Transformer winding fault diagnosis method based on Hemma-convolutional neural network

Abstract

The invention relates to a transformer winding fault diagnosis method based on a Hemma-convolutional neural network, and the method comprises the steps: analyzing the representation form of a transformer winding fault, constructing a finite element model under the normal and fault states of a transformer, and arranging the position of a sensor; analyzing magnetic flux leakage characteristics of the transformer winding in a swelling and compression state, and extracting magnetic flux leakage intensity characteristics to identify a winding fault type; analyzing magnetic flux leakage differences of faults at different positions when the transformer winding is compressed, and extracting corresponding magnetic flux leakage characteristics to locate a fault winding and a fault position; analyzing magnetic leakage weak signal characteristics when the transformer winding bulges, identifying a bulge fault and a normal operation state through wavelet-skewness, and positioning a bulge position; and transformer winding fault diagnosis is realized based on the Hereum-convolutional neural network model. The method solves the problems that the slight winding fault accuracy of the transformer is low and the fault winding is difficult to locate.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer winding fault diagnosis, and particularly relates to a transformer winding fault diagnosis method based on Hippopotamus-Convolutional Neural Network. Background Art

[0002] As an important energy transmission device in the power system, the healthy operation state of a power transformer will affect the safety and stability of the entire power system. With the continuous expansion of the scale of China's power grid, the short-circuit current generated when the power supply system fails will be larger than before, resulting in the transformer being likely to be subjected to electromagnetic force far exceeding the normal state in a certain direction during the power system fault, thus causing deformation faults in the transformer winding. Affected by the setting threshold, traditional differential protection is difficult to exclude minor winding faults of the transformer, and such faults are difficult to detect manually, resulting in the possibility that minor winding faults may evolve into major accidents. To improve the diagnostic ability for minor winding faults of the transformer, researchers have designed various transformer fault detection methods such as frequency response method, vibration signal method, parameter identification method, oil chromatographic analysis method, etc., which effectively reduce the risk of transformer faults developing into accidents, but these methods are still difficult to identify minor winding faults and are prone to misjudgment risks.

[0003] In view of the problem of low diagnostic rate for minor winding faults of the transformer, many scholars currently use the leakage magnetic difference reflected by early transformer faults to achieve online detection of transformer faults. As recorded in Document [1]: Liu Jianfeng, Li Zhiyuan, Zhou Yaru. Early Fault Diagnosis of Transformer Windings Based on Leakage Magnetic Field and ICOA-ResNet [J]. Power System Protection and Control, 2024. And as recorded in Document [2]: Deng Xiangli, Zhu Hongye, Yan Kang, etc. Research on Transformer Magnetic Balance Protection Based on Fiber Optic Leakage Magnetic Field Measurement [J]. Transactions of China Electrotechnical Society, 2024. These studies show that the influence of winding faults on the leakage magnetic distribution is obvious, but these articles do not study the influence of different winding faults on the leakage magnetic change, nor do they deeply explore the characteristics of leakage magnetic change, making it difficult to identify minor winding faults and locate the faulty winding. Summary of the Invention

[0004] In view of the problems existing in the existing transformer diagnosis scheme, such as low diagnostic rate for minor winding faults and difficulty in excluding the influence of different faulty windings, a transformer winding fault diagnosis method based on Hippopotamus-Convolutional Neural Network is proposed. This method is based on the leakage magnetic distribution characteristics of the transformer, analyzes the extraction effect of different characteristic quantities on the leakage magnetic distribution curve, and uses the Hippopotamus-Convolutional Neural Network to realize the fault diagnosis of the transformer winding, improving the accuracy of transformer winding fault diagnosis.

[0005] The technical solution adopted by the present invention is as follows:

[0006] The method for diagnosing transformer winding faults based on Hippopotamus-Convolutional Neural Network includes the following steps:

[0007] Step 1: Analyze the main manifestation forms of transformer winding faults, construct finite element models of the transformer under normal and faulty states, and arrange the positions of sensors.

[0008] Step 2: Analyze the leakage magnetic field characteristics when the transformer winding is bulged and compressed, and extract the leakage magnetic field intensity characteristics to identify the winding fault types.

[0009] Step 3: Analyze the leakage magnetic field differences of faults at different positions when the transformer winding is compressed, and extract the corresponding leakage magnetic field characteristics to locate the faulty winding and the fault position.

[0010] Step 4: Analyze the characteristics of weak leakage magnetic field signals when the transformer winding is bulged, and identify the bulging fault and the normal operating state through wavelet-skewness, and locate the bulging position.

[0011] Step 5: Realize the diagnosis of transformer winding faults based on the Hippopotamus-Convolutional Neural Network model.

[0012] In the above Step 1, the manifestation forms of transformer winding deformation are that the winding bulges due to the action of radial electromagnetic force, and compresses due to the action of axial electromagnetic force; according to different deformation positions, the transformer winding faults are divided into 6 typical faults: upper-end compression, lower-end compression, upper and lower compression, upper-end bulge, middle bulge, and lower-end bulge.

[0013] In Step 1, based on the actual transformer model parameters, according to the changes in the winding spatial structure of the transformer under normal and faulty states, construct finite element models of the transformer under different operating states and arrange sensors. The transformer faulty states are divided into winding bulge and compression. In the leakage magnetic field distribution under the faulty operating state of the winding, for the bulging deformation, there is only a leakage magnetic field change in the area near the fault position, and this fluctuation is significantly affected by the distance, and there is an obvious leakage magnetic field change only near the bulging deformation; for the compression deformation, the overall leakage magnetic field waveform changes significantly, and the magnetic field fluctuation range is large, and the change of the magnetic field can be clearly reflected from the observation line.

[0014] According to the leakage magnetic field distribution law of the transformer under the faulty operating state, the closer to the fault area, the more obvious the change characteristics of the leakage magnetic field intensity. Therefore, the gap between the high-voltage and low-voltage windings is the most ideal leakage magnetic field observation area.

[0015] In Step 1, arrange sensors between the two-phase windings.

[0016] In the above Step 2, according to the characteristics of the differences in leakage magnetic field intensity when the transformer winding is compressed and bulged, extract the fault characteristic quantities reflecting the leakage magnetic field intensity.

[0017] When there is a winding bulging fault, the change amplitude of the leakage magnetic flux is small, and the leakage magnetic flux characteristics in the case of slight bulging are similar to the normal state; when there is a winding compression fault in the transformer, the change amplitude of the leakage magnetic flux is large, which can be clearly distinguished from the winding bulging and the normal state. Effective distinction is made using characteristic quantities that can identify the leakage magnetic flux amplitude, namely the root mean square error maximum deviation logarithmic root mean square error Their characteristic expressions are shown in Formulas (1) to (3) as follows:

[0018]

[0019]

[0020] Among them, the maximum leakage magnetic flux curve during normal operation of the transformer is B Ln , B Ln,i represents the magnitude of the leakage magnetic flux measured by the i-th sensor on curve B Ln ; when there is a winding fault in the transformer, the maximum leakage magnetic flux curve is B Lf , B Lf,i represents the magnitude of the leakage magnetic flux measured by the i-th sensor on curve B Lf ; the fault leakage magnetic flux curve is defined as x L = B Lf - B Ln , x L,i represents the magnitude of the leakage magnetic flux measured by the i-th sensor on curve x L ; L = 1, 2, 3, corresponding to the data of observation lines line1 to line3 respectively; n represents the number of sensors on the observation line; max{|x L |} represents the maximum value among the absolute values of x L .

[0021] Through the above three characteristic quantities, the operating state of the transformer can be divided into two types: ①. Winding compression fault; ②. Winding bulging fault or normal state.

[0022] In step 3, when the winding compression fault of the transformer occurs in different regions, the leakage magnetic flux distribution law is different, which is mainly reflected in the destruction of the leakage magnetic flux symmetry. When there is a winding compression fault in the transformer, the fault position can be divided into upper winding compression, lower winding compression, and upper and lower compression. For the differences in leakage magnetic flux symmetry reflected by different position faults, two types of characteristic quantities are selected to extract the leakage magnetic flux characteristics, namely the distribution difference degree D and the asymmetry degree γ, and their characteristic expressions are shown in Formulas (4) and (5) as follows:

[0023]

[0024] The distribution difference degree D and the asymmetry degree γ can highlight the asymmetry characteristics of the leakage magnetic flux, so as to accurately judge the fault position.

[0025] Considering that there are 4 windings in one window of the transformer, it is difficult to locate the faulty winding with the above characteristic quantities. Since the leakage magnetic flux intensity induced by different sensors varies when the winding fails, the leakage magnetic flux difference measured between different sensors can be compared to locate the faulty winding. Therefore, two types of characteristic quantities are selected: the root mean square error ratio η R and the maximum deviation ratio η M . The characteristic expressions are shown in Equations (6) and (7).

[0026]

[0027] In the formula, respectively represent the root mean square errors of the faulty leakage magnetic flux curves x L of the observation lines line1 and line3; respectively represent the maximum deviations of the faulty leakage magnetic flux curves x L of the observation lines line1 and line3.

[0028] The root mean square error ratio η R and the maximum deviation ratio η M can effectively distinguish the faults of transformer windings in different phases. Combining the differences in leakage magnetic flux intensity reflected by sensors at different positions, the faulty winding can be accurately located.

[0029] In step 4, when the transformer winding is slightly bulged, the leakage magnetic flux signal reflecting the fault characteristics is weak. Wavelet-skewness is used to extract the weak leakage magnetic flux signal and identify the fault type.

[0030] When the transformer winding is bulged, the bulging positions are divided into upper-end bulging, middle bulging, and lower-end bulging. The characteristic expressions of wavelet transform W T and skewness S are shown in Equations (8) and (9):

[0031]

[0032] Among them, t represents time, f(t) is the signal input source, corresponding to the leakage magnetic flux data x L read by the sensor; ψ * represents the complex conjugate, a is the scale parameter, and b is the translation parameter; represents the average value of the faulty leakage magnetic flux curve x L . The wavelet transform can reflect the signal fluctuation. The wavelet transform results are summed by partition, as shown in Equation (10).

[0033]

[0034] Among them, W T is the wavelet curve after wavelet transform, W T1 and W T2Sum the first half and the second half of the wavelet data within the data window respectively, and use this as the feature quantity for subsequent calculations.

[0035] Wavelet transform features can effectively distinguish the normal situation of the transformer and the winding bulging fault, and at the same time can distinguish the three fault positions during the bulging fault. Skewness can enhance the characteristic differences of asymmetric faults during the bulging fault to highlight the characteristics of the upper bulging and the lower bulging. The combination of the two can effectively identify the winding bulging fault.

[0036] In step 5, the hyperparameters of the convolutional neural network are optimized by the Hippopotamus Optimization (HO) algorithm:

[0037] The Hippopotamus Optimization (HO) algorithm is a new type of metaheuristic algorithm, which realizes algorithm optimization by simulating three prominent behavior patterns in the life of hippopotamuses. The algorithm description is as follows.

[0038] 1) Population initialization

[0039] The Hippopotamus Optimization algorithm is a population-based optimization algorithm. The population is initialized in the entire search space, and the formula is shown in Equation (11).

[0040] x ij = l j + r·(u j - l j )(11);

[0041] In the formula: x ij represents the position of the i-th hippopotamus in the j-th dimension of the search space; u j , l j represent the upper and lower boundaries of the j-th dimensional variable respectively; r is a random number with a range of [0,1].

[0042] 2) Exploration stage:

[0043] a. Random exploration stage:

[0044] In this stage, the male members of the hippopotamus compete with each other. The positions of the male hippopotamuses in the group are shown in Equation (12).

[0045]

[0046] In the formula, x i represents the position vector of the i-th hippopotamus; x best represents the position vector of the dominant hippopotamus, I 1 is a random number of 1 or 2; represents the result of the first position update.

[0047] In this stage, the young hippopotamuses will move randomly, possibly leaving the population, and at the same time causing changes in the positions of the female hippopotamuses.

[0048]

[0049] In the formula, represents the result of the second position update. h is a random number in [-1, 3]; I 2 is a random number of 1 or 2; T represents the distance between the juvenile hippopotamus and its mother, and if it exceeds 0.6, it means that the juvenile hippopotamus has separated from its mother; MG i represents the average value of randomly selected hippopotamuses; x' is an intermediate variable, represented by Equation (14); u j , l j respectively represent the upper and lower boundaries of the j-th dimensional variable.

[0050] Combining the above two position changes, the optimal position is updated.

[0051]

[0052] Among them, represents the position update result, P = P 1 , P 2 , P 3 , P 4 ; represents substituting the position information into the objective function for solution, f(x i ) represents substituting the x i position information into the objective function for solution.

[0053] b. Defense stage:

[0054] When the hippopotamus is attacked or its territory is invaded, when the distance is relatively close, the hippopotamus will quickly move towards the predator and make it retreat. The position of the predator in space is shown in Equation (16).

[0055]

[0056] In the formula, represents the result of the third position update; y represents a random position within the variable range; is a random vector with the same dimension as the position variable, and the range of each random variable in the vector is [0, 1]; f(y) represents substituting the random position information y into the objective function for solution.

[0057] p, c, d, and g are all random numbers, but with different values, p ∈ [2, 4], c ∈ [1, 1.5], d ∈ [2, 3], g ∈ [2, 4];

[0058] is a random vector representing a Levy distribution.

[0059] Update the optimal position according to Equation (15).

[0060] 3) Development stage:

[0061] The hippopotamus escapes from the predator, and the position update is shown in Equation (17).

[0062]

[0063] In the formula, represents the result of the 4th position update; s is a random variable in [0, 1]; t is the number of iterations.

[0064] Update the optimal position according to Equation (15).

[0065] The convolutional neural network is a neural network with a deep structure, which has great advantages in dealing with classification problems. The network structure mainly includes convolutional layers, pooling layers, fully connected layers, etc., and is a typical deep learning algorithm.

[0066] 1) Convolutional layer: mainly extracts features similar to the convolutional kernel in the original data. During the calculation, the convolutional kernel matrix slides in the input matrix. Each time it slides, the convolutional kernel multiplies and sums the numbers at the corresponding positions in the input matrix, and one element in the feature matrix is obtained. The sliding stride can be adjusted, and thus the convolutional matrix is obtained. After weighting the convolutional matrix, it is input into the activation function, and the output result is passed to the next layer. When performing convolutional calculations, multiple convolutional kernels can obtain multiple output matrices, the number of convolutional kernels can be controlled, and the Sigmoid function is selected as the activation function.

[0067] 2) Pooling layer: mainly reduces the amount of data. The pooling calculation also has a sliding window. Each time it slides, the maximum value of the elements in the matrix at the corresponding position in the input matrix is used as the output. This process is called pooling calculation, and the output result is passed to the next layer.

[0068] 3) Fully connected layer: The input matrix is unfolded row by row in sequence, weighted and then transmitted to the classifier, and finally the classification result is output.

[0069] When constructing a transformer fault diagnosis model, the Hippopotamus Algorithm (HO) is used to optimize the hyperparameters of the convolutional neural network CNN. The optimized parameters include the number of convolutional kernels, the initial learning rate, etc. The number of convolutional kernels and the initial learning rate are used as the position parameters of the Hippopotamus Algorithm, and the training accuracy of the convolutional neural network is used as the objective function of the Hippopotamus Algorithm. After optimization, the optimal number of convolutional kernels and the initial learning rate are output, and the convolutional neural network is parameterized accordingly, as Figure 11 shown. And the HO-CNN model is constructed accordingly to improve the generalization ability of the model.

[0070] During model training, the dataset contains transformer winding deformation faults with different windings, different fault types, different fault locations, and different fault degrees. 80% is used as the training set, and 20% is used as the test set. The training set samples and the test set samples do not overlap. The effectiveness of the HO-CNN fault diagnosis model is reflected by the prediction effect of the test set samples.

[0071] A transformer winding fault diagnosis method based on Hippopotamus-Convolutional Neural Network according to the present invention has the following technical effects:

[0072] 1) Compared with the existing transformer leakage magnetic fault diagnosis methods, the method of the present invention adds new feature quantities and data measurement positions, which can better highlight various fault characteristics and effectively solve the problem of low fault diagnosis rate in the case of slight winding faults. Traditional methods can only judge the fault type and fault location of a single transformer winding. On this basis, the present invention can locate the faulty winding and can perform more subtle winding fault diagnosis.

[0073] 2) The method of the present invention more comprehensively analyzes the influence of different winding faults on the diagnosis result and effectively solves the problem that other methods cannot locate the faulty winding, and has better application value.

[0074] 3) By reasonably distributing the leakage magnetic extraction positions, the method of the present invention effectively highlights the leakage magnetic characteristics of winding faults, adopts a variety of leakage magnetic feature extraction methods to highlight the characteristics of various winding fault types, and combines the Hippopotamus-Convolutional Neural Network algorithm to improve the accuracy of transformer winding fault diagnosis. This method is simple to implement, has obvious fault characteristics, and has a high model diagnosis accuracy. Description of the Drawings

[0075] The present invention will be further described below in conjunction with the drawings and examples;

[0076] Figure 1 is the two-dimensional finite element model of the transformer.

[0077] Figure 2 is the spatial leakage magnetic distribution diagram in the normal state.

[0078] Figure 3 is the spatial leakage magnetic distribution diagram in the fault state.

[0079] Figure 4 is the leakage magnetic distribution diagram with different degrees of bulging deformation.

[0080] Figure 5 is the leakage magnetic distribution diagram with different degrees of compression deformation.

[0081] Figure 6 is the fault type feature analysis

[0082] Figure 7 is the asymmetric fault feature analysis under compression deformation

[0083] Figure 8 For the analysis of the characteristics of the faulty winding

[0084] Figure 9 is the characteristic curve of wavelet transform.

[0085] Figure 10 is the skewness characteristic curve graph.

[0086] Figure 11 is the structure diagram of the HO-CNN model.

[0087] Figure 12 is the fault diagnosis effect diagram of different characteristic quantities.

[0088] Figure 13 is the curve of the loss function changing with the number of training times. Specific implementation manner

[0089] For the transformer winding fault diagnosis method based on the hippopotamus-convolutional neural network, first, a finite element model of the transformer in normal and faulty states is constructed in combination with the main forms of transformer winding faults, and the positions of sensors are arranged; second, according to the difference in the leakage magnetic intensity during winding bulging and compression faults, appropriate characteristic quantities are selected to extract the fault intensity characteristics to identify typical winding fault types. Then, the leakage magnetic symmetry characteristics are extracted to identify the winding compression faults in different regions, and the wavelet-skewness characteristics are combined to highlight and extract the winding bulging fault characteristics in different regions; finally, the transformer winding fault diagnosis is realized based on the hippopotamus-convolutional neural network. This method solves the problems of low accuracy of minor transformer winding faults and difficulty in locating the faulty winding.

[0090] Based on the leakage magnetic distribution characteristics of the transformer, this invention analyzes the extraction effects of different characteristic quantities on the leakage magnetic distribution curve, and uses the hippopotamus-convolutional neural network to realize the fault diagnosis of the transformer winding, and verifies the effectiveness of the proposed method through simulation.

[0091] Step 1: Analyze the main manifestation forms of transformer winding faults, construct finite element models of the transformer in normal and faulty states, and arrange the positions of sensors. When an external short-circuit fault occurs in the transformer, the current in the winding increases rapidly, and the iron core quickly reaches the magnetic saturation state, resulting in a large amount of leakage magnetic around the transformer. Under the action of the fault leakage magnetic, the winding will be subjected to electromagnetic forces far beyond the normal range, thus resulting in winding deformation faults. The main manifestation forms of transformer winding deformation are that the winding bulges under the action of radial electromagnetic forces and compresses under the action of axial electromagnetic forces. According to the different deformation positions, the transformer winding faults can be divided into 6 typical faults: upper compression, lower compression, upper and lower compression, upper bulge, middle bulge, and lower bulge, and finite element models of the transformer in different fault states and different fault degrees are constructed accordingly.

[0092] The finite element model of the transformer in the normal state is as shown in Figure 1 . Analyze the leakage magnetic characteristics of the transformer in one of the windows, as shown in Figure 2 . During the normal operation of the transformer, the leakage magnetic flux shows a central symmetric distribution and is mainly distributed in the gaps between the windings. As shown in Figure 3 , it is the leakage magnetic flux distribution in the fault operation state of the winding. For the bulging deformation, there is only a change in the leakage magnetic flux in the area near the fault location, and this fluctuation is significantly affected by the distance. There is an obvious change in the leakage magnetic flux only near the bulging deformation; for the compression deformation, the overall leakage magnetic flux waveform changes significantly, and the magnetic field fluctuation range is large. The change of the magnetic field can be clearly reflected from the observation line. According to the leakage magnetic flux distribution law of the transformer in the normal and fault operation states, the closer to the fault area, the more obvious the change characteristics of the leakage magnetic flux intensity. Therefore, the gap between the high-voltage and low-voltage windings is the most ideal leakage magnetic flux observation area. Considering that it will be difficult to locate the faulty winding only by measuring the leakage magnetic flux between the high-voltage and low-voltage windings, additional leakage magnetic flux detection is added between two phases. As shown in Figure 1 , the dotted lines in it are the placement positions of the leakage magnetic flux sensors, which are named line1 to line3 from left to right, and the 4 windings in the transformer window are named winding1 to winding4 from left to right.

[0093] Step 2: Analyze the leakage magnetic characteristics of the transformer windings in the bulging and compression states, and extract the leakage magnetic flux intensity characteristics to identify the winding fault types. The leakage magnetic flux distribution when the transformer winding bulges is as shown in Figure 4 . The fault location is that there is a bulge at the upper end of winding 2. When the bulge deforms, the leakage magnetic flux mainly changes in the fault area, and the change amplitude is limited. The smaller the fault degree, the closer the leakage magnetic flux curve is to the leakage magnetic flux curve under normal conditions.

[0094] The leakage magnetic flux distribution when the transformer winding is compressed is as shown in Figure 5 . The fault location is that the upper end of winding 2 is compressed. When the compression deforms, the change characteristics of the leakage magnetic flux are obvious, and the leakage magnetic flux intensity changes in almost all areas on the observation line. The closer to the fault center position, the greater the change degree of the leakage magnetic flux. Compared with the bulge deformation, the change of the leakage magnetic flux intensity is obvious and the characteristics are significant during the compression deformation. Since the difference in the leakage magnetic flux intensity is obvious when the winding bulges and is compressed, the leakage magnetic flux intensity characteristics can be extracted to identify the two fault types. The selected characteristic quantities are the root mean square error the maximum deviation the logarithmic root mean square error . The characteristic expressions are as shown in Eqs. (1) - (3).

[0095]

[0096]

[0097] Among them, the maximum leakage magnetic flux curves on line1 - line3 during the normal operation of the transformer are B 1n and B2n and B 3n , when there is a winding fault in the transformer, the maximum leakage magnetic flux curves are B 1f and B 2f and B 3f . Define the fault leakage magnetic flux curve as x L = B Lf - B Ln , where L = 1, 2, 3.

[0098] Taking the leakage magnetic flux characteristics of the line2 observation line for the winding 2 fault as an example, the root mean square error, maximum deviation, and logarithmic root mean square error can all reflect the difference in amplitude between the two curves, as Figure 6 shown. The curves of the three characteristic quantities under 6 typical winding faults can better distinguish between the compression fault and the bulge fault in terms of amplitude. As the degree of the fault intensifies, the distinguishing effect becomes more obvious. Considering that the leakage magnetic flux characteristics during the normal operation of the transformer are similar to those during the bulge deformation, the bulge fault and the normal condition cannot be distinguished only by the amplitude characteristics. When analyzing the characteristics through the amplitude, the transformer state can be divided into 2 types: 1. Compression fault; 2. Bulge fault or normal condition.

[0099] Step 3: Analyze the leakage magnetic flux differences of different position faults during the compression of the transformer winding, and extract the corresponding leakage magnetic flux characteristics to locate the faulty winding and the fault position. The winding compression faults are divided into upper end compression, lower end compression, and upper and lower compression situations. When faults occur at different positions, the leakage magnetic flux symmetry is different. Select two characteristic quantities that can reflect the symmetric distribution characteristics, namely the distribution difference degree D and the asymmetry degree γ, and their characteristic expressions are shown in Eqs. (4) and (5).

[0100]

[0101] Taking the leakage magnetic flux characteristics of the line1 observation line as an example, the changes in the characteristic quantities of the distribution difference degree and the asymmetry degree when the adjacent windings 1 and 2 have compression faults are as Figure 7 shown. As the degree of the fault intensifies, the asymmetric faults show opposite trends of change, and the characteristic quantities of the symmetric faults hardly change. Thus, the three typical fault regions under the compression fault can be distinguished.

[0102] When the same type of fault occurs in different windings, for example, when upper end compression faults occur in windings 1 - 4 respectively, the 4 groups of measured leakage magnetic flux data will show similar characteristic distribution laws. It is difficult to locate the faulty winding through the characteristic quantities in the above steps. Therefore, select the root mean square error ratio η R and the maximum deviation ratio η M to reflect this difference, and their characteristic expressions are shown in Eqs. (6) and (7).

[0103]

[0104] Considering that the greater the magnetic leakage intensity of the fault, the more obvious the amplitude difference shown by different observation lines. To reflect the applicability of this characteristic quantity, taking the upper end bulge fault as an example, the difference in the root mean square error ratio when the upper end bulge fault occurs under different windings is analyzed, as Figure 8 shown. According to Figure 8 the characteristic quantity, it can be clearly seen that there are obvious differences in the characteristic quantity when the windings 1 and 2 fail and when the windings 3 and 4 fail, that is, the faulty windings of different phases can be accurately located. The effect reflected by the maximum deviation is similar, so it will not be elaborated much. Based on the extraction of the root mean square error ratio and the maximum deviation ratio, the amplitude difference reflected by the three sensor arrangement positions can effectively distinguish the faults of windings 1 and 2 or the faults of windings 3 and 4, that is, the location of the faulty winding can be realized by combining the two characteristics.

[0105] Step 4: Analyze the weak magnetic leakage signal characteristics when the transformer winding bulges, identify the bulge fault and the normal operation state through wavelet-skewness, and locate the fault position. When the winding bulges slightly, the magnetic leakage curve is very similar to the normal situation, and the fault signal is weak. It is difficult to accurately locate the bulge position only through symmetry. Therefore, wavelet transform and skewness characteristic quantities are selected to extract the characteristics of weak signals. When the transformer winding bulges, the bulge position can be divided into upper end bulge, middle bulge, and lower end bulge. The characteristic expressions of wavelet transform W T and skewness S are shown in Eqs. (8) and (9).

[0106]

[0107] Among them, f(t) is the signal input source, corresponding to the magnetic leakage data read by the sensor. The wavelet transform can reflect the signal fluctuation situation, and the wavelet transform results are summed by partition, as shown in Eq. (10).

[0108]

[0109] Taking the bulge faults of windings 1 and 2 as an example, the change of the wavelet characteristic quantity of the line1 observation line is as Figure 9 shown. According to the characteristics in the figure, the end fault situation of the transformer can be clearly distinguished. W T1 reflects the deformation of the upper end bulge of the winding, and W T2 reflects the deformation of the lower end bulge of the winding. When the corresponding fault occurs, the characteristic quantity is close to -1. Since W T1 and W T2 are both calculated through W T , the two groups of data can be obtained simultaneously. Under normal conditions of the transformer, the magnitudes of W T1 and W T2 are both close to -1. Therefore, the normal situation and the bulge fault can be distinguished according to the combined state of W T1 and W T2 .

[0110] Considering that four types of distinctions need to be realized simultaneously in wavelet transform, and the identification of the middle bulge fault type completely depends on the accurate identification of the upper and lower bulge characteristics, the skewness feature quantity is increased to enhance the identification effect of asymmetric faults. The skewness can reflect the position information of the fault leakage magnetic field and measure the symmetry of the curve. As Figure 10 shown. The wavelet transform features can effectively distinguish the normal situation of the transformer and the winding bulge fault, and at the same time can distinguish the three fault positions during the bulge fault. The skewness can enhance the characteristic differences of the asymmetric faults during the bulge fault to highlight the characteristics of the upper bulge and the lower bulge. The combination of the two can effectively identify the winding bulge fault.

[0111] In summary, for the feature quantities extracted in the above steps, the root mean square error ratio and the maximum deviation ratio are the feature quantities for fault winding positioning, and the remaining seven features are the feature quantities for the identification of six typical winding faults. It is necessary to extract features from the data at three leakage magnetic field measurement positions. Since the wavelet transform features are divided into W T1 and W T2 a total of two feature quantities, there are 24 typical winding fault identification feature quantities and 2 winding positioning feature quantities.

[0112] Step Five: Analyze the algorithm principle, build a HO-CNN model, and construct a transformer winding fault diagnosis model. The Hippopotamus Optimization (HO) algorithm is a new type of meta-heuristic algorithm, which mainly realizes the optimization of the algorithm by simulating three living behaviors of the hippopotamus. Meta-heuristic algorithms generally include an exploration stage and an exploitation stage. The exploration stage of the hippopotamus algorithm is mainly divided into two parts: the random exploration stage and the defense stage. During this stage, the hippopotamus population moves randomly within the activity range, and this movement process can help the algorithm find the position area where the optimal solution may appear within the parameter range and avoid the situation of local optimization; the exploitation stage corresponds to the escape process of the hippopotamus population after being attacked and the defense is ineffective. At this time, the hippopotamus quickly moves to the nearest safe position, corresponding to the algorithm quickly obtaining the local optimal positions of each group of parameters and obtaining the optimal position of this iteration based on this.

[0113] The Convolutional Neural Network (CNN) is a neural network with a deep structure and has great advantages in dealing with classification problems. The network structure of CNN mainly includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer mainly extracts features similar to the convolutional kernel in the data, and multiple groups of convolutional kernels can extract different features in the data. The pooling layer can effectively reduce the feature dimension and retain the more significant feature data. The fully connected layer is the last part of the network, responsible for expanding the data and transmitting it to the classifier, and finally outputting the classification result. Among them, the structural parameters such as the number of convolutional layers, the number of pooling layers, and the size of the convolutional kernel can be set by the user. This process completes one training process of the network. According to the comparison between the training result and the real result, the loss function is transmitted forward, the parameters of each network layer are modified correspondingly, and the next training is carried out until the number of training times reaches the set number. During the training process, parameter settings such as the initial learning rate are involved, and the settings of these parameters will affect the training effect of the network. Improper settings may lead to problems such as overfitting of the model.

[0114] The present invention optimizes the hyperparameters of the CNN model through the HO algorithm, and the structure of the HO-CNN model is as Figure 11 shown, where the hyperparameters for training include the number of convolutional kernels, the initial learning rate, etc. The process of transformer winding fault diagnosis is as follows:

[0115] 1) Extract the maximum radial leakage magnetic flux features of the three observation lines as the original data for fault diagnosis;

[0116] 2) Convert the original data into 26 fault feature quantities according to the content in steps two to four;

[0117] 3) Normalize the feature quantities and construct a feature quantity matrix, with each matrix corresponding to a sample;

[0118] 4) Divide the samples into a training set and a test set;

[0119] 5) Select the population size, the number of iterations of the HO algorithm, search space boundaries, and initialize the population;

[0120] 6) Input the training set into the CNN model, use the hyperparameters of the CNN model as the position variables of the HO algorithm, and retrain the CNN model for each iteration until the maximum number of iterations is reached, at which time the optimal hyperparameters are output;

[0121] 7) Set the optimal hyperparameters in the CNN model to establish a transformer winding fault diagnosis model;

[0122] 8) Input the test set into the trained model to obtain the faulty winding and fault type of the transformer winding.

[0123] To verify the effectiveness of the proposed transformer winding fault diagnosis model, a finite element model was constructed based on a three-phase three-column dry-type transformer, and the model parameters are shown in Tables 1 and 2.

[0124] Table 1 Main electrical parameters of the transformer

[0125]

[0126] Table 2 Structural parameters of the transformer

[0127]

[0128] A total of 25 transformer operating states were considered in the present invention, as shown in Table 3. It includes transformer winding deformation faults with different windings, different fault types, different fault positions, and different fault degrees, and the fault degree is 3% - 20%. There are a total of 875 groups of data in the simulation samples. Among them, according to the proportion, 80% are used as the training set and 20% are used as the test set. Each group of samples contains 26 feature quantities.

[0129] Table 3 Encoding of transformer operating states

[0130]

[0131] To fully verify the diagnostic effect of the present invention on slight bulging faults, a comparison was made with existing methods. When extracting features from leakage magnetic flux for the existing methods, the feature quantities are correlation coefficient, root mean square error, maximum deviation, asymmetry degree, and distribution difference degree. The fault types are set as 7 - 12 and 25 in Table 3 and are renumbered as 1 - 7 in the order in the table. In the method of the present invention, the feature quantities are selected as the relevant feature quantities processed from the leakage magnetic flux data of line2. To reflect the effectiveness of the present invention, the same diagnostic model, that is, the CNN model, is used for fault diagnosis in both schemes, and the diagnostic effect is as Figure 12 shown. It can be clearly found from the classification effect that the existing methods have a poor diagnostic effect on slight bulging faults. As the main reason for misjudgment, they can only effectively distinguish 3 fault types in compression faults. In contrast, the present invention can effectively solve the problem of diagnostic errors and has a good discrimination effect on slight bulging faults.

[0132] To fully verify the diagnostic effect of the present invention on the location of faulty windings, the feature quantities selected by the present invention were used to compare the effects of only extracting features from the leakage magnetic flux data of line2 in the traditional method and extracting features from the leakage magnetic flux data of line1 - 3 in the present invention. The fault degree is above 5% for both, and fault diagnosis is carried out through a convolutional neural network, as shown in Table 4.

[0133] Table 4 Fault diagnosis results of different methods

[0134]

[0135] Among them, Category 1 indicates that the fault occurs in Windings 1-4, but only 7 types including 6 typical faults and normal conditions are distinguished. Category 2 indicates the faults under 25 numbers as shown in Table 6. Accaracy is the accuracy rate, representing the proportion of samples correctly predicted by the model in the total samples; F1-Score is the harmonic mean of precision and recall; G-mean is the geometric mean of recall and precision.

[0136] As can be seen from Table 4, when the faulty windings are not distinguished, the faults of different faulty windings will also affect the judgment of the fault type by traditional methods. If the faulty windings are distinguished, the diagnostic accuracy rate is lower than 70%, and the results will no longer have reference value. In contrast, the fault diagnostic accuracy rates of the present invention in both categories are relatively high, which can effectively prevent the influence of faults in different windings on the diagnostic results.

[0137] To fully verify the diagnostic effect of the model selected by the present invention, compare the fault diagnostic effects of the HO-CNN model and other models on the same sample set. Figure 13 is the loss function of the CNN and HO-CNN models. The loss function value of the HO-CNN model drops faster, reflecting its faster convergence speed and better training effect. To reflect the effect of the model of the present invention in fault diagnosis, compare the diagnostic effects of various models such as SVM, ELM, and RF, as shown in Table 5.

[0138] Table 5 Fault Diagnostic Results of Different Models

[0139]

[0140]

[0141] It can be seen that the evaluation indicators of the HO-CNN model are relatively high, and the accuracy rate can reach 99.4%. SVM and ELM are typical methods in machine learning, and their evaluation indicators in transformer fault diagnosis are relatively poor. RF is a random forest model, and its evaluation indicators are around 95%, still not achieving a good diagnostic effect. GRU and CNN are deep learning methods with deep network structures, and their effects are significantly better than the previous models from the results in Table 5. The model used in the present invention is the HO-CNN model, and the evaluation indicators can reach 99.4%. It can be seen that its diagnostic effect is more prominent after optimizing the network hyperparameters.

Claims

1. A transformer winding fault diagnosis method based on Hippo-convolutional neural network, characterized in that The following steps are involved: Step 1: Analyze the main manifestations of transformer winding faults, construct finite element models of the transformer in normal and fault states, and arrange sensor locations; Step 2: Analyze the leakage magnetic characteristics of the transformer winding in bulging and compression states, and extract the leakage magnetic intensity characteristics to identify the winding fault type; Step 3: Analyze the leakage magnetic differences of faults at different positions when the transformer winding is compressed, and extract the corresponding leakage magnetic features to locate the faulty winding and fault position; Step 4: Analyze the weak magnetic leakage signal characteristics when the transformer winding is bulging, identify the bulging fault and normal operation status through wavelet-skewness, and locate the bulging position; Step 5: Implement transformer winding fault diagnosis based on the Hippo-convolutional neural network model.

2. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In the step 1, the transformer winding deformation is manifested in that the winding is subjected to radial electromagnetic force and bulges, and is subjected to axial electromagnetic force and compresses. According to the different deformation positions, the transformer winding faults are divided into 6 typical faults: upper end compression, lower end compression, upper and lower compression, upper end bulge, middle bulge, and lower end bulge.

3. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 2 is characterized in that: In the magnetic flux leakage distribution under the winding fault operation state, for the bulge deformation, there is a magnetic flux leakage change only in the area close to the fault position, and the fluctuation is significantly affected by the distance, and there is a significant magnetic flux leakage change only near the bulge deformation; for the compression deformation, the overall magnetic flux leakage waveform changes significantly, the magnetic field fluctuation range is large, and the change of the magnetic field can be clearly reflected from the observation line; According to the leakage magnetic distribution law of the transformer during fault operation, the closer to the fault area, the more obvious the leakage magnetic intensity change characteristics are. Therefore, the gap between the high and low voltage windings is the most ideal leakage magnetic observation area.

4. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In step 1, a sensor is arranged between two phase windings.

5. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In the step 2, according to the characteristics of the difference in leakage magnetic intensity when the transformer winding is compressed or bulged, a fault characteristic quantity reflecting the leakage magnetic intensity is extracted; When there is a winding bulge fault, the leakage flux changes slightly, and the leakage flux characteristics are similar to the normal state when there is a slight bulge; when there is a transformer winding compression fault, the leakage flux changes greatly, which can be clearly distinguished from the winding bulge and the normal state; the characteristic quantity that can identify the leakage flux amplitude is used to effectively distinguish them, which are the root mean square error maximum deviation Logarithmic Root Mean Square Error Its characteristic expression As shown in formula (1) to formula (3): Among them, the maximum leakage magnetic curve during normal operation of the transformer is B Ln , B Ln,i Curve B Ln The leakage magnetic magnitude measured by the i-th sensor above; when a winding fault occurs in the transformer, the maximum leakage magnetic curve is B Lf , B Lf,i Curve B Lf The leakage magnetic field measured by the i-th sensor on the upper left; define the fault leakage magnetic field curve as x L =B Lf -B Ln , x L,i Represents the curve x L The leakage magnetic field measured by the i-th sensor on the upper left; L = 1, 2, 3, corresponding to the data of observation lines line1 to line3 respectively; n represents the number of sensors on the observation line; max{x L |} represents x L The maximum absolute value of ; Through the above three characteristics, the transformer operating status can be divided into two types: ① winding compression fault; ② winding bulging fault or normal state.

6. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In step 3, when the transformer winding compression fault occurs in different areas, the leakage magnetic distribution law is different, which is reflected in the destruction of the leakage magnetic symmetry; when the transformer winding compression fault occurs, the fault position is divided into the upper end compression of the winding, the lower end compression, and the upper and lower compression. According to the difference in leakage magnetic symmetry reflected by faults at different positions, two types of feature quantities are selected to extract the leakage magnetic features, namely, the distribution difference degree D and the asymmetry degree γ. The feature expressions are shown in formulas (4) and (5): The distribution difference D and asymmetry γ can highlight the asymmetric characteristics of leakage flux, thereby accurately determining the fault location.

7. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 6 is characterized in that: When a winding fails, the leakage magnetic intensity that can be sensed by different sensors is different. The leakage magnetic difference measured by different sensors can be compared to locate the faulty winding. For this purpose, two types of feature quantities are selected: root mean square error ratio η R and the maximum deviation ratio η M , the characteristic expressions are shown in formula (6) and formula (7); In the formula, Respectively represent the fault leakage magnetic curve x of observation line line1 and line3 L The root mean square error of Respectively represent the fault leakage magnetic curve x of observation line line1 and line3 L The maximum deviation of R and the maximum deviation ratio η M It can effectively distinguish transformer winding faults of different phases, and combined with the difference in leakage magnetic intensity reflected by sensors at different positions, it can accurately locate the faulty winding.

8. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In step 4, when the transformer winding is slightly bulging, the leakage magnetic signal reflecting the fault characteristics is weak, and the weak leakage magnetic signal is extracted by wavelet-skewness to identify the fault type; When the transformer winding bulges, the bulge position is divided into upper bulge, middle bulge, and lower bulge. The wavelet transform W T The characteristic expressions of and skewness S are shown in equations (8) and (9): Where t represents time, f(t) is the signal input source, and the corresponding sensor reads the magnetic leakage data x L ψ * represents complex conjugate, a is the scale parameter, and b is the translation parameter; Indicates the fault leakage magnetic curve x L The average value of; wavelet transform can reflect the signal fluctuation, and the wavelet transform result is partitioned and summed, as shown in formula (10); Among them, W T is the wavelet curve after wavelet transformation, W T1 and W T2 The first half and the second half of the wavelet data in the data window are summed up respectively, and used as the feature quantity for subsequent calculations.

9. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 1 is characterized in that: In step 5, the convolutional neural network hyperparameters are optimized by the Hippo algorithm (HO): The Hippopotamus Algorithm (HO) achieves algorithm optimization by simulating three prominent behavior patterns in the life of hippos. The algorithm is described as follows: 1) Population initialization: The Hippo algorithm is a population-based optimization algorithm that initializes the population in the entire search space. The formula is shown in formula (11); x ij =l j +r·(u j -l j )(11); Where: x ij represents the position of the i-th hippopotamus in the j-th dimension of the search space; u j , l j Represent the upper and lower boundaries of the j-th dimension variable respectively; r is a random number in the range of [0,1]; 2) Exploration stage: a. Random exploration phase: In this stage, male hippopotamus compete with other males; the position of male hippopotamus in the group is shown in formula (12); In the formula, x i represents the position vector of the i-th hippopotamus; x best Represents the dominant hippo position vector, I1 is a random number of 1 or 2; Indicates the first location update result; During this stage, young hippos will move randomly and may become separated from the herd, causing the positions of female hippos to change. In the formula, Indicates the result of the second position update; h is a random number in [-1,3]; I2 is a random number of 1 or 2; T indicates the distance between the juvenile hippopotamus and the mother, and if it is greater than 0.6, it means that the juvenile hippopotamus has distanced itself from the mother; MG i represents the average value of randomly selected hippos; x' is an intermediate variable, expressed by formula (14); u j , l j Represent the upper and lower boundaries of the j-th dimension variable respectively; Combining the above two position changes, the optimal position is updated; in, Indicates the position update result, P = P1, P2, P3, P4; Indicates that Substitute the position information into the objective function to solve, f(xi) means to convert x i Substitute the position information into the objective function to solve; b. Defense phase: When a hippopotamus is attacked or its territory is invaded, it will quickly move toward the predator and make it retreat if the distance is close. The position of the predator in space is shown in formula (16); In the formula, Indicates the result of the third position update; y indicates a random position within the variable range; is a random vector of the same dimension as the position variable, and the range of each random variable in the vector is [0,1]; f(y) means substituting the random position information y into the objective function for solution; p, c, d, g are all random numbers; represents a random vector with Levy distribution; 3) Development phase: The hippopotamus escapes from the predator and its position is updated as shown in equation (17); In the formula, Represents the result of the fourth position update; s is a random variable in [0,1]; t is the number of iterations.

10. The transformer winding fault diagnosis method based on Hippo-convolutional neural network according to claim 9 is characterized in that: The Hippo algorithm (HO) is used to optimize the hyperparameters of the convolutional neural network (CNN). The parameters to be optimized include the number of convolution kernels and the initial learning rate. The number of convolution kernels and the initial learning rate are used as the position parameters of the Hippo algorithm, and the training accuracy of the convolutional neural network is used as the objective function of the Hippo algorithm. After optimization, the optimal number of convolution kernels and the initial learning rate are output to set the parameters of the convolutional neural network, and the HO-CNN model is constructed to improve the generalization ability of the model. During model training, the data set contains transformer winding deformation faults with different windings, different fault types, different fault locations, and different fault degrees. 80% is used as a training set and 20% is used as a test set. The training set samples and the test set samples do not overlap. The prediction effect of the test set samples reflects the effectiveness of the HO-CNN fault diagnosis model.

Images