Multi-channel anomaly detection using graphical models.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Anomaly detection in multivariate time-series is crucial for asset monitoring and safety.
• Existing methods often neglect inter-channel feature dependencies, limiting accuracy.
• Detecting anomalies across multiple channels in time-series data remains a challenge.

Purpose of the Study:

• To introduce G-BOCPD, a novel graphical model-based annotation method for anomaly detection in multi-channel multivariate time-series data.
• To address the limitations of existing methods by considering interrelations between features and across multiple channels.
• To automatically detect and annotate anomalous segments in complex time-series data.

Main Methods:

• G-BOCPD employs a hybrid approach combining the graphical lasso and expectation-maximization algorithms.
• It estimates the concentration matrix to represent variable dependencies, utilizing the graphical lasso.
• Minimal path clustering is used for segment annotation, identifying diverse behaviors and patterns.

Main Results:

• G-BOCPD effectively detects anomalies in multi-channel multivariate time-series data.
• The method was successfully applied to real-world data from train engines and doors.
• G-BOCPD demonstrated superior performance over existing approaches in precision, recall, and F1-score.

Conclusions:

• G-BOCPD offers a robust solution for anomaly detection in multi-channel multivariate time-series.
• The method enhances fault detection and diagnosis for critical systems.
• This approach improves asset condition monitoring, reducing downtime and enhancing safety.