A Diffusion-Based Time-Frequency Dual-Stream Contrastive Learning Model for Multivariate Time Series Anomaly...

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series anomaly detection is crucial for industrial, financial, and medical surveillance.
• Existing methods struggle with adapting to anomalies during training and generalizing due to non-stationary data.
• Distribution shifts between training and testing data impair model performance.

Purpose of the Study:

• To propose the TFCID model for robust multivariate time series anomaly detection.
• To address limitations of existing methods, including adaptation to anomalies and distribution shifts.
• To improve the accuracy and generalization of anomaly detection systems.

Main Methods:

• The TFCID model utilizes diffusion principles for precise imputation of missing time series data.
• It incorporates an unconditional diffusion model with imputation masking in the temporal stream.
• An amplitude-aware frequency-domain masked autoencoder captures periodic anomalies in the frequency stream.
• Adversarial contrastive learning minimizes discrepancies between temporal and frequency representations.

Main Results:

• TFCID effectively imputes missing data, preventing anomalies from interfering with training.
• The model accurately captures anomalies in both temporal and frequency domains.
• Experimental results on five benchmark datasets demonstrate significant improvements in detection accuracy (F1-Score).

Conclusions:

• The TFCID model offers a novel and effective solution for multivariate time series anomaly detection.
• Its innovative approach overcomes key challenges, leading to superior performance.
• TFCID demonstrates significant outperformance compared to state-of-the-art methods.