Safety-critical vehicle control systems require both accurate state estimation and reliable uncertainty quantification, particularly for vehicle sideslip angle estimation. In production vehicles, sensor suites are inherently heterogeneous and distributed, with inertial measurement units, wheel speed sensors, and optionally wheel force transducers possibly operating at different native frequencies and exhibiting asynchronous dropouts and rate degradation. However, most learning-based estimators assume idealised synchronous sampling, often relying on resampling strategies such as zero-order hold, which remove explicit information about sensor staleness and temporal irregularity.

This work investigates whether neural estimators can maintain both accuracy and calibrated uncertainty under both synchronous and asynchronous sensor rate degradation, using an extensive experimental vehicle dynamics dataset. To address the brittleness of learned temporal representations, a targeted data augmentation strategy simulating asynchronous dropouts using a Bernoulli dropout is proposed. This strategy is evaluated across three data-driven architectures that differ in how deeply time is embedded: a discrete-time recurrent network operating on fixed-grid resampled inputs, a time-augmented recurrent network explicitly including per-channel staleness features, and a closed-form continuous-time model that embeds inter-arrival time directly into its hidden-state dynamics. These are evaluated against a model-based unscented Kalman filter and a proposed continuous-discrete hybrid estimator, in which the neural network provides pseudo-measurements and uncertainty estimates to the filter.

The estimation pipelines are evaluated across extensive synchronous and asynchronous sensor rate degradation sweeps. Results reveal a clear hierarchy driven by how temporal information is encoded and its interaction with data augmentation. The impact of augmentation varies significantly with the depth of the temporal embedding. The time-augmented network is highly sensitive to distribution shifts and critically requires training on augmented data to maintain functional performance. The zero-order hold approach achieves strong mean accuracy but obscures temporal irregularity from its uncertainty mechanism and lacks native support for operating on event-driven timelines. The continuous-time model provides the most inherently balanced behaviour, combining competitive accuracy with robust continuous-time updates. However, all purely data-driven models exhibit increasing statistical overconfidence under severe degradation, with unbounded neural estimators prone to non-physical errors on out-of-distribution inputs.

The proposed hybrid framework addresses this limitation by combining neural pseudo-measurements with physically consistent filtering. The unscented Kalman filter provides a structured, physics-based fallback in the presence of observability loss, while the continuous-time network contributes high-fidelity predictions and learned uncertainty estimates. This fusion significantly reduces worst-case errors under moderate and severe sensor rate degradation (10 Hz and 5 Hz), decreasing the maximum estimation error beyond either standalone component, albeit with a trade-off in nominal accuracy.

Overall, the results demonstrate that while continuous-time neural models improve robustness to irregular sampling, safe deployment in real-world automotive systems requires explicit coupling with model-based constraints to ensure bounded worst-case behaviour under sensor sparsity. ... Read more