VASCONCELOS, Guilherme Augusto Vilas Boas; http://lattes.cnpq.br/8922649967544135
Resumo:
The growing demand for energy efficiency and cost reduction in manufacturing processes
has driven the need for optimization in operations such as end milling, especially when machining
complex materials like duplex stainless steel UNS S32205. Although this material
exhibits excellent mechanical strength and corrosion resistance, it has low machinability
due to its tendency to work harden and its low thermal conductivity—factors that
compromise process stability and directly affect surface finish quality. In the end milling
of this steel, the precise definition of cutting parameters becomes even more challenging
due to the presence of noise variables inherent to the process. Given this complexity, the
application of machine learning models emerges as a promising approach, provided it is
combined with appropriate techniques to construct robust models, that is, models capable
of maintaining high predictive performance even in the presence of noise. This thesis
aims to develop and validate machine learning models for predicting surface roughness
(Ra) in the end milling of duplex stainless steel UNS S32205, simultaneously considering
control variables and noise variables. The experiments were conducted based on a
central composite design, combining the input variables (cutting speed, feed per tooth,
width of cut, and depth of cut) with noise variables (tool flank wear, fluid flow rate,
and tool overhang). The study applied the concept of robust parameter design, response
surface methodology, and machine learning techniques. The models used included Support
Vector Machines (SVM), Decision Tree Regressors (DTR), Random Forests (RF),
and Artificial Neural Networks (ANN). Performance evaluation metrics included RMSE,
MAE, MSE, and cross-validation using average R2. The ANN models showed the best
performance, with the 7-20-14-1 architecture standing out, achieving an R2 of 0.92, RMSE
of 0.06, and average R2 of 0.90 in surface roughness prediction. The ANN model was then
used as the objective function in an optimization process conducted using the Particle
Swarm Optimization (PSO) algorithm, aiming to identify cutting parameters that minimize
roughness. Confirmation experiments, performed using a Taguchi L9 array, validated
the obtained results, demonstrating that the selected machine learning model was effective
in simulating and optimizing the real behavior of the process.