Milling tool wear state classification and prediction using machine learning, homogeneous and heterogeneous ensemble learning approach

Tool wear has a significant impact on product quality, process reliability, and overall productivity in machining operations. Although the prevailing data-driven approaches are dominantly dependent on a single dataset, traditional Pearson’s Correlation Coefficient (PCC)–based feature selection, or stand-alone machine learning models, their performances are often restricted by feature redundancy and limited model diversity. In this context, this article presents a dual-dataset validated, hybrid feature selection technique that combines PCC with Random Forest (RF) importance scores, along with a heterogeneous stacking ensemble paradigm for classification and regression tasks. The proposed approach has been evaluated using the NUAA Ideahouse dataset and an in-house experimental milling dataset based on comprehensive time-domain and frequency-domain feature extraction and preprocessing. The proposed Stacking Ensemble Learning Classifier (SELC) achieves accuracies of 0.92 and 0.95 for tool-state classification and outperforms stand-alone machine learning and ensemble learning baselines. In the case of tool wear prediction, the Stacking Ensemble Learning Regressor (SELR) achieves the highest R² values of 0.90 and 0.92 for the two datasets. The results demonstrate that the combined hybrid feature selection and stacking-based architecture has greatly enhanced generalization and prediction performance compared to conventional methods in the literature.