Meta Variational Memory Transformer for Anomaly Detection of Multivariate Time Series

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Anomaly detection in multivariate time series (MTS) is critical for financial fraud and equipment monitoring.
• Current unsupervised probabilistic models for MTS anomaly detection face challenges with high computational costs and limited adaptability due to fixed parameter mappings.
• There is a need for more adaptive and computationally efficient methods for MTS anomaly detection.

Purpose of the Study:

• To introduce a novel Meta Variational Memory Transformer (MVMT) for unsupervised anomaly detection in MTS.
• To enhance the adaptability and reduce the computational burden of existing MTS anomaly detection techniques.
• To develop a model capable of capturing diverse patterns across various MTS effectively.

Main Methods:

• Developed a Meta Variational Memory Transformer (MVMT) incorporating a meta memory attention (MMA) module to encode diverse MTS patterns into memory units.
• Introduced a memory-guided probabilistic generative model that uses learned memory units as priors for latent states, creating expressive MTS representations.
• Implemented a Transformer-based upward-downward variational inference for estimating latent variable posterior distributions.

Main Results:

• The proposed MVMT model demonstrated effectiveness in one-for-all anomaly detection tasks across six diverse datasets.
• The MMA module enabled the generation of various patterns by providing a diversified prior in the latent space.
• MVMT achieved more expressive MTS representations through memory-guided latent state priors.

Conclusions:

• MVMT offers a significant advancement in unsupervised anomaly detection for multivariate time series.
• The model's architecture enhances adaptability and efficiency compared to previous methods.
• MVMT shows strong potential for real-world applications in fraud detection and industrial monitoring.