Intelligent tire force estimation with dynamic domain adaptation for varying wear states

This work introduces a comprehensive framework for reliable tire–force estimation under real-world wear conditions, combining a newly assembled wireless acceleration dataset with a dynamic domain-adaptation model. The dataset, collected across multiple wear depths and load levels under longitudinal free rolling conditions, does not consider slip, cornering, or tire-pressure variations. It reveals systematic shifts in signal distributions induced by progressive abrasion. To address these distributional changes, we design a feature-alignment network that jointly aligns mean and variance statistics through a learnable weighting mechanism, while a residual-attention encoder captures the non-stationary dynamics of worn tires. Our approach learns a domain-invariant representation that adapts continuously to varying wear states, enabling accurate cross-condition predictions. In cross-wear evaluations, the proposed model cuts relative error by nearly forty percent versus traditional regressors. These results demonstrate that coupling targeted data collection with adaptive alignment strategies can substantially enhance the generalization of force-prediction models across evolving mechanical conditions, offering a powerful tool for tire health monitoring.