【作者】 冯长建；

【导师】 吴昭同； 丁启全；

【作者基本信息】 浙江大学 ， 机械制造及其自动化， 2002， 博士

【摘要】 本文以国家自然科学基金项目“基于隐Markov模型的旋转机械故障诊断新方法的研究”(编号：50075079)为基础，提出的博士学位论文题目为“HMM动态模式识别理论、方法以及在旋转机械故障诊断中的应用”。本文以大型旋转机械为研究对象，研究了HMM动态模式识别理论与方法在旋转机械故障诊断中的应用，开辟了旋转机械计算机辅助故障诊断的新途径。全文主要研究内容如下： 第一章：介绍了旋转机械振动监测和诊断的概况；综述了多变量动态模式识别理论的发展和研究现状；结合国家自然科学基金提出了本文的研究内容；最后，给出了本文的总体框架和创新之处。 第二章：介绍了Markov链基本理论，并通过一个简单的实例把它扩展到了隐Markov模型(HMM)；然后重点介绍离散HMM的基本理论、算法以及在实际应用中的改进措施。 第三章：根据离散HMM(DHMM)的基本理论，提出了旋转机械振动幅值谱矢量的标量量化方法，并在此基础上提出了基于DHMM的故障诊断方法；利用转子升速过程的振动模式验证了DHMM故障诊断方法的有效性。 第四章：在连续隐Markov模型(CHMM)的基本理论基础上，提出了直接利用振动信号AR系数特征矢量序列建立混合密度CHMM的故障诊断新方法，并对转子升速过程的振动模式进行了成功的识别；对DHMM和CHMM故障诊断方法进行了对比分析，指出DHMM方法具有算法稳定、计算速度快、识别精度高等特点，对于CHMM方法只要通过合理选择特征参数也能得到高的识别精度。 第五章：利用SOM神经网络对多传感器的振动信息特征进行降维、压缩与编码，首次提出了基于多通道振动信息融合的HMM故障诊断方法，拓展了监测对象的观测范围，从而能够对旋转机械的整体运行状态做出综合识别，并对提出的方法进行了实验研究。 第六章：首次把AR系数矢量引入到Kalman滤波器和HMM中，得到了两种自适应描述非平稳动态时间序列的在线模型—Kalman-AR和HMM-AR模型；描述连续状态变化的Kalman-AR滤波模型为描述离散状态变化的HMM-AR模型提供了一个良好的初始化方法和给出过程状态转移点的先验信息，从而使HMM-AR模型能够对非平稳过程进行状态分割和分类；仿真和实验结果表明，提出的该旋转机械运行状态在线监测方法，能够实现对旋转机械运行的状态进行成功地分割和分类。 第七章：基于多变量动态模式识别的理论和方法，在混合编程环境下开发了HMM的旋转机械故障诊断应用软件;介绍了软件系统的开发环境、开发工具以及Matlab和C++混合编程的接口实现方法;介绍了整个软件系统的基本组成和功能。 第八章:给出了全文研究内容的总结;并对HMM理论在旋转机械故障诊断方面的进一步应用提出了展望。更多还原

【Abstract】 Based on the "Application on Faults Diagnosis of Rotating Machine in Hidden Markov Models" (National Nature Science Fund Project, No: 50075079), the Hidden Markov Models (HMMs) dynamic pattern recognition theories and methods are studied, then proposed the applications in faults diagnosis of rotating machine by HMM methods and developed the faults diagnosis software based on HMM. The details are studied as follows:Chapter one briefly introduces the general situation of vibration monitoring and faults diagnosis of the rotating machine. The developing and the current situations of the multivariate dynamic pattern recognition theories are summarized. At last, the background, main contents, general structure scheme and innovation points of this dissertation are present.Chapter two introduces the basic ideas of Markov Chain theories briefly, and then extends it to Hidden Markov Models through a simple example. At last the theories and algorithms of Hidden Markov Models are studied. The modification algorithms of HMM are proposed. Therefore the basic theories of this dissertation are established.Chapter three proposes the basic concept of dynamical pattern recognition and introduces the implementation theories based on probability statistics. Based on the theory of the Discrete Hidden Markov Models (DHMM), the scalar quantization method of vibration spectrum vector of rotating machine is proposed, and then the fault diagnosis method based on DHMM is designed. The experiment of run-up process of rotor machine is made to verify the effect of the new faults diagnosis method.Chapter four introduces the basic theories of Continue Hidden Markov Models (CHMM). the new method of faults diagnosis based mixture density CHMMs directly by the vibration AR coefficients vectors of rotating machine is proposed, and then the dynamic patterns presented in run-up process of rotor machine are successfully recognized. At last compares the two faults diagnosis methods of DHMM and CHMM, and points out the advantages and disadvantages of the two methods.Chapter five studies the new method for faults diagnosis of rotating machine based on SOM and HMM through the integrated information of multi-sensor. First self organization clustering method for high dimension data is proposed, and then DHMM for faults diagnosis is designed by the low dimension features. Experiments verified that the method is effective.Chapter six propose a useful method based on HMM-AR for modelling the dynamictime series of rotating machine running process. It is verified through simulating data and experiments data.Chapter seven builds the HMM faults diagnosis software for rotating machine. The developing environments, tools and implementation methods of interface between Matlab and C++ mixture programming languages on this software are introduced. At last the basic components and functions of the software are illustrated.At last, in the eighth chapter, all of the work in this dissertation is summed up, and the future researches on applications of HMMs are prospected.更多还原