Based On Deep Learning Research On The Technology Of Digitized Weld Film Defect Recognition

It has become the consensus of domestic operators that radiographic inspection is frequently carried out on the welds of major oil and gas pipelines.With the rapid development of oil and gas pipeline construction,the workload of radiographic film evaluation has increased significantly,which has brought great challenges to manual film evaluation.Moreover,the subjectivity,non-continuous work and time-space constraints of manual film evaluation have increased the uncertainty of crater quality evaluation and on-site construction arrangement.In order to solve the problems and "pain points" existing in traditional industrial film evaluation,this thesis has carried out the following research around deep learning technologies such as convolutional neural network and computer vision:(1)Based on the foreign open source ray database GDXray and the actual weld film on the domestic construction site,a new weld film database is established,from which2000 films with defects are selected.The films were digitized using the MII-5000 LC digital scanner.The moving segmentation method and window sliding segmentation method are used to segment the defects of digital films,and a total of 10976 defective images are extracted.Then the defective images are rotated,flipped,gray-scale balanced,Gaussian noise added,image blur and other data enhancement operations.Finally,four kinds of films,strip,circular,incomplete and No defect,are obtained for convolution neural network training.(2)Based on the convolution neural network technology,two types(five kinds in total)of deep learning models are built.The first type of model is used to judge the type of defect,with the highest overall accuracy of 93.36%,and the recognition accuracy of circular defect,incomplete fusion defect and strip defect are 90.2%,88.5% and 97.3%respectively;The second type of model is used to judge whether there are defects in the image,and the highest accuracy is 98.36%.Through the cross validation of the two types of models,the film defect identification is realized.Based on window sliding technology and threshold segmentation technology,combined with convolution neural network,the accurate location of film defects is realized.(3)Based on VB(Visual Basic)programming software,the application software of automatic identification of weld radiographic film defects is developed.The software takes convolution neural network as the core algorithm,and has the functions of defect type judgment,defect location,manual marking,automatic identification report output,batch film processing,historical processing result viewing and so on.Through the field application test,the highest defect detection rate of the software is 97.5%,the highest accuracy rate is 94.5%,and the lowest false positive rate is only 19.5%.The automatic defect identification system built in this thesis can replace or assist manual welding quality rating,alleviate the pressure of manual film evaluation,and has a positive significance for the further development of circumferential weld radiographic testing and pipeline construction engineering.