【作者】 徐其惠；

【导师】 贾嵘； 李辉；

【作者基本信息】 西安理工大学 ， 电力系统及其自动化， 2008， 硕士

【摘要】 随着世界上装机容量的迅速增长,对供电可靠性的要求越来越高。对电力设备进行在线监测与故障诊断,是实现设备预知性维修的前提,是保证设备安全运行的关键,也是对传统的离线预防性试验的重大补充和拓展。本文以变压器状态在线监测与故障诊断系统为研究对象,在深入研究最小二乘支持向量机(LS-SVM)和经验模态分解(EMD)理论的基础上,将其引入到变压器故障诊断系统中,着重对油中溶解气体分析(DGA)和局部放电监测中的信号抗干扰两个方面进行了研究。DGA是变压器绝缘监督的一个重要手段,但放电和过热两类故障共存时的故障难分辨会导致诊断正判率较低。本文对多种故障类型作了详细分析后提出了两种故障识别方法:基于LS-SVM多分类的DGA方法,通过相关统计分析和数据预处理,选择变压油中典型气体作为LS-SVM的输入,然后利用典型故障气体的相对含量在高维空间的分布特性进行变压器故障类型诊断;基于邻域粒子群算法优化BP神经网络的DGA方法,选择油中典型气体作为神经网络的输入,然后利用训练好的邻域粒子群算法优化后的神经网络进行变压器故障类型诊断。试验结果表明,这两种方法具有很好的分类效果,较好地解决了变压器放电和过热共存时故障的难分辨问题,对故障诊断的正判率较高。近年来局部放电监测一直是国内外研究的热点问题,而干扰抑制一直是其中的难点。本文针对窄带干扰,提出了两种方法:基于LS-SVM回归的频域分析法,将原始时域信号变换到频域,利用LS-SVM对信号干扰频率最大幅值点附近的数据进行拓延,最后逆变换以得到抑制干扰后的信号,该方法有效地拓展了干扰频带的宽度,较好地抑制了窄带干扰残余;基于EMD和自适应噪声对消的方法,首先在频域中降低干扰幅值,接着利用EMD的分频特性将宽频带的窄带干扰分解到不同频带,各频带内的窄带干扰频率相差有限,然后进行自适应噪声对消。针对白噪声干扰,提出了一种全新的方法:基于EMD和固有模态函数(IMF)重构的方法,首先对含噪的局部放电信号进行EMD,得到含特征频率的IMF,然后对所得的IMF分量进行自适应阈值处理后重构,从而抑制噪声干扰,相比于常规的小波去噪算法,该方法具有自适应性强,不受小波函数和最佳小波分解层数选取的限制等优点,而且实现了阈值和固有模态函数阈值处理层数的自动选取。更多还原

【Abstract】 With the rapid growth of installed capacity in the world, the requirement for the reliability of power system becomes more and more urgent. Implement of state on-line monitoring and fault diagnosis of the power equipment is the precondition of predicting maintenance, is the key element of reliable run, and is the important supplement and updated development to the traditional off-line preventive maintenance.In this thesis state on-line monitoring and fault diagnosis of transformer are the research objects. Firstly Least Squares Support Vector Machine (LS-SVM) and Empirical Mode Decomposition (EMD) are deeply studied. After that, they are applied into the fault diagnosis system for transformer, furthermore, study results of both Dissolved Gas Analysis (DGA) and on-line monitoring of Partial Discharge (PD) are given.DGA is one of the mainly methods on insulation monitoring of transformer, but the accuracy is low because it is difficult to distinguish between failures when overheating and PD coexist. By analyzing all types of faults, two methods are raised. One is the DGA of transformer based on multi-classification LS-SVM. Based on correlation analysis and pretreatment, some key gases are selected as the inputs of LS-SVM, furthermore, fault diagnosis is accomplished according to the concentration distribution of typical fault gases in higher dimensional space. The other is the DGA of transformer by neural network based on particle swarm optimization (PSO) with neighborhood operator. Based on correlation analysis and pretreatment, the key gases are selected as the inputs of neural network, furthermore, fault diagnosis is accomplished by the neural network based on PSO with neighborhood operator. By discussing the experiment results, the methods of this paper have very good classification results, and figure out the problem that is difficult to distinguish between failures when overheating and PD coexist, meanwhile, the effectiveness and usefulness is proved.In recent years, on-line monitoring of PD has been a focus at home and abroad, and suppression of interference is the difficulty of fault diagnosis. Two methods of suppressing narrow-band interference are proposed in this paper. One is the improved frequency domain analysis based on data extension using least squares support vector regression. Firstly, the actual PD signals with noise are transformed from time domain to frequency domain, then extending the data near the maximum of interference frequency, ultimately the final signals can be obtained by Inverse Fourier Transform (IFT). This method enlarges the range of interference frequency band and removes effectively the interference residue. The other is a new adaptive algorithm based on EMD. First, the mid-signals can be produced by reducing the amplitude of narrow-band interference in frequency region, which is decomposed with EMD, and intrinsic mode functions (IMF) which contain specific frequency can be obtained, then for every IMF adaptive noise canceller is used to suppress narrow-band interference. A new method of denoising based on EMD and reconstruction of IMF is investigated. The PD signals with noise are decomposed with EMD, and IMFs can be obtained, through thresholding and reconstructing every EMF, then the noise can be depressed. Compared with the conventional wavelet-based denoising algorithms, the new algorithm is simpler, more flexible, and not limited by the selection of wavelet function and optimal decomposition level of wavelet, meanwhile, automatic selection of threshold value and thresholding layers of intrinsic mode functions are realized.更多还原