Hybrid Networks for Real-Time Vehicle Dynamic State Estimation: Simulation and Experimentation

Obtaining real-time and accurate estimates of tire-road friction coefficient (TRFC) and sideslip angle is crucial for effectively implementing active safety control algorithms in vehicles. To collectively acquire these two key parameters, we propose a hybrid network architecture comprising a set of lightweight Inception V3 road surface classification networks to output predictive information about the future TRFC range. After processing, using a time synchronization strategy, the continuous TRFC value and the sideslip angle values are regressed through two deep neural networks with multiple long short-term memory layers. Additionally, to ensure that the performance of the hybrid network is consistent across the simulation and real domains, a mixed training scheme combining simulated and real-vehicle data is proposed. In this scheme, multidimensional vehicle active excitation curves are designed to adequately stimulate the dynamic characteristics of vehicles. Finally, extensive simulations and real-vehicle tests are conducted to validate the effectiveness of the hybrid network. The results indicate that the estimation accuracy of the proposed hybrid network is higher than those of the baseline methods, reaching 93.94%. More importantly, the inference time of the hybrid network on in-vehicle control units is only 0.3 ms, which fulfills the real-time requirement.