For the automated vehicles, the user experience on comfort plays an important role for the market acceptance. Generally, for the experienced drivers who already form some certain driving preferences during the longtime driving, they will feel apparent discomfort if the automated vehicles drive very differently from them. Therefore, it is of great significance for comfort driving if the automated vehicles could learn the driving preferences of the users. Fortunately, we enter the era of traffic big data, from the cyber physical system (CPS) perspective, we almost can get whatever data we need to map human drivers from physical space to cyberspace. In this paper, we build a general driving model based on deep convolutional fuzzy systems (DCFS), and design an online driving preferences learning algorithm based on the high-dimensional on-board data. For the verification of the method, we apply this method to design a personalized lane keeping controller (PLKC) with considering the guaranteed stability. Fifteen volunteers participate in the experiments on the Prescan-based simulation platform, and the results show that the PLKC has the online learning ability to the fixed and the time-varying lateral driving preferences.

对于自动驾驶汽车而言，用户对舒适性的体验对于市场接受度起着重要作用。一般来说，对于在长时间驾驶中已经形成一定驾驶偏好的有经验的驾驶员来说，如果自动驾驶汽车的驾驶方式与他们有很大不同，他们会感到明显的不适。因此，如果自动驾驶汽车能够学习用户的驾驶偏好，对于舒适驾驶具有重要意义。幸运的是，我们进入了交通大数据时代，从网络物理系统（CPS）的角度来看，我们几乎可以获得我们需要的任何数据，将人类驾驶员从物理空间映射到网络空间。在本文中，我们建立了基于深度卷积模糊系统（DCFS）的通用驾驶模型，并设计了一种基于高维车载数据的在线驾驶偏好学习算法。为了验证该方法，我们应用该方法在考虑保证稳定性的情况下设计个性化车道保持控制器（PLKC）。15 名志愿者在基于 Prescan 的模拟平台上参与实验，结果表明 PLKC 具有对固定和时变横向驾驶偏好的在线学习能力。