Machine Learning-based Data Analysis and Modelling of the Seismic Response of Nearby Tunnels and Bridges to Seismic Events

In the last decades, it has become increasingly important to monitor structures and infrastructures to determine their behaviour during normal service life and during particular stressful events such as earthquakes. In this way, it is possible to define the state of health of the building, whether it is healthy or damaged, the level of damage, and its eventual evolution. At the same time, it is possible to learn more about the behaviour of the building, especially in dynamic conditions. This framework is called Structural Health Monitoring and is mainly aimed at supervising historical buildings or crucial ones. Nowadays, Machine Learning has become one of the most popular topics and it is adopted in a large number of different case studies. Also, in Structural Health Monitoring the usage of Machine Learning is starting to be implemented. Indeed, with a Machine Learning model, it is possible to obtain different types of outcomes that are useful for the user, such as defining some hidden patterns and features into the dataset given as input for the model, defining the features of the data and then label them, or it is possible to create new data that have the characteristics in terms of structure and pattern compliant with the original data. Specifically, the last entry in the list is of primary relevance because one of the most important drawbacks when Structural Health Monitoring and Machine Learning come together is the data availability. Often, the data available are limited, but the model needs a large amount of data to obtain an outcome that has good precision, so to obtain acceptable precision is necessary to perform the data augmentation. This work is based on the records coming from the monitoring system applied on the Bay Bridge, the Transbay Tube and the Caldecott Tunnel are important infrastructures located in San Francisco, California. This data coming from the monitoring system are the recorded accelerations that the buildings exhibit when hit by near-fault earthquakes: Berkeley Earthquake, Piedmont Earthquake and South Napa Earthquake. This document faced two main topics: the first is related to a time history analysis of a reduced portion of the recorded data, and the second and most important argument is the data augmentation adopting a Machine Learning technique. Firstly, considering Bay Bridge and Transbay Tube a first graphical analysis of the records has been performed focusing on the time domain, the frequency domain and the Arias intensity on a properly selected number of sensors. Secondly, it is adopted a particular type of Machine Learning technique, belonging to the subbranch of Deep Learning, called Conditional Generative Adversarial Network. The model receives as input a labelled dataset that corresponds with acceleration recorded from the monitoring system. The task of the model is to generate new data that are compliant with the features of each label class. With this method, if it works, it is possible to overcome the restriction of the limited dataset.