Longitudinal speed estimation in vehicles and road condition identification using Artificial Neural Networks

The longitudinal speed in vehicles is a fundamental dynamic parameter that is widely used to understand the behaviour of the vehicle: it is employed in order to control the dynamics of the vehicle and for this reason is one of the main inputs for Anti-Blocking System (ABS), Electronic Stability Program (ESP) and other intelligent devices, that are fundamental to guarantee safe driving conditions. Thus, the correct and precise estimation of the longitudinal speed is fundamental for the correct working conditions of these active control systems. The vehicle speed estimation is generally computed using only data coming from the speedometer mounted on-board, which represents the velocity computed as the mean speed of the vehicle’s wheels. This method is not sufficient, because of the possibility to produce an error in the speed estimation is really high. Actually, it is common that in some driving conditions the vehicle can lock one of the four wheels or vice versa one of them can start spinning. During these occurrences, the behaviour of the longitudinal speed of a single wheel can corrupt the vehicle speed estimation, causing a wrong input for the safety systems of the vehicle. Thus, there is the need to improve our estimation capability. The state of the art gives us different tools that can be used in order to improve the estimation of the longitudinal velocity, such as cameras, GPS-based systems, Extended Kalman Filters (EKF) or Artificial Neural Networks (ANNs). The last solution is investigated in the present work. The use of ANNs for vehicle speed estimation represents a new method, which exploits Machine Learning (ML) techniques. ANNs have been chosen to perform the vehicle speed estimation, since they have a lower cost of implementation and, at the same time, can guarantee a highly reliable and accurate results. The present thesis work consists of two different and interconnected stages. The first stage is the design of a Nonlinear Autoregressive Neural Network with exogenous input (NARX) for the estimation of the longitudinal speed of the vehicle. This is performed using inputs parameters computed by means of sensors already implemented on the vehicle such as accelerometers (lateral and longitudinal), yaw rate sensor, steering sensors and wheel speed sensors. Then, the target for the estimation process is provided by the vehicle speed signal, measured by means of an optical sensor. The estimation has been carried out for dry, wet and icy road conditions. The second stage is the design of a pattern recognition neural network, that must classify in real time the road condition. Thus, receiving this information, the Electronic Control Unit (ECU) can select the most accurate estimation among the three designed in the previous step for dry, wet and icy conditions. The investigated system results to be very accurate, proving the reduction in the estimation error with respect to other systems commonly adopted for the vehicle speed estimation.