Research And Implementation Of Welding Defect Detection System Based On Deep Learning

It is one of the important means to ensure the quality of welding products to use radiographic detection technology to detect the internal defects of welding products to evaluate and repair them.With the development of deep learning,the application of object detection algorithm to Non-Destructive Testing(NDT)of welding defects is the current mainstream research direction,which also improves the low efficiency of traditional manual evaluation and the mis-assessment and omission due to the subjective factors of the assessors.Through analyzing the research status of welding defect detection at home and abroad,and combining with the research results and the latest development in the field of object detection,the research ideas of this paper are determined.Due to the lack of open,large,and standardized welding defect datasets,this paper screened and sorted out the dataset containing five common industrial welding defects,and introduced the data augmentations of flipping and crop-paste to enhance the size of the dataset and improve the generalization of the algorithm.For the welding defect detection task,this paper proposes a welding defect detection algorithm based on the cascaded structure model.At present,when object detection algorithms based on a single threshold are used to locate and identify the defects in the digital image of welding radiographic inspection,the final detection accuracy will be low due to the distribution difference between the size and feature information of different types of defect samples.In this paper,the Cascade Mask R-CNN is improved by using deformable convolution,an efficient global context modeling,and Feature Pyramid Network to optimize and fuse features,and self-setting the aspect ratios of anchors according to the annotation of the dataset.The experimental results show that the improved Cascade Mask R-CNN has significantly improved the detection accuracy compared with other two-stage models,especially for small objects such as round flaws and cracks,which verifies that the improved model partially counteracts the effects of differences in the defect samples’ distribution.Based on the above algorithm,this paper designs and implements an automatic welding defect detection system integrating image operation and processing methods commonly used in welding defect detection,which can assist NDT personnel in defect detection and record management,and promote the application of related target detection algorithms in practical industrial detection.