Disentangled Dynamic Deviation Transformer Networks for Multivariate Time Series Anomaly Detection.

Area of Science:

• Machine Learning
• Time Series Analysis
• Anomaly Detection

Background:

• Multivariate time series anomaly detection commonly uses graph neural networks to model sensor dependencies.
• Existing methods often focus on fixed sensor dependencies, neglecting nonlinear and dynamic inter-sensor and temporal relationships, which leads to false alarms.

Purpose of the Study:

• To propose a novel disentangled dynamic deviation transformer network (D3TN) for enhanced anomaly detection in multivariate time series.
• To jointly model multiscale dynamic inter-sensor dependencies and long-term temporal dependencies for improved prediction accuracy.

Main Methods:

• A disentangled multiscale aggregation scheme is designed to represent hidden sensor dependencies and learn fixed inter-sensor dependencies based on static topology.
• A self-attention mechanism is employed to capture dynamic inter-sensor dependencies influenced by real-time situations and anomalies.
• Complex temporal correlations are processed in parallel across multiple time steps.

Main Results:

• The proposed D3TN effectively models both static and dynamic inter-sensor dependencies.
• The network captures long-term temporal correlations across multiple time steps.
• Experiments on three real datasets demonstrate significant performance improvements over state-of-the-art methods.

Conclusions:

• The D3TN architecture offers a robust approach to multivariate time series anomaly detection.
• By addressing dynamic and multiscale dependencies, D3TN enhances prediction accuracy and reduces false alarms.
• This method advances the field of anomaly detection for complex time series data.