Prototypical contrastive learning with patch-based spatio-temporal alignment for multivariate time series anomaly...

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series (MTS) anomaly detection is challenged by sensor interdependencies, noise, and long-range dependency modeling.
• Current methods struggle to balance computational efficiency with accurate normal pattern modeling.

Purpose of the Study:

• To propose P-ALIGN, a novel framework for efficient and accurate MTS anomaly detection.
• To address limitations in existing methods regarding noise suppression and anomaly discrimination.

Main Methods:

• P-ALIGN utilizes patch-based feature extraction for linear complexity long-term context capture.
• An EmbedPatch encoder learns normal prototypes for feature alignment, suppressing noise and preventing anomaly over-reconstruction.
• A Contrastive Fusion module enhances discrimination between normal and abnormal data distributions.

Main Results:

• P-ALIGN demonstrates superior performance across six real-world benchmarks.
• Achieved an 11% improvement in F1-score and a 12.23% increase in Normalized Affinity (NAff).

Conclusions:

• P-ALIGN offers an effective solution for MTS anomaly detection, balancing efficiency and accuracy.
• The framework shows significant potential for real-world applications requiring robust anomaly identification.