Machine Vision Based Superconducting Cavity Flange Surface Defect Detection Research

The superconducting RF cavity is the main acceleration unit of a science device gas pedal,and the superconducting RF cavity flange surface is the flange sealing surface that makes the superconducting RF cavities interconnected.The presence of defects such as longitudinal scratches and den ts on the flange surface of the superconducting cavity can cause the vacuum leakage rate of the superconducting cavity to exceed a reasonable range,which can have a serious impact on the experiment.During the production,transportation and assembly of th e flange,the defects on the flange face are still detected manually,so there are still shortcomings such as low detection efficiency and low precision.In order to guarantee the vacuum leakage rate of superconducting cavity,this paper combines the exper imental demand of superconducting cavity,takes the flange surface of superconducting cavity as the research object,and uses two-dimensional defect detection and three-dimensional defect detection technology based on machine vision to detect the defects o n the flange surface of superconducting cavity,and the main research contents are as follows:(1)An integrated residual attention convolutional neural network(IRA-CNN)is established to effectively extract the defect features of the superconducting cavit y flange surface.The neural network uses four integrated residual attention blocks(IRA-Blocks)to continuously extract features,and the feature information is sampled and compressed using a maximum pooling operation to detect defects more accurately without degrading the original image resolution.(2)The integrated attention module(IAM)is further constructed for the randomness of defect locations on the flange surface of superconducting cavities and integrated into the integrated residual attention blo ck.The integrated attention module is inspired by the convolutional block attention module(CBAM),which is a channel attention submodule cascaded with a spatial attention submodule.The integrated attention module enables the integrated residual attentio n convolutional neural network to have better classification performance and model interpretability,suppress background noise,highlight defect regions,and improve defect localization accuracy.(3)A weakly supervised learning framework is built around superconducting cavity flange surface defect recognition and defect segmentation.The Grad-CAM++algorithm is used to multiply the gradient weights of the feature map with the feature map to obtain the saliency map.It can indicate the probability of whether each pixel is a defect or not,and then segment the saliency image by the OTSU method,which improves the segmentation accuracy of superconducting RF cavity flange surface defects and simplifies the segmentation task of the image.(4)A new method for generating parallax from stereo images under radiation conditions is proposed.The method uses an improved histogram equalization technique to maintain the object shape and the average brightness of the input image.A spatial gradient model is used for adaptive extraction and weighting of the feature set.In addition,a cost matching method based on correlation and support weights preserves edges and achieves exact matching of stereo pairs.After all the features are used in the correlation-based matching technique,they are then used in the costing and aggregation phase with support weights to produce the parallax map.Finally,the obtained parallax maps are weighted median filtered to remove artifacts and normalize sharp edges,which improves the stereo image discrepancy caused by reflections from the flange surface of the superconducting cavity.