Logo des Repositoriums

Auflistung nach Schlagwort "Time Series"

1 - 4 von 4
  • Textdokument
    An architecture for detecting infrastructure anomalies at Germany’s Federal Employment Agency
    (INFORMATIK 2021, 2021) Ludsteck, Johannes; Sultanow, Eldar; Chircu, Alina; Herget, Gebhard; Seßler, Matthias
    The data centers of Germany’s Federal Employment Agency (FEA) provide an information technology (IT) infrastructure that is critical for both external stakeholders and internal processes. For FEA and many other organizations like it, it is essential that any IT infrastructure anomalies - deviations from normal behavior - are detected and their underlying causes are understood and, if appropriate, addressed. In this paper we develop a solution that can help increase the availability of an IT landscape such as FEA’s that is characterized by increasing technical complexity and increasing relevance of its applications. The solution detects IT service anomalies based on IT service access logs analyzed with time series methods. The solution also provides visualizations to support further analyses.
  • Konferenzbeitrag
    Benchmarking Univariate Time Series Classifiers
    (Datenbanksysteme für Business, Technologie und Web (BTW 2017), 2017) Schäfer, Patrick; Leser, Ulf
    Time series are a collection of values sequentially recorded over time. Nowadays, sensors for recording time series are omnipresent as RFID chips, wearables, smart homes, or event-based systems. Time series classification aims at predicting a class label for a time series whose label is unknown. Therefore, a classifier has to train a model using labeled samples. Classification time is a key challenge given new applications like event-based monitoring, real-time decision or streaming systems. This paper is the first benchmark that compares 12 state of the art time series classifiers based on prediction and classification times. We observed that most of the state-of-the-art classifiers require extensive train and classification times, and might not be applicable for these new applications.
  • Textdokument
    Infrastructure anomaly detection: A cloud-native architecture at Germany’s Federal Employment Agency
    (INFORMATIK 2022, 2022) Herget,Gebhard; Sultanow,Eldar; Chircu,Alina; Ludsteck,Johannes; Hammer,Sebastian; Koch,Christian; Reuter,Willy; Seßler,Matthias
    In prior research we explored the use of time series analysis methods to detect one class of information technology (IT) infrastructure anomalies - Distributed Denial of Service (DDoS) attacks. The results of this prior work were a mathematical model and a prototype implementation that were concretely trialed and operated in the data centers of Germany's Federal Employment Agency (FEA). With this paper, we go one step further and generalize as well as optimize the mathematical model and create higher performance and scalability for an updated prototype through targeted use of cloud technologies. The starting point of our generalization is the Exponential Smoothing (E-S) approach, which underlies, for example, the well-known Holt-Winters method. This method is used to predict univariate time series. To detect anomalies (such as DDoS attacks) in infrastructure data, we extend the E-S approach to enable it to forecast multivariate time series. In this optimization of our method and our prototype, we take an exploratory, agile approach. Furthermore, we present a cloud-native architecture stack which we pilot in Azure.
  • Konferenzbeitrag
    Visualization and Machine Learning for Data Center Management
    (INFORMATIK 2019: 50 Jahre Gesellschaft für Informatik – Informatik für Gesellschaft (Workshop-Beiträge), 2019) Chircu, Alina; Sultanow, Eldar; Baum, David; Koch, Christian; Seßler, Matthias
    In this paper, we present a novel tool for data center management that incorporates data visualization and machine learning capabilities. We developed the tool in the context of an action design research project conducted at a large government agency in Germany, which hosts three highly available data centers containing more than 10,000 servers. We derived the requirements for the tool from qualitative interviews with agency employees who are familiar with monitoring the data center infrastructure as well as from a review of existing data center and other large infrastructure monitoring solutions. We implemented a web-based 3D prototype for the tool as an Angular 6 application running on Node.js, and evaluated it with the same employees. Most participants preferred the new tool, which provided a significantly better option and enabled visualization of historical data for all server instances at the same time, as well as real-time charts. Planned improvements will take advantage of the full potential of machine learning for time series forecasting.