Hybrid Mechanistic and Data-driven Modeling Method of Compliant Assembly Variation Prediction for Train Body

As a typical welded thin-wall assembly structure, the train body has characteristics of low stiffness, easy deformation, and significant influence by welding process, which makes it difficult to control manufacturing accuracy, and affects the assembly quality and takt time. Assembly variation prediction for train body requires a comprehensive consideration of random geometric deviation, component’s global deformation and local welding deformation. Combined with compliant variation analysis theory and data - based welding deformation prediction method, a physical-data hybrid modeling method is proposed to predict the welding assembly dimensions of train body. First, the stiffness matrixes of components are extracted and reduced by using substructure method to construct the physical relation between loads and deformation. Then, a dataset of local welding shrinkage and angular distortion is obtained for typical welding form by using finite element analysis, which is used to construct the mapping relation between match gaps and local deformations with a data model by using BP neural network. Finally, a physical-data hybrid compliant variation analysis model is established for the welding structures with consideration of the effects of global gravity, local gap and welding distortion. The proposed method is validated by comparing predicted the height and profile variation with the actual measurement data. Simulation results show that the physical-data hybrid modeling method avoids the welding simulation process with random input deviations, and can greatly improve the computational efficiency of statistical deviation simulation by combining substructure method. © 2024 Chinese Mechanical Engineering Society. All rights reserved.