Sensor Fault Detection Using an Extended Kalman Filter and Machine Learning for a Vehicle Dynamics Controller

This paper describes a new sensor fault detection approach for a vehicle dynamics controller. The detection problem is divided into two parts. First, a model-based observer is used to incorporate the knowledge of the system into the fault detection. Next, a data driven classification algorithm based...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society s. 361 - 366
Hlavní autoři: Ossig, Daniel L., Kurzenberger, Kevin, Speidel, Simon A., Henning, Kay-Uwe, Sawodny, Oliver
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 18.10.2020
Témata:
automotive applications
binary classification problem
EKF
Fault detection
fault diagnosis
hybrid fault diagnosis
machine learning
Machine learning algorithms
Mathematical model
Observers
Roads
state estimation
Technological innovation
Vehicle dynamics
wavelet packet transform
ISSN:2577-1647
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:This paper describes a new sensor fault detection approach for a vehicle dynamics controller. The detection problem is divided into two parts. First, a model-based observer is used to incorporate the knowledge of the system into the fault detection. Next, a data driven classification algorithm based on kalman filter performance metrics is used. This machine learning algorithm is trained using real vehicle data and, therefore, able to handle model uncertainties and disturbances inherently. Due to the usage of a nonlinear observer, the fault detection is suitable up to the limits of handling. The presented structure offers the possibility to use the same classification algorithm for different vehicles as the vehicles' behavior is abstracted in the observer. Therefore, the need of extensive training data is reduced. This paper focuses on the development of features and gives a first proof of concept. The developed fault detection is validated with real car measurements.
ISSN:2577-1647
DOI:10.1109/IECON43393.2020.9254448