SIA OpenIR  > 机器人学研究室
基于局部线性嵌入映射的非线性传感器故障诊断研究
Alternative TitleResearch on Locally Linear Embedding Based Nonlinear Sensor Fault Diagnosis
张伟1,2
Department机器人学研究室
Thesis Advisor周维佳 ; 李斌
ClassificationTP212
Keyword故障诊断 局部线性嵌入映射 切空间 投影寻踪
Call NumberTP212/Z35/2009
Pages113页
Degree Discipline模式识别与智能系统
Degree Name博士
2009-05-30
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract当今社会,随着系统复杂度越来越高以及对系统可靠性、安全性要求的日益提高,故障诊断越来越受到人们的重视,在这其中,非线性系统由于自身的特点,一直是故障诊断中的一个热点和难点。本文针对这个问题,对非线性系统传感器故障诊断的若干关键技术进行了深入研究,旨在建立一种相对通用的非线性系统传感器故障诊断方法。本论文研究的主要内容包括系统数据故障特征提取、故障检测、故障识别等。对于故障数据的特征提取问题,在深入分析现有各种算法的基础上,结合非线性传感器数据的特点,将现有的局部线性嵌入映射算法(LLE)引入到故障诊断中来。LLE算法克服了传统线性方法如主元分析(PCA)法在非线性系统上应用的局限,同时也具有以下几个突出的优点:1、能够很好的表达数据的内在流形结构,保留数据特征,这点在故障诊断中有重要的意义,可以比较好的保留原有数据的故障特征;2、不存在局部最优解;3、算法本身参数的选择很少,适合于工程应用。特征提取的研究为后续的故障检测与故障诊断打下基础。同时对原有的LLE算法本身存在的一些缺陷做了一定的改进,完善了其在故障诊断中的应用。LLE算法假设每个数据点及其近邻是局部线性的,因此近邻的选择就不能改变数据的固有流形。从这个概念出发,引入切空间距离代替传统的欧式空间距离,使近邻点的选择更加符合局部线性的要求,从而能更好的提高数据的输入/输出映射质量。针对LLE算法中内在维数难以估计的问题,在算法中应用分形中相关维的计算方法,用线性拟和方法改进了G-P算法,实现了线性区域的自动选择,提高了内在维数估计的精度。最后,通过核函数估计的方法,建立一个数据投影模型,可以实现实时数据到特征空间的直接投影,避免重复的计算权重矩阵,减少了计算复杂度。特征提取之后,针对故障检测问题,提出了一种基于空间分布的故障检测方法,通过将系统数据与高维空间分布相结合,把同一状态下的系统数据近似成为同一分布下的样本采样,检测实时数据是否服从系统正常状态下的分布,从而完成数据的故障检测。同时,结合LLE算法中的核函数投影方法,实现了数据在投影的同时完成故障判别,减少了计算复杂度。在故障检测的基础上,提出了基于投影寻踪的方法进行故障识别。对原有的投影寻踪的应用方法做了一定的修改,通过对最优投影向量的模式匹配完成故障的识别。首先建立各种故障状态下与正常数据之间的最优投影向量的数据库,然后将实时数据的最优投影向量与历史故障进行模式匹配,从而完成故障识别。为了提高投影向量寻找的速度与精度,采用粒子群优化算法,这也可以同时避免算法陷入局部极值点。最后,在总结全文的基础上,讨论了基于局部线性嵌入映射算法在理论及应用上有待于进一步研究的若干问题。; Nowadays,systems become more and more complex and the requirement of reliability and security of the system is increasingly improved. Because of its own characteristics, nonlinear sensor fault detection is all along a difficult and fucus problem. Aiming at this problem, this paper does a thorough research on some key technicals of nonlinear sensor fault diagnosis. The main research contents include feature extraction, fault detection and fault identification. Based on the deep analysis of the theories of existing algorithms of feature extraction and combined with charactors of nonlinear sensor data, a fault diagnosis algorithm based on Locally Linear Embedding (LLE) is proposed in this work. LLE algorithm overcomes the limition of traditional linear methods such as Principal Component Analysis applying in nonlinear systems and it also has some advantages. Fistly, it can express the internal structure of the data well, and preserve the data characteristics. This is important for the fault diagnosis, because it can preserve the fault feature of the data comparatively well. Secondly, the algorithm can avoid the local optimal solution. Thirdly, the algorithm has few parameters, and so it is suitable for engineering application. Feature extraction lays a good foundation for the subsequent research of fault detection and fault identification. Some improvements are made to the LLE algorithm aiming at its drawbacks, and so it perfects the application in fault diagnosis. LLE assumes that the sample data and its neighbors are locally linearization, so the selection of neighbors is very important to preserve the natual manifold of datas.Taking the concept of LLE algorithm as a point of departure, neighbor selection method based on tangent space is induced to the algorithm to replace the traditional Euclid space distance. It can fulfill the demand of locally linearization much better, and improv the project quality of Input/Output data. Aiming at the problem of intrinsic dimension estimation of data, Correlation Dimension estimate method is used in the algorithm, and the G-P algorithm is improved by linear fitting. It implenments the automatic selection of the linear zone, and improves the precision of intrinsic dimension estimation. At last, a project model is established by kernel method, and it can project the real time data to the feature space directly, which avoids computing the weight matrix repeatly and reduces computational complexity. After feature extraction, a spatial distribution based algorithm is proposed aiming at fault detection problem. It combines system samples with spatial distribution in high dimension space, and data from the same system state are considered as samples from the same distribution. So the fault detection task can be completed by checking whether the data are submited to the distribution of normal state of the system. At the same time, combining with kenel method reffered before, it can complete fault discrimination when the data are projected to feature space, and so it can reduce computational complexity remarkably. Based on fault detection algorithm, a fault identification method based on Projection Pursuit (PP) is proposed. The application method of PP is modified, and fault identification task is completed by pattern matching of the optimal projection vector of fault data and normal data. Firstly, the database of optimal projection vector between fault data and normal data is established, and then pattern matching is applied to the real time projection vector and projection vectors of historical fault. To improves the speed and precision of the selection of optimal projection vector, Particle Swarm Optimization (PSO) algorithm is used, which can also avoid the algorithm fall into local optimal. At last, there are concluded with a summary and some problems need to be studied in the future is put forward.
Language中文
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/426
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院研究生院
Recommended Citation
GB/T 7714
张伟. 基于局部线性嵌入映射的非线性传感器故障诊断研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2009.
Files in This Item:
File Name/Size DocType Version Access License
10001_20051801472804(3504KB) 开放获取--Application Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张伟]'s Articles
Baidu academic
Similar articles in Baidu academic
[张伟]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张伟]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.