Research On Deep Learning Based Algorithms For Automatic Detection Of Weld Defects

At present,Welding has become one of the most important process technologies in manufacturing and is widely used in industry.This is why the detection of weld defects is of great importance.At present,the mainstream weld defect detection still relies on the analysis and evaluation of X-ray flaw images by professionals,which is not only inefficient,but also subjective,making the automation of weld quality assessment difficult to achieve.In order to achieve automated weld quality assessment on welding production lines,this paper constructs a standard weld data set and explores weld defect detection methods in terms of both detection speed and detection accuracy with the help of deep learning.The following is the major focus of this study.1.To address the problem of the lack of publicly available and data-rich weld seam datasets at home and abroad.This paper collects and collates X-ray weld inspection images and designs a pre-processing process to improve the quality of weld images.The main pre-processing processes include non-local mean filtering,Laplace sharpening enhancement and weld seam region extraction.Finally,the processed weld images are annotated,and then a weld dataset for deep learning is constructed.2.The CenterNet algorithm achieves a detection speed of 86.3 frames per second(FPS)on the test set.In order to balance the speed and accuracy of the algorithm,this paper improves the down-sampling structure of the backbone network according to the network structure characteristics of the CenterNet algorithm,then optimises the dimensionality reduction structure of the residual module,and finally adds a selfcorrecting attention mechanism in the residual module to improve the network generalisation capability.Compared with the original algorithm,the improved CenterNet algorithm improved the detection accuracy by 2.0% for m AP,1.0% for AP50 and 3.4% for AP75 with comparable detection speed.3.Despite the high detection speed of the improved CenterNet algorithm,the accuracy is not satisfactory.The Cascade RCNN algorithm achieved 67.6% detection accuracy AP on the test set.The Cascade RCNN algorithm was also optimised and improved in various ways according to the characteristics of weld defect images.To address the problem of irregular shape of weld defects,this paper adds deformable convolution to the backbone network to improve the ability of defect feature extraction.To address the problem that the information of small-sized defects in weld seams is easily lost,the Bi FPN structure is added after the backbone network to improve the feature fusion capability of the network.To address the problem of inaccurate detection due to partial overlapping of defects in the weld,Soft NMS is used instead of classical NMS to improve the accuracy of the network in detecting dense defects in the weld.The Cascade RCNN algorithm with multiple improvements improved the detection accuracy by 3.9% for m AP,0.4% for AP50 and 4.6% for AP75 on the test set.