A filter-augmented auto-encoder with learnable normalization for robust multivariate time series anomaly detection.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Existing multivariate time series (MTS) anomaly detection methods often fail when training data contains noise or anomalies.
• Current approaches struggle with accurate normal pattern learning due to reconstructing contaminated data.
• Adapting normalization schemes for complex, mixed-feature MTS datasets remains a challenge.

Purpose of the Study:

• To propose a robust anomaly detection method for multivariate time series (MTS) that addresses noise and mixed features.
• To develop a model capable of learning accurate normal patterns despite data contamination.
• To introduce a flexible normalization strategy adaptable to diverse MTS datasets.

Main Methods:

• Proposed NormFAAE, a filter-augmented auto-encoder with learnable normalization.
• Integrated a deep hybrid normalization module trained end-to-end for optimal scheme selection.
• Employed a dual-phase auto-encoder with a filter module to separate noise/anomalies, reconstructing only normal data.

Main Results:

• NormFAAE demonstrated superior performance compared to 17 baseline methods.
• The method achieved robust anomaly detection on five diverse, real-world industrial datasets.
• The learnable normalization module effectively handled mixed-feature MTS data.

Conclusions:

• NormFAAE offers a more accurate and robust approach to MTS anomaly detection, especially in contaminated environments.
• The filter-augmented auto-encoder design enhances the learning of normal data patterns.
• The proposed method provides a significant advancement for anomaly detection in complex industrial time series data.