Conditional Distribution Shift-Domain Adaptive Regression Network for Robotic Milling Force Prediction Considering Posture-Dependent Dynamic Properties

Cutting force prediction is significant for machining vibration reduction, process parameter optimization, and robot performance improvement. Because the dynamic characteristics of a robot are dependent on its poses, many prediction methods applied to machine tools cannot be successfully employed on robot. Thus, this article proposes a conditional distribution shift–domain adaptive regression network (CDS-DARN), which realizes cutting force prediction under different robot postures using acceleration and robot joint angles without conducting modal tests. First, source model parameters are transferred to the target model. Subsequently, the cutting force prediction of the target pose is realized by utilizing the conditional distribution embedding discrepancy loss between the source and target pose unlabeled data, and between the target pose labeled and unlabeled data, as well as the regression loss for the joint domain adaptation. Experimental results of robotic milling with varying process parameters under multiple poses indicate that the CDS-DARN can achieve excellent predictive performance. The root mean square errors of the selected trajectories are reduced by more than 65.0% compared with the network trained directly using the source data and by more than 36.0% compared with the network trained directly using different sets of processing parameters under the target pose.