【作者】 曾强；

【导师】 陈德旺；

【作者基本信息】 北京交通大学 ， 控制工程（专业学位）， 2015， 硕士

【摘要】 城市轨道交通的快速发展,使得相关部门越来越重视列车的运行效率、乘客安全等方面。近年来,由于地铁普遍加装车站屏蔽门,对于列车停车精度提出了新的要求,更加精确的车站停车已成为未来研究的重点方向。然而列车的停车精确性又受到了列车制动前的初始速度、测速不准、制动系统特性等因素的影响,因此对于列车停车过程的研究很有必要。首先,本文根据列车停车时的制动特性,建立相应制动模型,再根据制动模型和列车基本阻力特性搭建了车站停车仿真平台,能够用来模拟实际的城市轨道交通车站停车系统,执行相关的运算和控制列车制动等任务,并且起到车地通信中传输设备的作用。然后,通过不同的计算智能方法计算出5种不同值的制动策略。创新性地将制动策略作为增强学习的动作集合,选择动作的过程即选择制动策略,并且将列车位置和速度信息作为增强学习的状态集合,停车误差作为立即回报,将停车过程与增强学习良好结合。以上能将不同制动策略应用到一次停车过程当中,使得停车时制动率的选择组合是不同的,得到的停车结果也是不同的,我们通过随机最优选则或者Q学习从中找出较好的结果即可。有效利用每个制动策略的优势,也是本文的一大创新之处。其次,给出了两种实用性算法来寻找最优的停车策略：随机最优选择算法(OSSa)、模糊隶属Q学习算法(FQLa),存储算法所得到的仿真结果和最优停车策略。之后提出三种制动率融合生成方法来计算任意初始速度下的停车误差,并使之满足误差要求,它们是：线性制动率融合生成法(LRm)、拟合制动率融合生成法(FRm)、插值制动率融合生成法(IRm)。通过选取一定的评估指标作为评估算法的标准。最后,通过进行大量的仿真实验测试,对算法的结果分析得出：两种算法都能够通过重复的训练和学习过程逐步减小停车误差,经过算法训练后的停车误差都能保持在±30cm以内,这满足了城市轨道交通的精确停车要求,能够保证精确停车可靠性在99.5%以上。FQLa比OSSa有更快的收敛速度、更小的停车误差、更好的稳定性和智能性。对于三种制动率融合生成方法,都能够求出每个未训练速度的停车策略,并保证停车误差在±30cm以内,其中LRm得到的误差最大值和平均误差是三种方法中最小的。更多还原

【Abstract】 With the fast development of the urban rail transit, the efficiency and safety of subway systems are paid more and more attention. In recent years, precise train station parking (PTSP) has been acting as an important part in urban rail transit, since platform screen doors (PSD) are installed in most new established stations. However, there are some factors that may impair the accuracy of train station parking, such as the variable initial velocities, inaccurate positioning data and uncertain braking characteristics. Therefore, it is necessary to study the problem of PTSP for a subway system.First, we establish a train braking model, as well as an urban rail transit parking simulation platform according to the braking characteristics and basic resistances. The platform contains the train braking model and it can simulate the process of train station parking. In addition, we can test different parking strategies on this platform.Second, we propose five different braking policies by using soft computing method, which are treated as action vectors of the reinforcement learning (RL). The state vector of RL is defined as the location and velocity of the train while the reward of RL is defined by parking error. We first propose the idea to combine the train parking process with RL. Thus, the PTSP based on RL can be described as follows:we need to find out the optimal policy sequences that correspond to the minimum parking error by using stochastic selection, i.e., Q-learning algorithm.Third, two algorithms are developed to find out the optimal braking policy in urban rail transit, which are optimal stochastic selection algorithm (OSSa) and fuzzy function based Q-learning algorithm (FQLa). After that, we propose three braking rate fusion methods to approximate the parking errors under different initial velocities, which haven’t appeared in OSSa or FQLa. The three braking rate fusion methods are linear braking rate fusion method (LRm), fitting braking rate fusion method (FRm) and interpolating braking rate fusion method (IRm). We also evaluate the two algorithms and three methods by setting different performance evaluation indices.Finally, numerical experiments are developed to test the effectiveness of the algorithms on the parking simulation platform. The results indicate that, both OSSa and FQLa can minimize the parking errors by training and learning repeatedly on different initial velocities, which can keep the parking error within±30cm that meet the parking requirement of urban rail transit. In addition, the two algorithms can also ensure accurate parking probability in the required range of more than99.5%. Furthermore, the FQLa performs better than OSSa. By using braking rate fusion methods, we can calculate the parking errors at any initial velocity without training and keep the parking error within±30cm. The LRm performs the best among the three methods at average parking errors and maximum errors.更多还原