Authors:
Tharindu Kumarage
;
Nadun De Silva
;
Malsha Ranawaka
;
Chamal Kuruppu
and
Surangika Ranathunga
Affiliation:
University of Moratuwa, Sri Lanka
Keyword(s):
Anomaly Detection, Industrial Software Systems, Variational Autoencoder, VAE, DBSCAN.
Abstract:
Industrial software systems are known to be used for performing critical tasks in numerous fields. Faulty conditions
in such systems can cause system outages that could lead to losses. In order to prevent potential system
faults, it is important that anomalous conditions that lead to these faults are detected effectively. Nevertheless,
the high complexity of the system components makes anomaly detection a high dimensional machine learning
problem. This paper presents the application of a deep learning neural network known as Variational Autoencoder
(VAE), as the solution to this problem. We show that, when used in an unsupervised manner, VAE
outperforms the well-known clustering technique DBSCAN.Moreover, this paper shows that higher recall can
be achieved using the semi-supervised one class learning of VAE, which uses only the normal data to train the
model. Additionally, we show that one class learning of VAE outperforms semi-supervised one class SVM
when training data con
sist of only a very small amount of anomalous samples. When a tree based ensemble
technique is adopted for feature selection, the obtained results evidently demonstrate that the performance of
the VAE is highly positively correlated with the selected feature set.
(More)