Intelligent Data-driven Classification and Forecasting Processes for Complex Engineering and Social Systems
Complex engineering systems such as automobiles and chillers in heating, ventilation, and air-conditioning (HVAC) systems are being equipped with increasingly sophisticated electronic systems. Operational problems associated with degraded components, failed sensors, improper installation, poor maintenance, and improperly implemented controls affect the efficiency, safety, and reliability of the systems. Failure frequency increases with age and leads to loss of comfort, degraded operational efficiency, and increased wear and tear of system components. Out of the research directions of this thesis is to develop a data-driven scheme for fault diagnosis and severity estimation to HVAC systems. Most existing HVAC fault-diagnostic schemes are based on analytical models and knowledge bases. These schemes are adequate for generic systems. However, real-world systems significantly differ from the generic ones and necessitate modifications of models and/or customization of the standard knowledge bases, which can be labor intensive. To overcome such issues, we consider a data-driven approach for fault detection and diagnosis (FDD) of chillers in HVAC systems. The research on the faults of interest in the chiller could enable the building system operators to improve energy efficiency and maintain the desired comfort level at reduced cost. Another research direction of this thesis is to develop data reduction techniques for on-board implementation of data-driven classification techniques in memory-constrained electronic control units (ECUs) of automobiles. One of the problems with high-dimensional datasets (caused by multiple modes of system operation and sensor data over time) is that not all the measured variables are important for understanding the underlying phenomena of interest. While certain computationally expensive methods can construct predictive models with high accuracy from high-dimensional data, it is still of interest in many applications to reduce the dimension of the original data prior to any modeling of the data. Data-driven applications on reduced datasets could also be suitable for ECUs, which have memory capacity limitations due to cost constraints. We also develop innovative classifier fusion techniques so as to decrease classification error and reduce variability in diagnostic error. We show that fusing marginal classifiers can increase the diagnostic performance substantially. Furthermore, we could reduce the diagnostic errors by combining traditional fusion techniques (e.g., classifier selection, combining classifier outputs, sampling training data, manipulating classifier outputs, classifier feature selection, etc.) with our novel classifier fusion techniques. The data-driven framework can be beneficial not only to the engineering community in the diagnostics and prognostics of complex systems, but is also potentially useful in social science research. What if we apply the approaches to the rise and fall of a nation state (a social system)? The approaches could augment the human cognitive capacity via automated information extraction, as well as analytical capabilities via a generalized framework for instability analysis and forecasting models based on data-driven techniques. We apply a classification and forecasting framework to conflict and instability analysis, and the objectives are to: (1) present a generalized data-driven framework for conflict analysis and forecasting, (2) show that state-of-the-art pattern classification techniques provide significant improvements to forecasting accuracy, and (3) introduce classification problems arising in social sciences to the engineering community for further enhancement of analysis techniques. The effort in this thesis will help political decision makers to successfully intervene by identifying and forecasting the relative stability of a state. The final direction of research in this thesis is on integrating disparate diagnostic approaches into a rapid prototyping platform for analyzing engineering and social systems. The current approaches to FDD and forecasting are time-consuming and labor-intensive and are conducted via independent computing platforms. The integrated FDD and forecasting toolbox will provide a unified computing platform to solve diagnostic and forecasting problems, and our diagnostic algorithms (i.e., data preprocessing, data-driven classification, fusion, performance evaluation, and forecasting) in the toolbox will continue to evolve in the future.
|Year of publication:||
|Subject:||Electrical engineering | Political science|
|Type of publication:||Other|
Dissertations Collection for University of Connecticut
Saved in favorites