Transformer substation fault positioning method based on distributed grounding current decoupling analysis

Abstract

The invention discloses a transformer substation fault positioning method based on distributed grounding current decoupling analysis, and belongs to the technical field of fault diagnosis. Aiming at the problem of how to quickly identify and position a fault point in a complex equipment network with high coupling, accurate positioning of a substation equipment fault is realized by constructing a GCC model comprising a data processor, a feature extractor and a feature classifier. The data processor introduces a variational mode decomposition method to explicit different frequency characteristics in a non-stationary grounding current signal containing noise and highlight key frequency information; in the feature extractor, a channel coupling module is provided to decouple the grounding current in the transformer substation and extract the time sequence feature and the equipment coupling feature of the grounding current; in the feature classifier, a Mask module is designed, random inactivation of features is achieved through Dropout, feature suppression is conducted through frequency domain analysis, and the problem that the generalization ability of a model classifier is weak is solved; a weighted cross entropy loss function is adopted, and the problem that fault data and normal data are unbalanced is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of fault diagnosis, and in particular relates to a substation fault locating method based on distributed grounding current decoupling analysis. Background Art

[0002] The grounding grid is an important protection device in the substation, which plays an indispensable role in ensuring the safety of personnel, the integrity of equipment and the stable operation of the substation. A high-quality grounding system requires a small grounding resistance and an appropriate grounding current to maintain the reliability and safety of the system. In the substation system, due to the multi-point grounding design, the rational use of the analysis of the grounding current can effectively identify the unidirectional grounding fault of the equipment in the system. In this way, the fault location can be judged more accurately, and measures can be taken quickly for repair and recovery to ensure the stable operation of the substation system.

[0003] As deep learning technology becomes more popular, current-based fault location methods have gradually attracted the attention of researchers due to their robustness to external environmental interference and their ability to accurately represent equipment status information. Therefore, with the help of ground current analysis, it is possible to accurately locate faults in substation equipment. This makes it possible to quickly identify and locate fault points in complex equipment networks, thereby improving the efficiency and accuracy of equipment maintenance.

[0004] When a complex substation system is in operation, there are many types of equipment, which are affected by current harmonics, making feature extraction more difficult (Wang Qingliang, Song Xi, Wang Xudong, et al. Load harmonic current detection method under background harmonic voltage environment [J]. Power System Protection and Control, 2022, 50(16): 58-65.). Moreover, when a grounding fault occurs in the equipment, due to the interconnection characteristics of the protective grounding network, it will cause current changes in other equipment (Mou Longhua, Li Guoxin, Meng Qinghai. Analysis of the influence of interconnection of high and low voltage power grid protective grounding networks [J]. Journal of Coal Society, 2004, 29(2): 238-242.), making it difficult to directly extract the coupling relationship between devices through convolution. In addition, the substation fault dataset has the problem of small number and imbalance (Zhang Wei, An Yong, Li Haitao, et al. Research on grounding transformer fault diagnosis method based on unbalanced data expansion [J]. Electrical Engineering, 2023 (20): 146-148 + 154. DOI: 10.19768 / j.cnki.dgjs.2023.20.043.). These factors have brought great challenges to the precise positioning of substation fault equipment. Summary of the invention

[0005] In response to the problem of how to quickly identify and locate fault points in a complex equipment network with high coupling, the present invention provides a substation fault location method based on distributed grounding current decoupling analysis, focusing on the in-depth analysis of grounding current and applying a deep learning model to achieve accurate fault location in the substation system.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The GCC deep learning model architecture with graph convolutional network combined with attention mechanism as the core is adopted to achieve accurate positioning of substation equipment faults. The GCC model consists of three parts: data processor, feature extractor and feature classifier; among them, the data processor uses variational mode decomposition method to effectively visualize the different frequency characteristics in the non-stationary ground current signal containing noise, highlighting the key frequency information hidden in the complex signal; in the feature extractor, a channel coupling module is proposed to decouple the ground current in the substation and extract the timing characteristics and equipment coupling characteristics of the ground current; in the feature classifier, a Mask module is designed to use Dropout to achieve random inactivation of features, and feature suppression is performed through frequency domain analysis to solve the problem of weak generalization ability of the model classifier. At the same time, the weighted cross entropy loss function is used to effectively solve the problem of imbalance between fault data and normal data.

[0008] A substation fault location method based on distributed grounding current decoupling analysis includes the following steps:

[0009] Step 1: Establish a data set of unidirectional grounding faults of different devices; sort out the grounding current data of the generated device faults through the current acquisition device, and the collected data includes fault samples S, time dimension L, and channel dimension d;

[0010] Step 2: Construct a GCC deep learning model consisting of a data processor, a feature extractor, and a feature classifier, and use the GCC deep learning model to train and test the data set; the feature extractor includes a channel coupling module consisting of a graph similarity attention module, a self-attention feature extraction module, and a channel attention module;

[0011] Furthermore, in the data processor described in step 2, variational modal decomposition is performed on the time dimension data; in the feature extractor, the ground current data of different frequencies are downsampled in the time dimension through one-dimensional convolution to obtain downsampled data; the downsampled data is passed through a graph similarity attention module to extract the coupling features between channels; the downsampled data is passed through a self-attention feature extraction module to extract the temporal features of the data; the output of the graph similarity attention module and the output of the self-attention feature extraction module are input into the channel attention module for feature integration; at the end of the feature extractor, the time dimension features are expanded through a fully connected layer; in the feature classifier, a Mask module is used to suppress low-frequency feature information.

[0012] Furthermore, in the data processor of step 2, variational mode decomposition is performed on the time dimension data, and the specific operation is as follows:

[0013] The ground current data with the size of S×d×L is decomposed into S×d×K×L, where K represents the different frequency data obtained by variational mode decomposition.

[0014] In the feature extractor of step 2, the ground current data of different frequencies are downsampled in the time dimension through one-dimensional convolution to obtain downsampled data. The specific operation is:

[0015] Decompose the grounding current data with a size of S×d×L and the grounding current different frequency data with a size of S×d×K×L into each batch of data with a size of B, and obtain the grounding current data with a size of B×d×L and the grounding current different frequency data with a size of B×d×K×L;

[0016] The ground current data with a size of B×d×L and the ground current data of different frequencies with a size of B×d×K×L are downsampled in the time dimension through a one-dimensional convolution with a convolution kernel size of 10 and a step size of 10, and the sizes of B×d×l and The downsampled data.

[0017] In the data processor of step 2, the downsampled data is passed through the graph similarity attention module to extract the coupling features between channels. The specific operation is:

[0018] The downsampled data of size B×d×l is transformed into B×l×d, and the query vector Q, key vector K and value vector V of size B×l×d are calculated through the graph convolutional network respectively;

[0019] The cosine similarity is calculated for the query vector Q and the key vector K to evaluate the coupling relationship between channels, and the weight is used for softmax normalization to obtain the weight S. The cosine similarity calculation formula is as follows:

[0020]

[0021] Where S represents the cosine similarity weight; Similarity cos represents the adaptive cosine similarity calculation function; || || represents the modulus of the tensor; T represents the transposition operation;

[0022] Multiply the value vector V by the weight S to perform a weighted operation, and obtain a deeper coupling feature B×l×d through the second layer of graph convolutional network; then, transform its dimension to feature H of size B×d×l G , for subsequent feature integration.

[0023] The data processor of step 2 extracts the time series features of the data by passing the downsampled data through the self-attention feature extraction module, and the specific operation is as follows:

[0024] The size is The downsampled data of are passed through the fully connected layer respectively, and the sizes are Query vector Q′, key vector K′, and value vector V′;

[0025] Perform matrix multiplication on the query vector Q′ and the key vector K′, perform normalization operation through softmax, and obtain the weight; then multiply it with the value vector V′, and the output size is characteristics; finally The decomposition dimensions are summed and tied to obtain an output H of size B×d×l S , for feature integration.

[0026] The data processor in step 2 outputs H to the graph similarity attention module. G And the self-attention feature extraction module output H S , input channel attention module for feature integration, the specific operation is:

[0027] The two outputs are concatenated and the obtained features are input into a two-dimensional convolution with a kernel size of 1×1 and a stride of 1;

[0028] Perform global pooling on the features output by the convolutional layer and then output the features;

[0029] Through a layer of compressed fully connected layer, the channel dimension is compressed to the original

[0030] Through a layer of decompression fully connected layer, the channel dimension is restored;

[0031] Finally, the output of the convolutional layer is weighted with this value, and the timing features and coupling features are fused to associate the two.

[0032] In the feature classifier of step 2, the Mask module is used to suppress the low-frequency feature information. The specific operation is as follows:

[0033] First, the features are inactivated by Dropout; then, a specific frequency between the maximum frequency and the minimum frequency is selected to calculate the suppression length s; finally, a mask tensor is generated according to the length s, and then matrix multiplied with the original feature tensor to suppress its features to a certain extent. The formula is as follows:

[0034] H f =|fft(H out )| (2)

[0035]

[0036]

[0037] Where, fft represents the Fourier transform function; H out represents the hidden layer output; || represents the absolute value operation; H f It represents the frequency domain information after taking the absolute value; α represents the selection coefficient; f represents the frequency selection value.

[0038] In the field of time series classification, the entropy loss function is used to measure the accuracy of model prediction and reveal the difference between the predicted results and the actual results. Through data concentration, the one-way grounding faults generated by different devices are weighted to reduce the impact of different amounts of fault data of different devices on the training results. The calculation formula is as follows:

[0039]

[0040]

[0041] in, represents the predicted value; y represents the true value; b represents the sample index; c represents the classification index; B is the total number of single training samples; C represents the total number of categories; y b,c Represents the probability of the c-type sample of the b-th sample in the true value; represents the probability of the c-type sample of the predicted b-th sample; w c Represents the weight of class c in a single batch.

[0042] Step 3: Analyze the results of substation fault equipment location reasoning, conduct ablation experiments on the data set, and verify the network optimization capabilities of the channel coupling module, Mask module, and weighted cross entropy loss;

[0043] Step 4: GCC substation fault location is completed.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] The present invention discloses an innovative substation fault location technology, which is based on distributed ground current decoupling analysis and significantly improves the accuracy and efficiency of fault location. In the key feature extraction stage, the graph convolutional network and the attention mechanism are cleverly integrated. This innovative combination not only enhances the model's coupling feature extraction capability at the channel level, but also achieves more refined feature capture in the time dimension. This unique fusion strategy enables the method to more accurately identify and analyze fault features in ground current data, thereby providing strong technical support for substation fault location.

[0046] Compared with the current intelligent inspection method and the traditional fault location technology relied on by the smart substation, the advantages of the present invention are mainly reflected in the following aspects:

[0047] (1) High accuracy: By combining graph convolutional networks and attention mechanisms, the present invention can more accurately extract and analyze key features in ground current data, thereby achieving higher fault location accuracy.

[0048] (2) Real-time detection: It can capture and analyze ground current data in real time, so it can accurately locate the fault within a very short time after the fault occurs, greatly shortening the fault response time;

[0049] (3) Strong robustness: The present invention has good adaptability to different types of faults and complex operating environments, can maintain stable performance under various circumstances, and can achieve high accuracy on different data sets;

[0050] (4) Low maintenance cost: Compared with traditional inspection and fault location methods, the present invention is based on automated data processing and analysis, which can reduce the workload of manual inspection and maintenance, thereby reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a substation fault location method based on distributed ground current decoupling analysis;

[0052] Figure 2 This is a schematic diagram of the substation system structure;

[0053] Figure 3 It is a schematic diagram of the time series data after cleaning;

[0054] Figure 4 This is a schematic diagram of the overall architecture of the GCC model;

[0055] Figure 5 Schematic diagram of the graph similarity attention module;

[0056] Figure 6Schematic diagram of the self-attention feature extraction module;

[0057] Figure 7 Schematic diagram of the channel attention module;

[0058] Figure 8 This is a schematic diagram of the Mask module;

[0059] Fig. 9 This is a schematic diagram of transient ground current data before and after a substation fault. DETAILED DESCRIPTION

[0060] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0061] A substation fault location method based on distributed grounding current decoupling analysis includes the following steps:

[0062] Step 1: Establish a data set of unidirectional grounding faults of different devices; sort out the grounding current data of the generated device faults through the current acquisition device, and the collected data includes fault samples S, time dimension L, and channel dimension d;

[0063] The grounding current data of equipment failures in a 220kV substation in the north over the past five years were collated using current acquisition equipment. The substation system structure is as follows: Figure 2 As shown. The data set is taken from 11 devices on the main transformer line of the system, with a sampling frequency of 10kHz. Data sampling starts 0.1s before the fault occurs and ends 0.1s after the fault occurs. During the long-term operation of the substation, 11 devices produce unidirectional grounding faults of varying degrees, and the current data in the current substation system (protective grounding current of 11 devices, three-phase current of two current transformers at the transformer inlet and outlet, and working grounding current of two main transformers) are recorded, totaling 19 channels. As shown Figure 3 As shown in the figure, the collected data has three dimensions, namely, fault dimension, transient time dimension and channel dimension. The fault dimension B has a total of 262 data, the time dimension L is 2000, and the channel dimension d is 19. It covers the time series data under the influence of multiple factors, providing data support for subsequent model training.

[0064] Step 2: Build a GCC deep learning model consisting of a data processor, a feature extractor, and a feature classifier, such as Figure 4As shown, the GCC deep learning model is used to train and test the data set; wherein the feature extractor includes a channel coupling module consisting of a graph similarity attention module, a self-attention feature extraction module, and a channel attention module;

[0065] Furthermore, in the data processor described in step 2, variational mode decomposition is performed on the time dimension data; in the feature extractor, the ground current data of different frequencies are downsampled in the time dimension through one-dimensional convolution to obtain downsampled data; the downsampled data is passed through a graph similarity attention module to extract the coupling features between channels; the downsampled data is passed through a self-attention feature extraction module to extract the temporal features of the data; the output of the graph similarity attention module and the output of the self-attention feature extraction module are input into the channel attention module for feature integration; at the end of the feature extractor, the time dimension features are expanded through a fully connected layer; in the feature classifier, a Mask module is used to suppress unimportant low-frequency feature information.

[0066] Furthermore, in the data processor, variational mode decomposition is performed on the time dimension data, and the specific operation is:

[0067] The ground current data with a size of 262×19×2000 (B×d×L) is decomposed into 262×19×8×2000 (B×d×K×L), where K represents the different frequency data obtained by variational mode decomposition.

[0068] In the feature extractor, the ground current data with a size of 262×19×2000 (B×d×L) and the ground current different frequency data with a size of 262×19×8×2000 (B×d×K×L) are downsampled in the time dimension through a one-dimensional convolution with a convolution kernel size of 10 and a step size of 10, respectively. Downsampled data to improve the calculation speed and efficiency of the model.

[0069] like Figure 5 As shown in the figure, the downsampled data of size 262×19×200 (B×d×l) is passed through the graph similarity attention module to extract the coupling features between channels. The downsampled data dimension is transformed into 262×200×19 (B×l×d), and the query vector Q, key vector K and value vector V of size 262×200×19 (B×l×d) are calculated through the graph convolution network respectively;

[0070] The cosine similarity of the query vector Q and the key vector K is calculated to evaluate the coupling relationship between channels, and the softmax normalization is performed based on this weight to obtain a weight S of size 262×19×19; the cosine similarity calculation formula is as follows:

[0071]

[0072] Where S represents the cosine similarity weight; Similarity cos represents the adaptive cosine similarity calculation function; || || represents the modulus of the tensor; T represents the transposition operation;

[0073] The value vector V is multiplied by the weight S to perform a weighted operation, and a deeper coupling feature 262×200×19 (B×l×d) is obtained through the second layer of graph convolutional network; then, its dimension is transformed into a feature H of size 262×19×200 (B×d×l). G , for subsequent feature integration.

[0074] like Figure 6 As shown, the size is The downsampled data is passed through the self-attention feature extraction module to extract the temporal features of the data. The downsampled data is passed through the fully connected layer to obtain Query vector Q', key vector K' and value vector V'.

[0075] The query vector Q′ and the key vector K′ are matrix multiplied and normalized by softmax to obtain a weight of size 262×19×200×200. Then multiplied by the value vector V′, the output size is 262 Features; Finally The decomposition dimensions are summed and tied to obtain an output H of size 262×19×200 (B×d×l) S , for feature integration.

[0076] The output of the graph similarity attention module H G And the self-attention feature extraction module output H S , input as Figure 7 The channel attention module shown in the figure performs feature integration. First, the two outputs are concatenated to obtain a feature of size 262×38×200, which is input into a two-dimensional convolution with a convolution kernel size of 1×1 and a stride of 1.

[0077] First, the features of the convolutional layer output are globally pooled, and the output size is 262×38×1; secondly, a compression fully connected layer is used to compress the channel dimension to the original That is, 262×19×1; then, through a layer of decompression fully connected layer, its channel dimension is restored; finally, this is used as the weight to weight the output of the convolutional layer, and the timing features and coupling features are fused to associate the relationship between the two.

[0078] In the feature classifier, use Figure 8The Mask module shown in Figure 1 suppresses unimportant low-frequency feature information to enhance the generalization ability of the classifier. The specific operations are as follows:

[0079] First, the features are inactivated by Dropout; then, a specific frequency between the maximum frequency and the minimum frequency is selected to calculate the suppression length s; finally, a mask tensor is generated according to the length s, and then matrix multiplied with the original feature tensor to suppress its features to a certain extent. The formula is as follows:

[0080] H f =|fft(H out )| (2)

[0081]

[0082] Where, fft represents the Fourier transform function; H out represents the hidden layer output; || represents the absolute value operation; H f It represents the frequency domain information after taking the absolute value; α represents the selection coefficient; f represents the frequency selection value.

[0083] In the field of time series classification, the loss function is used to measure the accuracy of model predictions and reveals the difference between the predicted results and the actual results. Through the data set, the one-way grounding faults generated by different devices are weighted to reduce the impact of different amounts of fault data of different devices on the training results. The calculation formula is as follows:

[0084]

[0085]

[0086] in, represents the predicted value; y represents the true value; b represents the sample index; c represents the classification index; B is the total number of single training samples; C represents the total number of categories; y b,c Represents the probability of the c-type sample of the b-th sample in the true value; represents the probability of the c-type sample of the predicted b-th sample; w c Represents the weight of class c in a single batch.

[0087] Step 3: Analysis of the inference results of substation fault equipment location. Ablation experiments are conducted on the data set to verify the network optimization capabilities of the channel coupling module, the Mask module, and the weighted cross entropy loss. × in the table indicates that the full connection layer, the dropout layer, and the cross entropy loss module are used to replace the corresponding original modules for testing. The ablation experiment results are shown in Table 1. The experimental results show that the introduction of the channel coupling module, the Mask module, and the weighted cross entropy loss can effectively improve the location effect of substation fault equipment, and the classification accuracy can reach 79.6%.

[0088] Table 1. Comparison results of ablation experiments

[0089]

[0090] As shown in Table 2, the performance comparison of the method GCC in the present invention and other mainstream target detection methods is compared. It can be seen that the classification accuracy of the GCC method in the present invention is significantly better than other mainstream detection methods, and the classification accuracy is improved by 19.5% compared with the classification method in TimesNet.

[0091] Table 2 Comparison of public datasets, self-built data and experiments

[0092]

[0093] Step 4: GCC substation fault location is completed.

[0094] like Fig. 9 The figure shows the transient ground current data of a substation before and after a fault. The last line is the data classification probability.

[0095] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A substation fault location method based on distributed grounding current decoupling analysis, characterized in that: The method comprises the following steps: Step 1: Establish a unidirectional grounding fault data set for different devices; sort out the grounding current data of the generated device faults through the current acquisition device, and the collected data includes fault samples S, time dimension L, and channel dimension d; Step 2: Construct a GCC deep learning model consisting of a data processor, a feature extractor, and a feature classifier, and use the GCC deep learning model to train and test the data set; the feature extractor includes a channel coupling module consisting of a graph similarity attention module, a self-attention feature extraction module, and a channel attention module; Step 3: Analyze the results of substation fault equipment location reasoning, conduct ablation experiments on the data set, and verify the network optimization capabilities of the channel coupling module, Mask module, and weighted cross entropy loss; Step 4: GCC substation fault location is completed.

2. A substation fault location method based on distributed ground current decoupling analysis according to claim 1, characterized in that: The step 2 constructs a GCC deep learning model including a data processor, a feature extractor and a feature classifier, and uses the GCC deep learning model to train and test the data set; wherein the feature extractor includes a channel coupling module consisting of a graph similarity attention module, a self-attention feature extraction module and a channel attention module; wherein: In the data processor, variational mode decomposition is performed on the time dimension data; In the feature extractor, the ground current data of different frequencies are downsampled in the time dimension through one-dimensional convolution to obtain downsampled data; The downsampled data is passed through the graph similarity attention module to extract the coupling features between channels; The downsampled data is passed through the self-attention feature extraction module to extract the temporal features of the data; The output of the graph similarity attention module and the output of the self-attention feature extraction module are input into the channel attention module for feature integration; At the end of the feature extractor, the time dimension features are expanded through a fully connected layer; In the feature classifier, the Mask module is used to suppress low-frequency feature information.

3. A substation fault location method based on distributed ground current decoupling analysis according to claim 2, characterized in that: In the data processor, the specific operation of performing variational mode decomposition on the time dimension data is: The ground current data with the size of S×d×L is decomposed into S×d×K×L, where K represents the different frequency data obtained by variational mode decomposition.

4. A substation fault location method based on distributed ground current decoupling analysis according to claim 3, characterized in that: In the feature extractor, the ground current data of different frequencies are downsampled in the time dimension through one-dimensional convolution to obtain the downsampled data in the following specific operations: Decompose the grounding current data with a size of S×d×L and the grounding current different frequency data with a size of S×d×K×L into each batch of data with a size of B, and obtain the grounding current data with a size of B×d×L and the grounding current different frequency data with a size of B×d×K×L; The ground current data with a size of B×d×L and the ground current data of different frequencies with a size of B×d×K×L are downsampled in the time dimension through a one-dimensional convolution with a convolution kernel size of 10 and a step size of 10, and the sizes of B×d×l and The downsampled data.

5. A substation fault location method based on distributed ground current decoupling analysis according to claim 4, characterized in that: The specific operation of extracting the coupling features between channels by passing the downsampled data through the graph similarity attention module is: The downsampled data of size B×d×l is transformed into B×l×d, and the query vector Q, key vector K and value vector V of size B×l×d are calculated through the graph convolutional network respectively; The cosine similarity is calculated for the query vector Q and the key vector K to evaluate the coupling relationship between channels, and the weight is used for softmax normalization to obtain the weight S. The cosine similarity calculation formula is as follows: Where S represents the cosine similarity weight; Similarity cos represents the adaptive cosine similarity calculation function; || || represents the modulus of the tensor; T represents the transposition operation; Multiply the value vector V by the weight S to perform a weighted operation, and obtain a deeper coupling feature B×l×d through the second layer of graph convolutional network; then, transform its dimension to feature H of size B×d×l G .

6. A substation fault location method based on distributed ground current decoupling analysis according to claim 5, characterized in that: The specific operation of extracting the temporal features of the downsampled data through the self-attention feature extraction module is as follows: The size is The downsampled data of are passed through the fully connected layer respectively, and the sizes are Query vector Q′, key vector K′, and value vector V′; Perform matrix multiplication on the query vector Q′ and the key vector K′, perform normalization operation through softmax, and obtain the weight; then multiply it with the value vector V′, and the output size is characteristics; finally The decomposition dimensions are summed and tied to obtain an output H of size B×d×l S .

7. A substation fault location method based on distributed ground current decoupling analysis according to claim 6, characterized in that: The specific operation of integrating the output of the graph similarity attention module and the output of the self-attention feature extraction module into the channel attention module is: The outputs of the two modules are concatenated and the obtained features are input into a two-dimensional convolution with a kernel size of 1×1 and a stride of 1; Perform global pooling on the features output by the convolutional layer and then output the features; Through a layer of compressed fully connected layer, the channel dimension is compressed to the original Through a layer of decompression fully connected layer, the channel dimension is restored; Finally, the output of the convolutional layer is weighted with this value, and the timing features and coupling features are fused to associate the two.

8. A substation fault location method based on distributed ground current decoupling analysis according to claim 7, characterized in that: In the feature classifier, the Mask module is used to suppress the low-frequency feature information. The specific operation is: First, the features are inactivated by Dropout. Then, a specific frequency between the maximum frequency and the minimum frequency is selected to calculate the suppression length s. Finally, a mask tensor is generated according to the length s, and then matrix multiplied with the original feature tensor to suppress its features. The formula is as follows: H f =|fft(H out )| (2) Where, fft represents the Fourier transform function; H out represents the hidden layer output; || represents the absolute value operation; H f It represents the frequency domain information after taking the absolute value; α represents the selection coefficient; f represents the frequency selection value.

9. A substation fault location method based on distributed ground current decoupling analysis according to claim 8, characterized in that: In step 3, the analysis of the reasoning results of the substation fault equipment location is carried out, and an ablation experiment is performed on the data set to verify the specific operations of the channel coupling module, the Mask module, and the network optimization capabilities of the weighted cross entropy loss: The entropy loss function is used to measure the accuracy of model prediction and reveal the difference between the predicted results and the actual results. Through data concentration, the one-way grounding faults generated by different devices are weighted to reduce the impact of different amounts of fault data of different devices on the training results. The calculation formula is as follows: in, represents the predicted value; y represents the true value; b represents the sample index; c represents the classification index; B is the total number of single training samples; C represents the total number of categories; y b,c Represents the probability of the c-type sample of the b-th sample in the true value; represents the probability of the c-type sample of the predicted b-th sample; w c Represents the weight of class c in a single batch.

Images