【作者】 汪江；

【导师】 陆颂元；

【作者基本信息】 东南大学 ， 动力机械与工程， 2005， 博士

【摘要】 电力工业的迅速发展,对汽轮发电机组状态监测与故障诊断技术提出了高的要求。本论文就机组振动故障特征自动提取、振动异常检测、振动故障趋势预测、多征兆故障诊断以及机组远程监测诊断等几个影响故障诊断技术发展的关键技术进行了详细研究,取得了一定的成果。分析了发电机组振动故障及征兆的特点,研究了故障征兆自动提取方法,采用Mann-Kendall检验、EMD分解、相关系数法、不变矩理论对趋势型征兆、相关性征兆以及转子轴心轨迹图形类征兆进行了提取。将统计学习理论的研究成果-支持向量机(SVM)应用于机组振动异常检测、振动趋势预测及故障诊断中。分析了One-Class SVM异常检测的基本原理,提出了发电机组振动异常检测One-Class SVM方法,仅通过对正常运行状态样本的学习,就可以达到对机组振动异常的准确识别,解决了生产中机组异常振动训练样本缺乏问题;与Fuzzy ART1及ARTMAP神经网络的比较试验结果进一步证实了该方法的有效性;提出了一种核聚类二叉树型多类SVM算法,采用遗传算法对模型参数进行选择,将其应用在发电机组多征兆故障诊断中,试验表明该方法具有较高的故障识别率,缩短了模型学习需要的时间,并解决了one-against-one等多类学习方法存在的拒分类问题;利用One-Class SVM的一类样本学习能力,提出一种新的机组振动故障诊断动态模型,解决了其它故障诊断系统模型结构固定,无法对新增故障进行诊断的问题,实现了对未知新故障和复合故障的正确识别和学习,扩充了故障诊断系统的功能;研究了支持向量回归方法及最小二乘支持向量回归(LS-SVR)在时间序列预测的应用,通过仿真实验得出了LS-SVR在有噪声环境下,可以取得比RBF更好的预测效果,适合于现场实际应用;将LS-SVR用于发电机组振动趋势预测,取得了较高的预测精度。采用JAVA语言,使用Applet、Socket、多线程、Web数据库技术,开发了一套汽轮发电机组Internet远程监测与故障诊断系统,实现了机组振动远程在线监测与故障诊断。更多还原

【Abstract】 With the rapid development of power industry, it asks more on the condition monitoring and fault diagnosis of turbo-generator. The paper made an intensive and valuable study on several key technologies associated with the development of fault diagnostics, which include attribute extraction of vibration fault, multi-symptom fault diagnosis, vibration condition monitoring, vibration fault trend forecasting and remote online monitoring and diagnosis.Characteristics of vibration fault and fault symptoms of turbine generators are studied. A few new methods are used for attribute extraction of vibration fault symptom. A Mann-Kendall test is used to extract trend symptom of vibration parameters; an EMD method and a correlation coefficient one are used to extract correlation symptom; an invariant moment method is used to extract graphical symptom of shaft center orbit.A new machine learning method-Support Vector Machines, has been used to detect vibration abnormality, predict vibration and fault trend, and to diagnose vibration fault.Based on analysis of One-Class SVM theory, a new vibration abnormality detection method has been put forward for steam turbine generator units. This One-Class SVM abnormality detection method, require only samples gathered when turbine generators are in good order, which finds a solution to lack of abnormal training samples. Comparison tests with Fuzzy ART1 and ARTMAP neural networks the method this thesis suggested have a good performance.A multi-class classifier has been bring forward and been used for vibration fault diagnosis, which shorten time when training classifier with fault samples, and get rid of the rejection classifying, which will be found in some multi-svm classifiers, such as the one-against-one SVM one. Another fault diagnosis model using One-Class SVM has also been made, which extends capability of diagnosis model by add new One-Class SVM classifier dynamically.The theory of support vector regression and the application of LS-SVM in time series prediction have been studied. An emulational test has been made which proves LS-SVM can have a good performance with a time series polluted by noise and get a better result than RBF neural networks. This method was used to predict vibration and fault trend for turbine generators, which can have a more accurate result. At last of the thesis, by using Java programming language, an Internet-based remote condition monitoring and fault diagnosis system was been developed, which takes full advantage of Java Applet, Socket, multithreads and Web database technology, and make true remote monitoring and fault diagnosis.更多还原