Generalized flow-based variational autoencoder networks for anomaly detection in multivariate time series

Time series are widely used in applications such as finance, robotics, telecommunications, astronomy, and many more. Detecting anomalies like robotic arm failures or server attacks is a valuable and important task. Recent research in anomaly detection in multivariate temporal data formulates the problem as one of variational inference. To solve this problem, such approaches have used variational autoencoders to try to learn the probability distribution of multiple time series. Variational autoencoders are used as a way to approximate intractable distributions, and methods to improve these approximations are explored through the use of normalizing flows. By applying normalizing flow transforms to the latent variables of a variational autoencoder, the true latent distribution can be more richly modeled and learned, thus enabling better metrics for anomaly detection. This thesis explores five different types of normalizing flow in the context of three multivariate datasets, and demonstrates the effectiveness, compared to prior research, of flows and convolutional networks for anomaly detection by improving popular metrics like the F1-score.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Shah, Raimi

Contributors dc:contributor

• Zhao, Zhizhen

• anomaly detection
• deep learning
• normalizing flow
• variational autoencoder
• convolutional neural network

Statement dc:rights

• Copyright 2021 Raimi Shah

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110543

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110543