Unsupervised anomaly detection in time-series

Audibert, Julien

Anomaly detection in multivariate time series is a major issue in many
fields. The growing complexity of systems and the explosion of the
quantity of data have made its automation essential. Methods based
on Deep Learning have shown good results in terms of detection but
do not meet the industrial requirements due to their long training and
limited robustness. To meet the industrial needs, this thesis proposes a
new unsupervised method for anomaly detection in multivariate time series
called USAD based on an auto-encoder architecture and adversarial
training. This method meets the requirements of robustness and speed
of training of the industrial world while achieving state-of-the-art performance
in terms of detection. However, deep neural network methods
suffer from a limitation in their ability to extract features from the data
since they only rely on local information. Thus, in order to improve the
performance of these methods, this thesis presents a feature engineering
strategy that introduces non-local information. This strategy increases
the performance of neural network based approaches without increasing
the training time. Given the good performance of deep learning methods
for anomaly detection in multivariate time series in recent years,
researchers have neglected all other methods in their benchmark, causing
the complexity of the proposed methods to explode in the current
publication. This lack of comparison with more conventional methods in
the literature does not allow to assert that the progress reported in the
benchmarks is not illusory and that this increasing complexity is necessary.
To address this issue, this thesis proposes a comparison of sixteen
methods for anomaly detection in multivariate time series grouped into
three categories: Conventional methods, machine-learning methods and
approaches based on deep neural networks. This study shows that there
is no evidence that deep neural networks are a necessity to address this

Data Science
Eurecom Ref:
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in and is available at :
See also:

PERMALINK : https://www.eurecom.fr/publication/6637