Efficient prediction of machine tool position-dependent dynamics based on transfer learning and adaptive sampling

Abstract

Large-scale machine tools usually exhibit pronounced position-dependent dynamic characteristics. Accurate prediction of machine tool position-dependent dynamics is crucial for efficient and high-precision machining. Theoretical modeling has mostly focused on small machine tools, whereas research on the position-dependent dynamics of large-scale machine tools mainly relies on experiments. However, the high cost of these experiments presents significant challenges for studying the dynamics of large machine tools. This paper aims to address the challenge of accurately predicting machine tool position-dependent dynamics with limited experimental data. By employing progressive neural network transfer learning, we utilize machine tool dynamic theoretical models with systematic errors to generate prior expert knowledge, thus resolving the issue of training convergence with small sample data. An adaptive sampling strategy suitable for gantry machine tool position-dependent dynamic prediction is proposed, which integrates prior knowledge and information from existing sampling points during the sampling process. This approach decreases the amount of sampling data and improves the efficiency of predicting machine tool position-dependent dynamics. Using a large gantry five-axis composite machine tool with a workspace of 6.5 m × 6 m× 2 m as an example, this paper predicts its position-dependent dynamic characteristics. These include natural frequencies, damping ratios, and modal shapes. The predictions are based on a dynamic model and small sample modal experimental data, which are validated through both simulation and experimentation. Compared to full-space modal experiments, the proposed method achieves an average error of 0.26 Hz in predicting the top three position-dependent modal frequencies of the machine tool across the entire workspace with 11 sampling points. Compared to traditional methods of fitting after random sampling, the accuracy is improved by 74.51 %, and the convergence speed is improved by 45 %.