Masked Graph Neural Networks for Unsupervised Anomaly Detection in Multivariate Time Series

Area of Science:

• Data Science
• Artificial Intelligence
• Machine Learning

Background:

• Multivariate time-series anomaly detection is crucial for applications like grid operations and machine fault detection.
• High-dimensional sensor data with latent relationships presents significant challenges for existing unsupervised methods.
• Current methods struggle to capture complex inter-sensor associations and leverage latent relationships effectively.

Purpose of the Study:

• To propose a novel unsupervised anomaly detection method for multivariate time-series data.
• To address the limitations of existing methods in capturing complex sensor correlations.
• To develop a model capable of learning unobserved causal structures among sensors.

Main Methods:

• Introduced Masked Graph Neural Networks for Unsupervised Anomaly Detection (MGUAD).
• Employed a masking and reconstruction strategy to learn temporal context from individual sensor data.
• Utilized graph-level masking within a GNN framework to learn relationships between sensors.

Main Results:

• MGUAD demonstrated superior performance in anomaly detection across three public datasets.
• The method effectively learns unobserved causal structures among sensors.
• Experimental findings indicate MGUAD outperforms state-of-the-art anomaly detection techniques.

Conclusions:

• MGUAD offers a robust solution for unsupervised anomaly detection in high-dimensional multivariate time-series.
• The proposed masking approach enhances the learning of both temporal and graph-level contexts.
• This method advances the field by enabling better utilization of latent relationships in sensor data.