4:30 PM - 4:45 PM
[SCG49-11] Seismic demand prediction using low-dimensional features of ground motions extracted by a hierarchical autoencoder
★Invited Papers
Keywords:Seismic risk assessment, dimensionality reduction, Seismic demand, Gaussian process regression, Bayesian optimization
Seismic risk assessment is crucial for implementing effective resilience measures against earthquake disasters, ensuring human safety, and minimizing economic losses and downtime. Typically, it utilizes fragility functions, which model the relationship between ground motion intensity and the probability of structural damage. Seismic demand represents the structural response (e.g., maximum interstory drift ratio and peak floor acceleration) to an earthquake and serves as an indicator for assessing damage levels. Thus, constructing fragility functions requires modeling the relationship between ground motion intensity and seismic demand.
Our research group is developing a simulation-based seismic risk assessment method instead of relying on fragility functions. Generating seismic demand for a large set of expected ground motions through structural response analysis incurs significant computational costs. To address this, we propose a method that generates a small dataset using simulations and expands it using machine learning. This paper presents a predictive model for seismic demand corresponding to a set of assumed ground motions.
As input ground motions for this study, we generated broadband (0.1-10 Hz) ground-motion time series for a total of 1,515 earthquake scenarios for an Mw6.9 event in a heterogeneous three-dimensional medium using a hybrid approach combining 3D finite-difference with a semi-empirical method.
To represent ground motion information as input, we developed a method for extracting low-dimensional features from ground motion waveforms. First, we constructed an autoencoder using a convolutional neural network (CNN) to generate low-dimensional features. To verify whether the extracted features preserved the original waveform characteristics, we evaluated reconstruction performance by decoding the features back into waveforms. Differences were observed in both low- and high-frequency bands in the frequency domain. To improve this, we developed a hierarchical autoencoder using a Laplacian pyramid, which enhanced waveform reconstruction accuracy.
Next, we developed a seismic demand prediction model using Gaussian process regression (GPR) combined with Bayesian optimization. The model is initially trained on a randomly selected dataset and refined through active learning with Bayesian optimization.
We applied this method to predict seismic demand for a five-story steel frame building, focusing on maximum interstory drift and peak floor acceleration at each story. Seismic response analyses were conducted using 100 ground motion records to generate training data, and the predictive model was trained accordingly. The trained model was then used to predict seismic demand for 1,515 ground motions.
First, we examined the impact of acquisition function choice and hyperparameter optimization on prediction performance. The results showed that using the Upper Confidence Bound (UCB) acquisition function led to poor prediction accuracy, whereas using Expected Improvement (EI) yielded favorable results. Furthermore, hyperparameter optimization improved model accuracy by enhancing its fit to actual values.
To validate the effectiveness of low-dimensional ground motion features, we compared the proposed model with one using three conventional intensity measures (PGA, PGV, and Sa(T)) as explanatory variables. The results showed that the conventional model had lower accuracy for the fourth and fifth stories, whereas the proposed model maintained high accuracy across all stories. Moreover, the proposed method outperformed the conventional approach in predicting both maximum interstory drift and peak floor acceleration at all stories.
Our research group is developing a simulation-based seismic risk assessment method instead of relying on fragility functions. Generating seismic demand for a large set of expected ground motions through structural response analysis incurs significant computational costs. To address this, we propose a method that generates a small dataset using simulations and expands it using machine learning. This paper presents a predictive model for seismic demand corresponding to a set of assumed ground motions.
As input ground motions for this study, we generated broadband (0.1-10 Hz) ground-motion time series for a total of 1,515 earthquake scenarios for an Mw6.9 event in a heterogeneous three-dimensional medium using a hybrid approach combining 3D finite-difference with a semi-empirical method.
To represent ground motion information as input, we developed a method for extracting low-dimensional features from ground motion waveforms. First, we constructed an autoencoder using a convolutional neural network (CNN) to generate low-dimensional features. To verify whether the extracted features preserved the original waveform characteristics, we evaluated reconstruction performance by decoding the features back into waveforms. Differences were observed in both low- and high-frequency bands in the frequency domain. To improve this, we developed a hierarchical autoencoder using a Laplacian pyramid, which enhanced waveform reconstruction accuracy.
Next, we developed a seismic demand prediction model using Gaussian process regression (GPR) combined with Bayesian optimization. The model is initially trained on a randomly selected dataset and refined through active learning with Bayesian optimization.
We applied this method to predict seismic demand for a five-story steel frame building, focusing on maximum interstory drift and peak floor acceleration at each story. Seismic response analyses were conducted using 100 ground motion records to generate training data, and the predictive model was trained accordingly. The trained model was then used to predict seismic demand for 1,515 ground motions.
First, we examined the impact of acquisition function choice and hyperparameter optimization on prediction performance. The results showed that using the Upper Confidence Bound (UCB) acquisition function led to poor prediction accuracy, whereas using Expected Improvement (EI) yielded favorable results. Furthermore, hyperparameter optimization improved model accuracy by enhancing its fit to actual values.
To validate the effectiveness of low-dimensional ground motion features, we compared the proposed model with one using three conventional intensity measures (PGA, PGV, and Sa(T)) as explanatory variables. The results showed that the conventional model had lower accuracy for the fourth and fifth stories, whereas the proposed model maintained high accuracy across all stories. Moreover, the proposed method outperformed the conventional approach in predicting both maximum interstory drift and peak floor acceleration at all stories.