Compact and Interpretable Transformer for Vehicle Yaw Rate Forecasting

Accurate and interpretable estimation of vehicle yaw rate is critical for stability control, trajectory prediction, and advanced driver-assistance systems (ADAS). While physical gyroscopes are widely used, they are susceptible to drift, failure, and cost constraints, motivating virtual sensing approaches based on standard onboard signals. This paper proposes a compact, encoder-only Transformer model for short-horizon yaw-rate forecasting using commonly available vehicle signals, including wheel speeds, steering angle, and inertial measurements. The model employs sinusoidal positional encoding and a lightweight self-attention mechanism, achieving high forecasting accuracy with a low parameter count. The approach is trained and evaluated on real-world driving data. Experimental results demonstrate accurate prediction up to 400 ms ahead, with R 2 = 0.991 and RMSE = 1.13 ∘ S − 1 . The achieved error level is within a practically relevant range for short-horizon vehicle-dynamics estimation. Attention analysis reveals a strong emphasis on recent inputs, consistent with short-term vehicle dynamics, while gradient-based and permutation-based analyses identify steering angle as the dominant explanatory feature. Robust performance is retained even when using reduced sets of input signals. These findings demonstrate that compact attention-based models provide an accurate, efficient, and interpretable alternative for virtual yaw-rate sensing, supporting their potential use in resource-constrained automotive virtual-sensing applications.