Loading...

Predicting Structural Response of Steel Building under Ground Motion Excitation using Deep Learning Networks

Karami Seyedabadi, Reza | 2023

106 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56547 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Mohtasham Dolatshahi, Kiarash; Yazdanpanah, Omid
  7. Abstract:
  8. This paper aims at producing surrogate models which can predict building structural response under ground motion loads. Rapid response prediction has a great influence on post-event decision-making. The current study follows mentioned purpose in two main sections. The first section proposed deep models, able at estimating displacement time-series response by using only ground motion and roof acceleration. By this point, different preprocessing methods and their effects are studied. Also, a novel loss function is introduced and a hybrid model consists of different deep layers utilized to gain accurate models. These models train and evaluate on two case-study buildings; a special moment frame and an eccentrically braced frame. Based on mean absolute error, the best model of this section is able at predicting displacement time series with mean error of 6.25 and 3.5 percent, respectively for moment frame and eccentrically braced frame case study. Finally, a graphical user interface is developed to use these networks without the need of deep learning knowledge. In the second section, attention-based deep models are developed which can be able at predicting engineering demand parameters of different structures. To this end, a huge database consist of incremental dynamic analysis responses of 25 moment frame building under 50 ground motion records, is collected. Data augmentation and kmean clustering are used to preprocess datasets. Based on importing different kinds of inputs to networks, three scenarios are created. The first one aims at using achievable input in practice to be a practical scenario. The second scenario aims at predicting non-linear dynamic response by having non-linear static analysis results and the last scenario utilized all inputs of mentioned scenarios. A hierarchical hyperparameter tuning method is used to decrease computational costs. Moreover, the necessity and influence of inputs are discussed. Results show that the practical scenario outperforms others and determines building damage state based on FEMA advised by 78 and 74 percent accuracy, respectively for drift-sensitive structural damage and non-structural acceleration-sensitive damage
  9. Keywords:
  10. Peak Floor Acceleration ; Residual Drift ; Eccentrically Braced Frame (EBF) ; Graphical User Interface Test ; Time History Analysis ; Steel Structures ; Deep Attention Layer ; Maximum Story Drift Ratio

 Digital Object List

 Bookmark

No TOC