Title: Big-Data in Monitoring Engineering Systems: Forecasting Multivariate Time-Series

09:00 ~ 10:45 a.m., 19 February 2024 (Monday), The Sukosol Hotel, Bangkok, Thailand

Pavel Loskot

Zhejiang University-University of Illinois at Urbana Champaign Institute

Abstract

Many engineering systems are nowadays monitored by plethora of sensors that continuously report various quantities about the system state. It results in large volumes of longitudinal data that must be processed in order to either identify the changes in the system state that have already occurred or that may occur in near future. The corresponding longitudinal data models must capture the short-term and long-term spatio-temporal patterns and dependencies. The traditional approach to modeling multivariate time-series assumes variants of autoregressive random processes under the Markov property. The caveat is that such models fully rely on Gaussianity and stationarity of samples as well as on choosing the right model order. More modern approaches for processing multivariate time-series leverage the recent progress in deep learning architectures including ensemble learning, geometric learning, zero-short and cold-start learning, transformers with modified attention mechanism and other. These methods enable auto-normalization of data, and in many cases, also automated feature selection. More importantly, these methods are finally starting to outperform the traditional methods of statistical filtering of stochastic signals while relaxing many of the previously required assumptions. The objectives of this tutorial are: (1) survey deep learning methods for forecasting longitudinal data, (2) compare these methods with traditional approaches involving statistical models of random processes in terms of performance, complexity, interpretability, application constraints and other such aspects, and, (3) identify and discuss the benchmark datasets, relevant data science competitions, key literature, main applications, and open research problems.

Bio

B. Pavel Loskot joined the ZJU-UIUC Institute, Haining, China, in January 2021 as Associate Professor after being 14 years a Senior Lecture at Swansea University in the UK. In the past 25 years, he participated in numerous collaborative research and development projects, and also held a number of paid consultancy contracts with various companies mainly in the area of designing and deploying wireless and optical communication systems. He holds PhD in Wireless Communications from the University of Alberta, Canada, and MSc in Radioelectronics and BSc in Medical Electronics from the Czech Technical University of Prague, Czech Republic. He published papers in various engineering, mathematical and biological journals, received 4 best paper awards, and delivered over 50 keynote talks and tutorials in international conferences. Pavel Loskot is a Senior Member of the IEEE, a Fellow of the Higher Education Academy in the UK, and was awarded the Recognized Research Supervisor title by the UK Council for Graduate Education. His current research interests are concerned with mathematical and probabilistic modeling, statistical signal processing and classical machine learning for multi-sensor data in biomedicine, and computational molecular biology, and exploring algebraic methods for beyond-calculus signal and data processing.