Decision-making basis analysis method, device and equipment for high-resistance grounding fault identification model

Abstract

The present application discloses a decision basis analysis method, device and equipment for a high-resistance ground fault identification model, the method comprising: training an initial fault identification model based on unlabeled zero-sequence current data to obtain a preset high-resistance ground fault identification model, the preset high-resistance ground fault identification model comprising an encoder and a decoder; analyzing the importance of the encoding features output by the encoder to the model identification result, and calculating the corresponding global Shapley value; normalizing the encoding vector using an instance normalization algorithm, and inputting it into the decoder for decoding analysis to obtain a decoded waveform; performing spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyzing the model decision basis based on the global Shapley value to obtain the model identification basis. The present application can solve the technical problems that the existing ground fault identification model has a large demand for labeled data and the model lacks interpretability, resulting in the lack of pertinence and reliability of model training.

Description

Technical Field

[0001] The present application relates to the technical field of distribution network fault analysis, and in particular to a decision basis analysis method, device and equipment for a high-resistance grounding fault identification model. Background Art

[0002] The structure of distribution network is becoming more and more complex, and the probability of failure is also increasing. Single-phase grounding fault is one of the most important fault types in distribution network. When a single-phase grounding fault occurs in distribution network, it is often accompanied by arcing and generates arc grounding overvoltage. The amplitude of this voltage is high. If the action time is long, with the increase of feeders and the increase of capacitive current, the overvoltage generated during the operation of high-resistance grounding fault can easily cause new grounding points in system equipment, which can easily cause phase-to-phase short circuit or two-point and multi-point grounding faults, causing the accident to further expand. Therefore, accurate fault identification is of great significance to the stable operation of distribution network.

[0003] The main identification methods for high-resistance grounding faults in power grids include time domain method, frequency domain method, time-frequency domain method and artificial intelligence method. The time domain method focuses on the unique characteristics of voltage and current signals in the time domain, which have obvious physical properties; the frequency domain method is based on the high and low frequency components of voltage and current signals to distinguish high-resistance grounding faults from external interference; and the high-resistance grounding fault HIF intelligent identification method, which combines signal processing technology with artificial intelligence algorithms, realizes the effective processing of massive data through adaptive learning of deep features, avoiding the problem of artificial extraction of fault features being limited by prior experience.

[0004] However, the AI-based HIF identification method not only requires a large amount of labeled data, but also the training process belongs to the black box mode, which leads to the lack of interpretability of the model. Therefore, it is difficult to make a scientific and reliable analysis of the high-resistance grounding fault identification, and it cannot provide effective decision-making guidance for the training of fault identification models in specific scenarios. Summary of the invention

[0005] The present application provides a decision basis analysis method, device and equipment for a high-resistance grounding fault identification model, which is used to solve the technical problems that the existing grounding fault identification model has a large demand for labeled data and the model lacks interpretability, resulting in a lack of pertinence and reliability in model training.

[0006] In view of this, the first aspect of the present application provides a decision basis analysis method for a high-resistance grounding fault identification model, including:

[0007] Training an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder;

[0008] Analyzing the importance of the encoding features output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculating the corresponding global Shapley value;

[0009] After normalizing the coding vector composed of the plurality of coding features using an instance normalization algorithm, the coding vector is input into the decoder for decoding analysis to obtain a decoding waveform;

[0010] The spectrum analysis is performed by comparing the decoded waveform with the original fault waveform, and the model decision basis is analyzed based on the global Shapley value to obtain the model identification basis.

[0011] Preferably, the training of the initial fault identification model based on the unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault identification model includes:

[0012] Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data;

[0013] Pre-training an initial encoder and an initial decoder using the unlabeled zero-sequence current data to obtain an encoder and a decoder;

[0014] A nonlinear mapping is established between the encoder, the decoder and a high-resistance grounding fault label of a preset fully connected network layer to generate a preset high-resistance grounding fault identification model.

[0015] Preferably, the method of using the unlabeled zero-sequence current data to pre-train an initial encoder and an initial decoder to obtain an encoder and a decoder further includes:

[0016] Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave;

[0017] Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave;

[0018] An objective function of the encoder is constructed according to the fitting error and the waveform similarity.

[0019] Preferably, analyzing the importance of the coding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model and calculating the corresponding global Shapley value includes:

[0020] Constructing additive explanatory models based on Shapley values ​​of game theory;

[0021] Analyzing the importance of each encoding feature of each sample output by the encoder to the model recognition result according to the additive interpretation model to obtain a local Shapley value;

[0022] The local Shapley values ​​corresponding to the coding features at the same position of all samples are averaged by absolute values ​​to obtain a global Shapley value.

[0023] Preferably, the analysis of the importance of the coding features output by the encoder to the model recognition result and calculation of the corresponding global Shapley value, after a plurality of the coding features form a coding vector, further includes:

[0024] Arrange the global Shapley values ​​of all samples in descending order to obtain a Shapley value sequence.

[0025] The second aspect of the present application provides a decision basis analysis device for a high-resistance grounding fault identification model, comprising:

[0026] A model generation unit, used for training an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder;

[0027] A feature analysis unit, used to analyze the importance of the coding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculate the corresponding global Shapley value;

[0028] A decoding processing unit, configured to normalize the coding vectors composed of the plurality of coding features using an example normalization algorithm, and then input the normalized vectors into the decoder for decoding analysis to obtain a decoding waveform;

[0029] The basis analysis unit is used to perform spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyze the model decision basis based on the global Shapley value to obtain the model identification basis.

[0030] Preferably, the model generating unit is specifically used for:

[0031] Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data;

[0032] Pre-training an initial encoder and an initial decoder using the unlabeled zero-sequence current data to obtain an encoder and a decoder;

[0033] A nonlinear mapping is established between the encoder, the decoder and a high-resistance grounding fault label of a preset fully connected network layer to generate a preset high-resistance grounding fault identification model.

[0034] Preferably, it also includes: a training function construction unit, which is specifically used to:

[0035] Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave;

[0036] Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave;

[0037] An objective function of the encoder is constructed according to the fitting error and the waveform similarity.

[0038] Preferably, the feature analysis unit is specifically used for:

[0039] Constructing additive explanatory models based on Shapley values ​​of game theory;

[0040] Analyzing the importance of each encoding feature of each sample output by the encoder to the model recognition result according to the additive interpretation model to obtain a local Shapley value;

[0041] The absolute average of the local Shapley values ​​corresponding to the coding features at the same position of all samples is calculated to obtain a global Shapley value.

[0042] A third aspect of the present application provides a decision basis analysis device for a high-resistance grounding fault identification model, the device comprising a processor and a memory;

[0043] The memory is used to store program code and transmit the program code to the processor;

[0044] The processor is used to execute the decision basis analysis method of the high-resistance grounding fault identification model described in the first aspect according to the instructions in the program code.

[0045] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0046] In the present application, a decision basis analysis method for a high-resistance grounding fault identification model is provided, including: training an initial fault identification model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault identification model, wherein the preset high-resistance grounding fault identification model includes an encoder and a decoder; analyzing the importance of the coding features output by the encoder to the identification result of the preset high-resistance grounding fault identification model, and calculating the corresponding global Shapley value; normalizing a coding vector composed of multiple coding features using an instance normalization algorithm, and inputting the code into a decoder for decoding analysis to obtain a decoded waveform; performing spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyzing the model decision basis based on the global Shapley value to obtain a model identification basis.

[0047] The decision-making basis analysis method of the high-resistance ground fault identification model provided in this application uses unlabeled zero-sequence current data to perform unsupervised pre-training on the encoder model, and based on this, builds a model for high-resistance ground fault identification. The model does not require a large amount of input labeled data, but can still meet the fault identification needs; in addition, the model is interpreted based on the Shapley value and instance normalization algorithm to determine the basis for the model to perform fault identification analysis; it can provide a reference theoretical support with pertinence and reliability for model training, thereby improving the accuracy and credibility of model identification. Therefore, this application can solve the technical problems that the existing ground fault identification model has a large demand for labeled data and the model lacks interpretability, resulting in a lack of pertinence and reliability in model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flowchart of a decision basis analysis method for a high-resistance ground fault identification model provided in an embodiment of the present application;

[0049] Figure 2 A schematic diagram of the structure of a decision basis analysis device for a high-resistance grounding fault identification model provided in an embodiment of the present application;

[0050] Figure 3 An example diagram of a preset high-resistance ground fault identification model structure provided in an embodiment of the present application;

[0051] Figure 4 A schematic diagram of the circuit structure of a radial distribution network model provided for the application example of this application;

[0052] Figure 5 Schematic diagram of the encoder network structure provided for the application example of this application;

[0053] Figure 6 A global Shapley value ranking diagram corresponding to different feature vectors provided in the application example of this application;

[0054] Figure 7 Schematic diagram of the overall framework of the explainable high-resistance fault identification model provided for the application example of this application. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0056] For easier understanding, see Figure 1, an embodiment of a decision basis analysis method for a high-resistance grounding fault identification model provided by the present application includes:

[0057] Step 101 : training an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder.

[0058] Furthermore, step 101 includes:

[0059] Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data;

[0060] The initial encoder and the initial decoder are pre-trained using unlabeled zero-sequence current data to obtain an encoder and a decoder;

[0061] A nonlinear mapping between the encoder, decoder and the high-resistance grounding fault label of the preset fully connected network layer is established to generate a preset high-resistance grounding fault recognition model.

[0062] Furthermore, the initial encoder and the initial decoder are pre-trained using unlabeled zero-sequence current data to obtain an encoder and a decoder, and further comprising:

[0063] Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave;

[0064] Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave;

[0065] The objective function of the encoder is constructed based on the fitting error and waveform similarity.

[0066] It should be noted that the preset high-resistance grounding fault identification model of the present embodiment is mainly composed of an encoder, a decoder and a preset fully connected network layer, and unsupervised training is performed using a pre-training method of model migration. During this process, a large amount of zero-sequence current data in the distribution network is relatively easy to obtain, so that the encoder can learn unsupervisedly and capture the key features of the zero-sequence current data of the distribution network, and generate a compact code; in this way, the model can have a general feature extraction capability for the relevant data of the distribution network; then, a nonlinear mapping between the encoding features and the high-resistance grounding fault label is established through the preset fully connected network layer, and the preset high-resistance grounding fault identification model can be generated.

[0067] The model's encoder is responsible for converting the input data into an encoding in the latent representation space, while the decoder attempts to restore the encoding to the original input. Through this process, the autoencoder achieves self-reconstruction of the input data, with the goal of minimizing the reconstruction error, that is, ensuring that the decoded data is as close to the original input as possible. By training the autoencoder with the goal of minimizing the difference between the input data waveform and its reconstructed waveform, the encoder's training objective function can be constructed:

[0068] F=SSE+SW

[0069] Among them, F is the difference between the overall input data and the reconstructed waveform, SSE is the fitting error, and SW is the waveform similarity. The calculation process of the two is:

[0070]

[0071]

[0072] Among them, i f 、i s are the input waveform data and the reconstructed waveform data, k, N T / 2 are the kth sampling point and the total number of sampling times within half the power frequency cycle, ω k is the calculation weight of the kth sampling point. The value range of SW is [0,1]. The value indicates the degree of similarity. The smaller the value, the higher the similarity of the two waveforms.

[0073] After using the autoencoder to generate compact codes to capture the key features in the zero-sequence current data, the model needs to establish a nonlinear mapping from the code features to the high-resistance grounding fault labels, which is completed in the preset fully connected network layer. Moreover, using a small amount of labeled data to supervise the training of the fully connected network layer can achieve a complex mapping between the code and the high-resistance grounding fault label, and obtain a preset high-resistance grounding fault recognition model with higher accuracy.

[0074] Since the construction and use of the preset high-resistance grounding fault identification model are both black-box features, the user cannot determine what data or characteristics of the distribution network the model is based on for fault analysis and identification results; naturally, it is impossible to make targeted model adjustments and flexible applications according to the characteristics of the actual scenario. Therefore, this embodiment provides an analysis scheme for interpreting the model, finding the decision basis of the model for high-resistance grounding fault identification, and better guiding the training and application of the model in specific scenarios.

[0075] Step 102: Analyze the importance of the encoding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculate the corresponding global Shapley value.

[0076] Further, step 102 includes:

[0077] Constructing additive explanatory models based on Shapley values ​​of game theory;

[0078] According to the additive interpretation model, the importance of each encoding feature of each sample output by the encoder to the model recognition result is analyzed to obtain the local Shapley value;

[0079] The absolute average of the local Shapley values ​​corresponding to the coding features at the same position of all samples is calculated to obtain the global Shapley value.

[0080] Furthermore, step 102 further includes:

[0081] Arrange the global Shapley values ​​of all samples in descending order to obtain a Shapley value sequence.

[0082] It should be noted that the Shapley value is used in game theory to evaluate the contribution to the benefits. In this embodiment, the Shapley value of the encoded feature vector extracted from each sample is calculated to analyze its importance to the model prediction result.

[0083] The core idea is to sample the additive explanation model g(x) constructed based on the Shapley value of game theory to fit the complex model f(x), that is, the preset high-resistance grounding fault identification model of this embodiment, and provide a feasible explanation scheme for the prediction results of the model. The specific expression is:

[0084]

[0085] Among them, n is the number of features of sample x, φ0 is the model's prediction benchmark value for the sample, which is also the mean of the model's prediction results for all samples, and φ j is the jth feature x of sample x j The Shapley value of is used to quantify the feature x j The degree of influence on the model recognition results.

[0086] Therefore, the prediction result of the model for any sample can be expressed as the sum of the prediction benchmark value and the Shapley value of all the coding features of the sample, and the coding feature vector of the sample can correspond to multiple influence values, which can be simply understood as multiple Shapley values. The Shapley value is used to measure the importance of the coding feature to the model prediction result, reflecting the contribution of the coding feature to the model prediction result. Its specific calculation principle can be expressed as:

[0087]

[0088] Among them, {x 1 ,x 2 ,......,x n} represents the encoded feature vector of the n-dimensional sample x, that is, the feature set, which contains n encoded features, and S does not contain the encoded feature x j The feature subset of |S| is the number of feature elements in set S, U is the union operation, and f x (S∪{x j})、fx (S) respectively represent the encoding features x j and does not contain the encoded feature x j The prediction results of the model under the condition of j The first part on the right side of the calculation formula represents the weight. Encoded feature x j The Shapley value of φ j Defined as the mean of its marginal contribution in different subsets of encoded features.

[0089] For any given sample, the Shapley value of each coded feature reflects the degree of influence of the feature on the prediction result and the positive or negative influence. That is, the feature with a larger absolute value of the Shapley value has a greater impact on the model prediction result. The positive or negative value of the Shapley value indicates that the feature will increase or decrease the output result of the model, which greatly increases the local interpretability of the complex model f(x). That is, the Shapley value calculated by a single coded feature belongs to the local Shapley value. Based on this, if a feature x of all samples is j The absolute value of the Shapley value is taken as the average value as an indicator to measure the global importance of the feature, that is, the complex model f(x) can be globally explained. Therefore, the calculation process of the global Shapley value is expressed as:

[0090]

[0091] Among them, φ j (x i ) is the i-th sample x i The j-th encoded feature x j The Shapley value, that is, the local Shapley value, Y j is the jth encoded feature x j The corresponding global Shapley value, m is the total number of samples.

[0092] Evaluating the influence of coding features on the model fault identification results based on Shapley values ​​can greatly increase the interpretability and credibility of the model. This embodiment can also sort the global Shapley values ​​in descending order to analyze the coding features that have the greatest impact on the model prediction results.

[0093] Step 103: After normalizing the coding vector composed of multiple coding features using an instance normalization algorithm, the coding vector is input into a decoder for decoding analysis to obtain a decoding waveform.

[0094] Using the instance normalization algorithm to process each vector of encoded features one by one, the mean of the corresponding encoded feature vector can be reduced to 0 and the variance can be normalized to 1. See Figure 3, the features expressed by the vector in the code can be removed after instance normalization, because the dimension of each vector represents a certain coding feature extracted by 1-DCNN. Assume Figure 3 The rectangular box in the middle is the third harmonic feature. The third harmonic feature of HIF will be obvious, while the third harmonic feature of Non-HIF will not be obvious. If the instance normalization is performed, these features can be removed. The instance normalization process independently normalizes each sample of the encoded feature vector. The specific instance normalization process can be expressed as:

[0095]

[0096]

[0097]

[0098] Where c is the number of vectors, N is the number of eigenvectors, μ c is the mean of the current eigenvector, σ c is the standard deviation of the current eigenvector, x ic is the current encoding feature vector, is the new encoded feature vector after instance normalization, and ε is a small constant, usually 1e -5 , or a smaller value. The new encoded feature vector after instance normalization is input into the decoder for decoding, and the output waveform of the decoder, that is, the decoded waveform, can be obtained.

[0099] Step 104 , performing spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyzing the model decision basis based on the global Shapley value to obtain the model identification basis.

[0100] The original fault waveform is the waveform data input to the encoder, which is used to reflect the data of the high-resistance grounding fault; the decoded waveform reconstructed by the decoder is compared with the original fault waveform to analyze the spectrum difference; and attribution analysis is performed based on the global Shapley value to determine the contribution of the decoding vector that does not pass the input preset fully connected network layer to the prediction and recognition results, and the fault characteristics in the frequency domain when the high-resistance grounding fault HIF test occurs are combined to explain the recognition decision basis of the pre-trained high-resistance grounding fault recognition model based on the feature extractor based on the autoencoder and the fully connected neural network classifier.

[0101] For ease of understanding, this application also provides an application example of the decision basis analysis method of the high-resistance grounding fault identification model, please refer to Figure 4 ,The radial distribution network model is established using EMTDC / PSCAD simulation software. The system frequency is 50Hz, the sampling rate is 4kHz, and the parameters of the cable lines and overhead lines are shown in Table 1.

[0102] Table 1 Line parameters

[0103]

[0104]

[0105] In addition, Table 2 describes the HIF or HIF interference events that occur at different fault locations (FP) and on different lines. The data in Table 2 are divided into unlabeled training sets and labeled training sets at a ratio of 9:1. 1800 unlabeled data samples are used to pre-train the autoencoder, and 200 labeled data are used to train and adjust the fully connected neural network. Among them, capacitor switching (CS) adopts a parallel three-phase capacitor model, excitation inrush current (IC) is simulated by a single-phase transformer without load, low impedance fault (LIF) is simulated by a low resistance model (5Ω-100Ω), load switching (LS) adopts a three-phase asymmetric load model, and the HIF model is the Emanuel model. In addition, in order to meet the actual engineering situation, the asynchronous closing of CS is also added in the experiment to simulate the transient situation of non-fault. In this study, three-phase asynchronous closing means that phase A is connected to the system first, and phases B and C are connected to the system at the same time with the same delay. The initial fault angle is set to 0°30°60°90°120°.

[0106] Table 2 HIF and disturbance event samples

[0107]

[0108] The structure of the high-resistance grounding fault identification model pre-trained in this application example is as follows: Figure 3 As shown, the fully connected neural network structure is 108×138×168×38×2. Specifically, the encoder is 1-DCNN, and its structure can be found in Figure 5 The structure of the decoder is symmetrical with that of the encoder, and its parameters are shown in Table 3.

[0109] Table 3 Autoencoder related parameters

[0110]

[0111]

[0112] This application example uses 1-DCNN to directly operate on the original zero-sequence current signal, which can effectively capture the local features in the original zero-sequence current signal, and learn the patterns and changes between adjacent time steps in the input signal through convolution operations. This enables the model to better understand the temporal relationship in the signal. 1-DCNN uses convolution kernels for sliding operations, and reduces the number of parameters that need to be learned by parameter sharing, which means that the model can learn the patterns and structures in the signal with a smaller parameter scale, improving training efficiency and generalization performance. After the autoencoder is trained, the encoder part is extracted as the feature extraction module of the recognition model, with an input of 200×1 zero-sequence current waveform and an output of 18×6 feature vectors. The input layer of the fully connected neural network has 108 neurons, connecting the encoding result of the autoencoder with the hidden layer. After many experiments, it was found that when the hidden layer is composed of 138×168×38, the fully connected neural network has a good recognition effect on the measured high-resistance data. There are 2 neurons in the output layer of the neural network, corresponding to high resistance and non-high resistance respectively.

[0113] Use Shapley value to analyze the influence of vector on the prediction result of the model. If the input vector of the fully connected neural network is 6-dimensional, please refer to Figure 6 In the figure, the horizontal axis represents the global Shapley value of the feature, that is, the average impact of the input vector on the model prediction results. Different colors represent the average impact of the feature on different prediction results of the model, and the vertical axis is the feature name. In short, the larger the global Shapley value of the feature, the greater the contribution of the corresponding input vector to the model prediction results and the higher the importance. Therefore, the input vectors D4, D5 and D6 contribute more to the model prediction results and are more important. Use instance normalization to process each input vector of the fully connected neural network one by one, reduce the mean of the corresponding input vector to 0 and standardize the variance to 1, and remove the features expressed by the vector in the encoding.

[0114] The encoded vector after instance normalization is input into the decoder of the original autoencoder to obtain a new output waveform. Then, the spectrum of the original waveform and the waveform after instance normalization is analyzed to compare the difference between the two in the frequency domain. For the overall process, please refer to Figure 7 Then, according to the global Shapley value attribution analysis, it is found that the D4 vector contributes the most to the prediction results of the fully connected neural network. The third harmonic component has the largest difference in the waveform spectrum after erasing D4. Combined with electrical expertise, it is explained that one of the biggest decision bases for detecting HIF by the recognition model based on the feature extractor of the autoencoder and the fully connected neural network classifier is to detect the difference in the third harmonic component.

[0115] The decision-making basis analysis method of the high-resistance ground fault identification model provided in the embodiment of the present application uses unlabeled zero-sequence current data to perform unsupervised pre-training on the encoder model, and based on this, constructs a model for high-resistance ground fault identification. The model does not require a high amount of input labeled data, but can still meet the fault identification requirements; in addition, the model is interpreted based on the Shapley value and instance normalization algorithm to determine the basis for the model to perform fault identification analysis; it can provide a reference theoretical support with pertinence and reliability for model training, thereby improving the accuracy and credibility of model identification. Therefore, the embodiment of the present application can solve the technical problems that the existing ground fault identification model has a large demand for labeled data and the model lacks interpretability, resulting in a lack of pertinence and reliability in model training.

[0116] For easier understanding, see Figure 2 , the present application provides an embodiment of a decision basis analysis device for a high-resistance grounding fault identification model, including:

[0117] A model generation unit 201 is used to train an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder;

[0118] The feature analysis unit 202 is used to analyze the importance of the coding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculate the corresponding global Shapley value;

[0119] The decoding processing unit 203 is used to normalize the coding vector composed of multiple coding features using an example normalization algorithm, and then input the normalized vector into a decoder for decoding analysis to obtain a decoded waveform.

[0120] The basis analysis unit 204 is used to perform spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyze the model decision basis based on the global Shapley value to obtain the model identification basis.

[0121] Furthermore, the model generation unit 201 is specifically used for:

[0122] Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data;

[0123] The initial encoder and the initial decoder are pre-trained using unlabeled zero-sequence current data to obtain an encoder and a decoder;

[0124] A nonlinear mapping between the encoder, decoder and the high-resistance grounding fault label of the preset fully connected network layer is established to generate a preset high-resistance grounding fault recognition model.

[0125] Furthermore, it also includes: a training function construction unit 205, which is specifically used to:

[0126] Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave;

[0127] Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave;

[0128] The objective function of the encoder is constructed based on the fitting error and waveform similarity.

[0129] Furthermore, the feature analysis unit 202 is specifically configured to:

[0130] Constructing additive explanatory models based on Shapley values ​​of game theory;

[0131] According to the additive interpretation model, the importance of each encoding feature of each sample output by the encoder to the model recognition result is analyzed to obtain the local Shapley value;

[0132] The absolute average of the local Shapley values ​​corresponding to the coding features at the same position of all samples is calculated to obtain the global Shapley value.

[0133] The present application also provides a decision basis analysis device for a high-resistance ground fault identification model, characterized in that the device includes a processor and a memory;

[0134] The memory is used to store the program code and transmit the program code to the processor;

[0135] The processor is used to execute the decision basis analysis method of the high-resistance grounding fault identification model in the above method embodiment according to the instructions in the program code.

[0136] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0137] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0138] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0139] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, English abbreviation: ROM), random access memory (full name in English: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program codes.

[0140] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. The decision basis analysis method of the high-resistance grounding fault identification model is characterized by: include: Training an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder; Analyzing the importance of the encoding features output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculating the corresponding global Shapley value; After normalizing the coding vector composed of the plurality of coding features using an instance normalization algorithm, the coding vector is input into the decoder for decoding analysis to obtain a decoding waveform; The spectrum analysis is performed by comparing the decoded waveform with the original fault waveform, and the model decision basis is analyzed based on the global Shapley value to obtain the model identification basis.

2. The decision basis analysis method of the high-resistance grounding fault identification model according to claim 1 is characterized in that: The initial fault identification model is trained based on the unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault identification model, including: Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data; Pre-training an initial encoder and an initial decoder using the unlabeled zero-sequence current data to obtain an encoder and a decoder; A nonlinear mapping is established between the encoder, the decoder and a high-resistance grounding fault label of a preset fully connected network layer to generate a preset high-resistance grounding fault identification model.

3. The decision basis analysis method of the high-resistance grounding fault identification model according to claim 2 is characterized in that: The method of using the unlabeled zero-sequence current data to pre-train the initial encoder and the initial decoder to obtain the encoder and the decoder further includes: Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave; Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave; An objective function of the encoder is constructed according to the fitting error and the waveform similarity.

4. The decision basis analysis method of the high-resistance grounding fault identification model according to claim 1 is characterized in that: The analyzing the importance of the encoding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculating the corresponding global Shapley value, includes: Constructing additive explanatory models based on Shapley values ​​of game theory; Analyzing the importance of each encoding feature of each sample output by the encoder to the model recognition result according to the additive interpretation model to obtain a local Shapley value; The absolute average of the local Shapley values ​​corresponding to the coding features at the same position of all samples is calculated to obtain a global Shapley value.

5. The decision basis analysis method of the high-resistance grounding fault identification model according to claim 4 is characterized in that: The analyzing the importance of the coding features output by the encoder to the model recognition result and calculating the corresponding global Shapley value, after the plurality of coding features form a coding vector, further includes: Arrange the global Shapley values ​​of all samples in descending order to obtain a Shapley value sequence.

6. A decision basis analysis device for a high resistance ground fault identification model, characterized in that: include: A model generation unit, used for training an initial fault recognition model based on unlabeled zero-sequence current data to obtain a preset high-resistance grounding fault recognition model, wherein the preset high-resistance grounding fault recognition model includes an encoder and a decoder; A feature analysis unit, used to analyze the importance of the coding feature output by the encoder to the recognition result of the preset high-resistance grounding fault recognition model, and calculate the corresponding global Shapley value; A decoding processing unit, configured to normalize the coding vectors composed of the plurality of coding features using an example normalization algorithm, and then input the normalized vectors into the decoder for decoding analysis to obtain a decoding waveform; The basis analysis unit is used to perform spectrum analysis by comparing the decoded waveform with the original fault waveform, and analyze the model decision basis based on the global Shapley value to obtain the model identification basis.

7. The decision basis analysis device for the high-resistance grounding fault identification model according to claim 6, characterized in that: The model generation unit is specifically used for: Obtain zero-sequence current data in the distribution network to obtain label-free zero-sequence current data; Pre-training an initial encoder and an initial decoder using the unlabeled zero-sequence current data to obtain an encoder and a decoder; A nonlinear mapping is established between the encoder, the decoder and a high-resistance grounding fault label of a preset fully connected network layer to generate a preset high-resistance grounding fault identification model.

8. The decision basis analysis device for the high-resistance grounding fault identification model according to claim 7, characterized in that: Also includes: Training function building blocks, specifically for: Calculate the fitting error between the input zero-sequence current wave and the decoded reconstructed wave; Calculate the waveform similarity between the input zero-sequence current wave and the decoded reconstructed wave; An objective function of the encoder is constructed according to the fitting error and the waveform similarity.

9. The decision basis analysis device for the high-resistance grounding fault identification model according to claim 6, characterized in that: The feature analysis unit is specifically used for: Constructing additive explanatory models based on Shapley values ​​of game theory; Analyzing the importance of each encoding feature of each sample output by the encoder to the model recognition result according to the additive interpretation model to obtain a local Shapley value; The absolute average of the local Shapley values ​​corresponding to the coding features at the same position of all samples is calculated to obtain a global Shapley value.

10. Decision basis analysis equipment for high resistance ground fault identification model, characterized in that: The device comprises a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the decision basis analysis method of the high-resistance grounding fault identification model described in any one of claims 1-5 according to the instructions in the program code.

Images