| Weld defect detection has long been a key topic in the field of industrial inspection and plays a crucial role in the quality control and safety reliability of welded parts.Digital Radiography(DR),as an important research direction in radiographic inspection,is able to automate the acquisition process of radiographic images with the assistance of mechanical and electrical equipment,which meets the requirements of industrial automation.It meets the requirements for defect detection technology in the era of industrial automation.However,there are two problems in the field of DR detection: first,the extreme lack of real defect data makes the process of extracting effective feature information very difficult,and the characteristics of multi-scale defects,which are not easily distinguished from background areas,often lead to low accuracy of recognition results;second,researchers at this stage focus most of their work on the design of algorithm models to improve the accuracy of defect recognition,while the recognition algorithm models are not as accurate as the DR image acquisition process,the defect detection process,and the defect detection technology.The second is that at this stage,researchers have focused most of their work on the design of algorithm models to improve the accuracy of defect recognition,but little research has been done on the effective combination of recognition algorithms,DR image acquisition process,and visual management of defect recognition results,which makes it difficult to truly automate the whole process of DR inspection.To this end,this thesis takes the variable diameter barrel ring welds of complex components as the research object and uses digital ray detection technology and semantic segmentation methods based on deep learning to achieve fully automated detection of defects in the welds of complex components.The specific work and research are summarized as follows.1.Based on the theory of radiographic inspection,the structural characteristics and process parameters of complex components are analyzed,and the intelligent DR imagingbased weld defect recognition system for complex components is designed in conjunction with the defect detection requirements;the hardware acquisition system scheme is designed,and the detection process and optimal parameters are determined in conjunction with the actual inspection tasks.2.To address the problems of high noise interference and low contrast in the original weld images acquired by the hardware system,a weighted mean filtering algorithm is used to eliminate Gaussian noise in the images,and a Laplace operator with global HE weighted fusion enhancement algorithm is designed to improve the contrast of the weld areas.In order to allow the weld seam images to be input to the subsequent defect recognition network,the above processed images are processed with resolution reduction and data augmentation.Based on this,a high-precision defect recognition network model is proposed in this thesis.The feature extraction stage combines two modules,the backbone feature network and the enhanced feature network,to extract multidimensional feature information of defects;the feature fusion stage uses the feature information after multiple downsampling to gradually recover the image resolution;the network training process uses a weighted fusion loss function for sample imbalance optimization.On the private weld image dataset,the F1 and m IOU evaluation metrics are improved by 0.2 and0.3 percentage points,respectively,compared with Deep Labv3+,and the number of model parameters is reduced by 90%.3.Based on the high-precision defect segmentation network model,the DR imagingbased automatic identification software is designed on the basis of a full understanding of user requirements,which realizes the automation,precision and informationization of the defect detection process with friendly and interactive operation,and provides a strong guarantee for the automatic detection of defects in the welds of complex components. |
PDF Full Text Request