A method for identifying distribution network fault types based on big data

Abstract

The present invention relates to a distribution network fault type identification method based on big data, comprising: obtaining a preprocessed fault training sample set of the distribution network; performing feature extraction on fault waveform data of the fault training samples to obtain time domain, frequency domain and entropy features of the fault training samples; performing whitening processing on the time domain, frequency domain and entropy features of the fault training samples to obtain whitened time domain features, whitened frequency domain and whitened entropy features of the fault training samples; filtering the whitened time domain features, whitened frequency domain features and whitened entropy features of the fault training samples according to feature variances and fault category labels of all fault training samples to obtain time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training samples; and predicting the fault type of the fault training samples using a fault identification model according to the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training samples. The present invention effectively improves the accuracy of distribution network fault prediction.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and big data technology, and in particular relates to a distribution network fault type identification method based on big data. Background Art

[0002] In modern power systems, the identification of distribution network fault types is one of the important issues that the power industry is concerned about. With the continuous expansion of the scale of power systems and the rapid development of smart grids, accurate and efficient identification of distribution network fault types is crucial to ensure the safe operation of the power grid and timely repair of faults.

[0003] Traditional distribution network fault diagnosis methods usually use sequence modeling based on current and voltage trend data. If the model is based only on data in a single domain of current and voltage, it cannot capture the full picture of the fault. At the same time, the potential forgetting nature of the sequence model indirectly leads to the model's inability to deeply understand the original complex data, thereby limiting the accuracy of fault diagnosis. Existing fault diagnosis methods based on deep learning have serious numerical stability and parameter overfitting problems. Summary of the invention

[0004] In order to solve the problems existing in the background technology, the present invention provides a method for identifying distribution network fault types based on big data, comprising:

[0005] S1: Obtain a preprocessed fault training sample set of the distribution network, wherein the fault training sample includes: fault waveform data of the fault point when the fault occurs and the corresponding fault category label;

[0006] S2: Extract features from the fault waveform data of the fault training sample to obtain the time domain, frequency domain and entropy features of the fault training sample;

[0007] S3: whitening the time domain, frequency domain and entropy features of the fault training sample respectively to obtain the whitened time domain features, whitened frequency domain and whitened entropy features of the fault training sample;

[0008] S4: filtering the whitened time domain features, whitened frequency domain features, and whitened entropy features of the fault training samples according to the feature variances and fault category labels of all fault training samples to obtain the time domain attribute features, frequency domain attribute features, and entropy attribute features of the fault training samples;

[0009] S5: predicting the fault type of the fault training sample using the fault identification model according to the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample;

[0010] S6: Construct a loss function according to the fault category label of the fault training sample and the prediction result of the fault training sample to train the fault recognition model, and perform fault recognition on the fault waveform data to be recognized through the trained fault recognition model.

[0011] The present invention has at least the following beneficial effects

[0012] The present invention can more comprehensively capture the intrinsic characteristics of the fault signal while compressing the data by performing multi-field features of time domain, frequency domain and entropy on the fault training samples, thereby improving the accuracy of fault classification and diagnosis; the problem of large data differences is avoided by performing whitening processing on the time domain, frequency domain and entropy features of the fault training samples respectively, which helps to accelerate model convergence and improve the generalization ability of the model; the whitened time domain features, whitened frequency domain features and whitened entropy features of the fault training samples are filtered to improve data quality and reduce training costs, avoid the model from falling into overfitting, and improve the fault diagnosis efficiency and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0014] Figure 2 It is a schematic diagram of the framework of the method flow of the present invention. DETAILED DESCRIPTION

[0015] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0016] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0017] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0018] See also Figure 1 and Figure 2 The present invention provides a method for identifying distribution network fault types based on big data, comprising:

[0019] S1: Obtain a preprocessed fault training sample set of the distribution network, wherein the fault training sample includes: fault waveform data of the fault point when the fault occurs and the corresponding fault category label;

[0020] Preferably, the fault waveform data includes: one or more of voltage time series data, current time series data and filter resistance time series data of the fault point; wherein the filter resistance is the resistance between the fault point and the ground.

[0021] In this embodiment, a pre-processed fault training sample set of the distribution network is used, and the training process of the fault training sample set is as follows:

[0022] S11: retain the fault waveform data from N cycles before the fault occurs to M cycles after the fault occurs, and remove the redundant fault waveform data;

[0023] In this embodiment, since the waveform data of the distribution network is generally periodic under normal conditions, and its waveform state will be affected and changed after a fault occurs, the fault waveform data from N cycles before the fault to M cycles after the fault occurs can be retained in this embodiment to accurately extract the characteristics of the fault occurrence. In this embodiment, considering the stability factor of model training, N and M are preferably set to 5. Similarly, those skilled in the art can adjust the values ​​of N and M according to actual conditions.

[0024] S12: inputting the fault waveform data processed in step S11 into a bandpass filter for filtering to obtain pre-processed fault waveform data;

[0025] In this embodiment, high-frequency noise and low-frequency trends in each time series data are filtered out by a bandpass filter. By filtering out these unnecessary signals, the clarity and accuracy of the fault characteristic signal can be guaranteed, thereby providing a more reliable data basis for subsequent fault feature extraction and diagnosis.

[0026] S2: Extract features from the fault waveform data of the fault training sample to obtain the time domain, frequency domain and entropy features of the fault training sample;

[0027] In this embodiment, the time domain features of the fault training sample include: peak value, mean value, rectified mean value, variance, standard deviation, kurtosis, skewness, root mean square, peak-to-peak value, waveform factor, peak factor, pulse factor and margin factor;

[0028] The calculation formula of the form factor CF is:

[0029]

[0030] Where PMS represents the root mean square of z;

[0031] The calculation formula of peak factor PF is:

[0032]

[0033] The calculation formula of margin factor MF is:

[0034]

[0035] Wherein, RMS represents the root mean square of the fault waveform data, and z represents the fault waveform data.

[0036] The frequency domain features of the fault training samples include: centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation and total harmonic distortion;

[0037] Center of gravity frequency f c The calculation formula is:

[0038]

[0039] The calculation formula of total harmonic distortion THD is:

[0040]

[0041] Among them, f i Indicates frequency, P i Indicates the power corresponding to the frequency, P 1 represents the fundamental frequency, and N represents the highest harmonic number considered.

[0042] The entropy characteristics of fault training samples include power spectrum entropy, singular spectrum entropy and energy entropy.

[0043] Power spectrum entropy H PS The calculation formula is:

[0044]

[0045] Where E represents the energy of the fault waveform data;

[0046] In this embodiment, when the fault waveform data is a plurality of time series data, the time domain, frequency domain and quotient features are extracted for each time series data respectively; for example, when the fault waveform data is voltage time series data and current time series data, the time domain features, frequency domain features and quotient features of the voltage time series data and the current time series data are extracted respectively, and the time domain, frequency domain and entropy features of the fault training sample include: the time domain features, frequency domain features and quotient features of the voltage time series data and the current time series data.

[0047] By extracting the above-mentioned time domain, frequency domain and quotient features from the fault training samples, the intrinsic characteristics of the fault signal can be more comprehensively captured while compressing the data, thereby improving the accuracy of fault classification and diagnosis.

[0048] S3: whitening the time domain, frequency domain and entropy features of the fault training sample respectively to obtain the whitened time domain features, whitened frequency domain and whitened entropy features of the fault training sample;

[0049] Preferably, the whitening processing of the time domain, frequency domain and entropy features of the fault training sample respectively comprises:

[0050]

[0051] Among them, X white represents the whitened time domain feature, whitened frequency domain feature or whitened entropy feature of a single fault training sample after whitening processing, x represents the time domain feature, frequency domain feature or entropy feature of a single fault training sample; μ x represents the mean value of all training samples in the fault training sample set in feature c; σ x Represents the standard deviation of all training samples in the fault training sample set on feature x.

[0052] In this embodiment, the time domain features of the fault training samples include multiple types. For example, for the peak value of the fault training samples, the mean and standard deviation of the peak features of all fault training samples are calculated, and then the peak features of the training samples are whitened. All types of time domain features are processed in turn to obtain the processed whitened time domain features; similarly, the frequency domain features and quotient features are processed in the same way to obtain whitened frequency domain features and whitened entropy features; by whitening the fault training samples, the problem of large data differences is avoided, which helps to accelerate model convergence and improve the model generalization ability.

[0053] It is particularly noted that since the training sample set and the test set have the same sample distribution, when whitening the test sample in practical applications, the feature x′ of the test sample is whitened using the mean of the feature x and the standard deviation of the feature x of all training samples in the fault training sample set. The feature x′ represents the time domain feature, frequency domain feature or entropy feature of the test sample.

[0054] S4: filtering the whitened time domain features, whitened frequency domain features, and whitened entropy features of the fault training samples according to the feature variances and fault category labels of all fault training samples to obtain the time domain attribute features, frequency domain attribute features, and entropy attribute features of the fault training samples;

[0055] Preferably, the step S4 comprises:

[0056] S41: Concatenate the fault category labels of all fault training samples into a matrix Y; concatenate the whitened features of all fault training samples into a matrix X w ; Calculate matrix Y and matrix X w The mutual information of the whitening feature includes: one of a whitening time domain feature, a whitening frequency domain feature and a whitening entropy feature;

[0057] S42: Calculate the variance of the whitened features of all fault training samples; if the variance of a whitened feature is greater than the set threshold θ 1 , and the matrix X corresponding to the whitened feature w The mutual information between the two is greater than the threshold θ. 2 , then use the whitened feature as the attribute feature of the fault training sample; traverse all the whitened features of the fault training sample and repeatedly execute steps S41 to S42 to select the time domain attribute feature, frequency domain attribute feature and entropy attribute feature of the fault training sample.

[0058] In this embodiment, the whitened time domain features, whitened frequency domain features and whitened entropy features of the fault training samples are filtered to remove features irrelevant to the model classification results, thereby reducing the amount of model calculation. By calculating the variance of the whitened features of all fault training samples, the smaller the variance of the whitened features between fault samples of different categories, the smaller the impact on the model classification results. By calculating the mutual information between the whitened features of all training samples and the fault category labels, the relationship between the whitened features and the classification results can be reflected. By setting two thresholds θ 1 and θ 2 The whitening features irrelevant to the model classification are filtered out, thereby improving data quality and reducing training costs. The remaining features are closely related to the model classification, that is, they serve as attribute features of fault training samples.

[0059] For example, the whitened time domain features of the fault training samples include multiple types, such as the whitened peak features of the fault training samples, then the whitened peak features of all fault training samples are concatenated into a matrix X w , calculate the matrix Y and matrix X w The mutual information of the fault training samples is calculated by calculating the variance of the whitened peak features of all fault training samples, and then the variance is calculated by two thresholds θ 1 and θ 2 Determine whether the whitened peak feature is the attribute feature of the fault training sample; similarly, perform the same processing on each type of whitened time domain feature of the fault training sample to obtain the time domain attribute feature of the fault training sample. Similarly, the frequency domain attribute feature and entropy attribute feature of the fault training sample can be obtained. It is particularly noted that after determining which whitened features belong to the attribute features of the fault training sample through the fault training sample set, in practical applications, since the training sample set and the test set have the same sample distribution, after extracting the whitened features of the test sample, it is only necessary to directly extract the corresponding attribute features.

[0060] S5: predicting the fault type of the fault training sample using the fault identification model according to the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample;

[0061] Preferably, the method of predicting the fault type of the fault training sample by using the fault identification model includes:

[0062] S51: Mapping the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample to the same dimension through a linear layer to obtain intermediate time domain attribute features, intermediate frequency domain attribute features and intermediate entropy attribute features;

[0063] S52: using a bilinear operator to combine the intermediate time domain attribute features, the intermediate frequency domain attribute features and the intermediate entropy attribute features in pairs to obtain six feature interaction vectors after interaction;

[0064] Preferably, the step of combining the intermediate time domain attribute features, the intermediate frequency domain attribute features and the intermediate entropy attribute features in pairs by using a bilinear operator comprises:

[0065]

[0066] Among them, t represents the time domain, p represents the frequency domain, e represents entropy, and X i,j represents the feature interaction vector; X i and X j Represents time domain attribute features, frequency domain attribute features or entropy attribute features; d i and d j Respectively represent X i and X j The dimension of ; Linear represents the linear layer; and represents the intermediate time domain attribute feature, the intermediate frequency domain attribute feature or the intermediate entropy attribute feature; k represents the dimension of the intermediate time domain attribute feature, the intermediate frequency domain attribute feature or the intermediate entropy attribute feature; W represents the weight parameter; ⊙ represents the Hadamard product operation.

[0067] In this embodiment, when i=t, j=p, X t,p Represents the feature interaction vector of the time domain attribute feature to the frequency domain attribute feature; X t represents the feature matrix obtained by concatenating all time domain attribute features of the fault training samples; X p represents the feature matrix obtained by concatenating all frequency domain attribute features of the fault training samples; Indicates that the dimension is d t Features of X t Mapped into features with dimension k; the same principle is used for the remaining feature interactions, which will not be elaborated on; by using the bilinear operator to combine the intermediate time domain attribute features, the intermediate frequency domain attribute features and the intermediate entropy attribute features in pairs to obtain combined features, the model can more easily learn the intrinsic connections between different domains, which is beneficial to improving the model fitting effect.

[0068] S53: Concatenate the six feature interaction vectors to obtain the feature interaction matrix X 0 , the feature interaction matrix X 0 Input the compressed sensing layer to perform deep feature extraction to obtain multi-domain interactive deep features;

[0069] Preferably, the feature interaction matrix X 0 Input the compressed sensing layer to perform deep feature extraction to obtain multi-domain interactive deep features including:

[0070] S531: Initialize l=1;

[0071] S532: Feature interaction matrix Each row of the compressed sensing layer is respectively related to the matrix of the lth layer Perform Hadamard product operation to obtain m first intermediate two-dimensional matrices Reshape the m first intermediate two-dimensional matrices into a three-dimensional matrix When l = 1, X 1 is a trainable parameter matrix, m=6 represents the number of feature interaction vectors;

[0072] S533: The matrix Q l1 Split along the last dimension into k second intermediate 2D matrices Build H l+1 parameter matrix Each parameter matrix W l2Perform vector dot multiplication operations with the k second intermediate two-dimensional matrices q to obtain k scalars, and concatenate the k scalars to obtain a k-dimensional vector; H l+1 The parameter matrix W l2 The corresponding H l+1 k-dimensional vectors are concatenated into a matrix

[0073] S534: Let l = l + 1, and repeat steps S532-S533 until l + 1 equals n, and obtain n matrices of all layers of the compressed sensing layer The n matrices of the compressed sensing layer are respectively input into the average pooling layer for dimensionality reduction, and the results after dimensionality reduction are concatenated to obtain multi-domain interactive deep-level features of length n*k, where n represents the number of compressed sensing layers.

[0074] In this embodiment, the compressed sensing layer can extract more important features from high-dimensional data, reduce the dimension of the data, and improve the model's ability to fit the training data, thereby ensuring the model's prediction performance.

[0075] S54: Input the multi-domain interactive deep features into the fully connected layer to predict the fault type of the fault training sample.

[0076] S6: Construct a loss function according to the fault category label of the fault training sample and the prediction result of the fault training sample to train the fault recognition model, and perform fault recognition on the fault waveform data to be recognized through the trained fault recognition model.

[0077] Preferably, the loss function includes:

[0078]

[0079] Where loss represents the loss function, α is the weight parameter, CE(·) is the cross entropy loss function, and y is the true category label of the fault training sample. Represents the prediction result of the fault identification model for the fault training sample.

[0080] In this embodiment, the cross entropy loss function with a reward term can more stably learn the intrinsic differences between positive and negative samples, thereby achieving a result in which the prediction value of fault data is larger and the prediction value of non-fault data is smaller, further improving the accuracy of fault identification.

[0081] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0082] In summary, the present invention can more comprehensively capture the intrinsic characteristics of the fault signal while compressing the data by performing Multi-field features of time domain, frequency domain and entropy on the fault training samples, thereby improving the accuracy of fault classification and diagnosis; by whitening the time domain, frequency domain and entropy features of the fault training samples respectively, the problem of large data differences is avoided, which helps to accelerate model convergence and improve the generalization ability of the model; by filtering the whitened time domain features, whitened frequency domain features and whitened entropy features of the fault training samples, the data quality is improved and the training cost is reduced, the model is prevented from falling into overfitting, and the fault diagnosis efficiency and accuracy of the model are improved.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0084] 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 the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying distribution network fault types based on big data, characterized in that: include: S1: Obtain a preprocessed fault training sample set of the distribution network, wherein the fault training sample includes: fault waveform data of the fault point when the fault occurs and the corresponding fault category label; S2: Extract features from the fault waveform data of the fault training sample to obtain the time domain, frequency domain and entropy features of the fault training sample; S3: whitening the time domain, frequency domain and entropy features of the fault training sample respectively to obtain the whitened time domain features, whitened frequency domain and whitened entropy features of the fault training sample; S4: filtering the whitened time domain features, whitened frequency domain features, and whitened entropy features of the fault training samples according to the feature variances and fault category labels of all fault training samples to obtain the time domain attribute features, frequency domain attribute features, and entropy attribute features of the fault training samples; S5: predicting the fault type of the fault training sample using the fault identification model according to the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample; The fault types of the fault training samples predicted by the fault identification model include: S51: Mapping the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample to the same dimension through a linear layer to obtain intermediate time domain attribute features, intermediate frequency domain attribute features and intermediate entropy attribute features; S52: using a bilinear operator to combine the intermediate time domain attribute features, the intermediate frequency domain attribute features and the intermediate entropy attribute features in pairs to obtain six feature interaction vectors after interaction; The bilinear operator is used to combine the intermediate time domain attribute features, the intermediate frequency domain attribute features and the intermediate entropy attribute features in pairs, including: Among them, t represents the time domain, p represents the frequency domain, e represents entropy, and X i,j represents the feature interaction vector; X i and X j Represents time domain attribute features, frequency domain attribute features or entropy attribute features; d i and d j Respectively represent X i and X j The dimension of ; Linear represents the linear layer; and represents the intermediate time domain attribute feature, the intermediate frequency domain attribute feature or the intermediate entropy attribute feature; k represents the dimension of the intermediate time domain attribute feature, the intermediate frequency domain attribute feature or the intermediate entropy attribute feature; W represents the weight parameter; ⊙ represents the Hadamard product operation; S53: concatenate the six feature interaction vectors to obtain a feature interaction matrix X0, and input the feature interaction matrix X0 into a compressed sensing layer to perform deep feature extraction to obtain multi-domain interaction deep-level features; S54: Input the multi-domain interactive deep features into the fully connected layer to predict the fault type of the fault training sample; S6: Construct a loss function according to the fault category label of the fault training sample and the prediction result of the fault training sample to train the fault recognition model, and perform fault recognition on the fault waveform data to be recognized through the trained fault recognition model.

2. A method for identifying distribution network fault types based on big data according to claim 1, characterized in that: The preprocessing process of the fault training sample set includes: S11: retain the fault waveform data from N cycles before the fault occurs to M cycles after the fault occurs, and remove the redundant fault waveform data; S12: Input the fault waveform data processed in step S11 into a bandpass filter for filtering to obtain pre-processed fault waveform data.

3. A method for identifying distribution network fault types based on big data according to claim 2, characterized in that: The fault waveform data includes: one or more of voltage time series data, current time series data and filter resistance time series data of the fault point.

4. The method for identifying distribution network fault types based on big data according to claim 1 is characterized in that: The time domain characteristics of the fault training sample include: peak value, mean value, rectified mean value, variance, standard deviation, kurtosis, skewness, root mean square, peak-to-peak value, waveform factor, peak factor, pulse factor and margin factor; The frequency domain features of the fault training samples include: centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation and total harmonic distortion; The entropy characteristics of fault training samples include power spectrum entropy, singular spectrum entropy and energy entropy.

5. The method for identifying distribution network fault types based on big data according to claim 1, characterized in that: The whitening processing of the time domain, frequency domain and entropy features of the fault training samples respectively includes: Among them, X whiten represents the whitened time domain feature, whitened frequency domain feature or whitened entropy feature of a single fault training sample after whitening processing, x represents the time domain, frequency domain or entropy feature of a single fault training sample; μ x represents the mean value of all training samples in the fault training sample set in feature x; σ x Represents the standard deviation of all training samples in the fault training sample set on feature x.

6. A method for identifying distribution network fault types based on big data according to claim 1, characterized in that: The step S4 comprises: S41: Concatenate the fault category labels of all fault training samples into a matrix Y; concatenate the whitened features of all fault training samples into a matrix X w ; Calculate matrix Y and matrix X w The mutual information of the whitening feature includes: one of a whitening time domain feature, a whitening frequency domain feature and a whitening entropy feature; S42: Calculate the variance of the whitened features of all fault training samples; if the variance of a whitened feature is greater than the set threshold θ1, and the matrix X corresponding to the whitened feature w If the mutual information of Y and Y is greater than the set threshold θ2, the whitened feature is used as the attribute feature of the fault training sample; all whitened features of the fault training sample are traversed and steps S41 to S42 are repeatedly executed to filter out the time domain attribute features, frequency domain attribute features and entropy attribute features of the fault training sample.

7. A method for identifying distribution network fault types based on big data according to claim 1, characterized in that: The step of inputting the feature interaction matrix X0 into the compressed sensing layer to extract deep features to obtain multi-domain interaction deep features includes: S531: Initialize l=1; S532: Feature interaction matrix Each row of the compressed sensing layer is respectively related to the matrix of the lth layer Perform Hadamard product operation to obtain m first intermediate two-dimensional matrices Reshape the m first intermediate two-dimensional matrices into a three-dimensional matrix Among them, when l = 1, X1 is the trainable parameter matrix, and m = 6 represents the number of feature interaction vectors; S533: The matrix Q l Split along the last dimension into k second intermediate 2D matrices Build H l+1 parameter matrix Each parameter matrix W l2 Perform vector dot multiplication operations with the k second intermediate two-dimensional matrices q to obtain k scalars, and concatenate the k scalars to obtain a k-dimensional vector; H l+1 The parameter matrix W l2 The corresponding H l+1 k-dimensional vectors are concatenated into a matrix S534: Let l = l + 1, and repeat steps S532-S533 until l + 1 equals n, and obtain n matrices of all layers of the compressed sensing layer The n matrices of the compressed sensing layer are respectively input into the average pooling layer for dimensionality reduction, and the reduced dimensionality results are concatenated to obtain multi-domain interactive deep features with a length of n*k.

8. The method for identifying distribution network fault types based on big data according to claim 1, characterized in that: The loss function includes: Where loss represents the loss function, α is the weight parameter, CE(·) is the cross entropy loss function, and y is the true category label of the fault training sample. Represents the prediction result of the fault identification model for the fault training sample.

Images