Ranking Causal Anomalies via Temporal and Dynamical Analysis on Vanishing Correlations

Area of Science:

• Computer Science
• Network Science
• Cyber-Physical Systems

Background:

• Real-time monitoring data from large-scale information and cyber-physical systems has increased significantly.
• Detecting system anomalies is crucial for security, fault management, and industrial optimization.
• Invariant networks characterize system behavior, with vanishing correlations indicating potential anomalies.

Purpose of the Study:

• To address limitations of existing anomaly detection methods in invariant networks.
• To propose a novel network diffusion framework for identifying and ranking causal anomalies.
• To improve the accuracy and robustness of anomaly detection in complex systems.

Main Methods:

• Developed a network diffusion based framework to model fault propagation.
• Performed joint inference on structural and time-evolving broken invariance patterns.
• Utilized invariant networks where nodes represent system components and edges represent stable interactions.

Main Results:

• The proposed framework effectively models fault propagation across the entire invariant network.
• It accurately identifies high-confidence causal anomalies responsible for vanishing correlations.
• The approach compensates for unstructured measurement noise, enhancing detection robustness.

Conclusions:

• The network diffusion framework offers a superior method for causal anomaly detection in invariant networks.
• It overcomes limitations of existing approaches by considering fault propagation and temporal patterns.
• Experimental results on synthetic and real-world datasets validate the framework's effectiveness.