Distribution network fault detection and location method and system based on transfer convolutional neural network

Abstract

The present invention provides a method and system for detecting and locating distribution network faults based on a transfer convolutional neural network, wherein the method of the present invention performs waveform preprocessing on distribution network fault signals by using an S-transformation method of a hyperbolic window; after the preprocessed data, further sample image construction is performed using a time-frequency matrix to form a block time-frequency spectrum as a feature quantity input to a transfer convolutional neural network; and the transfer convolutional neural network is used to perform fault analysis to detect the type of distribution network faults and faulty lines. The present invention obtains deep-level grid fault information by analyzing grid fault information based on a transfer convolutional neural network, thereby solving the influence of topological structure differences. At the same time, for situations where data is limited, faults can also be accurately detected and located, and the method has universality.

Description

Technical Field

[0001] The present invention belongs to the field of electric power technology, and in particular relates to a distribution network fault detection and positioning method and system based on a migration convolutional neural network. Background Art

[0002] With the development and changes of the economy and society, residents have higher and higher requirements for the quality and sustainability of power supply in the power system. The distribution network is an important part of the power grid, connecting the transmission network and power users. It undertakes the important task of providing safe and reliable power supply to users and is an important infrastructure for improving people's livelihood. The distribution network has many devices and complex connections. How to correctly identify and handle faults after a power grid failure will have a great impact on the power supply stability of the distribution network. For a long time, research on power system fault diagnosis has been focused on the transmission network, and research on the distribution network is relatively small. Therefore, research on distribution network fault diagnosis is of great significance for the dispatching center to improve the efficiency of handling distribution network faults.

[0003] At present, the methods for distribution network fault diagnosis at home and abroad mainly include traveling wave method, impedance method, method based on artificial neural network, method based on analytical model and method based on matrix. However, the existing methods all have some disadvantages, such as being restricted by topology or requiring a large amount of data support. In the above situation, it is very necessary to develop a fault diagnosis method that can quickly and accurately diagnose distribution network faults and has redundancy for the refusal or malfunction of protection and switches and the false alarm or omission of alarm signals for the safe and stable operation of the distribution network. Summary of the invention

[0004] In view of this, the present invention aims to solve the problem that the existing distribution network fault diagnosis method is easily restricted by the topology, or requires a large amount of data support and cannot quickly and accurately diagnose distribution network faults.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a distribution network fault detection and location method based on a migration convolutional neural network, comprising the following steps:

[0007] Obtain the fault waveforms of various signals during historical faults in the distribution network, and use the hyperbolic window S-transform method to preprocess the fault waveforms;

[0008] The preprocessed waveform data is used to construct a sample image using a time-frequency matrix to obtain a block time-frequency spectrum;

[0009] The block time-frequency spectrum is used as input and the fault type is used as output to train the transfer convolutional neural network;

[0010] The fault waveform to be tested when the distribution network fails is collected as the input of the trained transfer convolutional neural network for fault type identification, and the output with the highest probability is used as the final distribution network fault type, and the final distribution network fault type and the corresponding fault line are output.

[0011] Furthermore, the preprocessed waveform data is used to construct a sample image using a time-frequency matrix to obtain a block time-frequency spectrum, which specifically includes:

[0012] Constructing the time-frequency matrix of the preprocessed waveform data frequency bands through bandpass filtering;

[0013] Each frequency band sequence in the time-frequency matrix is ​​equally divided according to time to obtain a number of time-frequency blocks and the energy of each time-frequency block in each frequency band sequence is calculated;

[0014] The time-frequency matrix of the frequency division band is processed based on the energy of each time-frequency block to obtain a block time-frequency spectrum matrix.

[0015] Furthermore, the transfer convolutional neural network is obtained by transfer learning using the convolutional neural network that has completed the distribution network fault classification training, specifically including:

[0016] The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training;

[0017] Perform migration fine-tuning training on the convolutional neural network model. During the training process, freeze the network parameters of the hidden layer of the convolutional neural network model, reset the parameters of the fully connected layer, and do not update the network parameters of the hidden layer during back propagation. After the learning of the fully connected layer is completed, unfreeze the hidden layer and then train the convolutional neural network model as a whole.

[0018] The convolutional neural network model that has undergone transfer fine-tuning training is defined as a transfer convolutional neural network. The distribution network fault data is used to verify the fault line selection effect of the transfer convolutional neural network. The line selection accuracy is used to evaluate the line selection effect and performance of the transfer convolutional neural network.

[0019] Furthermore, the convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training, specifically including:

[0020] Initialize the structure of the convolutional neural network to be trained;

[0021] Preprocess all the block time-frequency spectra and divide all the block time-frequency spectra samples into training set and test set;

[0022] The network is continuously trained using the training set until the output error meets the accuracy requirement or reaches the upper limit of the iteration. The parameters of each layer are adjusted each time the network is trained.

[0023] Save the latest parameters of each layer at the end of training;

[0024] The test set is labeled according to the fault type and classified and identified using the trained convolutional neural network until the recognition rate reaches the set requirement, thus obtaining a convolutional neural network model that completes the distribution network fault classification training.

[0025] Furthermore, the hyperbolic window S-transform method performs waveform preprocessing on the fault waveform through a hyperbolic window function, and the hyperbolic window function is specifically:

[0026]

[0027] Where X is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters and is the falling edge parameter of the function, is the rising edge parameter of the function, λ HY is the positive curvature parameter, and f represents the frequency.

[0028] In a second aspect, the present invention provides a distribution network fault detection and positioning system based on a migration convolutional neural network, comprising:

[0029] A data preprocessing unit is used to obtain the fault waveforms of various signals during historical faults in the distribution network, and to perform waveform preprocessing on the fault waveforms using an S-transformation method using a hyperbolic window;

[0030] An imaging unit is used to construct a sample image of the preprocessed waveform data using a time-frequency matrix to obtain a block time-frequency spectrum;

[0031] The recognition model training unit is used to train the migration convolutional neural network using the block time-frequency spectrum as input and the fault type as output;

[0032] The fault identification unit is used to collect the fault waveform to be tested when the distribution network fails as the input of the trained transfer convolutional neural network to identify the fault type, take the output with the highest probability as the final distribution network fault type, and output the final distribution network fault type and the corresponding fault line.

[0033] Furthermore, the imaging unit is specifically used for:

[0034] Constructing the time-frequency matrix of the preprocessed waveform data frequency bands through bandpass filtering;

[0035] Each frequency band sequence in the time-frequency matrix is ​​equally divided according to time to obtain a number of time-frequency blocks and the energy of each time-frequency block in each frequency band sequence is calculated;

[0036] The time-frequency matrix of the frequency division band is processed based on the energy of each time-frequency block to obtain a block time-frequency spectrum matrix.

[0037] Furthermore, the transfer convolutional neural network in the recognition model training unit is obtained by transfer learning using the convolutional neural network that has completed the distribution network fault classification training, specifically including:

[0038] The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training;

[0039] Perform migration fine-tuning training on the convolutional neural network model. During the training process, freeze the network parameters of the hidden layer of the convolutional neural network model, reset the parameters of the fully connected layer, and do not update the network parameters of the hidden layer during back propagation. After the learning of the fully connected layer is completed, unfreeze the hidden layer and then train the convolutional neural network model as a whole.

[0040] The convolutional neural network model that has undergone transfer fine-tuning training is defined as a transfer convolutional neural network. The distribution network fault data is used to verify the fault line selection effect of the transfer convolutional neural network. The line selection accuracy is used to evaluate the line selection effect and performance of the transfer convolutional neural network.

[0041] Furthermore, the convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training, specifically including:

[0042] Initialize the structure of the convolutional neural network to be trained;

[0043] Preprocess all the block time-frequency spectra and divide all the block time-frequency spectra samples into training set and test set;

[0044] The network is continuously trained using the training set until the output error meets the accuracy requirement or reaches the upper limit of the iteration. The parameters of each layer are adjusted each time the network is trained.

[0045] Save the latest parameters of each layer at the end of training;

[0046] The test set is labeled according to the fault type and classified and identified using the trained convolutional neural network until the recognition rate reaches the set requirement, thus obtaining a convolutional neural network model that completes the distribution network fault classification training.

[0047] Furthermore, the hyperbolic window S-transform method performs waveform preprocessing on the fault waveform through a hyperbolic window function, and the hyperbolic window function is specifically:

[0048]

[0049] Where X is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters and is the falling edge parameter of the function, is the rising edge parameter of the function, λ HY is the positive curvature parameter, and f represents the frequency.

[0050] In summary, the present invention provides a method and system for detecting and locating distribution network faults based on a transfer convolutional neural network, wherein the method of the present invention performs waveform preprocessing on distribution network fault signals by using the S-transformation method of a hyperbolic window; after the preprocessed data, further sample image construction is performed using a time-frequency matrix to form a block time-frequency spectrum as a feature quantity input to a transfer convolutional neural network; and a transfer convolutional neural network is used to perform fault analysis to detect the type of distribution network faults and faulty lines. The present invention obtains deep-level grid fault information by analyzing grid fault information based on a transfer convolutional neural network, which solves the impact of topological structure differences. At the same time, it can accurately detect and locate faults in situations where data is limited, and has universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 A schematic diagram of a flow chart of a distribution network fault detection and location method based on a migration convolutional neural network provided in an embodiment of the present invention;

[0053] Figure 2 A basic structural diagram of a convolutional neural network provided by an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of a convolution operation provided by an embodiment of the present invention;

[0055] Figure 4 A schematic diagram of a pooling operation provided by an embodiment of the present invention;

[0056] Figure 5 A schematic diagram of the structure of the fully connected layer and the output layer provided in an embodiment of the present invention;

[0057] Figure 6 A training flow chart of a CNN model for distribution network fault classification provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The research on distribution network fault diagnosis is an important research direction of distribution network. The realization of fast and accurate distribution network fault research is of great significance to improving the power supply quality of distribution network. Since the research on power system fault diagnosis has been focused on the transmission network for a long time, there are more and more mature methods for transmission network fault diagnosis, while the research on distribution network is relatively less. At the same time, the distribution network has more components and a larger scale, which also makes the difficulty of distribution network fault diagnosis much greater than that of transmission network. After the distribution network fails, the action of protection and switch will be transmitted to the dispatching center in the form of alarm signals, and the dispatching personnel will analyze and diagnose the fault. At present, the methods for distribution network fault diagnosis at home and abroad mainly include traveling wave method, impedance method, method based on artificial neural network, method based on analytical model and method based on matrix. However, the existing methods have more or less some shortcomings, such as being restricted by topology, or requiring a large amount of data support, etc. In the above situation, the development of fault diagnosis methods that can quickly and accurately diagnose distribution network faults and have redundancy for the refusal or malfunction of protection and switch and the false alarm or omission of alarm signals is of great significance for the safe and stable operation of distribution network and improving the power supply quality for power users.

[0060] In the existing research on distribution network fault diagnosis and location, the diagnosis and location technology faces many problems. The distribution network has a complex structure, multiple branches, and multiple comprehensive loads. Its harmonic quality, voltage fluctuations and other factors will affect the positioning accuracy. At the same time, unlike the traditional single-power radial distribution network, the access of distributed power sources makes the power flow bidirectional. Most of the methods based on traditional protection relays are no longer applicable. Therefore, it is necessary to explore new methods suitable for changes in the distribution network structure. At present, various methods for locating distribution network faults have more or less inherent defects. The impedance method and graph theory method require the known distribution network topology structure, and sometimes the known grounding impedance. When the distribution network structure changes, the correlation matrix needs to be reconstructed; the traveling wave method is not suitable for distribution networks with very short lines, and the traveling wave measurement equipment has extremely high accuracy requirements; artificial intelligence algorithms are widely faced with the contradiction between convergence and rapidity, and are prone to local fault points.

[0061] Based on this, the present invention provides a distribution network fault detection and positioning method and system based on a migration convolutional neural network.

[0062] The following is a detailed introduction to an embodiment of a distribution network fault detection and positioning method based on a migration convolutional neural network of the present invention.

[0063] See also Figure 1 This embodiment provides a distribution network fault detection and positioning method based on a migration convolutional neural network, comprising:

[0064] S100: Obtain the fault waveforms of various signals during historical faults in the distribution network, and perform waveform preprocessing on the fault waveforms using the S-transformation method of a hyperbolic window.

[0065] In this embodiment, seven fault waveforms, including bus three-phase voltage, zero-sequence voltage, and three-phase current on the low-voltage side of the main transformer, are collected. The waveform length is one cycle before the fault and two cycles after the fault. The sampling frequency is 10kHz, and the simulation time is 2s, that is, the number of sampling points is 20,000. The waveform data of one cycle before the fault and two cycles after the fault of each waveform with a total length of 0.06s are uniformly intercepted. Then, the seven fault waveforms are preprocessed, and then, according to the needs of the actual convolutional neural network, as the input of the convolutional neural network, the training samples and test samples are selected according to the conditions such as the initial phase angle of the fault, the fault resistance, and the fault point.

[0066] The S transform is obtained by replacing the wavelet basis in the wavelet transform with a Gaussian window function whose resolution can change with frequency. Since the Gaussian window cannot meet the admissible conditions of the wavelet, it is not a continuous wavelet transform in the simple sense. The equation of the known wavelet transform is shown in equation (1).

[0067]

[0068]

[0069]

[0070] Wherein, equation (2) is the wavelet basis. It is replaced by equation (3), and the scale factor of S transform is d=1 / |f|, that is, the formula of S transform is obtained. S transform is defined as:

[0071]

[0072] In the S transform, the expansion and contraction of the window function, that is, the window width, is controlled by 1 / |f|, and the translation of the window, that is, the position of the window on the time axis, is controlled by τ. It can be seen that the translation and expansion of the window are separate and do not interfere with each other. When the window is translated on the time axis, the window width can be unchanged, and the shape of the window function will not change; similarly, when the window shape changes, the expansion and contraction of the window will not affect the movement of the window position.

[0073] Compared with the kernel function of Fourier transform, the kernel function of S transform has an additional window function. The width of the window function is inversely proportional to the frequency domain, that is, the width of the window changes during the signal analysis process. When f becomes smaller, the window becomes larger, and when f becomes larger, the window becomes smaller, which overcomes the disadvantage that the width of the Fourier transform window cannot be changed once it is determined. S transform is similar to Fourier transform, so the calculation process is relatively simple, and it can be calculated quickly with a fast calculation speed. The result obtained by calculation is the relationship between time and frequency, which is more intuitive and accurate than the relationship between time and scale obtained by wavelet transform. The time-frequency characteristics of the local signal represented by S transform, if the final result is integrated over time, the Fourier transform of the signal can be obtained:

[0074]

[0075] The time domain window width of the S transform is

[0076]

[0077] The frequency domain window width of the S transform is

[0078]

[0079] The S transform is similar to the wavelet transform using Morlet as the wavelet basis, except that the phase factor e is added to the S transform. -j2πft , so that it retains the phase information in Fourier transform and can accurately identify parameters. Therefore, the S transform inherits the advantages of wavelet multi-resolution and retains the Fourier phase information, combining the advantages of both, and can be more accurate in frequency identification.

[0080] In order to ensure the reversibility of the S transform, the area of ​​the window function is required to be 1, that is:

[0081]

[0082] The inverse transform of S transform is derived from the Fourier transform formula, which has the property of lossless inverse transform and will not lose the energy of the signal. The inverse transform of S transform is defined as:

[0083]

[0084] Compared with other signal processing methods, S transform has the following advantages: it is a linear transform; it retains the signal phase information; it retains the characteristics of multi-resolution analysis; the result obtained by S transform is the relationship between time and frequency, avoiding the conversion of scale and frequency in wavelet transform. However, it still has some problems. The window width of S transform is inversely proportional to the frequency. The width of the window is suitable in some frequency ranges of the signal, but it is not suitable in other ranges. For example, the width of the window increases too fast in the low-frequency range, and the window width is too large, making the resolution no longer suitable for signal analysis. At high frequencies, if the window function becomes too narrow, it may be smaller than the wavelength of the signal, which will cause distortion during signal analysis. However, the change trend of the window cannot be changed. In order to make the size and change trend of the window more in line with the needs of the signal and make the calculated time-frequency graph aggregation achieve better results, a method to improve S transform is proposed.

[0085] The improvement method is mainly divided into two ideas. The first is to improve the size and change trend of the window function of the S transform. The main method is to add parameters to the window function and change the window width by controlling the size of the parameter value. The second is to adjust the type of window function to overcome the disadvantages of the Gaussian window function in the S transform.

[0086] The improved S transform is called the generalized S transform. The generalized S transform is obtained by replacing the Gaussian window function in the S transform with a generalized window function. The expression of the generalized S transform is:

[0087]

[0088] Among them, w(t,f,p) represents the improved window function. The parameters added to the window function are uniformly represented by P. The width of the window can be adjusted by the parameter P, and the resolution at each time point and frequency point can be controlled, so that the resolution at each frequency point of the signal is appropriate. Once the width of the window function is determined at each frequency f, it will no longer change. The window function moves on the time axis, and the shape and size of the window do not change. The position of the window on the time axis is determined by the parameter t.

[0089] From the properties of the S transform, we can see that when improving the S transform, the following conditions must be met:

[0090]

[0091] This condition indicates that the window function becomes When the window is changed, the area inside the window cannot be changed and must be 1.

[0092] Integrating the generalized S transform in the time direction gives the Fourier transform of the signal x(t):

[0093]

[0094] Signals x(t) and x(f) are inverses of each other. x(f) is obtained by Fourier transform, and x(t) is obtained by inverse Fourier transform. Therefore, it can be concluded from the above formula that x(t) and S(t,f,p) are also reversible. S(t,f,p) is obtained by generalized S transform of x(t). S(t,f,p) is integrated in the time domain to obtain Fourier transform, and the result of Fourier transform is subjected to inverse Fourier transform to obtain signal x(t). Therefore, x(t) can also be obtained by inverse transform of s(t,f,p).

[0095] In order to apply the fast Fourier transform to the generalized S transform, improve the calculation speed, and apply the convolution theorem, the generalized S transform formula is deformed to obtain a new form:

[0096]

[0097] Where X(α+f) and W(α,f,p) represent the Fourier transform results of the signal and window function respectively:

[0098]

[0099]

[0100] In the formula, α is the same as f, both of which are frequency variables. When the signal is transformed from the time domain to the frequency domain, it is multiplied by e. -j2παt , so in the frequency domain X(f) becomes X(α+f).

[0101] The advantage of the generalized S transform is that it has good local signal processing capabilities and a high resolution effect in adjacent frequencies. It can identify the information of signal changes with time and frequency as accurately as possible. Therefore, the window function should be as narrow as possible in both domains. Like the window in the S transform, it must satisfy the uncertainty theorem, that is, the product of the width Δt in the time domain and the width Δω in the frequency domain has a minimum value.

[0102]

[0103] The generalized S transform retains the invertibility of the original S transform and its relationship with the Fourier spectrum. Therefore, the generalized S transform can be used for time-frequency filtering using a band-limited filter that is activated only within a specific time range.

[0104] In the above, the parameters introduced in the improvement of S transform only change the height and width of the window, but not the shape. Changing the symmetric window in S transform into an asymmetric window function makes the window more flexible and enhances the control of the window function, so as to expand the application scope of S transform and enhance the accuracy of time-frequency analysis results.

[0105] In some cases, it is necessary to determine the time information of the signal when the frequency changes. It is necessary to narrow the window function width in the time domain to enhance the time recognition ability. For example, the γ GS Take a smaller value, ω GS The time domain window width will become very small, which can better identify the time when the frequency changes, and can also well identify the characteristics of the changes, with better discrimination effect, but it will inevitably cause the width of the window in the frequency domain to become larger, and the frequency resolution effect will deteriorate, resulting in some losses during frequency identification, so it will have a certain impact on the information identification of the signal in the global scope. Therefore, when using the window function mentioned above, in order to change the time resolution, the window width can only be made smaller, which affects the frequency resolution. Conversely, in order to take into account the frequency resolution, the window width in the frequency domain is reduced, and the time resolution cannot meet the requirements. When picking up signal information, there are always some parameters that are inaccurate. In order to solve this problem, the present embodiment uses an S transform composed of an asymmetric window, which is intended to improve the initial time resolution during signal analysis of the Gaussian S transform.

[0106] However, one problem with using a Gaussian function is that the time resolution in the time-frequency spectrum is reduced due to the long front taper. To address this problem, the present embodiment uses a generalized S transform, which includes asymmetric windows, including windows with complex scaling characteristics, including frequency-dependent shapes. First, the amplitude distribution of the signal under a given window shape and width is calculated using a generalized S transform as a function of time and frequency. The location of the maximum amplitude is then determined, and the instantaneous phase at that location is calculated. The estimated value of the component is then synthesized using the same window and amplitude, phase, frequency and time values ​​and subtracted from the original signal. By changing the shape and width of the window, a combination can be found that minimizes the energy of the residual signal, thereby representing the best match with the real component. The process is repeated within a given number of iterations, or until the amplitude of the residual signal is below a preset threshold, thereby decomposing the original signal into a set of components and a residual signal of low amplitude. The window function library studied includes Gaussian functions, simple exponential decay functions, cosine function amplitude modulation, and cosine function phase modulation, and this method can be extended to include any other suitable functions. This generalized S transform retains the original reversibility.

[0107] The process of calculating the generalized S transform of a signal is as follows:

[0108] Step (1): Calculate the forward Fourier transform of the signal using the fast Fourier transform function;

[0109] Step (2): Set the amplitude of negative frequencies to zero and double the amplitude of positive frequencies, thereby generating a spectrum of the analysis signal;

[0110] Step (3): Generate the Fourier transform of the selected window function for the desired frequency f and window parameters;

[0111] Step (4): Shift the spectrum of the analysis signal by f in the negative frequency direction;

[0112] Step (5): multiplying the displacement spectrum by the complex conjugate of the Fourier transform of the window function;

[0113] Step (6): Calculate the inverse Fourier transform using the inverse Fourier transform function;

[0114] Step (7): Calculate the amplitude and phase of the generated complex signal;

[0115] Step (8): Repeat steps (3) to (7) for different frequency values ​​within the desired range.

[0116] The process of implementing the decomposition of a signal using the generalized S transform is as follows:

[0117] Step (1): Calculate the total energy of the original signal.

[0118] Step (2): Select a window function from the available window library.

[0119] Step (3): Select one or more parameters for the window function.

[0120] Step (4): As described above, the time-frequency distribution of the signal is calculated using the generalized S-transform of the signal.

[0121] Step (5): Find the maximum amplitude of the time-frequency distribution, find the frequency and time at which it occurs, and then calculate the instantaneous phase at that point.

[0122] Step (6): Use amplitude, phase, frequency and time, and have the same window function and window parameters.

[0123] Step (7): Subtract the synthetic component from the signal.

[0124] Step (8): Add the energies of all synthesized components and the residual signal. If this sum, i.e., the total energy, is significantly different from the total energy of the original signal, a warning is flagged.

[0125] Step (9): Repeat steps (3) to (8) for a series of different window parameters until a parameter or set of parameters is found that minimizes the residual signal energy.

[0126] Step (10): Repeat steps (2) to (9) for each different window function, and select the function that minimizes the residual signal energy.

[0127] Steps (2) through (10) are repeated either for a predetermined number of passes or until the remaining signal contains less than a specified energy. The nesting of loops in steps (9) and (10) to find the optimal window function and window parameters will multiply the number of times the time-frequency distribution must be calculated in step (4). Some windows may have multiple parameters.

[0128] One way to solve this problem is to use an asymmetric window. When performing signal analysis, people generally pay more attention to the time when the event occurs. However, in a symmetrical window function, the time resolution effect is poor due to the wide front edge of the window, and the time parameter cannot be well identified. Therefore, a symmetrical window function is used to make the curve steeper in the first half of the window to increase the resolution effect, and to make the second half of the window flatter, which plays the role of window function compensation. The wider second half will reduce the judgment of the end time of the event. In order to improve the better analysis of the occurrence time of the signal event, the resolution of the end time of the event is sacrificed. When analyzing, it is necessary to analyze the characteristics of the signal. The characteristic refers to an identifiable area on the time-frequency plane, which has a limited amplitude and is surrounded by areas with negligible amplitudes, or surrounded by other areas with limited amplitudes. The characteristics of these areas allow some regularities and characteristics to be visually distinguished from the area in question. At high frequencies, the window is narrow and the time resolution is good, so a more symmetrical window should be used. At low frequencies where the window is wide and the frequency resolution is not critical, the frequency resolution is not important, and a more asymmetric window can be used to prevent the event from appearing too far in the S transform. Based on this concept, a "hyperbolic" window is designed. The hyperbolic window is a pseudo-Gaussian window, and the window function is defined as:

[0129]

[0130] in

[0131]

[0132] In equation (18), x is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters: and The function falling edge parameter is The larger the value, the greater the slope of the curve trailing edge and the steeper the curve trailing edge. The rising edge parameter of the function is The slope of the front of the curve is positively correlated with it. The larger the value, the larger the absolute value of the slope, and the steeper the curve. (Assuming ) and the positive curvature parameter λ HY , which determines the symmetry of the curve. The larger its value, the faster the symmetry of the window function changes with frequency. The translation of ζ ensures that ω HY The peak value of appears at (τ-t) = 0; ζ is defined as

[0133]

[0134] Although the mathematical expression of the hyperbolic S-transform is more complicated than that of the Gaussian S-transform, the hyperbolic S-transform is not difficult to compile.

[0135] Near f = 0, ω HY is very asymmetric, but as f increases, ω HY The shape of GS The shape of the window converges (i.e., symmetric). The blurring effect of horizontal and vertical expansion ω HY This occurs when f changes, making the frequency-dependent asymmetry difficult to discern visually. To address this issue, each Gaussian window is scaled up horizontally and vertically so that they all have a maximum height of 1 and a standard deviation of 1. After this rescaling, it can be seen that ω GS At f = 2, it is less symmetrical than at f = 0.5. If you choose a larger As f increases, the shape change from an asymmetric window to a symmetric window will be faster. It is empirically found that when and The best working HY The value is slightly less than 1 / f p , where f p is the peak frequency of important events in the time series.

[0136] S200: constructing a sample image using the preprocessed waveform data using a time-frequency matrix to obtain a block time-frequency spectrum.

[0137] Convolutional neural networks are currently mostly used for image recognition, and their input objects are images. In order to facilitate the use of transfer convolutional neural networks for fault classification and recognition, this embodiment uses a time-frequency matrix to further construct sample images to form a block time-frequency spectrum as an input feature quantity to the transfer convolutional neural network.

[0138] In order to more intuitively show the characteristics of each fault waveform in the time domain and frequency domain, the time-frequency matrix of each fault waveform is constructed by bandpass filtering. Assume that the fault waveform is decomposed into m sub-bands after bandpass filtering, the number of sampling points of each waveform is n, and the data points of each sub-band waveform are a ij (i=1,2,...,m;j=1,2,...n), then the time-frequency matrix can be obtained as

[0139]

[0140] Among them, the rows of the time-frequency matrix A represent the reconstructed waveform data of each sub-band after the fault waveform is bandpass filtered, and the columns represent the sampling time of the fault waveform. The time-frequency matrix A fully represents the time-frequency characteristics of the fault waveform in each sub-band, includes the transient steady-state information in the fault waveform, and separates them by frequency band, which is conducive to the extraction of feature quantities.

[0141] According to the singular value decomposition theory, for a real matrix A m×n , there must be two orthogonal matrices V m×m , U n×n and a diagonal matrix Make Established, of which are the singular values ​​of a real matrix A, whose rank is r(A)=min(m,n), V=(v1,v2,...,v m ) is AA T The unit eigenvector of n ) is A T The unit eigenvector of A is

[0142]

[0143] Singular value decomposition can be understood as a matrix A of rank r n×n Decomposed into a weighted sum of r m×n matrices of rank 1, each of which consists of two eigenvectors u i and v i Multiply them together to get the weight matrix A i The singular values ​​of Formula (23) indicates that after the singular value decomposition of matrix A, a series of sub-matrices A are obtained: i and the corresponding singular values These two can reflect the amount of time-frequency information contained in the matrix and characterize the inherent characteristic patterns of the matrix to a certain extent. The sum of the squares of the changes in the singular values ​​of the matrix does not exceed the norm of the perturbation matrix. When the elements of matrix A are disturbed or distorted, the singular values ​​of the matrix change little and have good stability. Therefore, in pattern recognition problems, the singular values ​​of the matrix are often used as characteristic quantities to characterize the signal.

[0144] The time-frequency matrix A reconstructed by bandpass filtering of the fault waveform is subjected to singular value decomposition. The time-frequency matrix corresponding to each fault waveform can be decomposed into N singular values, 1≤N≤min(m,n), where the number of non-zero singular values ​​represents the number of different characteristic modes contained in the time-frequency matrix A of the fault waveform, and the size of the value represents the proportion of the corresponding characteristic mode in the total mode, characterizing the characteristics of the fault waveform. As a stable matrix algebraic parameter, the singular value has good robustness and generalization ability, and can show strong adaptability to the multiple and complex fault conditions of the distribution network.

[0145] The time-frequency matrix is ​​reconstructed using the band waveform data and singular value decomposition is performed to obtain the 10th-order singular values ​​of the zero-sequence voltage. The proportion of the singular values ​​of the zero-sequence voltage is mainly concentrated in the first 4 singular values. In order to extract the main eigenvector, the singular value cumulative contribution rate k is defined, and the first s singular values ​​after the cumulative contribution rate reaches 85%, that is, k>85%, are selected as the main singular values.

[0146]

[0147] Where: i is an element in a set of singular values, and r is the total number of singular values ​​in this set.

[0148] At different fault points and fault types, when k>85%, the value of s is not the same. The first five singular values ​​corresponding to the seven fault waveforms (a total of 35 singular values) are uniformly selected as the feature quantities for distribution network fault type identification, and these singular values ​​are combined into a 1×35 feature vector matrix X. After using bandpass filtering and time-frequency matrix singular value decomposition on the seven fault waveforms in the above AG fault example, the extracted singular value feature quantities are used as comparison to distinguish the fault type.

[0149] Assume that the time-frequency matrix of a fault waveform is as shown in equation (22), and the fault waveform is decomposed into m frequency bands, each with n data points. In order to fully present the localized characteristics of the fault signal, the frequency bands (each row) in the time-frequency matrix are divided into Z equal parts according to time, and Z time-frequency regions are obtained, called time-frequency blocks. Each time-frequency block contains P data points (P = n / Z). Then the energy of the zth time-frequency block in the mth frequency band is defined as

[0150]

[0151] Where: The amplitude Δt of the yth data point in the zth time-frequency block of the mth frequency band represents the time interval between two data points, that is, the sampling period.

[0152] After processing the entire time-frequency matrix shown in formula (22) according to formula (25), the block time-frequency matrix can be obtained as follows:

[0153]

[0154] In the block time-frequency spectrum matrix E, each element represents the energy of a time-frequency block, which is arranged according to frequency band and time sequence, and each row has a total of Z elements. The selection of the Z value has a great influence on the fault recognition accuracy: if Z is too small, the energy of the obtained time-frequency block will not be able to express subtle local features; if Z is too large, the image of the block time-frequency spectrum will be larger, increasing the computational complexity of the convolutional neural network. After comprehensive consideration and verification by a large number of experiments, the Z value is taken as 60, which can better avoid the above two problems. The energy of the time-frequency block is similar to the pixels in a digital image. If the elements of E are plotted in the order of their own arrangement, a block time-frequency spectrum diagram can be obtained. If the block time-frequency spectrum diagrams corresponding to the 7 fault waveforms are spliced ​​from top to bottom in the order of three-phase voltage, three-phase current, and zero-sequence voltage, the total block time-frequency spectrum E can be obtained. total , its matrix expression is

[0155]

[0156] Similarly, the time-frequency matrix corresponding to the fault signal of the above fault instance is used for further feature quantity construction. According to the above-mentioned construction method of block time-frequency spectrum, the block time-frequency spectrum image corresponding to the group of fault signals is extracted. Each row in the block time-frequency spectrum represents a certain frequency band of the fault waveform. There are 7 fault waveforms in total, and each waveform corresponds to 10 frequency bands. Therefore, there are 70 rows in total, so the size of the block time-frequency spectrum image is 70×60. After a large number of experimental verifications, in most of the block time-frequency spectra, the block energy of the last 5 frequency bands of each fault signal accounts for a very small proportion of the energy of all frequency bands. Therefore, the block energy of the last 5 frequency bands of each fault signal is discarded, and the image size of the block time-frequency spectrum is reduced to 35×60. At the same time, considering that inputting a square image (with the same length and width) into the convolutional neural network is more convenient for the convolution and downsampling of the convolutional neural network, the blocked time-frequency spectrum is further cropped: the first 10 and last 15 energy blocks of each waveform band are discarded, and the middle 35 energy blocks are retained. The time range falls within half a cycle before the fault and one and a half cycles after the fault. After cropping, the transient steady-state characteristics of the fault signal before and after the fault can still be effectively retained, and the influence of the boundary effect can be further suppressed.

[0157] S300: The block time-frequency spectrum is used as input and the fault type is used as output to train the migration convolutional neural network.

[0158] The basic structure of a convolutional neural network is as follows Figure 2 As shown in the figure. The transfer-based convolutional neural network is mainly composed of input layer, convolution layer, pooling layer, fully connected layer, output layer and transfer learning. Figure 2It can be seen that the hidden layer is composed of convolutional layers and pooling layers. The convolutional layer is responsible for extracting features from the input data, and the pooling layer is responsible for reducing the dimension of the output of the convolutional layer. The convolutional layer and the pooling layer are alternately connected to complete the feature mining of the original data. The two-dimensional pooling layer data is expanded as the input of the fully connected layer to complete the data classification task. The specific steps are as follows:

[0159] Step 1: Convolutional layer.

[0160] The convolution operation of the convolutional neural network is equivalent to the inner product operation of the input matrix and the convolution kernel matrix. Figure 3 As shown in the figure, the size of the convolution kernel is 2×2. In the convolutional neural network, there are multiple convolution kernels in the same layer, and each convolution kernel has the same size. The convolution kernel moves on the input matrix in a certain order, and multiplies and sums each element of the matrix corresponding to its position. When the convolution kernel traverses all positions of the input matrix, the output matrix after the convolution operation is obtained. For example, Figure 3 In the example, the size of the circled matrix in the input matrix is ​​the same as the size of the convolution kernel. The matrices in the two positions are convolved with the convolution kernel to obtain the matrix in the output matrix. Different sliding strides have different effects on the convolution operation. By adjusting the sliding stride, the size of the output matrix can be changed. Padding the input samples can maintain the boundary information.

[0161] In order to select appropriate convolution layer parameters, it is necessary to understand the relationship between the input matrix size x×x, the convolution kernel size k×k, stride(s), padding(p), and the output matrix size y×y. The relationship between the five is shown in formula (20).

[0162]

[0163] The result of the convolution operation cannot be directly used as the input of the next layer. It is necessary to use an activation function to activate the result of the convolution operation as the input of the next layer. The activation functions include Sigmoid, Tanh, ReLU, etc. This patent selects ReLU as the activation function of the convolution layer.

[0164] Step 2: Pooling layer.

[0165] Pooling the data output by the convolutional layer can reduce the size of the data and prevent overfitting. The main pooling methods include maximum pooling and average pooling. Figure 4 As shown in the figure, with Stride = 2, the same input matrix is ​​subjected to maximum pooling and average pooling operations in a 2×2 area, and different output matrices are obtained. Maximum pooling selects the maximum value of the elements in the corresponding area as the output. For example, Figure 4In the figure, there are four elements in the upper left corner, 1, 3, 4, and 8, and the maximum value is 8, so the output element is 8. By performing the same operation on all regions, the output matrix is ​​obtained. Average pooling selects the average value of the elements in the corresponding region as the output. Figure 4 In , the average of the four elements in the upper left corner area is 4, and the output is 4. The relationship between the input matrix size x×x, the pooling area size n×n, the step size stride(s), and the output matrix size y×y is shown in formula (21).

[0166] Step 3: Fully connected layer and output layer.

[0167] The output of the pooling layer is multiple two-dimensional matrices. Each element of each matrix is ​​connected to the neurons of the fully connected layer to achieve the transformation from matrix data to one-dimensional data. The convolution layer and pooling layer perform feature mining on the data, and the fully connected layer and output layer classify the features. The structure of the fully connected layer and the output layer is as follows: Figure 5 shown.

[0168] Step 4: Transfer Learning

[0169] Transfer learning can use the labeled data to ensure the accuracy of the model on the new task, and can be classified according to the feature space, transfer scenario and transfer method. The model-based transfer method in the classification method mainly transfers network parameters and models to help train new models under small sample sets. The combination of transfer learning and deep learning has been maturely applied in the field of image recognition, providing new ideas for target recognition under small-scale data sets. The present invention will use model-based transfer learning to realize the combination of the two.

[0170] After determining the transfer learning method used, it is critical to correctly transfer the network model in actual application. According to the transfer learning principle, since the present invention can solve the problem of small fault data set, there is no need to retrain the network model. It is only necessary to use the CNN model to extract features from the fault data, and then randomize the fully connected layer parameters of the CNN, redefine the output layer related to the target task, and finally retrain the network under a small sample data set to obtain a new fully connected layer, so as to obtain a new model suitable for the target task.

[0171] CNN is mainly used in the field of image recognition and has good feature learning ability for images. However, the learning effect of CNN is affected by the network structure and the amount of data. The design flexibility of the network structure is high and there is no theoretical guidance. The current research is mainly based on existing experience. There are solutions to the problem of the impact of data volume on network learning performance. The combined application of transfer learning and CNN can realize its application under small sample data sets. Based on this, the present invention introduces a transfer learning method to solve the problem of insufficient fault data volume in network training, and applies transfer learning to the convolutional neural network to achieve the target task. The specific steps are as follows:

[0172] Step (1): Model pre-training

[0173] The full training of the CNN model is inseparable from a large number of training samples, so that the network can obtain better feature learning capabilities. However, in the fault line selection task, the sample data available for training is very limited and cannot reach the data scale of CNN model training. Therefore, the present invention selects the CNN model that completes the distribution network fault classification as the pre-training model for research. The source domain data of the model has a strong similarity with the target domain data studied in this paper, which is conducive to the migration of the pre-training model.

[0174] Step (2): Model transfer fine-tuning training

[0175] Since the pre-trained CNN model has the ability to extract features of power grid data images, it is very suitable as a feature extractor for fault time-frequency images to extract the main feature information. During the migration and fine-tuning training of the model, it is necessary to freeze the network parameters of the hidden layer, reset the parameters of the fully connected layer, and do not update the parameters of the hidden layer during back propagation. When the parameters of the fully connected layer are almost learned, unfreeze the hidden layer and then train the entire network as a whole to make the network suitable for fault line selection tasks.

[0176] Step (3): Fault line selection

[0177] In this embodiment, the CNN model that has undergone transfer fine-tuning training is defined as a transfer convolutional neural network (T-CNN). Small sample data is used to verify the line selection effect of the T-CNN model, and the line selection accuracy analysis is used to evaluate the line selection effect and performance of the T-CNN model.

[0178] In addition, a pre-trained model for migration convolutional neural network. The training of the CNN model for distribution network fault classification is to perform bandpass filtering on the fault signal, construct a block time spectrum, obtain a series of training samples and test samples, and then use the convolutional neural network to classify and identify the distribution network fault type. The output layer outputs a 10×1 type discrimination vector. Each element value in the output result is in [0,1]. The position number of the maximum value element is taken as the final fault type number, and the corresponding relationship is 1: AG; 2: BG; 3: CG; 4: ABG; 5: ACG; 6: BCG; 7: AB; 8: AC; 9: BC; 10: ABC. The CNN output result represents the probability that the input image belongs to a certain fault type, and the type with the highest probability is taken as the final judgment result.

[0179] like Figure 6 As shown, the specific training process is as follows:

[0180] Step (1): Initialize the CNN network structure, including the convolution kernel size, number, and number of output feature maps of each convolution layer (the convolution step size defaults to 1), the sampling window size, downsampling method, downsampling step size, etc. of each downsampling layer, and set the number of samples per batch iteration, the upper limit of iterations, etc.;

[0181] Step (2): adjust the image size of all block time-frequency spectrum samples to 35×35, which is the input image size required by the input layer; divide all samples into training set and test set;

[0182] Step (3): Initialize the parameters, initialize the weights ω of each layer, the bias term b, etc. to random numbers close to 0; initialize the hyperparameters a and λ to sufficiently small random numbers, and set the adjustment strategy for these two parameters as the number of training times increases;

[0183] Step (4): define the correlation matrix between each layer;

[0184] Step (5): Start training the network, perform a forward propagation, calculate the activation value of each layer in sequence, and then calculate the error between the actual output value of the output layer and the given type value;

[0185] Step (6): back-propagate the error value calculated in step (5), calculate the adjustment amount of each layer's weight and bias term, and determine whether the hyperparameters need to be updated;

[0186] Step (7): Adjust the weights and bias items of each layer according to the adjustment amount calculated in step (6);

[0187] Step (8): Repeat steps (5) to (7) until the error meets the accuracy requirement or reaches the upper limit of the number of iterations;

[0188] Step (9): After the training is completed, save the latest parameters at the end of the training.

[0189] Then test the fault sample data with fault type as label. When extracting features from the original data through the convolution layer, mark the test samples according to the fault type, and use the trained CNN to classify and identify the test samples. The classification and identification rules are shown in Table 1, where Y1~Y 10 10 elements representing the type output vector of the CNN.

[0190] Table 1 Fault type classification and identification rules

[0191]

[0192] The transfer convolutional neural network is trained according to the above rules.

[0193] S400: Collect the fault waveform to be tested when the distribution network fails as the input of the trained transfer convolutional neural network to identify the fault type, take the output with the highest probability as the final distribution network fault type, and output the final distribution network fault type and the corresponding fault line.

[0194] This embodiment provides a distribution network fault detection and location method based on a transfer convolutional neural network. The waveform of the distribution network fault signal is preprocessed by using the S transform method of the hyperbolic window; after the preprocessed data, the time-frequency matrix is ​​used to further construct a sample image to form a block time-frequency spectrum as a feature quantity input to the transfer convolutional neural network; the transfer convolutional neural network is used to perform fault analysis to detect the distribution network fault type and fault line. The present invention analyzes the information of the power grid fault based on the transfer convolutional neural network to obtain deep-level power grid fault information, which solves the influence of topological structure differences. At the same time, for situations with limited data, it can also accurately detect and locate faults, and has universality.

[0195] The above is a detailed introduction to an embodiment of a distribution network fault detection and positioning method based on a migration convolutional neural network of the present invention. The following is a detailed introduction to an embodiment of a distribution network fault detection and positioning system based on a migration convolutional neural network of the present invention.

[0196] This embodiment provides a distribution network fault detection and positioning system based on a migration convolutional neural network, including:

[0197] The data preprocessing unit is used to obtain the fault waveforms of various signals during historical faults in the distribution network, and to perform waveform preprocessing on the fault waveforms using the S-transformation method of the hyperbolic window.

[0198] It should be noted that the hyperbolic window S-transform method performs waveform preprocessing on the fault waveform through a hyperbolic window function, and the hyperbolic window function is specifically:

[0199]

[0200] Where X is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters and is the falling edge parameter of the function, is the rising edge parameter of the function, λ HY is the positive curvature parameter, and f represents the frequency.

[0201] The imaging unit is used to construct a sample image of the preprocessed waveform data using a time-frequency matrix to obtain a block time-frequency spectrum.

[0202] It should be noted that the imaging unit is specifically used to construct a time-frequency matrix of the preprocessed waveform data frequency bands through bandpass filtering; each frequency band sequence in the time-frequency matrix is ​​equally divided according to time to obtain a number of time-frequency blocks and the energy of each time-frequency block in each frequency band sequence is calculated; the time-frequency matrix of the frequency bands is processed based on the energy of each time-frequency block to obtain a block time-frequency spectrum matrix.

[0203] The recognition model training unit is used to train the migration convolutional neural network by taking the block time-frequency spectrum as input and the fault type as output.

[0204] It should be noted that the transfer convolutional neural network in the recognition model training unit is obtained by transfer learning using the convolutional neural network that has completed the distribution network fault classification training, specifically including:

[0205] The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training;

[0206] Perform migration fine-tuning training on the convolutional neural network model. During the training process, freeze the network parameters of the hidden layer of the convolutional neural network model, reset the parameters of the fully connected layer, and do not update the network parameters of the hidden layer during back propagation. After the learning of the fully connected layer is completed, unfreeze the hidden layer and then train the convolutional neural network model as a whole.

[0207] The convolutional neural network model that has undergone transfer fine-tuning training is defined as a transfer convolutional neural network. The distribution network fault data is used to verify the fault line selection effect of the transfer convolutional neural network. The line selection accuracy is used to evaluate the line selection effect and performance of the transfer convolutional neural network.

[0208] Furthermore, the convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training, specifically including:

[0209] Initialize the structure of the convolutional neural network to be trained;

[0210] Preprocess all the block time-frequency spectra and divide all the block time-frequency spectra samples into training set and test set;

[0211] The network is continuously trained using the training set until the output error meets the accuracy requirement or reaches the upper limit of the iteration. The parameters of each layer are adjusted each time the network is trained.

[0212] Save the latest parameters of each layer at the end of training;

[0213] The test set is labeled according to the fault type and classified and identified using the trained convolutional neural network until the recognition rate reaches the set requirement, thus obtaining a convolutional neural network model that completes the distribution network fault classification training.

[0214] The fault identification unit is used to collect the fault waveform to be tested when the distribution network fails as the input of the trained transfer convolutional neural network to identify the fault type, take the output with the highest probability as the final distribution network fault type, and output the final distribution network fault type and the corresponding fault line.

[0215] It should be noted that the fault detection and positioning system provided in this embodiment is used to implement the detection and positioning method provided in the aforementioned embodiment, and the specific configuration of each unit is subject to the complete implementation of the method, which will not be repeated here.

[0216] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. 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 invention.

Claims

1. A distribution network fault detection and location method based on a migration convolutional neural network, characterized in that: The steps include: Obtaining fault waveforms of various signals during historical faults of the distribution network, and performing waveform preprocessing on the fault waveforms using an S-transformation method of a hyperbolic window; The preprocessed waveform data is used to construct a sample image using a time-frequency matrix to obtain a block time-frequency spectrum; The block time-frequency spectrum is used as input and the fault type is used as output to train a migration convolutional neural network; The fault waveform to be tested when the distribution network fails is collected as the input of the trained migration convolutional neural network to identify the fault type, the output with the highest probability is used as the final determined distribution network fault type, and the final determined distribution network fault type and the corresponding fault line are output.

2. The distribution network fault detection and positioning method based on migration convolutional neural network according to claim 1 is characterized in that: The preprocessed waveform data is used to construct a sample image using a time-frequency matrix to obtain a block time-frequency spectrum, which specifically includes: Constructing the time-frequency matrix of the preprocessed waveform data frequency bands through bandpass filtering; Each frequency band sequence in the time-frequency matrix is ​​equally divided according to time to obtain a number of time-frequency blocks and the energy of each time-frequency block in each frequency band sequence is calculated; The time-frequency matrix of the sub-frequency bands is processed based on the energy of each time-frequency block to obtain a block time-frequency spectrum matrix.

3. The distribution network fault detection and positioning method based on migration convolutional neural network according to claim 1 is characterized in that: The transfer convolutional neural network is obtained by transfer learning using a convolutional neural network that has completed distribution network fault classification training, and specifically includes: The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training; Performing migration fine-tuning training on the convolutional neural network model, freezing the network parameters of the hidden layer of the convolutional neural network model during the training process, resetting the parameters of the fully connected layer, and not updating the network parameters of the hidden layer during back propagation, unfreezing the hidden layer after the learning of the fully connected layer is completed, and then training the convolutional neural network model as a whole; The convolutional neural network model that has undergone the migration fine-tuning training is defined as the migration convolutional neural network, and the fault line selection effect of the migration convolutional neural network is verified using distribution network fault data. The line selection accuracy is used to evaluate the line selection effect and performance of the migration convolutional neural network.

4. The distribution network fault detection and positioning method based on migration convolutional neural network according to claim 3 is characterized in that: The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training, including: Initialize the structure of the convolutional neural network to be trained; Preprocessing all the block time-frequency spectra, and dividing all the block time-frequency spectra samples into a training set and a test set; The network is continuously trained using the training set until the output error meets the accuracy requirement or reaches the upper limit of the iteration, and the parameters of each layer are adjusted each time the network is trained; Save the latest parameters of each layer at the end of training; The test set is marked according to the fault type and classified and identified using the trained convolutional neural network until the recognition rate reaches the set requirement, thereby obtaining the convolutional neural network model that completes the distribution network fault classification training.

5. The distribution network fault detection and positioning method based on migration convolutional neural network according to claim 1 is characterized in that: The hyperbolic window S-transform method performs waveform preprocessing on the fault waveform through a hyperbolic window function, and the hyperbolic window function is specifically: Where X is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters and is the falling edge parameter of the function, is the rising edge parameter of the function, λ HY is the positive curvature parameter, and f represents the frequency.

6. A distribution network fault detection and positioning system based on a migration convolutional neural network, characterized in that: include: A data preprocessing unit, used to obtain the fault waveforms of various signals during historical faults of the distribution network, and perform waveform preprocessing on the fault waveforms using an S-transformation method of a hyperbolic window; An imaging unit is used to construct a sample image of the preprocessed waveform data using a time-frequency matrix to obtain a block time-frequency spectrum; An identification model training unit, used for training a migration convolutional neural network by taking the block time-frequency spectrum as input and the fault type as output; The fault identification unit is used to collect the fault waveform to be tested when the distribution network fails as the input of the trained migration convolutional neural network to identify the fault type, take the output with the highest probability as the final determined distribution network fault type, and output the final determined distribution network fault type and the corresponding fault line.

7. The distribution network fault detection and positioning system based on transfer convolutional neural network according to claim 6 is characterized in that: The imaging unit is specifically used for: Constructing the time-frequency matrix of the preprocessed waveform data frequency bands through bandpass filtering; Each frequency band sequence in the time-frequency matrix is ​​equally divided according to time to obtain a number of time-frequency blocks and the energy of each time-frequency block in each frequency band sequence is calculated; The time-frequency matrix of the sub-frequency bands is processed based on the energy of each time-frequency block to obtain a block time-frequency spectrum matrix.

8. The distribution network fault detection and positioning system based on transfer convolutional neural network according to claim 6, characterized in that: The transfer convolutional neural network in the recognition model training unit is obtained by transfer learning using a convolutional neural network that has completed distribution network fault classification training, and specifically includes: The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training; Performing migration fine-tuning training on the convolutional neural network model, freezing the network parameters of the hidden layer of the convolutional neural network model during the training process, resetting the parameters of the fully connected layer, and not updating the network parameters of the hidden layer during back propagation, unfreezing the hidden layer after the learning of the fully connected layer is completed, and then training the convolutional neural network model as a whole; The convolutional neural network model that has undergone the migration fine-tuning training is defined as the migration convolutional neural network, and the fault line selection effect of the migration convolutional neural network is verified using distribution network fault data. The line selection accuracy is used to evaluate the line selection effect and performance of the migration convolutional neural network.

9. The distribution network fault detection and positioning system based on transfer convolutional neural network according to claim 8, characterized in that: The convolutional neural network is trained using distribution network fault samples to obtain a convolutional neural network model that completes distribution network fault classification training, including: Initialize the structure of the convolutional neural network to be trained; Preprocessing all the block time-frequency spectra, and dividing all the block time-frequency spectra samples into a training set and a test set; The network is continuously trained using the training set until the output error meets the accuracy requirement or reaches the upper limit of the iteration, and the parameters of each layer are adjusted each time the network is trained; Save the latest parameters of each layer at the end of training; The test set is marked according to the fault type and classified and identified using the trained convolutional neural network until the recognition rate reaches the set requirement, thereby obtaining the convolutional neural network model that completes the distribution network fault classification training.

10. The distribution network fault detection and positioning system based on transfer convolutional neural network according to claim 6, characterized in that: The hyperbolic window S-transform method performs waveform preprocessing on the fault waveform through a hyperbolic window function, and the hyperbolic window function is specifically: Where X is the hyperbola in (τ-t), and the left and right shapes of the curve depend on two parameters and is the falling edge parameter of the function, is the rising edge parameter of the function, λ HY is the positive curvature parameter, and f represents the frequency.

Images