Modeling the thermo-mechanical deformations of machine tool structures in CFRP material adopting data-driven prediction schemes

The thermo-mechanical effects in machine tools (MTs) are represented by complex models since they may produce non-linear distortions overtime, impacting significantly on the machining accuracy. This paper aims to model the deformation of CFRP (Carbon-Fiber-Reinforced-Polymers) structures using data-driven schemes to predict and compensate the structural thermo-mechanical behavior. A novel study is presented to investigate the thermally-induced distortions of CFPR structural materials, selecting and positioning sensors, simulating and validating models to compensate the error in real-time. Anisotropic materials are becoming an effective solution to reduce structure mass and increase damping of a MT, nevertheless their physical complexity and the different thermal-coefficients at the interface with conventional materials may generate undesired effects, limiting the obtained advantages. The proposed strategy is based on the evaluation of a set of data-driven models simultaneously, identifying the most suitable solution and comparing finite element simulations with machine learning approach. The study is developed on a vertical axis frame made of CFRP material. The experimental validation is executed on a commercial 5-axis machine tool by varying the temperature conditions and evaluating the structural thermo-mechanical deformation effect on the Tool-Tip-Point (TTP) displacement. The thermomechanical behavior is measured by fiber Bragg grating (FBG) sensing technology embedded in the CFRP structure. Data-driven lab tests are evaluated in operational conditions during 36 h, considering: i) trainingdeployment periods (875 min interval), ii) typical machining stresses and iii) environmental perturbations. The final selected data-driven model is able to reduce the detected error lower than 10 mu m range. In particular, the achieved results indicate a congruence between the TTP displacement measured and predicted with a residual error lower than 7.0 mu m (Y-direction) using the ANN-multilayer perceptron algorithm.