Disentangled Dynamic Deviation Transformer Networks for Multivariate Time Series Anomaly Detection

Abstract

Graph neural networks have been widely used by multivariate time series-based anomaly detection algorithms to model the dependencies of system sensors. Previous studies have focused on learning the fixed dependency patterns between sensors. However, they ignore that the inter-sensor and temporal dependencies of time series are highly nonlinear and dynamic, leading to inevitable false alarms. In this paper, we propose a novel disentangled dynamic deviation transformer network (D3TN) for anomaly detection of multivariate time series, which jointly exploits multiscale dynamic inter-sensor dependencies and long-term temporal dependencies to improve the accuracy of multivariate time series prediction. Specifically, to disentangle the multiscale graph convolution, we design a novel disentangled multiscale aggregation scheme to better represent the hidden dependencies between sensors to learn fixed inter-sensor dependencies based on static topology. To capture dynamic inter-sensor dependencies determined by real-time monitoring situations and unexpected anomalies, we introduce a self-attention mechanism to model dynamic directed interactions in various potential subspaces influenced by various factors. In addition, complex temporal correlations across multiple time steps are simulated by processing the time series in parallel. Experiments on three real datasets show that the proposed D3TN significantly outperforms the state-of-the-art methods.

Anomaly Detection in Time Series: Current Focus and Future Challenges

“…D3TN: Disentangled Dynamic Deviation Transformer Network (D3TN) [52] is highly effective system for multivariate time series anomaly detection. It considers both short-term and long-term temporal dependencies as well as complex intersensor dependencies.…”

TCF-Trans: Temporal Context Fusion Transformer for Anomaly Detection in Time Series

“…Although LSTM-based methods achieve satisfying performances in some anomaly detection tasks, their highly computationally expensive and inefficient long temporal pattern learning capability still limits their performances. Recently, innovations based on the vanilla Transformer [22] have been widely applied in various fields such as natural language processing (NLP) [28,29], computer vision (CV) [30,31,32,33] and time series applications [34,35,36]. The vanilla Transformer has four key modules: the attention module, feed-forward module, layer normalisation module, and positional encoding module [40].…”

Survey of Time Series Data Generation in IoT

“…Through the analysis of time series data, it is possible to reveal the underlying patterns and laws in the data, discover the correlation and periodicity between events, and then deeply understand the nature and mechanism of the event itself, providing strong support for research in related disciplines. Specifically, a deep understanding of time trends [ 14 ], periodicity [ 15 , 16 ], correlation [ 17 , 18 ], etc., can be gained and valuable information can be further extracted, such as anomaly detection [ 19 , 20 , 21 ], classification [ 22 , 23 , 24 ], clustering [ 25 , 26 ], etc. These studies require a large amount of time series data for experiments to test the effectiveness and practicality of different algorithms and techniques, optimize the parameters and structure of algorithms, evaluate the performance and accuracy of different techniques, and train machine learning models.…”