Power transformer winding loose state category identification method

Abstract

The invention provides a power transformer winding loose state category identification method. The method comprises the following steps: S1, acquiring transformer box surface vibration acceleration signals in a normal state and a loose state of a power transformer winding; s2, converting the one-dimensional time domain vibration signal into a two-dimensional time domain feature gray matrix by adopting a relative position matrix calculation method, and then performing pseudo-color processing on the two-dimensional time domain feature gray matrix to form a two-dimensional color feature map; s3, constructing a GoogLeNet network, and inputting the two-dimensional color characteristic spectrum into the GoogLeNet network for training until the training is completed; and S4, acquiring a vibration acceleration signal of the surface of the box body of the power transformer to be detected, then forming a two-dimensional color characteristic spectrum to be detected through the processing process in the step S2, and inputting the two-dimensional color characteristic spectrum to be detected into the trained GoogLeNet network for processing to obtain a loose state category identification result of the power transformer to be detected.

Description

Technical Field

[0001] The invention relates to a method for identifying faults of electric equipment, and in particular to a method for identifying the category of loose state of windings of a power transformer. Background Art

[0002] The power transformer is an important primary equipment in the power system. During the long-term operation of the power transformer, the loosening of the bolts on the pressure plate at the end of the winding will cause the winding to loosen, thus affecting the normal operation of the transformer. If the early fault is not discovered and handled in time, when the transformer encounters a short-circuit current shock, the winding coil will be subjected to a huge short-circuit electromagnetic force, generating complex mechanical stress. Due to the cumulative effect of this stress, the winding may be dislocated or even severely deformed under multiple short-circuit shocks, which may eventually lead to irreversible transformer accidents. Therefore, the state classification and identification of the power transformer winding is of great significance for judging the mechanical operation state of the transformer winding, monitoring and eliminating early transformer faults, and maintaining the normal operation of the transformer.

[0003] At present, the use of vibration signal detection method to realize the mechanical state classification of power transformer windings is mainly through extracting one-dimensional vibration signal characteristic values ​​(including peak-to-peak value, mean value, steepness value, fundamental frequency amplitude, main frequency) or two-dimensional characteristic spectrum information, and using machine learning algorithms or deep learning for classification and identification. Among them, machine learning algorithms include support vector machine, naive Bayes classification, decision tree classification, and k-nearest neighbor algorithm. These methods do not require the use of a large amount of data, and achieve the purpose of classification by mining the similarity between signal characteristic values ​​for data fitting. Deep learning algorithms are mainly manifested as neural network learning, which is a mathematical model constructed by a series of hierarchical functions that linearly combine variables and perform nonlinear transformations. For vibration signal classification and identification, its essence is to input the characteristic information of vibration signals in different states with labels into the neural network for training, and then input the test set into the trained model to realize the classification and identification of transformer windings in different states.

[0004] At present, in the research on transformer winding state recognition, neural network methods are mainly used to classify one-dimensional data feature information, including 1D-CNN, LSTM, BP neural network, etc. When extracting features from one-dimensional vibration signal data, the time correlation of the signal is often ignored, resulting in insufficient data information extraction, resulting in low accuracy of transformer winding state classification and recognition.

[0005] In recent years, convolutional neural networks (CNNs) have developed rapidly, mainly including LeNet networks, AlexNet networks, VGG networks, ResNet residual neural networks, etc. These traditional convolutional neural networks will have problems such as gradient explosion, overfitting or non-convergence during training, which will also lead to poor signal classification and recognition effects.

[0006] Therefore, in order to solve the above technical problems, it is urgent to propose a new technical means. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a method for identifying the category of loose state of power transformer windings, which converts the time series of vibration acceleration signals characterizing the loose state of the power transformer into a two-dimensional color feature map, and then uses the GoogLeNet network for training and classification and identification, which can extract vibration signal features at a deeper level and retain time series correlation. Compared with traditional convolutional networks, the GoogLeNet network can efficiently utilize data resources and extract more features under the same data resources. Therefore, using the GoogLeNet network to classify and identify the two-dimensional feature map of the vibration signal can obtain a higher accuracy.

[0008] The present invention provides a method for identifying the loose state category of a power transformer winding, comprising the following steps:

[0009] S1. Obtaining the transformer case surface vibration acceleration signal when the power transformer winding is in normal state and loose state;

[0010] S2. Using the relative position matrix calculation method, the one-dimensional time domain vibration signal is converted into a two-dimensional time domain feature grayscale matrix, and then the two-dimensional time domain feature grayscale matrix is ​​pseudo-colored to form a two-dimensional color feature map;

[0011] S3. construct a GoogLeNet network, which includes a convolution layer, a maximum pooling layer, an Inception module, an average pooling layer, a loss output module and a fully connected layer from input to output, and input the two-dimensional color feature map into the GoogLeNet network for training until the training is completed;

[0012] S4. Obtain the vibration acceleration signal of the surface of the power transformer box to be tested, and then form a two-dimensional color feature map to be tested through the processing process of step S2, and input the two-dimensional color feature map to be tested into the trained GoogLeNet network for processing to obtain the loose state category recognition result of the power transformer to be tested.

[0013] Further, step S2 specifically includes:

[0014] S21. Perform z-score normalization on the time series of surface vibration acceleration of the power transformer box;

[0015] S22. Performing data dimension reduction processing on the time series of the normalized vibration acceleration signal using a segmented aggregation approximation method;

[0016] S23. Constructing a relative position matrix, converting the time series of the vibration acceleration signal after dimensionality reduction into a two-dimensional matrix;

[0017] S24. Convert the relative position matrix into a gray value matrix, and then convert the gray value matrix into a color two-dimensional feature map using a pseudo color processing method.

[0018] Furthermore, the z-score normalization process of the time series of the surface vibration acceleration of the power transformer box specifically includes:

[0019] Determine the time series X of the vibration acceleration signal: X={x1,x2,…,x i ,…,x n}; xi represents the amplitude of vibration acceleration at the i-th moment;

[0020] The vibration acceleration signal is normalized by z-score to obtain the standard normal distribution form:

[0021]

[0022] Where: μ represents the average value of the vibration acceleration signal time series X, and σ represents the standard deviation of the time series X.

[0023] Furthermore, the time series of the normalized vibration acceleration signal is subjected to data dimension reduction processing using a segmented aggregation approximation method, specifically including:

[0024] Set the dimension reduction factor k and determine the new time series X': X' = {x'1, x'2, ..., x' i ,…,x' n}:

[0025]

[0026] Further, constructing the relative position matrix specifically includes:

[0027]

[0028] Further, the relative position matrix is ​​converted into a gray value matrix, and then the gray value matrix is ​​converted into a color two-dimensional feature map using a pseudo color processing method, specifically including:

[0029] The relative position matrix is ​​processed using the maximum and minimum normalization method:

[0030] Among them, M e,ij Represents the element in the i-th row and j-th column of the relative position matrix M, M e,max Represents the maximum value of the vibration acceleration signal in the relative position matrix, M e,minis the minimum value of the vibration acceleration signal in the relative position matrix, F ij Represents the gray value, represented by the gray value F ij Construct gray value matrix F;

[0031] The mapping rule is set, and each element in the gray value matrix F is used as an index value, and the corresponding R, G, and B intensity values ​​are searched in the color table with the index value.

[0032] The beneficial effects of the present invention are as follows: through the present invention, the time series of the vibration acceleration signal characterizing the loose state of the power transformer is converted into a two-dimensional color feature map, and then the GoogLeNet network is used for training and classification and recognition, which can extract the vibration signal features at a deeper level and retain the time series correlation. Compared with the traditional convolutional network, the GoogLeNet network can efficiently utilize data resources and can extract more features under the same data resources. Therefore, using the GoogLeNet network to classify and recognize the two-dimensional feature map of the vibration signal can obtain a higher accuracy rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 It is a flow chart of the present invention.

[0035] Figure 2 Distribution diagram of vibration acceleration sensors for power transformer load test.

[0036] Figure 3 Schematic diagram of the GoogLeNet network structure.

[0037] Figure 4 The figure is a graph showing the accuracy rate and loss of identifying loose windings of a power transformer according to the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described in detail below:

[0039] The present invention provides a method for identifying the loose state category of a power transformer winding, comprising the following steps:

[0040] S1. Obtain the vibration acceleration signal of the transformer case surface when the power transformer winding is in normal state and loose state; wherein, the acceleration sensor is pasted on the side of the transformer case, and the sensor is evenly distributed. The signal acquisition instrument is used to respectively collect the vibration signals of the case side surface corresponding to the normal and loose states of the transformer winding, and the vibration acceleration signal is transmitted to the PC through the signal acquisition system to summarize and store its vibration signal discrete time series data. The measurement point layout diagram of the power transformer acceleration vibration sensor is as shown in the attached figure. Figure 2 As shown;

[0041] S2. Using the relative position matrix calculation method, the one-dimensional time domain vibration signal is converted into a two-dimensional time domain feature grayscale matrix, and then the two-dimensional time domain feature grayscale matrix is ​​pseudo-colored to form a two-dimensional color feature map;

[0042] S3. construct a GoogLeNet network, which includes a convolution layer, a maximum pooling layer, an Inception module, an average pooling layer, a loss output module and a fully connected layer from input to output, and input the two-dimensional color feature map into the GoogLeNet network for training until the training is completed;

[0043] S4. Obtain the vibration acceleration signal on the surface of the power transformer box to be tested, and then form a two-dimensional color feature map to be tested through the processing process of step S2, and input the two-dimensional color feature map to be tested into the trained GoogLeNet network for processing to obtain the loose state category recognition result of the power transformer to be tested. Through the above method, the time series of the vibration acceleration signal characterizing the loose state of the power transformer is converted into a two-dimensional color feature map, and then the GoogLeNet network is used for training and classification recognition, which can extract the vibration signal features at a deeper level and retain the time series correlation. Compared with the traditional convolutional network, the GoogLeNet network can efficiently utilize data resources and can extract more features under the same data resources, so the GoogLeNet network can classify and recognize the two-dimensional feature map of the vibration signal to obtain a higher accuracy rate.

[0044] In this embodiment, step S2 specifically includes:

[0045] S21. Perform z-score normalization on the time series of surface vibration acceleration of the power transformer box;

[0046] S22. Performing data dimension reduction processing on the time series of the normalized vibration acceleration signal using a segmented aggregation approximation method;

[0047] S23. Constructing a relative position matrix, converting the time series of the vibration acceleration signal after dimensionality reduction into a two-dimensional matrix;

[0048] S24. Convert the relative position matrix into a gray value matrix, and then convert the gray value matrix into a color two-dimensional feature map using a pseudo color processing method.

[0049] Specifically: the z-score normalization process of the time series of the surface vibration acceleration of the power transformer box specifically includes:

[0050] Determine the time series X of the vibration acceleration signal: X={x1,x2,…,xi ,…,x n}; x i represents the amplitude of vibration acceleration at the i-th moment;

[0051] The vibration acceleration signal is normalized by z-score to obtain the standard normal distribution form:

[0052]

[0053] Where: μ represents the average value of the vibration acceleration signal time series X, and σ represents the standard deviation of the time series X.

[0054] The time series of the normalized vibration acceleration signal is processed by using the segmented aggregation approximation method to reduce the dimension of the data, which specifically includes:

[0055] Set the dimension reduction factor k and determine the new time series X': X' = {x'1, x'2, ..., x' i ,…,x' n}:

[0056]

[0057] Constructing the relative position matrix specifically includes:

[0058]

[0059] Among them: every two time stamps of the vibration signal time series are connected by the M matrix to obtain their relative positions. Each row and column of the M matrix takes a certain time stamp as a reference point and contains relevant information of the entire vibration signal time series.

[0060] The relative position matrix is ​​converted into a gray value matrix, and then the gray value matrix is ​​converted into a color two-dimensional feature map using a pseudo color processing method. Specifically, the following steps are performed:

[0061] The relative position matrix is ​​processed using the maximum and minimum normalization method:

[0062] Among them, M e,ij Represents the element in the i-th row and j-th column of the relative position matrix M, M e,max Represents the maximum value of the vibration acceleration signal in the relative position matrix, M e,min is the minimum value of the vibration acceleration signal in the relative position matrix, F ij Represents the gray value, represented by the gray value F ij Construct gray value matrix F;

[0063] The mapping rule is set, and each element in the gray value matrix F is used as an index value, and the corresponding R, G, and B intensity values ​​are searched in the color table with the index value.

[0064] like Figure 3 As shown, the GoogLeNet network of the present invention includes a convolutional layer, a maximum pooling layer, an Inception module, an average pooling layer, a loss output module and a fully connected layer from input to output;

[0065] Among them, the convolution layer uses convolution operation to extract features from the two-dimensional feature map;

[0066] The maximum pooling layer is used to perform the maximum pooling operation on the extracted features, thereby compressing and fusing information; the Inception module is used to perform convolution decomposition, splitting a larger n×n two-dimensional convolution into two smaller one-dimensional convolutions n×1 and 1×n to form a feature tensor, and the average pooling layer is used to reduce the dimension of the feature tensor. The loss output module and the fully connected layer are used to further process the information after the dimensionality reduction process and finally output the classification result. This is the existing technology.

[0067] The essence of the convolution operation on the two-dimensional feature map is to extract feature information from the image. The operation process is that the convolution kernel matrix of a specific size slides on the original input according to the set step size. Each slide performs a two-dimensional convolution operation with the area on the input feature map. The result of the operation is to project the information in the receptive field to a pixel point in the output feature map. The two-dimensional convolution operation is as follows:

[0068]

[0069] Where A is the input matrix and B is the convolution kernel matrix, with dimensions of M×N and m×n respectively.

[0070] After the convolution operation, the two-dimensional feature map is input to the maximum pooling layer for downsampling. For downsampling of an A matrix, it needs to be divided into blocks first, and the ijth block can be expressed as:

[0071]

[0072] then The maximum downsampling of is defined as:

[0073]

[0074] The average pooling layer requires average downsampling calculation, which is defined as:

[0075]

[0076] The softmax function is defined as:

[0077]

[0078] In this embodiment, GoogLeNet V3 convolutional neural network is selected for vibration signal two-dimensional graph training, and the specific network parameters are set as shown in Table 1:

[0079] Table 1 GoogLeNet V3 network parameter settings

[0080]

[0081]

[0082] The two-dimensional feature maps of the vibration signals of the transformer windings in the loose state and the normal state are divided into a training set and a test set according to a ratio of 4:1. The training set is input into the GoogLeNet V3 convolutional neural network, and the order of the two-dimensional feature images is shuffled. After the model is trained, the test set is input into the trained network to obtain the classification recognition accuracy and loss rate result graph, as shown in the attached figure. Figure 4 ; In the training process of the GoogLeNet V3 convolutional neural network model in this embodiment, the maximum number of iterations is 232 times, the training cycle is four rounds, the number of iterations in each round is 58 times, and the learning rate is 0.01. Different one-dimensional signal conversion two-dimensional image processing methods are used in combination with different convolutional network training models to conduct comparative experiments. The classification and recognition accuracy of different training models is shown in Table 2 below:

[0083] Table 2 Classification and recognition accuracy of different types of model combinations

[0084]

[0085] It can be seen from the results in Table 2 that in terms of one-dimensional vibration signal data processing, compared with the Markov transition matrix (MFT) and the Gram angular field matrix (GADF), the relative position matrix (RPM) further retains the time correlation in converting one-dimensional time series to two-dimensional images, and improves the image recognition effect. Compared with the ResNet50 residual neural network, the GoogLeNet V3 convolutional neural network is more precise and effective in extracting image information, while reducing the data redundancy, and its classification and recognition accuracy reaches 98.52%.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying the loose state category of a power transformer winding, characterized in that: The following steps are involved: S1. Obtaining the transformer case surface vibration acceleration signal when the power transformer winding is in normal state and loose state; S2. Using the relative position matrix calculation method, the one-dimensional time domain vibration signal is converted into a two-dimensional time domain feature grayscale matrix, and then the two-dimensional time domain feature grayscale matrix is ​​pseudo-colored to form a two-dimensional color feature map; S3. construct a GoogLeNet network, which includes a convolution layer, a maximum pooling layer, an Inception module, an average pooling layer, a loss output module and a fully connected layer from input to output, and input the two-dimensional color feature map into the GoogLeNet network for training until the training is completed; S4. Obtain the vibration acceleration signal of the surface of the power transformer box to be tested, and then form a two-dimensional color feature map to be tested through the processing process of step S2, and input the two-dimensional color feature map to be tested into the trained GoogLeNet network for processing to obtain the loose state category recognition result of the power transformer to be tested.

2. The method for identifying the loose state category of a power transformer winding according to claim 1 is characterized in that: Step S2 specifically includes: S21. Perform z-score normalization on the time series of surface vibration acceleration of the power transformer box; S22. Performing data dimension reduction processing on the time series of the normalized vibration acceleration signal using a segmented aggregation approximation method; S23. Constructing a relative position matrix, converting the time series of the vibration acceleration signal after dimensionality reduction into a two-dimensional matrix; S24. Convert the relative position matrix into a gray value matrix, and then convert the gray value matrix into a color two-dimensional feature map using a pseudo color processing method.

3. The method for identifying the loose state category of a power transformer winding according to claim 2 is characterized in that: The z-score normalization process of the time series of the surface vibration acceleration of the power transformer box specifically includes: Determine the time series X of the vibration acceleration signal: X={x1,x2,…,x i ,…,x n }; xi represents the amplitude of vibration acceleration at the i-th moment; The vibration acceleration signal is normalized by z-score to obtain the standard normal distribution form: Where: μ represents the average value of the vibration acceleration signal time series X, and σ represents the standard deviation of the time series X.

4. The method for identifying the loose state category of a power transformer winding according to claim 3 is characterized in that: The time series of the normalized vibration acceleration signal is processed by using the segmented aggregation approximation method to reduce the dimension of the data, which specifically includes: Set the dimension reduction factor k and determine the new time series X': X' = {x'1, x'2, ..., x' i ,…,x' n }:

5. The method for identifying the loose state category of a power transformer winding according to claim 4 is characterized in that: Constructing the relative position matrix specifically includes:

6. The method for identifying the loose state category of a power transformer winding according to claim 5, characterized in that: The relative position matrix is ​​converted into a gray value matrix, and then the gray value matrix is ​​converted into a color two-dimensional feature map using a pseudo color processing method. Specifically, the following steps are performed: The relative position matrix is ​​processed using the maximum and minimum normalization method: Among them, M e,ij Represents the element in the i-th row and j-th column of the relative position matrix M, M e,max Represents the maximum value of the vibration acceleration signal in the relative position matrix, M e,min is the minimum value of the vibration acceleration signal in the relative position matrix, F ij Represents the gray value, represented by the gray value F ij Construct gray value matrix F; The mapping rule is set, and each element in the gray value matrix F is used as an index value, and the corresponding R, G, and B intensity values ​​are searched in the color table with the index value.

Images