基于深度学习和分类集成的高速列车工况识别研究

发布时间：2018-10-08 11:53

【摘要】：中国高速铁路快速发展,目前已成为世界高速铁路的引领者。然而高速列车长时间的高速运行,使得列车走行部性能下降,这为列车的安全运行带来了巨大的隐患。走行部是高速列车的关键组成部分,对保障列车的安全性和乘客的舒适度起到重要作用。通过在列车走行部上安装传感器,采集并分析反映其运行状况的振动信号,是监测列车运营状态的主要技术之一。如何有效的从高速列车监测数据中挖掘出有用的特征信息,并实现典型工况的有效识别具有重要的研究价值。列车振动信号是非平稳、非线性信号,具有特征信息复杂、难辨识等特点。而传统的工况识别方法存在特征提取不完备和识别性能不精确的问题。本文设计了一种多视图特征提取方法,并首次引入分类集成技术,提出了多视图分类集成(Multi-view Classification Ensemble,MV-CE)的列车工况的识别方法。该方法首先提取FFT系数、小波能量、EEMD模糊熵,并对FFT系数进行Fisher比率特征选择,从而得到列车振动信号三个视图的特征。然后利用K最近邻分类器和最小二乘支持向量机分别对三个视图进行初步识别。最后通过分类熵投票策略集成多个分类器的输出结果。通过实验对比说明该方法可以提取出完备的特征,并验证了具有多样性集成模型的有效性。深度信念网络(DeepBeliefNetwork,DBN)可以自动的学习原始数据的特征,为高速列车工况识别的研究开拓了新的思路。结合深度学习与分类集成技术的优点,本文提出了一种DBN层次集成模型对高速列车工况进行识别。首先提取列车振动信号的FFT系数作为模型的可视层输入。利用DBN自动学习信号的层次特征。然后利用每一层特征训练支持向量机、K最近邻、RBF神经网络三种分类器。最后分别采用多数投票法、分类熵投票策略、胜者全取三种集成策略进行集成。实验结果表明,该模型的识别效果高于10种对比方法,并且其性能受网络层数和隐藏层单元数目变化的影响远小于传统DBN模型。列车走行部不同通道的振动信号既存在互补性又存在冗余性。为了充分利用多通道振动信号的互补信息,提出了基于相似度比率的通道筛选方法,并构建了一种多通道深度信念网络模型(Multi-channel Deep Belief Network,MDBN)进行多通道的工况识别。首先提取所有通道振动信号的FFT系数。然后,计算每个通道FFT特征的相似度比率,并选取相似度比率较大的若干通道。最后,构建MDBN模型对所筛选的多通道数据进行特征学习,利用MDBN的共联层实现多通道特征的融合,并进行分类识别。实验结果表明,MDBN的特征提取能力优于DBN模型,并且MDBN的工况识别率高于DBN和DBN层次集成模型。
[Abstract]:With the rapid development of Chinese high-speed railway, it has become the leader of high-speed railway in the world. However, the high speed train running for a long time makes the train running performance decline, which brings a huge hidden trouble for the safe operation of the train. The running part is the key component of the high speed train and plays an important role in ensuring the safety of the train and the comfort of the passengers. It is one of the main techniques to monitor the train running state by installing the sensor on the train running part and collecting and analyzing the vibration signal which reflects the running condition of the train. How to effectively mine useful feature information from high-speed train monitoring data and realize effective recognition of typical working conditions has important research value. Train vibration signal is non-stationary and nonlinear. It has complex characteristic information and is difficult to identify. However, the traditional working condition recognition methods have the problems of incomplete feature extraction and inaccurate recognition performance. In this paper, a multi-view feature extraction method is designed, and the classification integration technique is introduced for the first time, and a multi-view classification integration (Multi-view Classification Ensemble,MV-CE) method for train operating condition identification is proposed. The method firstly extracts FFT coefficient, wavelet energy and EEMD fuzzy entropy, then selects the Fisher ratio feature of FFT coefficient, and obtains the characteristics of three views of train vibration signal. Then K nearest neighbor classifier and least square support vector machine are used to identify the three views. Finally, the output of multiple classifiers is integrated by the classification entropy voting strategy. The experimental results show that the proposed method can extract complete features and verify the effectiveness of the diversity integration model. Deep belief Network (DeepBeliefNetwork,DBN) can automatically learn the characteristics of the original data, which opens up a new idea for the study of high-speed train condition identification. Combined with the advantages of deep learning and classification integration technology, this paper presents a DBN hierarchical integration model to identify the operating conditions of high-speed trains. First, the FFT coefficient of train vibration signal is extracted as the visual layer input of the model. DBN is used to automatically learn the hierarchical features of signals. Then three kinds of classifiers of support vector machine (SVM) nearest neighbor RBF neural network are trained by each layer feature. At last, the majority voting method, the classified entropy voting strategy and the winner integration strategy are adopted respectively. The experimental results show that the recognition effect of this model is higher than that of 10 comparison methods, and its performance is much less affected by the number of network layers and the number of hidden layer units than the traditional DBN model. The vibration signals in different channels of train running are both complementary and redundant. In order to make full use of the complementary information of multi-channel vibration signals, a method of channel selection based on similarity ratio is proposed, and a multi-channel depth belief network model (Multi-channel Deep Belief Network,MDBN) is constructed to identify multi-channel operating conditions. First, the FFT coefficients of all channel vibration signals are extracted. Then, the similarity ratio of FFT features of each channel is calculated, and several channels with high similarity ratio are selected. Finally, the MDBN model is constructed to learn the features of the filtered multi-channel data, and the co-layer of MDBN is used to realize the fusion of multi-channel features, and the classification and recognition are carried out. The experimental results show that the feature extraction ability of MDBN is better than that of DBN model, and the recognition rate of MDBN is higher than that of DBN and DBN hierarchical integration model.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U270.7;TP391.41

【参考文献】