Research On The Measurement System Of Video Extensometer Based On Deep Learning

In the field of sheet metal material forming,mastering the mechanical properties of metal materials is the basis for realizing forming.For exploring the forming mechanism of anisotropic materials,it is an important topic to use a biaxial tensile testing machine to obtain biaxial tensile mechanical properties under different loading paths.In order to obtain accurate mechanical performance parameters of materials,the strain measurement of the biaxial tensile testing machine has become one of the cores of equipment development.In this subject,the goal is to design and develop a noncontact video extensometer with a biaxial tensile testing machine’s strain measurement control.The use of deep learning methods and image recognition can achieve highprecision strain field measurement.The video extensometer hardware consists of an industrial camera,a fill light and a bracket.On the basis of the self-developed biaxial tensile testing machine,the Solid Works software was used to complete the structure design of the bracket,and the fixture design was optimized to meet the tensile test of various specimens.The image obtained by the video extensometer during measurement has a lot of noise and different brightness in the image due to the difference in ambient light and temperature,which makes the traditional pattern recognition method less stable in image recognition.In order to reduce the influence of ambient light,the video extensometer bracket adopts a closed space structure and uses LED light sources inside to provide a stable light environment for image collection.In order to enhance the robustness of image recognition,deep learning image segmentation methods are used to improve the accuracy of image recognition through data set training in different environments.Since the video extensometer marks the deformation area of the specimen before the experiment,the deformation amount is obtained by recording the position change of the mark area.This topic combines two experiments of unidirectional stretching and bidirectional stretching,of which the marking distance of the marking area is 50 mm in the unidirectional stretching experiment.Two marking methods are used in the biaxial stretching experiment.One is the traditional marking,in which the marking area is a square area of 60mm×60mm.The other is the speckle marking method.By spraying speckle paint on the surface of the test piece,several local feature regions and global feature regions are made to complete the acquisition of local deformation and global deformation.Aiming at the acquisition of the marked area of the test piece,this topic is based on the deep learning image segmentation algorithm to complete the acquisition and segmentation of the marked deformation,through the convolutional neural network in the down-sampling area to complete the feature acquisition and encoding of the input image,and through the up-sampling area Bilinear interpolation and deconvolution complete the expansion and decoding of the specific features of the image,and then obtain the image segmentation result.In order to prevent the loss of some features in the feature expansion,the method of layer jump connection is adopted,and the shallow features of the same dimension in the down-sampling area are copied,and the current deep features are merged to form the current data,ensuring that the data is not lost and improving the computational efficiency of the model.After training,the function of obtaining the strain area of the specimen through image segmentation is completed.