Open AccessDEVELOPMENT OF NEURAL NETWORKS FOR FORECASTING ROUGHNESS WHEN MILLING VARIOUS MATERIALS
Author(s) -
Е. Ерыгин,
T. A. Duyun
Publication year - 2020
Publication title -
vestnik bgtu im. v.g. šuhova
Language(s) - English
Resource type - Journals
ISSN - 2071-7318
DOI - 10.34031/2071-7318-2020-5-11-113-124
Subject(s) - artificial neural network , surface finish , surface roughness , materials science , tool wear , mechanical engineering , computer science , backpropagation , identifier , machining , artificial intelligence , metallurgy , engineering , composite material , programming language
The article presents a methodology for the development and testing results of artificial neural networks for predicting roughness in finishing milling. Experimental data of various researchers in the processing of materials with different physical and mechanical properties are used as the initial data base for the creation and training of neural networks. The value of material hardness is taken as the main identifier of physical and mechanical properties. In addition to the hardness of the material, the input parameters of the nets are also the cutting modes: tool feed, depth and cutting speed. The data of the obtained roughness are used for several groups of materials: non-ferrous metals, structural and stainless steels, heat-resistant alloys and tool steel. Nine highly specialized neural networks have been created that predict roughness when milling a certain material, a number of combined networks by combining several databases, including a broad-based neural network for several groups of materials. A comparative analysis of the results of testing the developed neural networks by the criterion of relative error is carried out. Most of the presented neural networks have a satisfactory error not exceeding 10%. Individual neural networks have higher accuracy, showing an error within 5 %.