Learning the feature distribution similarities for online time series anomaly detection

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection in multi-dimensional sequential data is vital for system performance.
• Traditional contrastive methods face challenges in efficiency and interpretability.
• Existing methods often struggle with complex objective functions and module design.

Purpose of the Study:

• Introduce SimDetector, a novel Local-Global Multi-Scale Similarity Contrast network.
• Improve efficiency and interpretability in anomaly detection.
• Address limitations of current contrastive learning approaches.

Main Methods:

• Utilize a restructured GRU module for generalized local feature extraction, capturing long-term trends.
• Employ a multi-scale sparse attention module for efficient global feature extraction with pattern information.
• Propose a symmetric absolute KL divergence tailored for time series anomaly detection, emphasizing distribution differences.

Main Results:

• SimDetector achieves state-of-the-art or competitive results on real-world and synthetic datasets.
• Significantly reduces Multiply-Accumulate Operations (MACs) and memory usage compared to existing methods.
• Demonstrates improved accuracy and efficiency in identifying anomalies.

Conclusions:

• SimDetector offers an effective and efficient solution for anomaly detection in multi-dimensional sequential data.
• The proposed framework balances performance with reduced computational overhead.
• Highlights the potential of multi-scale similarity contrast for complex data analysis.