A data-driven rolling optimization method for trajectory tracking error prediction of CNC machine tools

The dynamic performance of the feed-drive system in CNC machine tools directly influences the accuracy of machined parts. To enhance the motion control performance of CNC machine tools, a high-precision model of the feed-drive system is critical. However, current modeling methods for feed-drive systems seldom consider time-varying factors such as loads, wear, and lubrication. As a result, the model accuracy degrades when the system characteristics are affected by these time-varying factors. In this paper, a rolling optimization method with partial weights frozen is developed to realize quick iterative learning of a data-driven model for a feed drive system with time-varying characteristics using a small amount of data. First, the long short-term memory fully connected (LSTM-FC) network is built and divided into feature extraction and output fitting parts based on their functions. Then, a weight freezing-based rolling optimization method is applied. The weights in the feature extraction part are frozen, which preserves the learned common knowledge and patterns by solidifying the way that high-dimensional features are extracted from the input. By adjusting the weights in the output fitting part, the extracted high-dimensional features are remapped to the new data distribution changed by time-varying factors. Finally, the performance of the developed rolling optimization method is confirmed by experiments. The results show that the proposed rolling optimization method reduces the maximum prediction errors by 49.5% and the total training time by 96.3% compared with existing methods, which demonstrates that the proposed method can restore model accuracy when the system characteristics change due to time-varying factors, and significantly accelerate the optimization process by rolling optimization.