Advancements in Time Series Analysis: from Regular to Anomalous, Cloud-based to On-Device, and Univariate to Multivariate

The School of EECS is hosting the following PhD Thesis Review Seminar:

Advancements in Time Series Analysis: from Regular to Anomalous, Cloud-based to On-Device, and Univariate to Multivariate

Speaker: Yuting (Skye) Sun
Host: Prof. Hongzhi Yin 

Abstract: Time series data surrounds us and shapes our world at every moment. In the era of big data, time series analysis has surged forward with help of vast datasets and advanced computing. This research explores the latest advancements in time series analysis by examining three key aspects: pattern recognition, on-device applications, and complex multivariate time series modelling.

Most industrial time series data typically exhibits various trends, cycles, and irregularities over time. Detecting abnormal patterns in these cyber-physical systems is crucial for industrial operations. Despite numerous efforts to address this task in recent years, challenges persist due to label scarcity and the diversity of anomalies. Also, recent discussions have shown that flawed benchmarks and ill-posed evaluation create an illusion of superior performance for deep learning models and mask their actual shortcomings. In this case, we revisit debates about deep learning failures and chart a new path for designing robust deep learning-based methods in time series anomaly detection. Additionally, since most industrial time series applications (e.g., forecasting, anomaly detection) expect timely responses or updates, a low-latency solution is required to offload these prediction models from the cloud servers to sensor devices. However, on device real-time time series applications face many challenges given the limited memory and storage budget. Thus, we propose a framework to deploy off-the-shelf time series models on tiny Industrial Internet of Things (IIoT) devices, also known as microcontrollers (MCUs). While industrial time series often exhibit seasonality and periodicities, multivariate time series modeling poses a significant challenge, especially when dealing with noise and dynamic correlations. This research will also provide an overview of the potential solutions and emerging trends in addressing the complexities of this problem.

Bio: Yuting (Skye) Sun is a Ph.D. student from the School of EECS at The University of Queensland under the supervision of Prof. Hongzhi Yin, Dr. Thomas Taimre, and Dr Slava Vaisman. She received her master’s degree in Data Science from UQ. Her research interests include time series anomaly detection, spatiotemporal modelling, and on-device learning.