Decomposition-based multi-scale transformer framework for time series anomaly detection.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Time series anomaly detection is vital for system stability.
• Existing methods struggle with complex patterns and noise sensitivity.
• Mean squared error optimization can degrade performance on noisy data.

Purpose of the Study:

• To propose a transformer-based framework (TransDe) for multivariate time series anomaly detection.
• To address limitations in modeling complex dependencies and noise in time series data.
• To improve the accuracy and efficiency of anomaly detection systems.

Main Methods:

• A transformer-based framework (TransDe) integrating time series decomposition.
• A multi-scale patch-based transformer architecture to capture dependencies in decomposed components.
• A contrastive learning paradigm using KL divergence for aligning normal pattern representations.
• A novel asynchronous loss function with a stop-gradient strategy for efficient optimization.

Main Results:

• TransDe demonstrates superior performance compared to twelve baseline methods.
• The framework effectively learns complex patterns in normal time series data.
• Achieved state-of-the-art results, particularly in F1 score, across five public datasets.

Conclusions:

• TransDe offers an effective solution for multivariate time series anomaly detection.
• The proposed methods overcome challenges posed by complex patterns and noisy data.
• TransDe provides a computationally efficient and high-performing approach for anomaly detection.