Addressing Integration Challenges in Vehicle Roll and Pitch Estimation Using Neural Networks

This study introduces a novel multioutput estimator that utilizes longitudinal and lateral accelerations to simultaneously estimate a vehicle's roll and pitch angles. Training data were obtained by performing laps around the Nurburgring track in both directions and a slalom maneuver using CarMaker software. Ornstein-Uhlenbeck noise was applied to the accelerations to simulate real-world signals. Subsequently, the neural network was tested during a double lane change maneuver, with Ornstein-Uhlenbeck noise added to the input signals. A comparison between the neural network's estimations and those obtained through the integration of roll rate and pitch rate signals demonstrates the advantages of the neural network approach. By employing algebraic calculations instead of signal integration, the neural network provides consistent estimations over time. Evaluation using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) confirm the accuracy of the developed estimator.