中文摘要：

利用脉冲频率响应法在线检测电力变压器绕组变形故障时，传统的频率响应曲线并不能刻画暂态信号的时域特性，且快速Fourier变换的局限性可能造成有用信息的缺失。为了克服以上缺陷，提出利用连续小波变换处理脉冲在线注入获得的暂态信号，绘制检测信号小波时频图，以矩阵相似度作为故障评判的量化指标。首先从理论推导入手，构建了单绕组仿真模型，开展了各种故障类型仿真分析；其次，进行了故障绕组试验，分析了信号时频特性，另外，采用矩阵相似度量化评判健康与故障绕组检测信号的关联程度。结果表明：仿真不同故障类型对应的小波时频图与健康绕组的时频图差异明显，且时频图在各频段差异表现出和频率响应相似的规律；与健康绕组相比，试验故障绕组的小波时频图在0．6MHz以上高频段出现较大偏差，表现出绕组电容性故障应有的特性，而在0～0．6MHz频段，试验绕组小波时频图的矩阵相似度为0．9280，大于采用快速Fourier变换获得频响曲线的关联系数0．8003，证实了小波变换的优越性。仿真分析与实验测试的数据处理结果，均初步证实该方法的可行性。

英文摘要：

When impulse frequency response analysis is applied to online detect transformer winding deformation, the time domain characteristic of transient signal can not be depicted by conventional frequency response curves, and the limitations of fast Fourier transform （FFT） can result in deletion of available information. To overcome above problem, continuous wavelet transform was used to process transient signal, in which transient signal was obtained through online impulse signal injection, the wavelet time-frequency spectrum was plotted and matrix similarity was chosen as an indica- tor to diagnose fault. Firstly, a simulation model of single winding was constructed after theoretical deduction, the simulation analysis of different fault types was performed. Secondly, the faulty winding test was conducted, the time-frequency characteristic of detection signal was analyzed. In addition, the matrix similarity was used to quantize the relevance of detection signal between healthy winding and fault winding. Results show that the wavelet time-frequency spectrum under different fault types is significantly different from that of healthy winding, in which the deviations of spectrum in different frequency bands present similar regularity with the frequency response. Compared with healthy transformer, wavelet time-frequency spectrum of tested faulty winding shows significant difference beyond 0.6 MHz, which shows rightful characteristic of winding capacitive fault. In 0-0.6 MHz, matrix similarity of tested winding is 0.928 0, larger than the correlation coefficient 0.800 3 between signatures obtained by FFT, which shows priority of wavelet transform. Both the simulation analysis and experimental studies show the feasibility of the proposed method.