Abstract
Detecting early enough the anomalous behavior of technical systems facilitates cost savings thanks to avoiding system downtimes, guiding maintenance, or improving performance. The novel framework proposed in this paper processes event streams originating from system monitoring for anomaly detection purposes. Therefore, statistical models characterizing the normal behavior of the monitored system are learned from the events. Instead of having one coarse normal model for all operational states, the proposed framework contains a mechanism for automatically detecting different conditions of the system allowing for fine-tuned models for every condition. The performance of the framework is demonstrated by means of a real-world application, where the log files of a large-scale printing machine are analyzed for anomalies.
Zusammenfassung
Die rechtzeitige Erkennung eines abweichenden Verhaltens von technischen Systemen ermöglicht Kosteneinsparungen, da Ausfälle vermieden, Wartungen zielgerichtet durchgeführt oder Leistungsparameter gesteigert werden können. Das in diesem Papier vorgestellte neuartige Framework nutzt Ereignisströme einer Prozessüberwachung zwecks der Erkennung von Anomalien. Hierzu werden statistische Modelle, welche das Normverhalten des überwachten Systems widerspiegeln, aus den Ereignisdaten gelernt. Anstelle eines einzelnen, groben Normalmodells für alle Betriebszustände, nutzt das vorgeschlagene Framework einen Mechanismus zur automatischen Erkennung verschiedener Zustände, um so für jeden Zustand ein passendes Modell bereit zu stellen. Die Leistungsfähigkeit des Frameworks wird anhand einer Realweltanwendung demonstriert, bei welcher die Logdateien einer großformatigen Druckmaschine nach Anomalien durchsucht werden.
About the author
Dr. Marco Huber was born 1980 in Germany. He received his diploma degree and Ph.D. degree in computer science from the Karlsruhe Institute of Technology (KIT), Germany, in 2006 and 2009, respectively. From 2009 until 2011 he was with Fraunhofer IOSB, Karlsruhe, Germany, where he was leading a research group on computer vision and information fusion. From 2011 until 2015, he was Senior Researcher with AGT International, Darmstadt, Germany. Since 2015 he is Senior Consultant and Data Scientist with USU Software AG, Karlsruhe, as well as Adjunct Professor (Privatdozent) with the KIT. His research interests include machine learning, Big Data analytics, non-linear Bayesian estimation, probabilistic planning, and optimization.
USU Software AG, Rüppurrer Str. 1, 76137 Karlsruhe, Germany
Acknowledgement
This work was partially supported by the BMWi project SAKE (Grant No. 01MD15006A).
©2017 Walter de Gruyter Berlin/Boston
