Training of artificial neural networks with reduced-order models for the prediction of thermal errors on machine tools

Abstract Thermal errors are among the most significant factors affecting the positioning accuracy of a machine tool. This paper demonstrates the advantages of combining reduced-order model (ROM) finite element method (FEM) simulations with data-driven error prediction algorithms, such as artificial neural networks (ANN) or characteristic diagrams. First, the thermal behavior of a five-axis milling machine (DMU 80 eVo) was modelled in ANSYS, after experimental determination of the relevant thermal boundary conditions. The FE model was then validated with measurements taken from the machine tool with movements of all five axes under different operating conditions, such as dry air-cutting, air-cutting with coolant and stand-by. The validated thermo-elastic FE model can then be used for training prediction algorithms (ANN, char. diagrams) for online thermal error compensation. The use of ROMs overcomes the long computation times of the high-dimensional simulation models by reducing them to equivalent low-dimensional models enabling much faster thermo-elastic simulations while preserving a specified result accuracy. An ANSYS ACT extension was utilized to export the FE model data in the form of an input-output system, which is required for the application of the model order reduction (MOR) techniques. This inclusion of MOR tremendously speeds up the process of generating training data for the prediction algorithms. ANN based prediction algorithms trained by the ROM simulation data are finally validated with simulated independent thermal test scenarios.