Wing shape estimation with Extended Kalman filtering and KalmanNet neural network of a flexible wing aircraft

Bence Zsombor Hadlaczky, Noémi Friedman, Béla Takarics, Balint Vanek
Proceedings of The 5th Annual Learning for Dynamics and Control Conference, PMLR 211:1429-1440, 2023.

Abstract

The dynamic behaviour, stability, and the effects of the aerodynamic drag of a large-wingspan aircraft are mainly influenced by the structural flexibility and shape of the wings during flight. Therefore, utilizing a wing shape controller that minimizes the effects of drag can greatly improve the behaviour and fuel consumption of the aircraft. However, such a controller requires the measurement of the dynamics of the wing, more precisely, the modal coordinates which describe the structural and dynamic changes of the wing. For estimating the modal coordinates and reconstructing the wing shape a state observer is necessary because the direct and accurate measurement of these states is not feasible. It is demonstrated in this paper that machine learning-based methods can approach the accuracy of traditional model-based Kalman filtering in wing shape estimation. First, the model-based method Extended Kalman Filtering (EKF) is presented, using a Linear Parameter Varying (LPV) system model. Second, we present a machine learning-based approach based on the new KalmanNet architecture with two different recurrent neural network configurations: one with linear layers and one with one-dimensional convolutional layers. The results are evaluated on the T-Flex aerial demonstrator aircraft and compared using the LPV-based EKF as a reference. It is shown that the learning-based approach provides comparable results to the model-based method while using fewer design parameters.

Cite this Paper

BibTeX
@InProceedings{pmlr-v211-hadlaczky23a, title = {Wing shape estimation with Extended Kalman filtering and KalmanNet neural network of a flexible wing aircraft}, author = {Hadlaczky, Bence Zsombor and Friedman, No\'emi and Takarics, B\'ela and Vanek, Balint}, booktitle = {Proceedings of The 5th Annual Learning for Dynamics and Control Conference}, pages = {1429--1440}, year = {2023}, editor = {Matni, Nikolai and Morari, Manfred and Pappas, George J.}, volume = {211}, series = {Proceedings of Machine Learning Research}, month = {15--16 Jun}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v211/hadlaczky23a/hadlaczky23a.pdf}, url = {https://proceedings.mlr.press/v211/hadlaczky23a.html}, abstract = {The dynamic behaviour, stability, and the effects of the aerodynamic drag of a large-wingspan aircraft are mainly influenced by the structural flexibility and shape of the wings during flight. Therefore, utilizing a wing shape controller that minimizes the effects of drag can greatly improve the behaviour and fuel consumption of the aircraft. However, such a controller requires the measurement of the dynamics of the wing, more precisely, the modal coordinates which describe the structural and dynamic changes of the wing. For estimating the modal coordinates and reconstructing the wing shape a state observer is necessary because the direct and accurate measurement of these states is not feasible. It is demonstrated in this paper that machine learning-based methods can approach the accuracy of traditional model-based Kalman filtering in wing shape estimation. First, the model-based method Extended Kalman Filtering (EKF) is presented, using a Linear Parameter Varying (LPV) system model. Second, we present a machine learning-based approach based on the new KalmanNet architecture with two different recurrent neural network configurations: one with linear layers and one with one-dimensional convolutional layers. The results are evaluated on the T-Flex aerial demonstrator aircraft and compared using the LPV-based EKF as a reference. It is shown that the learning-based approach provides comparable results to the model-based method while using fewer design parameters.} }
Endnote
%0 Conference Paper %T Wing shape estimation with Extended Kalman filtering and KalmanNet neural network of a flexible wing aircraft %A Bence Zsombor Hadlaczky %A Noémi Friedman %A Béla Takarics %A Balint Vanek %B Proceedings of The 5th Annual Learning for Dynamics and Control Conference %C Proceedings of Machine Learning Research %D 2023 %E Nikolai Matni %E Manfred Morari %E George J. Pappas %F pmlr-v211-hadlaczky23a %I PMLR %P 1429--1440 %U https://proceedings.mlr.press/v211/hadlaczky23a.html %V 211 %X The dynamic behaviour, stability, and the effects of the aerodynamic drag of a large-wingspan aircraft are mainly influenced by the structural flexibility and shape of the wings during flight. Therefore, utilizing a wing shape controller that minimizes the effects of drag can greatly improve the behaviour and fuel consumption of the aircraft. However, such a controller requires the measurement of the dynamics of the wing, more precisely, the modal coordinates which describe the structural and dynamic changes of the wing. For estimating the modal coordinates and reconstructing the wing shape a state observer is necessary because the direct and accurate measurement of these states is not feasible. It is demonstrated in this paper that machine learning-based methods can approach the accuracy of traditional model-based Kalman filtering in wing shape estimation. First, the model-based method Extended Kalman Filtering (EKF) is presented, using a Linear Parameter Varying (LPV) system model. Second, we present a machine learning-based approach based on the new KalmanNet architecture with two different recurrent neural network configurations: one with linear layers and one with one-dimensional convolutional layers. The results are evaluated on the T-Flex aerial demonstrator aircraft and compared using the LPV-based EKF as a reference. It is shown that the learning-based approach provides comparable results to the model-based method while using fewer design parameters.
APA
Hadlaczky, B.Z., Friedman, N., Takarics, B. & Vanek, B.. (2023). Wing shape estimation with Extended Kalman filtering and KalmanNet neural network of a flexible wing aircraft. Proceedings of The 5th Annual Learning for Dynamics and Control Conference, in Proceedings of Machine Learning Research 211:1429-1440 Available from https://proceedings.mlr.press/v211/hadlaczky23a.html.

Related Material