基于信息几何与集成学习的列车转向架故障诊断技术研究

To deal with the lack of generalization ability of fault diagnosis for bogies caused by typical factors, this research combines the Information Geometry and Ensemble Learning, aiming at those typical factors, to study the generalized fault diagnosis for bogies from the data layer, the feature layer and the diagnosis layer..By using information geometry theory, on the data layer, this research employs the information manifold hypothesis to the nonlinear noise reduction and preprocess, so as to solve the problems of nonlinear noise and data corruption; on the feature layer, this research employs the topology analysis to extract features effectively and also reduce the dimensionality..By using ensemble learning structure, firstly, this research employs the information geometry to optimize the machine learning and learning machine generation algorithms; then, ① to deal with the varying work condition, this research employs the depth structure, the selection and ensemble algorithms of learning machines, to enhance the generalization ability of algorithms, so as to diagnose faults under the varying work condition; ② to deal with the lack of training data, this research employs the heterogeneous ensemble learning, combined with Incremental Learning, and update the knowledge continuously, so as to solve the lack of training data and improve the generalization ability of fault diagnosis..Finally, typical bogies are used to verify the proposed generalized fault diagnosis methods. This research will be meaningful and beneficial to the development and application of fault diagnosis technology.

为突破典型因素影响下的列车转向架故障诊断方法泛化能力不足的难题，本项目融合信息几何的理论优势与集成学习的泛化性优势，围绕转向架诊断泛化性制约因素，从数据层、特征层及诊断层开展转向架强泛化故障诊断研究。.在信息几何理论下，数据层上利用信息流形假设，研究非线性降噪及预处理，突破非线性噪声干扰、数据破损污染等问题；特征层上利用流形拓扑结构分析，有效提取故障特征，并对其降维优化。.在集成学习框架下，首先用信息几何自适应优化元学习算法；在此基础上，①针对转向架工况多维波动，揭示基学习器生成机理，探索融合深度结构，通过学习器集成及选择策略，充分挖掘泛化性能，实现工况多维波动下的故障诊断；②针对初始数据匮乏，利用异质学习器集成，结合增量学习，在线持续更新知识，克服初始数据匮乏的不利影响，提升转向架诊断的泛化性能。.最后，以现场实际数据为对象试验验证，形成具有宽泛适用性的强泛化高精度转向架故障诊断方法体系。

本项目针对列车转向架故障诊断技术研究及应用中的强泛化性需求，聚焦由于非线性噪声干扰、数据破损污染、特征提取困难、运行工况波动、初始数据匮乏等典型因素影响下的转向架故障诊断泛化能力不足的难题，深入研究列车转向架故障诊断中的非线性信号的流形几何表征机理及建模分析、流形拓扑特征提取及表征能力优化及面向泛化能力增强的转向架故障诊断算法构建及优化等关键技术，从数据层、特征层、诊断层等全面提升了转向架故障诊断方法的泛化性能，符合“需求牵引，突破瓶颈”属性。.1、针对非线性信号的流形几何表征机理及建模分析的难题，提出了基于邻域分布信息约束的局部投影非线性降噪、基于自适应分解与tile-coding谱编码的信号消噪预处理方法。.2、针对转向架信号的流形拓扑特征提取及表征能力优化的难题，提出了基于流形学习与奇异值分解的转向架特征分析、基于奇异值完全分解的转向架早期故障特征分析、基于子信号趋势分析的转向架早期故障特征提取等方法。.3、针对面向泛化能力增强的转向架故障诊断算法构建及优化的难题，提出了基于信息几何与支持向量机的元故障分类器构建、基于固定字典极限学习机的转向架变工况故障诊断、基于自适应多分类马氏田口系统的转向架轴承变工况故障诊断、基于Fast Kurtogram和深度残差学习的转向架变工况故障诊断、基于图论半监督学习的转向架轴承少样本故障诊断、基于流形对比表示学习的转向架轴承在线更新故障诊断、基于无标度收缩宽度学习的转向架轴承在线更新故障诊断、基于混沌集成在线框架的转向架故障状态估计器构建等方法。.负责人已发表学术论文6篇，其中IEEE Transactions 等高水平SCI/EI期刊论文5篇，EI会议论文1篇；授权专利4项，受理专利2项，软件著作权1项。项目组培养了博士毕业生4人，硕士毕业生4人，成功晋升副教授2人次；参与组织承办了2021中国工程院国际工程科技战略高端论坛“高速铁路安全保障与治理”等国际学术会议7次。.技术成果已经成功应用到CRH380、复兴号等高速列车上，支撑有效缩短了检修周期：定期检修效率提高20%以上；延长了使用寿命：关键设备使用寿命延长10%；提升了列车可用性：可用性提高10%；支撑实现的“高速列车一体化主动安全保障系统”入选国家“十三五”科技创新成就展，并被列为重点展项。