Seismic Intensity Measure Selection Incorporating Interaction Effects for Damage Assessment Across Different Structural Sensitive Regions
Zi-Nan Wu,Zhan-Quan Li,You Dong,Xiao-Lei Han,Guoqiang Zhang,Ruiwei Feng,Kuanjun Zhu
DOI: https://doi.org/10.1016/j.istruc.2024.106917
IF: 4.1
2024-01-01
Structures
Abstract:The selection of seismic intensity measures (IMs) has been important step within the performance-based earthquake engineering. Previous studies, primarily focusing on binary relationships between individual IMs and seismic damage, quantified the impacts of multiple IMs while neglecting their interaction effects. This led to excessive IM integration, culminating in overfitting and diminishing interpretability in damage assessment. In addressing these limitations, this study conducted a comprehensive sensitivity analysis using Random Forest (RF) algorithm, a method that incorporates interaction effects among IMs. Utilizing 1100 horizontal ground motion records, nonlinear time history analyses were performed on several multi-degree-of-freedom (MDOF) systems with fundamental periods ranging from 0.25 s to 9.85 s, generating a dataset of 9900 maximum interstory drift ratios. Upon this dataset, a RF-based sensitivity analysis of 17 widely used IMs was conducted to quantify their impacts, followed by forward feature selection based on their sensitivity rankings. It led to a quantitative categorization of structural periods into acceleration-, velocity-, and displacement-sensitive regions. Across these regions, Sa,T1 was identified as a consistently critical IM, with ASI and MIV in acceleration-sensitive, VSI in velocity-sensitive, and PGV in displacement-sensitive regions. This targeted selection of critical IMs proved robust in damage assessment across MDOF systems with various hysteretic rules and storey models, demonstrating effectiveness comparable to complex models integrating a wider range of IMs. Moreover, damage assessment based on these critical IMs indicated that multiple linear regression with logarithmic transformations tended to underestimate damage, while the RF algorithm showcased superior predictive accuracy and lower uncertainty, especially for structures in displacement-sensitive regions.