A Unified Unsupervised Anomaly Detection Framework with Score-based Generative Modeling for Multivariate Time Series
Loading...
Authors
Sarker, Prithul
Issue Date
2024
Type
Dissertation
Language
en_US
Keywords
Anomaly Detection , Diffusion Model , Multivariate Time Series , Score-based Generative Modeling , Stein Score , Virtual Reality
Alternative Title
Abstract
The challenge in unsupervised anomaly detection is the unknown nature of anomalous data points. This task requires the identification of abnormal patterns within the data, even when we lack prior knowledge about what those anomalous patterns might explicitly suggest. Existing unsupervised anomaly detection methods have attempted to address this issue by focusing on limited aspects of the overall problem. These methods can be broadly categorized into four main approaches: reconstruction-based, density estimation-based, boundary description-based, and explicit data characteristic modeling-based. Although each of these methodologies has its own advantages, they are also limited by inherent weaknesses that restrict their effectiveness yielding to sub-optimal results. In this research, I present a novel methodological framework, Unified Unsupervised Anomaly Detection (U2AD), that comprehensively addresses the problem of anomaly detection in multivariate time series. This approach provides a deeper understanding of anomalies within the data distribution space while elucidating the dynamics of non-anomalous data. The framework integrates previous techniques while offering a fresh, holistic perspective. This allows for the creation of customized solutions for various applications by increasing adaptability in selecting appropriate components needed for accurate, robust, and efficient anomaly detection in multivariate time series. Utilizing score-based generative modeling in conjunction with reengineered time-dependent score network and novel training objectives further enhances comprehension of anomalies. Additionally, reconstruction is achieved through the sampling method with deterministic numerical ordinary differential equation solver. Extensive experiments demonstrate that this methodology not only improves anomaly detection precision but also identifies anomalies at earlier stages than current state-of-the-art methods.