A novel end-to-end deep learning model for predicting the full stress field of tensioned membrane structures

To overcome the constraints of conventional finite element analysis, particularly its inefficiency in handling real-time feedback due to repetitive modeling and calculations. A deep learning model consisting of two essential modules was presented. The stress encoding-decoding module employs convolut...

Full description

Saved in:
Bibliographic Details
Published in:KSCE Journal of Civil Engineering Vol. 29; no. 4; pp. 100073 - 12
Main Authors: Xu, Junhao, Sheng, LingYu, Zhang, Yingying, Fei, Shuhuan, Zhao, Ziang
Format: Journal Article
Language:English
Published: 대한토목학회 01.04.2025
Subjects:
ISSN:1226-7988, 1976-3808
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:To overcome the constraints of conventional finite element analysis, particularly its inefficiency in handling real-time feedback due to repetitive modeling and calculations. A deep learning model consisting of two essential modules was presented. The stress encoding-decoding module employs convolutional autoencoders (CAE) to compress high-dimensional stress tensors into a set of scalar data known as stress codes. The nonlinear mapping module utilizes a deep neural network (DNN) to map the boundary conditions to the stress code, thereby facilitating the prediction of the full stress field in the tensioned membrane structure. The results show that the maximum structure similarity index measure (SSIM) between original stress images and reconstructed stress images is 0.905 and the minimum is 0.879. The determination coefficient of the nonlinear mapping module is as high as 0.991 after 5-fold cross-verification. The trained CAE-DNN model can generate stress tensor based on tensile load and displacement within 1 second and the mean values of normalized mean absolute error (NMAE) and normalized absolute error (NAE) for all testing data were 9.58 % and 10.43 % respectively. The novel end-to-end deep learning model is an efficient and accurate prestress analysis tool for the field of membrane structure, solving the problem of real-time calculation and repeated calculation of structure, and providing technical innovation for perceiving the overall mechanical state of membrane structure in real-time digital twin model. KCI Citation Count: 0
Bibliography:https://doi.org/10.1016/j.kscej.2024.100073
ISSN:1226-7988
1976-3808
DOI:10.1016/j.kscej.2024.100073