基于深度学习的高铁轮轨系统故障诊断研究

The high-speed rail system directly affects the safety and operation quality of high-speed trains. This project carries out research on fault diagnosis of high-speed rail wheel train system based on deep learning. Explore how to establish a wheel-rail intelligent fault diagnosis model from the actual wheel-rail detection data. A fault diagnosis method for high-speed railway wheel and rail based on improved depth residual network was proposed. Aiming at the problem that the network model constructed by low-dimensional wheel-rail data guides is too shallow and too narrow, an improved method of deep residual network is proposed (using asymmetric convolution kernels to reduce model parameter quantities, parallel convolution kernels to broaden the model breadth And so on), and then build an improved deep residual network model that deepens and widens. A high-speed rail wheel-rail fault diagnosis method based on deep capsule network was proposed. An integrated capsule network model including a fault diagnosis module and a data reconstruction module is constructed to clarify the information exchange and route update mechanism of high and low-level capsules, construct and solve the mathematical optimization problem of dynamic routing algorithms, and then provide dynamics between multi-layer capsules. The implementation of the routing algorithm, the construction of data reconstruction module encoding wheel and rail training sample data. The research plan for optimizing the parameters of the wheel rail depth learning model is given. Developed a validation method for the effectiveness of a deep learning model for high-speed railway wheel-rail fault diagnosis. The project research will provide theoretical support for intelligent fault diagnosis of wheel-rail systems.

高铁轮轨系统直接影响高速列车安全性和运行品质。本项目开展基于深度学习的高铁轮轨系统故障诊断研究，探索从实际轮轨检测数据中建立轮轨智能故障诊断模型的方法。提出基于改进深度残差网络的高铁轮轨故障诊断方法。针对由于低维数轮轨数据导构建的网络模型太浅太窄的问题，提出深度残差网络的改进方法（采用不对称卷积核以减少模型参数量、将卷积核并联以拓宽模型广度等），进而构建加深拓宽的改进深度残差网络模型。提出基于深度胶囊网络的高铁轮轨故障诊断方法。构建包含故障诊断模块和数据重构模块的胶囊网络整体模型，阐明高、低层胶囊进行信息交互和路由更新机制，构造并求解动态路由算法的数学优化问题，进而给出多层胶囊之间的动态路由算法的实现方案，构建数据重构模块编码轮轨训练样本数据。给出轮轨深度学习模型参数优化的研究方案。开展面向高铁轮轨故障诊断的深度学习模型的有效性检验方法。课题研究将为实现轮轨系统智能故障诊断提供理论支撑。

高铁轮轨系统直接影响高速列车安全性和运行品质。本项目开展基于深度学习的高铁轮轨系统故障诊断研究，探索从实际轮轨检测数据中建立轮轨智能故障诊断模型的方法。..提出基于改进深度残差网络的高铁轮轨故障诊断方法。针对由于低维数轮轨数据导构建的网络模型太浅太窄的问题，提出深度残差网络的改进方法(采用不对称卷积核以减少模型参数量、将卷积核并联以拓宽模型广度等)，进而构建加深拓宽的改进深度残差网络模型。..提出基于深度胶囊网络的高铁轮轨故障诊断方法。构建包含故障诊断模块和数据重构模块的胶囊网络整体模型，阐明高、低层胶囊进行信息交互和路由更新机制，构造并求解动态路由算法的数学优化问题，进而给出多层胶囊之间的动态路由算法的实现方案，构建数据重构模块编码轮轨训练样本数据。..给出轮轨深度学习模型参数优化的研究方案。开展面向高铁轮轨故障诊断的深度学习模型的有效性检验方法。..开展基于深度残差网络(DRN)的故障诊断研究，可从大量轮轨检测数据中自动提取故障特 征。针对传统深度神经网络训练困难以及梯度消失、维数爆炸等问题，采用快捷连接方式构建网络模型，并利用批归一化(BN)方法进行模型训练。进一步探究了模型层数、BN 技术和残差结构对故障诊断率的影响，通过 t 分布随机邻域嵌入(t-SNE)方 法对网络部分层的输出进行可视化。..课题研究将为实现轮轨系统智能故障诊断提供新的方法和视角。