Anomaly Detection and Structural Analysis in Industrial Production Environments

Abstract

—Detecting anomalous behavior can be of critical importance in an industrial application context. While modern production sites feature sophisticated alarm management systems, they mostly react to single events. Due to the large number and various types of data sources a unified approach for anomaly detection is not always feasible. One prominent type of data are log entries of alarm messages. They allow a higher level of abstraction compared to raw sensor readings. In an industrial production scenario, we utilize sequential alarm data for anomaly detection and analysis, based on first-order Markov chain models. We outline hypothesis-driven and description-oriented modeling options. Furthermore, we provide an interactive dashboard for exploring and visualization of the results.