Probabilistic Predictive Analytics for Collaborative Systems

Seokhyun Chung
Department of Systems and Information Engineering, University of Virginia

Time: 2023-11-01, 12:00 - 13:00 ET
Location: Rice 540 and Zoom

Abstract

Unprecedented connectivity across tangible physical units enabled by Internet of Things (IoT) technologies has set forth a new collaborative paradigm for predictive analytics. In this paradigm, physical units can borrow knowledge from each other, allowing for significantly enhanced model learning capabilities. Despite such a key advantage, it introduces crucial challenges: (i) the sheer amount of data collected from physical units is beginning to overwhelm traditional centralized computing resources, (ii) personalized analytics is instrumental as physical units often operate under different environments or settings, (iii) real-time data collection necessitates adaptive models in collaborative analytics.

This talk presents two collaborative approaches that address the challenges above. First, motivated by the recent revolutionary increase of the computing power in physical units, I will discuss a distributed framework for learning Multi-output Gaussian processes (MGP). MGPs have seen great success in data analytics due to their intrinsic capability to transfer knowledge across units by evaluating their correlations, along with key advantages such as uncertainty quantification and flexibility. By exploiting the natural hub-and-spoke structure of connected systems with a central cloud, the proposed method can distribute the effort to learn an MGP to physical units, circumvent the need for sharing their raw data, and enable personalized predictions. Next, I will discuss connected systems that feature real-time condition monitoring (CM). Specifically, I will introduce an approach that extrapolates the future trajectory of a CM signal from an in-service unit using knowledge from other historical units. Within the regime of real-time data collection, the method can instantaneously adapt its prediction to the incoming data stream. In addition to rapid adaptation, the model can integrate available qualitative data of historical units to enhance its predictive power. Real-world applications in reliability engineering highlight the advantageous features of the proposed approaches.

Bio

Seokhyun Chung is an assistant professor in the Department of Systems and Information Engineering at the University of Virginia. Seokhyun received his Ph.D. in Industrial & Operations Engineering in 2023 from the University of Michigan. Seokhyun’s current research explores collaborative and distributed analytics where Internet of Things-enabled entities (e.g., smartphones, electric vehicles, and wearable devices) exploit their edge computing power to build smart analytics collaboratively. Seokhyun was a finalist in the Quality, Statistics, & Reliability (QSR) Best refereed paper competition at INFORMS 2022 and the Quality Control & Reliability Engineering (QCRE) Best student paper competition at IISE 2021.