A series of forecasting models for seismic evaluation of dams based on ground motion meta-features

Abstract

Uncertainty quantification (UQ) due to seismic ground motions variability is an important task in riskinformed condition assessment of infrastructures. Since performing multiple dynamic analyses is computationally expensive, it is valuable to develop a series of forecasting models based on the unique ground motion characteristics. This paper discusses the application of six different machine learning techniques on forecasting the structural behavior of gravity dams. Various time-, frequency-, and intensity-dependent characteristics are extracted from ground motion signals and used in machine learning. A large set of about 2,000 real ground motions are used, each includes about 35 meta-features. The major outcome of this study is to show the applicability of meta-modeling-based UQ in seismic safety evaluation of dams. As an intermediary result, the advantages of different machine learning algorithms, as well as meta-feature selection possibility is discussed for the current dataset. This paper proposes a feasibility study to reduce the computational costs in UQ of large-scale infrastructural systems.

Neural Networks and Imbalanced Learning for Data-Driven Scientific Computing With Uncertainties

“…This is in line with Cloud analysis, where a large data set of signals are applied to the structural system [74]. For each ground motion, 31 intensity measure (IM) parameters are extracted to develop a side information matrix [45]. These 31 IM parameters are selected from a comprehensive list of over 70 IM parameters found in [14].…”

“…In addition to logistic regression and artificial NNs, other widely used classification algorithms that work directly on the input parameter space include support vector machines (SVMs) and the random forest (RF) algorithm [44], [45]. The former aims to find an optimal separating hyperplane or decision boundary that maximizes the geometric margin for producing a classification rule.…”

Kernel Matrix Approximation on Class-Imbalanced Data With an Application to Scientific Simulation

“…In order to be used in the context of training classifiers, we need to convert the stochastic nature of the ground motion signal into a manageable number of scalar features. Such a conversion is already discussed in [54], [55] for generic infrastructural systems. For each ground motion, we extract 31 intensity measure (IM) parameters, including all peak values, intensity-, frequency-, and durationdependent parameters.…”

Multistep prediction of dynamic uncertainty under limited data

“…NNs are composed of multiple layers, allowing them to learn complex nonlinear relationships. Bayesian deep learning (BDL) and variations thereof have been widely applied to forecast future events given existing data and update when presented with new data [29][30][31][32][33][34][35][36][37]. Deep learning models are only as accurate as the data they are trained on and, as such, typically require large datasets with defined trends over time [37].…”