An online milling deformation prediction method for thin-walled features with domain adversarial neural networks under small samples

Thin-walled machining features are extensively utilized in the aerospace industry, where the milling deformation caused by their weak rigidity has been the most common quality concern. Efficient control of milling deformation for thin-walled features is essential for enhancing quality. However, the high cost and time-consuming nature of data collection for aviation parts, leading to a limited availability of process data, which presents a significant challenge for predicting deformation in aerospace components. To address this issue, this study aims to develop a high-precision milling deformation prediction method by fully leveraging the small-sample data from machining experiments and simulation data. This paper first constructs a thin-walled features deformation prediction framework by integrating Domain Adversarial Neural Networks (DANN) with a digital twin process model. Secondly, the DANN method is adopted to achieve online prediction of milling deformation for thin-walled features. A small quantity of experimental deformation data serves as the target domain for training dataset, whereas milling simulation data produced by finite element software serves as the source domain. Milling deformation is accurately predicted using adversarial training based on the DANN structure for domain regression and domain classification. The best results show that the proposed method achieves better goodness of fit under limited sample conditions, with a 5 % increase in the Coefficient of Determination (R²) and a 15 % reduction in Mean Absolute Error (MAE) compared to five baseline methods. In the end, the DANN approach was integrated into the digital twin system for the milling process, and a prototype system was constructed to demonstrate the viability of the suggested approach.