Attention-Based LSTM Network for Unknown Road Input Estimation and Sensor Selection for Applications in Vehicle Suspension Control

This study proposes a novel deep learning method of estimating the unknown road input profile for applications in vehicle suspension control. The vehicle suspension control unit serves as a critical component to the vehicle system, ensuring steering stability and sound ride quality. For effective realization of model-based control strategies, having foreknowledge of the road input profile is critical. A problem arises as the road input profile cannot be practically measured, which calls for a need of an observer system to estimate, rather than measure the desired profile. Herein, data-driven approach is considered, and a novel LSTM-based observer with temporal-variable attention mechanism is proposed. Virtual sensor signals are obtained via vehicle simulation in the CarSim environment for training and validating the neural network. After validating the trained models, input sensor selection is conducted by observing variable attention values to select appropriate sensors. Lastly, input optimized model is evaluated, and results are compared with the model-based EKF-UI (Extended Kalman Filter with Unknown Inputs) as the baseline comparison model.