【作者】 刘一；

【导师】 赵奇；

【作者基本信息】 河北工程大学 ， 计算机应用技术， 2010， 硕士

【摘要】 汽轮发电机组是电力生产的重要设备,由于其设备结构的复杂性和运行环境的特殊性,汽轮发电机组的故障率一直比较高,故障危害性也很大。因此,汽轮发电机组的故障诊断一直是故障诊断技术应用的一个重要方面。汽轮机振动信号中一般含有大量的噪声,要求对振动信号进行消噪。在研究了Donoho阈值消噪的基础上,提出了阈值量化的新定义,并实现了一种基于遗传算法的信号消噪方法,通过引入了γ估计因子,对估计因子的遗传优化来实现提高信噪比的目的,比较检测效果,可以看到基于遗传算法的消噪效果要比传统的Donoho阈值消噪效果更好。小波包作为一种时频分析手段引入到振动信号分析,小波包系数可以非常灵活地提供信号在时域和频域的信息。通过实验分析,基于小波包分解算法的汽轮机故障特征提取相比FFT频谱分析算法,同样能够完全满足振动信号分析的要求,并且可以获得振动信号的能量在频率上的分布,这为贝叶斯网络的构建提供了基础。实验证明,该方法用于信号的特征提取是非常有效和切实可行的。将Bently转子实验室获取的能量-频率表离散化,来构建贝叶斯网络模型。并将Bently实验台上得到的碰磨数据进行消噪、特征提取,获得碰磨故障下的故障征兆,结合贝叶斯网络模型及专家经验确定的先验概率,来实现贝叶斯网络对故障的分类。实验结果证明,本文基于专家经验的贝叶斯网络模型与振动信号的消噪及特征提取技术,能根据振动信号,准确的判断故障类型。在山西某电厂实现了厂级信息监控系统(SIS)及该厂机组的仿真机系统,SIS能够实现生产流程监控及机组性能计算的功能;仿真机系统能完全并真实反映该厂机组的情况,为该厂提供了虚拟的技术平台。更多还原

【Abstract】 Turbine vibration signal generally contains a lot of noise, so requires de-noising the vibration signal.Based on the Donoho threshold de-noising, proposed a new definition of quantitative threshold, and implements a signal de-noising method based on genetic algorithm, Through the introduction of the estimated factorα,to achieve the purpose of improving signal to noise ratio by genetic optimization the estimated factor. Comparing test results, we can see based on genetic algorithms de-noising effect is much better than the traditional threshold of Donoho de-noising.Wavelet packet as a time-frequency domain analysis tools introduced into the vibration signal analysis, wavelet packet coefficients can be very flexible to provide the information of signal in time domain and frequency domain. Through experimental analysis, feature extraction of turbine fault based on wavelet packet decomposition algorithm is compared with FFT spectrum analysis algorithm, also be able to fully meet the requirements of vibration signal analysis.Also get the distribution of vibration signal energy in the frequency, this provide the basis for the construction of Bayesian networks.Experiments show that the method used to signal feature extraction is very effective and practical.Discretization the energy-frequency table which obtained in Bently Rotor laboratory, to build a Bayesian network model.At the same time, de-noising the Rubbing data from Bently experimental table, feature extraction, obtained the fault symptoms under the rub fault,combined with Bayesian network model, and the expertise to determine a priori probability, to achieve the fault classification by the Bayesian networks. Experimental results show that according to turbine vibration signals,the expertise-based Bayesian network model and vibration signal de-noising and feature extraction techniques, can accurately determine what the fault it is.更多还原