Detection Framework for Milling Chatter Based on Improved Coati Optimization Algorithm

Milling chatter is a self-excited vibrational phenomenon that leads to deterioration of machined surface quality, reduction in tool life, and decreased processing efficiency. Current chatter detection methods suffer from three major limitations: 1) signal noise interference results in reduced detection precision; 2) high-dimensional features cause extended training time; and 3) deep learning hyperparameter selection relies on experience. To address these critical issues, this article proposes a comprehensive chatter detection framework based on an improved coati optimization algorithm (ICOA). The ICOA effectively enhances the algorithm’s global optimization capability and convergence speed through the introduction of refraction-based reverse learning, a reinforced Lévy flight mechanism, and a variable spiral search mechanism. A combined noise reduction method of variational mode decomposition-wavelet threshold (VMD-WT) optimized by ICOA was employed, which improved the signal-to-noise ratio (SNR) and correlation coefficient by 16.8% and 22.8%, respectively. This effective removal of noise from vibration signals enhanced processing state detection accuracy by 2.16% compared to the original signal. In feature selection, ICOA demonstrated exceptional feature screening capability, enabling the selection of fewer, yet more effective, feature subsets in both University of California Irvine (UCI) datasets and chatter detection applications. For chatter detection, 11 critical features were optimally selected from a 20-D feature set, reducing training time by 12.5% while maintaining classification performance. A chatter detection model iTransformer-BiLSTM [(iTBM)] was constructed based on iTransformer and bidirectional long short-term memory network (BiLSTM) architectures, with hyperparameters optimized through ICOA. Compared to other advanced methods, the proposed approach exhibited significant improvements in metrics such as accuracy. Moreover, the proposed method requires only 38.88 ms to process 40 samples in practical prediction, fully satisfying real-time monitoring requirements. This research provides a novel technical solution for enhancing the accuracy and real-time performance of chatter detection in milling processes.