Partial discharge pulse clustering separation method based on multi-band energy three-dimensional mapping

Abstract

In order to solve the problem of spectrum confusion caused by the fact that a traditional method cannot achieve effective physical-level separation of frequency domain energy aliasing partial discharge pulses, the invention relates to a partial discharge pulse clustering separation method based on multi-band energy three-dimensional mapping, and the method comprises the following steps: collecting broadband signals corresponding to a plurality of pulse units in the partial discharge pulses, and obtaining a plurality of pulse units in the partial discharge pulses; determining a pulse energy integration interval corresponding to the corresponding pulse unit; in a single pulse energy integration interval, a corresponding broadband signal is divided into three sub-signals based on three preset characteristic frequency bands, energy scalar values are calculated after band-pass filtering processing is performed on the three sub-signals in parallel, and three-dimensional coordinates of a pulse unit are established based on the energy scalar values. Mapping the corresponding pulse units into pulse points in the three-dimensional energy characteristic space; clustering and separating all the pulse points based on an adaptive density clustering algorithm to form a pulse cluster; and performing atlas reconstruction on all the pulse clusters.

Description

Technical Field

[0001] The present invention relates to the field of partial discharge measurement, and in particular to a partial discharge pulse clustering and separation method based on multi-band energy three-dimensional mapping. Background Art

[0002] Partial discharge pulses typically refer to pulse signals triggered by insulation defects in power equipment. However, in actual discharge scenarios, these pulses often contain other pulses aliased in the frequency domain or pulses from multiple sources. Therefore, when detecting partial discharge pulses, their features often become confused in traditional spectra. Therefore, accurately separating partial discharge pulses and reconstructing their spectra is crucial for accurate insulation condition diagnosis.

[0003] Traditional separation methods only have single-dimensional phase amplitude analysis capabilities and are unable to effectively physically separate partial discharge pulses with frequency domain energy overlap, resulting in the inability to generate independent discharge maps that can be classified and identified.

[0004] In view of this, there is an urgent need to provide a partial discharge pulse clustering separation method based on multi-band energy three-dimensional mapping to ensure the physical level separation of phase-mixed partial discharge pulses and reconstruct a high-resolution spectrum, which is a problem that needs to be solved urgently. Summary of the Invention

[0005] Based on the above analysis, the main purpose of the present invention is to provide a partial discharge pulse clustering separation method based on multi-band energy three-dimensional mapping, so as to solve the problem that traditional methods cannot achieve effective physical-level separation of partial discharge pulses with frequency domain energy aliasing, resulting in spectrum confusion.

[0006] In this regard, the present invention provides a method for clustering and separating partial discharge pulses based on multi-band energy three-dimensional mapping, comprising: a. collecting broadband signals corresponding to multiple pulse units within a partial discharge pulse, and determining the pulse energy integration interval of the corresponding pulse unit based on the broadband signals; b. within a single pulse energy integration interval, dividing the corresponding broadband signal into three sub-signals based on three preset characteristic frequency bands, performing bandpass filtering on the three sub-signals in parallel, and calculating energy scalar values ​​for each sub-signal; establishing the three-dimensional coordinates of the corresponding pulse unit based on the energy scalar values ​​corresponding to the three sub-signals, and mapping the corresponding pulse unit into pulse points in a three-dimensional energy feature space according to the three-dimensional coordinates; c. clustering and separating all pulse points within the three-dimensional energy feature space to form pulse clusters based on an adaptive density clustering algorithm; and performing spectrum reconstruction on all pulse clusters.

[0007] Preferably, step a includes: detecting the rising edge of the pulse of the broadband signal to determine the starting time of the corresponding pulse unit; taking the starting time as the interception starting point, intercepting the time period covering the effective pulse width of the pulse unit as the pulse energy integration interval of the corresponding pulse unit.

[0008] Preferably, in step b, there are overlapping frequency regions between the three preset characteristic frequency bands.

[0009] Preferably, the step b specifically includes: b1. performing bandpass filtering on the three sub-signals in parallel to obtain filtered signals respectively, calculating the energy integral values ​​corresponding to the filtered signals in the pulse energy integral interval where the three sub-signals are located, and converting the energy integral values ​​into energy scalar values ​​corresponding to the three sub-signals; b2. performing logarithmic transformation on the energy scalar values ​​corresponding to the three sub-signals respectively, and arranging the results of the logarithmic transformation as the three-dimensional coordinates of the pulse units corresponding to the three sub-signals.

[0010] As a further preferred embodiment, it also includes: b11. In the corresponding pulse energy interval, respectively extracting the amplitudes corresponding to the three sub-signals, continuously accumulating the squares of the absolute values ​​of the individual amplitudes to obtain corresponding energy integral values; multiplying the energy integral values ​​with the conversion coefficients to obtain corresponding energy scalar values; b21. performing logarithmic transformation on the energy scalar values ​​to obtain a set of corresponding transformation results, arranging the transformation results according to the dimensions of the three-dimensional energy feature space, and obtaining the three-dimensional coordinates of the corresponding pulse units, so as to map the corresponding pulse units into pulse points in the three-dimensional energy feature space according to the three-dimensional coordinates.

[0011] Preferably, in step c, all pulse points are clustered and separated to form pulse clusters based on the adaptive density clustering algorithm, including: c1. calculating the Euclidean distance set of corresponding pulse points based on all three-dimensional coordinates, and determining the neighborhood radius corresponding to all pulse points based on the median of the Euclidean distance set; obtaining the total number of pulse points and calculating the minimum density threshold of all pulse points; c2. calculating the density value of each pulse point, and dividing the points to be assigned from all pulse points based on the density value and the minimum density threshold; selecting the centroid from the points to be assigned based on the density value, and determining the pulse cluster based on the centroid and the neighborhood radius corresponding to the centroid; until all points to be assigned have been assigned to the corresponding pulse cluster; c3. calculating the Euclidean distance of the center of mass of two adjacent pulse clusters, and if the Euclidean distance of the center of mass is less than the neighborhood radius corresponding to the two centroids, the two pulse clusters are merged.

[0012] As a further preferred embodiment, in step c, the spectrum reconstruction of all pulse clusters includes: c4. calculating the phase angle set corresponding to all pulse clusters under a synchronous clock; calculating the normalized amplitude set corresponding to all pulse clusters based on the relationship between all pulse clusters and the corresponding pulse energy integration intervals; c5. reconstructing the PRPD spectrum and PRPS spectrum of the corresponding pulse cluster based on the phase angle set and the normalized amplitude set.

[0013] As a further preferred embodiment, step c2 specifically includes: using a Gaussian kernel function to calculate the density value of each pulse point, arranging all pulse points from high to low based on the density value, and listing the pulse points whose density values ​​are greater than the minimum density threshold as points to be allocated; taking the point to be allocated with the highest current density value as the center of mass, recursively absorbing all points to be allocated within the neighborhood radius corresponding to the center of mass to form a pulse cluster; canceling the identity of all pulse points in the formed pulse cluster as points to be allocated.

[0014] As a further preferred embodiment, the step c4 specifically includes: locking the power frequency cycle zero point corresponding to all pulse clusters based on the synchronous clock, obtaining the phase angle corresponding to each pulse cluster based on the time difference between the pulse leading edge moment of the corresponding pulse unit in the pulse cluster and the power frequency cycle zero point, and generating a phase angle set corresponding to all pulse clusters; extracting the signal peak amplitude within the pulse energy integration interval corresponding to the pulse cluster, and calibrating the signal peak amplitude to obtain a normalized amplitude set corresponding to all pulse clusters.

[0015] As a further preference, in the step c5, the reconstructing of the PRPD spectrum and PRPS spectrum of the corresponding pulse cluster specifically includes: based on the phase angle set and normalized amplitude set corresponding to the single pulse cluster, establishing a two-dimensional coordinate system with the phase angle set as the horizontal axis and the normalized amplitude set as the vertical axis, and generating the PRPD spectrum corresponding to the single pulse cluster through the discrete point distribution in the two-dimensional coordinate system; based on the zero point of the power frequency cycle, calculating the corresponding period number for each pulse unit in the single pulse cluster, establishing a three-dimensional coordinate system with the phase angle set as the X-axis, the period number as the Y-axis, and the normalized amplitude set as the Z-axis, and generating the PRPS spectrum corresponding to the single pulse cluster through color gradient mapping of amplitude intensity.

[0016] The present invention provides a partial discharge pulse clustering and separation method based on multi-band energy three-dimensional mapping, which has the following beneficial effects: First, unlike traditional methods that only have single-dimensional phase amplitude analysis capabilities, the present invention collects the broadband signal corresponding to the pulse unit in the partial discharge pulse, and divides the single broadband signal into three sub-signals based on three preset characteristic frequency bands within a single pulse energy integration interval, calculates the energy scalar value corresponding to each sub-signal, and realizes the parallel analysis and quantitative extraction of multi-band energy, providing frequency domain energy distribution information for the subsequent precise physical-level separation of aliased pulses and multi-source pulses in the three-dimensional energy feature space.

[0017] Secondly, unlike traditional methods that cannot effectively separate pulses with frequency domain energy aliasing, the present invention maps the three characteristic frequency band energy scalar values ​​of each pulse unit into the three-dimensional coordinate values ​​of the pulse point in the three-dimensional energy feature space, constructing a three-dimensional energy feature space that reflects the internal frequency domain energy structure of the pulse, so that pulses from different sources or frequency domain aliasing can be naturally distinguished in this space due to their inherent energy distribution differences, thereby realizing effective physical-level separation of local discharge pulses with frequency domain energy aliasing.

[0018] Thirdly, unlike the spectrum confusion caused by traditional methods, the present invention uses an adaptive density clustering algorithm to cluster and separate all pulse points in the constructed three-dimensional energy feature space to form independent pulse clusters, and realizes effective physical-level separation of aliased pulses and multi-source pulses. At the same time, spectrum reconstruction is performed based on the separated pure pulse clusters, ensuring that the reconstructed spectrum has high discrimination and can clearly reflect the characteristics of different discharge sources. In addition, this method separates based on the essential characteristics of frequency domain energy, effectively avoiding the confusion of traditional spectrum features, and providing a reliable basis for accurate diagnosis of insulation status. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a partial discharge pulse cluster separation method according to an embodiment of the present invention; Figure 2 Schematic diagram of the distribution of pulse points in a three-dimensional energy feature space according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be described in more detail below with reference to the accompanying drawings. It should be noted that the following description of the present invention with reference to the accompanying drawings is only illustrative and not restrictive.

[0021] Where possible, the various embodiments described below may be recombined with each other to form other embodiments not shown in the following description; the various technical features described below may also be recombined with each other to form other embodiments not shown in the following description.

[0022] Please refer to the attached Figure 1 ~Attachment Figure 2 .

[0023] Example: To address the problem of spectral confusion caused by the inability of traditional methods to effectively separate partial discharge pulses with frequency-domain energy aliasing at the physical level, this embodiment provides a partial discharge pulse clustering separation method based on multi-band energy three-dimensional mapping. The partial discharge pulse clustering separation method mainly includes the following steps: Step a: Acquire broadband signals corresponding to several pulse units within a partial discharge pulse, detect the broadband signals, and determine the pulse energy integration interval corresponding to each pulse unit based on the broadband signal detection results. In this step, by detecting the broadband signals corresponding to the pulse units within the partial discharge pulse, time domain information is obtained, and the pulse energy integration interval corresponding to each pulse unit is determined. This prevents energy interference between adjacent pulse units.

[0024] Step b: Within a single pulse energy integration interval, the corresponding broadband signal is divided into three sub-signals based on three preset characteristic frequency bands, and then the three sub-signals are subjected to bandpass filtering in parallel and the energy scalar values ​​corresponding to the three sub-signals are calculated respectively. In the above steps, the energy information corresponding to each sub-signal is obtained by performing three-way parallel bandpass filtering operations and combining the filtering results to form a three-dimensional feature representation in the form of a non-spectral vector, which can enable pulse units of the same type to exhibit clustering characteristics in the three-dimensional energy feature space, laying the foundation for clustering operations. The three-dimensional coordinates of the corresponding pulse unit are established based on the energy scalar values ​​corresponding to the three sub-signals, so that they are mapped into the form of pulse points in the three-dimensional energy feature space.

[0025] Step c: In the three-dimensional energy feature space, all pulse points are clustered and separated according to the adaptive density clustering algorithm to form pulse clusters. At this time, pulse points of the same type show a clustering trend, while pulse points of different types show a separation trend. Subsequently, spectrum reconstruction is performed on all pulse clusters to achieve physical analysis of aliased partial discharge pulses. The relevant spectrum is reconstructed based on the significantly increased energy characteristics after clustering and physically separating the pulse points, thereby finally outputting a highly discriminative fingerprint spectrum.

[0026] The method of this embodiment does not rely on time-domain pulse separation. Its principle is that different discharge sources, such as internal discharge and surface discharge, have inherent energy distribution characteristics in specific frequency bands due to differences in physical mechanisms. For the same time-domain aliased signal, scalar quantization of energy in three characteristic frequency bands and non-spectral analysis are used to construct a three-dimensional space, allowing pulses from different sources to form separable spatial clusters. This achieves effective physical-level separation of partial discharge pulses with frequency-domain energy aliasing, ultimately effectively avoiding the confusion of traditional spectral features.

[0027] To better apply the methods of this embodiment in practical scenarios, some embodiments can be based on the following physical foundations: pulse detection uses standard frontier recognition techniques to determine the pulse energy integration interval; a three-channel filter bank can implement real-time parallel processing based on an FPGA; clustering parameters are adaptively generated using global statistics to ensure robustness to different discharge scale scenarios. In engineering, the distance parameter is fixed based on the sensor installation location, which does not affect phase distribution pattern recognition and fundamentally solves the spectral confusion problem caused by frequency domain energy aliasing.

[0028] To accurately extract the mixed pulse characteristics of partial discharge pulses, in a preferred embodiment, step a includes: detecting the rising edge of the pulse of the broadband signal to determine the starting time of the corresponding pulse unit, then using the starting time as the interception starting point, intercepting the time period covering the effective pulse width of the pulse unit as the pulse energy integration interval. The relevant steps in this embodiment use precise time domain positioning to ensure that the integration interval completely contains the full pulse width energy of a single discharge event, while avoiding crosstalk between adjacent pulses, providing a pure time domain basis for independent feature extraction of multi-source mixed signals. In other related embodiments, the three characteristic frequency bands can be set to 300-500 MHz, 500-800 MHz, and 1000-1500 MHz, respectively. The design reason is that these three frequency bands cover the core response band of ultra-high frequency partial discharge, and the internal discharge energy is concentrated in the 300-500 MHz frequency band, while the surface discharge response is significant in the 1000-1500 MHz frequency band. 500-800 MHz serves as a transition frequency band to enhance the energy distribution contrast of different discharge sources. The division of the characteristic frequency bands mentioned above makes the energy distribution differences of multi-source aliasing signals in the same time period geometrically separable in the three-dimensional energy feature space. In order to more clearly display the distribution of pulse points in the three-dimensional energy feature space in a single-sensor scenario, the original AD waveform of each detected partial discharge pulse is first processed through a three-channel independent RF filter (characteristic frequency bands: 300-500MHz, 500-800MHz, 1000-1500MHz), and its energy scalar value in the three characteristic frequency bands is extracted. The energy scalar value corresponding to each pulse is then transformed and mapped to the three-dimensional energy feature space, as shown in Figure 2. Figure 2 As shown, After filtering the three sub-signals respectively, the three dimensions corresponding to the energy scalar values ​​in the three corresponding characteristic frequency bands are also the three dimensions of the three-dimensional energy feature space.

[0029] To compensate for bandpass filter attenuation, in a preferred embodiment, in step b, three pre-set characteristic frequency bands are independently configured, with adjacent characteristic frequency bands having overlapping frequency regions. This frequency domain overlap between adjacent frequency bands effectively compensates for signal attenuation within the filter's transition band, ensuring that broadband signals are not lost or distorted during the segmentation process. In other embodiments, adjacent frequency bands can have a 5 MHz overlap (e.g., a 500 MHz junction) to compensate for filter transition band attenuation.

[0030] To achieve accurate construction of a three-dimensional energy feature space, in a preferred embodiment, step b specifically includes the following steps: b1. Performing bandpass filtering on the three sub-signals in parallel and filtering the signals separately, calculating the energy integral values ​​of the filtered signals within the pulse energy integral intervals corresponding to the three sub-signals, and obtaining corresponding energy scalar values. In this step, parallel filtering and independent integration ensure that the signals in each frequency band do not interfere with each other, improving the accuracy and stability of energy extraction and providing a reliable data foundation for subsequent three-dimensional mapping. b2. Performing logarithmic transformation on the energy scalar values ​​corresponding to each of the three sub-signals, and mapping the logarithmic transformation results to the coordinates of the pulse points of the corresponding pulse units in the three-dimensional energy feature space. In this step, logarithmic transformation compresses the dynamic range of high-energy pulses while enhancing the distinguishability between low-energy pulses, making the distribution of discharge pulses of different intensities more balanced in three-dimensional space and improving the geometric separability between different discharge sources.

[0031] When combined with the previous characteristic frequency band embodiment, accurate construction of the three-dimensional characteristic space is achieved through the following sub-steps: Parallel filtering and energy integration are performed, that is, bandpass filtering is performed on the three sub-signals (300-500 MHz, 500-800 MHz, and 1000-1500 MHz frequency bands) in parallel to obtain filtered signals , , ; Calculate the pulse energy integration interval of each filtered signal The energy integral value within: ; In the above steps, the aliased signal is converted into an energy feature related to the physical source intensity through time-frequency domain energy scalarization, avoiding the dimensional redundancy of the traditional spectrum vector; Next, logarithmic transformation and three-dimensional mapping are performed, that is, the energy scalar value is logarithmically transformed, and the coordinates of the pulse point are obtained. : ; In the above steps, logarithmic transformation compresses the dynamic range of energy values ​​and improves the spatial linear separability of pulses from different sources; three-dimensional coordinateization makes pulses of the same type exhibit geometric clustering characteristics.

[0032] In this embodiment, different types of local discharge sources differ in their physical mechanisms, resulting in different energy distributions radiated in the UHF signal. The energy integral value reflects this energy radiation characteristic in three characteristic frequency bands: the 300-500 MHz band is more sensitive to discharges in the internal air gaps of insulation, the 1000-1500 MHz high-frequency band is more suitable for characterizing surface creepage discharges, and the intermediate 500-800 MHz transition frequency band helps enhance the energy contrast between different types of discharges. Even if multiple discharge signals are aliased in the time domain, their energy differences in each frequency band can still reflect their respective physical natures. Considering the large dynamic range of discharge energy, directly using a linear scale will cause high-energy pulses to mask low-energy signals. Therefore, a logarithmic transformation is performed on the energy values ​​in each frequency band to compress the spatial offset of high-energy points and improve the resolution of low-energy points, making the distribution of various types of discharge sources in three-dimensional space more balanced. In the final constructed three-dimensional feature space, different discharge types will form their own concentrated point clusters, for example, internal discharges are concentrated in high, In low areas, surface discharge is Low, The energy distribution is high, while noise sources such as corona interference are distributed near the spatial diagonal. This spatial clustering and separability based on energy distribution provides a solid foundation for subsequent density clustering and physical separation.

[0033] In order to make the obtained energy scalar value and coordinate more accurate, in a further preferred embodiment, the following steps are further included: b11. Extract three filtered sub-signals respectively , , In the corresponding pulse energy integration interval Amplitude within 、 、 , continuously accumulate the square of the absolute value of each amplitude to obtain the energy integral values ​​corresponding to the three sub-signals: ; ; ; The corresponding energy scalar value is obtained based on the product of the energy integral value and the conversion coefficient, where: 、 、 The channel conversion coefficient, the cumulative sum is equivalent to the effect of continuous integration. Finally, energy scalarization is achieved by accumulating the squares of discrete amplitudes, and the energy integral value is multiplied by the coefficient 1 to obtain the energy scalar value.

[0034] b21. Perform logarithmic transformation on the energy scalar values ​​to obtain a set of corresponding transformation results: ; Map the conversion result into the dimensional coordinate axis in the three-dimensional energy feature space to obtain the coordinates of the pulse point of the corresponding pulse unit in the three-dimensional energy feature space: .

[0035] In this embodiment, the X-axis, Y-axis, and Z-axis correspond to different characteristic frequency bands. In other embodiments, the X-axis corresponds to the 300-500 MHz energy logarithm, the Y-axis corresponds to the 500-800 MHz energy logarithm, and the Z-axis corresponds to the 1000-1500 MHz energy logarithm. This step establishes a strong correlation between the pulse physical source and the spatial geometric position through quantitative mapping.

[0036] This embodiment effectively improves the calculation accuracy and consistency of the energy scalar value by adopting a continuous accumulation method of discrete amplitude squares during the energy integration process and normalizing the energy of each frequency band in combination with the channel conversion coefficient. This makes the representation of partial discharge signals of different frequency bands and different intensities in the subsequent feature space more realistic and reliable. By performing logarithmic transformation on the energy scalar value and mapping it to the three-dimensional space coordinate axes, a three-dimensional energy feature space with clear physical meaning is constructed, in which the X-axis, Y-axis, and Z-axis correspond to the logarithmic energy values ​​of different characteristic frequency bands, respectively. This quantitative mapping method not only compresses the dynamic range of high-energy pulses and enhances the discrimination between low-energy pulses, but also establishes a strong correlation between the physical source of the partial discharge pulse and its geometric position in the three-dimensional space, providing a stable and interpretable feature basis for subsequent cluster analysis and physical separation.

[0037] In order to achieve robust separation of pulse aliasing, in a preferred embodiment, in step c, all pulse points are clustered and separated to form pulse clusters based on an adaptive density clustering algorithm, including: c1. Adopt the pulse point classification method based on adaptive density clustering, based on all pulse points The Euclidean distance set of coordinates , that is, the Euclidean distance between each two pulse points is aggregated into a set. The neighborhood radius corresponding to all pulse points is determined based on the median of the Euclidean distance set , and get the total number of pulse points to calculate the minimum density threshold of all pulse points: ,in is the total number of pulses.

[0038] In this step, parameters are dynamically generated based on the data distribution to adapt to density differences in different discharge scenarios. This allows the clustering process to adapt to data structures with different density distributions without relying on fixed parameters.

[0039] c2. During the clustering process, the density of each pulse point is first calculated. The points to be assigned are then divided from all the pulse points based on the density value and the minimum density threshold. The centroid is then selected from the points to be assigned based on the density value. Pulse clusters are then determined based on the centroid and its corresponding neighborhood radius. This process continues until all points to be assigned are assigned to their corresponding pulse clusters. This ensures that all points are properly classified, improving the stability and physical interpretability of the clustering results.

[0040] c3. Introduce a merging judgment mechanism based on centroid distance between clusters, that is, for two adjacent pulse clusters 、 , calculate the corresponding two centroids 、 The Euclidean distance of the centroid of the pulse cluster (the mean of the coordinate points in the pulse cluster) is less than the neighborhood radius of the two centroids. , then merge the two pulse clusters 、 This method effectively avoids the over-segmentation problem caused by local density differences. It not only enhances the ability to separate complex mixed partial discharge signals, but also improves the algorithm's adaptability and accuracy when dealing with different discharge source distribution characteristics. It provides high-quality pulse cluster segmentation results with clear physical meaning for subsequent spectrum reconstruction and type identification.

[0041] In order to improve the accuracy of spectrum reconstruction and feature expression ability after partial discharge pulse cluster separation, in a further preferred embodiment, in step c, spectrum reconstruction is performed on each pulse cluster, including the following steps: calculating the phase angle set corresponding to all pulse clusters under a synchronous clock, and calculating the normalized amplitude set corresponding to all pulse clusters based on the correspondence between all pulse clusters and pulse energy integration intervals; generating the PRPD spectrum and PRPS spectrum of the corresponding pulse cluster based on the phase angle set and the normalized amplitude set.

[0042] In this embodiment, phase synchronization and amplitude normalization processing are independently performed on each pulse cluster, and the phase angle information of each pulse is accurately obtained in combination with the power frequency synchronous clock. The corresponding normalized amplitude is calculated based on its pulse energy integration interval, thereby generating PRPD and PRPS spectra with high resolution and physical consistency. This method not only effectively preserves the statistical differences between different discharge sources in the phase and amplitude dimensions, enhancing the visual distinction between various discharge modes, but also avoids the spectrum blurring problem caused by aliasing signal interference in traditional methods, significantly improving the accuracy and reliability of partial discharge type identification, and providing high-quality, structured feature input for subsequent deep learning-based intelligent diagnosis.

[0043] In order to improve the robustness of the clustering process and the accuracy of pulse cluster division, in a further preferred embodiment, step c2 specifically includes: The Gaussian kernel function is used to perform weighted calculation on the local density of each pulse point in the three-dimensional energy feature space, that is, to calculate the density value of each pulse point so that the density value can more realistically reflect the aggregation characteristics of the point cloud distribution: ; Based on density value Arrange the pulse points from high to low, and list the pulse points with density values ​​greater than the minimum density threshold as points to be allocated. This step sorts the points from high to low according to the density value and filters out the points to be allocated that are higher than the minimum density threshold.

[0044] The screening conditions for the points to be allocated are: .

[0045] Take the current highest density point As the center of mass, absorb its For all points to be assigned in the neighborhood, recursively perform neighborhood absorption on the newly included points. When no new points are added, a pulse cluster is generated. Next, the new centroid with the highest density value is found to generate the next pulse cluster. This step ensures that homologous pulses belong to the same pulse cluster through density priority and topological continuity, ensuring that subsequent clustering focuses on valid data areas and improving the algorithm's ability to suppress noise.

[0046] This embodiment uses the above-mentioned technical means to gradually generate pulse clusters with clear structures and clear boundaries, which not only enhances the ability to separate complex aliased signals, but also improves the stability and physical interpretability of clustering results, providing a high-quality data foundation for subsequent spectrum reconstruction and discharge type identification. In order to improve the accuracy and physical consistency of the atlas reconstruction, in a further preferred embodiment, step c4 specifically includes: Lock the power frequency cycle zero point corresponding to all pulse clusters based on the synchronous clock , and based on pulse clusters The phase angle corresponding to each pulse cluster is calculated by the time difference between the pulse leading edge moment of the corresponding pulse unit and the zero point of the power frequency cycle, thereby achieving high-precision capture of the discharge phase information: ,in is the power frequency zero point, is the pulse leading edge moment, is the power frequency period.

[0047] At the same time, in the pulse energy integration interval Extract the peak amplitude of the signal of each pulse cluster: ; Normalization calibration is performed based on factors such as sensor distance to obtain a comparable normalized amplitude set: ,in is the distance from the sensor to the discharge source, that is, the distance from the PD pulse receiving position to the discharge position.

[0048] This embodiment effectively eliminates amplitude deviations caused by differences in signal propagation paths or detection positions, enhances the feature differentiation of various types of partial discharge signals in PRPD and PRPS spectra, and provides more stable, accurate, and interpretable data support for subsequent intelligent recognition based on spectra features.

[0049] In order to further improve the feature expression capability and visualization effect of the PD pulse cluster, in a further preferred embodiment, in step c5, generating the PRPD spectrum and PRPS spectrum corresponding to the pulse cluster specifically includes: By independently generating PRPD and PRPS spectra for a single pulse cluster, a refined characterization in the phase-amplitude space is achieved. Specifically: Based on the phase angle set corresponding to a single pulse cluster and the normalized amplitude set , with phase angle set is the horizontal axis, normalized amplitude set Establish a two-dimensional coordinate system for the vertical axis and place the single pulse cluster All points within The three-point distribution is drawn as an example, and the corresponding PRPD spectrum is generated to clearly reflect the phase distribution law and amplitude statistical characteristics of the discharge event within the power frequency cycle.

[0050] At the same time, based on the zero point of the power frequency cycle, the corresponding cycle number is calculated for each pulse unit in a single pulse cluster: ; Establish a phase angle set X-axis, period number Y-axis, normalized amplitude set The Z-axis is a three-dimensional coordinate system, and the PRPS spectrum is generated by color gradient mapping the amplitude intensity. For example, blue, yellow, and red correspond to low, medium, and high respectively, thereby realizing the dynamic evolution display of the discharge behavior in the time dimension.

[0051] This embodiment not only enhances the separability and recognizability of different types of discharge sources on the spectrum, but also provides a data foundation with a clear structure and physical meaning for subsequent intelligent classification and diagnosis based on spectrum features.

[0052] It should be understood that the embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A method for clustering and separating partial discharge pulses based on multi-band energy three-dimensional mapping, characterized in that: include: a. collecting broadband signals corresponding to a plurality of pulse units in a partial discharge pulse, and determining the pulse energy integration interval of the corresponding pulse unit based on the broadband signals; b. within a single pulse energy integration interval, dividing the corresponding broadband signal into three sub-signals based on three preset characteristic frequency bands, and performing bandpass filtering on the three sub-signals in parallel and calculating energy scalar values ​​for each sub-signal; Establishing three-dimensional coordinates of corresponding pulse units based on the energy scalar values ​​corresponding to the three sub-signals, and mapping the corresponding pulse units into pulse points in the three-dimensional energy feature space according to the three-dimensional coordinates; c. clustering and separating all pulse points in the three-dimensional energy feature space to form pulse clusters based on an adaptive density clustering algorithm; Perform spectral reconstruction on all pulse clusters.

2. The method for clustering and separating partial discharge pulses according to claim 1, characterized in that: The step a comprises: Detecting the rising edge of the pulse of the broadband signal to determine the starting time of the pulse unit corresponding to the pulse; Taking the starting moment as the interception starting point, the time period covering the effective pulse width of the pulse unit is intercepted as the pulse energy integration interval of the corresponding pulse unit.

3. The method for clustering and separating partial discharge pulses according to claim 1, characterized in that: In the step b, there are overlapping frequency regions between the three preset characteristic frequency bands.

4. The method for clustering and separating partial discharge pulses according to claim 1, characterized in that: The step b specifically includes: b1. Performing bandpass filtering on the three sub-signals in parallel to obtain filtered signals, respectively calculating the energy integral values ​​corresponding to the filtered signals within the pulse energy integration intervals of the three sub-signals, and converting the energy integral values ​​into energy scalar values ​​corresponding to the three sub-signals; b2. Perform logarithmic transformation on the energy scalar values ​​corresponding to the three sub-signals respectively, and arrange the logarithmic transformation results as the three-dimensional coordinates of the pulse units corresponding to the three sub-signals.

5. The method for clustering and separating partial discharge pulses according to claim 4, characterized in that: Also includes: b11. In the corresponding pulse energy interval, the amplitudes corresponding to the three sub-signals are extracted respectively, and the squares of the absolute values ​​of the single amplitudes are continuously accumulated to obtain the corresponding energy integral value; Multiplying the energy integral value by the conversion coefficient to obtain a corresponding energy scalar value; b21. Perform logarithmic transformation on the energy scalar values ​​to obtain a set of corresponding transformation results, arrange the transformation results according to the dimensions of the three-dimensional energy feature space, and obtain the three-dimensional coordinates of the corresponding pulse units, so as to map the corresponding pulse units into pulse points in the three-dimensional energy feature space according to the three-dimensional coordinates.

6. The method for clustering and separating partial discharge pulses according to claims 1 to 5, characterized in that: In step c, clustering and separating all pulse points to form pulse clusters based on an adaptive density clustering algorithm includes: c1. Calculate the Euclidean distance set of the corresponding pulse points based on all three-dimensional coordinates, and determine the neighborhood radius corresponding to all pulse points based on the median of the Euclidean distance set; obtain the total number of pulse points and calculate the minimum density threshold of all pulse points; c2. Calculate the density value of each pulse point, and divide all the pulse points into points to be allocated based on the density value and the minimum density threshold; select the centroid from the points to be allocated based on the density value, and determine the pulse cluster based on the centroid and the neighborhood radius corresponding to the centroid; until all the points to be allocated have been allocated to the corresponding pulse cluster; c3. For two adjacent pulse clusters, calculate the Euclidean distance between the two corresponding centroids. If the Euclidean distance between the two centroids is less than the neighborhood radius corresponding to the two centroids, merge the two pulse clusters.

7. The method for clustering and separating partial discharge pulses according to claim 6, characterized in that: In step c, performing spectrum reconstruction on all pulse clusters includes: c4. Calculate the phase angle set corresponding to all pulse clusters under the synchronous clock; calculate the normalized amplitude set corresponding to all pulse clusters based on the relationship between all pulse clusters and the corresponding pulse energy integration interval; c5. Reconstructing the PRPD spectrum and PRPS spectrum of the corresponding pulse cluster based on the phase angle set and the normalized amplitude set.

8. The method for clustering and separating partial discharge pulses according to claim 6, characterized in that: The step c2 specifically includes: Calculating the density value of each pulse point using a Gaussian kernel function, arranging all pulse points from high to low based on the density value, and listing the pulse points whose density value is greater than the minimum density threshold as points to be allocated; Taking the point to be allocated with the highest current density value as the centroid, recursively absorbing all the points to be allocated within the neighborhood radius corresponding to the centroid to form a pulse cluster; The identities of all pulse points in the formed pulse cluster as points to be allocated are cancelled.

9. The method for clustering and separating partial discharge pulses according to claim 7, characterized in that: The step c4 specifically includes: Lock the power frequency cycle zero point corresponding to all pulse clusters based on the synchronous clock, obtain the phase angle corresponding to each pulse cluster based on the time difference between the pulse leading edge moment of the corresponding pulse unit in the pulse cluster and the power frequency cycle zero point, and generate a phase angle set corresponding to all pulse clusters; The signal peak amplitude within the pulse energy integration interval corresponding to the pulse cluster is extracted, and the signal peak amplitude is calibrated to obtain a normalized amplitude set corresponding to all pulse clusters.

10. The method for clustering and separating partial discharge pulses according to claim 9, characterized in that: In step c5, reconstructing the PRPD spectrum and PRPS spectrum corresponding to the pulse cluster specifically includes: Based on the phase angle set and the normalized amplitude set corresponding to the single pulse cluster, a two-dimensional coordinate system is established with the phase angle set as the horizontal axis and the normalized amplitude set as the vertical axis, and a PRPD spectrum corresponding to the single pulse cluster is generated by distributing discrete points in the two-dimensional coordinate system; Based on the zero point of the power frequency cycle, the corresponding period number is calculated for each pulse unit in a single pulse cluster, and a three-dimensional coordinate system is established with the phase angle set as the X-axis, the period number as the Y-axis, and the normalized amplitude set as the Z-axis. The PRPS spectrum corresponding to the single pulse cluster is generated by color gradient mapping the amplitude intensity.

Images