Research On Grinding Surface Roughness Detection Based On Machine Vision

With the rapid development of manufacturing industry,roughness,as one of the evaluation indicators of machined surface quality,is an indispensable part of the inspection.It directly affects the performance and life of the parts in the future.Therefore,the detection of roughness has great significance for the development of manufacturing industry.Traditional roughness detection technology cannot meet the intelligent and efficient needs,especially in mass production and roughness requirements for full inspection,so many scholars at home and abroad have been committed to the research of new detection methods for many years.However,after different processing methods,the surface texture of parts will show different layout.The texture of grinding surface is mainly random and undirected,which is a difficult point in the research of grinding surface recognition.For many years,the research of roughness detection methods has been limited by their own equipment and environment requirements,and cannot be applied accordingly;in the study of new detection methods,most of the extracted features are based on the linear relationship between roughness and surface texture features,and are not analyzed from the importance of features for recognition models;for the number of extracted features,there is no reasonable planning,which easily leads to the problems of complex learning of recognition model and the low recognition accuracy.Therefore,the subject takes grinding surface as the research object and by base of the linear relationship between the recognition features and roughness,and study from the importance of the recognition accuracy of the model and the number of extracted features on the BP neural network model.Firstly,five groups of grinding surface images with different roughness values are obtained through grinding experiments and testing experiments.Because the images contain noise due to the reasons of the equipment itself and the environment,the source and properties of noise are analyzed in this paper.Through the analysis,we know that the image is Gauss white noise.By the evaluation of PSNR and SSIM for different filtering methods,the bilateral filtering method is finally determined.On the basis of previous studies,21 relevant features are extracted from the filtered images through gray level co-occurrence matrix and Tamura texture feature.Because the previous research,which was based on the linear relationship of mathematical model,did not consider the importance for recognition model,and then did not pay attention to the differences in the direction of gray level co-occurrence matrix,this topic will aim at the shortcomings and the advantages of SBS algorithm and random forest algorithm to combine with the two algorithms.Then,we select 12 more important features for recognition model,and reduce the learning complexity of the model and analyze the differences between features.Finally,the BP neural network model is established to identify the extracted features.The experiment optimizes the parameters such as hidden layer number,node number,L2 regularization coefficient,number of training pictures and iteration times,which improves the recognition rate of the model to 98.07% and reduces the training time to 37.4s.Through the above research,finally,this topic has developed an efficient,fast and accurate detection method based on BP neural network model,which is suitable for grinding surface,and can eliminate noise and select features through feature analysis.