(Georgia Institute of Technology, 2022-04-28)
Gao, Zhenyu
Catalyzed by advances in data quantity and quality, effective and scalable algorithms, and high-performance computation, the data-intensive transformation is rapidly reframing the aerospace industry. The integration of data-driven methods brings many new opportunities, such as (1) streamlining the aerospace design, testing, certification, and manufacturing process, (2) driving fundamental advancements in the traditional aerospace fields, and (3) enhancing the business and operations side of the industry. However, modern aerospace datasets collected from real-world operations, simulations, and scientific observations can be massive, high-dimensional, heterogeneous, and noisy. While sometimes being beyond people's capacity to store, analyze, and archive, these large datasets almost always contain redundant, trivial, and irrelevant information. Because the design and analysis of complex aerospace systems value computational efficiency, robustness, and interpretation, an additional procedure is required to process large datasets and extract/refine a small amount of representative information for in-depth analysis. This dissertation utilizes improved representations of operations data and aircraft models for efficient, accurate, and interpretable air transportation system analysis.
Under the overall scope of representative data and models, this dissertation consists of three main parts. Part I, representative operations data, considers the problem of selecting a small subset of the operations data from a large population for more efficient yet accurate probabilistic analyses. This is tackled by a novel distributional data reduction method called Probabilistic REpresentatives Mining (PREM), which is consistent in generating small samples with the same data distribution. Part II, representative aircraft model portfolios, considers the problem of selecting a small proportion of representative aircraft models to sufficiently cover the richness and complexity of a large population when the modeling of every participant in the complex system is infeasible. This is tackled through a clustering framework which optimizes for the minimax criterion and can conduct a trade-off between multiple criteria of the selected portfolio's representativeness. Part III, representative aircraft model features, considers the problem of obtaining improved aircraft model representations for environmental impacts modeling. This is accomplished through using a combination of large-scale computer experiment and multi-level feature representation and selection. The proposed methodologies are demonstrated and tested on four selected experiments through data visualization and quantitative metrics. Overall, this dissertation aims to contribute to both the general methodologies and the solutions to specific aerospace applications.