Doctor of Philosophy (PhD)
The Department of Civil & Environmental Engineering
Structural health monitoring (SHM) has been widely used for structural damage diagnosis and prognosis of a wide range of civil, mechanical, and aerospace structures. SHM methods are generally divided into two categories: (1) model-based methods; (2) data-driven methods. Compared with data-driven SHM, model-based methods provide an updated physics-based numerical model that can be used for damage prognosis when long-term data is available. However, the performance of model-based methods is susceptible to modeling error in establishing the numerical model, which is usually unavoidable due to model simplification and omission. The major challenge of data-driven SHM methods lies in data insufficiency, e.g., lack of data covering as many as possible damage states, especially for large-scale structures. Hence, multi-site damage identification using data-driven methods can be more challenging as pattern recognition theoretically requires sufficient data from each damage scenario.
The main objectives of this dissertation are to: (1) integrate model-based and data-driven SHM methods so that their shortcomings can be weakened while their respective merits can be preserved when implementing damage identification; (2) improve the accuracy of data-driven methods for multi-site damage identification with limited measured data.
To achieve the first research objective, physics-guided machine learning (PGML) is proposed to improve the performance of pattern recognition in data-driven SHM with insufficient measured data. The results of model-based SHM (i.e., FE model updating) are taken as an implicit representation of physics underlying the monitored structure, which is incorporated into the learning process of a neural network model with the physics guidance introduced into the loss function. In addition to PGML, transfer learning (TL) is used to bridge the gap between the numerical and experimental domains of SHM. The distribution difference and manifold discrepancy between the two domains is minimized through TL as a means of domain adaptation.
To improve the performance of multi-site damage identification in data-driven SHM, multi-label classification (MLC) and constrained independent component analysis(cICA) methods are applied to investigate the correlations between damage cases sharing common damaged sites. Finally, as a case study, a two-step strategy of identifying structural damage of offshore wind turbines via FE model updating is proposed.
Zhang, Zhiming, "Data-Driven and Model-Based Methods with Physics-Guided Machine Learning for Damage Identification" (2020). LSU Doctoral Dissertations. 5312.
Available for download on Friday, June 25, 2021