A method for predicting the skidding and slipping of tracked vehicles based on physics-informed machine learning

The dynamic characteristics of tracked vehicles are closely related to the slip and skid phenomena of their tracks. This paper proposes a slip prediction model for tracked vehicles based on physics-informed machine learning. The model preserves the overall structure of the classical physics-based theoretical model while introducing a longitudinal velocity correction factor kv and a yaw rate correction factor kw to compensate for prediction biases caused by variable track slip. A six-dimensional feature vector containing vehicle states such as longitudinal velocity, yaw rate, and total driving force is constructed. Gaussian Process Regression (GPR) is employed to learn the nonlinear mapping relationship from the vehicle state to the optimal correction factors based on test data. Experimental results under five typical conditions (including straight-line driving, in-place turning, medium-radius turning, and random driving) on both snow-covered and unpaved roads demonstrate that the proposed fusion model significantly improves the prediction accuracy of key states compared to the traditional theoretical model. Specifically, the prediction error for longitudinal velocity is reduced by approximately 70% on average, the error for yaw rate is reduced by about 80% on average in steering conditions, and the slip rate prediction error is reduced by an average of 65%. This research provides an effective solution for accurate slip estimation and state prediction of tracked vehicles under complex operating conditions.