EECS Seminar: From Sensors to "Structures as Sensors" - Combining Sensing and Learning in Dirty Cyberphysical Systems

Abstract: This talk introduces physics-aided approaches to improve learning in cyberphysical systems (CPS). Learning has become a useful tool for data-rich problems. However, its use in CPS has been limited because of its need for a large amount of well-labeled data for each application and deployment. This is especially challenging and often impossible due to the high number of variables that can affect data distribution in CPS (e.g., weather, time, persons, etc.). This talk introduces combinational techniques that incorporate physical models and hardware characteristics to enable learning in CPS with “small data.” We 1) improve sensed data through actuation of the sensing system, 2) incorporate physical characteristics to guide learning and 3) combine and transfer data from other domains using the physical understanding. This talk illustrates these approaches through two types of CPS. First, a citywide mobile system that actuates taxi fleets to optimize multiple goals. Then I will introduce Structures as Sensors, where a building acts as the physical elements of the sensor; and the structural response is interpreted to obtain information about the occupants. I will present the result of these applications through our real-world deployments in the city and medical care facilities.

Bio: Pei Zhang is an associate professor in the Department of Electrical Engineering and Computer Science at the University of Michigan, Ann Arbor. He received his bachelor's degree from the California Institute of Technology in 2002, and his doctorate in electrical engineering from Princeton University in 2008. His early work ZebraNet is considered one of the seminal works in sensor networks, for which he received the SenSys Test-of-Time Award in 2017. His recent work focuses on cyberphysical systems that utilize the physical properties of vehicles and structures to discover surrounding physical information. His work combines machine learning-based data models, physics-based models, as well as heuristic models to improve learning using a small amount of labeled sensor data. His work is applied to fields of medicine, farming and smart retail and is part of multiple startups. His work has been featured in popular media including CNN, CBS, NBC, Science Channel, Discovery Channel, Scientific American, etc. In addition, he has received various best paper awards, the NSF CAREER award (2012), SenSys Test of Time Award (2017), Google faculty award (2013, 2016), and he was a member of the Department of Defense Computer Science Studies Panel.