GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization

Abstract

This invention discloses a GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization, comprising the following steps: acquiring raw partial discharge pulse signals and performing signal processing to output a first processed signal; performing adaptive robust baseline determination on the first processed signal, detecting the wavefront position based on multi-feature fusion and dynamic threshold voting mechanism, and outputting the wavefront arrival time; matching the wavefront arrival time in time, calculating and outputting preliminary location results based on the time difference of arrival method; accumulating several preliminary location results, optimizing them using kernel density estimation, and outputting the final partial discharge fault location. This invention effectively suppresses impulse noise and white noise, providing a stable signal for subsequent processing, significantly improving the detection accuracy of wavefront arrival time, solving the problem of false detections and missed detections caused by noise interference in traditional fixed threshold or single feature methods, and significantly reducing the random error of a single location from an average of 0.33m to within 0.07m, improving the overall location accuracy by 80%.

Description

Technical Field

[0001] This invention relates to the field of partial discharge location technology, and in particular to a GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization. Background Technology

[0002] As one of the most important high-voltage devices in the power grid, the accurate location of partial discharge faults in GIS (Gas Insulated Switchgear) is a crucial direction in GIS fault diagnosis technology research and a significant challenge in building an intelligent monitoring system for GIS operation. During the transportation and installation of ultra-high-voltage GIS equipment, factors such as inadequate process control can easily lead to the accumulation of tiny metal particles inside the GIS and damage to insulation components. Different insulation defects can generate different types of partial discharges, including point discharge, internal discharge, particulate discharge, and floating discharge. These partial discharge defects accelerate the deterioration of insulation materials, and long-term accumulation can lead to insulation breakdown, causing short circuits or even explosions. Statistics show that over 85% of power equipment faults are related to insulation deterioration, and 70% of these faults show abnormal partial discharges six months prior to the fault.

[0003] The ultra-high frequency (UHF) method is one of the main methods for locating partial discharge defects in GIS. When UHF signals propagate inside the fully enclosed metal shell of GIS, they undergo multiple reflections and refractions, resulting in multipath effects such as wavefront "overlapping and blurring" and distortion, causing them to lose their original characteristics. At the same time, the wavefront rising edge of the UHF pulse signal is extremely steep, with a short duration, low amplitude, and low energy density, resulting in a low signal-to-noise ratio. The "random superposition" of external interference and noise further reduces the signal-to-noise ratio, making it difficult to distinguish the wavefront from the interference.

[0004] To address the aforementioned issues, this application introduces kernel density estimation (KDE) to further optimize the positioning results, thereby enhancing the stability of the positioning results, reducing the deviation from the actual location, and achieving accurate positioning of GIS partial discharge faults. Summary of the Invention

[0005] The purpose of this invention is to provide a GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization, which can solve the problems of being unable to quickly locate the location of GIS partial discharge, large location result error and low stability.

[0006] Technical solution

[0007] A GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization includes the following steps: acquiring raw partial discharge pulse signals and performing signal processing to output a first processed signal; performing adaptive robust baseline determination on the first processed signal, detecting the wavefront position based on multi-feature fusion and dynamic threshold voting mechanism, and outputting the wavefront arrival time, wherein the multi-features include: window slope, window curvature, and relative rate of change of window amplitude; matching the wavefront arrival time in time, calculating and outputting preliminary location results based on the time difference of arrival method; accumulating several preliminary location results, optimizing them using kernel density estimation, and outputting the final partial discharge fault location;

[0008] The signal processing employs dual filtering for noise reduction using median filtering and SG smoothing filtering. First, the amplitude sequence of the original partial discharge pulse signal is subjected to median filtering, and then SG smoothing filtering is used to output the first processed signal.

[0009] The adaptive robust baseline determination includes: setting several candidate baseline windows, selecting the optimal baseline window by evaluating the stability score of the absolute slope value sequence, and fusing multiple statistics to generate the robust baseline;

[0010] The kernel density estimation method includes: preprocessing the preliminary location results, cross-validating the kernel density bandwidth, calculating the probability density curve, generating a kernel density analysis map, and selecting the point with the highest probability density as the final partial discharge fault location.

[0011] Beneficial effects

[0012] The present invention has the following significant technical effects:

[0013] 1. Dual-filter noise reduction effectively suppresses impulse noise and white noise, providing a stable and reliable signal foundation for subsequent adaptive robust baseline generation, and solving the problem of baseline drift in complex field environments using traditional single-filter methods;

[0014] 2. The adaptive robust baseline determination combined with the multi-feature dynamic threshold voting mechanism avoids false detection / missed detection under noise interference by fixed thresholds or single features, and significantly improves the detection accuracy at wavefront arrival time.

[0015] 3. Based on TDOA dual-end positioning, kernel density estimation optimization is introduced, which reduces the random error of a single positioning result from an average of 0.33m to less than 0.07m, improving the overall positioning accuracy by about 80%, and achieving a leap from "meter-level" to "decimeter-level" accuracy;

[0016] 4. The computational load of the entire method is controllable, fully meeting the real-time requirements of on-site online monitoring. It provides a reliable technical means for operation and maintenance personnel to quickly and accurately locate the partial discharge fault within the GIS, significantly improving the safe operation level of GIS equipment and possessing outstanding industrial application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the positioning method steps of the present invention;

[0018] Figure 2 This is a schematic diagram of the original pulse waveform of partial discharge in the GIS of this invention;

[0019] Figure 3 This is a schematic diagram illustrating the median filtering effect of the present invention;

[0020] Figure 4(a) is a schematic diagram of the dual filtering effect of the present invention;

[0021] Figure 4(b) is a schematic diagram comparing the different filtering effects of the original pulse waveform of the present invention;

[0022] Figure 5 This is a schematic diagram illustrating the wavehead detection effect of the present invention;

[0023] Figure 6 This is a schematic diagram of the dual-end positioning principle based on TDOA of the present invention;

[0024] Figure 7 This is a schematic diagram of the partial discharge signal localization results of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings.

[0026] Definitions:

[0027] Partial discharge (PD) is a discharge phenomenon that occurs in a localized area of ​​the insulation medium of high-voltage electrical equipment, and it is an early sign of insulation degradation.

[0028] Ultra-high frequency partial discharge (UHF PD) generates electromagnetic wave signals with frequencies in the range of 300MHz to 3GHz. It has the characteristics of high sensitivity and strong anti-interference, and is the main detection method for fault location in GIS.

[0029] The raw pulse signal is an unprocessed, noisy UHF partial discharge time-domain pulse waveform directly acquired by GIS field sensors.

[0030] Median filtering is a non-linear filtering technique that replaces the value of a point in a signal with the median of all sample points in its neighborhood. It is mainly used to filter out isolated impulse noise while protecting the signal edges from blurring.

[0031] Savitzky-Golay Filter (SG filter) is a filtering method that smooths data through local polynomial least squares fitting. While removing white noise, it can better preserve the peak value, width, and rising edge trend of the signal.

[0032] The stability score is an indicator that quantifies the degree of slope fluctuation within the baseline window. It is calculated by weighting the slope standard deviation and range. The smaller the score, the more stable the baseline.

[0033] A slope absolute sequence is a sequence of absolute values ​​of the ratio of the amplitude difference to the time difference between adjacent data points, used to reflect the rate of change of a signal.

[0034] Wave Head Position, the starting point of the rising edge in the UHF partial discharge pulse signal, is the core time reference for TDOA positioning.

[0035] The Time Difference of Arrival (TDOA) method is a two-end positioning method that calculates the location of the discharge source by measuring the time difference between the arrival of the same discharge pulse at both ends of the sensor and combining the sensor spacing and propagation speed.

[0036] The time synchronization module, a high-precision time synchronization module, is used to ensure that the time base is consistent among multiple acquisition channels, reducing the impact of clock deviation on positioning accuracy.

[0037] like Figure 1 As shown, a GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization includes the following steps:

[0038] Step 1: Acquire the original partial discharge pulse signal and perform signal processing to output the first processed signal. The first processed signal is the partial discharge pulse signal after double filtering and noise reduction, which provides a stable input for subsequent baseline determination and wavefront detection.

[0039] In a further embodiment of this application, the signal processing uses median filtering and SG (Savitzky-Golay) smoothing filtering for dual filtering noise reduction. First, the amplitude sequence of the original partial discharge pulse signal is subjected to median filtering, and then SG smoothing filtering is used to output the first processed signal. The median filtering is set to third-order filtering, the SG smoothing filtering is set to third-order filtering, and the window length is 5.

[0040] Considering the complex operating environment of GIS sites, the raw partial discharge pulse signals collected often contain a large amount of interference or noise signals, such as... Figure 2 As shown, especially for white noise and impulse noise, noise reduction is necessary to find a stable baseline.

[0041] Median filtering is a non-linear digital signal processing technique that eliminates isolated noise points by replacing the value of a point in a signal with the median of all its neighboring sample points. Median filtering is effective at filtering impulse noise, and in particular, it can preserve the edges of the signal while filtering out noise, preventing them from becoming blurred.

[0042] SG smoothing is a method that estimates a smoothing value by fitting a local polynomial to data points, thereby achieving signal noise reduction and smoothing. Compared with methods such as moving average and Gaussian filtering, SG smoothing is more advantageous in preserving the trend characteristics of the signal. The basic principle is: the original data points within the window are fitted with a k-1 degree polynomial, and the calculation formula is:

[0043] (1)

[0044] In the formula, i = (-m, -m+1, ..., 0, 1, ..., m-1, m). Therefore, all n original data points within this window can be written as one of the above polynomials. These n polynomials constitute a system of k linear equations, and the k fitting parameters need to be solved. Generally, the selected filter window width n is greater than or at least equal to k.

[0045] When n=k, the fitting parameters are solved using linear algebra; if n>k, the least squares method is used. The smoothing window width of the SG smoothing filter is set to 2m+1, which is also the number of original data points n in the window. n=2m+1, the window is selected from the original measurement data, not the entire dataset, and the window width is an odd number.

[0046] Based on the aforementioned characteristics of median filtering, noise reduction is performed on the pulse noise in the original pulse of partial discharge in GIS, such as... Figure 3 As shown, this "flattens" the random spikes in the original pulse waveform of partial discharge, effectively suppressing pulse noise.

[0047] The pulse waveform after median filtering is further denoised using SG smoothing filter, as shown in Figure 4(a). To more intuitively compare the effects of median filtering and double filtering, they are placed in the same figure as the original pulse waveform, as shown in Figure 4(b). This is a waveform before the rising edge of the pulse. The stable baseline and wavefront position are mainly determined based on this waveform.

[0048] After SG smoothing filtering and noise reduction, the white noise of the pulse waveform is suppressed to a certain extent. More importantly, the pulse data becomes relatively smooth but the trend of change remains unchanged, which is conducive to generating a stable baseline. Moreover, the inflection point at the rising edge is more prominent, which makes it easier to accurately detect the wave head position in the future.

[0049] In a further implementation of this embodiment, step two: adaptive robust baseline determination is performed on the first processed signal, wavefront position is detected based on multi-feature fusion and dynamic threshold voting mechanism, and wavefront arrival time is output. Specifically, this includes:

[0050] S1. Set several candidate baseline windows, and select the optimal baseline window from these candidate baseline windows by evaluating the stability score of the absolute slope value sequence. The baseline is the benchmark for judging whether the signal changes abruptly, and the stability of the baseline directly affects the wavefront detection accuracy. The candidate baseline window is a series of continuous data segments that may serve as the baseline in the buffered data before the rising edge. The most stable segment is selected by traversing and evaluating. In step one, reliable data is provided for baseline evaluation by dual filtering and noise reduction, and then the optimal baseline window is adaptively selected by multi-window stability evaluation.

[0051] In a further implementation of this embodiment, a window constraint range is set for the candidate baseline window. The range of the candidate baseline window is determined by the number of data points buffered before the rising edge of the first processing signal. The minimum value of the candidate baseline window is not less than 80% of the number of data points buffered before the rising edge of the first processing signal. The maximum value of the candidate baseline window is the smaller of the maximum preset number of baseline points and the integer value of one-third of the number of data points buffered before the rising edge. The maximum preset number of baseline points does not exceed 200% of the number of data points buffered before the rising edge of the first processing signal. The traversal step size of the candidate baseline window is 3, which is used to reduce the amount of computation and ensure search accuracy.

[0052] In a further implementation of this embodiment, after determining the constraint range of the candidate baseline window in step S1, the absolute value sequence of the slope within each candidate baseline window is calculated. The stability of the candidate baseline window is quantified by a stability score. The absolute value sequence of the slope is the amplitude difference / time difference of the candidate baseline window. The stability score is obtained by calculating the weighted combination of the standard deviation and range of the absolute value sequence of the slope within the candidate baseline window. The standard deviation has a higher weight than the range, which better reflects the overall stability.

[0053] In a further implementation of this embodiment, the stability score is calculated using the following formula:

[0054] Stability score = λ1 × standard deviation of slope + λ2 × range of slope (2)

[0055] Wherein, the standard deviation represents the degree of slope dispersion, the range represents extreme fluctuations, and λ1 and λ2 are weighting coefficients, with λ1 taking a value of 0.7 and λ2 taking a value of 0.3.

[0056] When traversing the windows for the last time, the candidate baseline window with the lowest stability score is retained as the optimal baseline window; when the stability score is lower than a preset threshold, the window traversal is terminated early, and the optimal baseline window is output at this time.

[0057] In a further implementation of this embodiment, the preset threshold value can be 0.06, 0.08, 0.1, or 0.12. According to the stability score formula, the smaller the stability score, the more stable the system. Therefore, the larger the system threshold value, the wider the range.

[0058] S2: Based on the optimal baseline window, multiple statistics are fused to output a robust baseline;

[0059] In a further implementation of this embodiment, the multiple statistics include: baseline window mean, baseline window median, and robust mean. The robust mean is sorted based on the baseline window amplitude and the first and last values ​​are filtered out before taking the mean. The weighting ratio of the baseline window mean, baseline window median, and robust mean is 0.5:0.3:0.2.

[0060] S3: Based on the statistical characteristics of the robust baseline, the slope threshold and curvature threshold are dynamically generated. Within the sliding window of the robust baseline, the window slope, window curvature and the relative rate of change of window amplitude are voted on respectively. When the total number of votes reaches the preset threshold, the wavefront position is determined.

[0061] The essence of a wavefront is the abrupt change in amplitude where the amplitude rises rapidly from the baseline. This abrupt change is quantified using the slope of the first derivative and the curvature of the second derivative. Simultaneously, slope and curvature thresholds are dynamically generated based on the statistical characteristics within a robust baseline window, avoiding false detections or missed detections when fixed thresholds are used in the presence of noise fluctuations. In the wavefront detection algorithm, the slope and curvature thresholds are calculated based on data within the robust baseline window, where the slope threshold T... s The slope T of the curve is calculated using formula (3). c It is calculated by formula (4), specifically as follows:

[0062] (3)

[0063] (4)

[0064] Among them, S b S m S s These represent the robust baseline window slope, the mean of the window slope, and the standard deviation of the window slope, respectively. C m C s These are the mean and standard deviation of window curvature, respectively; for the upper and lower equations of formula (3), based on To determine the robust baseline slope, if it is consistently less than 0.3, the above formula is used; otherwise, the slope fluctuates greatly. Choose the following formula.

[0065] S4: Search for the starting point of the robust baseline window, set the preset voting threshold and the preset wavefront verification window size, compare the starting point feature value with multiple features and vote on them, count the total number of votes and compare it with the preset voting threshold, and output the wavefront position based on the voting situation.

[0066] In a further implementation of this embodiment, the starting point features are the starting point slope, the starting point curvature, and the relative rate of change of the starting point amplitude. The voting conditions are: starting point slope > slope threshold, 1 vote; starting point curvature > curvature threshold, 1 vote; starting point relative rate of change of amplitude > preset relative rate of change of amplitude threshold, 1 vote. The total number of votes is counted and compared with the preset voting threshold. When the total number of votes is greater than or equal to the preset voting threshold, the wavefront position is output. The wavefront position is the starting point + preset wavefront verification window size / 2. Figure 5 As shown, based on the above steps and methods, the wavefront position was located at t=140ns, which is consistent with the actual wavefront, demonstrating the feasibility of the positioning method of this application.

[0067] In a further implementation of this embodiment, step three is to match the arrival time of the wavefront, calculate and output the preliminary positioning result based on the time difference of arrival method;

[0068] In a further implementation of this embodiment, the time difference of arrival method is based on a dual-end positioning method using sensors at both ends. It utilizes a time synchronization module to synchronize the first processing signals between multiple acquisition channels to obtain the arrival time of the wavefront, and calculates the location of the partial discharge power source based on the distance between the sensors at both ends and the electromagnetic wave propagation speed.

[0069] The principle of dual-end positioning based on time difference of arrival, such as... Figure 6 As shown, when a partial discharge occurs at a voltage of u, it will generate electromagnetic waves with the same amplitude but opposite polarity as u, which will propagate rapidly towards both ends. The distances between the partial discharge point and the sensors at both ends are L and L, respectively. a and L b Based on the time synchronization module, the time error of the partial discharge signals acquired from different acquisition channels is controlled within the nanosecond level, and the times when the partial discharge electromagnetic waves first arrive at the two end sensors can be obtained as t. a and t b The distance between the two sensors is L, and the distance between the midpoint of the two sensors and the partial discharge point is... The propagation speed v of partial discharge electromagnetic waves in GIS can be obtained as follows:

[0070] (5)

[0071] (6)

[0072] The dual-end positioning principle based on time difference of arrival is simple, yet offers high positioning accuracy. Given the structural characteristics of high-voltage GIS (Gas-Insulated Gate Geological Array) with its long partitions and standard units, the electromagnetic voltage signal strength of GIS discharge faults is relatively high. Furthermore, considering that the partial discharge sensor is built into the GIS, ensuring high measurement accuracy, and that the propagation attenuation of the partial discharge electromagnetic voltage signal within the GIS has minimal impact on measurement and positioning, this application can meet the positioning requirements of high accuracy and high positioning precision.

[0073] In a further implementation of this embodiment, step four: accumulate several preliminary positioning results. Since a single preliminary positioning result is affected by noise and has a large random error, the kernel density estimation method is used for optimization, and the final partial discharge fault location is output.

[0074] Based on the above time-of-arrival (TOA) positioning method, the positioning result can be calculated quickly. However, in field applications, due to interference from complex environmental noise, the performance of monitoring devices, and the impact of data quality, the positioning of a single paired pulse still has a large error. To eliminate the influence of single positioning error, this application uses kernel density estimation (KDE) to optimize and analyze the positioning results accumulated over 15 minutes.

[0075] Kernel density estimation, a non-parametric method in statistics, is used to estimate the probability density function of a random variable. Given a set of sample data, it fits a smooth approximation of the overall distribution, rather than assuming a specific form of the distribution as in the normal distribution, and thus better reflects the overall trend and characteristics of the data.

[0076] The kernel density estimation method includes: preprocessing the preliminary location results, cross-validating the kernel density bandwidth, calculating the probability density curve, generating the kernel density analysis map, and selecting the point with the highest probability density as the final partial discharge fault location.

[0077] In a further implementation of this embodiment, the kernel density estimation data preprocessing includes: expanding the coordinate axis range to avoid the boundaries being truncated.

[0078] In a further implementation of this embodiment, the kernel density estimation method uses a Gaussian kernel function and a 5-fold cross-validation to determine the optimal bandwidth parameter. The 5-fold cross-validation divides the dataset into 5 mutually exclusive subsets, and uses 4 subsets for training and 1 subset for validation in turn. After repeating this process 5 times, the average performance is taken to select the bandwidth with the best generalization ability, thereby preventing overfitting and underfitting. Finally, the peak position of the probability density is selected as the final partial discharge fault position.

[0079] Based on the above method and steps, the partial discharge data of a power plant's GIS were located and analyzed. The results are as follows: Figure 7As shown, the known true position is 0.425m. During this process, the single positioning results are widely distributed, with errors reaching several meters or even hundreds of meters. After optimization by the kernel density estimation method, the final positioning result is 0.4711m, with an error of only 0.0461m, which fully meets the on-site positioning accuracy requirements.

[0080] In a further implementation of this embodiment, in order to further verify the effectiveness of the positioning method of this application, the positioning results are recalculated using data that has been cleared up over the years, as shown in Table 1.

[0081] Table 1 Comparative Analysis of On-site Data Location Results

[0082] Data file Actual fault location (m) Original localization result (m) Absolute error (m) Location results for this article (m) Absolute error (m) 637025061060000000.ms -0.95 -1.5 0.55 -1.1019 0.1519 636437430010000000.ms 0 -0.3 0.3 -0.108 0.108 636458859010000000.ms -0.48 -0.7 0.22 -0.467 0.013 637588527050000000.ms 0.425 0.1 0.325 0.4711 0.0461 636869674810000000.ms 4.8 4.6 0.2 4.6552 0.0552 636869680820000000.ms 1.9 1.5 0.4 1.8609 0.0395 total - - -

[0083] Table 1 presents a comparative analysis of the partial discharge location results of GIS systems at six power plants. It can be seen that the average absolute error of the original monitoring device's location results was 0.3325m, generally within the range of 0.5m. The average absolute error of the location results recalculated using the location method of this application was 0.06895m, generally within the range of 0.1m. Compared to the original monitoring device's location results, the overall positioning accuracy has improved by approximately 80%, demonstrating a significant improvement. In the event of a partial discharge fault, this method can help maintenance personnel quickly locate the fault, greatly improving the efficiency of online partial discharge monitoring and providing strong support for the stable and safe operation of the on-site GIS system.

[0084] In a further implementation of this embodiment, this application proposes a GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization, based on dual-filter noise reduction, and obtains the following main conclusions:

[0085] (1) After double filtering and noise reduction, the pulse noise and white noise in the GIS partial discharge pulse signal are effectively suppressed. Moreover, the pulse waveform becomes smoother while the trend of change remains unchanged, which is conducive to generating a stable baseline. At the same time, the inflection point at the rising edge is more prominent, which makes it easier to accurately detect the wave head position in the future.

[0086] (2) In the wave head detection algorithm, an adaptive robust baseline is generated by dynamic candidate baseline window search and stability evaluation to avoid errors under different noise levels with a fixed baseline length; at the same time, dynamic thresholds of slope and curvature are generated based on the statistical characteristics of robust baseline features to adapt to different noise levels of data; finally, the three dimensions of window slope, window curvature and window amplitude relative change are fused to generate wave heads by sliding window voting, reducing the limitations of a single feature and further improving the accuracy of wave head detection.

[0087] (3) By using kernel density optimization to optimize and analyze the accumulated positioning results over a period of time, the positioning accuracy can be controlled from 0.5m to 0.1m compared with the original positioning, and the positioning accuracy can be improved by about 80%.

[0088] (4) The range of single positioning results is large. Further research will be conducted on related noise reduction technology, feature extraction and fusion technology, etc., to improve the accuracy and stability of wave head detection and further improve positioning accuracy.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization, characterized in that, Includes the following steps: The original pulse signal of partial discharge is acquired and processed, and the output is the first processed signal. The first processed signal is subjected to adaptive robust baseline determination. The wavefront position is detected based on multi-feature fusion and dynamic threshold voting mechanism, and the wavefront arrival time is output. The multi-features include: window slope, window curvature and relative rate of change of window amplitude. The wavefront arrival time is matched with time, and the preliminary positioning result is calculated and output based on the time difference of arrival method. Several preliminary positioning results are accumulated and optimized using kernel density estimation method to output the final partial discharge fault location. The signal processing employs dual filtering for noise reduction using median filtering and SG smoothing filtering. First, the amplitude sequence of the original partial discharge pulse signal is subjected to median filtering, and then SG smoothing filtering is used to output the first processed signal. The adaptive robust baseline determination includes: setting several candidate baseline windows, selecting the optimal baseline window by evaluating the stability score of the absolute slope value sequence, and fusing multiple statistics to generate the robust baseline; The kernel density estimation method includes: preprocessing the preliminary location results, cross-validating the kernel density bandwidth, calculating the probability density curve, generating a kernel density analysis map, and selecting the point with the highest probability density as the final partial discharge fault location.

2. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The multiple statistics include: baseline window mean, baseline window median, and robust mean. The robust mean is based on the amplitude of the candidate baseline window, which is sorted and the first and last values ​​are filtered out before the mean is taken.

3. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The stability score is calculated as follows: Stability score = λ1 × standard deviation of slope + λ2 × range of slope, where λ1 and λ2 are weighting coefficients.

4. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The multi-feature fusion and dynamic threshold voting mechanism includes: dynamically generating slope threshold and curvature threshold based on the statistical characteristics of the robust baseline; voting on the window slope, window curvature and relative change rate of window amplitude within the sliding window of the robust baseline; and determining the wavefront position when the total number of votes reaches a preset threshold.

5. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The range of the candidate baseline window is determined by the number of data points buffered before the rising edge of the first processing signal. The minimum value of the candidate baseline window is not less than 80% of the number of data points buffered before the rising edge. The maximum value of the candidate baseline window is the smaller of the maximum preset number of baseline points and the integer value of one-third of the number of data points buffered before the rising edge. The maximum preset number of baseline points does not exceed 200% of the number of data points buffered before the rising edge of the first processing signal. The traversal step size of the candidate baseline window is set to 3.

6. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The stability score is obtained by calculating a weighted combination of the standard deviation and range of the absolute slope sequence within the candidate baseline window, wherein the standard deviation has a higher weight than the range; the window traversal is terminated early when the stability score is lower than a preset threshold. The absolute slope value sequence is the amplitude difference / time difference of the candidate baseline window.

7. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 2, characterized in that, The weighted ratio of the baseline window mean, the baseline window median, and the robust mean is 0.5:0.3:0.

2.

8. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 3, characterized in that, The value of λ1 is 0.7, and the value of λ2 is 0.

3.

9. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The arrival time difference method is based on a dual-end positioning method using sensors at both ends. It uses a time synchronization module to align the first processed signals between multiple acquisition channels to obtain the arrival time of the wavefront, and calculates the location of the partial discharge power source based on the distance between the sensors at both ends and the electromagnetic wave propagation speed.

10. The GIS partial discharge fault location method based on multi-feature fusion and kernel density estimation optimization according to claim 1, characterized in that, The kernel density estimation data preprocessing includes: expanding the coordinate axis range; The kernel density estimation method uses a Gaussian kernel function and a 5-fold cross-validation to determine the optimal bandwidth parameter, and selects the probability density peak position as the final partial discharge fault position.

Images