Physics-constrained dual-path acoustic-emission monitoring of material removal and surface roughness for robotic CFRP belt grinding

Abstract

In-process monitoring of robotic CFRP belt grinding is hampered by low-contrast acoustic-emission (AE) spectrograms and signal-distribution drift under belt wear, with fewer than 100 labelled samples per campaign. This paper presents a physics-constrained dual-path framework that jointly predicts material-removal depth and surface roughness (Ra, Rq, Rz) from AE spectrograms, combining gradient-based spectrogram enhancement, an embedded Preston-equation prior anchoring removal prediction, a ratio-coupled roughness head, and a label-free annealed Markov-chain partition. A factorial ablation quantifies a division of labour: embedded physics governs cross-batch removal accuracy (main effect ＋0.479 R 2 ), while gradient-enhanced AE drives roughness (＋0.116). On the 15-sample cross-batch partition R 2 is 0.956 and 0.963 with bootstrap intervals of ± 0.04–0.06; across nineteen held-out folds removal accuracy is retained (0.918 ± 0.021) while roughness accuracy holds only within the calibrated parameter range.