Abstract
Reliable tool-wear monitoring is essential for maintaining machining quality and preventing unscheduled downtime in manufacturing. This investigation presents a sound-based classification framework for identifying wear states in the turning of AISI 316L stainless steel using advanced gradient-boosting models. Acoustic signals were recorded under constant cutting parameters to eliminate process-induced variability, and each recording was divided into standardized 2 s segments. A total of 540 multidomain features-including RMS, ZCR, spectral descriptors, Mel-spectrogram statistics, MFCCs and their derivatives, and discrete wavelet energies-were extracted to capture both stationary and transient characteristics of tool-workpiece interactions. Feature selection was performed using a three-stage pipeline comprising Boruta, LASSO, and SHAP analysis, resulting in a compact subset of highly informative descriptors. LightGBM, XGBoost, and CatBoost classifiers were trained using stratified 10-fold cross-validation across three wear states: Unworn, Slight wear, and Severe wear. LightGBM and XGBoost achieved the best performance, with mean accuracies above 0.96 and strong PRC-AUC and ROC-AUC values (0.98-1.00). Although Slight wear remained the most difficult class due to its transitional acoustic characteristics, all models showed clear separability for Unworn and Severe wear conditions. The results confirm that boosted decision-tree methods combined with SHAP-enhanced feature selection provide an effective, low-cost, and non-contact solution for tool-wear classification in 316L turning.