| dc.description.abstract | Engineering systems following disruptive events usually experience abrupt performance degradation over time. How to mitigate the disastrous effect of unanticipated events and restore the performance of the system to the original level is worthy of investigation. In this dissertation, we leverage machine learning and optimization techniques to investigate a variety of measures that can be taken before, during, and after the occurrence of extreme events to assess and enhance system resilience. Towards this end, six individual objectives are pursued: I. Prior to the occurrence of a hazardous event, we develop data-driven models to forecast when the hazardous event might occur in the future, thereby increasing stakeholder’s situation awareness; II. If a system malfunction has already happened, a hybrid model that blends multiple classification models is developed to predict the severity associated with the consequences in terms of their risk levels; III. Since operators’ experience and prior training plays a significant role in diagnosing and responding to off-nominal events, we develop a machine learning framework to measure the reliability of human operators in responding to malfunction events, based on multiple types of data collected from a human-in-the-loop experimental study; IV. Simulation data is used to characterize the performance of algorithmic response in managing an abnormal event; V. We investigate a design-for-resilience methodology, focusing on number and locations of service centers that respond to a disastrous event; VI. We also investigate a system reconfiguration strategy for resilient response to the increased demand caused by an extreme event.
<br />
<br />
A variety of machine learning and optimization models are investigated in this dissertation to accomplish the above objectives. Along the machine learning front, a multi-fidelity deep learning model is developed to forecast system behavior over time, thereby enabling early warning regarding the occurrence of system hazards; the model accounts for variability in the data and uncertainty in the prediction. In addition, a hybrid model that blends support vector machine and ensemble of deep neural networks is trained to predict the consequence of abnormal events. Finally, support vector machine-based models are constructed to assess the performance of human and algorithmic responses to hazardous events. With respect to system resilience enhancement, we leverage a multi-level cross-entropy algorithm to tackle the formulated NP-hard bi-level optimization problems, where the samples in the cross-entropy algorithm adaptively converge to near-optimal solution within limited time. This algorithm is used to optimize the design of a logistics service center distribution by accounting for the potential impact of natural disasters, as well as to optimize the reconfiguration of an already existing traffic network to mitigate the system-wide congestion caused by large-volume evacuation out of a disaster-hit area. | |
