Robust Transformer-Enhanced Tire Force Estimation: Combining Model-Based Observer and Deep Learning

This article proposes a novel hybrid framework for estimating longitudinal tire forces by combining a model-based observer and a Transformer neural network architecture. A key advantage of the proposed approach is its independence from wheel torque measurements, which are commonly used in vehicle force estimation methods. The proposed framework first employs a model-based observer using the Pacejka tire model to generate preliminary estimates of longitudinal tire forces. These estimates are then refined by a neural network featuring an encoder-only Transformer architecture augmented with dilated convolutions, which effectively captures long-term temporal dependencies in sensor data while compensating for model inaccuracies. In addition, our hybrid approach is able to estimate the longitudinal velocity and road friction coefficient, which are crucial for observer’s preliminary force estimate in extreme road and driving conditions. Experimental validation on an electric Equinox vehicle under diverse driving scenarios—including high-slip maneuvers on dry, wet, and icy roads—demonstrates the framework’s superior performance compared to model-based and neural network-only approaches. The proposed method achieves robust estimation accuracy even under sensor faults (e.g., biased acceleration signals) and harsh driving conditions, indicating its potential for enhancing vehicle safety systems, such as traction and stability control.