TF-F-GAN: A GAN-based model to predict the assembly physical fields under multi-modal variables fusion on vision transformer

Assembly is the final step in ensuring the precision and performance of mechanical products. Geometric variables, process variables, and other material or physical variables during the assembly process can all impact the assembly outcome. Therefore, the key for analyzing and predicting assembly results lies in establishing the mapping relationship between various assembly variables and the results. Traditional analysis methods typically consider the evolution of a single variable in relation to the assembly results and often focus on the value at a few nodes. Essentially, this approach constructs a value-to-value nonlinear mapping model, ignoring the coupling relationships between different variables. However, with the increase in assembly precision requirements and advancements in measurement equipment, assembly analysis has evolved from value-to-value prediction to field-to-field prediction. This shift necessitates the study of the assembly physical field results for specific regions rather than focusing on a few nodes. Therefore, this paper proposes an analysis framework, TF-F-GAN (Transformer-based- Field-Generative adversarial network), which is suitable for multi-source assembly variable inputs and physical field outputs. The framework draws inspiration from multimodal fusion and text-image generation models, leveraging the Vision Transformer (VIT) network to integrate multi-source heterogeneous data from the assembly process. The physical field data is color-mapped into a cloud image format, transforming the physical field prediction into a cloud image generation problem. The CFRP bolted joint structure assembly is used as a case study in this paper. Since assembly accuracy primarily focuses on geometric deformation, the deformation field of key regions in the CFRP bolted joint is taken as the output variable. In the case study, the geometric deviations of parts and mechanical behavior during the assembly process were considered. Data augmentation methods were used to construct the dataset. After training TF-F-GAN on this dataset, transfer learning was further conducted using experimental data. The final prediction error of TF-F-GAN relative to the experimental data was less than 15 %, with a computation time of less than 7 s. This prediction framework can serve as an effective tool for predicting the physical fields of general mechanical product assembly.