Data driven roll bending prediction via dynamic sliding window and basis function decomposition

Abstract

Roll bending is suitable for the production of high-precision and complex curved products in the aerospace field. During the dynamic forming process of variable curvature profiles, the bending moment distribution on the profile exhibits complex spatio-temporal coupling effects, leading to difficulties in accurately predicting the final shape. This study proposes a data-driven method for variable curvature roll bending prediction. Firstly, by analyzing influencing factors in roll bending process, a representation method using dynamic sliding windows is proposed to achieve temporal decoupling. Subsequently, a target region representation method based on basis function decomposition is introduced to mitigate the impact of outliers. Next, a Multi-Frequency Recurrent Neural Network (MFRNN) is proposed to model the relationship between control sequences and the final formed shape, where the MFRNN module learns features of different control sequence scales, and the frequency attention module learns the correspondence between scales and basis functions. Finally, a series of experiments demonstrate that the proposed method can accurately model the control relationships of variable curvature roll bending process, thereby providing effective solutions for the precision forming field.