Contribution
Tyre–Road Noise Prediction Using Multi-Task Learning and LOGO Cross-Validation
* Presenting author
Abstract:
This study introduces a comprehensive data-driven framework for the joint prediction of On-Board Sound Intensity (OBSI) and roadside noise levels through a multi-output regression approach. The objective is to model tyre–road acoustic behavior across various operating and environmental conditions. The dataset includes measurements from three different tyre types, multiple vehicles, and a wide range of speeds, accelerations, and tire pressures. Input features consist of vehicle dynamics, tyre and road temperature, and vibration signals. To enhance robustness and generalization, a multiscale feature extraction strategy is adopted, and performance is evaluated using Leave-One-Group-Out (LOGO) cross-validation, where each tyre–road condition is treated as an unseen test scenario. Hyperparameter tuning is applied to further optimize the best-performing model. The analysis investigates model accuracy across Lmax, narrowband (400–5000 Hz), and full-frequency ranges. Comparative results show that the proposed multiscale multi-output model consistently achieves lower RMSE and higher R² values than single-task regressors. Dominant features include vehicle speed and road temperature, while vibration signals capture transient variations. The framework provides a robust and physics-consistent basis for tyre–road noise prediction and supports data-driven optimization of tyre and pavement design.