GIS partial discharge intelligent diagnosis system and method based on one-dimensional ultrahigh frequency signal analysis

Abstract

The invention discloses a GIS partial discharge intelligent diagnosis system and method based on one-dimensional ultrahigh frequency signal analysis, and relates to the technical field of power electrical equipment intelligent monitoring, and the system comprises a signal collection and preprocessing module which is used for collecting ultrahigh frequency signals of GIS equipment and obtaining preprocessed signal data through a dynamic threshold algorithm; the discharge initial judgment module is used for performing multi-dimensional sequential judgment to eliminate interference discharge data so as to obtain effective discharge signal data; the feature extraction module is used for performing time domain kurtosis and pulse width analysis, frequency domain energy distribution analysis and time-frequency domain wavelet entropy calculation based on the multi-dimensional features of GIS partial discharge, and generating an optimized feature subset; and the type identification module is used for identifying the partial discharge type by using the integrated learning model to obtain a diagnosis result. According to the invention, the problem of unstable recognition accuracy caused by insufficient signal preprocessing, single feature representation and single classification algorithm in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of power electrical equipment, in particular to a GIS partial discharge intelligent diagnosis system and method based on one-dimensional ultra-high frequency signal analysis. BACKGROUND

[0002] In the power system, GIS (gas insulated switchgear) has become an indispensable key equipment in high-voltage and ultra-high voltage power grids due to its small footprint, high reliability, and small maintenance workload. However, due to factors such as manufacturing process, installation quality, and operating environment, internal partial discharge may occur in GIS equipment during long-term operation. This partial discharge is an important sign of insulation degradation of GIS equipment. If not discovered and handled in a timely manner, it can easily lead to equipment failure and even cause large-scale power outages, posing a serious threat to the safe and stable operation of the power system.

[0003] With the development of sensing technology and signal processing technology, GIS partial discharge diagnosis research has gradually formed various technical routes such as acoustic detection method, ultra-high frequency detection method, and ultra-high frequency detection method. Among them, the diagnosis method based on ultra-high frequency signal analysis has attracted widespread attention due to its non-invasiveness, high sensitivity, and strong anti-interference ability. The existing technology mainly uses sensors to obtain partial discharge signals, extracts basic features through simple signal processing methods, and then uses traditional pattern recognition methods to judge the discharge type, which has realized basic diagnosis function in some scenarios.

[0004] However, the existing GIS partial discharge diagnosis technology still has obvious deficiencies: first, the existing technology lacks effective noise suppression and interference identification mechanism in the signal preprocessing stage, making it difficult to accurately distinguish between effective signals and interference signals; second, traditional methods often only focus on single-dimensional signal characteristics in the feature extraction process, ignoring the multi-dimensional information of discharge signals, making the extracted features insufficient to fully represent different types of discharge characteristics; finally, in the discharge type recognition stage, the existing technology generally uses a single classification algorithm, which is difficult to cope with complex and variable actual discharge environments, resulting in unstable classification accuracy, especially when facing new or composite discharges, the recognition effect is poor; these problems seriously restrict the application effect and promotion value of GIS partial discharge diagnosis technology in actual engineering.

[0005] In view of the problems in the related art, no effective solution has been proposed so far. SUMMARY

[0006] To address the problems in related technologies, this invention proposes a GIS partial discharge intelligent diagnostic system and method based on one-dimensional ultra-high frequency signal analysis. It has the advantages of multi-dimensional sequential judgment and three-domain joint feature analysis, thereby solving the problems of insufficient signal preprocessing, single feature representation, and single classification algorithm in the existing technology, which leads to unstable recognition accuracy.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] According to one aspect of the present invention, a GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis is provided, the GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis includes:

[0009] The signal acquisition and preprocessing module is used to acquire one-dimensional UHF signals at the basin insulator of GIS equipment based on UHF sensors, and perform adaptive preprocessing through a dynamic threshold algorithm to obtain preprocessed signal data.

[0010] The initial discharge judgment module is used to perform multi-dimensional sequential judgment on the preprocessed signal data in order to eliminate interference discharge data and obtain effective discharge signal data.

[0011] The feature extraction module is used to perform time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation on effective discharge signal data based on the multi-dimensional characteristics of GIS partial discharge. It also generates an optimized feature subset by evaluating the feature contribution based on the differences in discharge type and recursively eliminating features.

[0012] The type identification module is used to identify the type of partial discharge based on an optimized feature subset using an ensemble learning model, and obtain diagnostic results.

[0013] Furthermore, adaptive preprocessing using a dynamic threshold algorithm yields preprocessed signal data, including: filtering the acquired one-dimensional UHF signal using an adaptive bandpass filter to suppress power frequency interference and wireless noise; performing dynamic threshold detection on the filtered one-dimensional UHF signal, with the dynamic threshold calculated using the signal mean, signal standard deviation, and an adaptive coefficient related to the signal noise level; calculating discharge pulse characteristics based on the dynamic threshold detection results, and combining the discharge pulse characteristics with time information to generate preprocessed signal data.

[0014] Furthermore, the multi-dimensional sequential judgment of the preprocessed signal data includes: pulse width judgment based on the full width at half maximum (FWHM) parameter of the signal; when the pulse width is greater than a preset pulse width threshold, it is judged as interference discharge data and discarded; frequency band judgment based on the signal data that passed the pulse width judgment using center frequency calculation; when the center frequency is not within the preset frequency band, it is judged as interference discharge data and discarded; wavelet entropy judgment based on the wavelet entropy value; when the wavelet entropy value is less than or equal to a preset wavelet entropy threshold, it is judged as interference discharge data and discarded; peak value ratio judgment based on the peak value to root mean square (RMS) ratio of the signal data that passed the wavelet entropy judgment; when the peak value ratio is less than or equal to a preset peak value ratio threshold, it is judged as interference discharge data and discarded, and the remaining signal data is taken as valid discharge signal data.

[0015] Furthermore, based on the multi-dimensional characteristics of GIS partial discharge, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation are performed on the effective discharge signal data. An optimized feature subset is generated through feature contribution evaluation and recursive elimination based on discharge type differences. This includes: time-domain features based on the transient characteristics of GIS discharge, solving for the kurtosis, peak-to-peak ratio, and pulse width parameters of the effective discharge signal data to construct a time-domain feature vector; obtaining the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data based on the distribution characteristics of partial discharge energy, and constructing a frequency-domain feature vector; using wavelet transform for time-frequency domain analysis, calculating the wavelet entropy and target subband energy ratio parameters of the effective discharge signal data to construct a time-frequency domain feature vector; and employing a recursive feature elimination algorithm to evaluate the importance of the time-domain feature vector, frequency-domain feature vector, and time-frequency domain feature vector, and performing dimensionality reduction to form an optimized feature subset.

[0016] Further, the kurtosis, peak-to-peak ratio (BPRR), and pulse width parameters of the effective discharge signal data are solved to construct a time-domain feature vector. This includes: extracting the time-domain sampling point sequence of the effective discharge signal data; calculating the mean, variance, and fourth moment of the time-domain sampling point sequence; using the ratio of the fourth moment to the fourth power of the standard deviation as the kurtosis to describe the sharpness of the signal waveform; searching for the maximum amplitude point of the effective discharge signal data; determining the peak value corresponding to the maximum amplitude point; using the ratio of the peak value to the mean of the sampling point sequence as the BPRR to characterize the prominence of the discharge pulse; detecting the start and end points where the amplitude of the effective discharge signal data exceeds half of the maximum amplitude; calculating the time interval between the start and end points as the pulse width; and correcting the pulse width based on the insulating gas pressure parameter; and setting weight coefficients for kurtosis, BPRR, and pulse width according to the feature contribution of different discharge types, and using the sum of the products of the weight coefficients and the corresponding feature parameters as the time-domain feature vector.

[0017] Further, the acquisition of the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data, and the construction of a frequency domain feature vector, includes: converting the effective discharge signal data from the time domain to the frequency domain, obtaining the power spectral density distribution curve to determine the total spectral energy and frequency distribution range; dividing the power spectral density distribution curve into different frequency band intervals, measuring the energy components within each frequency band interval, identifying the frequency band interval with the highest energy proportion as the main band, and calculating the ratio of the main band energy to the total spectral energy as the main band energy ratio; based on the power spectral density distribution curve, obtaining the first moment of the frequency point distribution as the spectral mean, calculating the third moment of the frequency point relative to the spectral mean, and normalizing the ratio of this third moment to the spectral standard deviation of the power spectral density distribution curve to obtain the spectral skewness; constructing a weighted integral model of the frequency and power spectral density of the power spectral density distribution curve, performing integration with frequency as the independent variable and power spectral density as the weight function, and determining the integration result as the center frequency; and assigning combined weights of the main band energy ratio, spectral skewness, and center frequency according to the frequency domain feature discrimination of different discharge types to generate a frequency domain feature vector.

[0018] Furthermore, by calculating the wavelet entropy and target sub-band energy ratio parameter of the effective discharge signal data, the time-frequency domain feature vector is constructed as follows: Wavelet multi-scale decomposition is performed on the effective discharge signal data to generate wavelet coefficient matrices for different sub-bands; energy calculation is performed on each sub-band of the wavelet coefficient matrix to obtain the energy distribution of each sub-band, and the energy of each sub-band is normalized; information entropy is calculated using the normalized sub-band energy values ​​to obtain the wavelet entropy, and an entropy threshold is set to distinguish different discharge types; the magnitudes of the normalized sub-band energy values ​​are compared, the sub-band with the largest energy value is determined as the target sub-band, the proportion of the target sub-band energy to the total energy is calculated as the target sub-band energy ratio, and the time-frequency domain feature vector is constructed by combining the wavelet entropy.

[0019] Furthermore, the recursive feature elimination algorithm is used to evaluate the importance of time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors. This includes: merging the time-domain, frequency-domain, and time-frequency-domain feature vectors into a unified feature matrix and normalizing it; constructing an evaluation model based on random forest, inputting the normalized unified feature matrix into the evaluation model, training the evaluation model, and extracting the contribution of each feature to the classification of GIS partial discharge types, and sorting them by contribution; setting up a recursive feature elimination loop based on the feature contribution ranking results, removing the feature with the lowest contribution in each recursive loop, and evaluating the classification performance score of the current feature subset through cross-validation; plotting the relationship curve between the number of features and the classification performance score, analyzing the inflection point of the relationship curve, and determining the importance evaluation level of each feature.

[0020] Furthermore, based on the optimized feature subset, an ensemble learning model is used to identify the type of partial discharge, and the diagnostic results are obtained by: constructing a multi-level heterogeneous ensemble learning framework, inputting the optimized feature subset into the base learner to generate a primary classification result vector, wherein the base learner includes at least two different types of classification algorithms; by designing a feature concatenation transformation mechanism, the primary classification result vector and the optimized feature subset are combined and mapped to form an enhanced feature matrix; based on the enhanced feature matrix, a multi-class logistic regression model with regularization constraints is trained as a meta-learner, and the original class score is output; the original class score is nonlinearly normalized to generate probability distribution vectors for tip discharge, suspended discharge, air gap discharge, surface discharge, normal state, and other states, and the discharge type corresponding to the maximum value in the probability distribution vector is extracted as the diagnostic result.

[0021] According to another aspect of the present invention, a smart diagnostic method for partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis is also provided, the method comprising:

[0022] S1. Based on the UHF sensor, one-dimensional UHF signal at the basin insulator of GIS equipment is collected, and adaptive preprocessing is performed through dynamic threshold algorithm to obtain preprocessed signal data.

[0023] S2. Perform multi-dimensional sequential judgment on the preprocessed signal data to eliminate interference discharge data and obtain effective discharge signal data;

[0024] S3. Based on the multi-dimensional characteristics of partial discharge in GIS, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis and time-frequency domain wavelet entropy calculation are performed on the effective discharge signal data. An optimized feature subset is generated by feature contribution evaluation and recursive elimination based on the differences in discharge type.

[0025] S4. Based on the optimized feature subset, the partial discharge type is identified using an ensemble learning model to obtain the diagnostic results.

[0026] The beneficial effects of this invention are as follows:

[0027] (1) By constructing a multi-dimensional sequential judgment mechanism and a three-domain joint feature analysis system, this invention can accurately identify and eliminate interference discharge data at the signal acquisition end. It can also comprehensively extract discharge features from three dimensions: time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation. Then, it can generate the optimal feature subset by using feature contribution evaluation and recursive elimination technology based on discharge type differences. This greatly improves the comprehensiveness and accuracy of GIS partial discharge feature representation and significantly reduces computational complexity and system cost.

[0028] (2) This invention cleverly integrates the advantages of different types of classification algorithms by designing a multi-level heterogeneous integrated learning framework and a feature cascade transformation mechanism. It combines the primary classification results of the base learner with the optimized feature subset to form an enhanced feature matrix. Then, it uses a multi-class logistic regression learner with regularization constraints to make the final decision and discrimination. Finally, it achieves high-precision identification of six types of discharges: tip discharge, suspension discharge, air gap discharge, surface discharge, normal state and other states. This enables the system to maintain stable identification performance when facing complex and ever-changing actual discharge environments, and provides reliable technical protection for the safe operation of GIS equipment. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of a GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to an embodiment of the present invention.

[0031] Figure 2 This is a detailed implementation diagram of the discharge initial judgment module in the GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to an embodiment of the present invention;

[0032] Figure 3 This is a detailed implementation diagram of the feature extraction module in the GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to an embodiment of the present invention;

[0033] Figure 4 This is a flowchart illustrating the intelligent diagnostic method for partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis according to an embodiment of the present invention. Detailed Implementation

[0034] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0035] According to embodiments of the present invention, a GIS partial discharge intelligent diagnostic system and method based on one-dimensional ultra-high frequency signal analysis is provided.

[0036] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis is provided. This GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis includes:

[0037] Signal acquisition and preprocessing module 1 is used to acquire one-dimensional UHF signals at the basin insulator of GIS equipment based on UHF sensors, and perform adaptive preprocessing through dynamic threshold algorithm to obtain preprocessed signal data.

[0038] The initial discharge judgment module 2 is used to perform multi-dimensional sequential judgment on the preprocessed signal data to eliminate interference discharge data and obtain effective discharge signal data.

[0039] Feature extraction module 3 is used to perform time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis and time-frequency domain wavelet entropy calculation on effective discharge signal data based on the multi-dimensional characteristics of GIS partial discharge. It also generates an optimized feature subset by evaluating the feature contribution based on the differences in discharge type and recursively eliminating it.

[0040] Type identification module 4 is used to identify the type of partial discharge based on the optimized feature subset using an ensemble learning model, and obtain diagnostic results.

[0041] In one embodiment, adaptive preprocessing using a dynamic threshold algorithm to obtain preprocessed signal data includes: filtering the acquired one-dimensional UHF signal using an adaptive bandpass filter to suppress power frequency interference and wireless noise; performing dynamic threshold detection on the filtered one-dimensional UHF signal, wherein the dynamic threshold is calculated using the signal mean, signal standard deviation, and an adaptive coefficient related to the signal noise level; calculating the discharge pulse characteristics based on the dynamic threshold detection results, and combining the discharge pulse characteristics with time information to generate preprocessed signal data.

[0042] Specifically, the signal acquisition and preprocessing module 1 uses an ultra-high frequency sensor installed at the basin insulator of the GIS equipment to acquire ultra-high frequency signals of 300MHz-3GHz;

[0043] Specifically, firstly, GIS UHF signal data is collected, and then the data is filtered using an adaptive bandpass filter with a passband of 300MHz-1.5GHz to suppress power frequency interference and wireless noise.

[0044] Specifically, dynamic threshold detection is then performed on the GIS UHF signal data to calculate the discharge pulse. The threshold formula is Thr = u + 3σ0, where u is the average noise floor value, σ0 is its standard deviation, and 3 is the initial value of the adaptive coefficient k.

[0045] Specifically, in the implementation of the dynamic threshold algorithm, the adaptive coefficient k is a key parameter calculated in real time based on the current operating environment and noise level of the GIS equipment. In this embodiment, the initial value is set to 3, and then dynamically adjusted according to the signal-to-noise ratio. When the system detects an increase in the ambient noise level, the k value can automatically increase to the range of 3.5 to 4.5 by calculating the background noise variance over a continuous 20ms. When the ambient noise is stable, the k value is maintained at the standard value of 3. When the ambient noise level is low, the k value can be reduced to the range of 2.5 to 2.8 to improve detection sensitivity. When calculating the discharge pulse characteristics, the system mainly extracts five key parameters: pulse amplitude (in millivolts), pulse width (in nanoseconds), rise time (in nanoseconds), pulse energy, and phase angle corresponding to the power frequency cycle. The pulse energy is obtained by integrating the discharge signal waveform, and the phase angle is determined by comparing it with the synchronously acquired power frequency reference signal. These parameters are combined with the acquisition timestamp to form a structured data packet, accurate to the microsecond level, as the preprocessed signal data output. The discharge pulse characteristics are calculated using wavelet transform and adaptive fitting algorithms, which can accurately describe the time-domain characteristics of the discharge pulse in noisy environments and provide high-quality input data for subsequent initial discharge judgment.

[0046] In one embodiment, multi-dimensional sequential judgment of preprocessed signal data includes: pulse width judgment based on the full width at half maximum (FWHM) parameter of the signal; when the pulse width is greater than a preset pulse width threshold, it is judged as interference discharge data and discarded; frequency band judgment based on the signal data judged by pulse width using center frequency calculation; when the center frequency is not within the preset frequency band, it is judged as interference discharge data and discarded; wavelet entropy judgment based on the signal data judged by frequency band using wavelet entropy value calculation; when the wavelet entropy value is less than or equal to a preset wavelet entropy threshold, it is judged as interference discharge data and discarded; peak value ratio judgment based on the peak value to root mean square (RMS) ratio of the signal data judged by wavelet entropy; when the peak value ratio is less than or equal to a preset peak value ratio threshold, it is judged as interference discharge data and discarded, and the remaining signal data is taken as valid discharge signal data.

[0047] Specifically, the initial discharge judgment module 2 eliminates interference from the preprocessed signal data to determine whether a discharge exists. The specific workflow is as follows: Figure 2 As shown.

[0048] Specifically, in this step, the preprocessed signal data obtained by acquiring 300MHz-3GHz UHF signal data using a UHF sensor installed at the basin insulator of the GIS equipment in the signal acquisition and preprocessing module 1 is used to make an initial discharge judgment.

[0049] Specifically, the multi-dimensional sequential judgment adopts a four-level serial architecture, with each level of judgment targeting specific signal characteristics, forming a progressive screening mechanism to ensure that the final obtained discharge signal data has high reliability.

[0050] Specifically, firstly, the pulse width (FWHM) of the collected UHF discharge signal data is determined using the formula: FWHM = t2 - t1, where t2 and t1 are the two time points corresponding to the signal amplitude decreasing to 50% of its peak value. When FWHM > 100 ns, it is marked as interference discharge. This invention considers that the full width at half maximum (FWHM) parameter is a key indicator of the time-domain characteristics of the discharge pulse. Real partial discharges within GIS equipment typically exhibit steep and sharp pulse waveforms, and their FWHM parameter is significantly smaller than that of external interference signals. The system accurately calculates FWHM by precisely detecting the rising and falling edges of the signal, combined with a peak localization algorithm. The FWHM threshold range is determined by training a logistic regression model with FWHM as the core feature, combined with other time-domain features, optimizing the threshold, and evaluating the classification performance under different thresholds using k-fold cross-validation. The threshold with the minimum sum of false positive and false negative rates is selected as the final FWHM. Here, the full width at half maximum (FWHM) parameter corresponds to the pulse width determination mentioned in the technical solution. A regression model trained on historical data dynamically adjusts the threshold parameters. Using operating condition parameters as input, it predicts the correction coefficient α (α∈[0.8,1.2]) for the Free Water Hygiene Model (FWHM), ultimately adjusting the threshold to the FWHM value. best =α×FWHM.

[0051] Specifically, next, the frequency band of the ultra-high frequency signal data that meets the discharge pulse width is determined, and its frequency band f c The calculation formula is:

[0052]

[0053] In the formula, f is the UHF discharge signal, and S(f) is the spectrum of the UHF discharge signal, calculated by Fast Fourier Transform, representing the distribution characteristics of the signal power in the frequency domain; |S(f)| 2 This represents the power spectral density of the ultra-high frequency discharge signal. When f c Signals outside the 0.5GHz to 1.5GHz range are marked as interference discharges. This frequency band determination formula calculates the center frequency of the UHF signal, corresponding to the center frequency calculation in the technical solution. This frequency band range is the optimal determination interval derived from statistical analysis of a large amount of GIS partial discharge experimental data. The signal energy concentration within this interval is high, and it is effectively distinguishable from common wireless communication frequency bands, greatly improving the system's ability to identify real discharge signals.

[0054] Specifically, then, wavelet entropy calculation is performed on the ultra-high frequency signal data that meets the discharge pulse width and frequency band requirements. The wavelet entropy calculation formula is as follows: In the formula, p j For normalized energy, Energy in each frequency band c j,k To perform wavelet decomposition on a UHF signal, the j-level decomposition coefficients are obtained, where k is the index for traversing all coefficients within the frequency band. The total energy of the entire frequency band is obtained by summing these coefficients. When the wavelet entropy WE ≤ 0.55, it is marked as interference discharge. Wavelet entropy is an important indicator for measuring the time-frequency characteristics of a signal, reflecting the degree of disorder in the energy distribution of the signal in the wavelet domain. Due to its transient impact characteristics, a real partial discharge signal exhibits a high entropy value in the wavelet domain, while regular periodic interference or steady-state noise exhibits a low entropy value.

[0055] Specifically, finally, the peak value ratio (PSR) is calculated and judged for the ultra-high frequency signal data that meets the requirements of discharge pulse width, frequency band, and wavelet entropy. Where f is the UHF discharge signal and rms is the root mean square value of the UHF discharge signal. When the peak-to-peak ratio (PTR) is ≤0.3, it is marked as interference discharge. The PTR reflects the sharpness of the signal waveform and is an effective parameter for distinguishing pulsed discharge signals from stable background noise. Real partial discharge signals exhibit a high PTR, while background noise or weak discharge signals exhibit a low PTR. The system employs a sliding window-based peak search algorithm and a root mean square calculation method to ensure the accuracy of the PTR parameter.

[0056] Specifically, after completing the four-level sequential judgment, the system marks the signal data that meets all judgment conditions as valid discharge signal data. This data is then processed through timestamp association and phase clustering to form a structured discharge signal feature set for use by the subsequent feature extraction module. By progressively filtering based on four dimensions—pulse width, bandwidth, wavelet entropy, and peak value ratio—the reliability of the final acquired valid discharge signal data is ensured. This multi-dimensional sequential judgment mechanism significantly improves the system's recognition rate of real discharge signals while effectively reducing the false alarm rate, laying the foundation for accurate diagnosis of partial discharge in GIS.

[0057] In one embodiment, based on the multi-dimensional characteristics of GIS partial discharge, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation are performed on the effective discharge signal data. An optimized feature subset is generated through feature contribution evaluation and recursive elimination based on discharge type differences. This includes: calculating the kurtosis, peak-to-peak ratio, and pulse width parameters of the effective discharge signal data based on the transient characteristics of GIS discharge, and constructing a time-domain feature vector; obtaining the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data according to the distribution characteristics of partial discharge energy, and constructing a frequency-domain feature vector; performing time-frequency domain analysis using wavelet transform, and constructing a time-frequency domain feature vector by calculating the wavelet entropy and target subband energy ratio parameters of the effective discharge signal data; and using a recursive feature elimination algorithm to evaluate the importance of the time-domain feature vector, frequency-domain feature vector, and time-frequency domain feature vector, and performing dimensionality reduction to form an optimized feature subset.

[0058] In one embodiment, solving for the kurtosis, peak-to-peak ratio (BPRR), and pulse width parameters of the effective discharge signal data to construct a time-domain feature vector includes: extracting the time-domain sampling point sequence of the effective discharge signal data; calculating the mean, variance, and fourth moment of the time-domain sampling point sequence; using the ratio of the fourth moment to the fourth power of the standard deviation as the kurtosis to describe the sharpness of the signal waveform; searching for the maximum amplitude point of the effective discharge signal data; determining the peak value corresponding to the maximum amplitude point; using the ratio of the peak value to the mean of the sampling point sequence as the BPRR to characterize the prominence of the discharge pulse; detecting the start and end points where the amplitude of the effective discharge signal data exceeds half of the maximum amplitude; calculating the time interval between the start and end points as the pulse width; and correcting the pulse width according to the insulating gas pressure parameter; setting weight coefficients for kurtosis, BPRR, and pulse width based on the feature contribution of different discharge types; and using the sum of the products of the weight coefficients and the corresponding feature parameters as the time-domain feature vector.

[0059] In one embodiment, acquiring the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data, and constructing a frequency domain feature vector includes: converting the effective discharge signal data from the time domain to the frequency domain, obtaining the power spectral density distribution curve to determine the total spectral energy and frequency distribution range; dividing the power spectral density distribution curve into different frequency band intervals, measuring the energy components within each frequency band interval, identifying the frequency band interval with the highest energy proportion as the main band, and calculating the ratio of the main band energy to the total spectral energy as the main band energy ratio; based on the power spectral density distribution curve, obtaining the first moment of the frequency point distribution as the spectral mean, calculating the third moment of the frequency point relative to the spectral mean, and normalizing the ratio of this third moment to the spectral standard deviation of the power spectral density distribution curve to obtain the spectral skewness; constructing a weighted integral model of the frequency and power spectral density of the power spectral density distribution curve, performing integration with frequency as the independent variable and power spectral density as the weight function, and determining the integration result as the center frequency; and assigning combined weights of the main band energy ratio, spectral skewness, and center frequency according to the frequency domain feature discrimination of different discharge types to generate a frequency domain feature vector.

[0060] In one embodiment, constructing a time-frequency domain feature vector by calculating the wavelet entropy and target sub-band energy ratio parameter of the effective discharge signal data includes: performing wavelet multi-scale decomposition on the effective discharge signal data to generate wavelet coefficient matrices for different sub-bands; calculating the energy of each sub-band of the wavelet coefficient matrix to obtain the energy distribution of each sub-band, and normalizing the energy of each sub-band; calculating the information entropy using the normalized sub-band energy values ​​to obtain the wavelet entropy, and setting an entropy threshold to distinguish different discharge types; comparing the normalized energy values ​​of each sub-band, determining the sub-band with the largest energy value as the target sub-band, calculating the proportion of the target sub-band energy to the total energy as the target sub-band energy ratio, and constructing a time-frequency domain feature vector by combining the wavelet entropy.

[0061] In one embodiment, the recursive feature elimination algorithm for evaluating the importance of time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors includes: merging the time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors into a unified feature matrix and normalizing it; constructing an evaluation model based on random forest, inputting the normalized unified feature matrix into the evaluation model, training the evaluation model, extracting the contribution of each feature to the classification of GIS partial discharge types, and sorting them by contribution; setting up a recursive feature elimination loop based on the feature contribution ranking results, removing the feature with the lowest contribution in each recursive loop, and evaluating the classification performance score of the current feature subset through cross-validation; plotting the relationship curve between the number of features and the classification performance score, analyzing the inflection point of the relationship curve, and determining the importance evaluation level of each feature.

[0062] Specifically, the feature extraction module 3 performs feature extraction on the UHF signal data of confirmed discharge, extracting time-domain features, frequency-domain features, and time-frequency-domain features.

[0063] Specifically, the feature extraction module employs a joint process of multi-dimensional feature extraction and adaptive optimization. Through detailed time-domain, frequency-domain, and time-frequency-domain feature calculation methods, combined with recursive feature elimination (RFE), it achieves feature dimensionality reduction and redundancy elimination, improving the accuracy and efficiency of subsequent classification. The specific workflow is as follows: Figure 3 As shown.

[0064] Specifically, the time-domain features in this invention include: kurtosis, peak-to-peak ratio, and pulse width; the frequency-domain features include: main band energy ratio, spectral skewness, and center frequency; and the time-frequency-domain features include: wavelet entropy and optimal sub-band energy ratio.

[0065] It should be noted that this invention is based on the multi-dimensional characteristics of GIS partial discharge signals. The feature extraction module achieves a comprehensive characterization of the effective discharge signal through three-domain joint analysis, constructing an optimal feature subset according to the principle of first comprehensively extracting and then optimizing and filtering. The three-domain feature extraction not only captures the unique characteristics of the discharge signal in each domain, but also demonstrates the advantages of mutual complementarity in the subsequent feature fusion process, making the identification of discharge types more accurate and reliable.

[0066] Specifically, firstly, the time-domain characteristics are calculated as follows: Kurtosis K: Due to the transient nature of partial discharge caused by insulation defects, the kurtosis of the UHF discharge signal is significantly higher than that of the interference signal. The calculation formula is as follows: μ is the mean, σ is the standard deviation, N is the sample length, and f i Here, is the value of the i-th sampling point of the UHF discharge signal; PAR is the ratio of the peak value to the mean value of the UHF discharge signal. Where f is the UHF discharge signal, and μ is the root mean square value of the UHF discharge signal. Its pulse amplitude is much higher than the background noise; pulse width: based on the full width at half maximum (FWHM) of the signal, two time points are marked when the signal amplitude drops to 50% of the peak value, and the time difference is calculated.

[0067] Specifically, the frequency domain characteristics are then calculated: the dominant frequency band energy ratio (MBER): the dominant frequency band is defined as 300MHz-1.5GHz, and the dominant frequency band energy ratio is the ratio of energy in this frequency band to the total energy, and its formula is:

[0068]

[0069] In the formula, X(f) is the Fourier transform result of the UHF discharge signal f; spectral skewness SS: Due to the fixed defect type and relatively concentrated energy distribution, the energy distribution of the UHF discharge signal is relatively low, and its calculation formula is as follows: fii For the ultra-high frequency discharge signal frequency point, μ f Its mean frequency in the spectrum, σ f Its spectral standard deviation, E(f ii ) is f ii Energy at a given location; Center frequency CF: Reflects the central location of energy concentration, and its calculation formula is:

[0070]

[0071] Specifically, the time-frequency domain characteristics are then calculated: wavelet entropy WE, which measures the disorder of energy distribution of a signal in different frequency bands. UHF discharge signals have a lower WE (<1.5) because their energy is concentrated in a few frequency bands; interference signals have a higher WE because their energy is dispersed across multiple frequency bands; target sub-band energy ratio OSER: the sub-band with the largest energy proportion. OSER is the ratio of the energy in that sub-band to the total energy. It is calculated by finding the sub-band j* with the largest energy proportion, and the formula is:

[0072]

[0073] In the formula, E j* The energy of the largest sub-band j*.

[0074] It should be noted that the feature subset is a combination of features selected from the original 8 features (3 time-domain features, 3 frequency-domain features, and 2 time-frequency-domain features) after processing by the recursive feature elimination algorithm, and is used for subsequent dimensionality reduction processing.

[0075] Specifically, finally, the model is recursively trained using RFE to eliminate unimportant features, narrowing down the feature subset. The eight features from the feature extraction module are used as the input feature set, and a random forest is selected as the base model to calculate the importance score for each feature. Based on the importance score, the feature with the lowest score is eliminated, establishing a new, optimized feature set. This invention employs a wrapped feature selection strategy, continuously training the random forest model and evaluating the contribution of each feature to classification accuracy to determine the feature importance ranking. During the recursive process, the system removes the least important feature each time and evaluates the performance of the new feature subset using five-fold cross-validation until the optimal feature combination is found. This recursive elimination mechanism can significantly reduce feature dimensionality and computational complexity while maintaining classification performance, improving the system's real-time performance and robustness.

[0076] Specifically, the recursive feature elimination algorithm used in the feature extraction module first merges and standardizes all features in the time domain, frequency domain, and time-frequency domain. Then, it constructs an evaluation model based on random forest to calculate the importance of each feature. During the iteration process, it gradually removes the features with the lowest contribution and evaluates the classification performance of the remaining feature subset through five-fold cross-validation. Finally, it determines the optimal feature combination by analyzing the inflection point of the curve showing the relationship between the number of features and classification performance. Experiments show that after selecting 4-5 core features from the original 8 features, the system's classification accuracy improves by 2-3 percentage points, and the computational efficiency is improved by about 40%, significantly enhancing the real-time performance and generalization ability of the GIS partial discharge diagnostic system.

[0077] In one embodiment, the identification of partial discharge types using an ensemble learning model based on an optimized feature subset to obtain diagnostic results includes: constructing a multi-level heterogeneous ensemble learning framework; inputting the optimized feature subset into a base learner to generate a primary classification result vector; the base learner includes at least two different types of classification algorithms; designing a feature concatenation transformation mechanism to combine and map the primary classification result vector with the optimized feature subset to form an enhanced feature matrix; training a multi-class logistic regression model with regularization constraints as a meta-learner based on the enhanced feature matrix and outputting the original category score; performing nonlinear normalization on the original category score to generate probability distribution vectors for tip discharge, suspended discharge, air gap discharge, surface discharge, normal state, and other states; and extracting the discharge type corresponding to the maximum value in the probability distribution vector as the diagnostic result.

[0078] It should be noted that ensemble learning models include base learners and meta-learners.

[0079] Specifically, after extracting features from the UHF discharge signal, the type recognition module 4 inputs the optimized feature subset into the ensemble learning model for partial discharge type identification. This ensemble learning model employs a multi-level heterogeneous ensemble framework. The base learner includes three different types of classification algorithms: support vector classification, random forest classification, and XGBoost classification. A feature cascade transformation mechanism is designed to map the primary classification result vectors of the base learner to the optimized feature subset, forming an enhanced feature matrix. Finally, a multi-class logistic regression model with L2 regularization constraints is trained as the meta-learner. This model performs probability scoring and normalization on various discharge types, generating probability distribution vectors for tip discharge, suspended discharge, air gap discharge, surface discharge, normal state, and other states. The discharge type corresponding to the highest probability value is extracted as the final diagnostic result.

[0080] Specifically, in this embodiment, firstly, the support vector classification algorithm maps low-dimensional inseparable data to a high-dimensional space through a kernel function, constructs an optimal hyperplane to achieve classification, and requires adjustment of the following variable parameters: Gaussian kernel function kernel = 'rbf', penalty coefficient C = 12.5, and kernel width gamma = 'scale'. As one of the base learners, this algorithm effectively handles the nonlinear relationship of discharge characteristics by utilizing kernel function technology, and shows good classification performance for types that are difficult to separate linearly, such as air gap discharge and suspension discharge.

[0081] Specifically, the random forest classification algorithm then achieves classification by integrating multiple decision trees through a voting mechanism. This requires adjusting variable parameters, including the number of trees, tree depth, minimum number of split samples, and enabling multi-core parallelism. By constructing multiple decision trees and using a voting method, the algorithm effectively reduces the risk of overfitting, exhibits high sensitivity to temporal features (such as kurtosis and peak-to-peak ratio) in the original optimized feature subset, and demonstrates strong robustness in dealing with noise interference and outliers in ultra-high frequency signals.

[0082] Specifically, the gradient boosting framework in the XGBoost classification algorithm achieves classification by optimizing the objective function using the second derivative. The variable parameters that need to be adjusted include: learning rate, number of trees, tree depth, sample sampling rate, feature sampling rate, and L1 regularization term. As a third base learner, this algorithm combines gradient boosting and regularization techniques, and performs well in handling frequency domain features and time-frequency domain features. In particular, it has high sensitivity to the optimal subband energy ratio feature and good adaptability to the problem of imbalanced samples (such as fewer samples in other state categories).

[0083] Specifically, in Stacking ensemble learning, the three classification algorithms in the type recognition module 4 are fused to construct a multi-level learner structure. The prediction results of multiple base learners are used as new features input to the meta-learner for final decision-making. Specifically, multi-class logistic regression with L2 regularization is used as the meta-learner, which is normalized by the softmax function to output the probability distribution of various discharge type states. Based on the feature cascade transformation mechanism, the primary classification result vectors generated by the three base learners are combined and mapped with the original optimized feature subset to form an enhanced feature matrix. This matrix retains the information of the original features while incorporating the judgment results of different classification algorithms on the samples, significantly improving the classification ability of the meta-learner. L2 regularization constraints prevent overfitting, and the softmax function converts the original class scores into probability distributions, facilitating subsequent decision analysis.

[0084] Specifically, through the above model construction, this invention can identify discharge types: tip discharge, suspended discharge, air gap discharge, surface discharge, normal, and others. The system extracts the type corresponding to the maximum value in the probability distribution vector as the final diagnostic result. Simultaneously, the probability value can also serve as a quantitative indicator of diagnostic reliability. This system demonstrates excellent anti-interference capability and generalization performance, providing reliable technical support for intelligent status monitoring and fault diagnosis of GIS equipment.

[0085] like Figure 4 As shown, according to another embodiment of the present invention, a smart diagnostic method for partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis is also provided. This smart diagnostic method for partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis includes:

[0086] S1. Based on the UHF sensor, one-dimensional UHF signal at the basin insulator of GIS equipment is collected, and adaptive preprocessing is performed through dynamic threshold algorithm to obtain preprocessed signal data.

[0087] S2. Perform multi-dimensional sequential judgment on the preprocessed signal data to eliminate interference discharge data and obtain effective discharge signal data;

[0088] S3. Based on the multi-dimensional characteristics of partial discharge in GIS, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis and time-frequency domain wavelet entropy calculation are performed on the effective discharge signal data. An optimized feature subset is generated by feature contribution evaluation and recursive elimination based on the differences in discharge type.

[0089] S4. Based on the optimized feature subset, the partial discharge type is identified using an ensemble learning model to obtain the diagnostic results.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis, characterized in that, include: The signal acquisition and preprocessing module is used to acquire one-dimensional UHF signals at the basin insulator of GIS equipment based on UHF sensors, and perform adaptive preprocessing through a dynamic threshold algorithm to obtain preprocessed signal data. The initial discharge judgment module is used to perform multi-dimensional sequential judgment on the preprocessed signal data in order to eliminate interference discharge data and obtain effective discharge signal data. The feature extraction module is used to perform time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation on effective discharge signal data based on the multi-dimensional characteristics of GIS partial discharge. It also generates an optimized feature subset by evaluating the feature contribution based on the differences in discharge type and recursively eliminating features. The type identification module is used to identify the type of partial discharge based on an optimized feature subset using an ensemble learning model, and obtain diagnostic results.

2. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 1, characterized in that, Adaptive preprocessing using a dynamic threshold algorithm yields preprocessed signal data including: An adaptive bandpass filter is used to filter the acquired one-dimensional ultra-high frequency signal in order to suppress power frequency interference and wireless noise. Dynamic threshold detection is performed on the filtered one-dimensional ultra-high frequency signal. The dynamic threshold is calculated using the signal mean, signal standard deviation, and an adaptive coefficient related to the signal noise level. The discharge pulse characteristics are calculated based on the dynamic threshold detection results, and the discharge pulse characteristics are combined with time information to generate preprocessed signal data.

3. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 1, characterized in that, Multi-dimensional sequential judgment of preprocessed signal data includes: Based on the full width at half maximum (FWHM) parameter of the signal, the pulse width of the preprocessed signal data is determined. When the pulse width is greater than the preset pulse width threshold, it is determined to be interference discharge data and is removed. The signal data determined by pulse width is judged by calculating the center frequency. When the center frequency is not within the preset frequency band, it is judged as interference discharge data and is removed. Wavelet entropy is used to determine the wavelet entropy of signal data that is judged by frequency band. When the wavelet entropy value is less than or equal to the preset wavelet entropy threshold, it is judged as interference discharge data and is removed. The peak value ratio is determined by the ratio of peak value to root mean square (RMS) of the signal data judged by wavelet entropy. When the peak value ratio is less than or equal to the preset peak value ratio threshold, it is judged as interference discharge data and is removed. The remaining signal data is taken as valid discharge signal data.

4. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 1, characterized in that, Based on the multi-dimensional characteristics of partial discharge in GIS, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis, and time-frequency domain wavelet entropy calculation are performed on effective discharge signal data. Furthermore, through feature contribution evaluation based on discharge type differences and recursive elimination, an optimized feature subset is generated, including: Based on the time-domain characteristics of GIS discharge transients, the kurtosis, peak-to-peak ratio, and pulse width parameters of the effective discharge signal data are solved to construct a time-domain feature vector. Based on the distribution characteristics of partial discharge energy, the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data are obtained, and a frequency domain feature vector is constructed. Time-frequency domain analysis is performed using wavelet transform. By calculating the wavelet entropy of the effective discharge signal data and the target subband energy ratio parameter, a time-frequency domain feature vector is constructed. A recursive feature elimination algorithm is used to evaluate the importance of time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors, and then dimensionality reduction is performed to form an optimized feature subset.

5. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 4, characterized in that, Solving for the kurtosis, peak-to-peak ratio, and pulse width parameters of the effective discharge signal data, and constructing the time-domain feature vector includes: Extract the time-domain sampling point sequence of the effective discharge signal data, calculate the mean, variance and fourth moment of the time-domain sampling point sequence, and use the ratio of the fourth moment to the fourth power of the standard deviation as the kurtosis to describe the sharpness of the signal waveform. Search for the maximum amplitude point of the valid discharge signal data, determine the peak value corresponding to the maximum amplitude point, and use the ratio of the peak value to the mean of the sampling point sequence as the peak ratio to characterize the prominence of the discharge pulse. The starting and ending points of the effective discharge signal data amplitude exceeding half of the maximum amplitude are detected. The time interval between the starting and ending points is calculated as the pulse width, and the pulse width is corrected according to the insulating gas pressure parameter. Based on the characteristic contribution of different discharge types, weighting coefficients for kurtosis, peak ratio, and pulse width are set, and the sum of the products of the weighting coefficients and the corresponding characteristic parameters is used as the time-domain feature vector.

6. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 4, characterized in that, To obtain the main band energy ratio, spectral skewness, and center frequency parameters of the effective discharge signal data, and to construct the frequency domain feature vector, the following parameters are used: The effective discharge signal data is converted from the time domain to the frequency domain to obtain the power spectral density distribution curve, so as to determine the total spectral energy and frequency distribution range. By dividing the power spectral density distribution curve into different frequency bands, measuring the energy components in each frequency band, identifying the frequency band with the highest energy proportion as the main frequency band, and calculating the ratio of the main frequency band energy to the total spectrum energy as the main frequency band energy ratio; Based on the power spectral density distribution curve, the first moment of the frequency point distribution is obtained as the spectral mean. The third moment of the frequency point relative to the spectral mean is calculated, and the ratio of the third moment to the spectral standard deviation of the power spectral density distribution curve is normalized to obtain the spectral skewness. Construct a weighted integral model of the power spectral density distribution curve with frequency and power spectral density as the independent variable, perform integral calculation with frequency as the independent variable and power spectral density as the weight function, and determine the integral result as the center frequency; Based on the frequency domain characteristic distinguishability of different discharge types, a combination weight of the main frequency band energy ratio, spectral skewness, and center frequency is assigned to generate a frequency domain feature vector.

7. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 4, characterized in that, By calculating the wavelet entropy and target subband energy ratio parameter of the effective discharge signal data, a time-frequency domain feature vector is constructed, including: Wavelet multi-scale decomposition is performed on the effective discharge signal data to generate wavelet coefficient matrices for different sub-bands; The energy of each sub-band of the wavelet coefficient matrix is ​​calculated to obtain the energy distribution of each sub-band, and the energy of each sub-band is normalized. The information entropy is calculated using the normalized sub-band energy values ​​to obtain the wavelet entropy, and an entropy threshold is set to distinguish different discharge types. The energy values ​​of each sub-band after normalization are compared, and the sub-band with the largest energy value is determined as the target sub-band. The proportion of the target sub-band energy to the total energy is calculated as the target sub-band energy ratio, and a time-frequency domain feature vector is constructed by combining wavelet entropy.

8. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 4, characterized in that, The recursive feature elimination algorithm is used to evaluate the importance of time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors, including: The time-domain feature vector, frequency-domain feature vector, and time-frequency-domain feature vector are merged into a unified feature matrix and then normalized. An evaluation model based on random forest is constructed. The normalized unified feature matrix is ​​input into the evaluation model to train the evaluation model and extract the contribution of each feature to the classification of GIS partial discharge types, and sort them according to the contribution. Based on the feature contribution ranking results, a recursive feature elimination loop is set up. In each recursive loop, the feature with the lowest contribution is removed, and the classification performance score of the current feature subset is evaluated by cross-validation. Plot the relationship curve between the number of features and the classification performance score, and analyze the inflection point of the curve to determine the importance assessment level of each feature.

9. The GIS partial discharge intelligent diagnostic system based on one-dimensional ultra-high frequency signal analysis according to claim 1, characterized in that, Based on the optimized feature subset, an ensemble learning model is used to identify the type of partial discharge, and the diagnostic results include: A multi-level heterogeneous ensemble learning framework is constructed, and an optimized feature subset is input into the base learner to generate a primary classification result vector. The base learner includes at least two different types of classification algorithms. By designing a feature concatenation transformation mechanism, the primary classification result vector is combined and mapped with the optimized feature subset to form an enhanced feature matrix; Based on the enhanced feature matrix, a multi-class logistic regression model with regularization constraints is trained as a meta-learner, and the original class scores are output. The original category scores are nonlinearly normalized to generate probability distribution vectors for tip discharge, suspension discharge, air gap discharge, surface discharge, normal state, and other states. The discharge type corresponding to the maximum value in the probability distribution vector is extracted as the diagnostic result.

10. A method for intelligent diagnosis of partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis, employing the intelligent diagnosis system for partial discharge in GIS based on one-dimensional ultra-high frequency signal analysis as described in any one of claims 1-9, characterized in that, The method includes: S1. Based on the UHF sensor, one-dimensional UHF signal at the basin insulator of GIS equipment is collected, and adaptive preprocessing is performed through dynamic threshold algorithm to obtain preprocessed signal data. S2. Perform multi-dimensional sequential judgment on the preprocessed signal data to eliminate interference discharge data and obtain effective discharge signal data; S3. Based on the multi-dimensional characteristics of partial discharge in GIS, time-domain kurtosis and pulse width analysis, frequency-domain energy distribution analysis and time-frequency domain wavelet entropy calculation are performed on the effective discharge signal data. An optimized feature subset is generated by feature contribution evaluation and recursive elimination based on the differences in discharge type. S4. Based on the optimized feature subset, the partial discharge type is identified using an ensemble learning model to obtain the diagnostic results.

Images