Mesure par apprentissage automatique des paramètres de rugosité des surfaces fraisées.

Abstract

Machine learning methods are gaining more and more space in basic sciences and engineering, the vast field of metrology is no exception because the challenge today is to integrate numerical methods of calculation in order to optimize the manufacturing and control process. The aim of this thesis is to set up methods for predicting the roughness of milled surfaces by machine learning. Machine learning models such as artificial neural network, multiple linear regression, vector support regression and decision tree regression have been developed. The cutting time, the percentage of carbon of the material machined and the nature of the tool were taken as input variables of the predictive systems. The Python programming environment was used for the development of the algorithms. The database used consisted of 3 datasets, one for high speed steel, another for tungsten carbide and a last combining the 2 previous datasets. The models were developed using 48 experiments for the combined dataset and 24 for the other 2 datasets distributed in 75% for the training phase, 25% for the test phase. The root mean square error (MSE) used to evaluate the performance of predictive systems gives a preferred choice to the ANN with the tungsten carbide tool as input with an MSE value = 0.017.