條件生成對抗網路於鋼筋混凝土柱遲滯迴圈預測之開發與應用;Development and Application of Conditional Generative Adversarial Networks for Predicting Hysteretic Loops of Reinforced Concrete Columns

NCUIR > Engineering College > Graduate Institute of Civil Engineering > Electronic Thesis & Dissertation > Item 987654321/94457

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/94457

Title:	條件生成對抗網路於鋼筋混凝土柱遲滯迴圈預測之開發與應用;Development and Application of Conditional Generative Adversarial Networks for Predicting Hysteretic Loops of Reinforced Concrete Columns
Authors:	陳漢興;Chen, Han-Xing
Contributors:	土木工程學系
Keywords:	鋼筋混凝土柱;遲滯迴圈;物理條件;條件生成對抗網路;Reinforced concrete columns;Hysteretic loops;Physical conditions;Conditional Generative Adversarial Networks
Date:	2024-07-23
Issue Date:	2024-10-09 14:44:33 (UTC+8)
Publisher:	國立中央大學
Abstract:	臺灣位於環太平洋地震帶上，受到板塊擠壓作用的影響導致地震頻繁，常有致災型地震發生。鋼筋混凝土(Reinforced Concrete，RC)建築是台灣主要的建築結構材料，其受震反應是一重要的研究課題，尤其是在反覆載重下RC柱能否有效地發揮其韌性，是房屋耐震能力的重要指標。為了解RC柱的遲滯行為，一般多採用反覆載重分析。然而反覆載重若採實驗所需的成本高、採用數值分析則多有數值不穩定的問題需要反覆的進行參數校正。為了更有效地了解RC柱的遲滯反應，本研究提出一種基於深度學習技術的條件生成對抗網路（Conditional Generative Adversarial Networks，CGAN）來預測不同設計條件下的RC柱遲滯迴圈。本研究共蒐集257組RC柱反覆載重試驗之數據，並分別訓練CGAN、CDCGAN、CWGAN-GP三種不同的條件生成對抗網路模型對不同破壞型式控制的試體進行預測，用以探討最適用之模型。研究結果顯示使用CWGAN-GP有最好的IoU（Intersection Over Union）。此外，本研究探討2種數值型標籤之物理條件所生成之遲滯迴圈，結果顯示採用10種設計RC柱設計參數較僅採用1種剪力強度比做為條件所生成的遲滯迴圈有較高的IoU及迴圈力學表現，因此本研究根據此結果，期望在未來能夠提供有需求的民眾或結構工程師一種快速且低成本的分析方法，使其無須進行反覆載重試驗，便能獲得準確遲滯迴圈。;Taiwan is located on the Pacific Ring of Fire, which results in frequent earthquakes and sometimes loss of life. Reinforced Concrete (RC) building is the most widely used building type in Taiwan. The behavior of RC columns under seismic loads is hence critical, especially their seismic capacity and energy dissipation performance. Cyclic loading tests are commonly used to understand the hysteretic behavior of RC columns. However, experimental test is costly, and numerical analysis, on the other hand, is easy to suffer from numerical instability issues and hence requires repeatedly calibration of parameters. To address it, this study proposes a deep learning technique based on the conditional Generative Adversarial Networks (CGAN) to predict the hysteretic behavior of RC columns. This study collects experimental data from 257 cyclic loading tests on RC columns. Three different conditional generative adversarial network models—CGAN, CDCGAN, and CWGAN-GP—are investigated to identify the optimal model for predicting the hysteretic loops with various failure modes. The result shows that CWGAN-GP has the highest IoU (Intersection Over Union) and hence is recommended. On the other hand, Different conditions are also assessed to optimize the prediction performance, and the result reveals that the general RC design parameters are more suitable than a single shear strength ratio. Based on these results, this study hopes to provide a rapid and low-cost analysis method for the public and structural engineers in the future. This method would allow them to obtain accurate hysteresis loop analysis without the need for repetitive load testing.
Appears in Collections:	[Graduate Institute of Civil Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	60	View/Open

社群 sharing

Loading...