| With the advent of ubiquitous sensing, robust communication and advanced computation, data-driven modeling is increasingly becoming popular for many engineering problems. Eliminating di?culties of physics-based modeling, avoiding simplifying assumptions and ad hoc empirical models are signi?cant among many advantages of data-driven approaches, especially for large-scale complex systems. While classical statistics and signal processing algorithms have been widely used by the engineering community, advanced machine learning techniques have not been su?ciently explored in this regard. This study summarizes various categories of machine learning tools that have been applied or may be a candidate for addressing engineering problems. While there are increasing number of machine learning algorithms, the main steps involved in applying such techniques to the problems consist in: data collection and pre-processing, feature extraction, model training and inference for decision-making. To support decision-making processes in many applications, hierarchical feature extraction is key. Among various feature extraction principles, recent studies emphasize hierarchical approaches of extracting salient features that is carried out at multiple abstraction levels from data. In this context, the focus of the dissertation is towards developing hierarchical feature extraction algorithms within the framework of machine learning in order to solve challenging cyber-physical problems in various domains such as electromechanical systems and agricultural systems. Furthermore, the feature extraction techniques are described using the spatial, temporal and spatiotemporal data types collected from the systems. The wide applicability of such features in solving some selected real-life domain problems are demonstrated throughout this study. |
