Research On Detection Algorithms Of Weld Surface Defects Based On Machine Vision For Assemble Structural Parts

As a common structural part on the harbor facilities,assembled structural parts are usually manufactured by welding techniques.The box girder structural part of the derrick tower is a typical assembled structural part,whose main function is to increase the torsional strength of the integrated assembled structural part which is made up of beam rib,web,bottom plate and so on while the main load passes through any position on the beam.However,in the process of manufacturing structural parts,due to the complexity of assembly and the existence of uncertain factors such as welding environment,it is easy to produce weld defects such as overlap,deviation,porosity,splash and so on.Therefore,it is necessary to detect the defects on weld morphology of assembled structural parts.In the traditional defect detection,it mainly relies on manual detection,which is easily limited by the experience of testing personnel and the unpredictability of the construction period and other deficiencies,and cannot meet the requirements of modern welding technology.Machine vision has been widely used in industrial inspection because of its high sensitivity,high precision and no contact with the workpiece.In this paper,it is proposed that a new method of the defect detection based on machine vision for the weld surface defects of assembled structural parts.The main research contents are as follows:(1)Aiming at the problem of the existence of noise and not much difference of the contrast in the gathered image of weld surface,an appropriate image processing algorithm is selected to remove noise by image filtering experiment,and the contrast between the parent metal and the weld area is enhanced by sin function.The binary image of the weld is obtained by using OSTU segmentation algorithm.Finally,the gray proportion of the weld defect area in the whole image is used to extract the weld area and the defect area.Experimental results show that this series of image preprocessing algorithms can be used to segment and extract the weld surface defect areas of assembled structural parts.(2)Considering that the batch algorithm of 2DPCA is unable to realize the online feature extraction and the drastic changes in weld defect images have very big impact on the convergence of feature vectors,a feature extraction algorithm of weld surface defects based on the GI2 DPCA is proposed.This algorithm can estimate online the real principal component in an iterative way with good convergence,and can reduce the impact of drastic changes of weld surface defects on the convergence of the principal component convergence.Experimental results show that the performances of this algorithm are better than 2DPCA,LBPH and LDA in terms of defect recognition and required CPU time and memory.(3)In order to meet the requirement of classification and recognition rate of weld defect detection,KNN,BP neural network and SVM are used as the classifiers for defect recognition.The defect feature is extracted by the GI2 DPCA algorithm which is as input vector of these three classifiers to achieve the defect recognition.Experimental results show that the performances of SVM are better than KNN and BP neural network in the aspect of weld defect recognition,it is the best choice to use the GI2 DPCA algorithm and the SVM classifier for the defect recognition.(4)In order to solve the problems of the new defect category and the update of feature space in the process of weld defect detection,the input of new defect category is monitored by the use of the reconstruction error between new input and reconstruction one using current eigenvectors,and the inner-class and inter-class distance is used to monitor the appearance of new positions of the known defect categories.The feature space is optimized online which can improves the information of defect detection.Experimental results show that the reconstruction error increases sharply after the input of new defect category.In the process of the detection,the feature space is constantly updated because of the appearance of new positions,the information of the online defect detection is gradually improved,and the recognition rate of the defect detection meets basically the requirement in actual industry.