Research On Weld Quality Inspection Technology Based On Vision

Today the welding quality inspection is mainly achieved by manual.This method has many disadvantages,for example higher labor costs,longer time,lower efficiency and strong subjectivity.Welding quality testing simple relying on workers has been unable to meet the requirements of modern welding technology.Machine vision has been widely used in industrial inspection because of its high sensitivity,high procision and no contact with the workpiece.In this paper,the machine vision technology is applied to the welding quality inspection,it can be used to measure the geometric shape of welding and identify and classify the defects on the surface of the shell-tube welding.The main research work and innovations are as follow:Aiming at the centerline extraction of structured light images representing weld forming dimensions,a method combining Hessian matrix method with centroid method is proposed.First,the Hessian matrix method is used to calculate the normal direction of the structured light stripe,and the extremum point of the gray level distribution is calculated in the normal direction to determine the central point.Then,the centroid method based on the square weight of the gray value is adopted to extract the centerline of the structured light.When the center line is obtained,the abnormal points are eliminated.Afterwards,the center point eliminated are filled by linear interpolation.So,the connectivity of central line is better.The experimental results show that this method has high precision,less time consuming and high stability.For the feature points extraction,the RANSAC algorithm is used to fit the structured light.Compared with the least square method,the method has the advantages of strong anti-interference ability and high accuracy.At the same time,this paper puts forward the method of halving the original data set to solve the problem of long fitting time,which can ensure the accuracy and greatly improve the fitting speed at the same time.For the extraction of feature points,the extreme points are detected to achieve rough location.Then a method based on moving vectors distance is used to locate the feature points precisely.By analyzing the typical defects on the weld surface of shell-tube welding,their unique characteristics are obtained.Then the machine vision is used to detect defects on the surface of shell-tube welding for the first time.A weighted two-dimensional Renyi entropy algorithm based on gray level-strength of gradient excitation co-concurrence matrix is proposed to solve the segmentation problem of shell-tube welding defects.The traditional two-dimensional entropy does not consider the influence of edge and noise points,thus the segmentation results obtained is sensitive to noise.In this paper,a two-dimensional Renyi entropy algorithm based on gray level-strength of gradient excitation co-concurrence matrix is proposed.This matrix takes full account of the main direction and gradient intensity of the neighborhood gradient of pixels,and uses a new two-dimensional entropy partitioning method to partition the matrix,which is robust to noise points.At the same time,a Renyi entropy algorithm for weighting local entropy and joint entropy is proposed in order to segment the image target more accurately,which has achieved good results in the segmentation of weld surface defects in shell-tube welding.In view of the classification and identification of weld defects in shell-tube welding,the geometric characteristics and shape characteristic parameters of each defect are calculated.In this paper,a special characteristic parameters,image acquisition method,is developed for the shell-tube welding,and the effective features of each defect can be selected from these features.For the defect classification,some traditional classification methods are attempted to complete the defect classification,such as decision tree,BP neural network and multiple classifiers method integrating them.Because the traditional classification methods have low recognition rate,a binary Tree SVM classification algorithm based on bat algorithm is proposed.For the task of classifying welding surface defects,a SVM classifier based on binary tree structure is established from the point of view of SVM multi classification algorithm.To optimize the parameters of SVM classifier,an optimization algorithm based on bat algorithm is proposed.The BA algorithm introduces the speed weight into the bat's flight speed,so that each individual can adjust the flight speed according to their fitness.At the same time,the global search ability of bat algorithm is improved by increasing the frequency range.The binary tree SVM classification algorithm based on bat algorithm has achieved higher recognition rate in the classification of weld surface defects in shell-tube welding.Finally,in view of the identification of weld defects in shell-tube welding,a new algorithm based on depth convolution neural network for surface defect recognition is developed.In order to train a network model with good classification performance and robust,defect database Welding Defect Dataset is carried out extended operation.After analyzing classification of surface defects in shell-tube welding,an improved Le Net-5 model with 10 layers is proposed.In this network,the maximum value method is used to replace the previous average method in pooling level,so that the network can further highlight the characteristics of the image.The activation function of each layer becomes Re LU function,which can suppress and activate the neuron.Meanwhile,it can reserve the useful features of the data to the greatest extent,accelerate the convergence speed of the training and solve the problem of “gradient disappearance”.By training the model,the accurate classification and identification of the surface defects of shell-tube welding are completed.The experimental results show that the improved Le Net-5 model presented in this paper has a good performance in the recognition and classification of weld defects,the correct rate is 96.34% and meanwhile the convergence effect is also very good.