Past Event: Oden Institute Seminar

Data-driven learning for engineering sensor-actuator placement and forecasting

Krithika Manohar, NSF postdoctoral fellow and von Karman instructor in Computing & Mathematical Sciences, California Institute of Technology

Abstract

The increasingly high-dimensional data generated by complex systems presents a tremendous challenge for efficient prediction, estimation and control. Data-driven learning provides a powerful tool in the analysis of systems strictly governed by physical laws, and can uncover heavily compressed representations of the underlying dynamics. The first part of my talk describes using linear dimensionality reduction, such as PCA and balanced model reduction, to efficiently solve the optimal sensor placement problem. This method bypasses the combinatorially complex brute-force search through all possible candidate placements, and generalizes to actuator placement for control using observability and controllability metrics. The resulting sensors are used to reconstruct flow fields and imaging data with many thousands of candidate locations, including an application to streamline the manufacturing process for the Boeing 787 using targeted measurement of gaps in the wing-body join structure. Next, I describe ongoing work using nonlinear dimensionality reduction with diffusion kernels to forecast macroscale dynamics in slow-fast systems, which are commonly observed in climate and oceanography. I will discuss future directions in designing observables and sensor placements for engineering systems when the dynamics are nonlinear, only partially observed, or unknown. Bio Krithika Manohar is NSF postdoctoral fellow and von Karman instructor in Computing & Mathematical Sciences at the California Institute of Technology. She received the dual B.S. degree in Mathematics and Computer Science from University of Massachusetts Lowell, and the Ph.D. degree in Applied Mathematics from University of Washington. She is a recipient of the Boeing Award for Excellence in Research for her work on data-driven sensor placement methods, and was awarded the NSF Mathematical Sciences Postdoctoral Research Fellowship in 2018. Her research leverages data-driven learning to solve prediction, estimation and control problems in high-dimensional complex systems.