An anomaly detection model for multivariate time series with anomaly perception.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Multivariate time series anomaly detection is vital for IT applications.
• Unsupervised methods dominate due to rare anomaly labels.
• Limited anomaly labels are feasible and offer valuable insights.

Purpose of the Study:

• To improve multivariate time series anomaly detection using limited anomaly samples.
• To propose a novel deep learning model, EDD (Encoder-Decoder-Discriminator).

Main Methods:

• Integrated Graph Attention Network and Long Short-Term Memory (LSTM) for spatial-temporal feature extraction.
• Mapped time series data into a latent space for anomaly identification.
• Utilized a specialized loss function to cluster normal data and disperse abnormal data.

Main Results:

• The EDD model demonstrated significant superiority in multivariate time series anomaly detection.
• Achieved an average F1-Score outperforming the second-best method by 2.7% and 73.4% across evaluation approaches.
• Validated the effectiveness of leveraging limited anomaly samples for robust detection.

Conclusions:

• The proposed EDD model is highly effective for anomaly detection in multivariate time series data.
• The integration of graph attention and LSTM enhances the capture of complex data patterns.
• The latent space mapping and loss function enable precise identification of anomalies.