Partial discharge waveform comprehensive identification system
By working together with signal reconstruction, frequency fitting, and verification and identification modules, the problems of difficulty in cross-site comparison of partial discharge waveforms in GIS and noise interference are solved, and accurate identification and reliable operation of partial discharge waveforms are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING TRANSMISSION LIANPU TECH CO LTD
- Filing Date
- 2026-04-10
- Publication Date
- 2026-05-29
AI Technical Summary
In multi-site monitoring scenarios of gas-insulated metal-enclosed switchgear (GIS), partial discharge waveform identification suffers from difficulties in cross-site comparison and is susceptible to random interference, leading to distorted identification results.
The system employs a signal reconstruction module for empirical mode decomposition and Hilbert transform, combined with a frequency fitting module for linear fitting and normalization, and uses compressed sensing reconstruction and high-pass filtering to remove noise. Finally, a verification and recognition module is used for statistical significance verification to ensure the accuracy of the recognition results.
It effectively solves the difficulty of identifying partial discharge waveforms across sites, improves the accuracy of identification, eliminates noise interference, and ensures the reliability and consistency of identification results.
Smart Images
Figure CN122109804A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of partial discharge detection technology, and more specifically to a comprehensive partial discharge waveform recognition system. Background Technology
[0002] In the operation and maintenance of gas-insulated metal-enclosed switchgear (GIS), partial discharge detection is a crucial means to ensure the reliability of equipment insulation. GIS equipment widely contains T-shaped branching structures. When a partial discharge occurs, the ultra-high frequency electromagnetic waves generated by the discharge propagate in the tangential and straight-through branches of the T-shaped branch. Due to the significant near-cutoff dispersion effect of the T-shaped branch on the propagation of electromagnetic waves, the partial discharge signals collected from the tangential and straight-through branches exhibit distinct differences in their instantaneous frequency evolution: the electromagnetic waves in the tangential branch attenuate more rapidly due to path geometry constraints, resulting in a high-frequency arrival followed by a monotonically decreasing instantaneous frequency. In contrast, the dispersion effect in the straight-through branch is relatively weaker, and the instantaneous frequency change tends to be more gradual. This is a typical characteristic of partial discharge signals at the T-shaped branch of GIS.
[0003] However, in actual GIS multi-site monitoring scenarios, the following core technical problems exist: Due to differences in geometric structure, the cutoff frequencies of the equivalent waveguides at different GIS sites' T-shaped branches vary. This results in different dimensional benchmarks for the instantaneous frequency slopes extracted directly from the tangential and straight branches, making direct comparative analysis across sites impossible. Simultaneously, the signal of a single partial discharge pulse is susceptible to random interference such as acquisition noise and subtle fluctuations in the pulse itself. If waveform identification relies solely on the characteristics of a single pulse, accidental errors can easily distort the identification results, making it difficult to accurately determine whether the partial discharge waveform at the T-shaped branch conforms to its inherent physical laws. Therefore, how to accurately identify the partial discharge waveform of GIS has become a critical problem that urgently needs to be solved. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention proposes a comprehensive partial discharge waveform recognition system, which solves the problem of how to accurately identify partial discharge waveforms in GIS.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a partial discharge waveform comprehensive identification system, comprising:
[0006] The signal reconstruction module decomposes the first-reach window waveform into intrinsic mode functions based on empirical mode decomposition, selects the intrinsic mode functions for signal reconstruction to obtain the reconstructed signal, performs Hilbert transform on the reconstructed signal to obtain the analytical signal, and calculates the instantaneous frequency sequence based on the phase of the analytical signal. The first-reach window is a window that contains the early high-frequency components of the partial discharge pulse.
[0007] The frequency fitting module obtains the instantaneous frequency sequences of the tangential branch and the through branch, performs linear fitting on the instantaneous frequency sequences of the tangential branch and the through branch respectively, sets fitting weights, and obtains the slope and intercept by minimizing the sum of squared residuals after weighting. It calculates the fitting residuals and determines whether the linear fitting is effective based on the fitting residuals. If effective, it obtains the slope of the tangential branch and the slope of the through branch.
[0008] The normalization module calculates the normalized slopes of the tangential branch slope and the through branch slope based on the TE mode cutoff frequency, respectively, to obtain the normalized slopes of the tangential branch and the through branch. The difference between the normalized slope of the tangential branch and the normalized slope of the through branch is marked as the normalized feature.
[0009] Preferably, the data acquisition module acquires the original partial discharge signal and the TE mode cutoff frequency, where the TE mode cutoff frequency is the cutoff frequency of the equivalent waveguide at the T-type branch of the gas-insulated metal-enclosed switchgear.
[0010] The time-frequency energy spectrum is obtained by performing a short-time Fourier transform on the original partial discharge signal. The window function of the short-time Fourier transform is the Hanning window. The time-frequency energy spectrum includes the sampling time and signal frequency of the original partial discharge signal.
[0011] The time interval of the first arrival window is determined by integrating the time-frequency energy spectrum.
[0012] The original partial discharge signal within the first arrival window time interval is obtained by extracting the original signal segment using a rectangular window function. The first arrival window signal is obtained by high-pass filtering the original signal segment. The cutoff frequency of the high-pass filter is 1.2 times the cutoff frequency of the TE mode.
[0013] The first-window waveform is obtained after reconstructing the first-window signal.
[0014] Preferably, the signal start time range and signal frequency range of the original partial discharge signal are obtained, and the time-frequency energy spectrum is integrated over the signal start time range and signal frequency range to obtain the total signal energy.
[0015] Using the signal start time range as the time axis, perform a sliding integral along the time axis on the signal frequency of the time-frequency energy spectrum to calculate the cumulative energy. Record the moment when the cumulative energy is ≥ 0.2 × the total signal energy, and mark this moment as the start of the first arrival window.
[0016] Taking the first arrival window as the first starting moment, record along the time axis the moment when the signal frequency of the time-frequency energy spectrum reaches the maximum signal frequency in the range of signal frequency. Mark this moment as the relay moment. The signal frequency at the relay moment is marked as the signal peak value.
[0017] Using the relay time as the second starting time, record along the time axis the time when the time-frequency energy spectrum is within the range of signal frequency ≤ 0.5 × signal peak value, and mark this time as the end of the first arrival window;
[0018] The time interval of the first arrival window is obtained by combining the starting point and the ending point of the first arrival window.
[0019] Preferably, the first-reaching window waveform is obtained by reconstructing and solving the first-reaching window signal based on the orthogonal matching pursuit algorithm in compressed sensing.
[0020] Preferably, the first-reaching window waveform is decomposed into n intrinsic mode functions (IMFs) based on empirical mode decomposition. i (t), where the value of i ranges from [1, n];
[0021] Select the IMF for i=[1,3]. i (t) performs signal reconstruction, specifically as follows:
[0022] IMF of i=[1,3] i The reconstructed signal is obtained by summing the t values.
[0023] Preferably, the reconstructed signal is obtained by performing a Hilbert transform on the reconstructed signal and then adding the reconstructed signal back to it.
[0024] Let the phase of the analytic signal be φ(t). Dividing 1 by 2π and then multiplying by the derivative of the phase φ(t) with respect to time t yields the instantaneous frequency sequence f. i (t), where i is the branch identifier, i=T represents the instantaneous frequency sequence of the tangential branch, and i=F represents the instantaneous frequency sequence of the through branch.
[0025] Preferably, the instantaneous frequency sequences of the tangential branch and the through branch are linearly fitted based on the fitting formula, which is: instantaneous frequency sequence f i (t) = slope s i × time t + intercept b i Where i is the branch identifier, the fitting process is as follows:
[0026] For the instantaneous frequency sequence f i For each time t of (t), the square of the instantaneous frequency amplitude corresponding to time t is taken as the energy density. The weight of time t is obtained by dividing the energy density of time t by the maximum energy density in the instantaneous frequency sequence. This weight is set as the fitting weight.
[0027] The fitting weights are used as the weights for the weighted least squares method, and the optimal slope s is obtained by solving the weighted least squares method. i and optimal intercept b i The specific solution process is as follows:
[0028] The instantaneous frequency sequence f i The instantaneous frequency at each time t in (t) is calculated by substituting the preset slope s into the fitting formula. i and intercept b i The difference in instantaneous frequencies calculated later is multiplied by the fitting weight at the corresponding time t, and then the square root is taken. The sum of these values is then calculated, and the optimal slope s is obtained by solving the summation. i and optimal intercept b i .
[0029] Preferably, the fitting residual is calculated based on the fitting formula and weights. The specific calculation process is as follows:
[0030] Calculate the instantaneous frequency and substitute it into the optimal slope s i and intercept b i The difference between the calculated values of the fitting formula is obtained by multiplying the squares of the differences at all times by the fitting weights at the corresponding times, summing the results, dividing the sum by the sum of the fitting weights at all times, and taking the square root of the quotient.
[0031] If the residual is less than the slope s i If the absolute value of the linear fit is multiplied by 0.05, the linear fit is considered valid; otherwise, it is considered invalid data.
[0032] When the linear fit is deemed valid, the tangential branch slope s is obtained. T and the slope s of the straight-through branch F .
[0033] Preferably, the tangential branch slope s T Dividing by the TE mode cutoff frequency yields the normalized slope of the tangential branch;
[0034] The slope s of the straight-through branch F Dividing by the TE mode cutoff frequency yields the normalized slope of the through branch;
[0035] The difference between the normalized slope of the tangential branch and the normalized slope of the through branch is marked as the normalized feature. The normalized feature is a dimensionless feature that reflects the relative contrast between the instantaneous frequency decline rates of the tangential and through branches.
[0036] The normalized features are statistically significant by using a verification and identification module.
[0037] Preferably, the verification and recognition module obtains the mean and standard deviation of the normalized features, determines whether the current normalized feature is negative, and if it is negative, defines the feature boolean value as 1, otherwise it is 0;
[0038] Determine if the mean of the normalized feature is negative; if it is negative, define the Boolean value of the mean as 1, otherwise as 0.
[0039] Determine whether the standard deviation of the normalized feature is less than 0.2 times the absolute value of the mean of the normalized feature. If it is less, define the standard deviation Boolean value as 1; otherwise, it is 0.
[0040] If the Boolean values for the feature, mean, and standard deviation are all 1, the statistical significance test result is considered to be passed; otherwise, it is considered to be failed.
[0041] If the statistical significance verification result is passed, the original partial discharge signal is determined to be the target waveform; otherwise, it is determined to be a non-target waveform.
[0042] Compared with existing technologies, it has the following advantages:
[0043] This solution proposes a comprehensive partial discharge waveform recognition system. Through multi-module collaborative operation, it effectively solves the problems of difficult cross-site comparison and susceptibility to random interference in partial discharge waveform recognition of gas-insulated metal-enclosed switchgear, significantly improving recognition accuracy. The data acquisition module dynamically determines the first-arrival window time interval through time-frequency energy integration. Combined with high-pass filtering at 1.2 times the TE mode cutoff frequency and compressed sensing reconstruction, it accurately preserves the early high-frequency components of the partial discharge pulse while filtering out low-frequency interference and noise from multipath reflections. This provides a high signal-to-noise ratio first-arrival window waveform for subsequent analysis, solving the problem of impure extraction of effective components from the original signal. The signal reconstruction module decomposes the first-arrival window waveform into intrinsic mode functions based on empirical mode decomposition. It selects the first three high-frequency components to reconstruct the signal, further stripping away noise and low-frequency redundancy. Then, it obtains the analytic signal through Hilbert transform and calculates the instantaneous frequency sequence, ensuring that the frequency evolution characteristics truly reflect the physical laws of dispersion and avoiding the frequency jump problem caused by directly analyzing noisy signals. The frequency fitting module sets the fitting weights based on the square of the instantaneous frequency amplitude as the energy density. It solves for the optimal slope and intercept using the weighted least squares method, and judges the fitting validity by combining the residual (less than 5% of the absolute value of the slope). This makes the slope more closely match the evolution trend of the high-frequency energy-dominated period, filters invalid data, and improves the reliability of the slope parameters. The normalization module divides the tangential and straight-through branch slopes by the TE mode cutoff frequency respectively, eliminating the dimensional differences in the reference values caused by geometric differences between different sites. The resulting normalized feature (the difference between the normalized slopes of the tangential and straight-through branches) is a dimensionless quantity, enabling cross-site comparison. The verification and identification module uses the mean, standard deviation, and sign of multiple samples for joint verification to ensure the statistical significance of the features, effectively filtering random interference from single pulses, and ultimately achieving accurate identification of the target waveform. This provides key technical support for the reliable operation of gas-insulated metal-enclosed switchgear. Attached Figure Description
[0044] Figure 1This is a schematic diagram of the system framework of the present invention; Detailed Implementation
[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0046] Please see Figure 1 This application provides a partial discharge waveform comprehensive identification system, including a data acquisition module, a signal reconstruction module, a frequency fitting module, a normalization module and a verification and identification module;
[0047] The data acquisition module acquires the original partial discharge signal x(t) and the TE mode cutoff frequency f of the gas-insulated metal-enclosed switchgear. c Specifically, the TE mode cutoff frequency is the cutoff frequency of the equivalent waveguide at the T-junction of a gas-insulated, metal-enclosed switchgear. It is determined by the geometry of the equipment, and its range needs to be determined through actual measurement or simulation. A typical range is 500MHz ≤ f c ≤1500MHz, the T-type branch is a common geometric structure in gas-insulated metal-enclosed switchgear, which allows the circuit to branch from a main path into two branches, similar to the shape of the letter T;
[0048] The time-frequency energy spectrum E(t,f) is obtained by performing a short-time Fourier transform on the original partial discharge signal x(t), where t is the sampling time of the original partial discharge signal and f is the signal frequency of the original partial discharge signal. The window function of the short-time Fourier transform is a Hanning window, and the window length L of the Hanning window satisfies L×f. s >1÷f c , where f s The sampling frequency of the original partial discharge signal;
[0049] The signal start time range and signal frequency range of the original partial discharge signal are obtained. The time-frequency energy spectrum E(t,f) is integrated over the signal start time range and signal frequency range to obtain the total signal energy E. total Specifically, the signal start time range is the time range from the sampling time of the first original partial discharge signal to the sampling time of the last original partial discharge signal when sampling the original partial discharge signal at the sampling frequency. The signal frequency range is the range from the minimum frequency to the maximum frequency of the original partial discharge signal collected within this time range.
[0050] Using the signal's initial time range as the time axis, the cumulative energy E is calculated by performing a sliding integral along the time axis with respect to the signal frequency of the time-frequency energy spectrum E(t,f). cum Record the cumulative energy E cum ≥0.2E total The moment of arrival is marked as the start of the first arrival window. Specifically, the first arrival window is a window that includes the early high-frequency components of the partial discharge pulse.
[0051] Taking the first arrival window as the first starting time, record along the time axis the moment when the time-frequency energy spectrum E(t,f) reaches the maximum signal frequency in the signal frequency range, mark this moment as the relay moment, and mark the signal frequency at the relay moment as the signal peak value;
[0052] Taking the relay time as the second starting time, record along the time axis the time when the time-frequency energy spectrum E(t,f) is within the range of signal frequency ≤ 0.5 × signal peak value. Mark this time as the end of the first arrival window. Combine the start and end of the first arrival window to obtain the time interval of the first arrival window.
[0053] Specifically, by dynamically determining the first-arrival window through the dual thresholds of time-frequency energy accumulation and attenuation, it can adaptively capture the early stage of high-frequency arrival in partial discharge pulses, avoid interference from low-frequency subsequent pulses of multipath reflection, and ensure that the first-arrival window includes the early high-frequency components of partial discharge.
[0054] The original signal segment is obtained by extracting the original partial discharge signal within the first arrival window time interval using a rectangular window function. Specifically, x(t) × rect = original signal segment, where rect is the rectangular window function. The original signal segment is then high-pass filtered to obtain the filtered signal, which is marked as the first arrival window signal. The cutoff frequency of the high-pass filter is the TE mode cutoff frequency f. c 1.2 times that of the first-reach window signal. Specifically, the first-reach window signal is the signal after filtering the original partial discharge signal within the first-reach window. The high-frequency component of the first-reach window signal is the core of reflecting the dispersion characteristics. The frequency of the multipath reflection wave is usually lower than that of the first-reach window signal. High-pass filtering can retain the high-frequency dominant component of the first-reach window signal.
[0055] Since the first-window signal is sparsity in the frequency domain, with the main energy concentrated in a few high-frequency components, the first-window waveform is obtained by reconstructing the first-window signal based on the orthogonal matching pursuit algorithm in compressed sensing.
[0056] Specifically, considering the high-frequency dominance and frequency-domain sparsity of partial discharge signals in gas-insulated metal-enclosed switchgear, high-pass filtering is used to initially filter multipath low-frequency interference from the first-line window signal, and compressed sensing sparse reconstruction further suppresses residual noise and redundant components, making the output first-line window signal purer. This results in a purified first-line window waveform, which more accurately reflects the true electromagnetic propagation characteristics of the early stage of the partial discharge pulse, providing high signal-to-noise ratio basic data for subsequent instantaneous frequency extraction.
[0057] The signal reconstruction module decomposes the first-reaching window waveform into n intrinsic mode functions (IMFs) based on empirical mode decomposition (EMD), denoted as IMFs. i (t), where the value of i ranges from [1, n]. Specifically, n is the total number of IMFs obtained by decomposition, which is determined by the time-frequency characteristics of the signal itself.
[0058] Since the high-frequency components of the first arrival window waveform of partial discharge in gas-insulated metal-enclosed switchgear are the core of reflecting the high-frequency first arrival dispersion characteristics, the intrinsic mode functions of the first three high frequencies are selected, i.e., the IMF of i=[1,3]. i (t) performs signal reconstruction, and the remaining IMF i (t) Removal: Specifically, the adaptive screening process of Empirical Mode Decomposition (EMD) will first extract the highest frequency oscillation component in the signal to generate the first IMF. The remaining signals will be screened again to extract the second highest frequency component to generate the second IMF, and so on. Therefore, in this example, the first three IMFs are selected for signal reconstruction. These IMFs contain the dominant high-frequency energy of the first-window waveform, while the remaining IMFs are mostly low-frequency IMFs containing noise, so they are removed.
[0059] IMF of i=[1,3] i The reconstructed signal x is obtained by summing the results (t). recon (t), Specifically, although the first-reach window waveform has undergone preliminary denoising, it may still contain residual noise or irrelevant low-frequency components. By selecting key high-frequency IMF components for superposition and reconstruction, noise and irrelevant low-frequency interference can be further removed, and the high-frequency dominant components reflecting the dispersion characteristics in the first-reach wave can be retained.
[0060] Specifically, the signal is decomposed into IMFs of different frequency scales using EMD, and then the high-frequency IMFs are selected for signal reconstruction to obtain the reconstructed signal x. recon (t) can effectively separate and remove noise and low-frequency interference components, making the reconstructed signal more focused on the high-frequency core features of the first window waveform, laying the foundation for subsequent instantaneous frequency calculation;
[0061] For the reconstructed signal x reconAfter performing a Hilbert transform on z(t), the reconstructed signal is added to obtain the analytic signal z(t). Specifically, the analytic signal is equal to the reconstructed signal plus the imaginary unit multiplied by the result of the reconstructed signal after the Hilbert transform. Since the real signal itself cannot directly describe the instantaneous frequency through phase change, the real-valued reconstructed signal is extended into a complex-valued analytic signal through the Hilbert transform, providing an analytical basis on the complex plane for extracting the instantaneous frequency from the phase change.
[0062] Let the phase of the analytic signal be φ(t). The analytic signal can be expressed as the magnitude of z(t) multiplied by the complex exponent e. jφ(t) Given the form f(t), where j is the imaginary unit, dividing 1 by 2π and then multiplying by the derivative of the phase φ(t) with respect to time t yields the instantaneous frequency sequence f. i (t), where i is the branch identifier used to distinguish the instantaneous frequency sequences of tangential and direct branches. i=T represents the instantaneous frequency sequence of the tangential branch, and i=F represents the instantaneous frequency sequence of the direct branch. Specifically, the rate of phase change of a signal directly reflects the time-varying characteristics of the frequency. The faster the phase changes with time, the higher the frequency. By differentiating the phase and dividing by 2π for scaling, the instantaneous frequency can be obtained. The instantaneous frequency describes the instantaneous oscillation frequency of the signal at a certain moment. The instantaneous frequency sequence is the set of instantaneous frequencies that change with time within the first arrival window time interval. The instantaneous frequency sequence can not only reflect the true frequency evolution of the first arrival window waveform, but also resist frequency jitter caused by noise.
[0063] Specifically, by combining the reconstructed signal obtained through EMD with the time-frequency analysis of Hilbert transform, a link is formed that first purifies the signal and then accurately extracts the frequency. This solves the problems of frequency jumps and ambiguity of physical meaning when directly performing Hilbert transform on noisy signals, and ensures that the instantaneous frequency can reliably reflect the physical characteristics of partial discharge pulse dispersion.
[0064] The frequency fitting module obtains the instantaneous frequency sequence f of the tangential branch. T (t) and the instantaneous frequency sequence f of the direct branch F (t) Specifically, the instantaneous frequency sequence of the tangential branch and the instantaneous frequency sequence of the through branch are the instantaneous frequency sequences of the partial discharge signals at two measurement points at the T-shaped branch of the gas-insulated metal-enclosed switchgear: the tangential branch measurement point and the through branch measurement point, respectively, within the first arrival window time interval. The instantaneous frequency sequence of the tangential branch can reflect the evolution law of the high frequency arriving first and the frequency monotonically declining in the first arrival window waveform of the tangential branch. The instantaneous frequency sequence of the through branch can reflect the characteristics of the frequency changing slowly with time due to the weak dispersion of the first arrival window waveform of the through branch.
[0065] Linear fitting is performed on the instantaneous frequency sequences of the tangential branch and the through branch, respectively. The fitting formula for the linear fitting is as follows:
[0066] instantaneous frequency sequence f i (t) = slope s i × time t + intercept b i Where i is the branch identifier, specifically, this fitting formula is used to describe the linear change trend of instantaneous frequency over time, with slope s. i Reflecting the rate of change of instantaneous frequency, for example, a tangential branch is expected to have a negative slope, representing a downward slope in instantaneous frequency, while a through branch is expected to have an approximately zero slope, representing a flat instantaneous frequency. The intercept b... i It is the initial value of the instantaneous frequency when time is 0;
[0067] Calculate the instantaneous frequency sequence f i Energy density E (t) f (t), specifically calculated as follows: for the instantaneous frequency sequence f i For each time t of (t), the energy density E is obtained by calculating the square of the amplitude of the instantaneous frequency at that time. f Specifically, the energy of a signal is proportional to the square of its amplitude. The magnitude of the instantaneous frequency amplitude reflects the intensity of the frequency component at that moment. Therefore, the square of its amplitude can characterize the energy density of the instantaneous frequency at that moment.
[0068] Based on energy density E f (t) Set the fitting weight w(t), specifically: divide the energy density at time t by the maximum energy density in the instantaneous frequency sequence to obtain the fitting weight w(t) at that time. Specifically, during the fitting process, the fitting weight w(t) is related to the energy density E of the instantaneous frequency. f The positive correlation between (t) and the high frequency energy ratio is intended to give greater weight to the time period with a high frequency energy ratio during the fitting process, so that the fitting result is more focused on the core time period that reflects the physical law that high frequency arrives first.
[0069] The fitting weights are used as the weights for the weighted least squares method, and the optimal slope s is obtained by solving the weighted least squares method. i and optimal intercept b i The specific solution process is as follows:
[0070] The instantaneous frequency sequence f i The instantaneous frequency at each time t in (t) is calculated by substituting the preset slope s into the fitting formula. i and intercept b i The difference in instantaneous frequencies calculated later is multiplied by the fitting weight at the corresponding time t, and then the square root is taken. The sum of these values is then used to solve for a set of slopes s that minimize the weighted sum of squared residuals. i and intercept b iSpecifically, by continuously fitting the formula until the instantaneous frequency calculated by the fitting formula approaches the actual instantaneous frequency, the optimal slope s is obtained. i and intercept b i ;
[0071] The fitting residual σ is calculated based on the fitting formula and weights. i The specific calculation process is as follows: calculate the instantaneous frequency and substitute it into the optimal slope s. i and intercept b i The difference between the calculated values of the fitting formula is taken as the sum of the squares of the differences at all times, multiplied by the fitting weight w(t) at the corresponding time. The square root of the quotient obtained by dividing the sum by the sum of the fitting weights w(t) at all times gives the fitting residual σ. i ;
[0072] Determine the fitting residual σ i Is it less than the slope s? i 5% of the absolute value, if the residual σ i <slope s i If the absolute value of the linear fit is multiplied by 0.05, the linear fit is considered valid; otherwise, it is considered invalid data.
[0073] When the linear fit is deemed valid, the tangential branch slope s is obtained. T Slope s of the straight-through branch F Tangential branch fitting residual σ T and the fitting residual σ of the through branch F ;
[0074] Specifically, weighted minimization fitting enhances the weight of high-frequency energy moments, making the slope more closely match the physical law that high frequencies arrive first and then decline. By verifying the fitting residuals, the reliability of the fitting results is ensured, and invalid slopes caused by noise or signal distortion are filtered out, providing accurate and reliable slope parameters for subsequent calculations.
[0075] The normalization module will normalize the tangential branch slope s. T Divide by the TE mode cutoff frequency f c The normalized slope of the tangential branch is then obtained;
[0076] The slope s of the straight-through branch F Divide by the TE mode cutoff frequency f c The normalized slope of the through branch is then obtained;
[0077] Specifically, the normalized slope obtained by dividing the branch slope by the TE mode cutoff frequency eliminates the influence of different cutoff frequency references caused by differences in geometric structure on the slope of different gas-insulated metal-enclosed switchgear sites, making the normalized slope a dimensionless relative slope with cross-site comparability. For example, if the normalized slope of the tangential branch of different sites is -0.3, it means that their decline relative to their own cutoff frequency is consistent.
[0078] The difference between the normalized slope of the tangential branch and the normalized slope of the through branch is marked as the normalized feature. Specifically, the normalized feature is a dimensionless feature that reflects the relative contrast between the instantaneous frequency decline rate of the tangential and through branches. By integrating the difference between the relative decline of the tangential branch and the relative flatness of the through branch through the normalized feature, a single cross-branch feature quantity is formed. The normalized feature not only eliminates the benchmark differences in dimensions and frequencies between stations, but also more intuitively reflects the core physical characteristics of the contrast between the instantaneous frequency decline rate of the first arrival window of the tangential and through branches at the T-shaped bifurcation.
[0079] Specifically, the slope of a single branch is made dimensionless by dividing the slope by the cutoff frequency, eliminating the benchmark differences between sites. By normalizing the slope difference, the slope differences of the two branches are integrated into a single feature. The two steps work together to provide normalized features that are comparable across sites and have clear physical meaning for feature verification and waveform recognition, ensuring the consistency of partial discharge waveform recognition in multi-site scenarios.
[0080] The verification and identification module performs statistical significance verification on the normalized features. The specific verification method is as follows: after performing no less than 30 normalized feature calculations on the measurement points at the T-type branch of the same gas-insulated metal-enclosed switchgear, the mean and standard deviation of the normalized features are obtained. Specifically, a single partial discharge pulse is easily affected by random interference, such as noise in the acquisition circuit, which can cause feature distortion. By performing no less than 30 statistical calculations, the mean and standard deviation are made more stable and can reflect the typical characteristics of the partial discharge pulse, rather than random errors.
[0081] Determine whether the current normalized feature is negative. If it is negative, define the feature Boolean value as 1; otherwise, it is 0. Specifically, according to physical laws, the instantaneous frequency of the tangential branch is expected to decline more significantly, while the direct branch is nearly flat. Therefore, the normalized feature obtained by subtracting the normalized slope of the direct branch from the normalized slope of the tangential branch should show a negative difference.
[0082] To determine if the mean of the normalized feature is negative, if it is, define the Boolean value of the mean as 1; otherwise, define it as 0. To determine if the standard deviation of the normalized feature is less than 0.2 times the absolute value of the mean of the normalized feature, if it is less, define the Boolean value of the standard deviation as 1; otherwise, define it as 0. Specifically, by judging the mean and standard deviation of the normalized feature, it can be ensured that the normalized feature of multiple pulses is concentrated around the negative mean, with low dispersion and statistical significance.
[0083] If the Boolean values for the feature, mean, and standard deviation are all 1, the statistical significance test result is considered to be passed; otherwise, it is considered to be failed.
[0084] Specifically, by jointly verifying the normalized feature sign, mean, and standard deviation, invalid features caused by single-impulse noise or anomalies are filtered out, ensuring that the features involved in the identification are statistically reliable and avoiding random errors from interfering with the identification conclusions.
[0085] If the statistical significance verification result is passed, the original partial discharge signal is determined to conform to the law that the instantaneous frequency of the tangential branch of the gas-insulated metal-enclosed switchgear at the T-type bifurcation point drops and the straight branch is nearly flat. The output identification conclusion is the target waveform. Otherwise, it is determined to be a non-target waveform, such as external noise and other types of discharge waveforms.
[0086] Specifically, by verifying the reliability of the normalized features through statistical methods and filtering out random errors, the accurate identification of the partial discharge waveform at the T-branch of gas-insulated metal-enclosed switchgear was finally achieved.
[0087] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A partial discharge waveform comprehensive recognition system, characterized in that, include: The signal reconstruction module decomposes the first-reach window waveform into intrinsic mode functions based on empirical mode decomposition, selects the intrinsic mode functions for signal reconstruction to obtain the reconstructed signal, performs Hilbert transform on the reconstructed signal to obtain the analytical signal, and calculates the instantaneous frequency sequence based on the phase of the analytical signal. The first-reach window is a window that contains the early high-frequency components of the partial discharge pulse. The frequency fitting module obtains the instantaneous frequency sequences of the tangential branch and the through branch, performs linear fitting on the instantaneous frequency sequences of the tangential branch and the through branch respectively, sets fitting weights, and obtains the slope and intercept by minimizing the sum of squared residuals after weighting. It calculates the fitting residuals and determines whether the linear fitting is effective based on the fitting residuals. If effective, it obtains the slope of the tangential branch and the slope of the through branch. The normalization module calculates the normalized slopes of the tangential branch slope and the through branch slope based on the TE mode cutoff frequency, respectively, to obtain the normalized slopes of the tangential branch and the through branch. The difference between the normalized slope of the tangential branch and the normalized slope of the through branch is marked as the normalized feature.
2. The partial discharge waveform comprehensive recognition system according to claim 1, characterized in that, Also includes: The data acquisition module acquires the original partial discharge signal and the TE mode cutoff frequency. The TE mode cutoff frequency is the cutoff frequency of the equivalent waveguide at the T-type branch of the gas-insulated metal-enclosed switchgear. The time-frequency energy spectrum is obtained by performing a short-time Fourier transform on the original partial discharge signal. The window function of the short-time Fourier transform is the Hanning window. The time-frequency energy spectrum includes the sampling time and signal frequency of the original partial discharge signal. The time interval of the first arrival window is determined by integrating the time-frequency energy spectrum. The original partial discharge signal within the first arrival window time interval is obtained by extracting the original signal segment using a rectangular window function. The first arrival window signal is obtained by high-pass filtering the original signal segment. The cutoff frequency of the high-pass filter is 1.2 times the cutoff frequency of the TE mode. The first-window waveform is obtained after reconstructing the first-window signal.
3. The partial discharge waveform comprehensive identification system according to claim 2, characterized in that, The time interval of the first arrival window is determined by integrating the time-frequency energy spectrum, including: The signal start time range and signal frequency range of the original partial discharge signal are obtained, and the time-frequency energy spectrum is integrated over the signal start time range and signal frequency range to obtain the total signal energy. Using the signal start time range as the time axis, perform a sliding integral along the time axis on the signal frequency of the time-frequency energy spectrum to calculate the cumulative energy. Record the moment when the cumulative energy is ≥ 0.2 × the total signal energy, and mark this moment as the start of the first arrival window. Taking the first arrival window as the first starting moment, record along the time axis the moment when the signal frequency of the time-frequency energy spectrum reaches the maximum signal frequency in the range of signal frequency. Mark this moment as the relay moment. The signal frequency at the relay moment is marked as the signal peak value. Using the relay time as the second starting time, record along the time axis the time when the time-frequency energy spectrum is within the range of signal frequency ≤ 0.5 × signal peak value, and mark this time as the end of the first arrival window; The time interval of the first arrival window is obtained by combining the starting point and the ending point of the first arrival window.
4. The partial discharge waveform comprehensive identification system according to claim 2, characterized in that, The first-reaching window waveform is obtained after reconstructing the first-reaching window signal, including: The waveform of the first-reaching window is obtained by reconstructing and solving the first-reaching window signal based on the orthogonal matching pursuit algorithm in compressed sensing.
5. The partial discharge waveform comprehensive identification system according to claim 2, characterized in that, The reconstructed signal is obtained by selecting the intrinsic mode function and reconstructing the signal, including: Based on empirical mode decomposition, the first-reach window waveform is decomposed into n intrinsic mode functions, denoted as IMF. i (t), where the value of i ranges from [1, n]; Select the IMF for i=[1,3]. i (t) performs signal reconstruction, specifically as follows: IMF of i=[1,3] i The reconstructed signal is obtained by summing the t values.
6. The partial discharge waveform comprehensive identification system according to claim 5, characterized in that, Methods for calculating instantaneous frequency sequences include: The analytic signal is obtained by performing a Hilbert transform on the reconstructed signal and then adding the reconstructed signal back to it. Let the phase of the analytic signal be φ(t). Dividing 1 by 2π and then multiplying by the derivative of the phase φ(t) with respect to time t yields the instantaneous frequency sequence f. i (t), where i is the branch identifier, i=T represents the instantaneous frequency sequence of the tangential branch, and i=F represents the instantaneous frequency sequence of the through branch.
7. The partial discharge waveform comprehensive identification system according to claim 6, characterized in that, Perform linear fitting, including: Linear fitting was performed on the instantaneous frequency sequences of the tangential branch and the through branch based on the fitting formula, which is: Instantaneous frequency sequence f i (t) = slope s i × time t + intercept b i Where i is the branch identifier, the fitting process is as follows: For the instantaneous frequency sequence f i For each time t of (t), the square of the instantaneous frequency amplitude corresponding to time t is taken as the energy density. The weight of time t is obtained by dividing the energy density of time t by the maximum energy density in the instantaneous frequency sequence. This weight is set as the fitting weight. The fitting weights are used as the weights for the weighted least squares method, and the optimal slope s is obtained by solving the weighted least squares method. i and optimal intercept b i The specific solution process is as follows: The instantaneous frequency sequence f i The instantaneous frequency at each time t in (t) is calculated by substituting the preset slope s into the fitting formula. i and intercept b i The difference in instantaneous frequencies calculated later is multiplied by the fitting weight at the corresponding time t, and then the square root is taken. The sum of these values is then calculated, and the optimal slope s is obtained by solving the summation. i and optimal intercept b i .
8. The partial discharge waveform comprehensive identification system according to claim 7, characterized in that, Determining the effectiveness of linear fitting based on the fitting residuals includes: The fitting residuals are calculated based on the fitting formula and weights. The specific calculation process is as follows: Calculate the instantaneous frequency and substitute it into the optimal slope s i and intercept b i The difference between the calculated values of the fitting formula is obtained by multiplying the squares of the differences at all times by the fitting weights at the corresponding times, summing the results, dividing the sum by the sum of the fitting weights at all times, and taking the square root of the quotient. If the residual is less than the slope s i If the absolute value of the linear fit is multiplied by 0.05, the linear fit is considered valid; otherwise, it is considered invalid data. When the linear fit is deemed valid, the tangential branch slope s is obtained. T and the slope s of the straight-through branch F .
9. A partial discharge waveform comprehensive identification system according to claim 8, characterized in that, Methods for calculating the normalized slope include: The tangential branch slope s T Dividing by the TE mode cutoff frequency yields the normalized slope of the tangential branch; The slope s of the straight-through branch F Dividing by the TE mode cutoff frequency yields the normalized slope of the through branch; The difference between the normalized slope of the tangential branch and the normalized slope of the through branch is marked as the normalized feature. The normalized feature is a dimensionless feature that reflects the relative contrast between the instantaneous frequency decline rates of the tangential and through branches. The normalized features are statistically significant by using a verification and identification module.
10. A partial discharge waveform comprehensive identification system according to claim 9, characterized in that, The verification and recognition module obtains the mean and standard deviation of the normalized features, determines whether the current normalized feature is negative, and defines the feature boolean value as 1 if it is negative, otherwise it is 0. Determine if the mean of the normalized feature is negative; if it is negative, define the Boolean value of the mean as 1, otherwise as 0. Determine whether the standard deviation of the normalized feature is less than 0.2 times the absolute value of the mean of the normalized feature. If it is less, define the standard deviation Boolean value as 1; otherwise, it is 0. If the Boolean values for the feature, mean, and standard deviation are all 1, the statistical significance test result is considered to be passed; otherwise, it is considered to be failed. If the statistical significance verification result is passed, the original partial discharge signal is determined to be the target waveform; otherwise, it is determined to be a non-target waveform.