Towards Design Principles for a Real-Time Anomaly Detection Algorithm Benchmark Suited to Industrie 4.0 Streaming Data

Abstract

The vision of Industrie 4.0 includes the automated reduction of anomalies in flexibly combined production machine groups up to a zero-failure ideal. Algorithmic real-time detection of production anomalies may build the basis for machine self-diagnosis and self-repair during production. Several real-time anomaly detection algorithms appeared in recent years. However, different algorithms applied to the same data may result in contradictory detections. Thus, real-time anomaly detection in Industrie 4.0 machine groups may require a benchmark ranking for algorithms to increase detection results' reliability. This paper makes a qualitative research contribution based on ten expert interviews to find design principles for such a benchmark ranking. The experts were interviewed on three categories, namely timeliness, thresholds and qualitative classification. The study's results can be used as groundwork for a prototypical implementation of a benchmark.

A Prototypical Dashboard for Knowledge-Based Expert Systems used for Real-Time Anomaly Handling in Smart Manufacturing

“…Sensors have become indispensable in digitized production environments as they developed from purely mechanical sensors with a specific field of application and time-delayed data transmission to internet-enabled multi-sensory devices that communicate data in real-time (Schütze et al, 2018) For KBES, sensors mean permanent data input for decision-making (Schütze et al, 2018). Domain knowledge as well as experience of experts are sources used for data analysis and decision making (Stahmann & Rieger, 2022). Historical data and analysis results can also serve as a basis for assessing anomalies in real-time.…”

“…In particular, results from predictive simulations can also be used to concretize expectations of upcoming production runs and identify deviations in real production data. In KBES, these data serve to acquire knowledge that may be used as basis for rules and cases (Beierle & Kern-Isberner, 2019;Stahmann & Rieger, 2022). Furthermore, explanations for anomalies and their elimination or prevention can be given, especially by using empirical knowledge.…”

“…If data exceed these limits, they are always considered anomalies. In their qualitative survey, Stahmann and Rieger (2022) identified thresholds as one of the most important means to detect anomalies in practice. Threshold values are set based on expert knowledge, historical data, or simulations.…”

“…DEs 1.1 and 1.2 show graphs that give a first visual impression of the current sensor data and the currently present anomalies with respect to a categorization of their criticality. The categorization is done from historical data, expert knowledge and previous simulation (Stahmann & Rieger, 2022). The two DEs meet the infrastructural requirements of real-time notification and evaluation on anomalies.…”