Correlation-Based Anomaly Detection Method for Multi-sensor System.

Area of Science:

• Industrial monitoring and sensor systems
• Machine learning for anomaly detection
• Time-series data analysis

Background:

• Multi-sensor systems are crucial for detecting abnormalities in industrial settings.
• Existing anomaly detection methods often overlook feature correlations and their sequential changes.
• Understanding feature correlation is key to identifying anomaly propagation.

Purpose of the Study:

• To propose a novel multi-sensor anomaly detection method that leverages feature correlations in time-series data.
• To address the limitations of current methods that do not consider sequential correlation dynamics.
• To enhance the accuracy and effectiveness of anomaly detection in complex industrial environments.

Main Methods:

• Analysis of correlation characteristics in multi-sensor systems to understand anomaly propagation.
• Conversion of multidimensional time-series data into temporal correlation graphs within defined time windows.
• Application of a structure-sensitive graph neural network for anomaly detection, treating it as a graph classification problem.

Main Results:

• The proposed method achieved superior performance compared to baseline methods on three real-world industrial multi-sensor systems.
• Experimental results showed a mean F1 score exceeding 0.90 and a mean AUC score exceeding 0.95.
• The method effectively detects anomalies in multidimensional time-series data by utilizing temporal correlation graphs.

Conclusions:

• The developed method successfully identifies anomalies in multi-sensor systems by analyzing temporal correlations.
• The graph-based approach combined with graph neural networks offers a powerful framework for anomaly detection.
• This technique provides a significant advancement in monitoring industrial equipment and environments for anomalies.