Deep Neural Network Based Data-Driven Virtual Sensor in Vehicle Semi-Active Suspension Real-Time Control

Abstract

This research presents a data-driven Neural Network (NN)-based Virtual Sensor (VS) that estimates vehicles’ Unsprung Mass (UM) vertical velocity in real-time. UM vertical velocity is an input parameter used to control a vehicle’s semi-active suspension. The extensive simulation-based dataset covering 95 scenarios was created and used to obtain training, validation and testing data for Deep Neural Network (DNN). The simulations have been performed with an experimentally validated full vehicle model using software for advanced vehicle dynamics simulation. VS was developed and tested, taking into account the Root Mean Square (RMS) of Sprung Mass (SM) acceleration as a comfort metric. The RMS was calculated for two cases: using actual UM velocity and estimations from the VS as input to the suspension controller. The comparison shows that RMS change is less than the difference threshold that vehicle occupants could perceive. The achieved result indicates the great potential of using the proposed VS in place of the physical sensor in vehicles.

Review of Integrated Chassis Control Techniques for Automated Ground Vehicles

“…Other measurable parameters were the steering angle and angular velocities of four wheels [ 182 ]. In vehicles with controllable suspensions, additional displacement and (or) acceleration sensors were installed on UMs [ 145 ], and in some industrial applications, additional acceleration sensors were installed above absorbers on sprung mass (SM) [ 183 ].…”

“…In the last few years, so-called data-driven virtual sensors for vehicle state estimation have been proposed for application in the automotive industry. These data-driven approaches have the potential to replace model-based methods [ 145 , 191 , 192 ]. Data-driven estimators employ artificial neural networks (ANN), which necessitate datasets for training, testing, and validation, but do not rely on mathematical models and lead to higher accuracy.…”

“…Sensor data pre-processing is now essential [ 110 ]. By combining sensors using kinematic and dynamic models in Kalman Filters and other adaptive filtering methods, the data from several sensors can be fused using the sensor fusion approach to achieve robustness, stability, and accuracy [ 144 , 145 , 146 ]. On the other hand, the data pre-processing step can involve virtual sensors that add additional measurements that cannot be provided by physical sensors or that are not available or costly to integrate.…”

Development and Evaluation of Artificial Neural Networks for Real-World Data-Driven Virtual Sensors in Vehicle Suspension

“…The study [22] used the KF for suspension state estimation. However, due to the nonlinearity of vehicle components, it can be challenging to use mathematical equations to create an accurate VS model [23]. On the other hand, the data-driven VS relies solely on recorded data obtained from observation of system operation [24].…”

“…Data-driven VS developed for vehicle suspensions has been investigated in previous studies [15], [23], [24], [25]. In [24], the authors presented a data-driven approach based on deep learning (DL) for estimating the road profile height and state variables of vertical displacement and velocity of vehicle UMs using onboard sensors.…”