【作者】 杨霁；

【导师】 孙才新；

【作者基本信息】 重庆大学 ， 电气工程， 2004， 博士

【摘要】 由于绝缘内部的局部放电是导致电气设备绝缘劣化的重要因素，而且放电模式与内部缺陷类型具有紧密的联系，因此，局部放电在线监测与模式识别技术是及时发现放电性故障并诊断出绝缘内部局部缺陷的重要手段，对于防止高压电气设备事故发生，提高高压电气设备运行的安全性和可靠性具有重要的意义。本文在分析国内外电力变压器局部放电在线监测与模式识别技术的基础上，系统地研究了抗干扰技术和图像分析中的小波多尺度变换的原理和方法，对局部放电监测中的小波去噪技术以及基于小波多尺度分析的局部放电图像识别技术进行了深入研究。本文主要研究内容如下： 1)系统地研究了小波多尺度分析的基本理论和具体算法，并对7种常用小波进行了对比分析；通过对Blocks和Doppler两种典型的仿真信号的小波多尺度分析，得到了信号的不同频率分量在小波分解各尺度上的分布规律；通过对Woman和Sinsin图像的小波多尺度分析，得到了各尺度上低频子图像和高频子图像的性质及特点，以及图像的位置信息和类别信息随尺度的变化规律。 2)在分析局部放电信号、白噪声的小波多尺度变换特性以及小波阈值去噪法原理，分析和总结Donoho提出的阈值估计及阈值处理方法的基础上，提出了一种应用于局部放电监测中去除白噪的自适应阈值选择方法：采用一种具有多阶导数的阈值函数，从而能够进行自适应迭代寻找到最优阈值。通过对Blocks、Bumps、Doppler、HeaviSin等四种典型仿真信号和仿真局部放电信号的去噪处理，证明该方法比软阈值法具有更有效的噪声抑制效果，同时具有较低的信号畸变率。 3)应用小波多尺度分析技术去除局部放电信号中白噪干扰的分析，发现了小波基的选取对去噪结果有很大的影响，并针对指数衰减型和衰减振荡型两种局部放电脉冲的特征，根据小波与局部放电脉冲波形相关性选取最佳小波基的原理，提出了一种新的最优小波基的自动选择算法；计算结果表明，在Daubechies小波系中，db2小波最适合于去除指数衰减型局部放电脉冲和衰减振荡型局部放电脉冲信号中白噪声。 4)通过周期性窄带干扰的小波多尺度分析以及对用小波阂值去噪法去除周期性窄带干扰效果的分析，发现小波阈值去噪法对去除周期性窄带干扰并不理想，提出采用自适应小波阈值法和二阶级联IIR格型陷波器的组合去噪方法，通过对仿真信号和变压器局部放电在线监测系统实测信号的分析表明，该方法对去除包含白噪声和窄带周期干扰的混合噪声具有良好的效果。 5)对设计的5种放电模型进行了大量实验，并根据测得的放电数据，设计了重庆大学博士学位论文局部放电灰度图像的构造方法和统一的数据文件格式;在对图像的小波多尺度分析方法基础上，提出了一种基于图像小波多尺度变换的局部放电图像识别方法。通过样本数据的识别计算和分析表明，分别采用局部放电图像的低频子图像、高频子图像、高低频子图像组合图像作相似度计算而得到识别结果不同，而且识别结果与图像的小波分解尺度有关;相比较而言，采用局部放电图像的低频子图像进行模式识别能够获得较高的识别正确率，而且对于本文中的5类放电样本，采用dbZ小波作4尺度小波分解的低频子图像进行识别时的识别正确率相对最高，其最低识别正确率高于90%。 以上研究表明，小波多尺度分析技术特有的时频特性与其他信号处理技术结合起来能有效、稳定地去除在线监测系统中混杂在局部放电信号中的各种干扰;同时小波多尺度分析在图像识别中的应用，提高了局部放电图像识别的能力，这些应用证明小波多尺度分析在局部放电监测中具有广阔的应用前景。关键词:局部放电，小波多尺度分析，最优小波，自适应阐值法，图像识别更多还原

【Abstract】 Partial discharge (PD) inside insulation is considered as one main reason that lead to electrical equipments damage, and discharge patterns are closed related to the types of internal defects. As one of the major methods to find out discharge faults and diagnose internal defects, the technologies of PD on-line monitoring and pattern recognition are of importance for the diagnosis of the quality of HV insulation system, and of importance to the safety and reliability of electrical equipment in service. The national and international techniques of transformer PD on-line monitoring and pattern recognition are investigated in this paper, based on which the techniques of noise-suppression and multi-resolution wavelet transform normally applied in image processing are introduced in to solve the problems of signal de-noising and pattern recognition for PD online measurement of transformers. The main contents are as follows:(1) The fundamental theory and detailed algorithm are presented in this paper. Comparison among seven kinds of standard wavelets is carried out. Two typical simulation signals, Blocks and Doppler, are analyzed using multi-resolution wavelet, through which the distribution rules about different frequency components of signal in different wavelet decomposition scales are achieved. The Woman and Sinsin pictures are analyzed using multi-resolution wavelet transform, through which the properties of low and high frequency sub-image are investigated and the principles of image position and pattern changing with wavelet scales are obtained.(2) In this paper, the properties about multi-resolution wavelet transform of PD and white-noise signals and the de-nosing method using thresholds estimation is developed, and a self adaptive thresholds selection method is then proposed to solve the problem of white-noise suppression: a new multiple derivative threshold function is introduced in self adapting iteration process to search for the optimal value. Four typical simulating PD signals are used for de-nosing research. It is proved to be more effective than soft-threshold when applied in noise-suppression.(3) The multi-resolution wavelet transform is applied in white-noise inference suppression. A new algorism for optimizing to choose wavelet is put forward in order to get nice effect for noise-suppression. The method for choosing the best wavelet using related function is studied and the searching algorism is made. According to the characteristics of two kinds of typical PD signal respectively attenuation model andattenuation and surging model, it has been demonstrated that DB2 is the best one in Daubechies for white noise suppression.(4) Multiple-resolution wavelet transform and threshold de-nosing methods are investigated to remove narrow-band noise while the effects are not so satisfied. Therefore, a new de-noising method based on self-adaptive thresholds combined with the 2~nd cascade IIR. notch digital filter is proposed. By comparing the simulated PD-signals with real PD -signals measured by online measuring system at a specific substation, it shows that the proposed method is an effective tool to remove noise and improve sensitivity of PD online measuring systems.(5) A large number of experiences are carried out using 5 artificial PD models. Unitized data file style and PD gray image constitution method are designed according to the measured PD data. A PD imagine recognition method based upon multi-resolution transform is proposed in this paper. It shows that different recognition results are obtained though the correlation parameters calculated respectively by low, high frequency and their combined sub-image. By cooperation, a higher discrimination can be obtained using low-frequency sub-image. Furthermore, the best discrimination is achieved through wavelet transform using DB2 at the 4 scale, and the lowest discrimination is more than 90%.In general, wavelet multi-scale analysis is a good way to restrain white-noise, and also narrow-band noise and others with modern signal更多还原