| dc.description.abstract | In the last few decades, with the introduction of machine learning, research on the design and monitoring of civil structures and mechanical systems has evolved into a new era. One of the main challenges in developing accurate learning models is the lack of training data covering a large range of operational conditions. While extensive training data can be generated with the aid of physics-based computer simulations, it is often difficult to represent the realistic behavior of the system with simulation data. As a result, simulated training data may have inherent modeling errors and may not describe the actual conditions appropriately. Typical data-driven learning approaches adopt a black-box approach with the assumption that the simulated training and experimental testing data are from the same underlying probabilistic distributions. However, this assumption is often unrealistic for many applications due to modeling limitations.
The divergence between training and testing data is often imminent and may lead to prediction errors and compromise the safety of the system.
This dissertation proposes a series of machine learning techniques to effectively close the knowledge gap between the simulation and experimental domains. First, this work proposes domain adaptation as a solution to structural health monitoring problems where a classifier has access to labeled training (source) and unlabeled test (target) domains. The proposed domain adaptation method forms a feature space to match the latent features of both source and target domains. Second, this dissertation proposes a physics-guided learning architecture that integrates physical parameters extracted from physics-based simulation data into the intermediate layers of the neural network time to constrain the learning process during training and to improve the generalization during inference time. Additionally, this study combines domain adaptation with physics-guided learning to further improve generalization when the network is trained with simulation data and tested with experimental data.
Another important component of this dissertation focuses on surrogate modeling of structural designs and its interpretability. While an appropriately trained black-box machine learning model may be capable of predicting outcomes with satisfactory accuracy, the model's inference process and its predictions are difficult to interpret, especially for deep-learning architectures. The prediction made by the model is often obscure to the end user due to the non-transparent nature of the architecture. This work proposes a deep learning-based physics-guided surrogate modeling architecture that allows the end user to explain the design predictions through physics-based parameters and improve the interpretability of the prediction. | |
