基于LCD与流形学习的监测数据分析研究

发布时间：2018-01-19 11:45

本文关键词： 走行部故障监测数据局部特征尺度分解集合局部特征尺度分解信息熵流形学习　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机技术与人工智能的发展,特征提取作为完整模式识别系统中的重要环节,已引起越来越多的重视。在面对高复杂度、非平稳信号时,选择合适的特征识别方法是能否挖掘有效信息的关键。由于提取到的高维特征空间中存在着相关性和冗余信息,利用流形学习这种非线性降维的机器学习方法可以达到维数简约的目的,挖掘原始数据的内在本质,实现数据的可视化。本文以列车走行部故障诊断与雷达辐射源信号识别为背景,探讨特征提取与降维在基于监测数据的信号处理领域的应用,并开展了以下研究工作:1.论文采用局部特征尺度分解(LCD)与信息熵结合的特征提取方法,以轴承故障标准数据集为研究对象,对信号数据做LCD分解并提取ISC分量的多种信息熵特征组成故障特征向量,仿真验证了 LCD信息熵特征在故障特征提取分析的有效性和可行性。2.针对LCD分解方法的不足,提出一种利用噪声辅助改进的集合局部特征尺度分解(ELCD)方法,仿真验证改进的算法能有效抑制模态混叠现象并具有高效的算法效率。针对走行部横向减振器部分失效工况数据,提取ELCD分解多种信息熵组成特征向量。由于原始特征向量包含有大量的冗余信息,采用流形学习的LLTSA算法对原始高维向量进行降维,然后运用Fisher比率分别对LLTSA降维前后的特征进行评价。实验结果表明:降维后特征对分类的贡献率更大,即LLTSA降维算法能够在最大程度保留本质特征,采用ELCD和LLTSA相结合的特征分析方法,横向减振器部分故障工况识别率更高。3.针对不同雷达辐射源由于调制方式不同和噪声影响引起瞬时频率变化中统计参数的差异,开展了基于ELCD和流形学习的雷达信号识别方法研究。首先,对辐射源信号做多重相位差分法求时频曲线,并制定调制识别的层次决策分类器模型识别信号的调制类型;然后,对雷达信号进行ELCD分解并提取Renyi熵,将调制识别结果、Renyi熵和PDW参数组成特征向量;最后,采用S-ISOMAP降维处理,对降维后的特征采用SVM分类识别。实验表明:1)所提取的特征能有效描述不同信号的脉内调变规律,总体分类率正确率为98.8%;2)运用S-ISOMAP算法对原始特征向量降维,所得低维特征的分类效果更好。
[Abstract]:With the development of computer technology and artificial intelligence, feature extraction, as an important part of the complete pattern recognition system, has attracted more and more attention in the face of high complexity and non-stationary signals. Choosing the appropriate feature recognition method is the key to mining the effective information because of the correlation and redundancy in the extracted feature space. The machine learning method of nonlinear dimensionality reduction by manifold learning can achieve the goal of dimensionality reduction and mine the intrinsic essence of the original data. Based on the background of train line fault diagnosis and radar emitter signal recognition, this paper discusses the application of feature extraction and dimension reduction in the field of signal processing based on monitoring data. The following research work is carried out: 1. The paper adopts the feature extraction method which combines local feature scale decomposition (LCD) with information entropy, and takes the standard data set of bearing faults as the research object. The signal data is decomposed by LCD and a variety of information entropy features of the ISC component are extracted to form the fault feature vector. Simulation results verify the effectiveness and feasibility of LCD information entropy feature in fault feature extraction and analysis. 2. Aiming at the shortage of LCD decomposition method. A new method of local feature scale decomposition (ELCD) based on noise assisted improvement is proposed in this paper. Simulation results show that the improved algorithm can effectively suppress modal aliasing and has high efficiency. ELCD decomposes a variety of information entropy to form a feature vector. Because the original feature vector contains a lot of redundant information, the LLTSA algorithm of manifold learning is used to reduce the dimension of the original high-dimensional vector. Then the Fisher ratio is used to evaluate the characteristics of LLTSA before and after dimensionality reduction. The experimental results show that the contribution rate of dimensionally reduced features to classification is greater. That is, the LLTSA dimensionality reduction algorithm can retain the essential features to the maximum extent, and adopts the feature analysis method which combines ELCD and LLTSA. For different radar emitter due to different modulation mode and noise influence, the statistical parameters of instantaneous frequency change are different. 3. The radar signal recognition method based on ELCD and manifold learning is studied. Firstly, the time-frequency curve of emitter signal is obtained by multi-phase difference method. A hierarchical decision classifier model for modulation recognition is developed to identify the modulation types of signals. Then, the radar signal is decomposed by ELCD and the Renyi entropy is extracted, and the modulation recognition result is composed of the Renyi entropy and the PDW parameter. Finally, S-ISOMAP is used to reduce the dimension, and the feature after dimension reduction is identified by SVM classification. The experiment shows that the extracted feature can effectively describe the pulse modulation law of different signals. The total classification rate is 98.8%. 2) using S-ISOMAP algorithm to reduce the dimension of the original feature vector, the classification effect of the low dimensional feature is better.
【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】：TP391.4

【相似文献】