Simplified feed forward neural network approach for seismic performance prediction
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
A framework is proposed as a tool to support resilience-based seismic design by providing an approach for quick, post-earthquake response prediction using neural networks, thereby providing a better understanding of the relationship between seismic design demands of structural and nonstructural components. For most structures, inter-story drift (IDR) and peak floor acceleration (PFA) are important parameters for structural resiliency and recovery as they gauge the seismic performance of structural and nonstructural components. Solving for IDR and PFA using time history analysis (THA) is costly in terms of time, skills, and computational resources. Past research has presented either simplified classic methods (i.e. shear and flexure beam) model, which can be cumbersome for complex three-dimensional structures, or time-series data-based methods that mimic THA and require large datasets for training or prediction. This study utilizes a data-based approach to provide quick predictions, for a building archetype, using a limited dataset of ground motion parameters. Feed-forward neural networks (NN) are trained, in a regression fashion, to predict nonlinear IDR and PFA for a five-story, steel moment-resisting frame, office building. The model performance varied with a coefficient of determination (R^2) from 0.90 on the lower floors to 0.70 on the upper floors. Similarly, the potential relationship between IDR and PFA was studied to bridge the gap between structural and nonstructural components' performance. IDR proved to be a good PFA predictor, with R^2 between 0.60 and 0.97. ☐ In addition to predicting PFA to identify structural vulnerabilities to inform post-recovery responses and retrofit, the study is extended to include the use of low-cost vibration sensing units for data collection and broader distribution of sensor networks for structural monitoring and evaluation. Laboratory testing was conducted to evaluate assembled low-cost vibration sensing units using open-source solutions to establish a community-based seismic sensing network that can fill the gaps in demand for data necessary to train machine learning (ML) models for resiliency and recovery. A parametric study is conducted to assess the gap between models and field measurements that results in PFA amplification. Diaphragm flexibility and torsional irregularity are varied, a parametric study, resulting in 242 unique structural layout combinations. The structures are subjected to a suite of eleven ground motions and the results of THA are presented. The applicability of the ASCE7-22 formula for PFA demand is examined against the study outcomes.
Description
Keywords
Earthquake, Low-cost sensors, Neural networks, Peak floor acceleration, Structural dynamics, Torsional effects