Thesis Title: Dynamic Robust Sparse Modelling and Sampling of High-Dimensional Data Streams for Online Process Monitoring

Advisors: Dr. Jan Shi and Dr. Yajun Mei

Committee members: 

Dr. Kamran Paynabar

Dr. Nagi Gebraeel  

Dr. Kaibo Liu (Department of Industrial and Systems Engineering, University of Wisconsin-Madison)

Date and Time: Wednesday, April 17, 2019, 10:00 AM

Location: Groseclose 304

Abstract:

This thesis contributes to the area of System Informatics and Control (SIAC) to develop systematic and dynamic methodologies for effective monitoring and change detection in complex systems. The proposed procedures facilitate (1) dynamic strategies for sampling in the event of resource constraints, (2) robust modelling complex data structures with sparse spatial dependencies, (3) adaptive updating of system models based on novel features extracted from online observations. This thesis ties advanced statistical methodologies and engineering knowledge to address practical applications in various areas such as advanced manufacturing service systems.

The thesis begins by addressing manufacturing and service systems with resource limitations. In Chapter 1, we investigate methods for sampling within data rich environments and propose a dynamic sampling strategy for monitoring these environments with restricted resource. A procedure called “Correlation based Dynamic  Sampling” (CDS) that leverages spatial dependencies within the data streams to improve decision making when deploying sensors in real time. 

Chapter 2 examines the system modelling aspect of data rich environments by exploring dimension reduction methods. We develop a dimension reduction method named “Robust Sparse Principal Component Analysis” (RS-PCA), that is designed to robustly estimate a lower dimensional subspace by exploiting the sparse structure that is typical in high-dimensional data. The probabilistic approach for modelling offers a direct medium for making inferences on system conditions. 

Subsequently, Chapter 3 extends the aforementioned RS-PCA procedure for implementation in dynamic systems. The proposed adaptive RS-PCA method reduces the false alarm rate that may result from implementing static procedures. The trade-off between learning from novel observations and overfitting is managed by proposing an adaptive robust learning rate through stochastic variational inference.

In summary, this thesis sheds the light on the challenges of modeling and sampling within data rich environments for the purpose of process control. Adaptive systematic modelling and sampling strategies are developed to address common challenges from these environments. Furthermore, these strategies are implemented on several exemplary systems to assess their capability for real time application in practical scenarios.