【作者】 刘鹏；

【导师】 鲍跃全；

【作者基本信息】 哈尔滨工业大学 ， 力学， 2020， 硕士

【摘要】 高速铁路建设对现代交通方式的发展发挥着重大作用,为有效控制路基沉降、节省土地资源、保证行车顺利,高铁桥梁中大多以桥代路。对于高铁桥梁,桥梁布跨在方案设计过程中是一个关键问题。高铁桥梁一般采用预制标准跨,传统的人工设计方式很容易产生碎跨,而且对墩高、工程量的考虑并不充分,尤其是针对长距离路段,设计人员很难同时将上百跨桥梁合理布置。因此,迫切需要研究采用人工智能的布跨设计代替传统人工方式,以提高效率、节省时间、节约成本。本文研究基于强化学习的高铁桥梁布跨设计方法。主要研究内容为:研究了高铁桥梁布跨基本原则并建立了布跨数据库。针对高铁桥梁布跨问题,概括了桥梁布跨基本原则以及曲线桥梁布置位置计算,然后利用C#语言进行CAD二次开发提取地形地物等桥梁布跨所需信息建立相应数据库。提出了一种基于强化学习中Q-learning算法的高铁桥梁布跨设计方法。在进行算法设计时,将将高铁线路离散化实现连续无限状态数到离散有限状态数的转变,避免因状态数目过多导致在Q表值计算时计算内存爆炸,保障了Q表值的正常更新计算;动作设计是根据梁库中可以选用的标准简支梁和连续梁类型进行设计,动作数目等于桥梁种类数目。根据桥梁布跨要求的优先等级对布跨条件进行权重配置,设计奖赏函数;基通过大量数值计算,分析参数对算法效果的影响,确定最终参数配置。通过实际案例,将Q-learning算法布跨方案与传统设计方案进行对比,发现新算法所得方案不仅效率快,而且降低了桥墩工程量。但是由于Q-learning算法自身一些缺点,该算法有可能陷入局部最优,导致部分区域桥墩设计侵入道路红线无法满足跨越要求。提出了基于模拟退火改进Q-learning算法的高铁桥梁布跨设计方法。针对Qlearning算法的缺点,设计出一种基于模拟退火Q-learning(SA-Q)的高铁桥梁布跨算法,并根据强化学习的探索利用关系,提出了一种新的根据Q值变化自适应变化的退火策略,使得算法前期进行充分的探索,保证布跨结果的多样性;同时利用模拟退火策略的突跳性,避免陷入局部最优解。通过两个实验算例对比传统设计方法、Q-learning和SA-Q三种方法的布跨方案,表明基于SA-Q算法在学习效率、收敛速度相比于Q-learning提升35%;与传统设计相比节省大量时间,所需工程量有所降低,并且能保证布跨合理性。更多还原

【Abstract】 The construction of high-speed railway plays an important role in the development of modern transportation mode.In order to effectively control subgrade settlement,save land resources,and ensure smooth driving,most road in high-speed railway are replaced by bridges.For high-speed railway bridges,the bridge span is a critical problem in the design process.High-speed railway bridges generally adopt prefabricated standard spans,the traditional manual design method is prone to produce broken spans.The pier height and the amount of work are also not fully considered,especially for long-distance sections,it is difficult for designers to arrange the poets with hundreds of spans reasonably at the same time.Therefore,it is urgent to study the use of intelligent cloth span design instead of the traditional manual way to improve efficiency,save time and cost.This paper studies the design method of high-speed railway bridge span based on reinforcement learning algorithm.The main contents are:The basic principles of high-speed railway bridge span arrangement are studied and the database of span arrangement is established.First,for the problem of high-speed railway bridge span,the basic principles of bridge span and the calculation of curving bridge layout position are investigated.Then,the secondary development of CAD using C# language is carried out to extract the information needed for the bridge span,such as terrain and surface features,and the corresponding database is established.A method for designing the span of high-speed railway bridge based on Q-learning algorithm of reinforcement learning is proposed.In the algorithm design,the high-speed railway line will be discretized to realize the discrete finite state number from the continuous infinite state,to avoid the excessive number of states resulting in the explosion of the calculation memory during the calculation of the Q table value,and ensure the normal update calculation of the Q table value.The action is designed according to the types of standard simply-supported beam and continuous beam that can be selected in the beam library.The number of actions is equal to the number of beams.According to the priority levels of bridge span requirements,the weights of different bridge span conditions are determined,and the reward function is designed.Based on a large number of numerical calculations,the influences of parameters on the algorithm accuracy are analyzed,and the final parameters are determined.By implementation of practical cases,the results show that the scheme obtained by the Q-learning is not only fast in efficiency,but also reduces bridge pier quantities compared with the traditional design scheme.However,due to some shortcomings of the Q-learning algorithm,the algorithm may fall into local optimization,resulting in the design of bridge piers in some areas invading the red line of the road and failing to meet the intersection requirements.A design method of high-speed railway bridge span based on Q-learning algorithm and simulated annealing.To overcome the shortcomings of Q-learning algorithm,a highspeed railway bridge spanning algorithm based on simulated annealing Q-learning(SAQ)is designed.According to the exploration and utilization relationship of reinforcement learning,a new annealing strategy based on the change of Q value is proposed,which enables the algorithm to make full exploration in the early stage and ensure the diversity of spanning results.At the same time,the jumping property of simulated annealing strategy is used to avoid local optimal solutions.By comparing the traditional design method,Q-learning method and SA-Q method,the results of two experimental examples show that the learning efficiency and convergence speed based on SA-Q algorithm are 35% higher than that of Q-learning.Compared with the traditional design,the proposed method can save time cost,require less work,and guarantee the rationality of the layout.更多还原