Carbon-fibre-reinforced polymer (CFRP) robotic grinding is strongly anisotropic and stage-dependent, hindering reliable surface-roughness prediction. This study proposes a physically constrained multimodal framework for predicting Sa by fusing Z-axis vibration and post-grinding surface texture. Experiments at nine grinding angles were segmented into Entry, Stable, Exit and Whole stages. Grey Relational Analysis was used to select informative vibration and texture features, and a fifth-order polynomial augmentation strategy expanded the dataset from 36 to 148 samples. An early-fusion artificial neural network was trained and compared with vibration-only and texture-only models. On an independent test set, the fusion model achieved an MAE of 0.111 μm, an RMSE of 0.131 μm and an R² of 0.898, reducing MAE by 56.6% and 45.6%, respectively. Within ±0.30 μm tolerance, it achieved a 100% pass rate, demonstrating robust and engineering-reliable Sa prediction for CFRP robotic grinding.
Shan et al. (Tue,) studied this question.