Predictive Modeling of Surface Roughness in Machining Processes Using Random Forest Regressor

Development of machine learning, specifically the Random Forest Regressor (RFR), to predict surface roughness (Ra) in machining processes. The study considered machining forces (Fx, Fy, Fz, and the resultant force F) and the main cutting parameters: depth of cut (ap), cutting speed (vc), and feed per revolution (f). The methodology involved a set of 612 experimental data points, distributed between Experiment 1 (New Tool) and Experiment 2 (Worn Tool). Statistical analyses, including Pearson correlation, were performed to assess the relationships between the variables. The model was trained and validated using statistical metrics widely applied in regression problems, along with graphical visualizations that aid in interpreting its predictive performance. The prediction algorithm achieved high coefficients of determination R², being 0.9534 for Experiment 1 and 0.9643 for Experiment 2. Furthermore, it showed low predictive errors, with MAE of 0.0406 (Exp1) and 0.0271 (Exp2), and MAPE of 6.43% (Exp1) and 4.63% (Exp2). The model's performance was evaluated using scatter plots, which demonstrated a strong alignment between the predicted values and the actual roughness values. The histograms of the residual errors showed an approximately normal distribution centered at zero, indicating a well-fitted model without significant biases. The analysis of variable importance revealed that feed per revolution (f) was the most influential factor in predicting roughness, especially in Experiment 1, where it represented 58.08% of the model's total contribution. In Experiment 2, due to tool wear, the influence of feed was reduced to 38.16%, while the machining forces, used as factors in the model, particularly the components of feed (Fy) and passive (Fz), had, collectively, a greater impact on the final roughness. The results demonstrate that RFR is an effective tool for predicting surface roughness.