Cable fault feature extraction method and device based on improved GAN algorithm, and medium

Abstract

The invention discloses a cable fault feature extraction method and device based on an improved GAN algorithm, and a medium, and the method comprises the steps: firstly collecting the three-phase current of a cable, and carrying out the optimization of an input potential variable of a training generator based on permutation entropy and approximate entropy; importing the potential variables into a training generator to obtain synthetic data, and expanding the small sample fault data set; importing the real data and the synthetic data into a training discriminator for discrimination, and correcting network parameters of the discriminator and the generator according to a discrimination result; when the discriminator cannot distinguish the synthetic data from the real data, outputting sample data sets with the same quantity in different states; feature values of the three-phase current of the cable in the time domain, the frequency domain and the time-frequency domain under the healthy and fault working conditions are taken, and a feature matrix is established; according to the technical scheme, the problem that the number of samples is unbalanced in existing cable fault diagnosis can be solved, the stability and accuracy of synthetic data are greatly improved, the sample expansion process is accelerated, and convenience is provided for extraction of fault features.

Description

Technical Field

[0001] The present invention belongs to the field of cable fault diagnosis, and relates to a fault feature extraction technology in cable fault diagnosis, and in particular to a cable fault feature extraction method, device and medium based on an improved generative adversarial network (GAN) algorithm. Background Art

[0002] As the scale of the power grid continues to expand, the use of cables as an important carrier of power transmission is gradually increasing. Cables are usually buried underground, and the operating environment is relatively harsh. In the later stage of operation, they are prone to early failures due to local insulation degradation. After repeated occurrence, early failures will evolve into permanent failures. Early failures occur at the beginning of the failure, last for a short time, and are difficult to detect. However, the voltage and current data recorded in real time in the power grid monitoring system can reflect the operating status of the cable. According to the corresponding relationship between the voltage and current changes and the cable status, an accurate mathematical model is established to realize the identification of early cable failures, which is conducive to the staff to repair or replace the cable in time before the permanent failure occurs in the cable, so as to ensure the stable and reliable operation of the power system.

[0003] In recent years, with the rise of deep learning technology, data-driven intelligent fault diagnosis technology has been widely studied and applied in various fields. However, most existing studies have failed to achieve the ideal performance level when actually applied to cable fault diagnosis. The most important reason is that the premise for ensuring the effectiveness of the intelligent diagnosis model is that the training set needs to have sufficient fault samples. However, in the actual production process, due to safety considerations, the cable cannot continue to operate in a faulty state, resulting in extremely scarce fault samples in the collected monitoring data and limited categories covered. Therefore, the actual available training data sets usually show skewed class distribution characteristics.

[0004] Data augmentation methods provide a simple but effective way to solve the imbalance problem: based on existing real data sets, synthesize new samples or delete original samples to make the originally imbalanced data sets balanced, thereby eliminating the negative impact of imbalanced data sets on diagnostic model training. Among them, data augmentation methods based on deep generative models, such as GAN and its variants, have shown extremely superior performance in learning complex high-dimensional data distributions, providing a new perspective for solving the imbalance problem in the field of fault diagnosis.

[0005] Related Literature:

[0006] 《Unbalanced_Research on Intelligent Diagnosis of Natural Gas Pipeline in Small Sample Scenarios_Wang Chuang》

[0007] Research on early fault identification of cables based on principal component analysis and incremental convolutional neural network_Tao Yunchun

[0008] 《Research on Transmission Line Fault Diagnosis Method Based on CNN-LSTM_Feng Yin》 Summary of the invention

[0009] The purpose of the present invention is to provide a cable fault feature extraction method, device and medium based on an improved GAN algorithm, which can solve the problem of sample number imbalance in existing cable fault diagnosis.

[0010] In order to achieve the above object, the solution of the present invention is:

[0011] A cable fault feature extraction method based on an improved GAN algorithm comprises the following steps:

[0012] Step A: Collect the three-phase current i of the cable abc And the corresponding time t, state information y, for the three-phase current i abc Processing is performed to obtain a sample current x;

[0013] Step B, calculating the permutation entropy and approximate entropy of the sample current x, and constructing a latent variable Q(t) based on the permutation entropy and approximate entropy as an input of the training generator;

[0014] Step C, import the latent variable Q(t) into the training generator G to obtain the synthetic data G{Q(t)};

[0015] Step D: import the real data x and the synthetic data G{Q(t)} into the training discriminator J for discrimination, and adjust the network parameters θ of the training discriminator J and the training generator G according to the discrimination results. J ,θ G make corrections;

[0016] Step E, repeat steps CD until the discriminator cannot distinguish between synthetic data and real data, and terminate the iteration, outputting the same number of sample data sets in different states;

[0017] Step F, extracting the eigenvalues ​​of the three-phase current of the cable in the time domain, frequency domain and time-frequency domain under healthy and fault conditions, and establishing a characteristic matrix A.

[0018] In the above step A, the state information y includes the following states:

[0019] State 1: Phase A ground fault

[0020] State 2: Phase B ground fault

[0021] State 3: Phase C ground fault

[0022] State 4: AB two-phase grounding fault

[0023] State 5: BC two-phase grounding fault

[0024] State 6: AC two-phase ground fault

[0025] State 7: Three-phase ground fault

[0026] State 8: Normal state

[0027] State 9: Arc fault.

[0028] In the above step A, the three-phase current i abc The formula for processing the sample current x is,

[0029]

[0030] In the formula, x min is the three-phase current i abc The minimum amplitude of x max is the three-phase current i abc The maximum amplitude.

[0031] In the above step B, the method for calculating the permutation entropy of the sample current x is:

[0032] Setting up the time series dataset in Represents the i-th true sample at time t, i = 1,…,n t , n t is the number of real samples, The sample current contains N sampling points; the permutation entropy of each sample is calculated with a sliding window of length L, L ≤ N, and reconstructed within the sliding window The phase space of is described as:

[0033]

[0034] Where m is the embedding dimension, τ is the time delay, and M = L-(m-1)τ is the number of vectors in the reconstructed image space. The j-th row of the reconstruction matrix is ​​arranged in ascending order to form a new sequence, which is expressed as:

[0035]

[0036] Where a1 represents the index of the minimum value in the jth row of the reconstruction matrix, a m represents the index of the maximum value in the jth row of the reconstruction matrix; the sequence thus composed has m! permutation possibilities; the permutation entropy is expressed as:

[0037]

[0038] Where P a represents the probability distribution of symbols, K≤m!.

[0039] In the above step B, the method for calculating the approximate entropy of the sample current x is:

[0040] Assume that the time series Z N ={z1,z2,…z N The length is N, and the sequence is reconstructed into m-dimensional phase space, m>2, and the sequence is obtained: Z m (i) = {z(k), z(k+1), …, z(k+m-1)}, k = 1, 2, …, N-m+1; the distance between the two dimensions is:

[0041] d=max|Z m (k)-Z m (j)|,j≠k,k≤j≤Nm

[0042] Where max is the maximum difference between the corresponding elements of the two-dimensional time series, and the similarity tolerance r=0.2*std is defined, where std is Z N Standard deviation; the number of vectors whose distance is less than the similarity tolerance r is the ratio of num to the total vector:

[0043]

[0044] Take the logarithm of the above ratio and find the average value of each k where m is a vector. The formula is:

[0045]

[0046] Change the phase space reconstruction dimension m = m + 1, repeat the above steps to obtain The approximate entropy obtained by calculation is:

[0047]

[0048] In the above step B, the latent variable Q(t) is constructed based on the permutation entropy and the approximate entropy according to the following formula:

[0049]

[0050] In the formula, iter is the current iteration number, iter max is the maximum number of iterations, and y is the current state.

[0051] In the above step C, the loss function of training generator G is,

[0052]

[0053] In the formula, G and J are the training generator and the training discriminator respectively, q represents the latent variable Q(t), P q (q) is the probability of the latent variable; J(·) is the probability that the input to the trained discriminator is real data rather than synthetic data.

[0054] In the above step D, the loss function for training the discriminator J is,

[0055]

[0056] In the formula, xˊ represents the real data, P x (x) is the distribution of real data.

[0057] In the above step D, the network parameters θ of the training discriminator J and the training generator G are adjusted according to the discrimination results. J ,θ G Correction is made, where the network parameters θ of the training discriminator J are J The corrected formula is,

[0058] θ J =1+αJ(G(q))

[0059] The network parameters θ of the training generator G G The corrected formula is,

[0060] θ G =1-αJ(G(q))

[0061] Where α is a constant; q represents the latent variable Q(t), and J(·) is the probability that the input to the discriminator is real data rather than synthetic data.

[0062] In step E above, the necessary condition for the discriminator to be unable to distinguish between synthetic data and real data is that J(G(q))≈1.

[0063] In the above step F, the time domain features in the feature matrix A include maximum value, minimum value, standard deviation, root mean square value and kurtosis, the frequency domain features include centroid frequency, average frequency, root mean square frequency and frequency standard deviation, and the time-frequency domain features include singular entropy, approximate entropy and fuzzy entropy.

[0064] A cable fault feature extraction device based on an improved GAN algorithm comprises a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the steps of the cable fault feature extraction method based on the improved GAN algorithm as described above are implemented.

[0065] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the steps of the cable fault feature extraction method based on the improved GAN algorithm as described above are implemented.

[0066] After adopting the above scheme, the present invention provides a solution to the problem of unbalanced number of samples in cable fault diagnosis: on the one hand, the present invention normalizes the cable fault phase current, so that it has universal applicability when dealing with different fault sample data sets; on the other hand, the present invention improves the training generator in the GAN network, and integrates the permutation entropy and approximate entropy as the composition of the latent variables of the training generator, which can more effectively characterize the time correlation of the data and capture more intrinsic connections between categories. In addition, in the composition of latent variables, by considering the influence of the number of iterations, the proportion of the one-hot vector gradually increases with the increase of the number of iterations, which can better capture the contextual information of the time series and gradually reduce the uncertainty of the latent category information. The present invention greatly improves the stability and accuracy of the synthetic data, accelerates the sample expansion process, and facilitates the extraction of fault features. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a cable fault feature extraction method based on an improved GAN algorithm of the present invention;

[0068] Figure 2 This is the basic framework diagram of the GAN network;

[0069] Figure 3 This is the schematic diagram of phase A fault grounding. DETAILED DESCRIPTION

[0070] The technical solutions and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0071] like Figure 1 As shown, the present invention provides a cable fault feature extraction method based on an improved GAN algorithm. Figure 2 The basic framework diagram of the GAN network is given. The method includes the following steps:

[0072] Step 1: Data collection: Collect the three-phase current i of the cable abc , record it as sample current x. Record time information t, state information y;

[0073] The status information y includes the following status:

[0074] y=1:Phase A ground fault

[0075] y=2:B phase ground fault

[0076] y=3:C phase ground fault

[0077] y=4:AB two-phase grounding fault

[0078] y=5:BC two-phase grounding fault

[0079] y=6:AC two-phase ground fault

[0080] y=7:Three-phase ground fault

[0081] y=8:Normal state

[0082] y=9:arc fault

[0083] In this embodiment, taking the feature extraction of phase A grounding fault as an example, Figure 3 A schematic diagram of a phase A fault grounding is given, where x is the phase A fault current and the state information y=1.

[0084] Step 2: Data normalization: normalize the sample current x to obtain the normalized data set D(xˊ,y);

[0085] The A phase fault current is normalized, and the formula is:

[0086]

[0087] Where xˊ is the normalized sample current; x min is the minimum amplitude of the sample current; x max is the maximum amplitude of the sample current.

[0088] Step 3: Based on the normalized permutation entropy and approximate entropy of the sample current, construct the latent variable Q(t) as the input of the training generator; optimize the latent variable: Based on the excellent ability of the permutation entropy and approximate entropy of the normalized sample current in capturing the dynamic anomalies and change laws of the time series, optimize the input latent variable Q of the training generator. t i Optimize

[0089] The permutation entropy and approximate entropy of the normalized A-phase fault current xˊ are calculated using the following method:

[0090] 1) Calculate the permutation entropy

[0091] Permutation Entropy (PE) is a dynamic anomaly detection method for time series. Its uniqueness lies in that it can accurately present the characteristic differences between normal points and fault points by amplifying small changes in the signal. Therefore, permutation entropy has excellent ability to capture abnormal situations in time series and can effectively distinguish normal data from faulty data. Its calculation method is as follows:

[0092] Consider a time series dataset in It is composed of N sampling points, which means the current data set composed of N sampling points in the i-th real sample current data at time t; n t is the number of real samples. In this embodiment, the real samples are the normalized sample currents in step 2. The permutation entropy of each sample is calculated using a sliding window of length L (L≤N), and reconstructed in the sliding window. The phase space of can be described as:

[0093]

[0094] Where m is the embedding dimension, τ is the time delay, and M = L-(m-1)τ is the number of vectors in the reconstructed image space. Arrange the jth row of the reconstruction matrix in ascending order to form a new sequence, which can be expressed as:

[0095]

[0096] Where a1 represents the index of the minimum value in the jth row of the reconstruction matrix, a m represents the index of the maximum value in the jth row of the reconstruction matrix. The sequence thus formed has m! possible permutations. The permutation entropy of the real cable phase current can be expressed as:

[0097]

[0098] 2) Calculate approximate entropy

[0099] Approximate entropy can describe the change pattern and unpredictability of time series data. The approximate entropy algorithm is as follows:

[0100] Time Series Z N ={z1,z2,…z N The length is N. Reconstructing the sequence in m (m>2) dimensional phase space can obtain the sequence: Z m (i) = {z(k), z(k+1), …, z(k+m-1)}, k = 1, 2, …, N-m+1. The distance between the two dimensions is:

[0101] d=max|Z m (k)-Z m (j)|,j≠k,k≤j≤Nm

[0102] Where max is the maximum difference between the corresponding elements of the two-dimensional time series, and the similarity tolerance r=0.2*std is defined, where std is Z N The number of vectors whose distance is less than the similarity tolerance r is the ratio of num to the total vector:

[0103]

[0104] Take the logarithm of the above ratio and find the average value of each k where m is a vector. The formula is:

[0105]

[0106] Change the phase space reconstruction dimension m = m + 1, repeat the above steps to obtain The approximate entropy of the real cable phase current can be obtained by calculation:

[0107]

[0108] The potential variable Q(t) of the fault current of phase A is obtained, and its calculation formula is:

[0109]

[0110] In the formula, iter is the current iteration number, iter max is the maximum number of iterations, and y is the current state.

[0111] Step 4: Obtain synthetic data for sample expansion: Import the latent variable Q(t) into the training generator G, obtain synthetic data G{Q(t)}, and expand the small sample fault dataset;

[0112] By importing the latent variable of phase A current into the training generator G, the training generator obtains synthetic data to expand the phase A ground fault data set. The loss function of the training generator G is:

[0113]

[0114] In the formula, G and J are the training generator and the training discriminator respectively, q represents the latent variable Q(t), P q (q) is the probability distribution of the latent variable. J(·) is the probability that the input to the trained discriminator is real data rather than synthetic data.

[0115] In the formula, Represents the loss function L for training the generator G G The expectation of [log(1-J(G(q)))] in q(q) Indicates that q obeys P q (q) distribution.

[0116] Step 5: Data discrimination and network parameter correction: The real data xˊ and the synthetic data G{Q(t)} are imported into the training discriminator J for discrimination, and the network parameters θ of the training discriminator J and the training generator G are adjusted according to the discrimination results. J ,θ G make corrections;

[0117] The real A-phase fault current and the A-phase fault current synthesized by the training generator G are introduced into the training discriminator J for discrimination. The loss function of the discriminator J is:

[0118]

[0119] In the formula, xˊ represents the real data, P x' (x') is the distribution of the real data. Represents the loss function of training the discriminator J expectations.

[0120] The network parameters θ of the discriminator J J The update formula is:

[0121] θ J =1+αJ(G(q))

[0122] The network parameter correction formula for training the generator G is:

[0123] θ G =1-αJ(G(q))

[0124] Where α is a constant.

[0125] Step 6: Repeat steps 4-5, and iterate to complete the data set expansion. Specifically, the training generator obtains synthetic data, and then the training discriminator compares the synthetic data with the real data. If the output discrimination result does not meet the termination condition, the network parameters are corrected and the process returns to step 4. The previous process is repeated until the training discriminator cannot distinguish between synthetic data and real data, and the iteration is terminated, and the sample data sets in different states are output in equal quantities;

[0126] Among them, the necessary condition for the discriminator to be unable to distinguish between synthetic data and real data is J(G(q))≈1.

[0127] Step 7: Extract the eigenvalues ​​of the three-phase current of the cable in the time domain, frequency domain, and time-frequency domain under healthy and fault conditions, and establish the characteristic matrix A.

[0128] The composition of the feature matrix A is:

[0129] Time domain characteristics:

[0130] 1) Maximum value: i max =max(i j );

[0131] 2) Minimum value: i min =min(i j );

[0132] 3) Standard Deviation:

[0133] 4) Root mean square value:

[0134] 5) Kurtosis:

[0135] Frequency domain characteristics:

[0136] The frequency components of the signal are obtained through spectrum analysis, and the frequency domain features of the signal are extracted. The obtained frequency domain features reflect the state of the signal in the frequency domain, thereby forming a fault identification feature vector. The selected frequency domain feature expression is as follows, where s(k) represents the spectrum of the original phase current signal i(t), k = 1, 2, ... K, K is the number of spectrum lines, and f(k) represents the frequency value of the kth spectrum.

[0137] 1) Center of gravity frequency:

[0138] 2) Average frequency:

[0139] 3) Root mean square frequency:

[0140] 4) Frequency standard deviation:

[0141] Time-frequency domain features:

[0142] 1) Singular entropy:

[0143] Among them, λ1, λ2, …, λ r are the singular values ​​of the matrix X, and their size relationship satisfies the relationship: λ1≥λ2≥…λ r , and r≤k. The matrix X satisfies:

[0144] X=UNV T

[0145] Among them, U∈R m×m , V∈R m×m , and the matrix N satisfies:

[0146]

[0147] Among them, S = diag (λ1, λ2, ..., λ r ).

[0148] 2) Approximate entropy

[0149] Approximate entropy can describe the change pattern and unpredictability of time series data. The approximate entropy algorithm is as follows:

[0150] Time Series Z N ={z1,z2,…zN The length is N. Reconstructing the sequence in m (m>2) dimensional phase space can obtain the sequence: Z m (i) = {z(k), z(k+1), …, z(k+m-1)}, k = 1, 2, …, N-m+1. The distance between the two dimensions is:

[0151] d=max|Z m (k)-Z m (j)|,j≠k,k≤j≤Nm

[0152] Where max is the maximum difference between the corresponding elements of the two-dimensional time series, and the similarity tolerance r=0.2*std is defined, where std is Z N The number of vectors whose distance is less than the similarity tolerance r is the ratio of num to the total vector:

[0153]

[0154] Take the logarithm of the above ratio and find the average value of each k where m is a vector. The formula is:

[0155]

[0156] Change the phase space reconstruction dimension m = m + 1, repeat the above steps to obtain The approximate entropy can be obtained by calculation:

[0157]

[0158] 3) Fuzzy entropy

[0159] The fuzzy membership function is introduced on the basis of entropy to reduce the impact of time series fluctuations by calculating the mean. The continuity of the fuzzy membership function ensures that the fuzzy entropy value will not change greatly with the change of the sequence. The algorithm of fuzzy entropy is as follows:

[0160] Time Series Z N ={z1,z2,…z N The length is N. Reconstructing the sequence in m (m>2) dimensional phase space can obtain the sequence: Z m (i) = {z(k), z(k+1), ..., z(k+m-1)}, k = 1, 2, ..., N-m+1. Similarity tolerance r = 0.2*std, std is Z N Standard deviation. Compute the mean of each dimension of the time series:

[0161]

[0162] A new m-dimensional time series can be obtained:

[0163] H m(k)={z(k),z(k+1),...,z(k+m-1)}-z0(k)

[0164] The distance d between two-dimensional time series is calculated as the maximum difference between the corresponding elements of the two:

[0165] d=max||z(k+s)-z0(k)|-|z(j+s)-z0(j)||

[0166] The similarity of these two-dimensional time series is calculated using the fuzzy membership function:

[0167]

[0168] Where r is the similarity tolerance, calculate:

[0169]

[0170] Finding the average value yields:

[0171]

[0172] Changing the dimension to m=m+1, repeating the above steps, the calculation formula of fuzzy entropy is:

[0173] FE=lnψ m (r)-ψ m+1 (r)

[0174] An embodiment of the present invention also provides another cable fault feature extraction device based on an improved GAN algorithm, comprising a processor and a memory configured to store a computer program that can be run on the processor; wherein, when the processor is configured to run the computer program, the method steps in the aforementioned embodiment are executed.

[0175] In practical applications, the processor includes a field programmable gate array (FPGA), and the processor may be a central processing unit (CPU) or a digital signal processor (DSP). It is understandable that for different devices, the electronic device used to implement the function of the processor may be other, and the embodiment of the present invention does not specifically limit it.

[0176] The above-mentioned memory can be a volatile memory (volatile memory), such as a random access memory (RAM); or a non-volatile memory (non-volatile memory), such as a read-only memory (ROM), a flash memory, a hard disk (HDD) or a solid-state drive (SSD); or a combination of the above-mentioned types of memory, and provide instructions and data to the processor.

[0177] In an exemplary embodiment, an embodiment of the present invention further provides a computer-readable storage medium for storing a computer program.

[0178] Optionally, the computer-readable storage medium can be applied to any one of the methods in the embodiments of the present invention, and the computer program enables the computer to execute the corresponding processes implemented by the processor in each method in the embodiments of the present invention. For the sake of brevity, they are not described here.

[0179] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. 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 can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0180] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The schemes in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0181] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0182] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0183] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0184] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0185] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A cable fault feature extraction method based on an improved GAN algorithm, characterized in that The steps include: Step A: Collect the three-phase current i of the cable abc And the corresponding time t, state information y, for the three-phase current i abc Processing is performed to obtain a sample current x; Step B, calculating the permutation entropy and approximate entropy of the sample current x, and constructing a latent variable Q(t) based on the permutation entropy and approximate entropy as an input of the training generator; Step C, import the latent variable Q(t) into the training generator G to obtain the synthetic data G{Q(t)}; Step D: import the real data x and the synthetic data G{Q(t)} into the training discriminator J for discrimination, and adjust the network parameters θ of the training discriminator J and the training generator G according to the discrimination results. J ,θ G make corrections; Step E, repeat steps CD until the discriminator cannot distinguish between synthetic data and real data, and terminate the iteration, outputting the same number of sample data sets in different states; Step F, extracting the eigenvalues ​​of the three-phase current of the cable in the time domain, frequency domain and time-frequency domain under healthy and fault conditions, and establishing a characteristic matrix A.

2. The method according to claim 1, characterized in that: In step A, the state information y includes the following states: State 1: Phase A ground fault State 2: Phase B ground fault State 3: Phase C ground fault State 4: AB two-phase grounding fault State 5: BC two-phase grounding fault State 6: AC two-phase ground fault State 7: Three-phase ground fault State 8: Normal state State 9: Arc fault.

3. The method according to claim 1, characterized in that: In step A, the three-phase current i abc The formula for processing the sample current x is, In the formula, x min is the three-phase current i abc The minimum amplitude of x max is the three-phase current i abc The maximum amplitude.

4. The method according to claim 1, characterized in that: In step B, the method for calculating the permutation entropy of the sample current x is: Setting up the time series dataset in Represents the i-th true sample at time t, i = 1,…,n t , n t is the number of real samples, The sample current contains N sampling points; the permutation entropy of each sample is calculated with a sliding window of length L, L ≤ N, and reconstructed within the sliding window The phase space of is described as: Where m is the embedding dimension, τ is the time delay, and M = L-(m-1)τ is the number of vectors in the reconstructed image space. The j-th row of the reconstruction matrix is ​​arranged in ascending order to form a new sequence, which is expressed as: Where a1 represents the index of the minimum value in the jth row of the reconstruction matrix, a m represents the index of the maximum value in the jth row of the reconstruction matrix; the sequence thus composed has m! possible permutations; the permutation entropy is expressed as: Where P a represents the probability distribution of symbols, K≤m!.

5. The method according to claim 1, characterized in that: In step B, the method for calculating the approximate entropy of the sample current x is: Assume that the time series Z N ={z1,z2,…z N The length is N, and the sequence is reconstructed into m-dimensional phase space, m>2, and the sequence is obtained: Z m (i) = {z(k), z(k+1), …, z(k+m-1)}, k = 1, 2, …, N-m+1; the distance between the two dimensions is: d=max|Z m (k)-Z m (j)|,j≠k,k≤j≤N-m Where max is the maximum difference between the corresponding elements of the two-dimensional time series, and the similarity tolerance r=0.2*std is defined, where std is Z N Standard deviation; the number of vectors whose distance is less than the similarity tolerance r is the ratio of num to the total vector: Take the logarithm of the above ratio and find the average value of each k where m is a vector. The formula is: Change the phase space reconstruction dimension m = m + 1, repeat the above steps to obtain The approximate entropy obtained by calculation is:

6. The method according to claim 1, characterized in that: In step B, a latent variable Q(t) is constructed based on the permutation entropy and the approximate entropy according to the following formula: In the formula, iter is the current iteration number, iter max is the maximum number of iterations, and y is the current state.

7. The method according to claim 1, characterized in that: In step C, the loss function of training generator G is, In the formula, G and J are the training generator and the training discriminator respectively, q represents the latent variable Q(t), P q (q) is the probability of the latent variable; J(·) is the probability that the input to the trained discriminator is real data rather than synthetic data.

8. The method according to claim 1, characterized in that: In step D, the loss function of training the discriminator J is, In the formula, xˊ represents the real data, P x (x) is the distribution of real data.

9. The method according to claim 1, characterized in that: In step D, the network parameters θ of the training discriminator J and the training generator G are adjusted according to the discrimination results. J ,θ G Correction is made, where the network parameters θ of the training discriminator J are J The corrected formula is, θ J =1+αJ(G(q)) The network parameters θ of the training generator G G The corrected formula is, θ G =1-αJ(G(q)) Where α is a constant; q represents the latent variable Q(t), and J(·) is the probability that the input to the discriminator is real data rather than synthetic data.

10. The method according to claim 1, characterized in that: In step E, the necessary condition for the discriminator to be unable to distinguish between synthetic data and real data is J(G(q))≈1.

11. The method according to claim 1, characterized in that: In step F, the time domain features in the feature matrix A include maximum value, minimum value, standard deviation, root mean square value and kurtosis, the frequency domain features include centroid frequency, average frequency, root mean square frequency and frequency standard deviation, and the time-frequency domain features include singular entropy, approximate entropy and fuzzy entropy.

12. A cable fault feature extraction device based on an improved GAN algorithm, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that: When the processor executes the computer program, the steps of the cable fault feature extraction method based on the improved GAN algorithm as described in any one of claims 1 to 11 are implemented.

13. A computer-readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, the steps of the cable fault feature extraction method based on the improved GAN algorithm as described in any one of claims 1 to 11 are implemented.

Images