Evaluation of Deep Learning Algorithms in Predicting Seismic Response of a Reinforced Concrete Structure

Abstract

This thesis presents an evaluation of the performance of three well-established deep learning algorithms in predicting the response of a six-story instrumented reinforced concrete hotel in California to seismic excitation. Given the increasing availability of strong-motion data and expanded usage of deep learning in structural health monitoring, this thesis seeks to evaluate the predictions of purely data-driven and physics-informed architectures using processed instrumentation data in order to more accurately predict structural response for use in structural health monitoring and performance-based design applications.

By employing a variety of results metrics previously used in the literature, including correlation coefficients, normalized error distributions, and peak errors, this thesis examines different components of the models’ capabilities to learn more about patterns in the data learned by the computational mechanisms of each architecture, and exploring the feasibility of a generalized approach for further application in structural response prediction.

Findings from the work show the data-driven Long Short-Term Memory (LSTM) network performing the most accurately, but not consistently outperforming the other algorithms. Some trends in the data could be evidence of how different architectures may be better equipped in predicting different mode shapes and frequency contents. For example, the data-driven and physics-guided LSTM models predicted the third floor’s response more accurately than the roof, whereas the physics-guided convolutional neural network (CNN) was the opposite, showing a contrast between the two base architectures. This thesis also contributes to this growing field by documenting the experimental setup in detail to allow for the replication of results and for the facilitation of future application by structural engineers.

As structural engineering research in deep learning continues to gain popularity, this thesis provides an experimental basis of a case study that can be followed and replicated to motivate future experimentation, as well as offering compelling different directions that future work could be directed to further the usage of deep learning in structural response prediction and structural health monitoring as a whole.