Power distribution network high-resistance grounding fault identification method based on deep feature fusion

Abstract

The invention discloses a power distribution network high-resistance grounding fault identification method based on deep feature fusion, and belongs to the technical field of power distribution network fault diagnosis and artificial intelligence deep learning. Comprising the following steps: performing deep mining on the waveform of a high-resistance fault signal by using discrete wavelet transform, and constructing a multi-scale time-frequency feature map and a global statistical feature matrix to enhance the expression of high-resistance fault features; designing a deep feature extraction network in combination with a lightweight residual network structure and a self-attention mechanism to realize fusion extraction of local and global time-frequency features; metric element learning is introduced to calculate a class prototype and the distance between the class prototype and the sample for the characteristics of various samples in a metric space, so that the construction of a high-resistance fault classifier is realized. According to the method, high-resistance fault identification tasks under different complex power distribution network operation conditions can be achieved, a good classification effect is still achieved when field data samples are few, and theoretical technical support is provided for application of the high-resistance fault identification method based on data driving in an actual system.

Description

Technical Field

[0001] The present invention relates to the field of distribution network fault diagnosis and artificial intelligence deep learning technology, and in particular to a distribution network high-resistance grounding fault identification method based on deep feature fusion. Background Art

[0002] Medium-voltage distribution networks are often affected by factors such as the natural environment and short overhead lines, which often form high-resistance grounding faults that contact concrete floors, grass, sand, and other media. Due to the high impedance characteristics of the grounding medium, the fault current is small and the fault characteristics are weak, making it difficult to be effectively identified and removed by traditional zero-sequence protection schemes. Once a high-resistance fault exists in the distribution network for a long time, it is easy to damage the insulation of the distribution line, which will indirectly cause fire and electric shock hazards. Therefore, detecting and eliminating high-resistance faults as soon as possible is of great significance to maintaining the safe and reliable operation of the distribution network.

[0003] Existing high-resistance fault identification methods are mainly divided into physical model methods and data-driven methods. The high-resistance fault identification method based on the physical model proposes a basis for fault detection by analyzing the changes in electrical quantities of high-resistance faults, and realizes the determination of high-resistance faults by comparing the index thresholds. Common fault electrical quantity information includes active power, zero-sequence voltage and zero-sequence current. This type of method has a simple structure and clear physical meaning, but the threshold setting is mostly based on expert experience, and it is impossible to fully utilize the key information contained in existing fault cases on site. On the other hand, the normal operating conditions of the distribution network such as capacitor switching, load switching and excitation inrush current present similar electrical characteristics to high-resistance faults, which makes it easy for the fault identification scheme to confuse the difference between faults and disturbances, resulting in insufficient reliability of the high-resistance fault identification scheme.

[0004] With the development of computer science and information technology, machine learning technology has brought opportunities for distribution network fault diagnosis, and emerging data-driven related algorithms have been increasingly used in high-resistance fault identification. However, in most existing methods, the expression and extraction of high-resistance fault information mostly only considers the local or global features of the sample, ignoring the deep fusion of fault signal features, which limits the improvement of high-resistance identification accuracy. On the other hand, most of the existing high-resistance fault identification models are trained on sufficient samples. In the actual high-resistance identification task, it is difficult to obtain sufficient samples to support the construction of classification detection models. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion, which solves the problem that the expression and extraction of high-resistance fault information mostly only consider the local features or global features of the samples, ignores the deep fusion of the features of the fault signal, and limits the improvement of the accuracy of high-resistance identification. On the other hand, most of the existing high-resistance fault identification models are trained on sufficient samples, and it is difficult to obtain sufficient samples to support the construction of classification detection models in actual high-resistance identification tasks.

[0006] To achieve the above object, the present invention provides a method for identifying high-resistance grounding faults in a distribution network based on deep feature fusion, comprising the following steps:

[0007] S1. Construct a high-resistance fault information representation structure. Based on the discrete wavelet transform analysis of the high-resistance fault signal waveform, deeply mine the fault information to obtain a multi-scale time-frequency feature map and a global statistical feature matrix, which together enhance the expression ability of the high-resistance fault information.

[0008] S2. Establish an expression network that integrates local and global time-frequency features, consisting of a lightweight residual module and a self-attention module, and construct a feature extraction network from the perspectives of graphic information and discrete information respectively;

[0009] S3. Introduce metric meta-learning to design high-resistance fault classifier.

[0010] Preferably, in step S1, based on the analysis of the high-resistance fault signal waveform by discrete wavelet transform, the fault information is deeply mined to obtain a multi-scale time-frequency feature map and a global statistical feature matrix, which together constitute an enhanced expression capability of the high-resistance fault information. The specific features include a multi-scale time-frequency feature map and a global statistical feature matrix.

[0011] Preferably, the steps for obtaining the multi-scale time-frequency feature map are as follows:

[0012] Perform 6-layer wavelet decomposition on the high-resistance electrical signal, and use the 3rd, 4th, 5th and 6th layer detail components after wavelet decomposition as fault features;

[0013] After denoising the detail components of each layer, a 32×32 feature map representing the temporal sequence relationship is constructed through the Gram angular field transform;

[0014] The four feature maps are concatenated to obtain a 64×64 multi-scale time-frequency feature map for the input of the subsequent local feature extraction network.

[0015] Preferably, the steps for obtaining the global statistical feature matrix are as follows:

[0016] Amplitude, mean, standard deviation, kurtosis, mean energy, energy entropy and skewness are used as statistical features to characterize the electrical waveform signal;

[0017] The waveform of the six-layer detail signal after discrete wavelet decomposition of the original signal is reconstructed and the upper and lower envelopes of the six branch signals are calculated;

[0018] The 7 statistical features of each branch signal of the reconstructed waveform and the upper and lower envelopes are calculated as the global statistical features of the high-resistance fault signal. Specifically, a feature of the 6 signals is packaged into a 6-dimensional vector to obtain the statistical feature matrix W TM ∈R 7*6 , similarly, we get the statistical characteristic matrix U of the upper and lower envelopes TM ∈R 7*6 and L TM ∈R 7*6 ; The matrix W TM , U TM and L TM Splicing, as the input of the subsequent global feature dimensionality reduction network.

[0019] Preferably, in step S2, the lightweight residual module:

[0020] High-resistance fault signals are usually represented in different time-frequency segments in the zero-sequence current signal, that is, the high-resistance fault features are scattered in the multi-scale time-frequency feature map. Since the image size of the multi-scale time-frequency feature map is small, a lightweight residual network structure is improved and designed based on the residual network model to avoid the network being too complex and resulting in a decrease in the generalization ability of the model, while reducing network parameters and computational complexity, making it easier to deploy and apply subsequent recognition algorithms. The improvements are mainly in terms of model reduction, using a convolution kernel of size 3×3, the convolution step size is set to 1, and the padding is set to 1 or 0. At the same time, the lightweight residual network only uses 3 convolution stacked residual unit structures, and eliminates the pooling operation of the intermediate layer. The global average pooling layer is retained at the end of the output layer, and a fully connected layer Y is formed. LRN ∈R 1*128 This is to facilitate the subsequent concatenation of feature vectors.

[0021] Preferably, in step S2, the self-attention module:

[0022] Unlike convolutional and recurrent neural networks that integrate context information through stacking, the SAM network obtains the dependency between sequence data by calculating the correlation coefficient between any two positions in the input data. By linearly transforming the input sequence X, we get the query Q = XW Q , key K = XW K The sum V = XW V Then, Q and K are calculated by scaling the dot product and using the SoftMax function. T, to obtain the self-attention weight matrix A (called attention score). Matrix A describes the correlation between all elements in the input feature sequence as a whole, that is, global information, which is used to enhance the expression of dominant features and reduce the interference of irrelevant redundant features. Finally, V is multiplied by the attention matrix A to obtain the output O, which represents the feature sequence weighted by global information. The present invention uses a self-attention module to extract key feature information from the global statistical features of the waveform signal for subsequent high-resistance fault identification.

[0023] The self-attention module network has four layers, each of which consists of a "multi-head" self-attention module, a feed-forward layer, an addition layer, and a normal layer. The input sequence of the self-attention module, where L is the length of the sequence, D I is the dimension of the sequence elements. Input sequence X a Through trainable linear projection Get sequence Output sequence of multi-head self-attention layer It represents the feature sequence obtained by weighted multiplication of the attention score and the sequence value. The calculation method is as follows:

[0024]

[0025] Where: h i refers to the i-th attention head; Refers to multiple linear transformation operations; and Represents the embedding X a The input is linearly transformed and projected to obtain the query Q, key K and value V corresponding to the sequence.

[0026] According to the size of the input data dimension, set the fully connected layer Y after the SAM network. SAM ∈R 1*128 Extract global time-frequency features. Finally, connect the local feature information and the global feature information together to obtain a deep fusion feature vector, and use the result as the output of the deep feature extraction network.

[0027] Preferably, in step S3, the high-resistance fault classifier is a prototype-updated classifier proposed in combination with the idea of ​​metric meta-learning to solve the problem that the prototype network model trained under scarce data distribution is prone to category center error, resulting in inaccurate classification and susceptibility to noise. The details are as follows:

[0028] S31. Metric Meta-Learning Classification Strategy:

[0029] In order to correct the fitting deviation of clustering feature points in the prototype network, find the prototype feature points in the class that cause the error, and use the unlabeled sample information in the query set to offset and update the prototype feature points of the existing dataset. Specifically, predict the category label for each query set sample, set the category with the highest prediction probability for each query set sample as the query set category, and add the pseudo label of this category to the query set. The calculation process is as follows:

[0030]

[0031] Where: is the sample x in the query set m The probability of belonging to category j; c j For the prototype point; is the embedding function; n K is the number of sample categories.

[0032] A reorganized support set is constructed by assigning label values ​​in the query set according to the "pseudo label" strategy, and then the modified prototype network is used for small sample high resistance fault identification. The new prototype point of each category is calculated using the new support set, and the prediction probability P j The calculation process is as follows:

[0033]

[0034] Where: Indicates the number of samples of class j in the support set; Map the function for the new dataset.

[0035] S32, classifier loss function:

[0036] Considering that the more dispersed the feature distribution of the same type of data samples in the metric space is, the more likely it is that the samples will overlap with other types, which may lead to unclear prototype boundaries. The present invention designs two different feature loss representation functions to improve the distribution of sample feature points in the feature space.

[0037] The first loss is called the nearest neighbor loss, which aims to reduce the distance between the samples within the class and the prototype cluster feature points, so that the samples of the same category are more concentrated in the feature space. The operation to achieve this goal is to minimize the variance between the sample and the prototype point, and calculate the prototype cluster feature point p marked as i iThe Euclidean distance between all feature points (called the nearest neighbor distance) is then used to obtain the global nearest neighbor loss value by superimposing the distance between all samples. The second loss is called the boundary loss, which aims to increase the distance between class feature points in the prototype dataset so that there are no samples of other categories near the prototype in the class. The operation to achieve this goal is to maximize the distance between samples of different categories and the class prototype, and calculate the class prototype p marked as i. i The Euclidean distance between all features marked as j (called the intrusion distance) is calculated, and then the error calibration function is constructed to calculate the difference between the nearest neighbor distance and the intrusion distance. If the error calibration is less than 0, it means that no samples of different categories have invaded the representation boundary of the class prototype; if the error calibration is greater than 0, it means that samples of different categories have invaded the representation range of the samples of this class.

[0038] In the model loss function training, the neighbor loss and boundary loss are superimposed to form the neighbor boundary loss to update and correct the prototype features of the sample. The final loss function L NB The specific mathematical expression is as follows:

[0039]

[0040] Where: y il is the type label; F(·) is the calculation of the Euclidean distance; p i is the prototype of the class with label i; is the high-dimensional feature vector of the support set samples with labels i and l; [A] + =max(A,0) indicates error calibration.

[0041] When i=l,y il =1, calculate the first loss (neighbor loss), when i≠l, y il = 0, the second loss (boundary loss) is calculated, which can more clearly clarify the prototype representation boundary and make the classification model more stable.

[0042] Therefore, the present invention adopts a distribution network high-resistance grounding fault identification method based on deep feature fusion using the above structure, which has the following beneficial effects:

[0043] The method of the invention designs a high-resistance fault recognition model with strong feature expression based on the time-frequency domain feature differences of the high-resistance fault signal after discrete wavelet decomposition. The model integrates the intrinsic information of the field samples to update the fault recognition strategy, has the ability to handle small sample diagnosis of the real distribution network, and improves the generalization performance and field application possibility of power artificial intelligence.

[0044] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of a 10kV power distribution simulation system including DG access according to a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion of the present invention;

[0046] Figure 2 It is a real-type test site and a brief topological schematic diagram of a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion in the present invention;

[0047] Figure 3 It is a schematic diagram of a small sample high-resistance fault identification application process of a distribution network high-resistance grounding fault identification method based on deep feature fusion of the present invention;

[0048] Figure 4 It is a schematic diagram of test results of different methods and sample numbers of a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion according to the present invention;

[0049] Figure 5 It is a schematic diagram of visualization results of t-distributed stochastic neighbor embedding (t-SNE) of different methods in the same type of identification method of a distribution network high-resistance grounding fault identification method based on deep feature fusion of the present invention;

[0050] (a) is a schematic diagram of the characteristic visualization results of the prototype network (method 2) of the same type of recognition method;

[0051] (b) is a schematic diagram of the characteristic visualization results of the prototype correction network (method 3);

[0052] (c) is a schematic diagram of the characteristic visualization results of the boundary loss prototype network (method 4);

[0053] (d) Schematic diagram of the characteristic visualization results of the same type of recognition method combined with the prototype correction classification strategy and boundary loss;

[0054] Figure 6 This is a schematic diagram of the recognition results under different feature extraction network conditions of a distribution network high-resistance grounding fault recognition method based on deep feature fusion in the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0056] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] Example

[0058] The present invention provides a method for identifying high-resistance grounding faults in a distribution network based on deep feature fusion, comprising the following steps:

[0059] S1. Construct a high-resistance fault information representation structure. Based on the discrete wavelet transform analysis of the high-resistance fault signal waveform, deeply mine the fault information to obtain a multi-scale time-frequency feature map and a global statistical feature matrix, which together enhance the expression ability of the high-resistance fault information.

[0060] S2. Establish an expression network that integrates local and global time-frequency features, consisting of a lightweight residual module and a self-attention module, and construct a feature extraction network from the perspectives of graphic information and discrete information respectively;

[0061] S3. Introduce metric meta-learning to design a high-resistance fault classifier. The high-resistance fault classifier is a prototype-updated classifier proposed by combining the metric meta-learning idea to address the problem that the prototype network model trained under scarce data distribution is prone to category center error, resulting in inaccurate classification and susceptibility to noise.

[0062] Specifically, build on the MATLAB / Simulink simulation platform Figure 1 The 10kV radial distribution network model is shown in Table 1. The neutral point grounding method through arc suppression coil is widely used because of its advantages in rapid fault removal. Therefore, the neutral point operation mode of the test system is set to grounding through arc suppression coil, with an overcompensation degree of 5%, that is, the arc suppression coil inductance parameter is 0.1543H and the resistance parameter is 5.115Ω. The system capacitance current is 32.257A. The capacity of distributed power sources DG1 and DG2 is 2MVA, and they are connected to feeder L1 and feeder L6 respectively through isolation transformers.

[0063] Table 1 10kV distribution simulation system line parameters

[0064]

[0065] A zero-sequence current waveform information recording device is installed at the outlet of the distribution system feeder, and the sampling frequency is set to 10kHz. The high-resistance fault and normal disturbance working conditions are simulated and analyzed. In order to effectively record the recorded data before and after the working condition changes and avoid frequent detection startup, the present invention uses the zero-sequence voltage greater than 15% of the normal operating voltage as the algorithm startup criterion. The waveform data intercepts the zero-sequence current waveform of the first 2 and last 6 cycles of the detection algorithm startup as the information input of the recognition model, so as to construct a multi-scale time-frequency feature map and a global statistical feature matrix for feature extraction and category determination.

[0066] Previous studies have shown that the working conditions that are extremely difficult to distinguish from high-resistance faults are mostly three normal disturbances: capacitor switching, load switching, and excitation inrush current. Therefore, in the example test of the present invention, only four working conditions are distinguished and identified: high-resistance fault, capacitor switching, load switching, and excitation inrush current. In addition, considering that high-resistance grounding faults are often accompanied by arcing, arc characteristics should be considered when establishing a high-resistance grounding fault model. The present invention uses the Emanuel model as a high-resistance fault model, and changes the parameters in the Emanuel high-resistance model to change the transition resistance size, asymmetry, and nonlinearity of the high-resistance fault. In order to verify the effectiveness of the proposed algorithm, according to the different recording time, fault initial phase angle, fault location, and operating condition parameter values, Figure 1 A total of 2700 sets of simulation data samples were generated in the distribution system. Among them, the power value of capacitor switching is 1.2~6Mvar; the active power of load switching is 1.5~2.5MW, and the reactive power is 0.3~0.5Mvar; the connection mode of excitation surge transformer is YN / D1 and YN / D11. Table 2 shows the data distribution of each type of working condition in the simulation data set samples. 80% of the samples of each type of working condition are used for model training and the remaining 20% ​​of the samples are used for model testing, which are recorded as simulation training set and simulation test set respectively.

[0067] In order to verify the application value of the proposed method in the field distribution network, data were collected at a real-type test site of a distribution network of a certain power research institute in China as a real-type test set. Figure 2 As shown. The real test site consists of 4 feeders, which can model the real simulation of various high-resistance fault types and disturbances in different grounding systems. A fault recorder with a sampling frequency of 6.4kHz is installed at the head end of the line to simulate the sampling of the micro PMU. By setting different types of grounding methods, fault initial phase angles, high-resistance faults at fault locations, and disturbance data, a total of 150 groups were tested and verified. Table 2 shows the data distribution of each type of working condition in the real test set sample.

[0068] Table 2 Distribution of samples of different working conditions in the data set

[0069]

[0070] The proposed method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion under small sample conditions is as follows: Figure 3 When there is less labeled data, the flowchart of high-resistance fault identification using the proposed method can be divided into three parts: data preparation, model training, and test verification. The detailed description of each part is as follows. The specific process is:

[0071] 1) Raw data preparation

[0072] The goal of high-resistance fault recognition is to correctly predict the class label of each unlabeled sample. These class states are usually determined by the dataset Τ t The electrical information waveform is shown in FIG. t For each unknown category to which the samples in the dataset Φ belong, it is assumed that t A small number of labeled samples (even only one sample, usually served as an actual field dataset) are collected in the dataset. In addition, a large-scale labeled dataset Φ b Used for training and testing (usually served by a simulated data set). b Randomly select most of the samples and put them into the training data set Φ b-t The remaining samples are put into the validation data set Φ b-v middle.

[0073] 2) Metamodel training

[0074] Before model training begins, set the training rounds and test rounds, and initialize the meta-learner model parameters. Calculate the loss function in each training task, and update the network parameters through back propagation. In order to improve the test efficiency, set the test threshold parameter β, that is, perform a test every β training rounds. The test task is constructed in the same way as the training task. Then, calculate the ratio of the number of predicted maximum probability category labels that are the same as the actual label to the total number of query samples as the recognition accuracy.

[0075] 3) Update test verification

[0076] During the test, by t As the support set, t Then, the best meta-learner obtained by the meta-learning process is used to predict the test set Τ t The label of each sample in . If Τ is known t If the true labels of the data samples in , the test accuracy can be obtained to evaluate the proposed high-resistance fault recognition performance.

[0077] The proposed classification model (method 1) is compared and analyzed with the existing common small sample classification processing methods. The benchmark models for comparison are the prototype network (method 2), the prototype correction network (method 3) and the boundary loss prototype network (method 4). In order to ensure the fairness of the comparison, the feature extractors and parameter settings of all methods are the same as those of the present invention. At the same time, the domain adaptive migration network (method 5) and the generative adversarial network (method 6) that can process small samples are added as control groups for comparative analysis. In order to highlight the recognition performance of different methods under small sample conditions, a direct training method for field data (method 7) is also used for comparative testing. The diagnostic results of different methods and sample numbers on the field test set are shown in the figure. Figure 4 As shown in the figure. Among them, 1-shot, 5-shot, 10-shot, 15-shot and 20-shot refer to randomly selecting 1, 5, 10, 15 and 20 field data samples in the test set to update the model, and then testing the remaining samples after removing the updated samples from the 150 samples. The test process is carried out for 5 rounds, and the average of all results is taken as the test accuracy for display.

[0078] from Figure 4 It can be seen that the recognition accuracy based on metric meta-learning methods (methods 1, 2, 3 and 4) increases with the number of samples L. NB The value of K increases. N When the value is 10, it begins to stabilize. N The value of K hardly increases after reaching 15. In addition, in domain adaptive transfer learning (method 5), the recognition accuracy increases with K N The value of increases, but it is always much lower than metric meta-learning. At the same time, the accuracy of the generative adversarial network (method 6) in this study is less than 85%. Compared with the prototype network, the recognition accuracy of the method proposed in the present invention for 1-shot, 5-shot, 10-shot, 15-shot and 20-shot in the field test set is improved by 8.77%, 7.39%, 6.58%, 5.46% and 4.25% respectively. It can be found that the less data available on site, the greater the improvement in the recognition accuracy of the method proposed in the present invention, indicating that the proposed classification strategy has a better classification effect for tasks with fewer samples. Considering K N = 10, the accuracy of fault diagnosis has exceeded 95%. N The value of is increasing, but the improvement of recognition accuracy is not obvious compared with the previous one. At the same time, in the paradigm of small sample learning, high-precision recognition is more meaningful when the number of support sets is small. Therefore, considering the scarcity of field fault signals, K will be selected by default in subsequent tests. N =10 as the pre-analysis condition.

[0079] When K N=10, the recognition accuracy of the proposed method (method 1) in the field test set is 96.97%, which is 6.58%, 3.71%, 3.22%, 23.54%, 27.69%, and 45.62% higher than that of methods 2-7, respectively. Compared with the prototype correction network (method 3), the proposed method corrects the spatial distribution of samples by using the nearest boundary loss, clarifies the prototype representation boundary, and solves the sample distribution problem in the metric space to a certain extent. Compared with the boundary loss prototype network (method 4), the prototype correction classification strategy increases the query set data in the classification process, expands the information content of the model, and improves the fault recognition accuracy. At the same time, the performance of method 1, which combines the advantages of the two schemes, is higher than any of them, proving the application advantages of the proposed method.

[0080] In order to further demonstrate the advantages of the small sample prototype high resistance identification method proposed in this invention, Figure 5 The visualization results of t-distributed stochastic neighbor embedding (t-SNE) of different methods are shown. Due to the characteristics of the prototype network (method 2), its fault feature distribution shows obvious clustering phenomenon, but there are still classification sample confusion and superposition phenomena, such as Figure 5 The prototype correction network (method 3) combines the overall information of unlabeled samples in the feature offset query set to make the distribution of each fault category more clustered, as shown in (a). Figure 5 As shown in (b) in the figure. The boundary loss prototype network (method 4) clarifies the boundaries of the prototype representation and makes the boundaries between fault clusters more obvious, as shown in Figure 5 The proposed method combines the advantages of prototype correction classification strategy and boundary loss, better adjusts the spatial distribution of features, makes the distance within the domain more concentrated, and the distance between classes more dispersed, as shown in (c). Figure 5 As shown in (d) in the figure, there is an obvious separation effect even for the difficult-to-separate categories, so that the features of each fault type have the best cluster distribution.

[0081] In order to compare the advantages of the multi-feature fusion extraction proposed in this study as a feature extractor, the present invention selects stacked autoencoders (SAE), self-attention networks (SAM), convolutional neural networks (CNN), lightweight residual networks (LRN), CNN+SAE and LRN+SAM as feature extraction networks for comparison. Among them, the SAE and SAM networks focus on the global statistical matrix, and the CNN and LRN networks focus on multi-scale time-frequency feature maps. In the parameter setting of the feature extraction network, CNN, LRN and SAM have the same parameters as the multi-feature fusion extraction module of the present invention. Figure 6The recognition results under different feature extractors are shown. The recognition accuracy of the deep feature fusion network proposed in the present invention is 23.75%, 20.00%, 15.00% and 11.25% higher than that of SAE, SAM, CNN and LRN, respectively. This is mainly because the deep feature fusion network enhances the feature extraction capability and fully integrates local image features and global statistical features, thereby improving the overall recognition performance. From the perspective of standard deviation, the deep feature fusion network has the lowest standard deviation. Compared with CNN+SAE, the standard deviation is reduced by 0.04, and it has stronger robustness. The main reason is that the global statistical information of the input signal is fully learned through the attention mechanism, which is not restricted by local interactions. At the same time, LRN also solves the problem of limited receptive field of convolution kernels to a certain extent.

[0082] Therefore, the present invention adopts the above-mentioned method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion, and designs a high-resistance fault identification model with strong feature expression based on the time-frequency domain feature differences of the high-resistance fault signal after discrete wavelet decomposition. It integrates the intrinsic information of field samples to update the fault identification strategy, has the ability to handle small sample diagnosis of real distribution networks, and improves the generalization performance and field application possibility of power artificial intelligence.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion, characterized in that: The following steps are involved: S1. Construct a high-resistance fault information representation structure. Based on the discrete wavelet transform analysis of the high-resistance fault signal waveform, deeply mine the fault information to obtain a multi-scale time-frequency feature map and a global statistical feature matrix, which together enhance the expression ability of the high-resistance fault information. S2. Establish an expression network that integrates local and global time-frequency features, consisting of a lightweight residual module and a self-attention module, and construct a feature extraction network from the perspectives of graphic information and discrete information respectively; S3. Introduce metric meta-learning to design high-resistance fault classifier.

2. According to claim 1, a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion is characterized by: In step S1, the specific features include a multi-scale time-frequency feature map and a global statistical feature matrix.

3. According to claim 2, a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion is characterized in that: The steps for obtaining the multi-scale time-frequency feature map are as follows: Perform 6-layer wavelet decomposition on the high-resistance electrical signal, and use the 3rd, 4th, 5th and 6th layer detail components after wavelet decomposition as fault features; After denoising the detail components of each layer, a 32×32 feature map representing the temporal sequence relationship is constructed through the Gram angular field transform; The four feature maps are concatenated to obtain a 64×64 multi-scale time-frequency feature map for the input of the subsequent local feature extraction network.

4. According to claim 3, a method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion is characterized in that: The steps for obtaining the global statistical feature matrix are as follows: Amplitude, mean, standard deviation, kurtosis, mean energy, energy entropy and skewness are used as statistical features to characterize the electrical waveform signal; The waveform of the six-layer detail signal after discrete wavelet decomposition of the original signal is reconstructed and the upper and lower envelopes of the six branch signals are calculated; Pack a feature of the six signals into a six-dimensional vector and get the statistical feature matrix W TM ∈R 7*6 , similarly, we get the statistical characteristic matrix U of the upper and lower envelopes TM ∈R 7*6 and L TM ∈R 7*6 ; The matrix W TM , U TM and L TM Splicing, as the input of the subsequent global feature dimensionality reduction network.

5. The method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion according to claim 1 is characterized in that: In step S2, the lightweight residual module uses a convolution kernel of size 3×3, the convolution step is set to 1, and the padding is set to 1 or 0; At the same time, the lightweight residual network only uses 3 convolutional stacked residual unit structures; At the end of the output layer, the global average pooling layer is retained to form a fully connected layer Y LRN ∈R 1*128 This is to facilitate the subsequent concatenation of feature vectors.

6. The method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion according to claim 1 is characterized in that: In step S2, the self-attention module includes four layers, each of which consists of a multi-head self-attention module, a feedforward layer, an addition layer and a normal layer. The specific acquisition steps are as follows: set up The input sequence of the self-attention module, where L is the length of the sequence, D I is the dimension of the sequence elements; Input sequence X a Through trainable linear projection Get sequence Output sequence of multi-head self-attention layer It represents the feature sequence obtained by weighted multiplication of the attention score and the sequence value. The calculation method is as follows: Where: h i refers to the i-th attention head; Refers to multiple linear transformation operations; and Represents the embedding X a Perform linear transformation projection on the input to obtain the query Q, key K and value V corresponding to the sequence; According to the size of the input data dimension, set the fully connected layer Y after the SAM network. SAM ∈R 1*128 Extract global time-frequency features; The local feature information and the global feature information are connected together to obtain a deep fusion feature vector, and its result is used as the output of the deep feature extraction network.

7. The method for identifying high-resistance grounding faults in distribution networks based on deep feature fusion according to claim 1 is characterized in that: In step S3, the specific steps of introducing metric meta-learning to design a high-resistance fault classifier are as follows: S31, metric meta-learning classification strategy, predicting a category label for each query set sample, setting the category with the highest prediction probability for each query set sample as the query set category, and adding the pseudo label of the category to the query set; Prediction probability The calculation process is as follows: Where: is the sample x in the query set m The probability of belonging to category j; c j For the prototype point; is the embedding function; n K is the number of sample categories; By assigning label values ​​in the query set according to the "pseudo label" strategy to construct a reorganized support set, the new prototype point of each category is calculated using the new support set, and the prediction probability P is j The calculation process is as follows: Where: Indicates the number of samples of class j in the support set; Mapping functions for new datasets; S32, the classifier loss function includes neighbor loss and boundary loss. The neighbor loss and boundary loss are superimposed to form the neighbor boundary loss to update and correct the prototype features of the sample. The final loss function L NB The specific mathematical expression is as follows: Where: y il is the type label; F(·) is the calculation of the Euclidean distance; p i is the prototype of the class with label i; is the high-dimensional feature vector of the support set samples with labels i and l; [A] + =max(A,0) indicates error calibration.

Images