Resumo:
Prediction of tool life is important to ensure conformity of the product and to avoid
damage to the product itself and to the machine. Surface roughness is an essential requirement
for machined products. The main figure for that requirement is surface roughness. Accurate
forecasts of tool life may lead to cost reduction and increase in productivity. Prediction of
roughness may contribute to improve product quality and to reduce production times and
costs. To perform such predictions is a difficult task due to intrinsic non-linearity that
characterizes the processes involved. Neural networks are proven tools for prediction in
processes involving non-linearity, as it is the case for prediction of tool life and roughness.
This work is a study on the performance and variability of RBF (Radial Basis Function)
neural networks, projected with support of the methodology of Design of Experiments (DOE)
applied to prediction of tool life and surface roughness in the turning of a SAE 52100
hardened steel (55 HRC) with mixed ceramic tool. The factors employed in experimental
planning are parameters of project of the networks (number of radial units, algorithm for
determination of radial centers and algorithm for determination of the spread factor). Cutting
process parameters are employed as input for the networks. The output variable chosen to
measure network performance is the S. D. Ratio obtained during the testing phase of the
networks. Effects of project parameters and of size of training set on network performance are
investigated. In order to do that, experiments with training sets of different sizes are
conducted. Possible effects of interaction among factors are also investigated. Results are
expressed as network project parameters, for each training set employed. The best networks
obtained by means of the proposed method show accuracy and precision increasingly better as
the number of available examples for training grows up. Other results are the ranking of
relative importance of project factors in network performance. The work concludes that
interaction effects among levels of factors involved are statistically significant to the
performance of RBF networks in the proposed tasks. A comparison between RBF networks
and a linear optimization method points to superior performance of the networks in modeling
tool life and roughness. Conclusions suggest that the approach of using the methodology of
Design of Experiments (DOE) as a tool for project of networks for prediction of tool life and
surface roughness may constitute a better option than the trial and error approach or than the
strategy of varying one factor at a time in the search for high performance network
configurations.