Loading...

A Comprehensive Machine Learning-Based Approach to Seismic Loss and Repair Time Assessment of Special Steel Moment-Resisting Frame Structures

Fahimi, Alireza | 2024

2 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56967 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Rahimzadeh Rofooei, Fayaz
  7. Abstract:
  8. Assessing the seismic loss incurred by structures post-earthquake, or to evaluate structural performance vis-à-vis potential seismic events, constitutes an imperative and pivotal measure for informed decision-making geared towards enhancing community resilience. However, large-scale assessments targeting risk analysis at citywide levels necessitate significant temporal and computational resources. To expedite this process and mitigate associated costs, this study introduces a machine learning-based methodology for estimating seismic loss and repair time of special steel moment-resisting frames. Six distinct machine learning models were trained based on results derived from nonlinear time history analyses conducted on 621 structural models spanning 1 to 19 floors, encompassing diverse geometrical and loading configurations. These models were trained using a dataset comprising 240 earthquake records. In pursuit of a comprehensive and precise estimation of seismic loss, investigations were conducted on 28 different occupancy types and damage states of structural components, drift-sensitive nonstructural components, acceleration-sensitive nonstructural components, and damage states based on residual drift, considering the responses of inter-story drift ratios, absolute accelerations of stories, and residual drifts. To the best of the author’s knowledge, such a holistic approach has not been previously undertaken in the realm of machine learning applications within this field. By selecting four structural features and examining 13 earthquake record features as input variables, an optimal predictive combination of damage states was identified, and the contribution of each feature towards predicting structural states were determined. Structural characteristics, namely the first three natural periods of buildings and the number of stories, along with the root mean square of acceleration and the geometric average of spectral acceleration as earthquake record features, were identified as defining parameters. Additionally, the selection of data points specific to each machine learning training trend and hyperparameter tuning of the algorithms were vital contributors to optimal model performance. Addressing the challenge of dataset imbalance, seven algorithms for generating synthetic samples were comprehensively evaluated, with the SMOTE-Tomek algorithm selected for training the model related to the damage states of structural components and SVM-SMOTE for the remaining three damage states. In addition to employing cross-validation procedures to ascertain model performance metrics, the selected models were validated against new earthquake records outside the original dataset, comparing predictions with damage state categories resulting from additional nonlinear time history analyses. Ultimately, the Extremely Randomized Trees model emerged as the recommended choice, achieving accuracies of 97%, 96%, 93%, and 94% in predicting damage states for structural components, drift-sensitive nonstructural components, acceleration-sensitive nonstructural components, and damage states based on residual drift, respectively. To facilitate the practical application of the proposed approach, a software interface was developed. This research constitutes a positive stride in the realm of resilience engineering, significantly expediting seismic evaluations
  9. Keywords:
  10. Machine Learning ; Structural Damage Detection (SDD) ; Time History Analysis ; Seismic Damage ; Steel Special Moment Frame ; Seismic Repair Time ; Loss Estimation ; Non-Structural Damage

 Digital Object List

 Bookmark

No TOC