Ultrasonic Identification Of Weld Defects In T-joints Based On Multi-CNN Ensemble Learning

Welded T-joints are widely used in various steel structures.The weld quality in welded components is one of the main factors affecting the mechanical properties of components.The quality of welds is determined by the nature,distribution and size of defects in welds.Accurate qualitative identification of weld defects is the key to ensure the safety of engineering structures.In this thesis,the weld defects in T-tube joints are taken as the research object,and the following aspects are carried out based on deep learning and ensemble learning methods.A total of 70 T-joint specimens with weld defects were designed and welded.The ultrasonic flaw detector was used for ultrasonic testing.The collected ultrasonic echo signal images were used for training,verification and testing of the convolutional neural network.Eighteen butt steel plate specimens,25 butt steel tube specimens and18 butt steel tube cone head specimens were designed and welded.The ultrasonic echo signal images collected after ultrasonic testing were used to test the generalization ability of the integrated deep learning model.A data set containing 8800 ultrasound echo signal images was established as the object of network training,verification and testing.In the process of ultrasonic testing,the label is used to mark the defect position and record the data,and the defect type at the mark is verified by the cutting method.The one-to-one correspondence between the ultrasonic echo signal image and the four weld defect types of porosity,slag inclusion,incomplete penetration and incomplete fusion is established.The structure layer,parameters and weights of convolutional neural networks Mobile Net-v2,Alex Net,Google Net and VGG are adjusted.Four convolutional neural networks are trained by training set data,and four convolutional neural networks suitable for ultrasonic detection echo signal recognition of T-tube nodes are obtained.The average recognition accuracy of the four networks is 91 %,85 %,66 % and 45 %respectively.The recognition accuracy and recall rate are calculated as the evaluation indexes of the recognition performance of the adjusted convolutional neural network.Based on ensemble learning,the trained convolutional neural network is combined according to the voting method to establish an integrated deep learning model,and the influence of the combination of various types of convolutional neural networks and the size of ensemble size on the recognition accuracy of the integrated deep learning model is explored.It is found that the recognition accuracy of the integrated deep learning model formed by the combination of Mobile Net-v2,Alex Net and Google Net convolutional neural networks under the weighted voting method is the highest,and the increase of the ensemble size is beneficial to improve the recognition accuracy of the model.The recognition performance of the integrated deep learning network model composed of the absolute majority voting method,the relative majority voting method and the weighted voting method is compared.The best voting method is the weighted voting method,and the worst is the absolute majority voting method.It shows that in the integrated deep learning model,it is necessary to give more voting rights to the convolutional neural network with stronger recognition performance,and for the convolutional neural network with weaker recognition performance,its decision-making power should be appropriately reduced to improve the recognition performance of the entire integrated deep learning model.The generalization ability of the trained network is tested by using the ultrasonic detection echo signal images of butt steel plate,butt steel pipe and butt steel pipe cone specimens.The test results show that the trained network can also be used to identify the weld defect types of the above three structural types of components,but the recognition accuracy is reduced.