A Novel Supervised Chatter Detection Method Based on Noise-Added ResNet-AutoEncoder

Thin-walled components, widely employed in aerospace and high-precision industries for their lightweight and high specific strength, are highly susceptible to chatter—a self-excited vibration that severely degrades machining quality and tool longevity. Traditional chatter detection methods often rely on labor-intensive preprocessing, manual feature selection, and complex classifiers, limiting their practicality in real-time applications. This study proposes an innovative ResNet-AutoEncoder-based framework integrated with noise-added training to address these challenges. By directly processing raw multi-channel sensor data (triaxial force and acceleration signals) without complex preprocessing, the model leverages ResNet’s residual learning to stabilize gradient flow and enhance hierarchical feature extraction. A novel noise-added strategy during supervised training forces the model to disentangle noise-invariant core features of stable states, enabling robust differentiation from chatter using simple reconstruction metric thresholds (e.g., MSE). Experimental results on γ-TiAl alloy thin-walled components demonstrate exceptional performance: 98.47% chatter detection accuracy: 97.15% stable-state accuracy, and 0.0367-second average decision latency, meeting real-time industrial requirements. Theoretically, this work redefines noise as a dual-functional tool—simultaneously improving robustness and amplifying feature discriminability. Practically, the framework eliminates dependence on expert-driven feature engineering, offering a scalable, efficient solution for real-time chatter mitigation in smart manufacturing systems. Furthermore, we validated the proposed method across diverse machining environments and material types, achieving remarkable recognition accuracy of 96.10% for chatter states and 96.22% for stable states. These consistent results demonstrate the method's exceptional robustness under varying operational conditions.