Chatter-sensitive feature construction and dual-branch learning for machining state identification in micro-milling of glow discharge polymer

Glow discharge polymer (GDP) ablators with micro-milled modulated microstructures are increasingly used in inertial confinement fusion (ICF) applications. However, self-excited vibration/chatter, generated from an insufficient-stiffness manufacturing system, could degrade the resultant surface quality and process productivity during micro-milling. In this study, a physically guided data-driven framework was developed for chatter-sensitive feature construction and machining-state identification in micro-milling of GDP polymers. Tri-directional milling-force and acceleration signals were in-situ acquired during micro-milling with a single-edge diamond tool. Wavelet packet decomposition (WPD) was initially employed for noise suppression and signal reconstruction, with optimal parameters determined using an energy-weighted feature evaluation strategy. Variational mode decomposition (VMD) was then applied to isolate state-sensitive intrinsic mode functions (IMFs), and a state-discriminative energy–entropy criterion was developed to determine task-oriented decomposition parameters. Based on the selected chatter-sensitive modes, eight multi-domain features were extracted and organized into two heterogeneous feature tensors. A dual-branch Inception Residual network (DIR-Net) was constructed to independently encode force- and acceleration-related features before high-level feature fusion. Experimental results showed that WPD improved the signal-to-noise ratio (SNR) of the milling-force and acceleration signals by 65.8 % and 43.0 %, respectively. DIR-Net achieved an identification accuracy of 97.08 % and a macro-averaged F1-score of 97.02 % on the independent test set, outperforming conventional CNN and ResNet benchmarks. Feature visualization and online monitoring results further indicated improved separability among air-cut, stable cutting, and chatter states, as well as stable recognition performance under the unseen tool-overhang and tool-size conditions. The proposed research framework provides an interpretable solution for chatter identification and online monitoring in ultra-precision micro-milling of GDP.