Inhalt des Dokuments
Es gibt keine deutsche Übersetzung dieser Webseite.
Master Thesis: Deep Learning for Anomaly Detection in Time Series
Deep Learning for Anomaly Detection in Time Series
In this thesis, we perform anomaly detection with deep learning in time series data, namely inthe HDFS and Blue Gene/L super computer logs, and also simulated time series. We mainly focused on autoencoder based anomaly detection with deep, denoising, recurrent and variational variants. In order to compare our results to the state of the art, we also evaluated the accuracy of classical machine learning models on the HDFS logs. All our datasets are labeled, thus we did extensive evaluation of the detection performance at several anomaly thresholds. The implementation is written in Python, and we used Keras with a Tensorflow backend to design and train our deep neural networks. Computations with deep neural networks were accelerated by GPUs. We conducted experiments with various anomaly ratios in the training data to test the applicability of our deep learning models in practical settings, where labeled datasets are rarely available.
Supervisor: Tanja Deutsch, Boris Lorbeer
Type: Master Thesis
Duration: 6 months
10587 Berlin, Germany
Phone: +49 30 8353 58811
Fax: +49 30 8353 58409