Power transformer fault diagnosis method based on improved graph convolutional network

Abstract

The invention relates to the technical field of power equipment fault diagnosis, in particular to a power transformer fault diagnosis method and system based on an improved graph convolutional network, and the method comprises the steps: obtaining a power transformer fault data set, and carrying out the preprocessing; an improved graph convolution network structure is constructed, and calculation of a graph convolution layer is improved through a self-circulation and adjacent matrix normalization method; dividing the power transformer fault data set into a training set and a test set, and performing sparse processing on an adjacent matrix; initializing the structure and parameters of the graph convolutional network; and training the graph convolutional network model by adopting a back propagation algorithm, and inputting an optimized result into a test set for verification to obtain a power transformer fault diagnosis result. According to the invention, the accuracy and efficiency of fault diagnosis of the power transformer are improved, the similarity information of the historical data and the current sample is fully utilized, and the defects that the complex nonlinear relationship is difficult to capture and the historical data is not fully utilized in the existing method are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a power transformer fault diagnosis method and system based on an improved graph convolutional network. Background Art

[0002] Power transformers are key equipment in substations and power plants. Their operating status is directly related to the safety of the power system and the quality of power. Once a fault occurs, it may cause local or large-scale power outages, resulting in huge economic losses. Therefore, it is of great significance to accurately diagnose power transformer faults.

[0003] At present, most large transformers are oil-immersed transformers. When they fail, they will release a large amount of dissolved gas. Dissolved gas analysis (DGA) is an important indicator for diagnosing faults. Existing fault diagnosis methods based on DGA are mainly divided into two categories: distance-based methods and model-based methods. Distance-based methods, such as case-based reasoning, expert systems, k-nearest neighbors (KNN) and twin neural networks, diagnose by calculating the similarity between the samples to be classified and the historical samples. Although these methods use historical data, they are difficult to capture the complex nonlinear relationship between dissolved gas and fault type, and their accuracy is limited; model-based methods, such as support vector machines, multi-layer perceptrons, extreme gradient boosting trees and deep learning methods (such as convolutional neural networks, deep belief networks, capsule neural networks), can mine the complex nonlinear relationship between dissolved gas and fault type, but fail to fully utilize the similarity information between historical data and current samples, and the performance improvement is limited.

[0004] The information disclosed in this background technology section is only intended to deepen the understanding of the overall background technology of the present disclosure, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art known to those skilled in the art. Summary of the invention

[0005] The present invention provides a power transformer fault diagnosis method and system based on an improved graph convolutional network, thereby effectively solving the problems pointed out in the background technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A power transformer fault diagnosis method based on an improved graph convolutional network, the method comprising:

[0008] Acquire a power transformer fault data set and perform preprocessing, wherein the data set includes sample data of different operating states and fault types;

[0009] Construct an improved graph convolutional network structure and improve the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods;

[0010] Dividing the power transformer fault data set into a training set and a test set, and performing sparse processing on the adjacency matrix;

[0011] Initialize the structure and parameters of the graph convolutional network, and build a graph convolutional network model including at least one graph convolutional layer and at least one dense layer;

[0012] The back-propagation algorithm is used to train the graph convolutional network model, and the optimized results are input into the test set for verification to obtain the power transformer fault diagnosis results.

[0013] Further, the preprocessing includes:

[0014] Select characteristic gases and perform data preprocessing to obtain normalized characteristic values;

[0015] The dissolved gas content and new features are mapped to the specified interval using the min-max normalization method.

[0016] Furthermore, an improved graph convolutional network structure is constructed, and the calculation of the graph convolutional layer is improved through self-loop and adjacency matrix normalization methods, including:

[0017] Constructing an adjacency matrix, wherein the adjacency matrix represents a similarity measure between historical data and current samples;

[0018] Performing a self-loop operation in the adjacency matrix, each node contains its own feature information during calculation;

[0019] The adjacency matrix is ​​normalized, and the output features are calculated using the improved graph convolution layer through the input graph matrix and the normalized adjacency matrix.

[0020] Furthermore, the normalization processing formula is:

[0021]

[0022] Where A″ is the normalized adjacency matrix, D is the diagonal node degree matrix of the adjacency matrix A′, E is the diagonal node degree matrix of the adjacency matrix A, A′ is the result of the self-loop operation, A′=A+I, where I is the unit matrix.

[0023] Further, the structure and parameters of the graph convolutional network are initialized, and a graph convolutional network model including at least one graph convolutional layer and at least one dense layer is constructed, including:

[0024] Taking graph structure data as input, we use intrinsic topology to generate intermediate features, and use the weights represented to linearly combine the feature vector of each node with the feature vectors of adjacent nodes to obtain a mixed feature matrix.

[0025] In the first graph convolution layer, the output data of the first layer is obtained by multiplying the weight matrix and adding the bias vector, choosing ReLU as the activation function;

[0026] The output data of the first layer is input into the second graph convolution layer, and the output data of the second layer is obtained through linear combination and activation function processing;

[0027] Add two dense layers after the graph convolutional layer, check and process the data in the third layer, and use the weight matrix, bias vector and activation function to get the output data;

[0028] The Softmax function is used as the activation function in the fourth layer to process the output data of the third layer to obtain the fault type output by the improved GCN.

[0029] Furthermore, a back-propagation algorithm is used to train the graph convolutional network model, including: forward excitation propagation and reverse weight update. In the forward excitation propagation, the input variables are processed by multiple graph convolutional layers and then passed to the dense layer, and the labels of the samples are output.

[0030] Furthermore, the loss function is calculated using the fault diagnosis result and the actual result. The loss function is:

[0031]

[0032] Among them, L(θ) is the loss function, C is the number of categories, that is, the number of possible classifications in the classification problem, and v c is the weight of category C, y i is the actual label of the i-th sample, (y i =c) is the indicator function. When the actual label of the i-th sample is c, (y i =c) = 1, otherwise 0, s i is the probability that the i-th sample is predicted to be the correct category, x i is the score of the i-th sample in the input layer of the Softmax function, x c is the score of category c in the Softmax function input layer, It is the sum of the index scores for all categories.

[0033] Furthermore, after the loss function is established, the graph convolutional network is globally optimized, and the parameter θ is updated as follows:

[0034]

[0035] x l =∑θ·x l-1 +b I

[0036]

[0037] Among them, μ is the learning rate, Q is the number of samples in each layer, and x l is the forward propagation function, l is the number of layers of the neural network, E D is the Euclidean distance between the predicted and actual values ​​for each mini-batch.

[0038] Furthermore, the reverse weight updating includes:

[0039] The loss between the network output and the actual label is calculated through forward propagation, and the error is transferred from the output layer to the intermediate layer and input layer layer by layer using the chain rule to calculate the gradient of each layer;

[0040] Using the gradient descent method to update the weight matrix and the bias vector according to the calculated gradient;

[0041] Repeat the above process, iteratively train the model until the loss function converges, and use the test set to evaluate the performance of the graph convolutional network.

[0042] A power transformer fault diagnosis system based on an improved graph convolutional network, the system comprising:

[0043] A fault data set acquisition module is used to acquire a power transformer fault data set and perform preprocessing, wherein the data set includes sample data of different operating states and fault types;

[0044] The network structure building module builds an improved graph convolutional network structure and improves the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods;

[0045] A matrix sparse processing module, which divides the power transformer fault data set into a training set and a test set, and performs sparse processing on the adjacency matrix;

[0046] The model basic setting module initializes the structure and parameters of the graph convolutional network and builds a graph convolutional network model including at least one graph convolutional layer and at least one dense layer;

[0047] The diagnosis result output module uses the back-propagation algorithm to train the graph convolutional network model, and inputs the optimized results into the test set for verification to obtain the power transformer fault diagnosis results.

[0048] The technical solution of the present invention can achieve the following technical effects:

[0049] The accuracy and efficiency of power transformer fault diagnosis are improved, and the similarity information of historical data and current samples is fully utilized, which overcomes the shortcomings of existing methods that are difficult to capture complex nonlinear relationships and fail to fully utilize historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 It is a flowchart of a method for fault diagnosis of power transformer based on improved graph convolutional network;

[0052] Figure 2 Schematic diagram of the process of building an improved graph convolutional network structure and performing self-loop and adjacency matrix normalization;

[0053] Figure 3 Schematic diagram of the process of initializing the structure and parameters of the graph convolutional network. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0056] Embodiment 1

[0057] like Figure 1 As shown, the present invention provides a method for diagnosing power transformer faults based on an improved graph convolutional network, the method comprising:

[0058] S1: Obtain a power transformer fault data set and perform preprocessing. The data set includes sample data of different operating states and fault types;

[0059] Specifically, through cooperation with State Grid Corporation of China and related research institutions, sample data covering seven types of states of transformers with a voltage level of 220kV, including normal operation, low-temperature thermal fault, medium-temperature thermal fault, high-temperature thermal fault, partial discharge, low-energy discharge and high-energy discharge, were collected.

[0060] S2: Construct an improved graph convolutional network structure and improve the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods;

[0061] Specifically, by introducing an improved graph convolutional network structure, useful feature information can be extracted more effectively from the power transformer fault data set and the feature learning ability can be improved. The graph convolutional network is good at processing graph structure data and can mine the complex relationship between nodes (fault samples), thereby improving the model's learning and recognition ability of fault features. The introduction of the self-loop mechanism ensures that each node retains its own characteristics in the convolution calculation, prevents information loss, and improves the robustness and stability of the model. In addition, the normalization of the adjacency matrix eliminates the influence of different node degrees on the convolution operation, prevents the scale of the eigenvalue from changing during the convolution process, and ensures the stability and consistency of the calculation. The sparse processing of the adjacency matrix further reduces the computational complexity and improves the training efficiency of the graph convolutional network. Efficient calculation not only speeds up the training speed of the model, but also achieves good performance on large-scale data sets. These improvements enable the graph convolutional network to more accurately capture the intrinsic connection between power transformer fault samples, improve the accuracy and overall effect of fault diagnosis, and provide a more reliable and efficient solution for power transformer fault diagnosis in practical applications.

[0062] S3: Divide the power transformer fault data set into a training set and a test set, and perform sparse processing on the adjacency matrix;

[0063] Specifically, by dividing the model into training sets and test sets, we can ensure that the model only uses training data during training and performs independent evaluation on the test set to avoid data leakage, thereby objectively verifying the generalization ability of the model. Sparse processing of the adjacency matrix can significantly reduce computational complexity, retain only connections with higher similarity, reduce computing resource consumption, and improve training efficiency. This ensures that the model can be trained more quickly and focus on key features, thereby improving the robustness of the model and its performance in practical applications. At the same time, the data set is reasonably divided to ensure that samples of different fault types are evenly distributed in the training set and test set, avoiding the model from being biased towards a specific fault type, and comprehensively improving the accuracy and reliability of fault diagnosis.

[0064] S4: Initialize the structure and parameters of the graph convolutional network, and build a graph convolutional network model including at least one graph convolutional layer and at least one dense layer;

[0065] Specifically, the graph convolution layer extracts node features from the graph structure data through convolution operations, aggregates the information of each node and its neighboring nodes, thereby capturing the complex relationship between samples and effectively reflecting the intrinsic connection of the power transformer fault data. The dense layer is used to further integrate these features, perform nonlinear transformation and feature combination, and finally achieve the classification task and output the probability distribution of the fault type. This design ensures that the model can not only make full use of the graph structure characteristics for feature learning, but also realize efficient and accurate fault diagnosis through the classification layer, thereby improving the overall performance and accuracy of power transformer fault diagnosis.

[0066] S5: Use the back-propagation algorithm to train the graph convolutional network model, and input the optimized results into the test set for verification to obtain the power transformer fault diagnosis results.

[0067] Specifically, by adjusting the model parameters and minimizing the loss function, the model's prediction accuracy and performance are improved; by verifying on an independent test set, the model's generalization ability and actual application performance are evaluated to ensure its stability and robustness when processing unknown data; and finally, the power transformer fault diagnosis results are output to provide a reliable basis for power system maintenance and fault warning, while at the same time, model deficiencies are discovered through test feedback information.

[0068] The present invention improves the accuracy and efficiency of power transformer fault diagnosis, fully utilizes the similarity information of historical data and current samples, and overcomes the shortcomings of existing methods that are difficult to capture complex nonlinear relationships and fail to fully utilize historical data.

[0069] As a preference of the above embodiment, the preprocessing includes:

[0070] A10: Select characteristic gas and perform data preprocessing to obtain normalized characteristic value;

[0071] A20: Use the min-max normalization method to map the dissolved gas content and new features to the specified interval.

[0072] Specifically, the data was cleaned, records with more missing values ​​were deleted and some missing values ​​were filled, and outliers were detected and eliminated using box plots and standard deviation methods. Then, the contents of dissolved gases (H2, CH4, C2H6, C2H4, C2H2) were selected as original features, and four new features (C2H2 / C2H4, CH4 / H2, C2H4 / C2H6, C2H6 / H2) were constructed using Rogers ratio coding, for a total of nine features. All features were min-max normalized and their values ​​were mapped to the [0,1] interval to eliminate the dimensional differences between features. Finally, the preprocessed data was divided into training and test sets to ensure that samples of different fault types were evenly distributed in the training and test sets, thereby providing a high-quality data foundation for subsequent graph convolutional network model training.

[0073] As a preferred embodiment of the above, Figure 2 As shown, step S2 constructs an improved graph convolutional network structure and improves the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods, including:

[0074] S21: construct an adjacency matrix, which represents the similarity measure between historical data and current samples;

[0075] S22: Perform a self-loop operation in the adjacency matrix, and each node contains its own feature information during calculation;

[0076] S23: Normalize the adjacency matrix and use the improved graph convolution layer to calculate the output features through the input graph matrix and the normalized adjacency matrix.

[0077] Specifically, first, an adjacency matrix is ​​constructed to represent the similarity measure between historical data and current samples. This step generates a matrix by calculating the similarity between samples (such as Euclidean distance or cosine similarity), in which each element represents the degree of similarity between sample pairs. Sample pairs with high similarity will have higher values ​​in the matrix, indicating that they are closely connected in the graph structure; then, a self-loop operation is performed in the adjacency matrix, that is, 1 is added to the diagonal of the adjacency matrix so that each node contains its own feature information during calculation. This operation ensures that the node not only considers the features of neighboring nodes during the convolution process, but also retains its own features, thereby preventing information loss and improving the integrity of features and the accuracy of calculation; then, the adjacency matrix is ​​normalized Normalization is performed to eliminate the impact of different node degrees on the calculation. The unnormalized adjacency matrix may cause nodes with high connectivity to occupy too much weight in the convolution, resulting in feature imbalance. The normalization process allows the features of each node to be evenly considered in the convolution calculation, avoiding the scale change of the eigenvalue, thereby ensuring the stability and consistency of the calculation. Finally, the improved graph convolution layer uses the input graph matrix (including node features) and the normalized adjacency matrix for convolution calculation. The graph convolution layer aggregates and transfers the features of each node with the features of its neighboring nodes to generate new node features. This process combines the structural information of the adjacency matrix and the feature information of the node, which can more accurately capture the complex relationship between samples and improve the performance of the model in fault diagnosis.

[0078] As a preferred embodiment of the above embodiment, the normalization processing formula is:

[0079]

[0080] Where A″ is the normalized adjacency matrix, D is the diagonal node degree matrix of the adjacency matrix A′, E is the diagonal node degree matrix of the adjacency matrix A, A′ is the result of the self-loop operation, A′=A+I, where I is the unit matrix.

[0081] Specifically, first, an adjacency matrix A is constructed to represent the similarity measure between historical data and current samples. By calculating the similarity between samples (such as Euclidean distance or cosine similarity), an adjacency matrix is ​​generated, in which each element represents the degree of similarity between sample pairs. Then, a self-loop operation is performed, that is, a self-loop is added to the adjacency matrix so that each node contains its own feature information during calculation. The self-loop operation is completed by adding the unit matrix I to the adjacency matrix A, and the adjacency matrix after the self-loop is obtained. A′=A+I. This step ensures that the information of each node itself will not be lost during the convolution process, thereby enhancing the integrity of the features. Then, the adjacency matrix A′ is normalized to eliminate the influence of different node degrees on the calculation. The normalization process uses a symmetric normalization method to make the features of each node evenly considered in the convolution calculation, avoiding the scale change of the eigenvalues ​​and ensuring the stability and consistency of the calculation. Finally, the improved graph convolution layer is used to perform convolution calculations through the input graph matrix and the normalized adjacency matrix A″. The graph convolution layer aggregates and transfers the features of each node with the features of its neighboring nodes to generate new node features. This process combines the structural information of the adjacency matrix with the feature information of the node, which can more accurately capture the complex relationship between samples and improve the performance of the model in fault diagnosis.

[0082] As a preferred embodiment of the above, Figure 3 As shown, step S4, initializing the structure and parameters of the graph convolutional network, and constructing a graph convolutional network model including at least one graph convolutional layer and at least one dense layer, including:

[0083] S41: Take graph structure data as input, use intrinsic topology to generate intermediate features, and use the weights represented to linearly combine the feature vector of each node with the feature vectors of adjacent nodes to obtain a mixed feature matrix;

[0084] S42: In the first graph convolution layer, the output data of the first layer is obtained by multiplying the weight matrix and adding the bias vector, selecting ReLU as the activation function;

[0085] S43: Input the output data of the first layer to the second graph convolution layer, and obtain the output data of the second layer through linear combination and activation function processing;

[0086] S44: Add two dense layers after the graph convolution layer, check and process the data in the third layer, and use the weight matrix, bias vector and activation function to get the output data;

[0087] S45: The Softmax function is used as the activation function in the fourth layer to process the output data of the third layer to obtain the fault type output by the improved GCN.

[0088] Specifically, first, the graph structure data is used as input, and the intrinsic topology is used to generate intermediate features. The graph structure data includes nodes and their connection relationships (adjacency matrix). By using the weights represented, the feature vector of each node is linearly combined with the feature vectors of adjacent nodes to obtain a mixed feature matrix. This step is achieved through matrix operations, in which the weight matrix is ​​used to adjust the influence of each node and neighbor node features; then, in the first graph convolution layer, the mixed feature matrix is ​​processed by multiplying the weight matrix and adding a bias vector. Select ReLU (Rectified Linear Unit) as the activation function, the data after linear transformation is nonlinearly transformed to obtain the output data of the first layer. The introduction of ReLU activation function helps to increase the nonlinear expression ability of the model and capture more complex features; then, the output data of the first layer is input to the second graph convolution layer, and the second graph convolution layer is also processed by linear combination and activation function to further extract node features. The specific process includes multiplying the output data of the first layer with the new weight matrix, adding the bias vector, and then performing nonlinear transformation through ReLU activation function to obtain the output data of the second layer; after the graph convolution layer, two dense layers are added. The third layer is the first dense layer, which is used to The data output from the graph convolution layer is checked and processed. By using the weight matrix, bias vector and activation function, the data is further transformed and combined to obtain the output data of the third layer. This step helps to integrate the features and prepare for the final classification. Finally, the fourth layer is the second dense layer. The Softmax function is used as the activation function to process the output data of the third layer. The Softmax function converts the output data into a probability distribution, so that the features of each node can be expressed as the probability of belonging to different fault types, thereby obtaining the fault type output by the improved GCN. This result can be used for power transformer fault diagnosis to help identify specific fault types.

[0089] As a preferred embodiment of the above embodiment, a back propagation algorithm is used to train the graph convolutional network model, including: forward excitation propagation and reverse weight update, forward excitation propagation, the input variables are processed by multiple graph convolution layers and then passed to the dense layer, and the labels of the samples are output.

[0090] Specifically, using the back-propagation algorithm to train the graph convolutional network model can optimize model parameters, improve the model's prediction accuracy, effectively train deep networks, capture complex nonlinear features, automatically adjust parameters to achieve optimal convergence, enhance the model's generalization ability, and process complex data structures, thereby achieving efficient, accurate, and reliable fault identification and classification in power transformer fault diagnosis tasks.

[0091] As a preferred embodiment of the above, the loss function is calculated using the fault diagnosis result and the actual result, and the loss function is:

[0092]

[0093] Among them, L(θ) is the loss function, C is the number of categories, that is, the number of possible classifications in the classification problem, and v c is the weight of category C, y i is the actual label of the i-th sample, (y i =c) is the indicator function. When the actual label of the i-th sample is c, (y i =c) = 1, otherwise 0, s i is the probability that the i-th sample is predicted to be the correct category, x i is the score of the i-th sample in the input layer of the Softmax function, x c is the score of category c in the Softmax function input layer, It is the sum of the index scores for all categories.

[0094] Specifically, the model calculates the score of each sample in the Softmax input layer. These scores represent the confidence that the sample belongs to each category. The Softmax function is used to convert the score into a probability. The actual label and the predicted probability are used to calculate the loss value of each sample through the loss function formula. The loss value reflects the difference between the model prediction and the actual result. The back-propagation algorithm is used to calculate the gradient of the loss function to the model parameters, and the parameters are updated along the gradient direction to minimize the loss function, thereby optimizing the model.

[0095] As a preferred embodiment of the above, after the loss function is established, the graph convolutional network is globally optimized, and the parameter θ is updated as follows:

[0096]

[0097] x l =∑θ·x l-1 +b I

[0098]

[0099] Among them, μ is the learning rate, Q is the number of samples in each layer, and x l is the forward propagation function, l is the number of layers of the neural network, E D is the Euclidean distance between the predicted and actual values ​​for each mini-batch.

[0100] First, we need to calculate the gradient of the loss function L(θ) with respect to the parameter θ. This process involves the application of multiple chain rules, as shown in the following formula: By expanding the loss function and deriving it layer by layer, we can get the specific expression of gradient calculation; the expansion of the loss function: Then the loss function L(θ) is used for s i The gradient of is calculated: Then calculate s i x i The gradient is: Finally calculate x i The gradient of the parameter θ, which involves updating the weights in the network, is: i =∑θ·x i-1 +b i , combining the above gradients, we get the total gradient of the loss function L(θ) with respect to the parameter θ: According to the gradient descent algorithm, update the parameters The specific expressions of the gradient are combined together to obtain the update formula:

[0101] As a preferred embodiment of the above, the reverse weight updating includes:

[0102] B10: Calculate the loss between the network output and the actual label through forward propagation, and use the chain rule to transfer the error from the output layer to the intermediate layer and input layer layer by layer, and calculate the gradient of each layer;

[0103] B20: Use gradient descent to update the weight matrix and bias vector according to the calculated gradient;

[0104] B30: Repeat the above process, iteratively train the model until the loss function converges, and use the test set to evaluate the performance of the graph convolutional network.

[0105] Specifically, in the reverse weight update process, the loss between the network output and the actual label is first calculated through forward propagation. Specifically, the input data is passed through each layer of the graph convolutional network to obtain the output value, and the loss function (such as cross entropy loss) is used to calculate the error between the predicted result and the actual result; then, the chain rule is used to transfer the error from the output layer to the intermediate layer and input layer layer by layer, and the gradient of each layer is calculated; next, the gradient descent method is used to update the weight matrix and bias vector of each layer according to the calculated gradient. This process will be iterated continuously, repeating the steps of forward propagation, loss calculation, gradient calculation and parameter update until the loss function converges, that is, the error is no longer significantly reduced; finally, the trained model is evaluated using the test set, and the generalization ability and practical application effect of the model are verified by calculating the performance indicators of the model on the test set (such as accuracy, precision, recall and F1 score), thereby ensuring that the model parameters are optimized and improving the accuracy and stability of fault diagnosis.

[0106] Embodiment 2

[0107] Based on the same inventive concept as the power transformer fault diagnosis method based on an improved graph convolutional network in the aforementioned embodiment, the present invention also provides a power transformer fault diagnosis system based on an improved graph convolutional network, the system comprising:

[0108] A fault data set acquisition module acquires the power transformer fault data set and performs preprocessing, wherein the data set includes sample data of different operating states and fault types;

[0109] The network structure building module builds an improved graph convolutional network structure and improves the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods;

[0110] The matrix sparse processing module divides the power transformer fault data set into a training set and a test set, and performs sparse processing on the adjacency matrix;

[0111] The model basic setting module initializes the structure and parameters of the graph convolutional network and builds a graph convolutional network model including at least one graph convolutional layer and at least one dense layer;

[0112] The diagnosis result output module uses the back-propagation algorithm to train the graph convolutional network model, and inputs the optimized results into the test set for verification to obtain the power transformer fault diagnosis results.

[0113] The above-mentioned fault diagnosis system in the present invention can effectively implement the power transformer fault diagnosis method based on the improved graph convolutional network, and the technical effects that can be achieved are as described in the above-mentioned embodiments and will not be repeated here.

[0114] Although the present application has been described in conjunction with specific features and embodiments thereof, it is obvious that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the accompanying drawings are merely exemplary illustrations of the present application as defined therein, and are deemed to have covered any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A power transformer fault diagnosis method based on an improved graph convolutional network, characterized in that: include: Acquire a power transformer fault data set and perform preprocessing, wherein the data set includes sample data of different operating states and fault types; Construct an improved graph convolutional network structure and improve the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods; Dividing the power transformer fault data set into a training set and a test set, and performing sparse processing on the adjacency matrix; Initialize the structure and parameters of the graph convolutional network, and build a graph convolutional network model including at least one graph convolutional layer and at least one dense layer; The back-propagation algorithm is used to train the graph convolutional network model, and the optimized results are input into the test set for verification to obtain the power transformer fault diagnosis results.

2. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 1 is characterized in that: Preprocessing includes: Select characteristic gases and perform data preprocessing to obtain normalized characteristic values; The dissolved gas content and new features are mapped to the specified interval using the min-max normalization method.

3. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 1 is characterized in that: Construct an improved graph convolutional network structure and improve the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods, including: Constructing an adjacency matrix, wherein the adjacency matrix represents a similarity measure between historical data and current samples; Performing a self-loop operation in the adjacency matrix, each node contains its own feature information during calculation; The adjacency matrix is ​​normalized, and the output features are calculated using the improved graph convolution layer through the input graph matrix and the normalized adjacency matrix.

4. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 3 is characterized in that: The normalization processing formula is: Among them, A″ is the normalized adjacency matrix, D is the adjacency matrix A ′ E is the diagonal node degree matrix of the adjacency matrix A, A ′ is the result of the self-loop operation, A ′ =A+I, where I is the identity matrix.

5. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 4 is characterized in that: Initialize the structure and parameters of the graph convolutional network and build a graph convolutional network model containing at least one graph convolutional layer and at least one dense layer, including: Taking graph structure data as input, we use intrinsic topology to generate intermediate features, and use the weights represented to linearly combine the feature vector of each node with the feature vectors of adjacent nodes to obtain a mixed feature matrix. In the first graph convolution layer, the output data of the first layer is obtained by multiplying the weight matrix and adding the bias vector, choosing ReLU as the activation function; The output data of the first layer is input into the second graph convolution layer, and the output data of the second layer is obtained through linear combination and activation function processing; Add two dense layers after the graph convolutional layer, check and process the data in the third layer, and use the weight matrix, bias vector and activation function to get the output data; The Softmax function is used as the activation function in the fourth layer to process the output data of the third layer to obtain the fault type output by the improved GCN.

6. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 1 is characterized in that: The back propagation algorithm is used to train the graph convolutional network model, including: forward excitation propagation and reverse weight update. In the forward excitation propagation, the input variable is passed to the dense layer after being processed by multiple graph convolution layers, and the label of the sample is output.

7. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 6 is characterized in that: The loss function is calculated using the fault diagnosis results and the actual results. The loss function is: Among them, L(θ) is the loss function, C is the number of categories, that is, the number of possible classifications in the classification problem, and v c is the weight of category C, y i is the actual label of the i-th sample, (y i =c) is the indicator function. When the actual label of the i-th sample is c, (y i =c) = 1, otherwise 0, s i is the probability that the i-th sample is predicted to be the correct category, x i is the score of the i-th sample in the input layer of the Softmax function, x c is the score of category c in the Softmax function input layer, It is the sum of the index scores for all categories.

8. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 7 is characterized in that: After the loss function is established, the graph convolutional network is globally optimized, and the parameter θ is updated as follows: x I =θ·x I-1 +b I Among them, μ is the learning rate, Q is the number of samples in each layer, and x l is the forward propagation function, l is the number of layers of the neural network, E D is the Euclidean distance between the predicted and actual values ​​for each mini-batch.

9. The power transformer fault diagnosis method based on improved graph convolutional network according to claim 6 is characterized in that: The reverse weight updating includes: The loss between the network output and the actual label is calculated through forward propagation, and the error is transferred from the output layer to the intermediate layer and input layer layer by layer using the chain rule to calculate the gradient of each layer; Using the gradient descent method to update the weight matrix and the bias vector according to the calculated gradient; Repeat the above process, iteratively train the model until the loss function converges, and use the test set to evaluate the performance of the graph convolutional network.

10. A power transformer fault diagnosis system based on an improved graph convolutional network, characterized in that: The system comprises: A fault data set acquisition module is used to acquire a power transformer fault data set and perform preprocessing, wherein the data set includes sample data of different operating states and fault types; The network structure building module builds an improved graph convolutional network structure and improves the calculation of the graph convolution layer through self-loop and adjacency matrix normalization methods; A matrix sparse processing module, which divides the power transformer fault data set into a training set and a test set, and performs sparse processing on the adjacency matrix; The model basic setting module initializes the structure and parameters of the graph convolutional network and builds a graph convolutional network model including at least one graph convolutional layer and at least one dense layer; The diagnosis result output module uses the back-propagation algorithm to train the graph convolutional network model, and inputs the optimized results into the test set for verification to obtain the power transformer fault diagnosis results.

Images