Research On Algorithm For SSAW Defect Identification And Evaluation Based On Digital Radiography And Instance Segmentation

Large-diameter pipelines are extremely important oil/gas transportation infrastructure,and spiral submerged arc welding(SSAW)is the main method for producing such pipelines.However,welding defects can seriously affect the safety of pipelines.Therefore,defect detection in weld joints is crucial.Digital radiographic imaging technology is a commonly used non-destructive testing method,which has the advantages of high sensitivity,high resolution,and high reliability,and is suitable for detecting welding pipeline defects.Currently,the evaluation of digital radiography images of SSAW welds is mainly done through visual interpretation,which is inefficient and can be affected by subjective factors of the assessors Therefore,improving the automation level of non-destructive testing of weld defects has become a research trend.To address this issue,we propose an automatic evaluation algorithm for digital radiography images of SSAW welds based on deep learning instance segmentation.The specific workflow is as follows:(1)We collect images of SSAW weld defects using a digital radiography imaging device and analyze burn-through,crack and porosity defects from their causes and morphological features.To address issues such as different sample sizes for different types of defects,we perform data augmentation on the images using rotation,flipping,random cropping,and contrast variation to build a more complete dataset of SSAW weld defects.(2)We propose an improved instance segmentation algorithm called Weld Mask,which builds upon the Blend Mask algorithm to address the challenges posed by small-sized porosity defects with low pixel proportions,as well as the narrow and elongated shape of crack defects in SSAW welds.The Weld Mask algorithm combines several techniques,including Convolutional Block Attention Module and Deformable Convolutional Network incorporated into the backbone network,Bi-directional Feature Pyramid Network instead of Feature Pyramid Network,and the Generalized Intersection over Union loss function to improve regression loss.Experimental results demonstrate that the Weld Mask algorithm has better detection and segmentation performance than other classic instance segmentation algorithms on digital radiography images of SSAW welds.The average precision of the detection of welds and their defects is 93.41%,and the average precision of masks is 75.48%.Overall,the Weld Mask algorithm has high practical value for the steel pipe industry.(3)Based on the Weld Mask instance segmentation algorithm proposed in this thesis,we implemented an intelligent evaluation system for digital radiography images of SSAW welds,using Py Qt5 to construct the graphical user interface.Our system can effectively detect and segment the weld area and defects in SSAW weld digital radiography films and rate them according to predetermined quality rating standards based on the instance segmentation results,helping inspectors better complete the evaluation of SSAW weld films.