as to improve the accuracy of weld defect recognition and improve the degree of automation of recognition process.In defect detection process,first of all,regarding the full picture of SAWH pipe welding seam X-ray images as test object,using the method of wavelet filtering,image enhancement,Ostu segmentation,Prewitt edge detection and Hough transform and a series of digital image processing technology to fit the weld zone,and it is verified that the existence of undercut weld defects do not affect extraction of the weld edge.Secondly,according to the limitations of the common threshold based segmentation method,which is not suitable for the segmentation of small area defects,the defect and noise interference points of any size in the region are segmented with the Ordering Points to Identify the Clustering Structure(OPTICS).Then,the defects,noise and normal images without defects are extracted and the data enhancement and size normalization operation are performed,so as to complete the pre-processing of the weld images to build the sample library.In defect recognition process,first of all,the traditional method "artificial feature extraction + machine learning" is used to identify the defects.According to the external geometric features of the defect and noise,six shape characteristic parameters,such as Heywood diameter,were calculated,and the principal component analysis of the characteristic parameters was performed to reduce the dimension,so as to complete the feature extraction of weld defects,and then using the least square support vector machine(LS-SVM)and BP neural network based on genetic algorithm optitmization(GA-BP)to realize weld defect recognition based on artificial feature extraction,and the recognition accuracy was 96.33%and 95.33%respectively.Secondly,using the deep learning technique,combining the convolution neural network(CNN)and the Softmax classifier to train the CNN model,which uses the defect and noise as input samples,and the test results show that the designed CNN model is not only reliable,but also can accurately identify the defects in the adjacent location,and eliminate the complexity of the artificial feature extraction,and achieve the high recognition accuracy of 97.44%.Finally,training three CNNs to build the weld defect identification model and carry out the actual application experiment,the experimental results verify the effectiveness of the algorithm,and the method can reduce the labor quantity while improving the recognition accuracy,thus improving the automatic degree of detection and recognition.