Vehicle Condition Monitoring: Anomalous Vibration Detection with Accelerometers
Bitte benutzen Sie diese Kennung, um auf die Ressource zu verweisen:
https://doi.org/10.48693/523
https://doi.org/10.48693/523
Titel: | Vehicle Condition Monitoring: Anomalous Vibration Detection with Accelerometers |
Autor(en): | Schmidt, Jonas Frederic |
ORCID des Autors: | https://orcid.org/0000-0002-2946-8120 |
Erstgutachter: | Prof. Dr. Kai-Uwe Kühnberger |
Zweitgutachter: | Prof. Dr. Martina Juhnke-Kubitzke Prof. Dr. Julius Schöning |
Zusammenfassung: | Alongside electrification, autonomous driving is the primary factor of the ongoing transformation in the automotive industry. The success of this technology is closely linked to its safety and the trust users have in it. To enhance the safety of future autonomous vehicles, they must not only be able to perceive their external environment but also be capable of monitoring the condition of their internal parts, components, and functions for faults or failures. This task is performed by condition monitoring systems (CMSs), which extend autonomy from mere driving to the level of vehicle diagnostics. This work deals with computationally efficient and advanced anomaly detection methods for vibration-based CMSs focusing on automotive applications. In individual publications, the detection of low tire pressure, gearbox damage, and loose wheel bolts as anomalies were investigated using vibration data recorded with accelerometers. The proposed anomaly detection algorithms using features from the time, frequency, time-frequency, and graph domain were able to detect the anomalies reliably. To enhance the accuracy of the anomaly detection algorithms, domain knowledge and a careful selection of feature extraction hyperparameters were combined. The evaluations showed that small neural network architectures executable on microcontrollers and classical anomaly detection algorithms are sufficient to detect anomalies in vibration data accurately. Computational efficiency is decisive in guaranteeing low-latency predictions of CMSs on vehicle microcontrollers with limited computation capacity. For this purpose, an optimized algorithm for faster computation of horizontal visibility graphs that works efficiently on streamed data was introduced. This algorithm was applied for feature extraction in low tire pressure detection and is particularly suitable for processing data with a high sample rate, such as that from accelerometers. Furthermore, a method was developed and patented for faster inference of convolutional neural networks on microcontrollers that process streamed data with overlapping successive inputs. This is typically the case when dealing with spectrogram representations from streamed vibration or acoustic data. The method allows larger neural networks to be computed with low latency on microcontrollers so that the driver is quickly warned of condition changes. |
URL: | https://doi.org/10.48693/523 https://osnadocs.ub.uni-osnabrueck.de/handle/ds-2024022910985 |
Schlagworte: | Condition Monitoring; Vehicles; Anomaly Detection; Machine Learning; Accelerometer; Vibration |
Erscheinungsdatum: | 29-Feb-2024 |
Lizenzbezeichnung: | Attribution 3.0 Germany |
URL der Lizenz: | http://creativecommons.org/licenses/by/3.0/de/ |
Publikationstyp: | Dissertation oder Habilitation [doctoralThesis] |
Enthalten in den Sammlungen: | FB08 - E-Dissertationen |
Dateien zu dieser Ressource:
Datei | Beschreibung | Größe | Format | |
---|---|---|---|---|
thesis_schmidt.pdf | Präsentationsformat | 3,3 MB | Adobe PDF | thesis_schmidt.pdf Öffnen/Anzeigen |
Diese Ressource wurde unter folgender Copyright-Bestimmung veröffentlicht: Lizenz von Creative Commons