Machine Learning-Based Cutting Constant Estimation for Mechanistic Force Models of End Milling Operation

Abstract

A predictive cutting force model is essential for power requirement estimation, cutting tool design, surface error estimation and stability analysis during the end milling operation. Mechanistic model estimates cutting forces by correlating analytically computed chip geometry with lumped coefficients combining tool-work material properties through empirical relationships. Establishing reliable relationships through the statistical curve-fitting is demanding due to the need for several experiments, anomaly or noise in the experimental data, and process disturbances that deteriorate the goodness of fit. Machine learning models can effectively deal with such inherent uncertainties and serve as an alternative to the statistical curve-fitting. The present work proposes to improve the empirical relationship between instantaneous uncut chip thickness and cutting coefficients by employing a deep learning algorithm, namely Adaptive Moment Estimation (ADAM). The ADAM algorithm is augmented with decoupled weight decay and warm restart features for the improved performance. The decoupled weight decay assigns dynamic sensitivity values to the data points for outlier removal resulting in better model generalization, while warm restart allows better guesses through adaptive learning rates. The proposed approach has been implemented as a computational tool to determine improved coefficients values and empirical relationships. The cutting forces predicted using coefficient values determined using statistical curve fitting and ADAM-based machine learning are compared with experimentally measured data over an extensive range of cutting conditions. It is concluded that the augmentation of the ADAM approach enables the Mechanistic force model to effectively capture end milling process physics by estimating better coefficients values resulting in enhanced prediction abilities.