Estimating a vehicle's lateral velocity (or sideslip angle) and roll/pitch angles is essential for precise vehicle stability control and autonomous driving. This study develops two lateral velocity estimation models and two roll/pitch angle estimation...
Estimating a vehicle's lateral velocity (or sideslip angle) and roll/pitch angles is essential for precise vehicle stability control and autonomous driving. This study develops two lateral velocity estimation models and two roll/pitch angle estimation models, then proposes an integrated state estimator that fuses them using deep neural networks.
For lateral velocity estimation, a kinematic model provides high instantaneous accuracy based on sensor data but suffers from cumulative error, while a bicycle model offers stability through dynamics-based calculations but shows weakness during transient maneuvers. For roll/pitch angle estimation, a dynamic model excels in dynamic scenarios while a static model performs better in static conditions. Since each model pair compensates for the other's weaknesses, proper fusion can deliver improved estimation performance across diverse driving conditions. This research uses AI to automatically determine fusion ratios, overcoming the limitations of manual expert tuning.
The challenge of not having ground truth for fusion ratios in supervised learning was addressed by adopting the physics-informed neural network (PINN) concept. By incorporating estimator equations into the loss function, the AI learns fusion ratios through estimation accuracy rather than memorizing data. This approach induces physics-based learning and ensures interpretability. Knowledge distillation and transfer learning were applied to ensure production viability. Knowledge distillation enabled development of lightweight models for low-cost MCUs, while transfer learning demonstrated practical applicability by adapting models trained on one vehicle type to another.
The proposed AI-based integrated state estimator was validated through approximately 5,000 seconds of varied real-vehicle driving scenarios, confirming excellent estimation performance.