Attention-enhanced physics-informed long short-term memory for roll angle prediction of intelligent vehicle

The variation of roll angle in a multi-axle intelligent wheeled vehicle is a critical factor for the vehicle safety control system. However, classical dynamic models exhibit significant errors under non-steady-state conditions and are computationally complex due to their high degrees of freedom. Meanwhile, data-driven deep learning networks often struggle with poor recognition performance under data loss or zero-sample conditions. To improve roll angle recognition accuracy across various operating conditions, this study proposes an attention improved physics-informed long short-term memory (Attention-PI-LSTM). The proposed deep learning neural network also shows better performance of value prediction with sensor-noise or data loss, even with zero-sampling portion training data, the predict function of the Attention-PI-LSTM still maintains an acceptable accuracy level. The algorithm is validated through simulation using real-world data collected from a 4-axle intelligent wheeled vehicle. Firstly, a vehicle roll angle dynamics model considering axle load transfer during steering was established. Secondly, the physics-informed long short-term memory (PI-LSTM) network is developed with a physics-informed loss function derived from the dynamics model. Finally, adopted attention mechanism to the PI-LSTM algorithm to assign weights to sequential input data. Experiment results show that the proposed Attention-PI-LSTM algorithm improves the coefficient of determination (R²) by 27.4% compared to the regular deep learning neural networks. It also maintains recognition accuracy of 0.651 and 0.694 under data loss and zero-sample conditions, respectively.