Improving Machine Self-Diagnosis with an Instance-Based Selector for Real-Time Anomaly Detection Algorithms

Abstract

The diffusion of smart sensor technology in production enables real-time monitoring of production conditions. Machine self-diagnosis shall serve the analysis of these conditions by differentiating expected data from anomalies. Several algorithms have been developed in practice and academia to detect anomalies in real-time and to support machine self-diagnosis, so that counteractions can be taken. However, due to the algorithms' different functionalities, they yield different results when applied to the same data. Our research aims to leverage complementary potentials among these algorithms. To this end, we use a design science research approach to design and prototypically implement a real-time anomaly detection algorithm selector to support decision making regarding machine self-diagnosis. The selector decides in real-time for each sensor-emitted data point, which algorithm yields the most reliable result in terms of anomaly detection. We evaluate functionality and feasibility with two real-world case studies. The evaluation shows that the algorithm selector may outperform single algorithms and that it is applicable in practice. © 2022, Springer Nature Switzerland AG.