Deep Learning-Based Estimation of the Unknown Road Profile and State Variables for the Vehicle Suspension System

Abstract

The vehicle suspension control unit serves as a critical component to the vehicle system, as it ensures the steering stability and sound ride quality of the vehicle. To effectively realize control strategies, it is essential to foreknowledge the road profile and the suspension system's internal state variables. While the mentioned variables are not practically measurable using commercial sensors, it is necessary to estimate the desired variables by utilizing observer systems. Conventional means have mainly employed modelbased approaches, in which model uncertainties and high computational cost pose limitations for practical implementation. Herein, we propose a data-driven deep learning method as an alternative because no explicit physical modeling is required, and evaluation is computationally cheap. We first propose a novel encoderdecoder structured recurrent neural network model with a two-phase attention mechanism to estimate the unknown road profile and four state variables of the vehicle suspension system. Based on a simulated data set, we assess the proposed model's qualitative and quantitative results and demonstrate that our model can achieve highly accurate estimation results with fast computation time. Besides, we validate our black-box model's reliability by comparing its interpretation with the suspension system's actual physical characteristics. Furthermore, we compare the proposed model with existing baseline methods, and the results show that our proposed deep learning model significantly outperforms the baseline. Lastly, we experiment with our network's autoregressive capability and demonstrate the feasibility of estimating a sequence of future values, which has not been presented in previous works.INDEX TERMS Deep learning, long short term memory (LSTM), observer systems, recurrent neural networks (RNNs), vehicle suspension control.

Using Neural Networks by Modelling Semi-Active Shock Absorber

“…In this regard, a certain support can be provided by using artificial intelligence (AI) tools that can be confirmed with results of some recent studies, e.g. [6], [7], [8], [9]. The presented work will demonstrate the use of AI technique based on the reinforcement learning (RL).…”

“…The correlating dimensionless values that quantify the model quality are shown in Table II and consist of the Mean Squared Error MSE (7), the Mean Absolute Error MAE (8) and the R 2 -Value (9). Note: Although the R 2 value is generally not a suitable metric for evaluating a non-linear regression, it can provide information on the comparability of the results, therefore it was decided to include this parameter for validation purposes.…”

A study on a vehicle semi-active suspension control system based on road elevation identification

“…In Reference [18], a method for estimating the road height using the inverse model of the suspension system and the dynamic response of the suspension system was proposed. In Reference [19], a Kalman filter and neural network were jointly used to estimate road height, and its effectiveness depended largely on the quality of data training. The road height estimation method based on dynamic responses is limited by the vehicle state acquisition error, and is only suitable for outputting road grades; this method cannot achieve real-time output.…”

Simultaneous Estimation of Unknown Road Roughness Input and Tire Normal Forces Based on a Long Short-Term Memory Model

“…A previous study used a fuzzy model along with a static nonlinear autoregressive exogenous (NARX) model to successfully estimate the road roughness using simulated data [12], while a further study validated the performance of the NARX model with field-test results [13]. To estimate the unknown road roughness and four system states, Kim et al proposed a new encoder-decoder structured recurrent neural network (RNN) model with a twophase attention mechanism to better characterize the dynamic behavior of suspension systems [14]. Although the vehicle normal force of a road on a tire at the contact patch can be used in many vehicular control applications (e.g., active roll control to prevent the rollover), there were few types of research because the measurements of tire normal force are challenging.…”

“…Because previous studies [19][20][21][22][23] are associated with physical modelbased simultaneous estimation, physical model inaccuracy leads to low estimation performance. Recently, Kim et al estimated the road roughness and vehicle state based on model-free prediction such as a deep-learning-based observer system [14]. The sequence data calculation input to the offline pre-trained deep-learning model consists of only the calculated nonlinear activation function, resulting in a small computational burden compared with the Kalman filter.…”