Transformer bushing defect comprehensive diagnosis method based on oil chromatography and ultrahigh frequency

Abstract

The invention relates to a transformer bushing defect comprehensive diagnosis method based on oil chromatography and ultrahigh frequency, and the method comprises the steps: carrying out the gas chromatography analysis of a transformer bushing, and obtaining a gas chromatography result; performing ultrahigh frequency analysis on the transformer bushing to obtain an ultrahigh frequency result; and constructing a convolutional neural network, and inputting the gas chromatography result and the ultrahigh frequency result into the convolutional neural network to obtain a diagnosis result. According to the method, the influence of oil chromatography and ultrahigh frequency on the oiled paper sleeve and the harmfulness of defect discharge on the transformer are comprehensively considered. And the extremely strong data processing capability of the convolutional neural network is utilized to analyze real-time data and analyze and diagnose influence factors of defect discharge of the oil-impregnated paper bushing, so that a conclusion is obtained for workers to handle problems in time, and the toughness of the transformer is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of electrical control, in particular to a comprehensive transformer bushing defect diagnosis method based on oil chromatography and ultra-high frequency. Background Art

[0002] Energy and electricity are the foundation of economic and social development. Safe, efficient and clean electricity supply is an important guarantee for my country's modernization process. With the continuous growth of power load demand and the large-scale construction of ultra-high voltage power grids, higher requirements are placed on the insulation level of key equipment for power transmission and transformation. Oil-impregnated paper bushings (hereinafter referred to as bushings) are key equipment widely used in power transformer outlet devices or wire wall penetrations. They are the first trial-produced insulation structures for the development of ultra-high voltage power systems. With the increase in voltage levels and the expansion of power grid scale, the market demand for bushings is increasing, and their insulation reliability is very important for the normal operation of power systems. Accurately assessing the operating status of bushings and taking efficient operation and maintenance strategies in a timely manner are of great significance to the reliable operation of power grids and energy security.

[0003] Oil-impregnated paper bushing is a relatively fragile part of the key components of power transformers. Unlike other oil-paper insulation equipment, oil-impregnated paper bushing has the characteristics of slender structure, multiple paper layers, and multiple aluminum foils. The slender plug-in electrode structure makes the insulation field strength distribution of the bushing uneven. The multi-layered winding bushing core has a series of built-in aluminum foils, which improves the uniformity of the radial electric field distribution and increases the edge electric field effect of the aluminum foil. The outstanding barrier properties of aluminum foil make the bushing insulation moisture process more special. In actual operation, the slender bushing has to withstand high voltage, high current and strong mechanical load for a long time. At the same time, it has to withstand internal and external temperature changes, rain, fog, pollution and other environmental factors. Its insulation performance is subject to severe tests.

[0004] Bushing failure is the main cause of catastrophic failure of power transformers. The harm of bushing failure is not only the unplanned shutdown of itself, but also the surrounding important equipment. According to the accident statistics of China Electric Power Research Institute, the fire accident rate of power transformers caused by insulation discharge failure of oil-impregnated paper bushings is as high as 83%. The National Power Reliability Analysis Center pointed out that bushing failure is the primary cause of unplanned shutdown of 220kV and 500kV transformers. From 2008 to 2010, bushing failure accounted for 56.5% of the total unplanned shutdown time in the country. According to the statistics of CIGRE Working Group A2.37 in 2015, 37.3% of the fire or explosion failures of transformers were caused by bushing failure. Bushing failure exposes the vulnerability of power system equipment. There are latent insulation defects of bushings in the design, manufacturing, installation and operation and maintenance processes. If they are not discovered and eliminated in time, they will gradually develop and deteriorate after commissioning, resulting in discharge problems, which often lead to serious consequences such as bushing insulation breakdown and explosion. This is enough to show that bushing insulation quality control and efficient operation and maintenance are very important for the safe operation of the power system. The construction of a guarantee system for bushing equipment quality supervision and safe operation is imperative.

[0005] With the increase of online monitoring data and test data of bushing, deep learning models have absolute advantages over traditional shallow machine learning methods in data mining. In recent years, as an emerging method in the field of machine learning, deep learning has achieved brilliant results in the fields of image and speech recognition with its powerful automatic feature extraction capabilities. In the field of fault diagnosis and status assessment, deep learning has a strong feature extraction capability and can automatically extract features from a large amount of data, reducing the dependence on expert fault diagnosis experience and information processing technology. And by establishing a deep model, it can well characterize the complex mapping relationship between information and health status, which is suitable for the needs of diversity, nonlinearity, and high-dimensional detection data status analysis under the background of big data. A series of studies have been conducted on deep learning methods in the fault diagnosis and assessment of transformer bushings. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a comprehensive diagnosis method for transformer bushing defects based on oil chromatography and ultra-high frequency, which can comprehensively diagnose both oil chromatography and ultra-high frequency through neural networks.

[0007] The present invention solves the technical problem by adopting the following technical solutions:

[0008] The comprehensive diagnosis method of transformer bushing defects based on oil chromatography and ultra-high frequency includes the following steps:

[0009] Step 1, performing gas chromatography analysis on the transformer bushing and obtaining a gas chromatography result;

[0010] Step 2, perform UHF analysis on the transformer bushing to obtain UHF results;

[0011] Step 3: Construct a convolutional neural network, input the gas chromatography results and UHF results into the convolutional neural network, and obtain the diagnosis results.

[0012] Moreover, the specific implementation method of step 1 is:

[0013] Step 1.1, calculate the gas production according to the concentration of dissolved gas in oil;

[0014] Step 1.2, calculating the gas production rate according to the gas production in step 1;

[0015] Step 1.3: Extract characteristic gases according to the gas production rate and analyze them using characteristic gas analysis method to obtain gas chromatography results.

[0016] Moreover, the calculation method of the dissolved gas concentration in step 1.1 is:

[0017]

[0018] Among them, V gi is the volume of gas dissolved in the oil when the characteristic gas reaches saturation at 101 kPa, and V is the volume of transformer oil.

[0019] Moreover, the specific implementation method of step 1.2 is:

[0020]

[0021] Among them, γ a is the absolute gas production rate; c i2 The gas content is measured for the second sampling; c i1 is the gas content measured by the first sampling; Δt is the actual running time between two sampling intervals; G is the total oil volume of the equipment; ρ is the oil density.

[0022] Moreover, the specific implementation method of step 3 is:

[0023] The percentage of hydrogen in the total amount of hydrogen hydrocarbons:

[0024] H 2 % = H 2 / (H 2 + total hydrocarbons)

[0025] The percentage of hydrocarbon gas in total hydrocarbons:

[0026] CH 4 % = CH 4 / Total hydrocarbons

[0027] C 2 H4 % = C 2 H 4 / Total hydrocarbons

[0028] C 2 H 6 % = C 2 H 6 / Total hydrocarbons

[0029] C 2 H 2 % = C 2 H 2 / Total hydrocarbons.

[0030] Moreover, the specific implementation method of step 2 is: local discharge detection is achieved by detecting ultra-high frequency electromagnetic waves. When local discharge occurs in the power cable, a pulse current will be generated on its grounding down conductor or other ground potential connecting line; the high-frequency pulse current signal flowing through the grounding down conductor or other ground potential connecting line is detected by a high-frequency current sensor to achieve live detection of local discharge of the power cable.

[0031] Moreover, the convolutional neural network in step 3 includes an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer and an output layer. The convolutional neural network realizes feature extraction of the input data set and compresses the convolutional layer output data through the convolutional layer and the pooling layer. The layer-by-layer cascading of the two completes the deep mining of the input data; the fully connected layer and the output layer flatten the matrix data of the last layer of the pooling layer into data and perform a fully connected mapping to finally generate a fixed-dimensional vector according to the specific task. The dimension represents the type of task result, and completes the classification of the task input data by the convolutional neural network.

[0032] Moreover, the convolutional layer is:

[0033]

[0034] Among them, j represents the jth convolutional surface of the convolutional layer, x represents the input data, and w j Represents the weight coefficient of the convolution kernel of the jth convolution surface, represents convolution calculation, b is bias, f(x) represents the activation function, and nonlinear mapping is performed on the data after convolution calculation. j is the convolution surface obtained after one convolution.

[0035] Moreover, the activation function includes a sigmoid function and a ReLU activation function;

[0036] Among them, the Sigmoid function is:

[0037]

[0038] The ReLU activation function is:

[0039] g(x)=max(0,x).

[0040] Moreover, the output layer is:

[0041]

[0042] Among them, Fi(x) represents the i-th input of the sample pair softmax, and P(L|x) represents the probability that the sample x is classified into category L.

[0043] The advantages and positive effects of the present invention are:

[0044] The present invention performs gas chromatography analysis on the transformer bushing and obtains gas chromatography results; performs ultra-high frequency analysis on the transformer bushing and obtains ultra-high frequency results; constructs a convolutional neural network, inputs the gas chromatography results and ultra-high frequency results into the convolutional neural network, and obtains a diagnosis result. The present invention proposes to comprehensively consider the influence of oil chromatography and ultra-high frequency on the oil-paper bushing, as well as the harm caused by its defective discharge to the transformer. The convolutional neural network uses its extremely strong data processing capability to analyze real-time data, analyze and diagnose the influencing factors of the defective discharge of the oil-paper bushing, and draw conclusions so that the staff can deal with the problem in time, which helps to improve the toughness of the transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the automatic degassing oscillator of the present invention;

[0046] Figure 2 A schematic diagram of the Agilent 7890B GC series chromatograph of the present invention;

[0047] Figure 3 A schematic diagram of the change of discharge gas content in the casing defect of the present invention;

[0048] Figure 4 It is a schematic diagram of the ultra-high frequency sensor of the present invention;

[0049] Figure 5 This is a schematic diagram of the ultra-high frequency spectrum distribution of the casing defect discharge of the present invention;

[0050] Figure 6 This is the basic structure diagram of the convolutional neural network of the present invention;

[0051] Figure 7 This is the convolution principle diagram of the present invention;

[0052] Figure 8 This is a schematic diagram of edge filling of the present invention;

[0053] Fig. 9 is the Sigmoid function image of the present invention;

[0054] Fig.10 is the ReLU function image of the present invention;

[0055] Fig.11 Schematic diagram of maximum pooling of the present invention;

[0056] Fig.12 Schematic diagram of average pooling of the present invention;

[0057] Fig.13 Schematic diagram of the fully connected layer and output layer of the present invention;

[0058] Fig.14 It is a schematic diagram of local connection and weight sharing of the present invention;

[0059] Fig.15 Schematic diagram of the training set loss function of the present invention;

[0060] Fig.16 This is a schematic diagram of the accuracy of the verification set of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below with reference to the accompanying drawings.

[0062] The comprehensive diagnosis method of transformer bushing defects based on oil chromatography and ultra-high frequency includes the following steps:

[0063] Step 1: Perform gas chromatography analysis on the transformer bushing and obtain gas chromatography results.

[0064] Gas chromatography has the advantages of high efficiency, rapidity, sensitivity and easy automation. It is widely used in various analytical tests. By detecting the content of dissolved gas components in insulating oil of power equipment by gas chromatography, latent faults of power equipment can be diagnosed. Automatic degassing oscillator is used to separate dissolved gas from oil sample, and mechanical oscillation is used for degassing. Automatic degassing oscillator is used for example Figure 1 shown.

[0065] Gas chromatograph is an analytical instrument used to separate and analyze the content of each component in a complex mixture. Its working principle is to use gas as the mobile phase to separate the components in the sample in the chromatographic column, and then detect and record them through the detector, so as to achieve qualitative and quantitative analysis of the components in the sample, such as Figure 2 as shown.

[0066] 1. Gas production

[0067] The concentration of dissolved gas in oil is generally expressed in uL / L.

[52] Requirements for dissolved gas content in oil of newly-built, newly-commissioned and in-service equipment; H requirements for bushings in operation at 330 kV and above2 The content value is 500uL / L, CH 4 Note that the value is 100uL / L, C 2 H 2 Note that the value is 1uL / L. The concentration of dissolved gas in oil is defined as

[0068]

[0069] Among them, V gi is the volume of gas dissolved in the oil when the characteristic gas reaches saturation at 101 kPa, and V is the volume of transformer oil.

[0070] 2. Gas production rate

[0071] It is difficult to judge the severity and development trend of a fault by the absolute value of the gas content result. The development of a fault is generally divided into different stages from the beginning to the serious stage. The fault energy and severity are different in different stages. Therefore, it is necessary to analyze the development trend of the fault, that is, the gas production rate at the fault point. The fault gas production rate is directly related to the fault energy and location. The absolute gas production rate is used in this article, and the calculation method is as follows:

[0072]

[0073] where γ a is the absolute gas production rate, mL / h; c i2 The gas content measured by the second sampling, μL / L; c i1 is the gas content measured by the first sampling, μL / L; Δt is the actual running time between two sampling intervals, h; G is the total oil volume of the equipment, t; ρ is the oil density, t / m 3。

[0074] 3. Characteristic gas analysis method

[0075] The percentage of hydrogen in the total amount of hydrogen hydrocarbons

[0076] H 2 % = H 2 / (H 2 + total hydrocarbons) Formula 3

[0077] The percentage of hydrocarbon gas in total hydrocarbon

[0078] CH 4 % = CH 4 / Total hydrocarbons formula 4

[0079] C 2 H 4 % = C 2 H 4 / Total hydrocarbon formula 5

[0080] C2 H 6 % = C 2 H 6 / Total hydrocarbon formula 6

[0081] C 2 H 2 % = C 2 H 2 / Total hydrocarbon formula 7

[0082] The gas production law test of different electrode touch times at 1kV simulates the process of metal tube contact shunt discharge during operation. The gas content changes in this process are as follows: Figure 3 shown.

[0083] Step 2: Perform UHF analysis on the transformer bushing to obtain UHF results.

[0084] Each partial discharge in the insulating medium will cause neutralization of positive and negative charges, accompanied by a steep current pulse, and radiate electromagnetic waves to the surrounding. Studies have shown that the spectrum characteristics of the electromagnetic waves radiated by partial discharge are related to the geometry of the partial discharge source and the dielectric strength of the discharge gap. When the discharge gap is relatively small, the discharge process is relatively short, the steepness of the current pulse is relatively large, and the ability to radiate high-frequency electromagnetic waves is relatively strong; when the dielectric strength of the discharge gap is relatively high, the breakdown process is relatively fast, the steepness of the current pulse is relatively large, and the ability to radiate high-frequency electromagnetic waves is relatively strong. Partial discharge has extremely fast discharge time characteristics in the high-pressure SF6 gas inside the GIS, and the steep pulse along the partial discharge contains frequency components from low frequency to microwave frequency bands when it rises rapidly. Therefore, when partial discharge occurs in GIS, the partial discharge situation can be evaluated by detecting the signal in the 0.3GHz to 3GHz band in its electromagnetic wave, thereby avoiding interference problems such as corona discharge that are difficult to avoid in conventional electrical testing methods, so as to improve the signal-to-noise ratio of the detection system. Since the pulses measured by this monitoring method are in the GHz frequency band, it is called Ultra High Frequency (UHF) detection method. The UHF method generally measures electromagnetic waves higher than 300MHz. For 400kV GIS equipment, the upper limit frequency can be set to 1GHz, and for 132kV equipment, it can be expanded to 1.5GHz. The current UHF upper limit frequency range is 0.3 to 3GHz. If there is noise such as air corona, the interference can be greatly eliminated by setting a filter below 500MHz.

[0085] Partial discharge can be detected by detecting UHF electromagnetic waves. Here is an example of the UHF electromagnetic wave signal of partial discharge in SF6 gas and its spectrum distribution.

[0086] When partial discharge occurs in a power cable, a pulse current is usually generated in its grounding down conductor or other ground potential connection line. The high-frequency current sensor detects the high-frequency pulse current signal flowing through the grounding down conductor or other ground potential connection line to achieve live detection of partial discharge in the power cable. The UHF sensor is a highly sensitive device used to detect partial discharge. It captures the UHF electromagnetic wave signal generated in the power equipment to obtain the insulation status and potential fault information of the equipment, such as Figure 4 shown.

[0087] Commonly used high-frequency partial discharge detection devices include: sensors, signal processing units, signal acquisition units and data processing terminals. Figure 5 shown.

[0088] Step 3: Construct a convolutional neural network, input the gas chromatography results and UHF results into the convolutional neural network, and obtain the diagnosis results.

[0089] As one of the representative algorithms of deep learning, convolutional neural network translates and classifies input information through matrix convolution calculation and deep structure feedforward neural network. The main components of a convolutional neural network are: input layer, convolution layer, activation function, pooling layer, fully connected layer and output layer. Convolutional neural network realizes feature extraction of input data set and compresses convolution layer output data through convolution layer and pooling layer. The two are cascaded layer by layer to complete the deep mining of input data. The fully connected layer and output layer flatten the matrix data of the last pooling layer into data and perform a fully connected mapping. Finally, a fixed-dimensional vector is generated according to the specific task. This dimension represents the type of task result and completes the classification of task input data by the convolutional neural network. Figure 6 This is the basic structure diagram of the convolutional neural network. Next, this article will introduce the principles and characteristics of these most basic components.

[0090] 1) Convolutional Layer

[0091] The convolution layer is the most critical component of CNN, also known as the convolution kernel or filter. It performs matrix convolution operations through the convolution kernel to complete the feature extraction of the input data. The convolution layer performs convolution operations with the data of the previous layer through a series of trainable convolution kernels, and then convolves the calculation results within a certain range through a nonlinear excitation function transformation. This is why the convolutional neural network has nonlinear fitting capabilities. Each element in the convolution kernel consists of a weight coefficient w and a bias b. Therefore, the calculation formula of the convolution layer can be expressed by Formula 8.

[0092]

[0093] j represents the jth convolutional surface of the convolutional layer, x represents the input data, and wj Represents the weight coefficient of the convolution kernel of the jth convolution surface, represents convolution calculation, b is bias, f(x) represents the activation function, and nonlinear mapping is performed on the data after convolution calculation. j is the convolution surface obtained after one convolution.

[0094] Convolution operation is the most important step of the convolution layer and the core of feature extraction. Each neuron in the convolution layer is connected to the neurons in the corresponding area of ​​the previous layer. Different neurons in the convolution surface correspond to different areas of the previous layer. These areas are also called receptive fields. The size of the convolution kernel determines the size of the receptive field. The initial value of the weight parameter in the convolution kernel is randomly generated, and the final value is obtained through training. It is not designed manually. During the convolution operation, the weight parameter remains unchanged, and the previous layer data is scanned according to a certain rule until all data participate in the convolution operation. Taking the two-dimensional convolution operation as an example, the calculation principle is as follows: Figure 7 express.

[0095] Figure 7 In the figure, the input matrix is ​​4×4, where the red and blue boxes represent the range and receptive field, which are equal to the size of the convolution kernel (3×3). The convolution kernel is convolved with the matrix represented by the receptive field according to the sliding stride of 1, completing a convolution operation, and the final output matrix size is 2×2. In this process, it is not difficult to find that when the convolution kernel scans the input matrix according to the fixed sliding stride, the edge data is used less than the intermediate data. With the increase of the input dimension and the number of convolution layers, the edge data will be underutilized or even missing. To avoid this problem, edge padding is usually performed, and the input matrix is ​​filled with 0 values ​​to expand the input matrix dimension (such as Figure 8 As shown in Figure 2, this not only makes full use of edge information, but also improves the output matrix. Therefore, the size of the output matrix of the convolution layer is determined by the input matrix size, the convolution kernel size, the sliding step size, and the edge filling dimension.

[0096] 2) Activation Function

[0097] After the convolution layer, a nonlinear activation function transformation is required to limit the calculation results to a certain range. This is why convolutional neural networks have nonlinear fitting capabilities. Common activation functions include: Sigmoid function, tanh function, ReLU function

[0098] (1) Sigmoid function

[0099] like Fig. 9 As shown, the mathematical expression of the Sigmoid function is formula 9:

[0100]

[0101] The Sigmoid function is a commonly used activation function in neural networks. This function converts the input value into an output value between 0 and 1 through a specific mathematical function formula. It is often used to map and determine the output in neural networks. The advantage of the Sigmoid function is its smooth output capability, which makes the output of the network continuously variable, facilitating the calculation and optimization of the gradient descent algorithm of the model. It was widely used in early neural networks. However, the Sigmoid function does have an obvious problem, that is, its gradient becomes very small when the output is 0 or 1, which will cause the gradient to disappear in the back propagation algorithm of the neural network, resulting in the inability to effectively train the model. Therefore, with the continuous development of the neural network structure, the Sigmoid function has gradually been replaced by other activation functions (such as ReLU, Leaky ReLU, ELU, etc.).

[0102] (2) ReLU activation function

[0103] like Fig.10 As shown, the mathematical expression of the function is formula 10,

[0104] g(x)=max(0,x) Formula 10

[0105] The derivative of the ReLU (Rectified Linear Unit) activation function is a very simple step function. When the input is greater than or equal to 0, the derivative is 1, and when the input is less than 0, the derivative is 0. The derivative of this step function is very easy to calculate, so the weight parameters can be quickly updated during the back propagation of the neural network. Especially in deep neural networks, the use of the ReLU activation function can avoid the problem of gradient disappearance, thereby improving training efficiency and accuracy. Compared with the Sigmoid activation function, the neural network using the ReLU activation function is easier to optimize, because the derivative of the ReLU activation function is always 1 when the input is greater than 0, which makes the gradient more stable and efficient during the back propagation process, thereby improving the training speed and convergence effect of the network. In addition, the ReLU activation function does not map all inputs to the same output value, so it can increase the nonlinear expression ability of the network, better adapt to complex input patterns, and further improve the performance of the network.

[0106] Based on the above advantages, the ReLU activation function has been widely used in convolutional neural networks and has become a core component of many successful models, such as AlexNet, VGG, ResNet, etc. Due to the simplicity and effectiveness of the ReLU activation function, it has become one of the most commonly used activation functions in neural networks and is widely used in various tasks and scenarios.

[0107] 3) Pooling layer

[0108] The pooling layer is a commonly used layer in convolutional neural networks. It usually follows the activation function layer. Its main function is to reduce the size of the feature map by downsampling the convolution feature map and extract the most important information in the feature, thereby reducing the amount of calculation and the number of parameters, while enhancing the fault tolerance and generalization performance of the network. Maximum pooling and average pooling are the two most commonly used methods.

[0109] (1) Max Pooling

[0110] like Fig.11 As shown in the figure, the maximum pooling selects the largest value from each local area of ​​the feature map as the pooling result of the area, which is equivalent to extracting the most important feature in the local area.

[0111] (2) Average Pooling

[0112] like Fig.12 As shown in the figure, average pooling calculates the average value from the local area as the pooling result, which is equivalent to extracting the average features of the area.

[0113] Generally speaking, maximum pooling is more commonly used than average pooling because it is more capable of extracting the most significant information in the features and helps to enhance the invariance and generalization performance of the network.

[0114] In order to further enhance the translation invariance of the convolutional neural network, the "stride" method can also be used for downsampling operations. The stride refers to the step size of each movement of the pooling operation, which can control the downsampling ratio of the feature map by the pooling operation. By adjusting the size of the stride and the size of the pooled area, the size of the network output can be controlled without affecting the network performance, achieving the flexibility and adjustability of the network. In short, the pooling layer is an important part of the convolutional neural network. Its role is to further reduce the size of the feature map through downsampling operations and feature extraction operations, enhance the invariance and generalization performance of the network, and thus improve the performance and efficiency of the network.

[0115] 4) Fully connected layer and output layer

[0116] like Fig.13 As shown in Figure 1, the fully connected layer is a common layer in convolutional neural networks. It usually appears at the end of a convolutional neural network and is used to convert high-level feature maps into classification outputs. The main feature of the fully connected layer is that each output neuron is connected to all input neurons, which results in a large number of parameters and requires a lot of training to fit the data well.

[0117] The output of the fully connected layer can be seen as a global description of the input data, which is used to represent the category label of the image or the regression target value. In image classification tasks, the fully connected layer usually maps high-level features to classification categories and outputs the score or probability of each category. For example, in image generation tasks, the fully connected layer converts high-level features into pixel values ​​of the target image. In general, the role of the fully connected layer is to perform further classification or regression operations based on the convolutional feature map to achieve a complete description of the input.

[0118] The design of the fully connected layer needs to take into account the depth and width of the network. A fully connected layer that is too deep is prone to overfitting and gradient vanishing, while a fully connected layer that is too wide will lead to overfitting and high computational costs. Therefore, when designing the fully connected layer in a convolutional neural network, it is necessary to comprehensively consider factors such as network structure, computing resources, and model performance, and make reasonable trade-offs and adjustments.

[0119] The output layer is usually the last layer of the neural network, which is used to convert the network output into a probability distribution. The principle is to perform an exponential operation on the input and then normalize it so that all output values ​​are between 0 and 1 and the sum of the total mapping is 1. In image classification tasks, the output layer often uses a softmax function to predict the confidence that the input image belongs to different categories. Each dimension of the output represents the probability of the corresponding category. It is expressed as:

[0120]

[0121] Where Fi(x) represents the i-th input of the sample to softmax, and P(L|x) represents the probability that the sample x is classified into category L.

[0122] The Softmax function is usually combined with the cross entropy loss to form the cross entropy loss function, which is a measure of the gap between the probability distribution of the model's prediction and the probability distribution of the true label. By calculating the gap between the predicted output and the true label, the cross entropy loss function can measure the performance of the model and be used to optimize the network parameters. During the training process, the smaller the value of the cross entropy loss function, the closer the model's prediction result is to the true label, and the better the network training effect. Therefore, by combining the Softmax activation function and the cross entropy loss function, the classification task can be completed more efficiently. During the back propagation process, by deriving the loss function, the error gradient of each node can be calculated, and then the parameters of the neural network can be updated according to the back propagation algorithm, thereby improving the performance of the network and making it better adapted to the task being processed.

[0123] 5) Characteristics of Convolutional Neural Networks

[0124] like Fig.14As shown in the figure, the local connection and weight sharing of convolutional neural networks are their most important features, and are also the important reasons for their success in tasks such as speech recognition and image processing.

[0125] Local connection

[0126] In convolutional neural networks, local connection means that each neuron is only connected to some of its adjacent neurons (receptive field), and the feature correlation between adjacent neurons is relatively high. Therefore, the local connection method can more effectively extract useful feature information and is more suitable for processing high-dimensional data.

[0127] Weight sharing

[0128] Weight sharing means that during the convolution process, the weight parameters of the same convolution kernel are used to process different parts of the output matrix of the previous layer, that is, the convolution kernel uses the same weight for each input position, and shares a set of weights through data in different receptive fields to reduce the number of model parameters and the risk of overfitting. In the convolution layer, the convolution kernel can be regarded as a fixed window, which is convolved with different positions in the input features to obtain different convolution results. Due to weight sharing, the number of parameters is greatly reduced and the generalization ability of the model is improved. During the training process, the weight parameters of the convolution kernel can be automatically updated and adjusted through the back-propagation algorithm to improve the model performance.

[0129] The operation rules of convolutional neural networks give them the characteristics of local connection and weight sharing. Transient data is widely distributed and changes rapidly, so the use of convolutional neural networks for processing has certain advantages. By training the network, the mapping relationship between input data and stability is obtained, which is used to realize power system stability assessment, and certain results have been achieved in practice.

[0130] According to the above-mentioned comprehensive diagnosis method for transformer bushing defects based on oil chromatography and ultra-high frequency, a simulation model was built to conduct tests to verify the effect of the present invention.

[0131] 1. Build a simulation model, obtain raw data, and generate simulation samples

[0132] 2. Construct a convolutional neural network model and record 500 simulation data and simulation labels. The training set, validation set, and test set are composed in a ratio of 3:1:1. The loss function uses CrossEntropyLoss and Adam optimization algorithm, and the hyperparameter settings in Table 1 are used.

[0133] Table 1 Convolutional neural network model hyperparameter settings

[0134]

[0135] 3. Input the training set and labels, train the hidden layer parameters of the convolutional neural network, and calculate the model loss. And test the loss and accuracy of the trained model on the validation set and test set.

[0136] like Fig.15 and Fig.16 As shown in the figure, as the number of iterations increases, the losses of the training set and the validation set continue to decrease, and the accuracy of the validation set continues to increase. After the number of iterations reaches 60, the accuracy stabilizes at around 95.02%.

[0137] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. A comprehensive diagnosis method for transformer bushing defects based on oil chromatography and ultra-high frequency, characterized by: The following steps are involved: Step 1, performing gas chromatography analysis on the transformer bushing and obtaining a gas chromatography result; Step 2, perform UHF analysis on the transformer bushing to obtain UHF results; Step 3: Construct a convolutional neural network, input the gas chromatography results and UHF results into the convolutional neural network, and obtain the diagnosis results.

2. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and ultra-high frequency according to claim 1 is characterized in that: The specific implementation method of step 1 is: Step 1.1, calculate the gas production according to the concentration of dissolved gas in oil; Step 1.2, calculating the gas production rate according to the gas production in step 1; Step 1.3: Extract characteristic gases according to the gas production rate and analyze them using characteristic gas analysis method to obtain gas chromatography results.

3. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and ultra-high frequency according to claim 2 is characterized in that: The calculation method of the dissolved gas concentration in step 1.1 is: Among them, V gi is the volume of gas dissolved in the oil when the characteristic gas reaches saturation at 101 kPa, and V is the volume of transformer oil.

4. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 2 is characterized in that: The specific implementation method of step 1.2 is: Among them, γ a is the absolute gas production rate; c i2 The gas content is measured for the second sampling; c i1 is the gas content measured by the first sampling; Δt is the actual running time between two sampling intervals; G is the total oil volume of the equipment; ρ is the oil density.

5. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 1 is characterized in that: The specific implementation method of step 3 is: The percentage of hydrogen in the total amount of hydrogen hydrocarbons: H2% = H2 / (H2+total hydrocarbons) The percentage of hydrocarbon gas in total hydrocarbons: CH4% = CH4 / total hydrocarbons C2H4%=C2H4 / total hydrocarbons C2H6% = C2H6 / total hydrocarbons C2H2%=C2H2 / total hydrocarbons.

6. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 1 is characterized in that: The specific implementation method of step 2 is: local discharge detection is achieved by detecting ultra-high frequency electromagnetic waves. When local discharge occurs in the power cable, a pulse current will be generated on its grounding down conductor or other ground potential connecting line; the high-frequency pulse current signal flowing through the grounding down conductor or other ground potential connecting line is detected by a high-frequency current sensor to achieve live detection of local discharge of the power cable.

7. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 1 is characterized in that: In step 3, the convolutional neural network includes an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer and an output layer. The convolutional neural network realizes feature extraction of the input data set and compresses the convolutional layer output data through the convolutional layer and the pooling layer. The two are cascaded layer by layer to complete the deep mining of the input data; the fully connected layer and the output layer flatten the matrix data of the last pooling layer into data and perform a fully connected mapping to finally generate a fixed-dimensional vector according to the specific task. The dimension represents the type of task result, and completes the classification of the task input data by the convolutional neural network.

8. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 7 is characterized in that: The convolutional layer is: Among them, j represents the jth convolutional surface of the convolutional layer, x represents the input data, and w j Represents the weight coefficient of the convolution kernel of the jth convolution surface, represents convolution calculation, b is bias, f(x) represents the activation function, and performs nonlinear mapping on the data after convolution calculation. j It is the convolution surface obtained after one convolution.

9. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 7 is characterized in that: The excitation function includes a sigmoid function and a ReLU excitation function; Among them, the Sigmoid function is: The ReLU activation function is: g(x)=max(0,x).

10. The transformer bushing defect comprehensive diagnosis method based on oil chromatography and UHF according to claim 7, characterized in that: The output layers are: Among them, Fi(x) represents the i-th input of the sample pair softmax, and P(L|x) represents the probability that the sample x is classified into category L.

Images