A robust multi-scale feature extraction framework with dual memory module for multivariate time series anomaly...

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Existing multivariate time series anomaly detection (MTSAD) methods often fail with noisy training data, reconstructing anomalies and blurring distinctions.
• Probabilistic methods offer noise robustness but lack stable training and anomaly suppression.
• Memory module methods improve anomaly detection but compromise normal pattern reconstruction.

Purpose of the Study:

• To propose a robust multi-scale feature extraction framework for MTSAD that addresses limitations of existing methods.
• To enhance the extraction of both local and long-term temporal dependencies.
• To improve the distinction between normal and anomalous data patterns, even with contaminated training sets.

Main Methods:

• Utilizes consecutive neighboring windows as input for capturing local and long-term dependencies.
• Employs a dual memory-augmented encoder for extracting global typical patterns and local common features.
• Incorporates a multi-scale fusion module to integrate features from different semantic levels for reconstruction.

Main Results:

• The proposed framework effectively extracts multi-scale features, fusing diverse semantic information and temporal dependencies.
• The dual memory module ensures accurate reconstruction of normal data while suppressing anomalous generalization.
• Experimental results demonstrate superior performance over 16 baseline methods across five diverse datasets.

Conclusions:

• The proposed robust multi-scale feature extraction framework with a dual memory module significantly advances MTSAD.
• This approach offers improved accuracy and robustness, particularly in the presence of noisy or contaminated training data.
• The method's ability to handle multi-scale features and temporal dependencies provides a more comprehensive anomaly detection solution.