Partial discharge diagnosis model generation method

Through multi-dimensional eigenvalue extraction and multiple algorithm model training, the optimal diagnostic model is selected, which solves the problems of low local discharge recognition rate and high resource demand in the prior art, and improves the fault diagnosis efficiency and accuracy of high-voltage electrical equipment.

CN119089144APending Publication Date: 2024-12-06ZHUHAI YADO MONITORING TECH CO LTD

Patent Information

Application Number
CN202411210903.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing local discharge monitoring methods have low recognition rates, especially in the case of large-capacity samples. The demand for traditional support vector machine (SVM) recognizers has surged in computing time and space storage resources, resulting in inefficient diagnostics in high-voltage electrical equipment.

Method used

By obtaining the ground wave sampling signal of high-voltage electrical equipment, performing multi-dimensional preset feature value extraction, using multiple recognition algorithm models to train the feature value, selecting the optimal local discharge diagnosis model, and adapting to different application scenarios and system resource conditions. .

Benefits of technology

It improves the recognition accuracy of local discharge signals and model training accuracy, reduces the demand for computing resources, is suitable for a variety of application scenarios, and enhances the fault diagnosis capability of high-voltage electrical equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089144A_ABST
    Figure CN119089144A_ABST
Patent Text Reader

Abstract

The invention provides a partial discharge diagnosis model generation method. The partial discharge diagnosis model generation method comprises the steps that ground electric wave sampling signals of high-voltage electrical equipment in multiple signal modes are acquired; performing multi-dimensional preset feature value extraction on the ground electric wave sampling signal corresponding to each signal mode; carrying out partial discharge diagnosis model training on the preset characteristic value by utilizing multiple recognition algorithm models to obtain a training result of each recognition algorithm model; and selecting an optimal partial discharge diagnosis model in the current application scene according to a training result. By applying the partial discharge diagnosis model generation method provided by the invention, the recognition accuracy of the signal mode can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge monitoring, and in particular to a method for generating a partial discharge diagnosis model. Background Art

[0002] Partial discharge is mainly the discharge of internal insulation of high-voltage electrical equipment under the action of high voltage. It is one of the main reasons for failure of various high-voltage equipment. With the increasing requirements for power supply reliability, it is particularly important to timely grasp the health status of high-voltage equipment, formulate correct maintenance countermeasures, and avoid sudden accidents caused by quality problems of the equipment itself.

[0003] The existing partial discharge monitoring methods mainly include: pulse current monitoring method, gas monitoring method, geoelectric wave monitoring method, ultra-high frequency monitoring method and infrared thermal imaging monitoring method. Among them, the geoelectric wave monitoring method is to install the geoelectric wave sampling device on the equipment shell, which has no effect on the operation of high-voltage electrical equipment, and will not be affected by the complex electrical, magnetic, thermal and other interferences inside the high-voltage electrical equipment, which is convenient for accurate positioning of local defects. Therefore, the geoelectric wave monitoring method is an excellent method for partial discharge monitoring and is widely used.

[0004] At present, there are two main research methods for monitoring and diagnosing ground wave partial discharge signals of high-voltage electrical equipment. The first is to compare the collected signal with a certain empirical threshold after processing. If it is higher than this threshold, partial discharge occurs. Although this identification method based on a single characteristic parameter is convenient and simple, the recognition rate of partial discharge is low; the second is to extract the multi-dimensional characteristic parameters of the partial discharge signal, train a partial discharge identification model, and use the partial discharge identifier of the support vector machine (SVM) to identify the collected signal. This method effectively improves the recognition rate of partial discharge, but SVM has a disadvantage: as the number of samples increases, the required computing time and space storage resources will increase exponentially. When identifying large-capacity samples, the SVM partial discharge identifier often cannot work effectively.

[0005] Traditional partial discharge insulation fault diagnosis methods are mainly based on neural networks and support vector machines, but due to the lack of phase information, their feature extraction and pattern recognition techniques are mainly based on applied voltage, discharge amount, discharge time interval and their statistical characteristics. The accuracy of traditional error-correcting output codes (ECOC) classifiers applied to insulation fault diagnosis of high-voltage electrical equipment is not high.

[0006] Therefore, it is necessary to find a more suitable and reliable classifier algorithm model. Summary of the invention

[0007] The main purpose of the present invention is to provide a method for generating a partial discharge diagnosis model which can improve the accuracy of signal pattern recognition.

[0008] In order to achieve the above main objectives, the partial discharge diagnostic model generation method provided by the present invention includes: obtaining the ground wave sampling signals of high-voltage electrical equipment under multiple signal modes; extracting multi-dimensional preset feature values ​​of the ground wave sampling signals corresponding to each signal mode; using multiple recognition algorithm models to train the partial discharge diagnostic model on the preset feature values ​​to obtain the training results of each recognition algorithm model; and selecting the optimal partial discharge diagnostic model in the current application scenario according to the training results.

[0009] As can be seen from the above scheme, the partial discharge diagnostic model generation method of the present invention obtains the ground wave sampling signals under various signal modes, performs multi-dimensional preset feature value extraction, and then uses various recognition algorithm models to train the partial discharge diagnostic model for the preset feature values, thereby selecting the optimal partial discharge diagnostic model under the current application scenario, so that the trained partial discharge diagnostic model is adapted to the application scenario, so that different algorithm models can be selected for application under different application scenarios and system resource conditions. At the same time, multi-dimensional preset feature value extraction is performed to improve the recognition accuracy of the signal mode.

[0010] In a further scheme, the step of extracting multi-dimensional preset characteristic values ​​of the ground wave sampling signal corresponding to each signal mode includes: preprocessing the ground wave sampling signal to obtain waveform signal data after noise suppression and interference filtering; and extracting preset characteristic values ​​based on the waveform signal data.

[0011] It can be seen that since the sampling hardware circuit of the geoelectric wave sampling device itself has a certain basic noise, these noise and interference signals are very unfavorable for subsequent data processing and diagnosis of partial discharge types, and should be removed or suppressed as much as possible. Therefore, when performing multi-dimensional preset feature value extraction, the geoelectric wave sampling signal is subjected to noise suppression and interference filtering to effectively suppress and filter out the signal background noise, external power frequency harmonics and high-frequency interference signals, so that the partial discharge signal can be extracted and identified more clearly, thereby greatly improving the signal-to-noise ratio of the signal and improving the accuracy of model training.

[0012] In a further scheme, the steps of preprocessing the ground wave sampling signal to obtain waveform signal data after noise suppression and interference filtering include: performing data reduction on the ground wave sampling signal to obtain an original ground wave signal array; arranging the data in the original ground wave signal array in time sequence to form a column matrix; folding the column matrix with a preset number of cycles to form a ground wave signal matrix; obtaining the average value of each row of data in the ground wave signal matrix to form a column mean array; performing polynomial fitting on the mean array to obtain a mean fitting array; subtracting the ground wave signal matrix from the mean fitting array to obtain a PRPD image data denoising array; arranging the data in the PRPD image data denoising array in time sequence to form waveform signal data.

[0013] It can be seen that when the ground wave sampling signal is preprocessed to obtain the waveform signal data after noise suppression and interference filtering, the amount of data analysis can be reduced by performing data reduction on the ground wave sampling signal. At the same time, taking the mean of the folded array for polynomial fitting can effectively remove background noise, harmonics and nonlinear trend states.

[0014] In a further solution, the step of extracting preset characteristic values ​​based on the waveform signal data includes: acquiring basic characteristic values ​​in the time domain, frequency component characteristic values ​​and ringing characteristic values ​​of the waveform signal data.

[0015] It can be seen that by obtaining the basic eigenvalues ​​in the time domain, the eigenvalues ​​of the frequency component and the ringing eigenvalues, the characteristics of the partial discharge signal can be obtained in multiple dimensions, which can better indicate the existence and status of the partial discharge signal in the switch cabinet.

[0016] In a further scheme, the basic time domain eigenvalues ​​include amplitude peak, effective value and waveform factor; the frequency component eigenvalues ​​include 50Hz correlation, 100Hz correlation and correlation factor; the ringing eigenvalues ​​include ringing count and ringing kurtosis.

[0017] It can be seen that by obtaining the characteristic values ​​of eight dimensions, namely, amplitude peak value, effective value, waveform factor, 50Hz correlation, 100Hz correlation, correlation factor, ringing count and ringing peakness, it can have sufficient representativeness of the partial discharge signal characteristics and is suitable for a variety of intelligent classification algorithm models with good recognition ability.

[0018] In a further scheme, the step of obtaining the frequency component characteristic value of the waveform signal data includes: performing an upper envelope acquisition on the waveform signal data to obtain a signal envelope; performing a discrete Fourier transform on the signal envelope to obtain envelope spectrum data; obtaining a preset 50 Hz correlation frequency band, a preset 100 Hz correlation frequency band and an irrelevant frequency band according to the envelope spectrum data, wherein the irrelevant frequency band is the remaining frequency band of the envelope spectrum data except the preset 50 Hz correlation frequency band and the preset 100 Hz correlation frequency band; obtaining 50 Hz correlation, 100 Hz correlation and correlation factor according to the preset 50 Hz correlation frequency band, the preset 100 Hz correlation frequency band and the irrelevant frequency band.

[0019] It can be seen that analyzing the 50 Hz correlation, 100 Hz correlation and correlation factor can help extract the partial discharge signal from the background noise and identify and locate the potential fault source.

[0020] In a further embodiment, the 50 Hz correlation is obtained by dividing the square of the maximum value of the spectrum amplitude in the preset 50 Hz correlation frequency band by the square of the effective value of the spectrum amplitude in the irrelevant frequency band. The 100 Hz correlation is obtained by dividing the square of the maximum value of the spectrum amplitude in the preset 100 Hz correlation frequency band by the square of the effective value of the spectrum amplitude in the irrelevant frequency band.

[0021] In a further solution, the step of obtaining the ringing characteristic value of the waveform signal data includes: obtaining the number of data in the waveform signal data that exceeds a preset amplitude as a ringing count.

[0022] It can be seen that in the partial discharge signal, the effective partial discharge signal is usually much larger than the basic noise signal amplitude, but the number of partial discharge pulses is very small. The number of ringing can be determined by obtaining the number of data exceeding the preset amplitude in the waveform signal data.

[0023] In a further scheme, the step of obtaining the ringing characteristic value of the waveform signal data also includes: locating the position of the maximum amplitude point in each column of the PRPD image data denoising array; determining the adjacent area range corresponding to the maximum amplitude point based on the position of the maximum amplitude point in each column; obtaining the position serial number of the maximum amplitude point in each adjacent area to form a pulse point position set; and obtaining the ringing peak value based on the pulse point position set.

[0024] It can be seen that the peak of the waveform in the vicinity of the maximum ringing point reflects the concentration of the PD signal in the vicinity of the point. The greater the degree of concentration, the greater the possibility of the existence of PD at the point, and vice versa. The ringing peak is calculated using the PRPD image data denoising array, which can represent the peak statistics during this period of time.

[0025] In a further solution, the step of selecting the optimal partial discharge diagnostic model for the current application scenario according to the training results includes: confirming the verification accuracy according to the training results; and selecting the optimal partial discharge diagnostic model according to the verification accuracy and the required storage capacity of the recognition algorithm model.

[0026] It can be seen that according to the functional requirements of partial discharge type classification, the two indicators of verification accuracy and required storage capacity of the identification algorithm model are the most restrictive indicators for model evaluation, and the optimal partial discharge diagnosis model for the current application scenario can be selected. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of an embodiment of a method for generating a partial discharge diagnostic model according to the present invention.

[0028] Figure 2 It is a flow chart of the steps of extracting multi-dimensional preset characteristic values ​​of the ground wave sampling signal corresponding to each signal mode in the embodiment of the partial discharge diagnostic model generation method of the present invention.

[0029] Figure 3 It is a flow chart of the steps of preprocessing the ground wave sampling signal in the embodiment of the partial discharge diagnosis model generation method of the present invention to obtain the waveform signal data after noise suppression and interference filtering.

[0030] Figure 4 It is a waveform diagram of a sampling period of the original ground wave signal in the embodiment of the partial discharge diagnosis model generation method of the present invention.

[0031] Figure 5 It is a waveform diagram of a PRPD diagram of an original sampling signal of a ground wave in an embodiment of a method for generating a partial discharge diagnostic model of the present invention.

[0032] Figure 6 It is a waveform diagram of a mean phase PRPD diagram of an electric wave signal in an embodiment of a method for generating a partial discharge diagnostic model of the present invention.

[0033] Figure 7 It is a waveform diagram of a polynomial fitting line diagram of a mean value of an electric wave signal in an embodiment of a method for generating a partial discharge diagnostic model of the present invention.

[0034] Figure 8 It is a waveform diagram of a PRPD diagram after the ground wave signal is denoised and the trend signal is removed in an embodiment of the method for generating a partial discharge diagnostic model of the present invention.

[0035] Fig. 9 It is a waveform signal diagram after the original ground wave signal is denoised in the embodiment of the partial discharge diagnosis model generation method of the present invention.

[0036] Fig.10It is a waveform diagram of a signal envelope obtained by performing upper envelope acquisition on waveform signal data in an embodiment of a method for generating a partial discharge diagnostic model of the present invention.

[0037] Fig.11 It is a spectrum diagram of envelope spectrum data obtained after discrete Fourier transform is performed on the signal envelope in the embodiment of the partial discharge diagnosis model generation method of the present invention.

[0038] The present invention is further described below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0039] Partial discharge diagnostic model generation method embodiment:

[0040] The partial discharge diagnosis model generation method of the present invention is a computer program applied to a computer device, and is used to generate a partial discharge diagnosis model to perform partial discharge diagnosis on high-voltage electrical equipment.

[0041] like Figure 1 As shown, in this embodiment, when the partial discharge diagnostic model generation method is working, step S1 is first performed to obtain ground wave sampling signals of high-voltage electrical equipment under multiple signal modes.

[0042] The signal mode can be set accordingly according to the high-voltage electrical equipment. For example, when the high-voltage electrical equipment is a high-voltage switchgear, the partial discharge types of the high-voltage switchgear defects include: tip discharge, suspended discharge, and surface discharge; the external interference signals include: interference type A, interference type B, and interference type C, which represent different interference sources respectively; the situation without partial discharge and interference source is the normal mode, and there are seven signal modes in total. Therefore, the ground wave sampling signals under the seven signal modes of the high-voltage switchgear can be obtained.

[0043] Before obtaining the ground electric wave sampling signals of high-voltage electrical equipment under various signal modes, an experimental platform needs to be built, including: preparation of the experimental environment, target experimental equipment: high-voltage switchgear, three defect partial discharge generating devices, external interference signal sources, specifically, including high-voltage discharge guns, analog signal generators, high-frequency rectifier power supplies, electric motors, LED lamps or fluorescent lamps, etc.; ground electric wave (TEV) signal acquisition system, including ground electric wave sampling devices, wireless relay devices, system equipment, etc.

[0044] The ground wave sampling device is a sensor device for online monitoring, which is powered by batteries or self-collected energy, and is equipped with a low-power ARM-type single-chip microcomputer and a wireless communication device. The ground wave sampling device has certain data collection, data processing and computing capabilities, but cannot realize large-scale data computing and instant communication. In this embodiment, the sampling rate of the underlying ground wave signal of the ground wave sampling device used is 288000Hz, and each sampling operation lasts for 1 second, that is, the number of samples in 1 second is 288000 data points.

[0045] Experimental tests are carried out on the prepared experimental platform, and different signal modes are tested multiple times to obtain geomagnetic wave sampling signals of high-voltage electrical equipment under multiple signal modes. Each signal mode is tested at least 10 times and the data is recorded. Preferably, each signal mode is tested 50 times and the data is recorded.

[0046] After obtaining the ground wave sampling signals of the high-voltage electrical equipment under multiple signal modes, step S2 is executed to extract multi-dimensional preset feature values ​​for the ground wave sampling signals corresponding to each signal mode. In order to improve the recognition accuracy of the signal mode, multi-dimensional preset feature value extraction is required. The type of preset feature value can be set as needed. By extracting the feature value of the ground wave signal, the existence and state of the partial discharge signal of the high-voltage electrical equipment can be better indicated.

[0047] See also Figure 2 When performing multi-dimensional preset feature value extraction on the ground wave sampling signal corresponding to each signal mode, first execute step S21 to pre-process the ground wave sampling signal to obtain waveform signal data after noise suppression and interference filtering. Since the sampling hardware circuit of the ground wave sampling device itself has a certain basic noise, these noise and interference signals are very unfavorable for subsequent data processing and diagnosis of partial discharge types, and should be removed or suppressed as much as possible. Therefore, when performing multi-dimensional preset feature value extraction, the ground wave sampling signal is subjected to noise suppression and interference filtering to effectively suppress and filter out the signal background noise, external power frequency harmonics and high-frequency interference signals, so that the partial discharge signal can be extracted and identified more clearly, thereby greatly improving the signal-to-noise ratio of the signal and improving the accuracy of model training.

[0048] See also Figure 3 When the ground wave sampling signal is preprocessed to obtain the waveform signal data after noise suppression and interference filtering, step S211 is first executed to reduce the data of the ground wave sampling signal to obtain the original ground wave signal array.

[0049] When reducing the data of the ground wave sampling signal, every 80 consecutive points are set as a sampling segment. For each sampling segment, only one maximum value is retained as the representative value of the sampling segment, and other values ​​are discarded. Therefore, only 3600 sampling point data are retained every second, that is, the effective data volume for subsequent processing is 3600 points per second. The reduced sampling data is used as the original ground wave signal array, and each 20ms is a cycle period, and each cycle period contains 72 sampling data points. The waveform of the original ground wave signal in one sampling period is as follows: Figure 4 The ground wave sampling device can be set with a fixed collection interval, for example, a collection action is performed every ten minutes, so as to ensure that the ground wave sampling device can work continuously with low power consumption.

[0050] After obtaining the original ground wave signal array, step S212 is executed to arrange the data in the original ground wave signal array in time sequence to form a column matrix. For example, assuming that the original ground wave signal array is M, the data in the original ground wave signal array M is arranged in time sequence to form a column matrix, and each original ground wave signal array M has a total of 3600 data points.

[0051] Then, step S213 is performed to fold the column matrix by a preset number of cycles to form a ground wave signal matrix. The preset number can be preset according to the implementation data. For example, the original ground wave signal array M is folded by 50 cycles. In chronological order, every 72 data points are regarded as a separate column, and the original ground wave signal array M is divided into 50 columns in total, thus forming a ground wave signal matrix M2 with 72 rows and 50 columns.

[0052] Using the geomagnetic wave signal matrix M2, a PRPD diagram can be drawn. The horizontal axis of the diagram is 72 data points, each data point represents a phase angle of 5°. The point cloud diagram formed by dotting the data of each row in the order of the corresponding columns can obtain the PRPD diagram. The PRPD diagram of the original sampling signal of the geomagnetic wave is as follows: Figure 5 As shown in the PRPD diagram of the original sampling signal of the ground wave, it can be seen that the original sampling signal of the ground wave has obvious power frequency harmonic characteristics. The harmonic has a regular trend and a large background noise, so it needs to be further removed.

[0053] After obtaining the ground wave signal matrix, execute step S214 to obtain the average value of each row of data in the ground wave signal matrix to form a column of mean array. For example, when calculating the mean of the ground wave signal matrix M2, assume that the mean array of the ground wave signal matrix M2 is M2_mean, and the function expression is: M2_mean=mean(M2,2), that is, according to the order of columns in the ground wave signal matrix M2, calculate the arithmetic mean of each row of data, and connect the arithmetic mean of each row into a column of 72 data to form a column of mean array. After the ground wave signal matrix is ​​processed by mean, the ground wave signal mean phase PRPD diagram is obtained as shown in the figure. Figure 6 shown.

[0054] After obtaining the mean array, execute step S215 to perform polynomial fitting on the mean array to obtain the mean fitting array. For example, when performing polynomial fitting on the mean array M2_mean, assume that the mean fitting array formed after fitting is M2_mean_fit, and the fitting polynomial adopts a 6th-order polynomial. If the degree of the polynomial is greater than 6, it is easy to form an overfitting of the mean array, which will affect the detrending and deharmonic effects. On the contrary, if the degree of the polynomial is less than 6, the mean array is underfitted, and the detrending and deharmonic effects are not sufficient. The function expression is: M2_mean_fit=polyfit(M2_mean, 6). For example, the fitted polynomial is: p(x)=-0.95x 6 -1.84x 5 +4.78x 4 +6.69x 3 -5.92x 2 -4.35x+14.1. After polynomial fitting of the mean value array, the mean polynomial fitting line of the geo-wave signal is obtained as follows: Figure 7 shown.

[0055] After obtaining the mean fitting array, execute step S216 to subtract the ground wave signal matrix from the mean fitting array to obtain the PRPD image data denoising array. For example, the ground wave signal matrix M2 is subtracted from the mean fitting array M2_mean_fit by column to obtain the PRPD image data denoising array M3. The function expression is: M3 = M2-M2_mean_fit. After subtracting the ground wave signal matrix from the mean fitting array, the ground wave signal can be denoised and detrended. Figure 8 As shown in the PRPD diagram after the ground wave signal is denoised and the trend signal is removed, it can be clearly seen that the background noise has been significantly removed, and the harmonics and trend states have been basically eliminated.

[0056] After obtaining the PRPD image data denoising array, execute step S217 to arrange the data of the PRPD image data denoising array in chronological order to form waveform signal data. The obtained PRPD image data denoising array is rearranged into a column array in chronological order to form a section of signal data, which is called waveform signal data, that is, the signal data after the original signal preprocessing, noise suppression, and interference filtering. The waveform signal diagram of the original ground wave signal after denoising is as follows: Fig. 9 shown.

[0057] After the waveform signal data is obtained, step S22 is executed to extract preset feature values ​​according to the waveform signal data.

[0058] In this embodiment, the step of extracting preset characteristic values ​​according to the waveform signal data includes: obtaining basic characteristic values ​​in the time domain, frequency component characteristic values, and ringing characteristic values ​​of the waveform signal data. The basic characteristic values ​​in the time domain include the amplitude peak value, the effective value, and the waveform factor; the frequency component characteristic values ​​include the 50Hz correlation, the 100Hz correlation, and the correlation factor; and the ringing characteristic values ​​include the ringing count and the ringing kurtosis.

[0059] Peak Value: The maximum amplitude in the ground wave signal, reflecting the intensity of the partial discharge event. RMS Value: The root mean square value of the ground wave signal, used to measure the average energy of the signal. Shape Factor: The ratio of the peak value to the effective value, reflecting the waveform characteristics of the signal. 50Hz correlation (R50Hz): The ratio of the energy in the 50Hz frequency band to the average energy in other frequency bands. 100Hz correlation (R100Hz): The ratio of the energy in the 100Hz frequency band to the average energy in other frequency bands. Correlation Factor (R Factor): The ratio of the 50Hz correlation to the 100Hz correlation, reflecting the spectral correlation characteristics of the signal. Pulse Count: In each set of data, the number of points that exceed the specified amplitude threshold represents the total number and duration of ringing, which is used to judge the continuity and stability of partial discharge. Pulse Kurtosis: The kurtosis of the waveform in the vicinity of the maximum ringing point (i.e., the peak point) reflects the concentration of the PD signal in the vicinity of the point. The greater the concentration, the greater the possibility of the existence of PD at the point, and vice versa. Since the PD signal is directly related to the power frequency cycle phase, it is difficult to obtain statistical information within a period of time by calculating the kurtosis of the maximum ringing point in a single cycle. Therefore, this embodiment uses the PRPD map data denoising array to calculate the kurtosis, which represents the statistical results of the kurtosis within this period of time.

[0060] In this embodiment, when the basic characteristic value of the waveform signal data is obtained in the time domain, the amplitude peak value and the effective value can be obtained by performing corresponding peak value and effective value calculation on the waveform signal data, and the waveform factor is equal to the ratio of the peak value to the effective value. This is a technology well known to those skilled in the art and will not be described in detail here.

[0061] In this embodiment, when the frequency component characteristic value of the waveform signal data is obtained, the upper envelope of the waveform signal data is first obtained to obtain the signal envelope. When the waveform signal data is subjected to envelope calculation, the upper envelope is obtained by the peak value method, and the sampling rate is set to 6. After the envelope operation, the amount of data and the noise value in the signal are greatly reduced, but the pulse signal representing the partial discharge event is completely retained. The obtained signal envelope diagram is as follows: Fig.10 shown.

[0062] After obtaining the signal envelope, the signal envelope is discretely transformed to obtain the envelope spectrum data. The time domain signal is converted into a frequency domain signal using the Fourier transform analysis method, and then the frequency components are analyzed. After the signal envelope is discretely transformed, the spectrum diagram obtained is as follows: Fig.11 From the example diagram, we can clearly distinguish the high energy state of 50Hz, 100Hz and their harmonics.

[0063] After obtaining the envelope spectrum data, the preset 50Hz correlation frequency band, the preset 100Hz correlation frequency band and the irrelevant frequency band are obtained according to the envelope spectrum data, wherein the irrelevant frequency band is the remaining frequency band of the envelope spectrum data except the preset 50Hz correlation frequency band and the preset 100Hz correlation frequency band. In this embodiment, the preset 50Hz correlation frequency band: the actual selected frequency band range is 49-51Hz, 50Hz is a specific value, which is a point in the frequency domain. When doing frequency energy statistics, it is necessary to set a range domain to calculate. Selecting the 49-51Hz range is relatively a larger range, because considering that the actual deviation of the signal sampling rate exists, and there is also a frequency deviation of the power frequency in the project. Another reason is that taking a larger range does not affect the reliability of the final result, because the effective partial discharge signal in the partial discharge signal is usually closely related to 50Hz or 100Hz. If there is a partial discharge event in this frequency band, the maximum energy in this frequency band is concentrated on the maximum amplitude point, and the energy of other non-partial discharge events is very small. Preset 100Hz correlation frequency band: The actual selected frequency band range is 99-101Hz, the principle is the same as above. Unrelated frequency band: The remaining frequency bands in 1-400Hz except the 50Hz and 100Hz correlation frequency bands.

[0064] After obtaining the preset 50Hz correlation frequency band, the preset 100Hz correlation frequency band and the irrelevant frequency band, the 50Hz correlation, the 100Hz correlation and the correlation factor are obtained according to the preset 50Hz correlation frequency band, the preset 100Hz correlation frequency band and the irrelevant frequency band. The algorithm of the 50Hz correlation (R50Hz) is as follows: the square of the maximum value of the spectrum amplitude in the preset 50Hz correlation frequency band is divided by the square of the effective value of the spectrum amplitude in the irrelevant frequency band. The algorithm of the 100Hz correlation (R100Hz) is as follows: the square of the maximum value of the spectrum amplitude in the preset 100Hz correlation frequency band is divided by the square of the effective value of the spectrum amplitude in the irrelevant frequency band. The algorithm of the correlation factor (R_Factor) is as follows: the ratio of the 50Hz correlation to the 100Hz correlation. In particular, when the 100Hz correlation = 0, let R_Factor = 0.

[0065] The meaning of ringing in partial discharge signals and the role of judging ringing: Ringing in partial discharge signals refers to a specific signal pattern generated during the partial discharge process, which is manifested as a series of rapidly repeated pulses or oscillating waveforms. Ringing contains important information about partial discharge, such as the intensity, frequency and position of the discharge. Accurate identification and analysis of ringing is of great significance for understanding partial discharge phenomena, fault warning and diagnosis, evaluating the insulation status of equipment, and identifying discharge patterns. Because the duration of a single sampling is fixed in engineering applications, the frequency of ringing is represented by counting the number of ringings in a sampling duration; the kurtosis of the ringing is calculated to characterize the intensity and shape of the ringing; the position of the ringing needs to correspond to the phase of the power frequency, so it is difficult to simply count. In other characteristic values, such as the statistics of the correlation between 50Hz and 100Hz, the ringing position feature can be replaced.

[0066] Principle of partial discharge ringing judgment: The main factor affecting the judgment of ringing is the noise level of the signal. Therefore, it is necessary to take appropriate methods and measures to improve the judgment of noise.

[0067] In this embodiment, the step of obtaining the ringing characteristic value of the waveform signal data includes: obtaining the number of data exceeding the preset amplitude in the waveform signal data as the ringing count. In the partial discharge signal, the effective partial discharge signal is usually much larger than the basic noise signal amplitude, but the number of partial discharge pulses is small. In a sampling signal, most of the data is noise data, a small amount is partial discharge pulse data, and the calculated effective value is approximately equal to the effective value of the noise. Due to the randomness of the noise data, its energy distribution probability is approximately Gaussian distribution. Through the probability calculation of Gaussian distribution, it is obtained that: the probability of data greater than 1.732 times the effective value is 4.16%; the probability of data greater than 2*1.732 times the effective value is 0.03%. When the preset amplitude is set to 2*1.732 times the effective value, about (1-0.03%) = 99.97% of the noise data in the signal can be shielded. Of course, in engineering, the setting of this preset amplitude can be adaptively adjusted according to the application scenario and the situation of the monitoring equipment.

[0068] The amplitude of the background noise or partial discharge ringing in the signal will change over time, and an adaptive dynamic threshold method is needed to dynamically adjust the threshold according to the real-time characteristics of the signal, which can improve the accuracy and reliability of partial discharge ringing identification. A specific algorithm for setting the amplitude dynamic threshold is: take a set of P_rms times the effective value of the waveform signal, preferably, take P_rms = 2*1.732.

[0069] In this embodiment, the step of acquiring the ringing characteristic value of the waveform signal data also includes: locating the position of the maximum amplitude point in each column of the PRPD image data denoising array; determining the adjacent area range corresponding to the maximum amplitude point based on the position of the maximum amplitude point in each column; obtaining the position sequence number of the maximum amplitude point in each adjacent area to form a pulse point position set; and obtaining the ringing kurtosis based on the pulse point position set. The PRPD image data denoising array is used to calculate the ringing kurtosis to represent the statistical results of the ringing kurtosis in this waveform signal.

[0070] For example, when calculating the ringing kurtosis, locate the position Imax of the maximum amplitude point in each column of the PRPD image data denoising array M3. Imax is the sequence number of the corresponding phase sequence, that is, the sequence number of the maximum amplitude point in the 72 data points of the cycle. Select a range within a cycle (0-360°) as the adjacent area range of the maximum amplitude point. Because within a cycle (0-360°), the maximum amplitude point may appear at any position (usually the range where the partial discharge signal appears is at the rising edge of the first quadrant of the cycle or the falling edge of the third quadrant), so specific regulations need to be made for different positions. Preferably, with the maximum amplitude point as the center, 9 data points on each side are the adjacent area range (72 data points in one cycle). If the maximum amplitude point is near the two ends of a cycle, the adjacent area range needs to be appropriately selected. Specifically, if the maximum amplitude point is near the left end of the cycle and there are less than 9 data on its left side, all the data points on the left and the 9 data points on the right are taken as the adjacent area range; if the maximum amplitude point is near the right end of the cycle and there are less than 9 data on its right side, all the data points on the right and the 9 data points on the left are taken as the adjacent area range.

[0071] In one embodiment, in the adjacent area range P_range, the position sequence number of the maximum value in each period array is extracted to form a pulse point position set P_row, and P_row has a total of P_number data. Preferably, P_number=50.

[0072] In another embodiment, in the adjacent area range P_range, the position serial numbers of the largest P_number amplitudes in the adjacent area range P_range of the PRPD map data are combined to form a pulse point position set P_row.

[0073] After obtaining the pulse point position set, the ringing kurtosis is obtained according to the pulse point position set. For the above pulse point position set P_row, the ringing kurtosis is calculated with the maximum amplitude point position as the mean. The specific calculation formula is:

[0074] CL4 = (sum(P_row-Imax)) 4 / P_number;

[0075] CL2=((sum(P_row-Imax)) 2 / P_number) 2 ;

[0076]

[0077] Among them, CL4 is the 4th order center distance, CL2 is the 2nd order center distance, and Kurtosis is the ringing kurtosis. In particular, when the 2nd order center distance = 0, let the ringing kurtosis = 10.

[0078] After extracting the multi-dimensional preset characteristic values ​​of the ground wave sampling signal corresponding to each signal mode, step S3 is executed to train the partial discharge diagnosis model for the preset characteristic values ​​using multiple recognition algorithm models to obtain the training results of each recognition algorithm model. In order to improve the recognition effect of the algorithm model on the partial discharge characteristic values, multiple algorithm models are used, tested and comprehensively evaluated respectively to find and verify the best recognition algorithm model. The recognition algorithm model includes two or more of the recognition algorithm models such as decision trees algorithm model, discriminant analysis algorithm model, ensemble algorithm model, Gaussian kernel model, Gaussian process regression (GPR) algorithm model, generalized additive model (GAM), k-nearest neighbors algorithm model, linear model, multi-class error correction output code model (ECOC), naive Bayes model, neural network model and support vector machines (SVM) algorithm model. Optionally, multiple recognition algorithm models may be used in combination.

[0079] After obtaining the training results of each recognition algorithm model, step S4 is executed to select the optimal partial discharge diagnosis model in the current application scenario according to the training results. There are multiple evaluation indicators for the verification and evaluation of the recognition algorithm model, such as error cost, prior probability and observation weight, required storage capacity of the recognition algorithm model, verification accuracy, classification operation speed, training time, memory usage and other parameters as evaluation indicators.

[0080] In this embodiment, the step of selecting the optimal partial discharge diagnosis model in the current application scenario according to the training results includes: confirming the verification accuracy according to the training results; and selecting the optimal partial discharge diagnosis model according to the verification accuracy and the required storage capacity of the identification algorithm model. According to the application scenario, two of the most restrictive indicators are selected as evaluation indicators of the model: the required storage capacity and verification accuracy of the identification algorithm model, and other evaluation indicators can be ignored because they have no decisive influence on the performance of the actual application product.

[0081] By comparing the verification accuracy of multiple recognition algorithm models and the required storage capacity of the recognition algorithm models, the optimal partial discharge diagnosis model for the current application scenario can be selected. For example, for portable or online terminal equipment, small-sized algorithm models should be preferred, while taking into account the accuracy of classification. For background servers or cloud system equipment, the system capacity and computing resources are large. When applying, the model classification accuracy is selected first, followed by the model capacity and computing speed.

[0082] After selecting the optimal partial discharge diagnosis model, the model is trained and verified multiple times using experimental data, and the parameters are tuned to ultimately determine the optimal model parameters. After model training and verification, the model needs to be packaged and debugged, and ultimately deployed on system devices, which can be handheld terminals or backend servers.

[0083] As can be seen from the above, the partial discharge diagnostic model generation method of the present invention obtains the ground wave sampling signals under various signal modes, performs multi-dimensional preset feature value extraction, and then uses various recognition algorithm models to train the partial discharge diagnostic model for the preset feature values, thereby selecting the optimal partial discharge diagnostic model under the current application scenario, so that the trained partial discharge diagnostic model is adapted to the application scenario, so that different algorithm models can be selected for application under different application scenarios and system resource conditions. At the same time, multi-dimensional preset feature value extraction is performed to improve the recognition accuracy of the signal mode.

[0084] It should be noted that the above are only preferred embodiments of the present invention, but the design concept of the invention is not limited thereto, and any non-substantial modifications made to the present invention using this concept also fall within the protection scope of the present invention.

Claims

1. A method for generating a partial discharge diagnostic model, characterized in that: include: Obtain geomagnetic wave sampling signals of high-voltage electrical equipment in various signal modes; Extracting multi-dimensional preset feature values ​​from the ground wave sampling signal corresponding to each signal mode; Using a plurality of recognition algorithm models to train a partial discharge diagnosis model for the preset characteristic values, and obtaining a training result of each recognition algorithm model; An optimal partial discharge diagnosis model for the current application scenario is selected according to the training results.

2. The method for generating a partial discharge diagnostic model according to claim 1, characterized in that: The step of extracting multi-dimensional preset feature values ​​of the ground wave sampling signal corresponding to each signal mode includes: Preprocessing the ground wave sampling signal to obtain waveform signal data after noise suppression and interference filtering; The preset feature value is extracted according to the waveform signal data.

3. The method for generating a partial discharge diagnostic model according to claim 2, characterized in that: The steps of preprocessing the ground wave sampling signal to obtain waveform signal data after noise suppression and interference filtering include: Performing data reduction on the ground wave sampling signal to obtain an original ground wave signal array; Arranging the data in the original ground wave signal array in time sequence to form a column matrix; folding the column matrix by a preset number of periods to form a geostationary wave signal matrix; Obtaining the average value of each row of data in the geomagnetic wave signal matrix to form a column of mean value array; Performing polynomial fitting on the mean array to obtain a mean fitting array; Subtracting the ground wave signal matrix from the mean value fitting array to obtain a PRPD image data denoising array; The data of the PRPD image data denoising array is arranged in time sequence to form the waveform signal data.

4. The method for generating a partial discharge diagnostic model according to claim 3, characterized in that: The step of extracting the preset characteristic value according to the waveform signal data comprises: The basic time domain characteristic values, frequency component characteristic values ​​and ringing characteristic values ​​of the waveform signal data are acquired.

5. The method for generating a partial discharge diagnostic model according to claim 4, characterized in that: The basic time domain characteristic values ​​include amplitude peak value, effective value and waveform factor; The frequency component characteristic values ​​include 50 Hz correlation, 100 Hz correlation and correlation factor; The ringing characteristic value includes a ringing count and a ringing kurtosis.

6. The method for generating a partial discharge diagnostic model according to claim 5, characterized in that: The step of acquiring the frequency component characteristic value of the waveform signal data comprises: Performing upper envelope acquisition on the waveform signal data to obtain a signal envelope; Performing discrete Fourier transform on the signal envelope to obtain envelope spectrum data; Acquire a preset 50 Hz correlation frequency band, a preset 100 Hz correlation frequency band and an irrelevant frequency band according to the envelope spectrum data, wherein the irrelevant frequency band is the remaining frequency band of the envelope spectrum data except the preset 50 Hz correlation frequency band and the preset 100 Hz correlation frequency band; The 50 Hz correlation, the 100 Hz correlation and the correlation factor are obtained according to the preset 50 Hz correlation frequency band, the preset 100 Hz correlation frequency band and the irrelevant frequency band.

7. The method for generating a partial discharge diagnostic model according to claim 6, characterized in that: The 50 Hz correlation is obtained by dividing the square of the maximum value of the spectrum amplitude in the preset 50 Hz correlation frequency band by the square of the effective value of the spectrum amplitude in the uncorrelated frequency band. The 100 Hz correlation is obtained by dividing the square of the maximum value of the spectrum amplitude in the preset 100 Hz correlation frequency band by the square of the effective value of the spectrum amplitude in the uncorrelated frequency band.

8. The method for generating a partial discharge diagnostic model according to claim 5, characterized in that: The step of acquiring the ringing characteristic value from the waveform signal data further includes: Locate the position of the maximum amplitude point in each column of the PRPD image data denoising array; Determine the range of the adjacent area corresponding to the maximum value point according to the position of the maximum value point in each column; Obtain the position sequence number of the maximum amplitude point within each adjacent area to form a pulse point position set; The ringing kurtosis is obtained according to the pulse point position set.

9. The method for generating a partial discharge diagnostic model according to any one of claims 1 to 8, characterized in that: The step of selecting the optimal partial discharge diagnosis model in the current application scenario according to the training result comprises: confirming the verification accuracy according to the training results; The optimal partial discharge diagnosis model is selected according to the verification accuracy and the required storage capacity of the identification algorithm model.

10. The method for generating a partial discharge diagnostic model according to any one of claims 1 to 8, characterized in that: The recognition algorithm model includes two or more of a decision tree algorithm model, a discriminant analysis algorithm model, an integrated algorithm model, a Gaussian kernel model, a Gaussian process regression algorithm model, a generalized additive model, a K nearest neighbor algorithm model, a linear model, a multi-class error correction output code model, a naive Bayes model, a neural network model and a support vector machine algorithm model.

Citation Information

Patent Citations

  • Multi-classifier information fusion partial discharge diagnostic method

    CN103323749A

  • Partial discharge classification and identification method

    CN112763871A

  • Power equipment partial discharge identification method, system and equipment and storage medium

    CN114325256A

  • Method and system for optimising partial discharge detection

    WO2024074218A1

Cited By

  • Partial discharge monitoring and positioning system and method based on Internet of Things architecture

    CN119936575A