DA-LSTM-VAE: Dual-Stage Attention-Based LSTM-VAE for KPI Anomaly Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Key Performance Indicators (KPIs) are crucial for monitoring system operations.
• Anomaly detection in KPIs is vital for rapid fault identification and system maintenance.
• Existing methods may struggle with noise and long-term dependencies in time-series data.

Purpose of the Study:

• To propose a novel dual-stage attention-based LSTM-VAE model for enhanced KPI anomaly detection.
• To improve the accuracy and robustness of anomaly detection in system monitoring.
• To reduce manual intervention through an adaptive thresholding mechanism.

Main Methods:

• Utilized Long-Short-Term Memory (LSTM) units within a Variational Autoencoder (VAE) framework to capture temporal correlations.
• Implemented a dual-stage attention mechanism: time attention in the encoder and feature attention in the decoder.
• Developed an adaptive threshold method based on reconstruction probability for anomaly scoring.

Main Results:

• The proposed DA-LSTM-VAE model demonstrated superior performance compared to baseline methods on a public dataset.
• Time attention effectively filtered noise by adaptively weighting input sequences.
• Feature attention improved the capture of long-term dependencies in time-series data.
• The adaptive threshold method minimized false positives and negatives.

Conclusions:

• The DA-LSTM-VAE model offers a significant advancement in KPI anomaly detection.
• The dual-stage attention mechanism enhances the model's ability to handle complex time-series data.
• The adaptive thresholding provides a more reliable and automated approach to anomaly identification.