Accuracy Improvement in Robotic Milling Through Data-Driven Modelling and Control

Six degree of freedom (6-dof) articulated arm industrial robots are promising candidates for aerospace machining operations such as milling due to their low-cost and large workspace compared to Computer Numerical Control (CNC) machine tools. However, the instantaneous position accuracy of industrial robots during milling is dependent on the vibratory behavior of the end effector tool tip. Consequently, it is important to model and predict the robot’s tool tip vibration as the arm configuration changes over the workspace. This dissertation addresses the modeling, prediction, and control of instantaneous tool tip vibrations of a 6-dof industrial robot over its workspace using data-driven methods. First, a data-driven modeling approach utilizing Gaussian Process Regression (GPR) of data acquired from modal impact hammer experiments to predict the modal parameters of a 6-dof industrial robot as a function of its arm configuration is presented. The GPR model is found to be capable of predicting the robot’s dominant natural frequency of vibration, stiffness, and damping coefficient in its workspace with root mean squared errors of 3.31 Hz, 150 KN/m, and 810 Ns/m, respectively. The predicted modal parameters are used to predict the average peak-to-valley vibrations of the tool tip during robotic milling. The results show that the average peak-to-valley vibrations predicted by the model follow the experimental trends with a maximum error of 0.028 mm. The prediction errors are attributed to the fact that the model only predicts the modal parameters corresponding to the dominant mode of vibration instead of the entire Frequency Response Function (FRF) of the robot. The GPR model is also used to create a Linear Quadratic Regulator (LQR) based pose-dependent optimal controller to suppress tool tip vibrations of a 6-dof industrial robot during milling. Robotic milling experiments show that the LQR controller reduces tool tip vibration amplitudes by an average of 47%. However, offset mass experiments show that the optimal controller has a bandwidth limitation of 24 Hz due to an intrinsic delay in the robot controller response to control commands. Finally, a hybrid statistical modelling approach that augments the GPR model of the robot’s pose-dependent FRF derived from experimental modal analysis, i.e. impact hammer tests, with the robot’s FRF derived from operational modal analysis, which utilizes milling forces and tool tip vibrations to compute the FRF, is presented. The hybrid model augmentation approach is demonstrated to be an efficient method to improve the prediction accuracy of the robot’s FRF with minimal optimization iterations. Specifically, the hybrid model is shown to reduce the root mean squared errors in predicting the FRF by 34% and the number of optimization iterations by 50%.Ph.D