Curvature Prediction Of Multi-pass Roll Bending For Aerospace Thin-walled Parts

Aerospace thin-walled parts are the key components of the launch vehicle storage tank,accounting for more than 60% of the total coverage area of the launch vehicle,and have the characteristics of multi-variety and small-batch production.Three-roll symmetrical roll forming is a key process link in the production process of aerospace thinwalled parts,and its forming curvature directly affects the subsequent welding quality.Predicting the forming curvature of thin-walled parts in advance can quickly find the phenomenon that the curvature of thin-walled parts does not meet the standard in the production process,effectively avoid the time loss caused by subsequent calibration,and is of great significance to improve the efficiency and accuracy of roll forming.However,the rolling process of aerospace thin-walled parts is affected by the mutual coupling of complex factors such as process parameters and size parameters,and there is a strong temporal correlation between the previous and the following passes,which makes it difficult to explore the direct relationship between various factors and the forming curvature.In addition,due to the difference in size and material of different varieties of thin-walled parts,the roll bending performance varies greatly,so it is difficult to establish a curvature prediction model for aerospace thin-walled parts across varieties.Therefore,this paper first studies the curvature prediction model of single-variety aerospace thinwalled parts to solve the multi-factor coupling and strong timing problems in the process of rolling and bending;A transfer learning problem with batch characteristics.The main work is as follows:1.Aiming at the characteristics of coupling of multiple influencing factors and strong timing in the rolling process of aerospace thin-walled parts,a single-variety aerospace thin-walled part based on graph attention network and long short-term memory network(GAT-LSTM)was proposed.The method for predicting the curvature of pass roll bending includes a graph reconstruction module,a feature extraction module and a curvature prediction module.The graph reconstruction module transforms the thin-walled rolling forming data samples from the traditional Euclidean space to the non-European space graph structure.The nodes on the graph represent the influencing factors of rolling forming;the feature extraction module designs a multi-head attention mechanism to mine the relationship between the influencing factors.By synthesizing the feature extraction results of multiple attention heads,the global characteristics of the rolling forming process of thin-walled parts are obtained;The curvature prediction module uses the long and short-term memory network to learn the time sequence between the passes before and after the roll-bending process of aerospace thin-walled parts,and achieve accurate prediction of the roll-bending curvature.In terms of experiments,simulation data and field data are used to verify the effectiveness of the proposed method.2.Aiming at the characteristics of aerospace thin-walled parts with multiple varieties and small batches,as well as the differences in size,material and rolling performance between different varieties of thin-walled parts,a multi-variety aerospace thin-walled part rolling based on domain confrontation transfer learning is proposed.Forming Curvature Adaptive Prediction Method(IDANN).First,the GAT-LSTM network is used to replace the feature extractor and label predictor of the original domain adversarial network,and a multi-layer single-channel one-dimensional convolution is designed to mine the key features of the material curve of a variety of thin-walled parts,increase the data dimension,and combine the multi-kernel maximum mean difference the algorithm(Multi Kernal Maximum Mean Discrepancy,MK-MMD)further reduces the distribution differences between domains,and realizes the efficient mining of cross-domain invariant features by the feature extractor;Secondly,two independent double-layer convolutional neural networks are designed to replace the fully connected layer of the domain classifier to improve the learning ability of the domain classifier for the source domain and target domain features;When the model is updated,the gradient reversal layer is used to implement the confrontation training between the domain classifier and the feature extractor,and the label predictor is used to promote the feature extractor to learn more general cross-domain invariant features,and finally realize the rolling and bending of various aerospace thin-walled parts Adaptive prediction of forming curvature;In terms of experiments,simulation data and field data are used to verify the effectiveness of the proposed method.3.Taking the sheet metal workshop of an aerospace equipment manufacturing enterprise in Shanghai as the case background,a prototype system for predicting the curvature of aerospace thin-walled parts by multi-pass roll bending is designed and developed,and the application verification of the research work is carried out,so as to provide efficient and high-precision roll forming for aerospace thin-walled parts Provide reliable tools.