Research On Key Technology Of Welding Surface Parameter Detection For Pressure Vessel Based On Machine Vision

In order to improve the safe operation level of pressure vessels and improve the detection efficiency of pressure vessel weld surface quality parameters.The title of this paper is"Research on the Key Technology of Surface Quality Parameter Detection of Pressure Vessel Welds Based on Machine Vision",aiming at the detection of surface quality parameters of pressure vessel welds.We focus on key technologies such as single quality parameter feature point analysis and calculation under multiple defects on the surface of active vision sensing welds,deep learning based pressure vessel weld surface quality parameter feature point extraction method,active vision pressure vessel weld quality parameter feature point extraction real-time performance enhancement method.This is of great practical significance and academic value to ensure the safety of equipment and human beings,and to improve the inspection technology and equipment level of pressure vessel weld surface quality parameters.The paper discusses the status of domestic and international research from the analysis and calculation of pressure vessel surface quality parameter feature points,deep learning parameter feature point extraction methods,and feature point extraction real-time performance enhancement methods.We point out that the active vision weld detection method detects shape information including weld surface,cross-sectional observation surface size information,etc.,and can detect all five weld quality parameters;due to the current pressure vessel weld surface quality parameters standard for weld single parameter measurement index definition,for weld defect parameters under the coexistence of feature points do not exist and difficult or impossible to calculate the problem;different diameter pressure vessel welds vary greatly The classical image feature point extraction algorithm for structural light vision sensing of weld parameters requires the design of different feature point detection algorithms.For deep learning image feature point extraction algorithm network structure deepening,need to further study real-time enhancement methods,etc.The paper investigates the selection method and measurement calculation method of weld quality parameters under the coexistence of single or multiple defects.The analysis of the active visual parameter feature points under the coexistence of multiple defects on the pressure vessel weld surface is carried out as a case study,laying the foundation for the study of the best feature point estimation method based on the corresponding parameter profile curve.In view of the coexistence of multiple defects on the surface of the pressure vessel weld using the standard recommended predetermined diameter sample method does not facilitate online measurement of the peaking h_peaking problem,from the definition of the angle of Class A and B welds,to study the active visual angle measurement calculation method under the coexistence of multiple defects on the surface of the pressure vessel weld.Using actual pressure vessel welds and joints as objects,it is verified that this paper’s method of selecting and measuring and calculating quality parameters under multiple defects on the surface of active vision sensing welds can achieve the detection and calculation of all five quality parameters of class A longitudinal welds and class B girth welds of pressure vessels.The paper investigates a deep learning-based method for extracting surface parameters of pressure vessel welds.Based on the idea of"multilayer CNN downsampling+deconvolution upsampling structure",we propose the structure design,principle and derivation of loss function calculation formula for the encoding-decoding image feature point extraction network（EDE-net）and the high-resolution module image feature point extraction network（HRM-net）.The proposed method of quality parameter image enhancement under the coexistence of multiple defects on the surface of pressure vessel welds for the problem of manual labeling of feature point location coordinates in each image,which has a large workload,time consuming,and sometimes a limited number of samples.Using actual pressure vessel welds as objects,it is verified that this paper’s active vision sensing weld surface quality parameter depth feature point extraction network,training image enhancement method for parametric weld profile simulation,and correction method to improve feature point extraction accuracy can achieve pressure vessel class A longitudinal weld and class B ring weld quality parameter feature point extraction.The paper investigates the active vision pressure vessel weld quality parameter feature point extraction real-time performance enhancement method.Adopting KD+structured pruning idea to compress the pressure vessel feature point extraction network weights and construct a framework for real-time performance enhancement of active vision pressure vessel weld seam quality parameter feature point extraction.Study the implementation method of deep feature point extraction network compression based on knowledge distillation KD,and fuse the output feature map of one stage of Resnet50 with the output feature map data of the later stage only using the adjacent stage feature map data fusion algorithm.To study the structured channel pruning method for deep feature point extraction network,to study the optimal compression method for each layer of convolutional weights given the overall compression rate CR of the network,to find the optimal compression rate CR^t for each layer of weights,and then to prune the convolutional weights based on CR^t to complete the compression of single layer of convolutional weights.Taking the pressure vessel class B ring weld as the object,we verify the effectiveness and correctness of the active vision pressure vessel weld quality parameter feature point extraction real-time performance enhancement method proposed in this paper-KD+structured channel pruning method.Combined with the former General Administration of Quality Supervision,Inspection and Quarantine Science and Technology Program Project"Development and Research of Portable Laser Vision-based Weld Seam 2D Morphometry"（No.2017QK105）and the State Administration of Market Supervision and Administration Science and Technology Program Project"Weld Seam Surface Morphology AI for Special Equipment+detection system development"（No.2019MK143）,to carry out general vision imaging weld seam two-dimensional morphology measurement experimental device,high-precision laser profiler weld seam three-dimensional parameter detection device development,and for the detection of pressure vessel weld surface quality parameters verification,the results show that this paper based on deep learning machine vision pressure vessel weld parameters detection technology,with validity,applicability and accuracy.