Online Learning with Big Data
Big data not only refers the size of data but also refers the stream of heterogeneous types of data collected continuously (or, sequentially). Many statistical and engineering models focus on characterizing a system (or, process) with historical data and predicting the future observations based on the developed static model. However, social and mechanical systems evolve over time due to a wide range of quantifiable/not-quantifiable, measurable/non-measurable inputs, interactions among input variables and other complex factors. We develop methods to track system’s dynamics and to use the results for prediction and control. Handling a stream of data might increase the computational complexity, so we also consider the question of how to relieve the computational burden and data storage requirement while not losing the information that the data contains.
Reliability Assessment, uncertainty quantification with Stochastic Simulations
The purpose of this research area is to provide computationally efficient methods to evaluate reliability using stochastic simulations. As simulation models become more realistic and their degrees of freedom increase, reliability evaluation remains challenging, because each simulation replication is computationally expensive. There has been a rich body of studies to run simulations efficiently to obtain estimates of interest. These studies have been limited to the cases where all of the random components in the simulation can be controlled (or, sampled). However, controlling all of the components inside a simulator is difficult, if not impossible, when a simulator models complicated processes in high dimensional spaces. We are concerned with reliability evaluation using stochastic simulations which generate random outputs given a fixed input. An optimal importance sampling method, which minimize the estimator variance, has been devised and validated using aeroelastic simulators.
Fault Diagnosis, Condition Monitoring, Structural Health Monitoring, and Classifications
We develop the methodology to detect abnormal patterns of systems by examining the variations in the operational responses. An integrative framework has been developed to define the decision boundaries that distinguish faulty conditions from normal conditions have been devised. The framework includes (1) modeling of nonlinear relationship between system’s response and operating conditions and (2) modeling heterogeneous variability of system responses in a range of operating conditions.
Operations and Maintenance (O&M) Optimization in Wind Power Systems
Wind turbines operate under highly variable loading conditions, and maintenance is constrained by stochastic weather conditions that can disallow or disrupt repair activities. Using a partially observed Markov decision process, I devised condition-based maintenance models including (1) a static model where the optimality structure can be analytically attained as a closed form under the assumption that the weather conditions remain stationary over the time; (2) dynamic model where the O&M policy is dynamically adapted to season-dependent weather conditions; (3) a tractable approximation of dynamic decision-making in a large-scale wind farm. The developed models has been integrated with discrete-event simulations for validations.