CN114386499A

CN114386499A - Multi-source partial discharge signal data stream clustering separation method based on GIS

Info

Publication number: CN114386499A
Application number: CN202111663965.5A
Authority: CN
Inventors: 陈昌川; 刘凯; 刘仁光; 代少升; 张天骐
Original assignee: Chongqing University of Post and Telecommunications
Current assignee: Chongqing University of Post and Telecommunications
Priority date: 2021-12-30
Filing date: 2021-12-30
Publication date: 2022-04-22

Abstract

The invention relates to a GIS (geographic information system) -based multi-source partial discharge signal data stream clustering separation method, which belongs to the technical field of partial discharge detection of high-voltage electrical equipment. According to the method, a natural neighborhood is adopted to create the KD tree to improve the efficiency of inquiring neighbors, namely, self-adaptive neighborhood radius and area density are obtained through the characteristics of stream data, so that local search and cluster formation can be realized, and real-time online separation of various local discharge sources is realized. The advantages of the method are verified in the manual data set and the real data set, and the method is applied to mode recognition of the fault of the gas insulated substation.

Description

Multi-source partial discharge signal data stream clustering separation method based on GIS

Technical Field

The invention belongs to the technical field of partial discharge detection of high-voltage electrical equipment, and particularly relates to a GIS (geographic information system) -based multi-source partial discharge signal data stream clustering separation method.

Background

In densely populated cities, the requirement for compact design and small size of substations makes the installation of Gas Insulated Substations (GIS) a prerequisite. Due to the compact design, low maintenance requirements and reliable operation of GIS, it has recently become widely used in power utilities. Like other high voltage devices, GIS device insulators may have various latent insulation defects causing damage in an internal region under the action of a strong electric field, resulting in different types of Partial Discharges (PDs). Insulation degradation mechanisms reflected by different types of PDs are different, and damage degrees to GIS equipment are also different. The PD type can be identified to provide a basis for diagnosis and maintenance of the transformer, so that safe and stable operation of the power system is ensured. The mode identification is one of main contents of fault diagnosis of the gas insulated substation, data of fault information is automatically processed and identified by a mathematical method, effective information is extracted, and therefore data points of faults are clustered and separated. PD signals are collected through a PD signal monitoring system, and characteristic information which can reflect the PD of the gas insulated substation in data is identified through a pattern recognition method, so that the discharge type of the PD can be judged. If the gas insulated substation has faults and the PD phenomenon is generated, the PD type can be judged, and certain technical guidance is provided for maintenance. PD types are roughly classified into: point discharge, hole discharge, floating electrode discharge, free metal particle discharge. These defects are mainly caused by mechanical vibrations of moving parts such as circuit breakers.

Disclosure of Invention

The invention relates to a GIS (geographic information system) -based multi-source partial discharge signal data stream clustering separation method, which realizes effective separation and identification of various partial discharge sources and adopts the following 4 parts.

The Phase Resolved Partial Discharge (PRPD) mode is the most commonly used method in PD measurement and analysis, and the feature extraction method used herein is a statistical feature method. And monitoring the PD signal by using an ultrahigh frequency method, and analyzing the discharge characteristic of the PD signal. And extracting characteristic parameters capable of reflecting the defects of the GIS equipment, wherein the characteristic parameters comprise skewness, steepness, rise time, fall time, pulse width and the like. The characteristic parameters are required to have high identification degree, the running state of the GIS is associated with each characteristic parameter to analyze and warn potential faults of the GIS equipment according to the result, and signals of the different characteristic parameters are separated by using a data stream clustering algorithm. For uncertain data stream characteristics, Cao F et al propose a DenStream clustering algorithm that introduces core micro-clusters to summarize clusters of arbitrary shape, while proposing potential core micro-clusters and abnormal micro-cluster structures to maintain and distinguish potential clusters and abnormal values. The traditional density-based clustering algorithm is expanded, and the problem of data stream clustering in any shape is mainly processed. The DenStream algorithm has the disadvantages that the number of the core micro-clusters is not limited, and a method for deleting or reducing the core micro-clusters does not exist, so that a large amount of memory overhead is caused. Chairukwattana R et al propose an SE-Stream clustering algorithm that improves the performance of the algorithm by reducing the execution time and improving the quality of the micro-clusters, determining an appropriate subset of dimensions for each active micro-cluster to express a particular characteristic of the micro-cluster in the data Stream. Supporting the change of the micro-cluster structure with time, including appearance, disappearance, self-evolution, combination and division of the micro-cluster. The SE-Stream algorithm has the defects that more parameters need to be initialized and defined, and the clustering effect in the later period is influenced by the initialized parameters. The algorithm aims to extract the best set of selected dimensions in each micro-cluster, and it cannot be guaranteed whether these dimensions are redundant. Aiming at the problem that various real-time PD signal data are changed constantly, a system needs to be capable of separating various PD signals in real time, and the efficient EAOStream is provided. After the clusters are formed, the radius can be increased or decreased along with the time, and some micro clusters can be split or combined along with the time according to the dynamic change of the data stream structure. Experimental test results show that the algorithm provides an effective solution for micro-clustering in any shape and has higher classification precision.

(1) Feature extraction: the separation of different PD fault signals is realized, and the characteristic quantity energy which must be extracted firstlyCan reflect the time domain characteristics of the PD signal. Thereby representing the PD signal by the characteristic value extracted by the uhf. Due to the fact that different discharging mechanisms, discharging defect positions and discharging signal propagation paths of various PD signals can show different differences in characteristic values, various PDs can be separated according to different time domain distribution characteristics of the different PD signals, the PD characteristic quantity selection method for type identification is mainly focused on a PD phase analysis mode, and statistical characteristic parameters are described by the characteristics of PRPD. This text extracts skewness S in signal features_kAbruptness K_uPhase phi, and discharge quantity Q.

(2) And (3) natural neighborhood algorithm: self-adapting to the radius neighborhood and the area density, initializing a natural neighborhood, putting data into a KD tree for nearest neighbor search, finding K-neighbor and inverse K-neighbor of each data point, and the number of the neighbors of each data point, and judging whether a stable termination balance condition is achieved.

(3) Parameter selection: the method comprises two stages, wherein a natural neighborhood algorithm is introduced into the first stage, a data set consisting of the first n data points is processed through the natural neighborhood algorithm to obtain a natural characteristic value and a minimum micro-cluster threshold value M, and then the average value of each data point is calculated after the minimum preset distance of each data point is calculated. Setting the neighborhood radius gamma (epsilon) of the algorithm, and obtaining the required region density and neighborhood radius by introducing a natural neighborhood algorithm, thereby not needing to initialize parameter values, continuously entering through data points and updating M and gamma (epsilon) in a self-adaptive manner.

(4) Clustering separation: and distributing the core micro-cluster, judging whether the data sample belongs to the current micro-cluster when a new data point arrives, and if not, creating a new micro-cluster. If the data is in the current micro-cluster, the data is further checked to be within the core radius or shell radius of the micro-cluster. If the data point is judged to fall within the shell radius area, the center position of the micro-cluster is updated. Deleting micro-clusters that decay to a minimum threshold: when all the nanocluster lifetimes have decreased to the delta, the nanoclusters are removed and the edges connected to them are deleted. Updating the cluster: the updating of the cluster map occurs at the position of the center point of the existing micro-cluster and has changed; the micro-clusters move or new micro-clusters are generated; the lifetime of the micro-clusters decays to a set threshold.

Compared with other clustering separation methods, the method has the advantages that: 1. and (3) the KD tree is created by adopting a natural neighborhood to improve the efficiency of inquiring the neighbor, namely, the adaptive neighborhood radius and the area density are obtained through the characteristics of stream data, so that the adaptive purpose is achieved. 2. And an online clustering separation method is adopted, so that online separation of multi-source partial discharge signals can be realized in real time.

Drawings

FIG. 1 is a general flow chart of a data flow clustering separation method based on multi-source partial discharge according to the present invention

FIG. 2 adaptive data stream clustering Process and results graph

FIG. 3 is a diagram of an ultrahigh frequency defect discharge simulation signal generator

FIG. 4 hardware System diagram

FIG. 5 is a comparison diagram of the effect of three clustering algorithms of multi-source partial discharge signals

FIG. 6PRPD clustering effect graph

Detailed Description

The invention is used for providing a GIS-based multi-source partial discharge signal data stream clustering separation method, and in order to make the technical scheme and effect of the invention clearer and clearer, the following describes the specific implementation mode of the invention in detail with reference to the attached drawings.

The flow chart of the clustering separation method for the multi-source partial discharge signal data stream shown in fig. 1 is shown. The parameter selection utilizes a natural neighborhood algorithm to obtain a natural characteristic value lambda of the first n data, and the natural characteristic value lambda is set as a density value D in the algorithm. And then calculating the sum of the neighbor distances of each data point D, averaging, and setting the obtained value as the neighborhood radius R of the algorithm, thereby being capable of obtaining the neighborhood radius and the region density in a self-adaptive manner. The initialization parameters are mainly neighborhood radius, regional density and attenuation value, and the algorithm uses the first data point to initialize the micro-cluster and sets the attribute of the micro-cluster to the initial value. The initialization parameters are mainly neighborhood radius, regional density and attenuation value, and the algorithm uses the first data point to initialize the micro-cluster and sets the attribute of the micro-cluster to the initial value. Deleting micro-clusters that decay to a minimum threshold: this portion reduces the lifetime of the micro-clusters, which are deleted when their lifetime is below zero. All the nanocluster lifetimes are reduced to a decaying amount, the nanoclusters are removed, and the edges connected to it are deleted. Updating the cluster: the updating of the cluster map occurs at the position of the center point of the existing micro-cluster and has changed; the micro-clusters move or new micro-clusters are generated; the lifetime of the micro-clusters decays to a set threshold.

1. Feature extraction

The separation of different discharge fault signals is realized, and the characteristic quantity which must be extracted firstly can reflect the time domain characteristic of the partial discharge signal. Thereby representing the partial discharge signal by the characteristic value extracted by the ultrahigh frequency. Due to the fact that different discharging mechanisms, discharging defect positions and discharging signal propagation paths of various partial discharging signals can show different differences in characteristic values, the various partial discharging signals can be separated according to different time domain distribution characteristics of the different partial discharging signals, the PD characteristic quantity selection method for type identification is mainly focused on a PD phase analysis mode, and statistical characteristic parameters are used for feature description of PRPD. This text extracts skewness S in signal features_kAbruptness K_uCharacteristic quantities such as power frequency phase phi and discharge quantity Q.

2. Multi-source partial discharge signal separation

And forming a series of two-dimensional or three-dimensional maps according to skewness Sk, steepness Ku, power frequency phase phi and discharge quantity Q in statistical parameters in the PRPD map, and classifying various partial discharge signals according to the statistical characteristics of the maps. According to the traditional clustering separation algorithm, a certain problem exists, partial discharge signals received by the ultrahigh frequency sensor are changed in real time and continuously, a complete power frequency period needs to be ensured, and an off-line clustering algorithm such as DBSCAN cannot meet the requirement. Therefore, an efficient self-adaptive online data stream clustering algorithm is provided.

The efficiency of inquiring neighbors is improved by creating the KD tree, the whole data set is traversed, and the KD tree is accessed in a recursion mode from the root node. K neighbors and inverse K neighbors of each data point are found. Definition of NaN method: given a set of data points ρ₁，ρ₂，ρ₃，…，ρ_NAll data points ρ_iAnd ρ_jSimilarity between s_ijThe purpose of (a) is to find the natural neighborhood of these points in the dataset, computed and stored in a distance matrix, one of the most popular choices to measure this distance is the euclidean distance. The condition for reaching the natural steady state is that the number of data points in the data set with zero neighbor number no longer changes or all objects have inverse neighbors. For data points, if the point ρ is simultaneously applied_iIs regarded as rho_jAnd point ρ_iIs regarded as rho_jThen ρ_iIs the point ρ_jThen ρ_iIs the point ρ_jOne of the natural neighbors. The natural stable structure of the data points was formulated as follows:

the multi-source partial discharge signal separation method is a self-adaptive complete online clustering method and comprises two stages. In the first stage, a natural neighborhood algorithm is introduced, a data set consisting of the first n data points is processed through the natural neighborhood algorithm to obtain a natural characteristic value lambda, a minimum threshold value D in the algorithm is set, then the distance between each data point and the minimum density D is calculated, and then the average value of the data points is calculated. Setting a neighborhood radius gamma (epsilon) of the algorithm, and introducing a natural neighborhood algorithm to obtain a required minimum threshold value and a required neighborhood radius, so that initialization parameter values are not needed, and the neighborhood radius and the minimum threshold value are updated in a self-adaptive manner by continuously entering data points. The calculation formula of the neighborhood radius of the algorithm is as follows:

the second stage is an intersecting micro cluster, the micro cluster is divided into a shell area and a core area, and micro clusters are grouped by considering the core area intersecting with the micro cluster shell, so that micro clusters at the edge can be automatically determined. Outliers will exist for micro-clusters that do not have the minimum threshold, and each micro-cluster contains a graph that demonstrates the intersection of the micro-clusters. The calculations required to separate the micro-clusters upon their rupture or eventual death can be minimized by the application of a graph structure. And obtaining a clustering result by adopting a mode of updating the graph structure in real time, calculating the accessibility of a plurality of connected micro-clusters around the modified micro-cluster after the data point arrives, and ensuring the effectiveness of micro-cluster division because the rest points do not need to be modified. Cluster formation process and results the different colored micro-clusters represent different categories as shown in fig. 2.

Step 1, parameter selection: and obtaining a natural characteristic value lambda from the first n data by utilizing a natural neighborhood algorithm, and setting the natural characteristic value lambda as a medium density value D of the algorithm. And then calculating the sum of the neighbor distances of each data point D, averaging, and setting the obtained value as the neighborhood radius R of the algorithm, thereby being capable of obtaining the neighborhood radius and the region density in a self-adaptive manner.

Step 2, initializing the micro-cluster: the initialization parameters are mainly neighborhood radius, regional density and attenuation value, and the algorithm uses the first data point to initialize the micro-cluster and sets the attribute of the micro-cluster to the initial value.

Step 3, distributing core micro-clusters: when a new data point is reached, it is determined whether the data sample belongs to any current micro-cluster. If not, a new micro-cluster is created. If the data is in the current micro-cluster, the data is further checked to be within the core radius or shell radius of the micro-cluster. If the data point is judged to fall within the shell radius area, the center position of the micro-cluster is updated.

Step 4, deleting the micro-clusters decaying to the minimum threshold: this portion reduces the lifetime of the micro-clusters, which are deleted when their lifetime is below zero. All the nanocluster lifetimes are reduced to a decaying amount, the nanoclusters are removed, and the edges connected to it are deleted.

Step 5, updating the cluster: the updating of the cluster map occurs at the position of the center point of the existing micro-cluster and has changed; the micro-clusters move or new micro-clusters are generated; the lifetime of the micro-clusters decays to a set threshold.

Under the above circumstances, the edge list of the algorithm may be changed, the number of micro clusters needs to be updated, and first, a proper neighborhood radius and region density may be calculated through a natural neighborhood algorithm. The micro-clusters that have been modified to have recently reached the threshold or moved their center positions will be addressed. At this time, the pattern edge of the micro-cluster is modified, and the number of the generated micro-cluster is also changed.

3. Clustering separation field actual measurement of multiple partial discharge source signals

In order to verify the feasibility and the effectiveness of the clustering separation of the time domain characteristics of the partial discharge signals, a partial discharge monitoring system is established with a school enterprise cooperation project, a system test is carried out on the partial discharge types mixed with different types in a laboratory, a physical diagram of an ultrahigh frequency defect discharge simulation signal generator is shown in figure 3, the collected signals are input into a hardware system board as shown in figure 4, the hardware system board comprises a signal collecting unit and a signal processing unit, a clustering algorithm is transplanted into the hardware system board, and the data after real-time online processing is uploaded to an upper computer for display. Selecting skewness S in statistical parameters in PRPD atlas_kAnd abruptness K_uThe separation of various partial discharge signals is realized by the formed two-dimensional map. In a laboratory, a plurality of partial discharge signals are collected to form a data set, effects of three clustering algorithms are compared with those of fig. 5, categories are judged according to colors, clusters with the same color represent one category, five categories of signals can be known from fig. 5, the effects of Denstoream and SE-Stream algorithms on separation of a first category and a second category in the diagram are not influenced, the first category of signals and the second category of signals are close to each other and are easily judged as one category of signals, and colors of the first category of signals and the second category of signals shown in fig. 5(b-c) are judged as one color by the algorithms. The algorithm herein is capable of accurately separating the two signals.

After real-time data are processed in a hardware system board, marking bits are provided for signals of different types and uploaded to an upper computer, and the data comprise three parameters including phase, discharge amplitude and the marking bits. As shown in fig. 6, the display result of the PRPD map is represented by a phase on the abscissa, a discharge assignment on the ordinate, and a discharge frequency mapped to the color space, and each distribution position may be superimposed, so that a color identifier may be obtained, and the color identifier may be divided into 6 color classes according to the accumulated frequency. In the experiment, two signals are mixed and input, namely a suspension electrode discharge signal and a point discharge signal. Because the discharge signal of the suspended electrode in the laboratory has no fixed phase, the phase can shift along with the time, and finally a green line is formed. The point discharge signal is generated by a high-voltage test transformer and an ultrahigh frequency defect discharge analog signal generator in a laboratory, and the color of the middle part can be always darker and finally tends to be stable due to the fact that the phase is fixed and is accumulated along with time.

Claims

1. A GIS-based multi-source partial discharge signal data stream clustering separation method is characterized by comprising the following steps:

step 1: and (4) feature extraction, wherein feature quantities such as skewness Sk, steepness Ku, phase phi, discharge quantity Q and the like in the signal features are extracted.

Step 2: and (3) a natural neighborhood algorithm, which is self-adaptive to the radius neighborhood and the area density, initializes the natural neighborhood, puts the data into a KD tree for nearest neighbor search, finds K-nearest neighbors and inverse K-nearest neighbors of each data point, and judges whether a stable termination balance condition is achieved or not according to the number of the nearest neighbors of each data point.

And step 3: parameter selection: the method comprises two stages, wherein a natural neighborhood algorithm is introduced into the first stage, a data set consisting of the first n data points is processed through the natural neighborhood algorithm to obtain a natural characteristic value lambda and a minimum micro-cluster threshold value M, and then after the minimum preset distance between each data point and the minimum preset distance is calculated, the average value of the data points is calculated. Setting the neighborhood radius gamma (epsilon) of the algorithm, and obtaining the required region density and neighborhood radius by introducing a natural neighborhood algorithm, thereby not needing to initialize parameter values, continuously entering through data points and updating M and gamma (epsilon) in a self-adaptive manner.

And 4, step 4: clustering separation: and distributing the core micro-cluster, judging whether the data sample belongs to the current micro-cluster when a new data point arrives, and if not, creating a new micro-cluster. If the data is in the current micro-cluster, the data is further checked to be within the core radius or shell radius of the micro-cluster. If the data point is judged to fall within the shell radius area, the center position of the micro-cluster is updated. Deleting micro-clusters that decay to a minimum threshold: when all the nanocluster lifetimes have decreased to the delta, the nanoclusters are removed and the edges connected to them are deleted. Updating the cluster: the updating of the cluster map occurs at the position of the center point of the existing micro-cluster and has changed; the micro-clusters move or new micro-clusters are generated; the lifetime of the micro-clusters decays to a set threshold.

2. The GIS multi-source partial discharge signal data stream clustering separation method based on claim 1, which is used for realizing separation of different discharge fault signals, wherein the feature quantity which must be extracted firstly can reflect the time domain feature of the partial discharge signal. Thereby representing the partial discharge signal by the characteristic value extracted by the ultrahigh frequency. Due to the fact that different discharging mechanisms, discharging defect positions and discharging signal propagation paths of various partial discharging signals can show different differences in characteristic values, the various partial discharging signals can be separated according to different time domain distribution characteristics of the different partial discharging signals, the PD characteristic quantity selection method for type identification is mainly focused on a PD phase analysis mode, and statistical characteristic parameters are used for feature description of PRPD. The method extracts characteristic quantities such as skewness Sk, steepness Ku, power frequency phase phi, discharge quantity Q and the like in signal characteristics.

3. The GIS-based multi-source partial discharge signal data stream clustering separation method according to claim 2, wherein: and initializing a natural neighborhood, putting data into the KD tree for nearest neighbor search, and judging whether a stable termination balance condition is achieved.

4. The GIS-based multi-source partial discharge signal data stream clustering separation method according to claim 3, wherein: the algorithm processes a data set consisting of the first n data points through a natural neighborhood algorithm to obtain a natural characteristic value lambda, sets a minimum threshold value D in the algorithm, calculates the distance between each data point and the minimum density D, and then calculates the average value of the data points. Setting a neighborhood radius gamma (epsilon) of the algorithm, and obtaining the required minimum threshold value and the required neighborhood radius by introducing a natural neighborhood algorithm, so that the neighborhood radius and the minimum threshold value are updated in a self-adaptive manner by continuously entering data points without initializing parameter values.

The calculation formula of the neighborhood radius of the algorithm is as follows:

wherein d is_λ(i) Denotes the distance of the lambda neighbor of data point i and n represents the number of data points used to obtain the minimum threshold. In order to address the natural law of data set distribution, for the area with sparse or dense data point distribution, the searched radius can be Weighted by using Weighted multi-width Gaussian Kernel with Multiple width (WGKMW), and the neighborhood radius of the micro-cluster is dynamically set. The larger the density value D is, the larger the neighborhood radius is, otherwise, the smaller the neighborhood radius is. The weighting formula is as follows:

where n represents the number of data points in the micro-cluster, M is the minimum threshold for the micro-cluster, and M Gaussian kernels of different widths are linearly combined into a weighted multi-width Gaussian kernel. R^m-sAre the weighting coefficients on m gaussian kernels of different widths. Constant factor R^mThe linear translation of the distance between data points in the micro-clusters is amplified, the difference of samples is enlarged in the characteristic interval, and clustering of the micro-clusters with weak difference can be better realized. Thus, the radius and density of the micro-clusters can be adaptively changed through a natural neighborhood algorithm. Then the micro-clusters are adjusted, the algorithm adopts a simple linear aging determination method, so that the service life of the micro-clusters is shortened, the micro-clusters completely disappear as unused micro-clusters, not only can the aging replacing technology be applied, but also the micro-clusters can be activated by adding more data points, and the spectrum of the micro-clusters is updated. When no data is available for reception, the micro-clusters will fade away. When this phenomenon is extensive, the lifetime of the micro-clusters will gradually reach zero, thereby eliminating the micro-clusters.

5. The GIS-based multi-source partial discharge signal data stream clustering separation method of claim 4, which is an intersecting micro cluster, wherein the micro cluster is divided into a shell area and a core area, and micro clusters are grouped by considering the core area intersecting with the micro cluster shell, so that micro clusters at the edge can be automatically determined. Outliers will exist for micro-clusters that do not have the minimum threshold, and each micro-cluster contains a graph that demonstrates the intersection of the micro-clusters. The calculations required to separate the micro-clusters upon their rupture or eventual death can be minimized by the application of a graph structure. And obtaining a clustering result by adopting a mode of updating the graph structure in real time, calculating the accessibility of a plurality of connected micro-clusters around the modified micro-cluster after the data point arrives, and ensuring the effectiveness of micro-cluster division because the rest points do not need to be modified. Under the above circumstances, the edge list of the algorithm may be changed, the number of micro clusters needs to be updated, and first, a proper neighborhood radius and region density may be calculated through a natural neighborhood algorithm. The micro-clusters that have been modified to have recently reached the threshold or moved their center positions will be addressed. At this time, the pattern edge of the micro-cluster is modified, and the number of the generated micro-cluster is also changed.