Bienefeld, Christoph (2024)
A Contribution on the Transferability of Data-Driven Models for Bearing Fault Diagnosis.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00027475
Ph.D. Thesis, Primary publication, Publisher's Version
Text
Dissertation_Bienefeld.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (12MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | A Contribution on the Transferability of Data-Driven Models for Bearing Fault Diagnosis | ||||
Language: | English | ||||
Referees: | Kirchner, Prof. Dr. Eckhard ; Klingauf, Prof. Dr. Uwe | ||||
Date: | 7 June 2024 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | 116, XXXIII Seiten | ||||
Date of oral examination: | 7 May 2024 | ||||
DOI: | 10.26083/tuprints-00027475 | ||||
Abstract: | In the automotive industry, the electrification of powertrains is steadily advancing. Additionally, in the context of increasingly automated driving functions, the requirements for safety against unforeseen failures of the vehicle subsystems are growing. Since a significant proportion of failures of electric machines are caused by rolling bearing faults, both their prevention and the early detection of emerging rolling bearing faults are important aspects of current research. A promising approach to meet the increasing safety requirements is the implementation of data-driven fault diagnosis. The foundation of such fault diagnosis is the application of condition monitoring based on at least one suitable measurement quantity. A particularly common variable for monitoring the condition of electric machines is vibration, which can be measured by mounting accelerometers at the structure of the machine. Based on the acquired data, Machine Learning (ML) methods enable fault diagnosis in an automated manner. To build models based on supervised learning, training data is required, which has to comprehensively represent the faults to be detected. Collecting this data involves considerable effort, which is why data-driven approaches at the current state of research are only commercially viable for a limited proportion of applications. The profitability would be significantly increased if the models trained using data from one machine type could be transferred to other machine types. Obstacles to this transferability are the changed properties of the sensor signals when switching to a different machine type. Given this motivation, the present thesis examines the idea that the structural dynamic properties, which can vary between different machine types, can have a decisive influence on the measured vibrations. To investigate the influence of differences in structural dynamics on the accuracy of the vibration-based, data-driven fault diagnosis, an experimental data set is acquired. For this purpose, artificially damaged rolling bearings are installed in an electric machine. During the experiments, the vibration acceleration is measured simultaneously at several sensor positions. Based on the acquired data set, various ML algorithms and feature generation methods are investigated to optimize the prediction accuracy of the fault diagnosis models. Using the resulting configurations, the transferability of the vibration-based fault diagnosis models to different sensor positions is evaluated, which represents the transfer to a different structural dynamic behavior. Based on this, two novel approaches are introduced in the present work, both of which aim to improve the previously assessed transferability of the models by incorporating domain knowledge from structural dynamics. The prediction accuracies achieved with these novel approaches show major improvements in terms of transferability. Accordingly, this work demonstrates novel methods to improve the transferability of data-driven fault diagnosis models between different systems with respect to varying structural dynamic properties. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-274757 | ||||
Classification DDC: | 600 Technology, medicine, applied sciences > 620 Engineering and machine engineering | ||||
Divisions: | 16 Department of Mechanical Engineering > Institute for Product Development and Machine Elements (pmd) | ||||
Date Deposited: | 07 Jun 2024 12:02 | ||||
Last Modified: | 10 Jun 2024 05:26 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/27475 | ||||
PPN: | 518988015 | ||||
Export: |
View Item |