Details
Originalsprache | Englisch |
---|---|
Qualifikation | Doktor der Ingenieurwissenschaften |
Gradverleihende Hochschule | |
Betreut von |
|
Datum der Verleihung des Grades | 18 Dez. 2024 |
Erscheinungsort | Hannover |
Publikationsstatus | Veröffentlicht - 3 Feb. 2025 |
Abstract
Neural Networks die Genauigkeit von Fahreridentifikationsmodellen signifikant verbessern kann. Die Evaluierung dieser Modelle erfolgte hauptsächlich in experimentellen Umgebungen, wodurch jedoch die Erklärbarkeit der Ergebnisse eingeschränkt wird. Um die Zuverlässigkeit der Modelle sicherzustellen, ist eine umfangreiche, qualitativ hochwertige und vielfältige Datensammlung sowie sorgfältige Validierung erforderlich. In dieser Arbeit umfassen wir den gesamten Prozess der Entwicklung und Implementierung eines Fahreridentifikationssystems. Wir bewerten modernste Deep-Learning-
Netzwerke für die Fahreridentifikation unter Verwendung eines umfangreichen realen Datensatzes, der über einen Zeitraum von zwei Jahren gesammelt wurde. Dieser Datensatz umfasst rohe Fahrdaten von über 709,000 Kilometern mit unterschiedlichen Fahrstilen und 29,000 Fahrstunden. Zusätzlich erweitern wir den Datensatz durch die Integration von Kontextinformationen wie Route, Fahrzeug, Verkehr und Wetterbedingungen. Durch die Einbeziehung von Kontextinformationen stellen wir fest, dass neuronale Netzwerke hohe Genauigkeitsraten erzielen, wenn sie auf denselben Fahrtbedingungen trainiert
und getestet werden. Die Genauigkeit nimmt jedoch ab und die Bedeutung einzelner Signale variiert, wenn auf unterschiedlichen Fahrtbedingungen getestet wird, obwohl bewährte Verfahren wie Stratifikation und Cross-Validierung angewendet wurden. Unsere Ergebnisse betonen die Notwendigkeit, moderne Deep-Learning-Identifikationsmodelle auf umfassenden und vielfältigen Datensätzen zu trainieren und zu evaluieren. Wir bewerten neuronale Netzwerke in Entwicklungs- und Implementierungsphasen und
zeigen, dass die Leistung bestehender Lösungen signifikant abnimmt, wenn sie auf neue Fahrer angewendet werden, die während der Entwicklungsphase des Netzwerks nicht berücksichtigt wurden. Zusätzlich schlagen wir ein Deep-Neural-Network-Modell vor, das die aktuellen state-of-the-art-Lösungen in beiden Phasen übertrifft. Schließlich demonstrieren wir die Möglichkeit, Feinabstimmung zu vermeiden, wenn das trainierte Modell auf neue Fahrer angewendet wird, indem wir die Triplet Loss-Funktion anpassen, um die Fahreridentifikationsproblematik auf Fahrzeugsensordaten anzuwenden.
Zitieren
- Standard
- Harvard
- Apa
- Vancouver
- BibTex
- RIS
Hannover, 2025. 116 S.
Publikation: Qualifikations-/Studienabschlussarbeit › Dissertation
}
TY - BOOK
T1 - Deep learning-based driver identification using vehicular time-series data
AU - Zeng, Li
PY - 2025/2/3
Y1 - 2025/2/3
N2 - As technology progresses and electric vehicles become more prevalent, integrating sensors into vehicles is becoming standard. This offers benefits like improved vehicle performance, enhanced safety, and increased comfort for drivers and passengers. Crucially, this integra- tion allows recognition of driver behaviour, with wide-ranging applications. For instance, driver authentication, crucial for security, ensures only authorised individuals access and operate the vehicle. Additionally, sensors also permit vehicle function customisation based on personal preferences, thus improving the driving experience. Furthermore, they aid in re- fining fleet management processes, enabling effective monitoring and optimisation of vehicle usage. The growing volume of sensor data and advancements in deep learning research have sparked a revolution in various research fields, including driver identification. Studies have shown that utilising deep neural networks can significantly improve the accuracy of driver identificationmodelswhenusingrawdatagatheredfromin-vehiclesensorsviatheController Area Networks (CAN) system. However, these models have primarily been validated in experimental setups using datasets collected from either simulators or real vehicles in a controlledenvironment. Modelsgeneratedbydeepneuralnetworksareoftentreatedasblack boxes, making it difficult to explain and interpret their results. This limited interpretability restricts the reliability of the learned models, which heavily relies on the quantity, quality, and diversity of the training dataset, as well as the validation scenarios. Furthermore, numerous questions regarding the deployment of these models remain unanswered. In this thesis, we encompass the entire end-to-end process of developing and deploying a driver identification system, aiming to address the last-mile question of bringing the driver identification system into production vehicles. We assess state-of-the-art deep learning net- works for driver identification using an extensive real-world dataset collected over a span of two years. This dataset comprises raw driving data encompassing over 709,000 kilometres of diverse driving styles and 29,000 driving hours. Additionally, we enhance the dataset by incorporating contextual information such as route, vehicle, traffic, and weather conditions. With the inclusion of contextual information, we observe that neural networks achieve high levels of accuracy when trained and tested on the same driving conditions. However, accuracy decreases and the importance of individual signals varies when testing on dif- ferent driving conditions, despite implementing best practices like stratification and cross- validation. Our findings underscore the necessity of training and evaluating state-of-the-art deep learning identification models on comprehensive and diverse datasets. Furthermore, we emphasise the importance of evaluating state-of-the-art neural net- works during both the development and deployment phases. Our evaluation shows that the performance of existing solutions drops significantly when applied to new drivers that have not been seen by the networks in the development phase. Additionally, we propose a deep neural network that surpasses current state-of-the-art solutions in both phases. Finally, we demonstrate the feasibility of avoiding fine-tuning when applying the trained model to new drivers by adapting the triplet loss to vehicular sensor data for the driver identification problem.
AB - As technology progresses and electric vehicles become more prevalent, integrating sensors into vehicles is becoming standard. This offers benefits like improved vehicle performance, enhanced safety, and increased comfort for drivers and passengers. Crucially, this integra- tion allows recognition of driver behaviour, with wide-ranging applications. For instance, driver authentication, crucial for security, ensures only authorised individuals access and operate the vehicle. Additionally, sensors also permit vehicle function customisation based on personal preferences, thus improving the driving experience. Furthermore, they aid in re- fining fleet management processes, enabling effective monitoring and optimisation of vehicle usage. The growing volume of sensor data and advancements in deep learning research have sparked a revolution in various research fields, including driver identification. Studies have shown that utilising deep neural networks can significantly improve the accuracy of driver identificationmodelswhenusingrawdatagatheredfromin-vehiclesensorsviatheController Area Networks (CAN) system. However, these models have primarily been validated in experimental setups using datasets collected from either simulators or real vehicles in a controlledenvironment. Modelsgeneratedbydeepneuralnetworksareoftentreatedasblack boxes, making it difficult to explain and interpret their results. This limited interpretability restricts the reliability of the learned models, which heavily relies on the quantity, quality, and diversity of the training dataset, as well as the validation scenarios. Furthermore, numerous questions regarding the deployment of these models remain unanswered. In this thesis, we encompass the entire end-to-end process of developing and deploying a driver identification system, aiming to address the last-mile question of bringing the driver identification system into production vehicles. We assess state-of-the-art deep learning net- works for driver identification using an extensive real-world dataset collected over a span of two years. This dataset comprises raw driving data encompassing over 709,000 kilometres of diverse driving styles and 29,000 driving hours. Additionally, we enhance the dataset by incorporating contextual information such as route, vehicle, traffic, and weather conditions. With the inclusion of contextual information, we observe that neural networks achieve high levels of accuracy when trained and tested on the same driving conditions. However, accuracy decreases and the importance of individual signals varies when testing on dif- ferent driving conditions, despite implementing best practices like stratification and cross- validation. Our findings underscore the necessity of training and evaluating state-of-the-art deep learning identification models on comprehensive and diverse datasets. Furthermore, we emphasise the importance of evaluating state-of-the-art neural net- works during both the development and deployment phases. Our evaluation shows that the performance of existing solutions drops significantly when applied to new drivers that have not been seen by the networks in the development phase. Additionally, we propose a deep neural network that surpasses current state-of-the-art solutions in both phases. Finally, we demonstrate the feasibility of avoiding fine-tuning when applying the trained model to new drivers by adapting the triplet loss to vehicular sensor data for the driver identification problem.
U2 - 10.15488/18489
DO - 10.15488/18489
M3 - Doctoral thesis
CY - Hannover
ER -