【作者】 王静；

【导师】 顾大强；

【作者基本信息】 浙江大学 ， 机械设计及理论， 2004， 硕士

【摘要】 铁谱分析是一种被广泛应用在机械设备磨损故障诊断和磨损状态监测中的技术手段，磨粒识别是铁谱分析技术中的关键问题。随着计算机技术和人工智能技术的迅猛发展，将计算机视觉技术、专家系统、人工神经网络、模糊理论等引入铁谱分析技术中，实现磨粒识别的智能化已成为铁谱技术研究领域中的热点和难点问题。 本文首次将支持向量机技术引入铁谱分析技术中，进行磨损模式识别方法研究。支持向量机是一种基于统计学习理论框架下新的通用机器学习方法，它不但可以较好地解决以往很多学习方法存在的小样本、过学习、局部最小等实际难题，而且具有很强的泛化能力。 本文的主要研究工作有： 1、综合国、内外有关文献，对磨粒分析技术的发展和现状进行综述；提出本文的研究思路和主要内容； 2、分析论述磨损的产生机理与分类，磨粒的分类及特征；阐述了基本磨粒类型、磨损类型、特征、产生机理与设备运行状态之间的内在联系； 3、研究磨粒图像的预处理方法和磨粒形态特征的提取方法；论述了基于人工神经网络和模糊理论的两种磨粒智能识别方法，并指出其中的难点和不足；研究在有限样本下的基于统计学习理论的支持向量机技术，探讨支持向量机的分类机理，建立基于支持向量机的磨粒识别系统框架； 4、将支持向量机应用于磨损模式识别，设计磨粒分类器；进行基于支持向量机的磨粒分类器的细节设计，包括数据样本的建立、训练算法、多分类模式、核函数等；分析分类器中的主要数据结构、类和函数的功能，并给出程序运行时的界面； 5、采用100个磨粒样本的四个形态特征量：圆形度、细长度、散射度和凹度作为支持向量机分类器的输入，输出为滑动磨损、切削磨损、正常磨损和疲劳点蚀四种磨损形式，研究支持向量机中的核参数对磨粒分类器的性能影响；选择适当的分类器参数对分类器进行仿真实验，得到了96％的分类准确率，验证分类器的有效性； 6、从理论和仿真实验两方面来比较基于支持向量机与基于BP神经网络的磨粒分类器的性能优劣研究，以相同的磨粒样本、特征和磨损形式作为分类器的输入、输出，结果表明前者比后者高出6％的识别准确率，说明基于支持向量机的磨粒分类器有一定的优势，并进行了原因分析。 本文提出的基于支持向量机的磨损模式识别方法为磨损故障诊断和状态监测以及铁谱分析技术智能化发展提供了一条新的思路和途径。 本项目受国家自然科学基金项目资助(项目批准号：50375141)更多还原

【Abstract】 As a technical method, ferrography analysis is used widely in wear fault diagnosis and condition monitoring. The recognition of wear particle is regarded as a key point in ferrography analysis technology. With the rapid development of the computer technology and artificial intelligence technology, some methods were applied in the ferrography technology, such as computer vision technology, expert system, artificial neural network and fuzzy theory. Intelligentization of the recognition of wear particle is a important and difficult issue.The thesis introduces a new method for ferrography analysis based on Support Vector Machine (SVM) for the first time. SVM is a kind of novel machine learning methods, which based on Statistics Learning. It becomes a study hotspot in the international machine learning comparing with other methods because of its excellent performance, e.g., limited samples, overfitting and local convergence problems. Moreover, it has better generalization ability.This paper is summarized as follows:1. The development and up-to-date status of wear particle analysis technology at home and abroad are evaluated synthetically. The study plan and main content are presented.2. Wear mechanism and classification including wear particle classification and characters are analyzed and discussed. The inner relations between each basic kind of wear particles and wear type, wear particle characters, wear mechanism, machine running status are analyzed and expatiated.3. The method about pretreatments of debris image and extracting morphologic characters of wear particles is studied. The popular intelligent recognition methods based on artificial neural network and fuzzy theory are studied, and their difficulties and shortcomings are pointed out. The SVM technique based on the Statistics Learning Theory and limited sample is studied. The classifying mechanism of SVM is discussed, and the framework of wear particle recognition system based on SVM is built.4. The SVM is applied to wear pattern recognition, and the wear particle classifier is designed. The detailed design of SVM wear particle classifier is performed including the build-up of wear sample data, the training algorithm, the multi-class pattern, the kernel function. The main data structures, classes and functions of the SVM wear particle classifier are analyzed, and the interface of procedures when running is presented.5. 100 wear particle samples are cliosen, and their morphologic characters are taken as the input of SVM wear particle classifier including roundness, slightness, scatter andconcavity, the sliding, cutting, normal, fatigue erosion wear are taken as the output of SVM wear particle classifier. The effect when applying different kernel parameters for SVM wear particle classifier is studied. Choosing the appropriate parameters to prove the validity by experiment, the correct of the classifier Is up to 96%.6. The performances of the classifiers based on SVM and on the BP neural network are tested by using the same wear particle samples. The result indicates that the correct ratio based on SVM is beyond BP neural network 6%. The superiorities of SVM classifier are presented by the comparing experiment and analyzed from theory.The wear pattern recognition method based on SVM and proposed by this thesis, gives a new way to wear fault diagnosis, wear condition monitoring and intelligentization of the ferrography technology.This project is supported by a grant from the National Natural Science Foundation of China which the authorizing number is 50375141.更多还原