Authors:
Moritz Fehsenfeld
1
;
Johannes Kühn
2
;
Zygimantas Ziaukas
1
and
Hans-Georg Jacob
1
Affiliations:
1
Leibniz University Hannover, Institute of Mechatronic Systems, An der Universität 1, Garbsen, Germany
;
2
Lenze SE, Hameln, Germany
Keyword(s):
Fault Diagnosis, Machine Learning, Industrial Application, Belt Drives, Mechatronics Systems.
Abstract:
Machine learning (ML) has received a lot of attention in solving fault diagnosis (FD) tasks. As a result, more and more advanced machine learning algorithms have been developed to increase accuracy. But the system’s excitation has likewise a high impact on the diagnosis performance and applicability. For this purpose, we describe different industrial application scenarios and the related set trajectory. They are divided into passive FD, where normal operation data serves as the input, and active FD, where an optimized excitation is injected. All scenarios are investigated concerning achievable accuracy and data requirement based on comprehensive measurements. We demonstrate that in active scenarios a high accuracy of 97:6% combined with a small number of measurements are obtained by very basic algorithms like a one-nearest neighbor with Euclidean distance. In passive scenarios, where the FD task is generally harder, the demand for large datasets and more advanced ML methods increases
. In this way, we illustrate how intelligent use of an optimized excitation strategy leads to feasible, reliable, and accurate fault diagnosis with a broad industrial application spectrum.
(More)