Memory-guided mask reconstruction with central contrastive learning for robust multivariate time series anomaly...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Unsupervised multivariate time series anomaly detection (MTSAD) methods often struggle with semantic biases from masking and overfitting to contaminated data.
• Existing contrastive learning approaches in MTSAD face challenges in separating feature distributions, limiting the capture of high-level semantic information.

Purpose of the Study:

• To propose a novel Memory-guided Mask Reconstruction with Central Contrastive Learning (MMR-CCL) method for unsupervised MTSAD.
• To address semantic distortions and overfitting issues in mask reconstruction-based MTSAD.
• To improve the extraction of high-level semantic information in MTSAD using contrastive learning.

Main Methods:

• Developed a global correlation-aware memory module to guide the mask reconstruction process and capture global commonality.
• Introduced a memory regulation factor to quantify and suppress overfitting to contaminated data.
• Implemented a memory-anchoring central contrastive learning strategy using unmasked data as anchors, masked data as positive samples, and noise-filled data as negative samples.

Main Results:

• The proposed MMR-CCL method effectively mitigates semantic distortions and enhances temporal context comprehension.
• The memory regulation factor successfully suppresses overfitting to contaminated data with strong local correlations.
• Extensive experiments on six public datasets show MMR-CCL outperforms 22 existing MTSAD methods.

Conclusions:

• MMR-CCL provides a robust framework for unsupervised multivariate time series anomaly detection.
• The integration of memory-guided reconstruction and central contrastive learning significantly improves detection performance.
• MMR-CCL establishes a more rational discrimination boundary, aligning reconstruction with normal pattern spaces.