Research On Automatic Detection And Recognition Of Weld Defect Based On X-ray Inspection Images

With the rapid development of the economy and technology,special pressure equipment has been gradually applied in all aspects of society.While the connection between production and life has become increasingly close,it has also caused hidden dangers to the safety of life and property of the public.Due to the particularity and complexity of its application field,such equipment has rigorous requirements on the welding quality.And in the welding process,various defects will inevitably appear in the weld,so it is essential to carry out Non Destructive Testing before it is put into use.Because of its shortcomings,such as low detection efficiency and great influence by subjective factors,the method of detecting defects by manual evaluation is gradually being replaced by computer-aided detection technology.The automatic detection method of weld defects considering deep learning could overcome the disadvantages of manual evaluation and improve the detection accuracy and efficiency.The development of deep learning technology needs a lot of data as research support.Due to the lack of public and accurate datasets in the field of weld defects,the development of defect detection technology based on deep learning is seriously restricted.Aiming at the problems of lack of datasets and low defect detection accuracy in the field of automatic detection and identification of weld defects,this paper firstly designs a preprocessing strategy to clarify the original X-ray image to obtain clear texture and geometric features of weld defects.Subsequently,the strategy of combining traditional data augmentation with improved generative adversarial networks is used to expand the dataset.The generated samples were sorted and labeled to facilitate the subsequent detection and identification of weld defects.Then,the YOLO-V5 network,which has a relatively balanced detection speed and accuracy in the object detection field,is used as the defect detection model.The model is optimized by introducing Attention Mechanism,Adaptively Spatial Feature Fusion(ASFF),and Weighted Boxes Fusion(WBF).And set up multiple sets of comparative experiments to test the welding defect detection effect of different improvement measures,and verify the rationality of the optimized object detection algorithm.Finally,based on the optimized object detection model,the automatic detection software of weld defects is designed and developed.The weld defect detection software is independent of the deep learning development environment,and the preliminary application of the image preprocessing algorithm and deep learning object detection model is realized.The main contents of this paper mainly include the following aspects:(1)Adding Res Net structure to optimize the WGAN-GP network and sorting and labeling the generated X-ray weld defect samples helps solve the problem of lack of weld defect samples in the field of industry and deep learning.(2)By adding Attention Mechanism,Adaptively Spatial Feature Fusion,and Weighted Boxes Fusion to optimize the YOLO-V5 object detection model,the speed and accuracy of weld defect detection are improved,and the m AP reaches 91.90%.(3)Develop and design the automatic weld defect detection software based on the optimized YOLO-v5 model,realize the integrated packaging of user login,image processing,weld defect detection,and other functions,and serve the weld defect detection personnel more conveniently.