Mirror milling chatter monitoring method for large ultra-thin parts based on attention representation enhancement, chatter fundamental frequency feature extraction and improved capsule network

Mirror milling of large ultra-thin curved components is a typical time-varying dynamic process, in which material removal, support position, and support force easily induce chatter, resulting in surface quality degradation and tool wear. Existing offline prediction methods are costly and difficult to apply in real time, while traditional machine-learning and common deep-learning models show limited generalization under variable stiffness and changing machining conditions. To address these issues, this study proposes an Adaptive Chatter Monitoring Network (AOC-Net) based on attention representation enhancement, chatter fundamental frequency feature extraction, and an improved capsule network. The attention representation-enhanced encoder combines channel-wise and temporal-wise attention to enhance chatter-sensitive information in multi-channel sensor signals. An operational modal analysis (OMA)-based state-space model is then established to adaptively extract chatter fundamental frequency features from the singular values of the Hankel matrix, thereby representing vibration intensity and time-varying dynamic characteristics. Finally, a capsule network with vector neurons and dynamic routing is used for chatter state classification. Experiments on 2A12 aluminum-alloy curved thin-walled components yielded 6000 labeled samples for model training and testing. Ablation studies show that the attention encoder, OMA-based feature extraction, and capsule network improve test accuracy by 9.10%, 4.80%, and 1.97%, respectively. Compared with five mainstream methods, AOC-Net achieves an average accuracy of 99.09% in the primary mirror-milling scenarios and maintains superior performance under small-sample conditions. Further validations involving cross-condition, cross-material, cross-system, coupled domain-shift, noise-injection, and cutting-fluid flow-rate tests demonstrate the transferability and robustness of the proposed method. The results indicate that AOC-Net can effectively capture chatter-sensitive time-varying dynamic features and provides a reliable solution for real-time chatter monitoring in mirror milling of large ultra-thin components.