Power distribution network fault adaptive spectrum analysis method based on variable time window cyclic search

Abstract

The invention discloses a power distribution network fault adaptive spectrum analysis method based on variable time window cyclic search, which comprises the following steps of: extracting a signal when a power distribution network has a fault in real time, extracting an initial time-frequency matrix A of a detection frequency band required by the fault signal, smoothing by using a filter to obtain a time-frequency matrix B, and calculating a time-frequency ridge line of the time-frequency matrix B; a variable time window parameter is initialized to obtain a time-frequency matrix n, then a variable time window cyclic search algorithm is constructed, the variable time window parameter is continuously updated to search fault signals until the time-frequency matrix n converges or reaches the maximum number of iterations, a time-frequency matrix C containing accurate frequency spectrum characteristics of the power distribution network fault signals is obtained, sliding energy analysis is carried out on the time-frequency matrix C, and the power distribution network fault signals are obtained. And the characteristic spectrum is extracted to realize spectrum characteristic analysis of the power distribution network fault signal. The method can accurately decompose the signal frequency spectrum containing the multi-frequency characteristics generated when the power distribution network fails, adaptively searches the frequency spectrum characteristics of different frequency bands of the fault signal, and achieves the frequency spectrum analysis of the power distribution network fault signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network fault processing, and particularly relates to a method for adaptive spectrum analysis of distribution network faults with variable time window cyclic search. Background Art

[0002] With the development of urban infrastructure construction, the distribution network lines are gradually changing from overhead lines to cables. Underground cables are easily affected by factors such as external forces, moisture, and chemical pollution. Coupled with possible defects in themselves, they often lead to insulation damage, and long-term live operation is prone to cause faults. There are many signal frequency bands in the distribution network, with signals ranging from direct current (0 Hz) to traveling waves (about 1 MHz). For the extraction of fault characteristics, it is necessary to adapt to the frequency components in different signals to maximize the guarantee that the spectrum decomposition in the frequency band required for fault analysis is not distorted. Therefore, establishing a spectrum feature decomposition and search method for adaptive multi-frequency characteristic signals, and extracting and analyzing their characteristics from the decomposed spectrum is of great significance for improving power quality and ensuring the safe and stable operation of the distribution system.

[0003] Currently, distribution network faults are mainly handled by analyzing the time-domain and frequency-domain characteristics of voltage and current, refining one or more electrical quantity thresholds from these signals, and implementing fault disposal based on experience or manually set thresholds. Although this method can achieve the analysis of fault signals to a certain extent, due to the uncertainty of the transient time-domain characteristics of various distribution network faults, the feasibility of actual on-site traveling wave measurement is relatively low, and signals below the traveling wave frequency band cannot be applied to the traveling wave theory. The high-frequency characteristics of distribution network fault signals can better reflect the characteristics of faults, but the existing spectrum analysis methods for high-frequency signals are difficult to adapt to the frequency characteristics of fault signals, cannot accurately decompose the characteristic spectrum of distribution network fault signals, and lack the objective authenticity of signal spectrum analysis. In recent years, with the development of artificial intelligence technology and the development of information fusion technology, the research on the analysis of distribution network fault signals has been further developed.

[0004] Currently, the problems existing in the distribution network fault handling methods mainly include: (1) The signals of some distribution network faults are weak, and the existing protection devices are difficult to determine the fault occurrence time; (2) The traveling wave method has high requirements for the sampling frequency of transformers and high costs. Although the theory is relatively mature, it is difficult to apply to the actual site; (3) The existing spectrum analysis methods have relatively fixed windows, and it is difficult to ensure the accuracy of different frequency spectrum decompositions during faults. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for adaptive spectrum analysis of distribution network faults with variable time window cyclic search, which can accurately decompose the signal spectrum containing multi-frequency characteristics generated during distribution network faults, adaptively search the spectrum characteristics of different frequency bands of fault signals, and realize the spectrum analysis of distribution network fault signals.

[0006] To achieve the above object, the present invention provides a method for adaptive spectrum analysis of distribution network faults with variable time window cyclic search, including the following steps:

[0007] S1 Extract the voltage and current signals in real time when the distribution network fails. Using the short-time Fourier transform, extract the initial time-frequency matrix A of the detection frequency band required for the fault signal, and smooth the initial time-frequency matrix A using a filter.

[0008] S2 According to the smoothed time-frequency matrix B obtained in S1, calculate its time-frequency ridge line, and initialize the variable time window parameter. Through the time-frequency ridge line data, make the variable time window translate, frequency shift and change in size according to the frequency characteristics of the signal, so as to adaptively obtain the spectrum of the fault signal and obtain the time-frequency matrix n.

[0009] S3 According to the time-frequency matrix n obtained in S2, calculate its time-frequency ridge line, construct a variable time window cyclic search algorithm, continuously update the variable time window parameter to search for the fault signal until the time-frequency matrix n converges or reaches the maximum number of iterations, and finally obtain the time-frequency matrix C containing the accurate spectrum characteristics of the distribution network fault signal.

[0010] S4 According to the time-frequency matrix C containing the spectrum characteristics of the distribution network fault signal obtained in S3, perform sliding energy analysis on it, and use the energy change of the characteristic frequency band at each time point to determine whether an abnormality occurs. If so, select an effective data window and extract its characteristic spectrum.

[0011] S5 According to the characteristic spectrum in the effective data window of the abnormal signal extracted in step S4, extract its spectrum characteristics to realize the spectrum characteristic analysis of the distribution network fault signal.

[0012] As a further solution of the present invention: the setting of the variable time window in S2:

[0013] S21 In view of the characteristic differences of the operating signals in the distribution network in multiple frequency bands, based on the traditional Gaussian window and the traditional S transform, design a variable signal analysis time window that adapts to the frequency characteristics of the signal:

[0014]

[0015] Among them, t is time, f is frequency, and f(t) is the time-frequency ridge line of the signal. is the frequency shift parameter of the window, which is divided into two items: linear frequency shift and non-linear frequency shift. Among them, f(t) is the time-frequency ridge line, and the standard deviation k and ∈ are adjustment variables, and the non-linear frequency shift term parameter

[0016] S22 Calculate the time-frequency ridge line f(t) using the smoothed time-frequency matrix B according to the variable time window in S21:

[0017]

[0018] where t is time, f is frequency, and S(t, f) is the time-frequency matrix;

[0019] Initialize the standard deviation of the variable time window parameter and the non-linear frequency shift term parameter, and perform spectral analysis on the signal:

[0020]

[0021] Finally, obtain the time-frequency matrix n, where S a (t, f) is the result of the variable time window S transform, x(t) is the distribution network fault signal, and τ is the continuous signal time shift.

[0022] As a further solution of the present invention: the variable time window cyclic search algorithm in S3:

[0023] S31 Through the first search of the signal by the variable time window, utilize the window transform to initially adaptively capture the multi-frequency characteristics of different time periods of the signal, and obtain the time-frequency ridge line of the time-frequency matrix n. Update the variable time window parameters, and then use the variable time window to search the signal again until the obtained time-frequency matrix n converges or reaches the maximum number of iterations;

[0024] S32 According to the convergence of the time-frequency matrix n in S31, the condition is:

[0025] ||vec(|S n+1 (t, f)|)-vec(|S n (t, f)|)||2 < α

[0026] where vec(X) is the matrix vectorization operation, which expands the matrix into a column vector by columns, S(t, f) is the time-frequency matrix, |S n (t, f)| is the amplitude of the time-frequency matrix, ||X||2 is the Euclidean norm of the vector, and α is the convergence threshold.

[0027] As a further solution of the present invention: the effective data window selection method in S4:

[0028] Consider the time-frequency characteristics of the characteristic frequency band, perform sliding energy analysis on the time-frequency matrix C containing the spectral characteristics of the distribution network fault signal, which is expressed by the following formula:

[0029]

[0030] In the formula, a and b are respectively the upper and lower limits of the detected frequency band; |S a (ti , f)| 2 is the time-frequency energy density at time t i and describes the instantaneous power distribution of the signal at frequency f at time point t i ;

[0031] Use double-threshold mutation detection to locate the effective window: high threshold T h : Mark the starting point of the energy sudden increase (E total (t i ) > 3σ), and intercept half a cycle forward from the energy mutation point; low threshold T l : Extend the window until the energy drops to T l (E total (t i ) < 1.2σ), and intercept half a cycle backward from the energy drop point; where σ is the energy constant for normal system operation; select the correct fault occurrence time through this method and extract the characteristic spectrum.

[0032] As a further solution of the present invention: The spectrum characteristics of the distribution network fault signal in S5 include but are not limited to:

[0033] Frequency band energy ratio FER: Divide the characteristic spectrum into multiple frequency bands and calculate the energy proportion of each frequency band;

[0034]

[0035] where m and n are the upper and lower limits of the divided frequency bands respectively, and a and b are the upper and lower limits of the frequency band for extracting the characteristic spectrum; |S a (t, f)| 2 is the time-frequency energy density, which describes the instantaneous power distribution at frequency f as the signal changes with time t;

[0036] Spectrum roll-off point SR: Calculate the frequency corresponding to when the cumulative energy in the signal spectrum reaches a specified ratio, which is used to quantify the high-frequency energy attenuation characteristics of early faults;

[0037]

[0038] where f s is the spectrum roll-off point at time t i , and β is the spectrum roll-off coefficient;

[0039] Spectrum entropy SE: Used to quantify the complexity of the energy distribution of the signal spectrum. The higher the value, the more dispersed the energy distribution; the lower the value, the more concentrated the energy in a few frequencies;

[0040]

[0041] where f jis a frequency point in the spectrum, f j ∈(a, b), and f j = j; p(t, f j ) is the normalized energy probability at frequency f j , that is, the proportion of the energy at frequency f j in the total energy, and is used to calculate the spectrum entropy;

[0042] Spectral kurtosis SK: Describes the kurtosis of the spectral energy distribution, that is, the degree of concentration of the spectral energy near the spectral centroid;

[0043]

[0044] where f j is a frequency point in the spectrum, f j ∈(a, b), and f j = j; μ is the spectral centroid, and σ 2 is the variance of the spectral energy; |S a (t, f j )| is the amplitude of the time-frequency vector.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] It can effectively realize the spectral analysis of the distribution network containing multi-frequency characteristic signals. By using the variable time-window cyclic search method, it can adaptively search for the spectral characteristics of the fault signal and perform spectral analysis, with good universality, reliability and objectivity, and can be widely applied to various engineering practices. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is the flow chart of the present invention.

[0048] Figure 2 is the distribution network topology diagram built in PSCAD / EMTDC.

[0049] Figure 3 is the flow chart of the variable time-window cyclic search algorithm.

[0050] Figure 4 is the time-frequency diagram of the harmonic signal obtained by Fourier transform, traditional S transform and variable time-window adaptive S transform.

[0051] Figure 5 is the time-frequency diagram of the frequency-varying signal obtained by Fourier transform, traditional S transform and variable time-window adaptive S transform.

[0052] Figure 6 is the spectral characteristics of the arc fault extracted by the variable time-window adaptive S transform. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The present invention will be further described below through embodiments.

[0054] As Figure 1 shown, a method for adaptive spectrum analysis of distribution network faults with variable time window cyclic search includes the following steps:

[0055] (1) Extract the voltage and current signals in real time when the distribution network fails. Using the short-time Fourier transform, extract the initial time-frequency matrix A of the required detection frequency band of the fault signal, and use the Savitzky-Golay filter to smooth the initial time-frequency matrix A.

[0056] (2) According to the smoothed initial time-frequency matrix B obtained in step (1), calculate its time-frequency ridge line f(t):

[0057]

[0058] where t is time, f is frequency, and S(t,f) is the time-frequency matrix.

[0059] (3) Based on the characteristic differences of the operating signals in multiple frequency bands in the distribution network, design a variable signal analysis time window that adapts to the signal frequency characteristics based on the traditional Gaussian window and the traditional S transform:

[0060]

[0061] where t is time, f is frequency, f(t) is the time-frequency ridge line of the signal, is the frequency shift parameter of the window, divided into two items: linear frequency shift and non-linear frequency shift. Among them, f(t) is the time-frequency ridge line, and the standard deviation k and ∈ are adjustment variables, and the non-linear frequency shift term parameter

[0062] (4) Initialize the standard deviation and non-linear frequency shift term parameter of the variable time window in step (3) according to the time-frequency ridge line obtained in step (2), so that the variable time window adapts to the frequency characteristics of the signal for translation, frequency shift, and size change, and adaptively performs spectrum analysis on the fault signal x(t):

[0063]

[0064] Finally, obtain the time-frequency matrix n and calculate its time-frequency ridge line. Among them, S a (t,f) is the result of the variable time window S transform, x(t) is the distribution network fault signal, and τ is the continuous signal time shift.

[0065] (5) Based on the time-frequency ridge line of the time-frequency matrix n obtained in step (4), fine-correct and update the variable time window parameters in step (3), and repeat step (4) until the obtained time-frequency matrix n converges or reaches the maximum number of iterations, realizing the variable time window cyclic search. The process is as shown in Figure 3 . Finally, obtain the time-frequency matrix C containing the accurate spectral characteristics of the distribution network fault signal. The condition for matrix convergence is:

[0066] ||vec(|S n+1 (t,f)|)-vec(|S n (t,f)|)||2<α

[0067] where vec(X) is the matrix vectorization operation, which expands the matrix by columns into a column vector, S(t,f) is the time-frequency matrix, |S n (t,f)| is the amplitude of the time-frequency matrix, ||X||2 is the Euclidean norm of the vector, and α is the convergence threshold.

[0068] (6) According to the time-frequency matrix C containing the accurate spectral characteristics of the distribution network fault signal obtained in step (5), considering the time-frequency characteristics of the characteristic frequency band, perform sliding energy analysis on this time-frequency matrix, which is expressed by the following formula:

[0069]

[0070] In the formula, a and b are respectively the upper and lower limits of the detection frequency band of the variable time window adaptive S transform; |S a (t i ,f)| 2 is the time-frequency energy density at time t i , describing the instantaneous power distribution of the signal at frequency f at time point t i .

[0071] Use double-threshold mutation detection to locate the effective window: high threshold T h : Mark the starting point of the sudden increase in energy (E total (t i )>3σ), and intercept half a cycle forward from the energy mutation point; low threshold T l : Extend the window until the energy drops to T l (E total (t i )<1.2σ), and intercept half a cycle backward from the energy drop point; where σ is the energy constant of the system under normal operation; select the correct fault occurrence time through this method and extract the characteristic spectrum.

[0072] (7) According to the effective characteristic spectrum extracted in step (6), the extracted characteristics include but are not limited to:

[0073] Frequency Band Energy Ratio (FER): The characteristic spectrum is divided into multiple frequency bands (such as low frequency, medium frequency, and high frequency), and the energy proportion of each frequency band is calculated.

[0074]

[0075] where m and n are the upper and lower limits of the divided frequency band respectively, and a and b are the upper and lower limits of the frequency band for extracting the characteristic spectrum; |S a (t,f)| 2 is the time-frequency energy density, which describes the instantaneous power distribution at frequency f when the signal changes with time t.

[0076] Spectrum Rolloff Point (SR): Calculate the frequency corresponding to when the cumulative energy in the signal spectrum reaches a specified proportion, which is used to quantify the high-frequency energy attenuation characteristic of early faults.

[0077]

[0078] where f s is the spectrum rolloff point at time t i , and β is the spectrum rolloff coefficient.

[0079] Spectrum Entropy (SE): It is used to quantify the complexity of the energy distribution of the signal spectrum. The higher the value, the more dispersed the energy distribution; the lower the value, the more concentrated the energy in a few frequencies;

[0080]

[0081] where f j is a frequency point in the spectrum, f j ∈(a,b), and f j =j; p(t,f j ) is the normalized energy probability at frequency f j , that is, the proportion of the energy at frequency f j in the total energy, which is used to calculate the spectrum entropy.

[0082] Spectrum Kurtosis (SK): It describes the kurtosis of the spectrum energy distribution, that is, the degree to which the spectrum energy is concentrated near the spectrum centroid;

[0083]

[0084] where f j is a frequency point in the spectrum, f j ∈(a,b), and f j =j; μ is the spectrum centroid, σ 2 is the variance of the spectrum energy; |S a (t,f j )| is the amplitude of the time-frequency vector.

[0085] Finally, the spectrum analysis of the distribution network fault signal is realized.

[0086] Simulation verification:

[0087] In order to verify the effectiveness of the variable time-window cyclic search, signals containing multiple frequency harmonics are generated in MATLAB. The signals generated by the system are set to have a known frequency distribution. Through the spectrum decomposition of the known signals, two groups of spectrum decomposition comparison experiments are carried out on the method proposed in the present invention and the traditional time-frequency analysis method to verify the accuracy of the method.

[0088] 1. Three common distribution network harmonics are set respectively: 500 Hz low-frequency stable harmonic, 5000 Hz medium-frequency pulse harmonic, and 25 kHz high-frequency decaying harmonic. Let the three harmonics exist simultaneously. The sampling frequency is 60 kHz, the signal duration is 0.5 s, the low-frequency harmonic exists all the time, the medium-frequency harmonic signal pulses every 0.1 s, and the pulse duration is 0.01 s. The high-frequency harmonic appears from 0.05 s and decays exponentially. Compare the spectrum decomposition of different methods.

[0089] 2. Set a comparative experiment on the spectrum decomposition of frequency-variable signals. The signal is set as sin(400π*t*(1 + t)), the sampling frequency is 5000 Hz, and compare and analyze the effects of different spectrum decomposition methods. The spectrum decomposition of different harmonic scenarios and frequency-variable signals is compared as Figure 4 and Figure 5 shown.

[0090] From the spectrum decomposition results, it can be obtained that Figure 4 Among the three signal spectrum analysis methods, the decomposition on the time scale can be realized, and the occurrence and end moments corresponding to each harmonic can be accurately expressed. However, for the Fourier transform, the spectrum decomposition below 500 Hz is disordered, and due to its fixed time-window setting, the decomposition amplitude of high-frequency signals is low, and the signal spectrum cannot be accurately decomposed. For the S transform, the central frequency of its high-frequency harmonics is roughly accurate, but its frequency band radius is large. For the main spectrum radius of high-frequency decaying signals, it reaches 5 kHz, and the decomposed spectrum cannot accurately express the frequency components of the given signal. For the variable time-window adaptive S transform, due to the variable nature of its time window, the harmonic signals of the three frequency components can be accurately decomposed, and the decomposed frequency band radius is small. For the signal spectrum of 500 Hz, its radius is within 15 Hz, and for the main spectrum radius of high-frequency decaying signals, it is only within 1 kHz, which can more accurately describe the spectrum characteristics. Figure 5The frequency band radii of both the Fourier transform and the S transform are relatively large. The spectral radius decomposed by the Fourier transform is greater than 300 Hz, and the minimum spectral radius decomposed by the S transform is 100 Hz, and there is a phenomenon of frequency band diffusion, which makes the spectral decomposition result distorted. For the variable-time window adaptive S transform, the decomposed frequency band radius is smaller, and the decomposed spectral radius is limited within 25 Hz. Although there is also a diffusion phenomenon, the variable-time window parameter is affected by the time-frequency ridge line, reducing the spectral diffusion and enabling a more accurate decomposition of the signal spectrum.

[0091] To verify the effectiveness of this method for the spectral analysis and feature extraction of the distribution network, as Figure 2 shown, a distribution network simulation model was built in PSCAD / EMTDC. Taking line L10 as an example, considering a single-phase arc grounding fault with a relatively complex harmonic component, a single-phase arc grounding fault was set in line L10. In the fault scenario using the Cassie arc model, the simulation running time was set to 0.1 s, the fault phase was phase A, the fault occurrence time was 0.06 s, the duration was 0.025 s, and the sampling frequency was 40 kHz to analyze the zero-sequence current signal generated by the arc fault. Since the proportion of low-frequency signals is still relatively high when the arc fault occurs, resulting in a relatively low proportion of high-frequency bands in time-frequency analysis, when decomposing the waveform by the fast Fourier transform, the high-frequency attenuation characteristic frequency band of the arc fault was selected for analysis, that is, signals above 5 kHz were analyzed, and the low-frequency band (5k - 8 kHz), medium-frequency band (8k - 12 kHz), and high-frequency band (12k - 20 kHz) were divided. The spectral analysis result and feature calculation result of the fault window are as Figure 6 shown.

[0092] From the simulation results, it can be seen that when the arc fault occurs, using the spectral decomposition method proposed in the present invention can well show the high-frequency attenuation characteristics of different frequency bands of the fault and accurately select the fault window. The spectral feature waveform was extracted from the decomposed fault spectrum, realizing the adaptive spectral analysis of the distribution network fault with variable-time window cyclic search. The experiment shows that the method proposed in the present invention has certain advantages in the field of distribution network fault signal analysis, and the spectral analysis result can be further applied to the fault identification and fault location of the distribution network, having a good application prospect.

Claims

1. An adaptive spectrum analysis method for distribution network faults with variable time-window cyclic search, characterized in that, It includes the following steps: S1 Extract the voltage and current signals in real time when the distribution network fails. Using the short-time Fourier transform, extract the initial time-frequency matrix A of the detection frequency band required for the fault signal, and smooth the initial time-frequency matrix A using a filter; S2 According to the smoothed time-frequency matrix B obtained in S1, calculate its time-frequency ridge line, and initialize the variable time window parameters. Through the time-frequency ridge line data, make the variable time window translate, frequency shift, and change in size according to the frequency characteristics of the adaptive signal, so as to adaptively obtain the fault signal spectrum and obtain the time-frequency matrix n; S3 According to the time-frequency matrix n obtained in S2, calculate its time-frequency ridge line, construct a variable time window cyclic search algorithm, continuously update the variable time window parameters to search for the fault signal until the time-frequency matrix n converges or reaches the maximum number of iterations, and finally obtain the time-frequency matrix C containing the accurate spectrum characteristics of the distribution network fault signal; S4 According to the time-frequency matrix C containing the spectrum characteristics of the distribution network fault signal obtained in S3, perform sliding energy analysis on it, and use the energy change of the characteristic frequency band at each time point to determine whether an abnormality occurs. If so, select an effective data window and extract its characteristic spectrum; S5 According to the characteristic spectrum in the effective data window of the abnormal signal extracted in step S4, extract its spectrum characteristics to realize the spectrum characteristic analysis of the distribution network fault signal.

2. The adaptive spectrum analysis method for distribution network faults with variable time window cyclic search according to claim 1, wherein The setting of the variable time window in S2: S21 In view of the characteristic differences of the operating signals in multiple frequency bands in the distribution network, based on the traditional Gaussian window and the traditional S transform, design a variable signal analysis time window that adapts to the frequency characteristics of the signal: where t is time, f is frequency, and f(t) is the time-frequency ridge line of the signal. is the frequency shift parameter of the window, which is divided into two terms: linear frequency shift and non-linear frequency shift. Among them, f(t) is the time-frequency ridge line, and the standard deviation k and ∈ are adjustment variables, and the non-linear frequency shift term parameter S22 According to the variable time window in S21, use the smoothed time-frequency matrix B to calculate its time-frequency ridge line f(t): Among them, t is time, f is frequency, and S(t,f) is the time-frequency matrix; Initialize the standard deviation of the variable time window parameter and the non-linear frequency shift term parameter, and perform spectrum analysis on the signal: Finally, the time-frequency matrix n is obtained, where S a (t, f) is the result of the variable-time window S transform, x(t) is the distribution network fault signal, and τ is the time shift of the continuous signal.

3. The adaptive spectrum analysis method for distribution network faults with variable time window cyclic search according to claim 1, characterized in that, The variable time window cyclic search algorithm in S3: S31 Through the first search of the signal by the variable time window, use the time window transformation to initially adapt to the multi-frequency characteristics of different time periods of the signal, and obtain the time-frequency ridge line of the time-frequency matrix n. Update the variable time window parameters, and use the variable time window to search for the signal again until the obtained time-frequency matrix n converges or reaches the maximum number of iterations; S32 According to the convergence of the time-frequency matrix n in S31, its condition is: ||vec(|S n+1 (t,f)|)-vec(|S n (t,f)|)||2 < α Among them, vec(X) is the matrix vectorization operation that unfolds the matrix by columns into a column vector, S(t,f) is the time-frequency matrix, |S n (t,f)| is the amplitude of the time-frequency matrix, ||X||2 is the Euclidean norm of the vector, and α is the convergence threshold.

4. A method for adaptive spectrum analysis of distribution network faults with variable time window cyclic search according to claim 1, characterized in that, The method for selecting the effective data window in S4: Considering the time-frequency characteristics of the characteristic frequency band, perform sliding energy analysis on the time-frequency matrix C containing the spectrum characteristics of the distribution network fault signal. The energy calculation method is expressed by the following formula: where a and b are the upper and lower limits of the detected frequency band respectively; |S a (t i , f)| 2 is the time-frequency energy density at time t i , describing the instantaneous power distribution of the signal at frequency f at time point t i . Using double-threshold mutation detection to locate the effective window: high threshold T h : Mark the starting point of the energy sudden increase (E total (t i >) > 3σ), and intercept half a cycle forward from the energy mutation point; low threshold T l : Extend the window until the energy drops to T l (E total (t i ) < 1.2σ), and intercept half a cycle backward from the energy drop point; where σ is the energy constant of the system under normal operation; select the correct fault occurrence time through this method and extract the characteristic spectrum.

5. The self-adaptive spectrum analysis method for distribution network faults with variable time window cyclic search according to claim 1, wherein The spectrum characteristics of the distribution network fault signal in S5 include but are not limited to: Frequency band energy ratio FER: Divide the characteristic spectrum into multiple frequency bands and calculate the energy proportion of each frequency band; where m and n are the upper and lower limits of the divided frequency bands respectively, and a and b are the upper and lower limits of the frequency bands for extracting the characteristic frequency spectrum; |S a (t,f)| 2 is the time-frequency energy density, which describes the instantaneous power distribution at frequency f when the signal changes with time t; Spectrum roll-off point SR: Calculate the frequency corresponding to when the cumulative energy in the signal spectrum reaches a specified ratio, which is used to quantify the high-frequency energy attenuation characteristics of early faults; where f s is the spectral roll-off point at time t i , and β is the spectral roll-off coefficient; Spectrum entropy SE: Used to quantify the complexity of the energy distribution of the signal spectrum. The higher its value, the more dispersed the energy distribution; the lower the value, the more concentrated the energy in a few frequencies; where f j is a frequency point in the spectrum, f j ∈(a, b), and f j = j; p(t, f j ) is the normalized energy probability at frequency f j , that is, the proportion of the energy at frequency f j in the total energy, and is used to calculate the spectral entropy; Spectral kurtosis SK: Describes the kurtosis of the spectral energy distribution, that is, the degree of concentration of spectral energy near the spectral centroid; where, f j is a frequency point in the spectrum, f j ∈(a, b), and f j = j; μ is the spectrum centroid, σ 2 is the variance of the spectrum energy; |S a (t, f j )| is the amplitude of the time-frequency vector.

Images