Intelligent chatter detection in micro-milling integrating self-attention-CNN feature fusion and optimized ensemble SVM

Chatter detection in micro-milling is essential because of its detrimental impact on surface integrity, tool life, and machining efficiency. Although machine learning (ML) approaches have shown promise in traditional machining, the unique dynamics of micro-milling pose additional challenges. This study proposes a novel online chatter detection framework that integrates deep feature fusion with ensemble learning. The methodology begins with generating Short-Time Fourier Transform (STFT) spectrograms from three-directional cutting force signals. A three-stream convolutional neural network (CNN) enhanced by a self-attention mechanism is constructed to extract and fuse discriminative features. The self-attention module captures long-range dependencies and emphasizes critical regions, while a convolutional fusion block consolidates the multi-channel representations. Although the CNN-FC model achieves classification accuracy of above 90 %, its performance is not sufficiently robust due to the limited capability of fully connected layers in handling imbalanced datasets. To address this, the extracted features are fed into an ensemble support vector machine (SVM) classifier, optimized using Particle Swarm Optimization (PSO) and Atom Search Optimization (ASO). This ensemble strategy significantly enhanced classification robustness, achieving a peak accuracy of 99.30 %. The dataset comprises 300 micro-milling experiments under varying cutting parameters, with chatter states labeled via frequency-domain analysis and surface morphology inspection. Comparative evaluations confirm that the proposed fusion of self-attention CNN and optimized ensemble SVM outperforms conventional architectures, offering a reliable and scalable solution for real-time chatter detection in micro-milling.