Time Series Anomaly Detection Based on a Self-Supervised Forecasting Model

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Time Series Anomaly Detection Based on a Self-Supervised 
        Forecasting Model"

by

SHEN Che

Abstract:

This paper presents a self-supervised learning-based approach for time 
series anomaly detection. The study utilizes the characteristics of time 
series data to develop a forecasting model that can identify anomalies in 
time series. The objectives of the study are to utilize self-supervised 
learning methods to realize time series forecasting, apply the trained 
model to detect anomalies, and compare the performance of this method with 
existing ones. The approach trains the forecasting model using the part of 
the dataset without anomalies. The trained model is then applied to the 
dataset with potential anomalies, and the data that are far from the 
predicted values are identified as anomalies. The results show that the 
proposed approach outperforms existing methods in terms of anomaly 
detection accuracy in some specific datasets. This study provides a novel 
approach for time series anomaly detection and demonstrates the potential 
of self-supervised learning for real-world applications.


Date            : 4 May 2023 (Thursday)

Time            : 15:00 - 15:40

Venue           : Room 5562 (near lifts 27/28), HKUST

Advisor         : Prof. YEUNG Dit-Yan

2nd Reader      : Prof. TANG Chi-Keung