Improving Multivariate Time-Series Anomaly Detection in Industrial Sensor Networks Using Entropy-Based Feature...

Area of Science:

• Industrial IoT and Cyber-Physical Systems
• Complex Systems Analysis
• Machine Learning for Anomaly Detection

Background:

• Anomaly detection in multivariate time-series data is challenging for complex industrial systems like Cyber-Physical Systems (CPSs) and the Internet of Things (IoT).
• Interconnected sensors in these systems mean local anomalies can propagate, complicating detection due to implicit and complex relationships.
• Existing methods often struggle to systematically characterize these intricate system interdependencies.

Purpose of the Study:

• To develop an advanced anomaly detection method for complex industrial systems using multivariate time-series data.
• To formally represent and model the implicit relationships within these interconnected systems.
• To enhance the accuracy and systematic characterization of anomaly detection.

Main Methods:

• Utilized graph neural networks (GNNs) integrated with a structure-entropy-based attention mechanism.
• Developed a network-based structural model to represent implicit relationships in complex industrial systems.
• Implemented a method to distinguish the weights of high-order neighbor nodes based on their location and analyze system entropy to identify key elements.

Main Results:

• The proposed method demonstrated improved anomaly detection performance compared to baseline approaches.
• Validated effectiveness across multiple benchmark datasets including SMAT, MSL, SWaT, and WADI.
• Successfully modeled multi-element relationships and formally represented implicit system interactions.

Conclusions:

• The graph neural network with a structure-entropy-based attention mechanism offers a robust solution for anomaly detection in complex industrial systems.
• This approach provides a systematic way to characterize implicit relationships and enhance detection accuracy.
• The findings are applicable to diverse fields including Industrial Control Systems (ICSs), Intrusion Detection Systems (IDSs), and remote sensing.