Chatter Detection in Micro-Milling Using Stacking Ensemble

Existing chatter detection models are established based on single classifiers and homogeneous ensemble classifiers, but the lack of diversity in these models leads to weak feature capturing and limited generalization capabilities. This article proposes a micro-milling chatter detection model based on stacking ensemble learning with diverse classifiers, aiming to enhance the model’s generalization capability. The model captures data features from multiple perspectives to accurately classify machining states, including stable, slight, and severe chatter. The collection and processing of vibration signals are methodically established to obtain representative samples across different machining states. Base models are developed by training the samples through various classifiers: classification and regression tree (CART), k-nearest neighbor (KNN), feed-forward neural network (FNN), and gate recurrent unit (GRU). The final chatter detection model is constructed by integrating the outputs of these base models through a meta-classifier. The proposed model is rigorously validated through extensive micro-milling experiments, achieving a detection accuracy of $(98.8 \pm 0.4)\%$ , with a 95% confidence interval (CI) of $[98.4\%, 99.2\%]$ . Its generalization capability is further evaluated under four different conditions: the Al-7050 workpiece but with different machining parameters, the Ti–6Al–4V workpiece, Tool 1, and Tool 2, demonstrating accuracies of $(94.2 \pm 0.6)\%$ , $(90.3 \pm 1.2)\%$ , $(90.3 \pm 1.6)\%$ , and $(90.0 \pm 1.4)\%$ , respectively, which significantly outperform those of support vector machine (SVM)-based models (e.g., traditional SVM and AdaBoost-SVM).