An automation system for vehicle driveability evaluation using machine learning

Abstract

The drivability is one of the important aspects of vehicle dynamic performances. To ensure quality of the drivability performance, comprehensive screening evaluation is necessary by controlling both complicated driver operation and vehicle behavior. However, it is not straightforward for engineers to handle all combinations of test patterns with limited resources while considering complex control logic and program carefully. This paper proposes a novel automated drivability screening and exploration system by introducing Bayesian active learning (BAL). The proposed system is composed of two key elements such as automated evaluation sub-system and screening and exploring sub-system. In the automated evaluation sub-system, driver model is used to mimic various operation patters of expert drivers and corresponded drivability performance is evaluated automatically using Power-Train Virtual and Real Simulator (PT-VRS). The automated exploring sub-system is used to screen severe condition for drivability and explore feasible region of design space described by control parameters. An effective sampling strategy is introduced based on BAL. In the BAL, Gaussian Process models of drivability performance evaluated by PT-VARS is trained. Based on the posterior distributions of trained Gaussian Processes models, acquisition functions are evaluated and maximized to generate new sampling points. To show effectiveness of the proposed system, an example is demonstrated.