Ultrasonic Detection Welding Defect Image Recognition Based On Deep Learning Method

The ultrasonic non-destructive testing method has the advantages of high detection sensitivity,accurate defect positioning,and harmlessness to the human body,and has been widely used in the field of welding detection industry.As a weak area of the load-bearing structure,if the weld has defects in the weld,it will endanger the structural reliability.Therefore,it has important theoretical and practical value to effectively detect and identify weld defects.At present,weld defect detection is still mainly based on manual detection and identification,and the results are greatly influenced by the professional level and concentration attitude of the test personnel,and the detection efficiency is low.Under the requirements of industrial automation,information and intelligence,it is especially important to automatically identify defects in welds based on machine learning theory.Traditional ultrasonic inspection of weld defects is based on defect echo signals.With the emergence and application of imaging detection techniques such as ultrasonic phased array and C-scan,how to identify weld defects on the basis of ultrasonic images is particularly important.Based on the deep learning algorithm of convolutional neural network,relying on the YOLOv3 network structure model,the C-scan image of carbon steel butt weld is trained to establish an automatic C-scan image recognition model,which has the weld defect in C-scan image.Automatic identification and positioning capabilities.The optimal detection model is obtained by analyzing the influence of training times on the recognition accuracy.Further,the preliminary defect quantification and localization in ultrasonic nondestructive testing is realized by defect frame selection and position output.Using 70 C-scan images training models,the model loss value can reach 0.05 after 3000 training rounds on a PC,and the training takes about 90 minutes.Using 38 C-scan image verification sets for detection,the defect recognition accuracy of the model is 91.01%,and the single C-scan image detection time is only 0.046 seconds,which represents the quantitative index of the defect size and position(cross-over)reaching 88.46%.In summary,the detection model has the advantages of fast training speed,high detection efficiency and good recognition rate,and has a good application prospect in actual ultrasonic non-destructive testing.