Research On Visual Inspection And Automatic Defect Identification Of Weld Surface Shape

In the field of mechanical manufacturing,welding,as one of the most widely used processing technology,to a great extent to promote production and promote the industrial process.The welding process is highly susceptible by external environment,welding parameters and machining accuracy,and a series of weld surface defects often appear,which directly affect the quality of weldments.Weld surface quality monitoring can effectively improve production efficiency and ensure the quality of welding parts.Laser vision,as a high precision,simple structure of the detection method,has been research hotspot in the field of industrial inspection.In this paper,laser vision technology was used to detect the forming quality of weld surface,the laser stripe image obtained by laser vision sensor is used as the information source,to carry out research on weld surface shape dimension measurement and surface defect intelligent detection methods.In order to obtain the original laser stripe image representing the weld surface contour,this paper built a set of visual inspection device for weld surface formation with the help of laser vision sensor.The composition and structure of the visual inspection device for weld were discussed in detail,and the conversion model between image coordinates and world coordinates was established.The transformation relation between the image coordinates of the weld characteristic points and the world three-dimensional coordinates is obtained.With butt weld and fillet weld as the research object,according to the imaging characteristics of laser fringe image of weld seam,the image preprocessing process was combined and optimized.The median filtering,threshold-based feature enhancement,line break compensation,noise elimination and other pre-processing work were successively carried out.At the same time,a center line extraction algorithm combining wavelet multi-scale modulus maximum edge detection with weighted gray gravity center method is proposed,improving the universality and accuracy of the center line extraction algorithm.According to the stripe image characteristics of different weld types,the base metal profile is fitted by using Random Sampling Consension(RANSAC)algorithm,and the Angle threshold is set to identify the weld types.Aiming at butt welds,an extraction algorithm combining fitting method and corner point recursion method was proposed to extract weld feature points by multi-area detection from weld center data.For fillet weld,the position of the left and right feature points of the weld can be determined according to the slope distribution of the laser curve.and the remaining high feature points are searched in the left and right intervals to complete the location of the fillet weld feature points.Finally,the overall forming size of the weld was collected by scanning and the 3D reconstruction of the weld forming contour was realized.In order to identify and classify the common defects on the weld surface,machine learning algorithm and deep learning algorithm were applied to study the laser vision intelligent detection method of weld defects.The detection method of weld surface defects based on machine learning adopts the Histogram of Oriented Gradient(HOG)to extract the defect features of weld laser stripe image,and carries out optimization of SVM model parameters through five-fold cross-validation mesh search method.Finally,the recognition and classification model of weld surface defects based on HOG-SVM algorithm was established,and the overall recognition rate reached 97.86%.The detection method of weld surface defects based on deep learning takes transfer learning as the technical means,migrates Alex Net and Googlenet two network models,and combines Adam and SGD two optimization algorithms.After super parameter optimization,combination tests are carried out,and it was found that Alex Net network model with SGD optimization algorithm has the best performance.It has a good classification and recognition effect for weld defects(porosity,sag,bite edge,no defect),and the accuracy rate is as high as 99.07 %.