Conditional generative adversarial networks for the data generation and seismic analysis of above and underground infrastructures

Abstract

Estimating the resilience of civil infrastructures is crucial for disaster prevention (i.e. earthquakes), encompassing both above- and underground constructions. However, while below-ground infrastructures are generally acknowledged as less vulnerable than their over-ground counterparts, this aspect has not yet garnered widespread attention. Thus, noting the limited number of seismic response comparisons for underground structures and the virtual absence of comparative analysis between above- and below-ground infrastructures in the scientific literature, this work aims to address this research gap. Nevertheless, data scarcity strongly hampers this endeavour. Not only do very few tunnels have permanent dynamic monitoring systems installed, but even fewer recorded major earthquakes are in proximity to similarly instrumented bridges and viaducts. This study focuses on three infrastructures of the San Francisco Bay Area: the Bay Bridge, the Caldecott Tunnel and the Transbay Tube. The chosen infrastructures represent a unique combination of nearby, continuously monitored case studies in a seismic zone. Yet, even for these selected infrastructures, few comparable data are available – e.g., only one earthquake was recorded for all three. Hence, a Conditional Generative Adversarial Network (CGAN) technique is put forward as a strategy to build a hybrid dataset, thereby incrementing the available data and overcoming the data scarcity issue. The CGAN can generate new data that resemble the real ones while simultaneously comparing different datasets via binary classification. With this dual objective in mind, the CGAN algorithm has been applied to various cases, varying the input given in terms of selected acquisition channels, infrastructure pairs, and selected strong motions. In conclusion, each pair underwent a postprocessing phase to analyse the results. This research’s outcomes show that the classifications performed with the Support Vector Machine reached excellent results, with an average of 91.6% accuracy, 93.1% precision, 93.3% recall, and 92.9% F1 score. The comparison in the time and frequency domain confirms the resemblance.