A conditional generative model for end-to-end stress field prediction of composite bolted joints
Yong Zhao,Yuming Liu,Qingyuan Lin,Wei Pan,Wencai Yu,Yu Ren,Sheng Liu
DOI: https://doi.org/10.1016/j.engappai.2024.108692
IF: 8
2024-06-02
Engineering Applications of Artificial Intelligence
Abstract:Carbon Fiber Reinforced Polymer (CFRP) laminates, prized for their lightweight and high stiffness, are extensively used in aerospace and maritime applications. Bolted joints play a crucial role in connecting these laminates. However, manufacturing variations arise during the assembly process, impacting performance due to material-related factors. Predicting the assembly stress fields of Carbon Fiber Reinforced Polymer bolted joints is of great significance in design optimization, manufacturing process control, and structural health monitoring. The currently prevalent finite element analysis methods incur extremely high computational costs, failing to meet the requirements for real-time prediction of the assembly and multiparametric design of composite bolted joints. Proposing a methodological framework for rapidly predicting the assembly physical field is necessary. This paper introduces a stress prediction framework to enhance analysis and aid material parameter design. The framework is inspired by image processing and artificial intelligence drawing by analogizing the computed physical field results to the generated images. Therefore, the Bolted Tightening Generative Adversarial Network (BT-GAN), a cascaded generative model, is proposed in this paper to predict stress fields of the composite bolted joints during assembly. The model starts with data augmentation of the stress filed results from the finite element analysis in a super-resolution network, which realizes an integral interpolation mapping from coarse-grid to fine-grid results. Then, the results of the data enhancement are fed into the subsequent conditional generative adversarial network for learning. Similar to the text-guided image generation approach, the network learns to understand the physical mapping relationships between different parameters and assembly stress fields. Moreover, the network achieves higher accuracy in stress field prediction by extraction the understanding of multi-scale features through the skip connection and the attention mechanism. This method effectively learns the physical mapping relationship between multiple parameters and the stress field, applying a graph generation approach to end-to-end predictions of the field. Compared to the results of finite element analysis from the coarse-grid, the Structure Similarity Index Measure (SSIM) of the cascaded generative network proposed in this paper has been improved from 0.584 to 0.962 and the Peak Signal-to-Noise Ratio (PSNR) metric has been increased from 17.3 dB to 58.2 dB. What's more, the mean relative error on the maximum values of the stress field has reached 6.9%. The trained model takes only 6.1s to complete a single prediction, significantly improving the prediction efficiency compared with finite element analysis. It is compared with the other networks commonly used for physical field prediction and shows improvement in the metrics proposed in the article. By constructing such an end-to-end stress field prediction framework during assembly, efficient forecasting for the assembly of composite bolted joints can be achieved. This is advantageous for the digital twin modeling of the assembly lines and the effective control of assembly quality, providing a powerful tool for assembly design and analysis.
automation & control systems,computer science, artificial intelligence,engineering, electrical & electronic, multidisciplinary