Physics-Informed Machine Learning for Chatter Detection in Thin-Walled Cylinder Turning of 1.4301 Steel

In thin-walled cylinder turning, wall thickness decreases with each pass. It reduces workpiece stiffness and shifts the stability limit of the cutting process. In this study, wall thickness and axial segment position are established as important factors that affect chatter occurrence in this geometry. To the best of the authors’ knowledge, there are no prior studies that used these geometric features as machine learning input. The objective of this study is to present a physics-informed framework for chatter detection in the thin-walled cylindrical turning of 1.4301 austenitic stainless steel. Five physics-informed features were extracted per segment: RMS resultant force, kurtosis, dominant non-harmonic frequency, normalized segment position and wall thickness. Four classical classifiers—Random Forest, Logistic Regression, Support Vector Machine (SVM) and Neural Network (NN)—are evaluated on 135 segments of 15 machining passes. For validation, Leave-One-Pass-Out (LOPO) cross-validation was performed. It holds out segments of each pass to reflect deployment conditions. Random Forest and Logistic Regression achieved above 95% recall and accuracy, exceeding the 90% safety threshold. A feature ablation study was performed to measure the importance and impact of individual input features; the results suggest that using geometric features achieves 100% recall with three classifiers. In addition, three validation schemes, LOPO, forward chaining, and fixed split, were used to evaluate the generalizability of the developed model. The non-tree classifiers ranked wall thickness and RMS cutting force as strong predictors, whereas the tree-based ensembles relied mostly on RMS. The results suggest that physics-informed feature engineering with classical ML is a promising approach for chatter detection in thin-walled cylindrical turning.