Predicting the position-dependent dynamics of machine tools using progressive network

Abstract

The structural dynamics of a machine tool exhibit position-dependent behavior under operation due to the kinematic reconfiguration of its feed drives. Experimental identification of position-dependent dynamics can be time-consuming, and simulation models may lack accuracy due to modeling uncertainties. In this paper, the prediction of a machine tool's position-dependency is approached under the progressive network, a transfer learning technique to bridge the gap between the simulation and real-world domains. The dynamics simulation model may be somewhat inaccurate, but still describes a machine tool's overall position-dependency. The real-world domain is then characterized by sparse dynamics measurements made on the real machine tool. Under the progressive network architecture, a neural network is first trained in the simulation domain to extract valuable prior knowledge about the machine's position-dependency. A second network posed in the real-world domain then accesses such prior knowledge from the simulation domain via lateral connections to improve its dynamics predictions. Such transfer learning methodology may be most advantageous when the number of experimental data is limited, and the machine is unavailable for further testing due to production purposes. The proposed methodology was validated under numerical and experimental studies.