Identification of Power Transformer Winding Mechanical Fault Types Based on Online IFRA by Support Vector Machine
A transformer is the most valuable and expensive property for power utility, thus ensuring its reliable operation is a major task for both operators and researchers. Online impulse frequency response analysis has proven to be a promising technique for detecting transformer internal winding mechanical deformation faults when a power transformer is in service. However, as so far, there is still no reliable standard code for frequency response signature interpretation and quantification. This
... tries to utilize a machine learning method, namely the support vector machine, to identify and classify the winding mechanical fault types, based on online impulse frequency response analysis. Actual transformer fault data from a specially manufactured model transformer are collected and analyzed. Two feature vectors are proposed and the diagnostic results are predicted. The diagnostic results indicate the satisfied classifying accuracy by the proposed method. of resistance, capacitance and inductance in high frequency range, and its frequency response signature can represent the status of the winding . The first version of FRA, known as sweep frequency response analysis (SFRA), has adopted sweep frequency sinuous signal to excite transformer windings and measure the response signal in the frequency domain to construct a frequency response signature  . Afterwards, the other version of FRA, known as impulse frequency response analysis (IFRA), has emerged  . An impulse signal with wideband frequency spectrum has been used to excite the transformer windings and the response signal is measured in time domain, both signals are converted to frequency domain to construct a frequency response signature  . Recently, online SFRA and online IFRA have been burgeoned. Behjat realized the experimental application of online SFRA by a homemade noninvasive capacitor sensor (NICS), in combination with a network analyzer [13, 14] . Bagheri studied the effect of the bushing tap capacitance on the online SFRA signature by connecting variable capacitive bushings to the tested windings  . For the online IFRA, Leibfried used the transient signal, which was caused by the switching operation, to excite transformer windings  ; while Wang introduced the lightning over-voltage being the excitation signal of the online IFRA  . These two types of transient signals belong to the uncontrollable signal. Furthermore, by using controllable signals, Rybel proposed a method to online monitor the transformer by injecting the high frequency signal to the winding terminal through the bushing tap  . Then, Yao established a capacitive coupling sensor (CCS) and an other apparatus for injecting the nanosecond pulse to the transformer windings  . Besides, Zhao analyzed the impact of the capacitive coupling circuit on the online IFRA signature by theoretical analysis and experimental validation  . Compared with current SFRA, IFRA has the advantage of a high signal to noise ratio and smaller energy injection, which reaches the potential for online application. Although there are extensive publications characterizing the online FRA, most of them focus on the implementation of the method itself. As so far, there is still no reliable standard code for both SFRA and IFRA signature identification and quantification [21, 22] . FRA is a graphical analysis method; skilled personnel is required for the interpretation of FRA signatures. Moreover, there is little literature regarding the identification of winding faults in online IFRA. Rahimpour has divided the comparison algorithms of FRA signatures into four major categories: Algorithms based on exact calculations, algorithms based on estimation methods, algorithms based on electric circuit models and algorithms based on artificial methods  . Among these algorithms, artificial methods, such as genetic algorithm (GA), neural network (NN), support vector machine (SVM), are intelligent machine learning algorithms with high accuracy. Particularly, SVM has advantages of solving the classification problem of small-scale samples, non-linearity, and high-dimension. For the problem of classifying transformer winding mechanical deformation, it is not easy to collect actual FRA data for different winding deformation types, extent and location, especially for large power transformers, which makes SVM developing a suitable tool for solving this problem. Bigdeli proposed SVM to identify transformer winding faults based on the transfer function, and the verification process reveals the proposed method has a high accuracy  . However, this work involves offline FRA. With a view to the current status of online IFRA identification, this paper proposes the SVM algorithm to classify the transformer winding mechanical fault types, for the purpose of achieving the actual application of the online IFRA method. Brief Introduction of Online Impulse Frequency Response Analysis A simple diagram of the online IFRA method is depicted in Figure 1 . The transformer is in service, with the systematic power source energized in the HV side and the load in the LV side. The controllable high voltage nanosecond pulses are injected into the terminal of the transformer winding by using a CCS, which is a metal strip that is wrapped around the bushing external insulation layer. Particularly, CCS functions as a medium to couple the nanosecond pulse voltage into the winding terminal. The usage of CCS has erased the low frequency data of online IFRA signature, and this effect can be found in [20, 25] . The impact of CCS on bushing external insulation can be discovered in  . More detailed information about the diagnostic system and the practical application can be found in [19, 27] .