A distribution transformer fault diagnosis method based on vibration signals

Abstract

The present invention relates to a distribution transformer fault diagnosis method based on vibration signals, comprising the following steps: obtaining the distribution transformer vibration signal, processing the distribution transformer vibration signal using a combination of adaptive noise complete set empirical mode decomposition and Hilbert transform, respectively obtaining marginal spectra of different frequency bands to construct eigenvectors; constructing a Gaussian function-weighted undirected complete graph for the eigenvector matrix, obtaining an adjacency matrix, and building a multi-channel and multi-connected graph convolutional neural network model for mining deep features and fault classification; in the graph convolutional neural network model, optimizing the Gaussian kernel bandwidth using an improved Grey Wolf optimization algorithm with a perturbation factor of a sine function to obtain an optimal diagnostic model; and performing fault identification on the object to be identified using the obtained optimal diagnostic model. This method is advantageous for improving diagnostic accuracy and identifying unknown types of faults.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis, and in particular relates to a distribution transformer fault diagnosis method based on vibration signals. Background Art

[0002] In recent years, my country's power grid has undergone significant changes, evolving from its initial phase of low power, low voltage, and small-scale development to a modern phase characterized by large capacity, large units, ultra-high voltage, and large-scale operation. As the scale of the power system continues to expand, the voltage levels and capacity of power equipment have also increased, leading to an increased probability of power equipment failure. This inevitably places higher demands on the overall safety of the power system and the reliability and stability of power equipment. Distribution transformers are core equipment in the distribution network, and sudden failures can cause significant safety incidents and economic losses. Statistics show that countless accidents occur each year due to transformer failures. Therefore, monitoring the operating status of distribution transformers is essential to ensure the proper operation of the distribution network.

[0003] Transformer fault detection technology based on vibration signal analysis is a current research hotspot. This method involves attaching displacement, velocity, or acceleration sensors to the transformer's supports, winding sides, or oil tank surface. The vibration signals are used to analyze the transformer's operating status and even identify the type of fault. This method offers advantages such as ease of installation, flexibility, and the absence of electrical connections.

[0004] Existing distribution transformer fault diagnosis algorithms are diverse, mainly including two categories: manual feature extraction and artificial intelligence methods. Manual feature extraction is to decompose the transformer vibration signal, extract representative features, train a strong classifier or propose some diagnostic indicators, and perform classification in the form of thresholds. With the development of digital intelligence technology, deep learning methods have become one of the most popular research directions. Artificial intelligence methods directly input the vibration signal into the network, eliminating the feature extraction step and directly analyzing and identifying the data. During the actual operation of the distribution transformer, the operating conditions are not static. The fluctuation of the operating conditions leads to large differences in the vibration signals, and the features proposed by manual experience are not effective in distinguishing them. In addition, there are often some unknown types of faults (i.e., new faults) during the operation of the distribution transformer, and the existing fault diagnosis algorithms are not effective in identifying unknown types of faults. Summary of the Invention

[0005] The object of the present invention is to provide a distribution transformer fault diagnosis method based on vibration signals, which is conducive to improving the diagnosis accuracy and identifying unknown types of faults.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a distribution transformer fault diagnosis method based on vibration signals, comprising:

[0007] Step S1, collecting the vibration signal of the distribution transformer, and processing the vibration signal of the distribution transformer using a combination of adaptive noise complete set empirical mode decomposition and Hilbert transform, respectively obtaining marginal spectra of different frequency bands to construct feature vectors;

[0008] Step S2: construct a Gaussian function-weighted undirected complete graph for the eigenvector matrix, obtain the adjacency matrix, and build a multi-channel and multi-connected graph convolutional neural network model for mining deep features and fault classification;

[0009] Step S3: In the graph convolutional neural network model, the Grey Wolf Optimization Algorithm with a perturbation factor of a sinusoidal function is used to optimize the Gaussian kernel bandwidth to obtain the optimal diagnostic model;

[0010] Step S4: perform fault identification on the object to be identified using the obtained optimal diagnosis model.

[0011] Furthermore, the vibration signal of the distribution transformer box surface is collected, and the distribution transformer vibration signal is processed by combining adaptive noise complete set empirical mode decomposition with Hilbert transform, including the following steps:

[0012] Step S11: using a vibration signal acquisition device to obtain a vibration signal on the surface of a distribution transformer box;

[0013] Step S12: Process the vibration signal of the distribution transformer using the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) with Adaptive Noise. The specific process is as follows:

[0014] Define operator E j Indicates obtaining the j-order IMF component through EMD; n i represents Gaussian white noise with mean 0 and variance 1; M(.) represents the local mean operator; std(.) represents the standard deviation; x is the original signal; ε0 is the coefficient used to control the signal-to-noise ratio of the auxiliary noise to the original signal, and generates the adaptive coefficient β when calculating the kth modal component k-1 , controls the size of the noise added to the previous margin; when k = 1, β0 = ε0std(x) / std(E1(n i )), when k≥2, β k-1 =ε0std(r k );

[0015] By calculating x i =x+β0E1(n i ) and obtain the first margin r1:

[0016]

[0017] Where i=1,2,...,N, N is the number of noise additions;

[0018] In the first stage, that is, k = 1, the first modal component is calculated:

[0019] IMF1=x-r1 (2)

[0020] The second IMF2 is expressed as:

[0021]

[0022] Add adaptive noise signal r1+β1E2(n i );

[0023] Similarly, find the kth IMF k , where k = 3, 4, ..., m, and m is the total number of IMF components:

[0024] IMF k =r k-1 -r k (4)

[0025] Repeat the above steps until the remainder r meets the residual component termination condition; finally, the original signal x is decomposed into:

[0026]

[0027] Step S13: Use Hilbert transform to obtain marginal spectrum information of the m-order IMF component obtained by CEEMDAN decomposition of the distribution transformer vibration signal. The calculation process is shown in equations (6) to (11):

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] Where a k (t) represents the instantaneous amplitude function of the kth modal component; φ k (t) represents the corresponding instantaneous phase function; ω k(t) represents the corresponding instantaneous frequency; H(ω,t) represents the Hilbert spectrum; B(ω) represents the Hilbert marginal spectrum, which characterizes the amplitude distribution of the signal at each instantaneous frequency;

[0035] Step S14: construct a feature vector from the Hilbert marginal spectra of different frequency bands.

[0036] Furthermore, a Gaussian function-weighted undirected complete graph is constructed for the eigenvector matrix to obtain the adjacency matrix. A graph convolutional neural network model is then built to mine deep features and classify faults. The following steps are included:

[0037] Step S21: constructing a Gaussian function-weighted undirected complete graph for the eigenvector matrix composed of marginal spectrum information, and obtaining an adjacency matrix;

[0038] Take each sample as a vertex, assume that all vertices are connected by edges but with different edge weights. The edge weights are calculated by the Gaussian function as follows:

[0039]

[0040] Where A pq =A qp represents the connection weight between two vertices, η represents the Gaussian kernel bandwidth, X p ,X q Represents the eigenvectors of the pth and qth samples in the eigenvector matrix X;

[0041] Step S22: construct a multi-channel and multi-connected graph convolutional neural network model, namely an improved GCN model. The improved GCN model uses multiple independent graph convolution layers gc1, gc2 and gc4 to extract features from each channel. After the features of each channel are fused, they are connected to a graph convolution layer gc5, and the output layer is connected to a classifier. A graph convolution layer gc3 is added to the forward propagation network to extract marginal spectrum information from different scales and increase the diversity of node features in the gc4 layer. Finally, the cross entropy loss function is used to calculate the loss value, and the Adam optimizer is used to update the model parameters. The GCN forward propagation is shown in Equations (13) to (18):

[0042] Model input layer:

[0043] H (1) =σ[D -1 / 2 (A+I)D -1 / 2 XW (1) ] (13)

[0044] Model hidden layer:

[0045] H (2) =σ[D -1 / 2(A+I)D -1 / 2 (H (1) W (2) +XW (3) )] (14)

[0046] H (4) =σ[D -1 / 2 (A+I)D -1 / 2 H (2) W (4) ] (15)

[0047] Each channel feature fusion layer:

[0048] H (5) =[H 1(4) ,H 2(4) ,...,H h(4) ] (16)

[0049] Model output layer:

[0050] y=[D -1 / 2 (A+I)D -1 / 2 H (5) W (5) ] (17)

[0051] lg_softmax classification:

[0052]

[0053] Where A+I is the adjacency matrix with self-loop; I is the identity matrix; D is the degree matrix corresponding to A+I; W (l) represents the weight of the lth layer; H (5) is the feature fusion of each channel feature; h represents the number of channels; σ(.) represents the activation function, and this study chooses ReLU(.)=max(0,.); y=[y1,y2,...y n ] is the output layer feature, and its dimension is equal to the number of categories n; Y=[Y1,Y2,…Y n ] is the probability value output by the classifier.

[0054] Furthermore, in the graph convolutional neural network model, the Grey Wolf Optimization Algorithm is improved by using a perturbation factor with a sine function to optimize the Gaussian kernel bandwidth and obtain the optimal diagnostic model; the following steps are included:

[0055] Step S31: using a disturbance factor with a sinusoidal function to improve the gray wolf optimization algorithm, mathematically modeling the gray wolf's predation behavior as shown in equations (19) to (25);

[0056] The gray wolf optimization algorithm defines the top three wolves with the best fitness in the pack as α, β, and δ, and the rest as The three best solutions from each generation of wolves guide the target search and position update;

[0057]

[0058]

[0059] F=2γr1-γ (21)

[0060] C=2r2 (22)

[0061]

[0062]

[0063]

[0064] Where d represents the distance between the individual and the target; Indicates the update of the gray wolf's position; represents the target vector position; T is the current iteration number; represents the position vector of the gray wolf; F and C are vector coefficients; r1 and r2 are random numbers in the range [0,1]; γ is a perturbation factor with a sine function, which is characterized by slowing down the decay rate of the perturbation factor in the early stage of the algorithm to improve the global search capability; in the later stage of the algorithm, the decay rate of the perturbation factor is increased and a smaller value is obtained to avoid the optimal solution from being unstable and accelerate the convergence of the algorithm;

[0065] Step S32: Use the improved gray wolf optimization algorithm to optimize the Gaussian function kernel bandwidth, with minimizing the model validation set loss value as the objective function, and the search space is set to (0.1-5);

[0066] Step S33, iterate multiple times until the iteration stop condition is met, that is, the model validation set loss value no longer changes or reaches the set maximum number of iterations, and then end the optimization to obtain a better kernel bandwidth value.

[0067] Furthermore, based on the obtained optimal diagnostic model, a two-level classification method is used to identify faults of the identified objects: in the first level classification, the peak factor is calculated using the output result of the last level of the optimal diagnostic model. If it exceeds the threshold, it is judged as an unknown type of fault. Otherwise, the second level classification is performed; in the second level classification, lg_softmax is used to identify the specific fault. Specifically, the following steps are included:

[0068] Step S41: Using the optimal diagnostic model obtained in S3, calculate the peak factor of the characteristic vector output by the optimal model for a known type of fault sample (training set), and set the upper and lower limits of the peak factor;

[0069] Step S42: When testing each unknown fault sample, use the optimal diagnosis model to obtain the last layer output feature vector y and calculate its peak factor f;

[0070] Step S43: In the first level classification, determine whether the peak factor f obtained in step S42 is within the upper and lower limits of the peak factor of a known type of fault. If not, mark it as an unknown type of fault; otherwise, enter the second level classification;

[0071] Step S44: In the second level classification, lg_softmax is calculated for the output feature vector y of the last layer of the model to determine the specific fault category.

[0072] Compared with the existing technology, the present invention has the following beneficial effects: the present invention overcomes the problems of low diagnostic accuracy, difficulty in feature extraction, and inability to identify unknown types of faults in existing distribution transformer fault diagnosis methods. The adaptive noise complete set empirical mode decomposition is combined with the Hilbert transform to process the distribution transformer vibration signal, obtain the marginal spectrum to construct the feature vector, and combine the Hilbert marginal spectrum with the graph convolutional neural network. Then, the gray wolf optimization algorithm is used to search for the optimal diagnostic model. Compared with manual feature extraction and end-to-end deep learning technology, the proposed features can be made clearer, thereby improving the diagnostic accuracy of the model. On this basis, by calculating the peak factor of the model output layer feature vector and using the threshold method to realize the identification of unknown types of faults, the problem that traditional algorithms cannot identify unknown types of faults and usually misjudge them as the most similar type of model is overcome. Therefore, the present invention has strong practicality and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is a flowchart of a method implementation of an embodiment of the present invention;

[0074] Figure 2 This is a flow chart of distribution transformer vibration signal processing in an embodiment of the present invention;

[0075] Figure 3 is a topological structure diagram of an undirected weighted complete graph in an embodiment of the present invention;

[0076] Figure 4 Schematic diagram of the improved graph convolutional neural network model structure in an embodiment of the present invention;

[0077] Figure 5 This is a flowchart of an implementation of optimizing Gaussian kernel bandwidth using the Grey Wolf optimization algorithm in an embodiment of the present invention;

[0078] Figure 6 This is a flow chart of unknown type fault prediction in an embodiment of the present invention;

[0079] Figure 7The original waveform and spectrum diagram of each state in the embodiment of the present invention;

[0080] Figure 8 This is a diagram of the diagnosis results in an embodiment of the present invention;

[0081] Figure 9 4 is a peak factor distribution diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0082] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0084] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0085] like Figure 1 As shown, this embodiment provides a distribution transformer fault diagnosis method based on vibration signals, comprising the following steps:

[0086] Step S1: Obtain the vibration signal of the distribution transformer box, and use the adaptive noise complete set empirical mode decomposition combined with Hilbert transform to process the distribution transformer vibration signal, and obtain the marginal spectrum of different frequency bands to construct feature vectors. Figure 2 As shown, step S1 specifically includes the following steps:

[0087] Step S11: Use a vibration signal acquisition device to obtain a vibration signal on the surface of the distribution transformer, wherein the sampling frequency is set to 3 kHz and the sampling time is set to 0.1 s.

[0088] Step S12: Process the vibration signal of the distribution transformer using the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) with Adaptive Noise. The specific process is as follows:

[0089] Define operator E j Indicates obtaining the j-order IMF component through EMD; n irepresents Gaussian white noise with mean 0 and variance 1; M(.) represents the local mean operator; std(.) represents the standard deviation; x is the original signal; ε0 is the coefficient used to control the signal-to-noise ratio of the auxiliary noise to the original signal, and generates a new adaptive coefficient β when calculating the kth modal component k-1 , controls the size of the noise added to the previous margin; when k = 1, β0 = ε0std(x) / std(E1(n i )), when k≥2, β k-1 =ε0std(r k );

[0090] 1) By calculating x i =x+β0E1(n i ) and obtain the first margin r1:

[0091]

[0092] Where i = 1, 2, ..., N, where N is the number of times noise is added.

[0093] 2) In the first stage, that is, k = 1, calculate the first modal component:

[0094] IMF1=x-r1 (2)

[0095] 3) The second IMF2 is expressed as:

[0096]

[0097] Add adaptive noise signal r1+β1E2(n i ).

[0098] 4) Similarly, find the kth IMF k , where k = 3, 4, ..., m, and m is the total number of IMF components:

[0099] IMF k =r k-1 -r k (4)

[0100] 5) Repeat the above steps until the remainder r meets the residual component termination condition; finally, the original signal x is decomposed into:

[0101]

[0102] Step S13: Use Hilbert transform to obtain marginal spectrum information of the m-order IMF component obtained by CEEMDAN decomposition of the distribution transformer vibration signal. The calculation process is shown in equations (6) to (11):

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] Where a k (t) represents the instantaneous amplitude function of the kth modal component; φ k (t) represents the corresponding instantaneous phase function; ω k (t) represents the corresponding instantaneous frequency; H(ω,t) represents the Hilbert spectrum; B(ω) represents the Hilbert marginal spectrum, which characterizes the amplitude distribution of the signal at each instantaneous frequency.

[0110] Step S14: construct a feature vector from the Hilbert marginal spectra of different frequency bands.

[0111] Step S2: Construct a Gaussian function-weighted undirected complete graph for the eigenvector matrix, obtain the adjacency matrix, and build a graph convolutional neural network model for mining deep features and fault classification. Step S2 specifically includes the following steps:

[0112] Step S21: construct an undirected complete graph weighted by a Gaussian function for the eigenvector matrix composed of marginal spectrum information, and obtain an adjacency matrix.

[0113] In the graph G = (V, E, A) model, V and E represent the set of vertices and edges, respectively, and A represents the adjacency matrix of the graph. According to the connection relationship of the edges, graphs can be divided into undirected graphs and directed graphs; according to the weight relationship of the edges, graphs can be divided into weighted graphs and unweighted graphs. The topological structure of an undirected complete graph does not change with parameter changes, and there is no need to consider the direction of the edge connection between vertices, so the construction process is simple. This method uses an undirected weighted complete graph, with each sample as a vertex. It is assumed that all vertices are connected by edges but the edge weights are different. The weight of the edge connection is calculated by a Gaussian function, as follows:

[0114]

[0115] Where A pq =A qp represents the connection weight between two vertices, η represents the Gaussian kernel bandwidth, X p ,X q Represents the eigenvectors of the pth and qth samples in the eigenvector matrix X.

[0116] Step S22: Use the pytroch framework in spyder software to build a multi-channel and multi-connected graph convolutional neural network (improved GCN) model to achieve different scale feature mining and multi-channel information connectivity, thereby improving the quality of the extracted information. In order to mine effective information reflecting fault characteristics from the marginal spectra of different frequencies, the improved GCN model constructed in this embodiment is as follows: Figure 4 As shown in Figure 2. The improved GCN model constructed uses multiple independent graph convolution layers gc1, gc2, and gc4 to extract features from each channel. After the features of each channel are fused, they are connected to a graph convolution layer gc5, and the output layer is connected to a classifier. A graph convolution layer gc3 is added to the forward propagation network to extract marginal spectrum information from different scales and increase the diversity of node features in the gc4 layer. Finally, the cross entropy loss function is used to calculate the loss value, and the Adam optimizer is used to update the model parameters. The forward propagation of GCN is shown in Equations (13) to (18):

[0117] Model input layer:

[0118] H (1) =σ[D -1 / 2 (A+I)D -1 / 2 XW (1) ] (13)

[0119] Model hidden layer:

[0120] H (2) =σ[D -1 / 2 (A+I)D -1 / 2 (H (1) W (2) +XW (3) )] (14)

[0121] H (4) =σ[D -1 / 2 (A+I)D -1 / 2 H (2) W (4) ] (15)

[0122] Each channel feature fusion layer:

[0123] H (5) =[H 1(4) ,H 2(4) ,...,H h(4) ] (16)

[0124] Model output layer:

[0125] y=[D -1 / 2 (A+I)D -1 / 2 H (5) W (5) ] (17)

[0126] lg_softmax classification:

[0127]

[0128] Where A+I is the adjacency matrix with self-loop; I is the identity matrix; D is the degree matrix corresponding to A+I; W (l) represents the weight of the lth layer; H (5) is the feature fusion of each channel feature; h represents the number of channels; σ(.) represents the activation function, and this study chooses ReLU(.)=max(0,.); y=[y1,y2,...y n ] is the output layer feature, and its dimension is equal to the number of categories n; Y=[Y1,Y2,…Y n ] is the probability value output by the classifier.

[0129] Step S3: In the graph convolutional neural network model, the Grey Wolf Optimization Algorithm with a perturbation factor of a sine function is used to optimize the Gaussian kernel bandwidth to obtain the optimal diagnostic model; Figure 5 As shown, step S3 specifically includes the following steps:

[0130] Step S31: Use the disturbance factor with a sinusoidal function to improve the gray wolf optimization algorithm, and mathematically model the gray wolf's predation behavior as shown in equations (19) to (25).

[0131] The gray wolf optimization algorithm defines the top three wolves with the best fitness in the pack as α, β, and δ, and the rest as The target search and position update are guided by the three best solutions in each generation of wolves.

[0132]

[0133]

[0134] F=2γr1-γ (21)

[0135] C=2r2 (22)

[0136]

[0137]

[0138]

[0139] Where d represents the distance between the individual and the target; Indicates the update of the gray wolf's position; represents the target vector position; T is the current iteration number; represents the position vector of the gray wolf; F and C are vector coefficients; r1 and r2 are random numbers in the range [0,1]; γ is a disturbance factor with a sine function, which is characterized by slowing down the decay rate of the disturbance factor in the early stage of the algorithm execution to improve the global search capability; in the later stage of the algorithm, the decay rate of the disturbance factor is increased and a smaller value is obtained to avoid the turbulence of the optimal solution and accelerate the convergence of the algorithm.

[0140] Step S32: Use the improved grey wolf optimization algorithm to optimize the Gaussian function kernel bandwidth, with minimizing the loss value of the model validation set as the objective function, and the search space is set to (0.1~5).

[0141] Step S33, iterate multiple times until the iteration stop condition is met, that is, the model validation set loss value no longer changes or reaches the set maximum number of iterations, and then end the optimization to obtain a better kernel bandwidth value.

[0142] Step S4: For each unknown fault object to be identified, a two-level classification method is used to identify the fault of the object to be identified based on the obtained optimal diagnostic model: in the first level classification, the peak factor is calculated using the output result of the last level of the optimal diagnostic model. If it exceeds the threshold, it is judged as an unknown type of fault. Otherwise, the second level classification is performed; in the second level classification, lg_softmax is used to identify the specific fault. Figure 6 As shown, step S4 specifically includes the following steps:

[0143] Step S41: Use S3 to obtain the trained optimal diagnostic model, calculate the peak factor of the model output feature vector of each sample in the training set of known type fault samples (training set) after the optimal model is passed, and set the upper and lower limits of the peak factor;

[0144] Step S42: When testing each unknown fault sample, use the optimal diagnosis model to obtain the last layer output feature vector y and calculate its peak factor f;

[0145] Step S43: In the first level classification, determine whether the peak factor f obtained in step S42 is within the upper and lower limits of the peak factor of a known type of fault. If not, mark it as an unknown type of fault; otherwise, enter the second level classification;

[0146] Step S44: In the second level classification, lg_softmax is calculated for the output feature vector y of the last layer of the model to determine the specific fault category.

[0147] In this example, the data samples obtained are from an S11-M-315 / 10 oil-immersed transformer. Fault simulations were performed on the distribution transformer. Vibration signals were collected under short-circuit and no-load conditions for four fault conditions: winding deformation, winding looseness, normal winding state, and core two-point grounding and core looseness. These faults are denoted as states 1-5 below. The server parameters used for the calculations were an Intel Core i5-10200H CPU and 16GB of RAM.

[0148] The original waveform and spectrum of each state are as follows Figure 7 The diagnostic results of this method are shown in Figure 8 As shown in Figure 1, the rated operating condition samples were divided into a training set of 1,200 and a validation set of 300 for model training. The test set contained five states, each with three different operating conditions, for a total of 3,000 samples. A confusion matrix was plotted for the diagnosis results, where 6 represents an unknown fault type. The proposed method accurately identified all five transformer states: normal, winding deformation, winding looseness, two-point core grounding, and loose core, with an accuracy rate of 97.73%. The diagnosis results for each state and different operating conditions are shown in Tables 1 and 2. The loose winding, normal, and two-point core grounding states were less affected by fluctuations in transformer operating conditions. Under no-load voltage fluctuations, the diagnostic accuracy for loose core faults was relatively low, ranging from 94.33% to 95%, while the diagnostic accuracy for two-point core grounding faults was higher. Observation revealed that samples with misclassified loose core faults were misclassified as two-point core grounding faults and unknown fault types. The diagnostic accuracy of winding deformation faults fluctuated significantly under varying load currents, with all misidentified samples misclassified as loose windings and unknown fault types. This demonstrates that the proposed method has a reasonable level of diagnostic accuracy for different operating conditions.

[0149] Table 1 Diagnostic results under different load currents

[0150]

[0151] Table 2 Diagnostic results at different no-load voltages

[0152]

[0153] In this example, to verify the effectiveness of this method in predicting unknown fault types, 300 sets of BC short-circuit fault samples were collected from the S11-M-315 / 10 transformer as unknown fault types and identified using the proposed two-level classification method. The diagnosis result was 92.76%, and the peak factor distribution of the output layer feature vector of the optimal diagnosis model is shown in the following figure: Figure 9The figure shows the difference in peak factors for 1200 known fault types (model training set samples) and 300 unknown fault types (newly emerging fault types). The peak factors for known fault types, after applying the optimal diagnostic model, are relatively concentrated, while the results for unknown fault types are dispersed. This means that by selecting appropriate thresholds, it is possible to accurately identify unknown fault types. Their boundaries can be calculated using statistical principles.

[0154] In summary, the distribution transformer fault diagnosis method based on vibration signals proposed in the present invention has achieved the following technical effects:

[0155] (1) Combining the Hilbert marginal spectrum with the graph convolutional neural network can make the proposed features clearer than manual feature extraction and end-to-end deep learning technology, thereby improving the diagnostic accuracy of the model.

[0156] (2) To address the problem that the convergence factor of the traditional gray wolf optimization algorithm does not match the actual gray wolf hunting behavior, a nonlinear perturbation factor with a sine function is proposed. This increases the convergence factor value in the early stage of the algorithm execution, obtaining sufficient perturbation to enhance the algorithm's global search capability, while reducing the convergence factor value in the later stage to avoid the turbulence of the optimal solution and accelerate convergence. At the same time, the structure of the GCN model is improved. A graph convolution layer with different sizes is added to the forward propagation network to realize multi-scale feature mining. At the same time, multiple parallel channels are designed to achieve multi-source information fusion.

[0157] (3) A two-level classification method is proposed to identify unknown type faults. In the first level classification, the peak factor of the model output layer feature vector is calculated and the threshold method is used to identify unknown type faults. In the second level classification, lg_softmax is used to identify known type faults, which overcomes the problem that traditional algorithms cannot identify unknown type faults and usually misjudge them as the type that is most similar to the model.

[0158] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A distribution transformer fault diagnosis method based on vibration signals, characterized in that: include: Step S1, collecting the vibration signal of the distribution transformer, and processing the vibration signal of the distribution transformer using a combination of adaptive noise complete set empirical mode decomposition and Hilbert transform, respectively obtaining marginal spectra of different frequency bands to construct feature vectors; Step S2: construct a Gaussian function-weighted undirected complete graph for the eigenvector matrix, obtain the adjacency matrix, and build a multi-channel and multi-connected graph convolutional neural network model for mining deep features and fault classification; Step S3: In the graph convolutional neural network model, the Grey Wolf Optimization Algorithm with a perturbation factor of a sinusoidal function is used to optimize the Gaussian kernel bandwidth to obtain the optimal diagnostic model; Step S4: perform fault identification on the object to be identified using the obtained optimal diagnostic model; Construct a Gaussian function-weighted undirected complete graph for the eigenvector matrix, calculate the adjacency matrix, and build a graph convolutional neural network model for mining deep features and fault classification. The following steps are included: Step S21: constructing a Gaussian function-weighted undirected complete graph for the eigenvector matrix composed of marginal spectrum information, and obtaining an adjacency matrix; Take each sample as a vertex, assume that all vertices are connected by edges but with different edge weights. The edge weights are calculated by the Gaussian function as follows: Where A pq =A qp represents the connection weight between two vertices, η represents the Gaussian kernel bandwidth, X p ,X q Represents the eigenvectors of the pth and qth samples in the eigenvector matrix X; Step S22: construct a multi-channel and multi-connected graph convolutional neural network model, namely an improved GCN model. The improved GCN model uses multiple independent graph convolution layers gc1, gc2 and gc4 to extract features from each channel. After the features of each channel are fused, they are connected to a graph convolution layer gc5, and the output layer is connected to a classifier. A graph convolution layer gc3 is added to the forward propagation network to extract marginal spectrum information from different scales and increase the diversity of node features in the gc4 layer. Finally, the cross entropy loss function is used to calculate the loss value, and the Adam optimizer is used to update the model parameters. The GCN forward propagation is shown in Equations (13) to (18): Model input layer: H (1) =σ[D -12 (A+I)D -12 XW (1) ] (2) Model hidden layer: H (2) =σ[D -12 (A+I)D -12 (H (1) W (2) +XW (3) )] (3) H (4) =σ[D -12 (A+I)D -12 H (2) W (4) ] (4) Each channel feature fusion layer: H (5) =[H 1(4) ,H 2(4) ,...,H h(4) ] (5) Model output layer: y=[D -12 (A+I)D -12 H (5) W (5) (6) lg_softmax classification: Where A+I is the adjacency matrix with self-loop; I is the identity matrix; D is the degree matrix corresponding to A+I; W (l) represents the weight of the lth layer; H (5) is the feature of the fusion of each channel feature; h represents the number of channels; σ(.) represents the activation function, ReLU(.)=max(0,.); y=[y1,y2,...,y n ] is the output layer feature, and its dimension is equal to the number of categories n; Y=[Y1,Y2,…,Y n ] is the probability value output by the classifier; Based on the obtained optimal diagnostic model, a two-level classification method is used to identify faults of the identified objects: in the first level classification, the peak factor is calculated using the output of the last level of the optimal diagnostic model. If it exceeds the threshold, it is judged as an unknown type of fault. Otherwise, the second level classification is performed; in the second level classification, lg_softmax is used to identify the specific fault. The specific steps include: Step S41: Using the optimal diagnostic model obtained in S3, calculate the peak factor of the characteristic vector output by the optimal model for a known type of fault sample, and set the upper and lower limits of the peak factor; Step S42: When testing each unknown fault sample, use the optimal diagnosis model to obtain the last layer output feature vector y and calculate its peak factor f; Step S43: In the first level classification, determine whether the peak factor f obtained in step S42 is within the upper and lower limits of the peak factor of a known type of fault. If not, mark it as an unknown type of fault; otherwise, enter the second level classification; Step S44: In the second level classification, lg_softmax is calculated for the output feature vector y of the last layer of the model to determine the specific fault category.

2. A distribution transformer fault diagnosis method based on vibration signals according to claim 1, characterized in that: The vibration signal of the distribution transformer box surface is collected and processed by combining adaptive noise complete set empirical mode decomposition with Hilbert transform, including the following steps: Step S11: using a vibration signal acquisition device to obtain a vibration signal on the surface of a distribution transformer box; Step S12: Process the vibration signal of the distribution transformer using the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) with Adaptive Noise. The specific process is as follows: Define operator E j Indicates obtaining the j-order IMF component through EMD; n i represents Gaussian white noise with mean 0 and variance 1; M(.) represents the local mean operator; std(.) represents the standard deviation; x is the original signal; ε0 is the coefficient used to control the signal-to-noise ratio of the auxiliary noise to the original signal, and generates the adaptive coefficient β when calculating the kth modal component k-1 , controls the size of the noise added to the previous margin; when k = 1, β0 = ε0std(x) / std(E1(n i )), when k≥2, β k-1 =ε0std(r k ); By calculating x i =x+β0E1(n i ) and obtain the first margin r1: Where i=1,2,...,N, N is the number of noise additions; In the first stage, that is, k = 1, the first modal component is calculated: IMF1=x-r1 (9) The second IMF2 is expressed as: Add adaptive noise signal r1+β1E2(n i ); Similarly, find the kth IMF k , where k = 3, 4, ..., m, and m is the total number of IMF components: IMF k =r k-1 -r k (11) Repeat the above steps until the remainder r meets the residual component termination condition; finally, the original signal x is decomposed into: Step S13: Use Hilbert transform to obtain marginal spectrum information of the m-order IMF component obtained by CEEMDAN decomposition of the distribution transformer vibration signal. The calculation process is shown in equations (6) to (11): Where a k (t) represents the instantaneous amplitude function of the kth modal component; φ k (t) represents the corresponding instantaneous phase function; ω k (t) represents the corresponding instantaneous frequency; H(ω,t) represents the Hilbert spectrum; B(ω) represents the Hilbert marginal spectrum, which characterizes the amplitude distribution of the signal at each instantaneous frequency; Step S14: construct a feature vector from the Hilbert marginal spectra of different frequency bands.

3. A distribution transformer fault diagnosis method based on vibration signals according to claim 1, characterized in that: In the graph convolutional neural network model, the Grey Wolf Optimization Algorithm with a sinusoidal perturbation factor is used to optimize the Gaussian kernel bandwidth and obtain the optimal diagnostic model. The following steps are included: Step S31: using a disturbance factor with a sinusoidal function to improve the gray wolf optimization algorithm, mathematically modeling the gray wolf's predation behavior as shown in equations (19) to (25); The gray wolf optimization algorithm defines the top three wolves with the best fitness in the pack as α, β, and δ, and the rest as The three best solutions from each generation of wolves guide the target search and position update; F=2γr1-γ (21) C＝2r2 (22) Where d represents the distance between the individual and the target; Indicates the update of the gray wolf's position; represents the target vector position; T is the current iteration number; represents the position vector of the gray wolf; F and C are vector coefficients; r1 and r2 are random numbers in [0,1]; γ is the perturbation factor with a sine function; Step S32: Use the improved gray wolf optimization algorithm to optimize the Gaussian function kernel bandwidth, with minimizing the model validation set loss value as the objective function, and the search space is set to (0.1-5); Step S33, iterate multiple times until the iteration stop condition is met, that is, the model validation set loss value no longer changes or reaches the set maximum number of iterations, and then end the optimization to obtain a better kernel bandwidth value.

Images