基于随机邻域嵌入的机械故障特征提取方法

发布时间：2018-07-07 09:32

本文选题：随机邻域嵌入 + Manhattan距离　；参考：《东南大学》2015年硕士论文

【摘要】：特征提取是机械故障诊断的基础,如何有效地获取故障特征信息是故障诊断领域的研究重点和热点。随着数据挖掘技术的不断发展,数据降维算法被引入到故障诊断领域用于信号的特征提取。本文针对故障诊断过程故障数据高维数、非线性化、复杂性等特点,研究基于随机邻域嵌入的机械故障特征提取方法,相关工作如下：针对欧氏距离在高维数据空间中不能提供较大的相对距离差,无法明显体现高维数据对象之间差异性的问题,提出一种基于Manhattan距离的随机邻域嵌入(Manhattan-SNE)算法。采用Manhattan距离衡量高维数据对象之间的相异度,得到高维空间和低维空间数据对象之间相似度的条件概率。UCI数据集和仿真故障信号分类识别验证所提改进算法的有效性。针对随机邻域嵌入算法无法利用现实数据中少量样本标记信息的问题,提出一种基于拉普拉斯正则化度量学习的半监督随机邻域嵌入(semi-supervised SNE, ss-SNE)算法。采用拉普拉斯正则化度量学习对距离矩阵进行半监督学习,利用已标记数据提供的信息,重新刻画数据点之间的距离,从而实现SNE算法的半监督改进。与其他半监督降维算法的对比分析表明所提改进算法的优越性。随机邻域嵌入算法是一种批处理方法,无法获取高维空间到低维嵌入空间的映射函数,由此导致该算法无法对新增数据进行增量式处理。本文构造一种随机邻域嵌入算法的增量形式(增量SNE算法),寻找新增样本点的K近邻,使得新增样本的K近邻的分布形式和K近邻对应的低维映射的分布形式尽可能匹配,实现新增样本的学习。最后,将上述方法应用于实际齿轮箱故障诊断,结果表明上述方法能够有效提高故障诊断精度,验证了算法在实际故障诊断应用中的可行性。
[Abstract]:Feature extraction is the basis of mechanical fault diagnosis. How to obtain fault feature information effectively is the focus and hotspot in fault diagnosis field. With the development of data mining technology, data dimensionality reduction algorithm is introduced to fault diagnosis for feature extraction. Aiming at the characteristics of high dimension, nonlinearity and complexity of fault data in fault diagnosis process, this paper studies a method of mechanical fault feature extraction based on random neighborhood embedding. The related work is as follows: aiming at the problem that Euclidean distance can not provide large relative distance difference in high-dimensional data space and can not manifest the difference between high-dimensional data objects, a Manhattan distance-based random neighborhood embedding algorithm is proposed. Manhattan distance is used to measure the dissimilarity between high-dimensional data objects and low-dimensional spatial data objects. The conditional probability of similarity between high-dimensional and low-dimensional spatial data objects. UCI dataset and simulation fault signal classification are used to verify the effectiveness of the proposed improved algorithm. Aiming at the problem that the random neighborhood embedding algorithm can not make use of a small number of samples in real data, a semi-supervised random neighborhood embedding (semi-supervised SNE-SNE) algorithm based on Laplace regularization metric learning is proposed. The distance matrix is semi-supervised by Laplace regularization metric learning, and the distance between data points is redescribed by using the information provided by marked data, so that the semi-supervised improvement of SNE algorithm is realized. The comparison with other semi-supervised dimensionality reduction algorithms shows the superiority of the improved algorithm. The random neighborhood embedding algorithm is a batch processing method, which can not obtain the mapping function from high-dimensional space to low-dimensional embedded space, which leads to the algorithm being unable to process the new data incrementally. In this paper, we construct an incremental form of random neighborhood embedding algorithm (incremental SNE algorithm) to find the K-nearest neighbor of the new sample point, so that the distribution form of the K-nearest neighbor of the new sample and the distribution form of the low-dimensional mapping corresponding to the K-nearest neighbor are matched as much as possible. The learning of new samples is realized. Finally, the above method is applied to the actual gearbox fault diagnosis. The results show that the above method can effectively improve the accuracy of fault diagnosis, and verify the feasibility of the algorithm in practical fault diagnosis.
【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TH17;TP277

【相似文献】