This study proposes an integrated vehicle state estimation model by fusing lateral velocity and roll/pitch angle estimation models using deep neural networks. Each estimation logic is designed to ensure robustness and meet performance requirements bas...
This study proposes an integrated vehicle state estimation model by fusing lateral velocity and roll/pitch angle estimation models using deep neural networks. Each estimation logic is designed to ensure robustness and meet performance requirements based on the measurements from a 6-DOF sensor. A novel deep neural network-based approach is introduced for model fusion.
The lateral velocity estimation integrates a kinematic model, which maintains high accuracy but suffers from integration error accumulation, with a dynamic model, which avoids such errors but exhibits reduced accuracy under non-linear tire dynamics. Similarly, the roll/pitch angle estimation combines a dynamic model, which suffers from integration error accumulation, with a static model, which is unaffected by integration errors but is less accurate under non-linear tire behavior.
State observer structures and Kalman filter frameworks are utilized for model fusion, replacing measurement updates in existing logics with outputs from alternative models. Appropriate fusion weights are dynamically determined using a deep neural network, trained with vehicle dynamic states as inputs and model weights as outputs. To address the lack of ground truth for weights, the loss function is designed to include the entire estimation process, comparing predicted states with measured states and incorporating errors over consecutive predictions to capture pure prediction characteristics.
The proposed methodology is implemented using MATLAB/Simulink and Python’s PyTorch library and validated through real vehicle tests. Training was conducted with 14 scenarios, totaling approximately 310 seconds, while validation employed five scenarios. Results demonstrate that the deep neural network-based fusion method, particularly State Observer Method 3, significantly outperforms conventional expert-designed lookup table-based methods in estimation accuracy across diverse driving scenarios.