Research On Roughness Recognition Technology Of Convolutional Neural Network Based On Microscopic Image Of Machining Surface

The micro-geometric features of the surface of the machined workpiece greatly affect the physical properties such as wear resistance,corrosion resistance,contact stiffness and fatigue resistance,as well as the functions of service life and reliability.The detection efficiency and accuracy of the surface roughness of the workpiece are increasingly higher.Since the traditional contact and non-contact measurement will leave scratches on the surface to be tested,the efficiency is low,the principle is complicated,and the operating environment is high,so the surface roughness measurement method based on machine vision has been widely used in recent years,but The machine vision-based measurement method requires manual design of image features during the analysis process,which is costly and cannot meet the actual needs.Therefore,this thesis proposes an image recognition method based on deep learning.The surface roughness of the workpiece is detected from the surface microscopic image of the workpiece.Through experimental research,The precise classification from the microscopic image of the workpiece surface to the roughness level and the recognition of the two-dimensional roughness parameter Ra and the three-dimensional roughness parameter Sq,Sa are realized.The structure and principle of convolutional neural networks are analyzed.The incomplete connection,weight sharing and pooled sampling structure of convolutional neural networks provide irreplaceable superiority in image classification and recognition,and its unique network structure improves the computing speed of the network.The working principle of convolutional neural network in image classification and recognition is studied.The convolutional neural network is composed of multiple network layers connected in series.The image features automatically extracted by the network are abstracted with the increase of network depth,and the depth feature information improves the invariance of the convolutional neural network to the geometric distortion of the image.In the network training analysis,the over-fitting problem and the solution method in the network training process are discussed.The improved LetNet-5 model is used to simulate the classification and recognition of handwritten digits,and the classification accuracy is obtainedMicroscopic images and topography of different roughness surfaces of each sample were obtained by microscope,and the two-dimensional image data was preprocessed to obtain two-dimensional grayscale images of each surface.According to the surface topography data,the contour curves of each surface are obtained,and the periodicity of the surface curve is analyzed to determine the size of the sample image.A convolutional neural network sample database is established,and the sample database is augmented with noise samples and rotated samplesA convolutional neural network model was established and the experimental environment and network performance evaluation criteria were analyzed.The variation of the network performance of the network model on the turning sample database with network parameters(network depth,filter size,number of filters,training batch and sparsity rate)is obtained by single factor experiment.According to the influence of network parameters on network performance,the optimal parameter model is obtained.The optimal network model is used to classify and identify the two-dimensional roughness level of each sample database,and the two-dimensional roughness classification accuracy of each sample database is obtainedA convolution regression network model is established,and a sample database with the two-dimensional roughness parameter Ra as the network output reference value is constructed.The convolution regression network is used to carry out the training test on each sample database to obtain the surface roughness of each sample.The recognition results show that the convolution regression network can effectively and accurately identify the two-dimensional roughness parameters of the sample surfaceA sample database with a three-dimensional roughness parameter Sq,Sa as a network output reference value is constructed.The training test was carried out on each sample database by using the constructed convolution regression network,and the three-dimensional roughness parameter recognition results corresponding to the surface images of each sample were obtained.The experimental results show that the convolution regression network can effectively and accurately identify the three-dimensional roughness parameters of the sample surface.