Research On Grinding Surface Roughness Measurement Methods Based On Transfer Learning

Surface roughness refers to workpiece surface unevenness,such as small spacing and micro peaks and valleys.Such unevenness directly affects equipment performance and life span.Accurate and efficient measurement is of great importance to modern industry.While the surface quality of the workpiece obtained by grinding is higher than that of other processing methods,As the final process of machining,the processing quality will directly affect the product aesthetics and roughness level,so the surface roughness of the grinding parts is accurate,efficient and reliable Measurement is of great significance to the promotion of manufacturing.In roughness measurement method,the merits of roughness measurement based on machine vision is obvious,such as high efficiency,non-contacting,good flexibility,high ratio of sexuality,high possibility of possession,etc.Many scholars in this study have achieved good results,but the roughness measurement method based on machine vision belongs to relative measurement cannot directly measure the specific value of roughness,it needs to build function relationship between the image features and surface roughness of the measured object,which requires a large number of samples with a wide range of known roughness at uniform intervals as input for training or fitting.However,it is difficult to meet the above requirements in the actual production and processing.So how to train the grinding surface roughness visual measurement model with superior performance under the condition of limited training resources is one of the frontier challenges.Based on the above-mentioned problems,this paper presents the method of measuring the surface roughness of the grinding surface,and the main research contents of this work are as follows:(1)The roughness measurement methods of stylus instrument and machine vision were introduced.The advantages and existing problems of the roughness machine vision measurement method were analyzed,According to the existing problems,a solution based on transfer learning is proposed.(2)A novel technical route of using simulation data as source domain and learning from it is presented to solve the problems that insufficient data in training model based on machine vision.A method of constructing surface roughness visual measurement simulation domain is proposed.In this method,firstly,we use the digital simulation of non-Gaussian surface and 3D solid modeling to generate the grinding surface virtual entity with arbitrary roughness.Then we can get the roughness image corresponding to the virtual entity by computer geometric optical simulation,and the feasibility of this method were proved through experiments.(3)A novel grinding surface roughness measurement method based on transductive transfer learning was proposed for the problems,especially when it is difficult to obtain a large amount of grinding samples with different roughness.Firstly,the labeled data in the simulation domain and the unlabeled data in the real domain were processed by the transfer kernel learning or transfer component analysis,and then the regression model was trained on the obtained kernel matrix.The experimental results show that the grinding surface roughness measurement method based on transductive transfer learning has a good measurement effect,and provides a new solution to solve this problem.(4)A novel grinding surface roughness measurement method based on inductive transfer learning was proposed to solve the problems on poor model generalization ability caused by lack of practical experimental data in machine vision measurement of grinding surface roughness,knowledge-leverage strategy was applied to learn the rich knowledge of data in simulation domain.The experimental results show that the simulation domain is rich in data resources,to some extent,can solve the problems that poor generalization ability caused by lack of practical experimental data,which provided a new technical route for continuous improvement of roughness machine vision measurement.