Anomaly Detection of Time Series With Smoothness-Inducing Sequential Variational Auto-Encoder

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Deep generative models excel at learning latent representations and modeling complex time series dependencies.
• Variational auto-encoders (VAEs) are effective generative models for time series.
• Existing Markov models often assume constant noise, limiting their applicability to nonstationary data.

Purpose of the Study:

• To develop a robust method for estimating and detecting anomalies in multidimensional time series.
• To address the fragility of flexible nonstationary models to anomalies.
• To enhance density estimation for improved anomaly detection.

Main Methods:

• A smoothness-inducing sequential variational auto-encoder (SISVAE) model is proposed.
• The model utilizes a recurrent neural network backbone for capturing temporal structures.
• A smoothness-inducing prior is introduced as a regularizer to penalize non-smooth reconstructions.
• A novel stochastic gradient variational Bayes estimator is employed for efficient learning.

Main Results:

• The SISVAE model demonstrates robust density estimation for multidimensional time series.
• The model effectively detects anomalies using reconstruction probability and reconstruction error.
• The proposed method shows strong performance on both synthetic and real-world benchmark datasets.

Conclusions:

• The SISVAE model provides a robust and effective approach for anomaly detection in multidimensional time series.
• The smoothness-inducing prior successfully mitigates model fragility caused by data flexibility.
• The method advances the capabilities of deep generative models in time series analysis.