Vehicle sideslip angle estimation using deep ensemble-based adaptive Kalman filter. (October 2020)

• Record Type: Journal Article
• Authors: Kim, Dongchan; Min, Kyushik; Kim, Hayoung; Huh, Kunsoo
• Is Part Of: Mechanical systems and signal processing. Volume 144(2020)
• Publication Date: 2020-10
• Subjects: Sideslip angle estimation -- Vehicle dynamics -- Kalman filters -- Deep ensemble -- Uncertainty; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.106862
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A novel sideslip angle estimation scheme is proposed combining DNN and EKF/UKF. The initial estimate and its uncertainty are obtained from deep ensemble. The outputs of deep ensemble are utilized in EKF/UKF for the final estimate. Combining the DNN with EKF/UKF improves the estimation performance significantly. The proposed method is validated under various road surface conditions. Abstract: This paper presents a novel sideslip angle estimation scheme combining deep neural network and nonlinear Kalman filters. The deep neural network contains a recurrent neural network with long short-term memory which is effective for analyzing sequential sensor data and deep ensemble which is used for robustness of the estimation and acquisition of the uncertainty of the estimate. The deep neural network is trained using input sets which consist of on-board sensor measurements (yawrate, velocity, steering wheel angle and lateral acceleration) and provides sideslip angle estimate and its uncertainty. The estimate of deep neural network is used as a new measure in the nonlinear Kalman filters and its uncertainty is used to make an adaptive measurement covariance matrix. The algorithm is verified through both simulation and experiment. The performance with the proposed method is analyzed in terms of the root mean squared error (RMSE) and maximum error (ME) as compared to the case where nonlinear Kalman filter or deep neural network is utilized individually. The results demonstrate the effectiveness of the proposed solution.