Physics-Informed Neural Network for Adaptive Road Roughness Recognition

Precisely recognizing road roughness can provide crucial prior information for active suspension control of intelligent vehicles, which in turn enhances vehicle handling stability and ride comfort. However, the existing road roughness recognition methods based on neural networks suffer from issues of high data demand and poor internal interpretability. To address these challenges, a novel method integrating physics-informed neural network (PINN) and dynamic equations of linear suspension systems is presented for adaptive road roughness recognition in this paper. The PINN architecture is designed in accordance with the suspension dynamic equations, and the forward propagation process is inherently provided with strong physical interpretability. In addition, the loss function of our proposed PINN model is also incorporated with a dynamic equation term, thus further enhancing the physical constraints imposed on the network learning. To validate our proposed PINN-based method, comprehensive simulation experiments with random road model have been conducted. It is shown that our presented PINN-based method is characterized by a low demand for training data and exhibits the strong capability of adaptive recognition, outperforming the CNN and LSTM methods for road roughness recognition.