Sensor-Derived Mechanism-Informed Prediction of Section-Level Residual Profile Error in Robotic Blade-Edge Finishing

Robotic belt finishing of turbine-blade edges is difficult to control because local edge radius, contact compliance, and the incoming profile state jointly affect the final residual profile error. This study develops a sensor-derived, mechanism-informed framework for predicting section-level root-mean-square (RMS) residual profile error. Online force measurements, robot and process records, CAD-derived edge geometry, and coordinate measuring machine (CMM) profiles are converted into interpretable section-level descriptors. Three coupled descriptors are introduced to represent the load-to-radius ratio, the force-radius-mismatch interaction, and the normalized radius mismatch. Four Gaussian process regression (GPR) configurations, a training-mean predictor, and a ridge-regression baseline are evaluated using a grouped leave-one-blade-out protocol on eight blades and 80 measured sections. The proposed descriptors show clear predictive value under blade-wise evaluation. Ridge-B3 achieves the best deterministic accuracy, with RMSE = 1.0285 µm and R² = 0.7759. The predefined GPR-B3 model does not provide the lowest point-prediction error, but it provides predictive intervals and descriptor-attribution information. These results indicate that descriptor construction is the primary source of deterministic accuracy, whereas GPR serves as an uncertainty-aware modeling layer for risk-aware blade-edge quality assessment.