.png){kind=icon}
.png){kind=icon}
Return
Bayesian optimization and early stopping improved a deep learning model for autonomous vehicle roll angle estimation
.png){kind=icon}
PRE
AI
X
G
邹
C
H
X
DOI：10.1016/j.displa.2026.103437.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To meet the demand for motion-state estimation in autonomous wheeled vehicles under high-mobility off-road conditions, while mitigating the slow computation of dynamic models and the fluctuations introduced by sensor measurements, this study proposed a vehicle roll angle identification model based on the Bayesian optimization and attention improved physics-informed long short-term memory algorithm. The proposed method is validated through simulation using in-vehicle data collected from a four-axle wheeled transport platform. To enhance the generalization capability and computational accuracy of the dynamic model under non-steady-state conditions, a vehicle roll angle dynamics model incorporating steering-induced axle load transfer is developed. Based on this model, a roll angle identification network is constructed using a physics-informed neural network, in which the loss function is constrained by the proposed dynamic model. This design strengthens the ability of the network to handle small-sample and zero-sample conditions. Subsequently, Bayesian optimization is employed to tune the network hyperparameters, and an early-stopping mechanism is introduced to alleviate over-fitting during the training process. Simulation results show that the proposed Bayesian-optimization physics-informed neural network achieves a coefficient of determination exceeding 0.85 for vehicle roll angle estimation, outperforming classical algorithms. In zero-sample scenarios, the model maintains robust performance, achieving a coefficient of determination greater than 0.65.
Keywords:
Rollangle
Dynamic model
Physics-informedneural network
Bayesian optimization
Earlystopping
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
Cited Papers