Method for removing partial discharge pulse interference noise based on density clustering

Abstract

The invention relates to the technical field of power equipment monitoring and noise suppression, in particular to a method for removing partial discharge pulse interference noise based on density clustering, and the method comprises the following steps: S1, collecting a noisy discharge signal, and carrying out the preprocessing of the noisy signal; phase-amplitude characteristics of the noisy signals are extracted, and a phase-amplitude distribution spectrogram is formed; s2, performing primary density clustering analysis on data points in the phase-amplitude distribution spectrogram to obtain a preliminary noise point set; s3, performing secondary density clustering analysis on the initial noise point set to obtain a final noise point set; and S4, performing denoising processing on the noisy discharge signal according to the final noise point set, and performing reverse normalization and reconstruction to obtain a denoised effective discharge signal. According to the technical scheme, the noise in the partial discharge signal can be efficiently and accurately recognized and removed, and the error caused by the fact that a threshold value is manually set in a traditional method is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring and noise suppression, and in particular to a method for removing partial discharge pulse interference noise based on density clustering. Background Art

[0002] Partial discharge (PD) is a key sign of insulation cracking during long-term operation of electrical equipment. It is a non-penetrating discharge phenomenon caused by local defects in the insulating medium. Local defects such as bubbles, cracks, suspended metal particles, and electrode burrs generated during the manufacturing and operation of high-voltage power equipment can cause PD failures in electrical equipment under certain operating conditions. If PD persists for a long time in the insulation structure of electrical equipment, the dielectric properties of the insulation may be severely damaged. If the PD failure remains undetected and untreated, it may ultimately lead to catastrophic failure of the electrical equipment. Ultra-high frequency (UHF) detection is currently an effective method for PD detection. This method uses antenna sensors to receive electromagnetic waves radiated by the PD process and detect the PD signal. However, external UHF PD detection systems are highly susceptible to interference from various signals in the external electromagnetic environment during detection. The probability of pulse interference signals in substations is as high as 67%. The mixing of noise signals complicates the original discharge signal, resulting in a decrease in the accuracy of the discharge information.

[0003] CN202211167595.0 discloses a local discharge pattern recognition method based on the combination of Kmeans and DBSCAN, which includes the following steps: data acquisition, continuously acquiring time domain data of a local discharge sensor, performing real-time processing on the acquired time domain data, and generating a PRPD spectrum; performing envelope waveform discharge / noise calculation on the PRPD spectrum; for a PRPD spectrum judged to be a discharge, performing white noise filtering on the PRPD spectrum based on the Kmeans method; clustering the filtered PRPD spectrum into PD clusters based on the hierarchical density clustering method, and performing feature extraction on the clustered clusters; establishing a model, training the model, and obtaining the credibility of the recognition model; inputting the measured data of the local discharge fault into the trained model for recognition, and obtaining the final recognition result. However, this existing technology has the following technical defects: First, the Kmeans algorithm requires the pre-setting of the number of clusters, which is difficult to accurately determine in practical applications, especially when the local discharge signal pattern is complex and changeable; second, the Kmeans algorithm is sensitive to noise and outliers, and it is easy to misjudge weak local discharge signals as noise and over-filter them, resulting in the loss of effective discharge information; in addition, the random pulse interference time is short and similar to the local discharge waveform. In the time domain, it appears as a pulse signal with a very short duration, and in the frequency domain, it appears as a broadband signal containing multiple frequency domain components, with time domain and frequency domain characteristics similar to the local discharge signal. The existing method based on simple clustering combination is difficult to effectively distinguish such highly similar interference signals, affecting the accuracy and reliability of local discharge detection. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems existing in the background technology, the present invention provides a method for removing partial discharge pulse interference noise based on density clustering, which can effectively remove noise interference in the partial discharge signal of power equipment and retain the effective partial discharge signal, thereby improving the accuracy of equipment fault diagnosis.

[0005] The technical solution of the present invention is implemented as follows: The present invention provides a method for removing partial discharge pulse interference noise based on density clustering, comprising the following steps:

[0006] S1. Collecting a noisy discharge signal, preprocessing the noisy signal, and extracting the phase-amplitude characteristics of the noisy signal to form a phase-amplitude distribution spectrum;

[0007] S2. Perform a first-level density cluster analysis on the data points in the phase-amplitude distribution spectrum to obtain a preliminary noise point set;

[0008] S3, performing secondary density cluster analysis on the preliminary noise point set to obtain the final noise point set;

[0009] S4. De-noising the noisy discharge signal according to the final set of noise points, and performing denormalization and reconstruction to obtain a de-noised effective discharge signal.

[0010] On the basis of the above technical solution, preferably, in step S1, the preprocessing of the noisy signal includes: filtering, amplifying and normalizing the collected original pulse signal, and mapping the phase and amplitude characteristics of the signal to a unified numerical range through normalization.

[0011] Based on the above technical solution, preferably, step S2 specifically includes the following steps:

[0012] S21. Setting the primary neighborhood radius and the minimum number of basic neighborhood points, and calculating the phase-sensitive distance based on the phase-amplitude distribution spectrum;

[0013] S22, performing adaptive core point determination on each data point based on the amplitude characteristics of the data point;

[0014] S23, marking the data points that meet the adaptive core point determination conditions as core objects, and grouping all data points that can be connected to the core objects into the same cluster through density reachability relationships to form a main partial discharge signal cluster;

[0015] S24 , taking all formed main partial discharge signal clusters as valid signals, and outputting data points that are not classified into any main partial discharge signal cluster as a preliminary noise point set.

[0016] Based on the above technical solution, preferably, in step S21, the calculation formula of the phase sensitive distance is:

[0017]

[0018] Where p i and p j is the data point in the phase-amplitude distribution spectrum; d PD (p i ,p j ) is the data point p i and p j The phase sensitive distance between i 、φ j is the data point p i and p j The normalized phase value, A i 、A j is the data point p i and p j is the normalized amplitude of , κ is the phase-amplitude weight ratio coefficient, and λ is the phase sensitivity enhancement coefficient.

[0019] Based on the above technical solution, preferably, in step S22, the conditions for adaptive core point determination are:

[0020]

[0021] Where, is the data point p i The set of neighbor points within the first-level neighborhood radius, The number of points in the neighbor point set; min _ samples1 is the minimum number of neighborhood points, μ is the amplitude attenuation adjustment coefficient, A ref is the reference amplitude; A i is the data point p i The normalized magnitude of .

[0022] Based on the above technical solution, preferably, step S3 specifically includes:

[0023] S31, creating an independent initial single point cluster for each data point in the preliminary noise point set to form an initial cluster set;

[0024] S32, calculating the similarity between any two clusters in the initial cluster set, wherein the similarity comprehensively considers the geometric distance between clusters and phase consistency;

[0025] S33, performing progressive cluster merging based on similarity, where the merging determination condition is that the similarity between clusters is greater than a set merging threshold and the total number of points of the merged cluster does not exceed a set maximum point limit, to form a candidate recovery cluster set;

[0026] S34: Perform feature verification on each cluster in the candidate recovery cluster set, calculate the comprehensive confidence, and determine the clusters with confidence lower than the set threshold as noise to form the final noise point set.

[0027] Based on the above technical solution, preferably, in step S32, the similarity calculation formula is:

[0028]

[0029] Where c i ,c j are two clusters in the initial cluster set; Similarity(c i ,c j ) is cluster c i and cluster c j The similarity between them is in the range of [0,1]; d min (c i ,c j ) is cluster c i and cluster c j The minimum phase sensitive distance betweenlocal is the local scale parameter; w phase (c i , c j ) is the phase consistency weight based on the average phase difference; Cluster c i and cluster c j The average phase value of σ φ is the tolerance parameter of phase difference.

[0030] Based on the above technical solution, preferably, in step S34, performing feature verification on each cluster in the candidate recovery cluster set specifically includes:

[0031] S341. For all candidate recovery clusters, calculate the cluster compactness ratio and phase consistency index;

[0032] S342: Classify candidate recovery clusters according to their point size, identifying clusters with fewer points than a set threshold as small clusters and clusters with no fewer points than the set threshold as large clusters, and further calculate local support indices for the large clusters;

[0033] S343. Calculate the comprehensive confidence based on the cluster classification result, compare the calculated comprehensive confidence with the set confidence threshold, and determine the cluster with a confidence higher than the threshold as a valid recovery cluster, and determine the cluster with a confidence lower than the threshold as a noise cluster.

[0034] On the basis of the above technical solution, preferably, in step S343, the comprehensive confidence is calculated based on the classification results of the cluster, including: the comprehensive confidence of the small cluster adopts a weighted combination of the compactness ratio and phase consistency, and the comprehensive confidence of the large cluster adopts a triple weighted combination of the compactness ratio, phase consistency and local support.

[0035] Based on the above technical solution, preferably, step S4 specifically includes:

[0036] S41, merging the main partial discharge signal cluster and the effective recovery cluster to form a complete effective partial discharge signal point set;

[0037] S42, performing denormalization processing on the valid signal point set, and restoring the normalized phase and amplitude data to the original physical dimensions;

[0038] S43, performing adaptive filtering enhancement on the original noisy discharge signal based on the phase-amplitude characteristics of the effective signal points, highlighting the effective discharge signal and suppressing the remaining noise component by phase window selection and amplitude threshold adjustment;

[0039] S44. Reconstruct the enhanced signal in the time domain, restore the time series characteristics of the discharge pulse according to the phase information, and obtain the effective discharge signal after denoising.

[0040] The method of removing partial discharge pulse interference noise based on density clustering of the present invention has the following beneficial effects compared with the prior art:

[0041] (1) The present invention establishes an adaptive density clustering algorithm system based on phase-sensitive distance, combined with a hierarchical processing mechanism of first-level coarse separation and second-level precise recovery, to achieve accurate identification and effective removal of pulse interference noise in partial discharge signals. By removing noise and retaining effective partial discharge signals, the accuracy of equipment fault diagnosis and early warning is significantly improved. Compared with traditional frequency domain filtering and threshold determination methods, the present invention does not require manual setting of complex parameters and can adapt to different noise environments and signal characteristics, effectively solving the technical problems of strong parameter dependence and low denoising accuracy in traditional methods. The present invention is not only suitable for denoising partial discharge signals, but can also be widely used in other types of signal denoising, such as health monitoring of electrical equipment, industrial detection and other fields.

[0042] (2) The two-level hierarchical clustering analysis technology established in the present invention achieves refined processing of the first-level clustering results through similarity-driven progressive cluster merging and multi-dimensional feature verification. It can effectively recover weak local discharge signals that are mistakenly identified as noise in the first-level clustering due to insufficient density. At the same time, the reliability of the recovered signal is ensured through comprehensive verification of compactness ratio, phase consistency and local support, avoiding erroneous recovery of noise and significantly improving the accuracy and robustness of the overall denoising algorithm.

[0043] (3) The hierarchical confidence calculation mechanism provided by the present invention adopts differentiated verification strategies for candidate recovery clusters of different sizes. It mainly relies on the intrinsic phase consistency characteristics to make judgments on small clusters, and adds local support verification to large clusters to provide more comprehensive statistical information. It effectively solves the problem of statistical instability of small samples, improves the accuracy of small-scale signal cluster judgment through the small sample correction factor, and ensures the statistical reliability and engineering applicability of the verification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 Flowchart of the method for removing partial discharge pulse interference noise based on density clustering of the present invention;

[0046] Figure 2 The collected pulse noise signal and partial discharge signal diagram of the present invention;

[0047] Figure 3 is a phase-amplitude distribution diagram of the normalized mixed signal of the present invention;

[0048] Figure 4 This is the phase-amplitude distribution diagram after PRPD analysis of the present invention;

[0049] Figure 5 A distribution diagram of the phase-amplitude results of the mixed signal obtained by the first-level density cluster analysis of the present invention;

[0050] Figure 6 The phase-amplitude distribution diagram of the mixed signal after the noise data is removed according to the present invention;

[0051] Figure 7 is a phase-amplitude distribution diagram of the denoised signal after denormalization of the present invention;

[0052] Figure 8 This is a three-dimensional spectrum of the PRPD phase-amplitude-discharge number of the discharge signal after denoising of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, the present invention provides a method for removing partial discharge pulse interference noise based on density clustering, comprising the following steps:

[0055] S1. Collecting a noisy discharge signal, preprocessing the noisy signal, and extracting the phase-amplitude characteristics of the noisy signal to form a phase-amplitude distribution spectrum;

[0056] S2. Perform a first-level density cluster analysis on the data points in the phase-amplitude distribution spectrum to obtain a preliminary noise point set;

[0057] S3, performing secondary density cluster analysis on the preliminary noise point set to obtain the final noise point set;

[0058] S4. De-noising the noisy discharge signal according to the final set of noise points, and performing denormalization and reconstruction to obtain a de-noised effective discharge signal.

[0059] The present invention achieves accurate identification and effective removal of local discharge pulse interference noise by establishing a two-level density clustering analysis architecture. The first-level density clustering analysis can quickly identify the main local discharge signal clusters and preliminarily separate suspicious noise points; the second-level density clustering analysis performs fine processing on the preliminary noise points. Through similarity calculation and feature verification mechanisms, it can effectively recover misjudged weak local discharge signals, avoiding the problem of traditional single-level clustering methods easily deleting weak discharge signals by mistake. Through multi-level noise verification and signal recovery mechanisms, the present invention can effectively remove random pulse interference while retaining true local discharge information to the greatest extent, significantly improving the signal-to-noise ratio and accuracy of local discharge detection.

[0060] Furthermore, in S1, if Figure 2 As shown, Figure 2 (a) is the impulse noise signal, Figure 2 (b) is a typical partial discharge signal. A high-frequency electromagnetic wave sensor is used to collect the partial discharge signal of the power equipment. The sampling frequency is in the MHz range, and the sampling time is 50 power frequency cycles, or 1 second. The noise signal is added to the discharge signal to obtain the noisy partial discharge signal. When collecting the signal, simple noise can be basically filtered out by setting the filter and sampling rate. The partial discharge signal of the power equipment is collected in real time using a high-precision sensor. The sampling time is 1 second. During the collection process, the partial discharge signal of the power equipment is usually affected by environmental noise, equipment noise, pulse interference generated by switching operations, and other electromagnetic interference signals. These interference signals have similar characteristics to the partial discharge signal in the time domain and frequency domain, which brings difficulties to subsequent signal analysis.

[0061] The preprocessing of the collected noisy signal includes filtering, amplification and normalization. The filtering process uses a bandpass filter. When collecting signals, the setting of the filter and the sampling rate can basically filter out simple noise and remove low-frequency power frequency interference and high-frequency noise components in the collected signal. The collected original pulse signal is processed to remove DC offset, bandpass filtering, low-frequency noise removal, etc. to ensure the improvement of signal quality. The amplification process is achieved through a variable gain amplifier, and the amplification factor is adaptively adjusted according to the signal strength to ensure that weak partial discharge signals can be effectively detected. The normalization process uses the Min-Max normalization method. Through normalization, the phase and amplitude characteristics of the signal are mapped to a unified numerical interval, and the data is normalized to the [0,1] interval. The specific implementation path is:

[0062]

[0063] Where x max is the maximum value of the sample data, xmin is the minimum value of the sample data, and the normalized result is as follows Figure 3 As shown, it is beneficial to the clustering of data.

[0064] After completing the signal preprocessing, it is necessary to extract the phase-amplitude characteristic parameters from the preprocessed signal. Perform phase-amplitude distribution (PRPD) analysis on the collected noisy discharge signal, and represent the amplitude and phase information of the signal in the form of a two-dimensional graph to form an amplitude-phase two-dimensional spectrum. The extraction of phase characteristics is achieved by calculating the phase angle of each local discharge pulse relative to the power frequency voltage. The phase angle reflects the voltage phase position where the local discharge occurs. The extraction of amplitude characteristics is achieved by detecting the peak amplitude of each local discharge pulse. The amplitude reflects the intensity of the local discharge. Through the PRPD graph, the regular characteristics of the local discharge signal can be identified more intuitively. The instantaneous amplitude and phase of the signal are extracted, and the local discharge signal is converted into a feature vector by combining time domain analysis and frequency domain analysis to form a data basis. Figure 4 As shown, Figure 4 is the phase-amplitude distribution diagram obtained through PRPD analysis and signal extraction, where Figure 4 (a) is the phase-amplitude distribution diagram of the impulse noise signal, Figure 4 (b) is the phase-amplitude distribution diagram of a typical partial discharge signal. Figure 4 (c) is the phase-amplitude distribution diagram of the mixture of noise and discharge signal.

[0065] Based on the extracted phase-amplitude characteristic parameters, a PRPD diagram is constructed. There are significant differences between the noise signal and the local discharge signal data in the phase-amplitude two-dimensional spectrum. The noise signal is irregularly scattered across the entire phase and amplitude range, while the local discharge signal is concentrated together to form dense clusters. The PRPD diagram uses the horizontal axis to represent the phase angle and the vertical axis to represent the discharge amplitude. Each data point in the diagram represents the phase-amplitude characteristic of a local discharge event. To facilitate subsequent density clustering analysis, the PRPD diagram is appropriately gridded to ensure that it can fully reflect the distribution characteristics of the local discharge signal. During the gridding process, the number of discharges in each grid is counted to form a two-dimensional distribution diagram with the number of discharges as the weight.

[0066] Furthermore, in step S2, accurate identification of the partial discharge signal and preliminary noise separation in the phase-amplitude distribution spectrum are achieved through phase-sensitive distance measurement and adaptive core point determination mechanism. Specifically, the following steps are included:

[0067] S21. Set the primary neighborhood radius and the minimum number of basic neighborhood points, and calculate the phase-sensitive distance based on the phase-amplitude distribution spectrum.

[0068] Specifically, in one embodiment, the value range of the primary neighborhood radius is usually set between 0.08 and 0.12. The specific value is adjusted according to the distribution density of the local discharge signal in the PRPD spectrum. For a denser signal distribution, the value can be appropriately reduced, and for a sparser signal distribution, the value can be appropriately increased. The typical value range of the basic minimum number of neighborhood points is 4 to 8. This parameter determines the minimum density requirement for forming a core point. If the value is too small, noise may be mistaken for a signal, and if the value is too large, weak local discharge signals may be lost. Figure 5 As shown, the clustering results when the first-level neighborhood radius is set to 0.1 and the basic minimum number of neighborhood points is 4 are as follows Figure 5 As shown, effective discharge clusters and noise points in the signal can be identified.

[0069] The calculation formula of phase sensitive distance is:

[0070]

[0071] Where p i and p j is the data point in the phase-amplitude distribution spectrum; d PD (p i ,p j ) is the data point p i and p j The phase sensitive distance between i 、φ j is the data point p i and p j The normalized phase value of A is in the range of [0,2π]; i 、A j is the data point p i and p j The normalized amplitude of the phase difference is in the range of [0, 1]. κ is the phase-amplitude weight ratio, which adjusts the relative importance of phase and amplitude features in distance calculation, with a typical range of [1.0, 2.0]. λ is the phase sensitivity enhancement factor, which enhances the effect of phase differences on distance calculation, with a range of [0.1-0.5]. By combining the basic distance with the phase sensitivity enhancement term through multiplication, the distance between data points with large phase differences is significantly amplified. Compared with the traditional Euclidean distance, this distance metric better reflects the true distribution characteristics of partial discharge signals in PRPD spectra and improves the accuracy of cluster analysis.

[0072] S22. Adaptively determine the core point of each data point based on the amplitude characteristics of the data point.

[0073] The conditions for adaptive core point determination are:

[0074]

[0075] Where, is the data point p i The set of neighbor points within the first-level neighborhood radius, is the number of points in the neighbor point set, that is, the number of data points that actually exist within the neighborhood radius; min _ samples1 is the minimum number of neighborhood points as the benchmark threshold for density clustering; μ is the amplitude attenuation adjustment coefficient, which controls the influence of amplitude on the core point judgment condition; A ref is the reference amplitude; A i is the data point p i The normalized amplitude is usually set to the median or 75th percentile of the normalized amplitudes of all points in the data set, which is used to normalize the exponential decay function.

[0076] For high amplitude partial discharge signals (A i Large), due to its strong signal strength, even if the number of neighboring points around it is relatively small, it should be identified as a valid core point. At this time, the exponential term Approaching 0, the right judgment threshold is reduced, making it easier to meet the core point condition; for low-amplitude partial discharge signals (A i Small), because its signal is weak and easily interfered by noise, a higher neighborhood density requirement is required to ensure its reliability. The threshold approaches 1, maintaining a relatively high decision threshold. Signal points with medium amplitudes are linearly interpolated between the two extremes based on their specific amplitudes. This adaptive mechanism fully accounts for the intensity differences of partial discharge signals, avoiding the potential problems of over-clustering of strong signals or misjudgment of weak signals that can occur with fixed threshold strategies. This improves the accuracy of the density clustering algorithm for identifying partial discharge signals of varying intensities.

[0077] S23. Mark the data points that meet the adaptive core point determination conditions as core objects, and classify all data points that can be connected to the core objects into the same cluster through density reachability relationships to form a main partial discharge signal cluster.

[0078] The establishment of density reachability relationship follows the following rules: for a core object, all data points within its neighborhood radius are defined as directly density reachable from the core object; if the data point p j Direct density reachable from the core object, and p j It is also a core object, then p j All points in the neighborhood are indirectly density-reachable from the core object. Through this transitive relationship, all data points that are mutually density-reachable can be classified into the same cluster.

[0079] The cluster formation process includes: starting with an unvisited core object, creating a new cluster and adding the core object to the cluster; searching all neighboring points of the core object and adding points that are directly density-reachable to the current cluster; recursively searching the neighboring points of the newly added core object to continue expanding the current cluster; repeating this process until the current cluster can no longer be expanded; and finally selecting the next unvisited core object to begin building a new cluster. This process forms multiple independent primary PD signal clusters.

[0080] During cluster formation, we also need to consider the handling of border points. Border points are data points that do not meet the core point criteria but are located within the neighborhood of a core point. While border points cannot serve as starting points for cluster expansion, they can be included in the corresponding cluster. If a border point is located within the neighborhood of multiple core points simultaneously, it is assigned to the cluster that was visited first. This approach ensures that each data point can belong to at most one cluster, avoiding overlap between clusters.

[0081] Through the establishment of the density reachability relationship and the cluster formation process, the data points with similar phase-amplitude characteristics and dense distribution in the PRPD spectrum can be effectively clustered together to form the main partial discharge signal cluster set C main ={c1, c2, ..., c k}.

[0082] S24 , taking all formed main partial discharge signal clusters as valid signals, and outputting data points that are not classified into any main partial discharge signal cluster as a preliminary noise point set.

[0083] Data points that are not classified into any major PD signal clusters constitute a preliminary set of noise points. These points may fall into two categories: true random noise points, which are randomly distributed in the PRPD spectrum and lack the clustering characteristics of PD signals; and misidentified weak PD signal points, which may not be correctly identified by the first-level density clustering due to weak signal intensity, sparse distribution, or distance metric deviation.

[0084] To ensure the quality of the first-level clustering results, a basic validity test is required for the main signal clusters formed. The validity test includes two aspects: cluster size test and cluster compactness test. The cluster size test requires that the number of points in each main signal cluster is no less than the set minimum cluster size threshold, usually set at 3 to 5 points. Clusters that are too small may be pseudo clusters formed by noise aggregation. The cluster compactness test evaluates the internal compactness of the cluster by calculating the average distance between data points within the cluster. Clusters with excessively large average distances may contain different types of signals and require further inspection. For pseudo clusters that fail the validity test, the data points therein are reclassified into the preliminary noise point set. For valid clusters that pass the test, they are confirmed as the main partial discharge signal clusters. Through this quality control mechanism, the reliability of the first-level density clustering results can be improved, providing high-quality input data for subsequent second-level refinement processing.

[0085] Furthermore, step S3 uses a two-level hierarchical clustering analysis technique to refine the preliminary noise point set output from step S2, and achieves effective recovery of weak partial discharge signals and accurate identification of real noise through similarity-driven progressive cluster merging and multi-dimensional feature verification mechanisms. Specifically, it includes:

[0086] S31. Create an independent initial single point cluster for each data point in the preliminary noise point set to form an initial cluster set. For the preliminary noise point set output in step S2, where each data point includes a normalized phase value and amplitude, create an independent initial single point cluster for each data point in the set to form an initial cluster set.

[0087] S32 , calculating the similarity between any two clusters in the initial cluster set, wherein the similarity comprehensively considers the geometric distance between clusters and the phase consistency.

[0088] For any two clusters c in the initial cluster combination i and c j , calculate their similarity, the similarity calculation formula is:

[0089]

[0090]

[0091] Where c i ,c j are two clusters in the initial cluster set; Similarity(c i ,c j ) is cluster c i and cluster c j The similarity between them is in the range of [0,1]; d min (c i ,c j ) is cluster ci and cluster c j The minimum phase sensitive distance between local It is a local scale parameter used to control the influence of distance on similarity. It is usually set to the standard deviation of the distances between all point pairs in the preliminary noise point set. The specific calculation method is:

[0092]

[0093] in is the average distance between all point pairs, n is the number of points in the initial noise point set, d PD (p i ,p j ) is the data point p i and p j The phase sensitive distance between them.

[0094] w phase (c i ,c j ) is the phase consistency weight based on the average phase difference; and Cluster c i and c j The average phase value, for a single-point cluster, is the phase value of the point; for a multi-point cluster, it is the arithmetic average of the phase values ​​of all points in the cluster; σ φ is the tolerance parameter of phase difference, which controls the influence of phase difference on similarity.

[0095] The above similarity calculation is mainly composed of the geometric distance term and the phase consistency weight term. The geometric distance term ensures that only clusters that are spatially close enough can be merged, avoiding the incorrect aggregation of points that are too far away. The phase consistency weight term reflects the physical property of partial discharge signals clustering in specific phase regions. When the average phase difference between two clusters is small, the phase weight term is close to 1, and the similarity is mainly determined by the geometric distance. When the phase difference is large, the phase weight term is significantly reduced, and even clusters with a close geometric distance find it difficult to achieve high similarity.

[0096] S33, performing progressive cluster merging based on similarity, wherein the merging judgment condition is that the similarity between clusters is greater than the set merging threshold and the total number of points of the merged cluster does not exceed the set maximum point limit, to form a candidate recovery cluster set.

[0097] Specifically, calculate the similarity of all current cluster pairs and construct a similarity matrix; then sort the cluster pairs in descending order of similarity, set the merge similarity threshold and the maximum number of points in a single cluster as the merge judgment condition, and then start the iterative merging process, starting from the cluster pair with the highest similarity and checking them one by one. For the currently checked cluster pair, determine whether it meets both the similarity constraint and the scale constraint. If it meets the merge condition, perform the merge operation and merge the two clusters into a new cluster containing all the data points of the original two clusters. Among them, the similarity constraint condition is: Similarity(c i ,c j )>T merge , where T merge To combine similarity threshold, the value is usually 0.6 to 0.8. In practical applications, the threshold can be optimized and adjusted according to the partial discharge characteristics of different types of power equipment. The scale constraint condition is: |c i |+|c j |≤N max , where |c i | and |c j |represent cluster c respectively i and cluster c j The number of points, N max The maximum number of points in a single cluster is limited, and the value range is 6-10.

[0098] Then update the attribute information of the new cluster, including calculating the weighted average phase and weighted average amplitude of the new cluster. The specific calculation formula is as follows:

[0099]

[0100] Where, is the average phase value of the new cluster, is the average amplitude of the new cluster, For cluster c i The average phase value of For cluster c j The average phase value of For cluster c i The average amplitude of For cluster c j The average amplitude of .

[0101] Then, the original two clusters are removed from the cluster set, the newly formed cluster is added, and the similarity between the new cluster and all other clusters is recalculated. The similarity matrix is ​​updated, the cluster pairs are re-sorted, and the next cluster pair with the highest similarity is checked. The above merging process is repeated until no cluster pair that meets the merging conditions can be found. Finally, the merged candidate recovery cluster set is output. Each cluster in this set represents a group of weak partial discharge signals that may be misjudged by the first-level density clustering.

[0102] S34: Perform feature verification on each cluster in the candidate recovery cluster set, calculate the comprehensive confidence, and determine the clusters with confidence lower than the set threshold as noise to form the final noise point set.

[0103] Specifically, it includes:

[0104] (1) For all candidate recovery clusters, calculate the cluster compactness ratio and phase consistency index.

[0105] Among them, the compactness ratio R(c) is used to evaluate the internal compactness and external separation of the cluster, and its calculation formula is:

[0106]

[0107] Where R(c) is the compactness ratio of cluster c, A r (c) is the average distance within the cluster, which represents the average of the phase-sensitive distances between all pairs of points within the cluster; A e (c) is the average distance between clusters, which represents the average distance between the points in cluster c and the nearest points in other clusters:

[0108]

[0109] Where p i and p j are different data points in cluster c.

[0110]

[0111] Where p i ' is the data point in cluster c; p j ' are other data points that do not belong to cluster c, and |c| is the number of data points contained in cluster c.

[0112] The phase consistency index P(c) adopts the small sample corrected phase concentration evaluation method, and its calculation formula is:

[0113]

[0114] Where P(c) is the phase consistency index of cluster c; σ φ (c) is the standard deviation of the phase within the cluster, and 45° is the reference standard for phase dispersion; σ φ (c) is calculated as follows:

[0115]

[0116] Where, Used to evaluate the concentration of phase and serve as a reference standard for phase dispersion; is the small sample correction factor, φ iis the phase value of the i-th data point in cluster c; is the average phase value of cluster c.

[0117] (2) Classify candidate recovery clusters according to their point size. Those with fewer points than the set threshold are identified as small clusters, and those with no less than the set threshold are identified as large clusters. For large clusters, the local support index is further calculated.

[0118] Based on their size, candidate clusters are divided into small and large clusters. A size classification threshold is set. Small clusters are classified as clusters with fewer points than the size classification threshold. Due to the limited number of samples, these clusters are primarily verified by the compactness ratio and phase consistency, and the local support index is not calculated. Large clusters are classified as clusters with at least the size classification threshold. These clusters have relatively sufficient samples. In addition to calculating the compactness ratio and phase consistency, the local support index is also calculated to provide more comprehensive verification information. In one embodiment, the size classification threshold is preferably set to 5.

[0119] For large cluster categories, the calculation formula of the local support index L(c) is:

[0120]

[0121] The support point set N(c) is defined as the data points that are within a certain range around cluster c but do not belong to the cluster:

[0122]

[0123] Where p is the data point that does not belong to cluster c, q is the data point in cluster c, eps1 is the first-level neighborhood radius, and 1.5 is the expansion coefficient of the support range.

[0124] (3) Calculate the comprehensive confidence based on the classification results of the clusters, and compare the calculated comprehensive confidence with the set confidence threshold. Clusters with confidence higher than the threshold are confirmed as valid recovery clusters, and clusters with confidence lower than the threshold are judged as noise clusters.

[0125] The comprehensive confidence of small clusters adopts a weighted combination of compactness ratio and phase consistency, that is,

[0126] C s (c)=0.4×R(c)+0.6×P(c)

[0127] The comprehensive confidence of large clusters adopts a triple weighted combination of compactness ratio, phase consistency and local support, that is,

[0128] C l (c)=0.3×R(c)+0.4×P(c)+0.3×L(c)

[0129] A unified confidence threshold is set, typically between 0.55 and 0.65. When the calculated comprehensive confidence is higher than the threshold, the cluster is confirmed as a valid recovery cluster and added to the final set of partial discharge signal points. When the confidence is lower than the threshold, the cluster is determined to be a noise cluster, and the data points within it are included in the final set of noise points.

[0130] The above-mentioned hierarchical verification mechanism can effectively distinguish weak partial discharge signals misjudged by the first-level density clustering from real random noise, achieve accurate signal recovery and noise suppression, and ultimately form a reliable final noise point set, providing an accurate data basis for subsequent signal reconstruction.

[0131] Furthermore, in one embodiment, step S4 specifically includes:

[0132] Based on the clustering results, the effective signal clusters are identified, and the main partial discharge signal cluster set obtained in step S2 is merged with the effective recovery cluster set output in step S3 to form a complete effective partial discharge signal cluster set. Then, based on the characteristics of a large number of partial discharge signal data points and a small number of noise signal data points, cluster size statistics are performed on all clustering results, the number of data points in each cluster is calculated, and a cluster size determination threshold is set, which is usually set to 1%-3% of the total number of data points. Clusters with a number of data points greater than the threshold are retained as effective signal clusters, and clusters with a number of data points less than the threshold are determined to be noise clusters and cleared. Finally, data points in all effective signal clusters are extracted to form a complete effective partial discharge signal point set. The denoising result is as follows: Figure 6 As shown, it can be seen that the noise data is obviously removed and the discharge signal is retained.

[0133] In order to solve the signal discontinuity problem that may occur after removing noise points, the interpolation method is used to smooth the effective signal point set. First, the effective signal points are sorted according to the phase value to form an ordered phase-amplitude sequence. Then, the cubic spline interpolation method is used to fill in the areas with large phase intervals. Then, the interpolated signal is subjected to moving average filtering, and finally the effective signal point set after smoothing is obtained. The effective signal point set after smoothing is denormalized, and the normalized phase and amplitude data are restored to the original physical dimensions. Finally, the denoised partial discharge signal is output. The signal maintains the original physical dimensions and can be directly used for subsequent signal analysis and equipment status evaluation. The optimized partial discharge signal is as follows Figure 7 As shown in , it is basically consistent with the original discharge signal distribution, indicating that the noise has been significantly eliminated. In addition, the denoised signal is analyzed by performing a PRPD phase-amplitude-discharge number three-dimensional spectrum, as shown in Figure 8 As shown, the spectrum characteristics belong to typical partial discharge signals, indicating that the technical solution of the present invention can achieve the purpose of eliminating mixed pulse interference in partial discharge signals and has a significant denoising effect.

[0134] In another embodiment, step S4 specifically includes:

[0135] S41. Merge the main partial discharge signal cluster and the effective recovery cluster to form a complete effective partial discharge signal point set.

[0136] S42, perform denormalization on the effective signal point set, and restore the normalized phase and amplitude data to the original physical dimensions; wherein, the phase denormalization adopts the linear transformation φ original (i) = φ norm (i)×φ scale +φ offset , where φ original (i) is the original phase value of the i-th data point, φ norm (i) is the normalized phase value, φ scale is the phase scaling factor, φ offset is the phase offset; the amplitude denormalization adopts A original (i) = A norm (i)×A scale +A offset , where A original (i) is the original amplitude of the i-th data point, A norm (i) is the normalized amplitude, A scale is the amplitude scaling factor, A offset The inverse normalization process ensures that the effective signal point set is restored to the original physical dimension, providing a correct parameter basis for subsequent adaptive filtering enhancement.

[0137] S43, performing adaptive filtering enhancement on the original noisy discharge signal based on the phase-amplitude characteristics of the effective signal points, highlighting the effective discharge signal and suppressing the remaining noise component by phase window selection and amplitude threshold adjustment;

[0138] First, the phase window selection strategy is determined according to the phase distribution of valid signal points. The phase domain is divided into M equal-width intervals. Then, the effective signal density is calculated for each phase window. Next, a phase window selection threshold is set, which is usually the average value of all window densities. Phase windows with density higher than the threshold are marked as valid windows. At the same time, an adaptive amplitude threshold adjustment is performed according to the amplitude distribution characteristics of the valid signal points. Finally, the original noisy signal is filtered to retain the signal components within the valid phase window and with amplitude higher than the dynamic threshold, while suppressing the noise components in other areas.

[0139] S44. Reconstruct the enhanced signal in the time domain, restore the time series characteristics of the discharge pulse according to the phase information, and obtain the effective discharge signal after denoising.

[0140] First, the correspondence between phase and time is established. Then, the filtered amplitude data is sorted according to the time series to form a time domain signal. Then, the time domain interpolation method is used to resample the non-equally spaced sampling points to generate an equally spaced time domain signal. The resampling frequency is determined according to the signal bandwidth and is usually set to 5-10 times the main frequency component of the partial discharge signal. Finally, the reconstructed signal is subjected to boundary smoothing processing, and the effective partial discharge signal after denoising is finally output.

[0141] In practical applications, the appropriate denoising method can be selected based on the specific application scenario and performance requirements. For online status monitoring of power equipment, the first method is preferred to ensure real-time performance; for equipment fault analysis and defect location, the second method is preferred to improve analysis accuracy. The two methods can also be combined, using the first method for rapid preprocessing and then the second method for refined processing of important signal segments, thus achieving the optimal balance between efficiency and accuracy.

[0142] 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 in the scope of protection of the present invention.

Claims

1. A method for removing partial discharge pulse interference noise based on density clustering, characterized in that: The following steps are involved: S1. Collecting a noisy discharge signal, preprocessing the noisy signal, and extracting the phase-amplitude characteristics of the noisy signal to form a phase-amplitude distribution spectrum; S2. Perform a first-level density cluster analysis on the data points in the phase-amplitude distribution spectrum to obtain a preliminary noise point set; S3, performing secondary density cluster analysis on the preliminary noise point set to obtain the final noise point set; S4. De-noising the noisy discharge signal according to the final set of noise points, and performing denormalization and reconstruction to obtain a de-noised effective discharge signal.

2. The method for removing partial discharge pulse interference noise based on density clustering according to claim 1, characterized in that: In step S1, the preprocessing of the noisy signal includes filtering, amplifying and normalizing the collected original pulse signal, and mapping the phase and amplitude characteristics of the signal to a unified numerical range through normalization.

3. The method for removing partial discharge pulse interference noise based on density clustering according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Setting the primary neighborhood radius and the minimum number of basic neighborhood points, and calculating the phase-sensitive distance based on the phase-amplitude distribution spectrum; S22, performing adaptive core point determination on each data point based on the amplitude characteristics of the data point; S23, marking the data points that meet the adaptive core point determination conditions as core objects, and grouping all data points that can be connected to the core objects into the same cluster through density reachability relationships to form a main partial discharge signal cluster; S24 , taking all formed main partial discharge signal clusters as valid signals, and outputting data points that are not classified into any main partial discharge signal cluster as a preliminary noise point set.

4. The method for removing partial discharge pulse interference noise based on density clustering according to claim 3, characterized in that: In step S21, the calculation formula of the phase sensitive distance is: Where p i and p j is the data point in the phase-amplitude distribution spectrum; d PD (p i ,p j ) is the data point p i and p j The phase-sensitive distance between φ i 、φ j is the data point p i and p j The normalized phase value, A i 、A j is the data point p i and p j is the normalized amplitude of , κ is the phase-amplitude weight ratio coefficient, and λ is the phase sensitivity enhancement coefficient.

5. The method for removing partial discharge pulse interference noise based on density clustering according to claim 3, characterized in that: In step S22, the conditions for adaptive core point determination are: Where, is the data point p i The set of neighbor points within the first-level neighborhood radius, is the number of points in the neighbor point set; min_samples1 is the minimum number of neighborhood points, μ is the amplitude attenuation adjustment coefficient, A ref is the reference amplitude; A i is the data point p i The normalized magnitude of .

6. The method for removing partial discharge pulse interference noise based on density clustering according to claim 3, characterized in that: Step S3 specifically includes: S31, creating an independent initial single point cluster for each data point in the preliminary noise point set to form an initial cluster set; S32, calculating the similarity between any two clusters in the initial cluster set, wherein the similarity comprehensively considers the geometric distance between clusters and phase consistency; S33, performing progressive cluster merging based on similarity, where the merging determination condition is that the similarity between clusters is greater than a set merging threshold and the total number of points of the merged cluster does not exceed a set maximum point limit, to form a candidate recovery cluster set; S34: Perform feature verification on each cluster in the candidate recovery cluster set, calculate the comprehensive confidence, and determine the clusters with confidence lower than the set threshold as noise to form the final noise point set.

7. The method for removing partial discharge pulse interference noise based on density clustering according to claim 6, characterized in that: In step S32, the similarity calculation formula is: Where c i ,c j are two clusters in the initial cluster set; Similarity(c i ,c j ) is cluster c i and cluster c j The similarity between them is in the range of [0,1]; d min (c i ,c j ) is cluster c i and cluster c j The minimum phase sensitive distance between local is the local scale parameter; w phase (c i , c j ) is the phase consistency weight based on the average phase difference; Cluster c i and cluster c j The average phase value of σ φ is the tolerance parameter of phase difference.

8. The method for removing partial discharge pulse interference noise based on density clustering according to claim 6, characterized in that: In step S34, feature verification of each cluster in the candidate recovery cluster set specifically includes: S341. For all candidate recovery clusters, calculate the cluster compactness ratio and phase consistency index; S342: Classify candidate recovery clusters according to their point size, identifying clusters with fewer points than a set threshold as small clusters and clusters with no fewer points than the set threshold as large clusters, and further calculate local support indices for the large clusters; S343. Calculate the comprehensive confidence based on the cluster classification result, compare the calculated comprehensive confidence with the set confidence threshold, and determine the cluster with a confidence higher than the threshold as a valid recovery cluster, and determine the cluster with a confidence lower than the threshold as a noise cluster.

9. The method for removing partial discharge pulse interference noise based on density clustering according to claim 8, characterized in that: In step S343, the comprehensive confidence is calculated based on the cluster classification results, including: the comprehensive confidence of small clusters adopts a weighted combination of compactness ratio and phase consistency, and the comprehensive confidence of large clusters adopts a triple weighted combination of compactness ratio, phase consistency and local support.

10. The method for removing partial discharge pulse interference noise based on density clustering according to claim 1, characterized in that: Step S4 specifically includes: S41, merging the main partial discharge signal cluster and the effective recovery cluster to form a complete effective partial discharge signal point set; S42, performing denormalization processing on the valid signal point set, and restoring the normalized phase and amplitude data to the original physical dimensions; S43, performing adaptive filtering enhancement on the original noisy discharge signal based on the phase-amplitude characteristics of the effective signal points, highlighting the effective discharge signal and suppressing the remaining noise component by phase window selection and amplitude threshold adjustment; S44. Reconstruct the enhanced signal in the time domain, restore the time series characteristics of the discharge pulse according to the phase information, and obtain the effective discharge signal after denoising.

Images