多维动态耦合特征下铁路主要技术标准深度学习优选

Railway main technology standards (MTSs) optimization is one of most important prerequisite decision-making in railway life cycle. Essentially, multiple environmental factors (MEFs) and MTSs construct a multidimensional, dynamic and coupling system. The key point to MTSs decision-making is to find the action mechanism and potential rules between MEFs and MTSs, furthermore to create mappings from MEFs to MTSs. Traditional methods usually suppose formulations formats firstly, and then regress the parameters according to a limited number of samples. These methods cannot comprehensively and accurately disclose the potential rules between MEFs and MTSs. In this research, a new idea in which model-centric evolving to data-centric is proposed as well as the deep learning approach based on big data is presented to replace the traditional statistical regression. First, big data for railway alignment cases are generated by combining the selected real cases with intelligently virtual cases. Then extracting, transforming and loading methods for cases data are designed. Furthermore, the high-relevant dimensions and suitable granularities for railway alignment cases data are selected out using association rule mining. On the basis of it, a new deep network model considering the multidimensional, dynamic and coupling features of the MTSs decision-making system is designed and a deep network training algorithm combining supervised with unsupervised learning is presented. In addition, to improve the computing efficiency in deep learning, a collaborative and parallel computing approach for multiple CPUs-GPUs in heterogeneous cluster is studied. Ultimately, the potential rules between MEFs and MTSs are recognized by deep learning big data of railway alignments. Breaking the idea and approach limitation of traditional methods recognizing complex laws through presetting models, the contribution of this project is trying to thoroughly solve the fundamental and theoretical problem for comprehensive optimization of railway MTSs.

主要技术标准优选是铁路全生命周期最重要的先决技术决策之一，其本质是在多维环境因素与主要技术标准构成的多维动态耦合系统中，探求输入端至输出端的作用机理与潜在规律并建立映射。传统方法需预先构造待求规律的模型表达式，再通过有限样本统计回归出模型参数，无法准确全面认知其中的规律。本项目提出以模型为中心进化到以数据为中心的思想，从有限样本统计回归演化至大数据深度学习的方法：基于实例采集与智能仿真生成线路案例大数据，提出案例大数据的自动抽取转换装载方法，通过关联规则挖掘建立适宜的维度与粒度模型；构建符合多维动态耦合特征的深度学习网络模型，提出有-无监督相结合的训练方法，并研制异构集群多CPU—GPU协同并行学习算法，从线路大数据中认知出多维环境因素—主要技术标准的作用机理与潜在规律。本课题力图突破传统模型认知复杂规律时的思路与方法局限，为主要技术标准综合优选解决基础理论问题。

主要技术标准优选是铁路全生命周期最重要的先决技术决策之一，本项目基于大数据深度学习方法对铁路主要技术标准综合优选问题开展研究：（1）提出了一种三维距离变换与分步粒子群优化相结合的线路优化方法，可智能生成优质的仿真铁路线路案例数据，采用“支持度-置信度-提升度”的Apriori关联规则挖掘算法，从案例数据中筛选出与标准优选强相关的环境因素，将强相关环境因素表征为多通道图像及数值型数据，并与主要技术标准数据一同构成实例+仿真的线路案例大数据样本库；（2）提出了“选区划分→单选区投票→选票汇总”选举投票策略，实现了兼顾局部与全局利益平衡的主要技术标准综合优选。参照人工决策经验分别设计了并行、串行及混合多任务深度神经网络，并开展对比实验，最终选定表现最佳的并行网络作为主要技术标准优选深度学习模型；（3）设计了基于OpenMP的并行计算架构（CPU并行），并提出基于CUDA并行计算平台的深度神经网络模型并行训练方法（CPU-GPU并行），实现了仿真案例的高效生成及深度神经网络模型的高效训练；（4）基于本研究提出的理论和方法开发的“铁路主要技术标准综合优选”原型系统已在中铁二院工程集团有限公司试用，为广湛铁路、兴泉铁路等铁路项目提供了技术标准决策建议，建立了从线路环境到技术标准的映射关系，可根据既有铁路线路环境为新建铁路主要技术标准决策提供科学、合理的决策依据，提高决策效率、提升决策质量。项目执行期间，发表科技论文22篇，其中SCI期刊检索14篇，EI期刊检索3篇，EI会议检索3篇，CSCD核心2篇；授权国家发明专利2项；获软件著作权1项；培养博士研究生4人，其中已毕业博士研究生2名，硕士研究生20名，已毕业6名；获行业协会科技进步特等奖1项、省级科技进步二等奖1项。