Rochester Institute of Technology
Novel Detection and Analysis using Deep Variational Autoencoders
Abstract
dc:description.abstractThis paper presents a Novel Identification System which uses generative modeling techniques and Gaussian Mixture Models (GMMs) to identify the main process variables involved in a novel event from multivariate data. Features are generated and subsequently dimensionally reduced by using a Variational Autoencoder (VAE) supplemented by a denoising criterion and a β disentangling method. The GMM parameters are learned using the Expectation Maximization(EM) algorithm on features collected from only normal operating conditions. A one-class classification is achieved by thresholding the likelihoods by a statistically derived value. The Novel Identification method is verified as a detection method on existing Radio Frequency (RF) Generators and standard classification datasets. The RF dataset contains 2 different models of generators with almost 100 unique units tested. Novel Detection on these generators achieved an average testing true positive rate of 97.31% with an overall target class accuracy of 98.16%. A second application has the network evaluate process variables of the RF generators when a novel event is detected. This is achieved by using the VAE decoding layers to map the GMM parameters back to a space equivalent to the original input, resulting in a way to directly estimate the process variables fitness.
Degree
thesis:*- Name thesis:degree_name
- Electrical Engineering (MS)
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Electrical Engineering (KGCOE)
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Graydon, Tucker B
- Contributors dc:contributor
-
- Ferat Sahin
Subjects
dc:subject × 5Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.rit.edu/theses/9897