Open AccessAPPLICATION OF ARTIFICIAL NEURAL NETWORKS FOR PREDICTING ROUGHNESS AT CLEANING AND GETTING POINT
Author(s) -
D. S. Baranov,
D. S. Baranov,
Татьяна Дуюн,
T. A. Duyun
Publication year - 2019
Publication title -
vestnik bgtu im. v.g. šuhova
Language(s) - English
Resource type - Journals
ISSN - 2071-7318
DOI - 10.34031/article_5d35d0b62dc823.22670125
Subject(s) - artificial neural network , sigmoid function , computer science , process (computing) , rake angle , surface roughness , rake , backpropagation , surface finish , convergence (economics) , point (geometry) , cutting tool , function (biology) , machining , algorithm , mechanical engineering , artificial intelligence , engineering , mathematics , materials science , geometry , evolutionary biology , economic growth , economics , composite material , biology , operating system
A technique for the development of artificial neural networks to predict the roughness of the treated surface during finishing and semi-finishing turning is presented. The back-propagation network architecture was adopted, having an input, hidden and output layers, a sigmoidal activation function for the hidden layer and a linear one for the output layer. To form a training sample, empirical expressions in the form of power functions were used, training of networks was carried out according to the Levenberg-Marquardt algorithm, which has fast convergence. Technological modes (cutting speed and depth of cut, tool feed), cutting tool geometrical parameters (main and auxiliary angles in terms of the tool, radius at the tip of the tool, rake angle), physicomechanical properties of the material being processed, each the training sample is formed from thousands of source data combinations. Separate networks have been developed that predict roughness during finishing and semi-turning turning, as well as a combined network that takes into account both types of processing. Analysis of the accuracy of the networks showed good results, the relative error of calculations does not exceed 1%. The proposed neural network models can be used in technological preparation of production, as well as in systems of adaptive control of the cutting process.