CN114386499A - A clustering and separation method of multi-source partial discharge signal data stream based on GIS - Google Patents

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

A clustering and separation method of multi-source partial discharge signal data stream based on GIS

technical field

The invention belongs to the technical field of partial discharge detection of high-voltage electrical equipment, in particular to a clustering and separation method based on GIS multi-source partial discharge signal data streams.

Background technique

In densely populated cities, the requirements for compact design and small size of substations make the installation of Gas Insulated Substations (GIS) necessary. Due to the compact design, low maintenance requirements and reliable operation of GIS, it has been widely used in electric utilities in recent years. Like other high-voltage equipment, under the action of a strong electric field, GIS equipment insulators may have various latent insulation defects that cause damage, resulting in different types of partial discharges (PD). Different types of PDs reflect different insulation degradation mechanisms and different degrees of damage to GIS equipment. Identifying the PD type can provide the basis for the diagnosis and maintenance of the transformer, thereby ensuring the safe and stable operation of the power system. Pattern recognition is one of the main contents of fault diagnosis of gas-insulated substations. It uses mathematical methods to automatically process and identify data with fault information, extract effective information, and then cluster and separate fault data points. The PD signal is collected by the PD signal monitoring system, and the characteristic information that can reflect the PD of the gas-insulated substation in the data is identified by the pattern recognition method, so that the discharge type of the PD can be judged. If the gas-insulated substation fails, resulting in the phenomenon of PD, the type of PD can be judged to provide certain technical guidance for maintenance. PD types are roughly divided into: tip discharge, hole discharge, suspended electrode discharge, and free metal particle discharge. These defects are mainly caused by mechanical vibrations of moving parts such as circuit breakers.

SUMMARY OF THE INVENTION

The content of the present invention is to realize a GIS-based multi-source partial discharge signal data stream clustering separation method, and realize the effective separation and identification of various partial discharge sources. The specific technical scheme includes the following four parts.

Phase Resolved Partial Discharge (PRPD) mode is the most commonly used method in PD measurement and analysis. The feature extraction method used in this paper is statistical feature method. The PD signal is monitored by the UHF method, and its discharge characteristics are analyzed. Extract the characteristic parameters that can reflect the defects of GIS equipment, these characteristic parameters include skewness, steepness, rise time, fall time and pulse width. These characteristic parameters should have a high degree of identification, correlate the operating status of the GIS with each characteristic parameter, analyze and use the results to warn the potential failure of GIS equipment, and use the data stream clustering algorithm to separate the signals of the unused characteristic parameters. come out. Aiming at the uncertain data flow characteristics, Cao F et al. proposed the DenStream clustering algorithm, which introduced core microclusters to generalize clusters of arbitrary shapes, and proposed potential core microclusters and abnormal microcluster structures to maintain and distinguish potential clusters and outliers. It extends traditional density-based clustering algorithms to focus on the problem of clustering data streams of arbitrary shapes. The DenStream algorithm has shortcomings. It does not limit the number of core microclusters, and there is no way to delete or reduce core microclusters, which will lead to a lot of memory overhead. Chairukwattana R et al. proposed the SE-Stream clustering algorithm, which improves the performance of the algorithm by reducing the execution time and improving the quality of microclusters, and determines the appropriate dimensional subset of each active microcluster to express the specific characteristics of the microclusters in the data stream. . Changes in microcluster structure over time are supported, including the emergence, disappearance, self-evolution, merging and splitting of microclusters. The SE-Stream algorithm has defects, it needs to initialize and define many parameters, and the later clustering effect is affected by the initialization parameters. The algorithm aims to extract the best set of selected dimensions in each microcluster, and there is no guarantee that these dimensions are redundant. Aiming at the problem of a variety of real-time PD signal data and changing problems, the system needs to be able to separate a variety of PD signals in real time. This paper proposes an efficient EAOStream, which uses the natural neighborhood (Natural Neighbor, NaN) method to create a KD tree (K-dimension Tree) to improve the efficiency of querying neighbors. This method calculates the adaptive neighborhood radius and area density through the eigenvalues of the stream data for local search, so as to support the best the clustering effect. After the cluster is formed, the radius can increase or decrease with time. According to the dynamic change of the data flow structure, some microclusters will be split or merged over time. The experimental test results show that the algorithm provides an effective solution for arbitrary-shaped micro-cluster clustering and has higher classification accuracy.

(1) Feature extraction: To achieve the separation of different PD fault signals, the feature quantity that must be extracted firstly can reflect the time domain characteristics of PD signals. Therefore, the PD signal is represented by the eigenvalues extracted by the UHF. A variety of PD signals will show different differences in eigenvalues due to different discharge mechanisms, discharge defect locations, and discharge signal propagation paths. Therefore, various PD signals can be separated according to the different time domain distribution characteristics of different PD signals. , the PD feature selection method for type identification mainly focuses on the PD phase analysis mode, and the statistical feature parameters are described by PRPD features. In this paper, the characteristic quantities such as skewness _Sk , steepness _Ku , phase Φ and discharge quantity Q are extracted from the signal features.

(2) Natural neighborhood algorithm: adaptive radius neighborhood and area density, initialize natural neighborhood, put data into KD tree for nearest neighbor search, and find its K-nearest neighbor and inverse K-neighbor for each data point. Neighbors, and the number of neighbors of each data point, determine whether a stable termination equilibrium condition is reached.

(3) Parameter selection: It is divided into two stages. In the first stage, the natural neighborhood algorithm is introduced, and the data set composed of the first n data points is processed by the natural neighborhood algorithm to obtain the natural feature value and the minimum threshold M of microclusters, and then After calculating the distance of each data point from the minimum preset, it is then averaged. Set the neighborhood radius of the algorithm γ(ε), by introducing the natural neighborhood algorithm, the required area density and neighborhood radius are obtained, so that there is no need to initialize the parameter values, and the data points are continuously entered, and M and γ(ε) are adaptively updated. .

(4) Cluster separation: Allocate core microclusters. When a new data point arrives, determine whether the data sample belongs to the current microcluster. If not, create a new microcluster. If the data is in the current microcluster, further check that the data is within the core radius or shell radius of the microcluster. If it is judged that the data point falls within the shell radius area, the center position of the microcluster is updated. Delete microclusters that decay to the minimum threshold: When all microclusters have reduced their lifespans to the decay amount, remove the microcluster and delete the edges connected to it. Update cluster: Cluster map update occurs when the position of the existing micro-cluster center point has changed; the micro-cluster moves or generates a new micro-cluster; the life of the micro-cluster decays to a set threshold.

Compared with other clustering separation methods, the advantages of the present invention are reflected in the following points: 1. Using natural neighborhoods to create KD trees to improve the efficiency of querying neighbors, that is, to obtain adaptive neighborhood radius and area through the characteristics of streaming data density, so as to achieve the purpose of self-adaptation. 2. The online clustering separation method is adopted, which can realize the online separation of multi-source partial discharge signals in real time.

Description of drawings

FIG. 1 is an overall flow chart of a method for clustering and separating data streams based on multi-source partial discharge according to the present invention.

Figure 2 Adaptive data stream clustering process and result diagram

Figure 3 UHF defect discharge analog signal generator diagram

Figure 4 Hardware system diagram

Figure 5 Comparison of the effects of three clustering algorithms for multi-source PD signals

Figure 6 PRPD clustering effect diagram

Detailed ways

The present invention is used to provide a method for clustering and separating data streams of multi-source partial discharge signals based on GIS.

Figure 1 is a flow chart of a method for clustering and separating data streams of multi-source partial discharge signals. The parameter selection uses the natural neighborhood algorithm to obtain the natural eigenvalue λ for the first n data, which is set as the density value D in the algorithm. Then calculate the sum of the distances of the neighbors of each data point D, and calculate the average value. The obtained value is set as the neighborhood radius R of the algorithm, so that the neighborhood radius and area density can be obtained adaptively. The initialization parameters are mainly neighborhood radius, area density and attenuation value. The algorithm uses the first data point to initialize the microcluster, and sets the properties of the microcluster to initial values. The initialization parameters are mainly neighborhood radius, area density and attenuation value. The algorithm uses the first data point to initialize the microcluster, and sets the properties of the microcluster to initial values. Remove microclusters that decay to a minimum threshold: This part reduces the lifetime of microclusters, removing them when their lifetime falls below zero. All microcluster lifetimes are reduced to the decay amount, the microcluster is removed, and the edges connected to it are deleted. Update cluster: Cluster map update occurs when the position of the existing micro-cluster center point has changed; the micro-cluster moves or generates a new micro-cluster; the life of the micro-cluster decays to a set threshold.

1. Feature extraction

To achieve the separation of different discharge fault signals, the feature quantity that must be extracted firstly can reflect the time domain characteristics of partial discharge signals. Therefore, the partial discharge signal is represented by the eigenvalues extracted by the UHF. Various partial discharge signals will show different differences in eigenvalues due to different discharge mechanisms, discharge defect locations, and discharge signal propagation paths. The PD feature quantity selection method for type identification mainly focuses on the PD phase analysis mode, and the statistical feature parameters are used for the feature description of PRPD. In this paper, the characteristic quantities such as skewness _Sk , steepness _Ku , power frequency phase Φ and discharge quantity Q are extracted from the signal features.

2. Multi-source partial discharge signal separation

According to the skewness Sk, the steepness Ku, the power frequency phase Φ and the discharge quantity Q in the statistical parameters of the PRPD spectrum, a series of two-dimensional or three-dimensional maps are formed, and the statistical characteristics of these maps are used to realize the detection of various partial discharge signals. Classification. According to the traditional clustering separation algorithm, there are certain problems. The PD signal received by the UHF sensor is real-time and constantly changing, and it is necessary to ensure a complete power frequency cycle. Offline clustering algorithms such as DBSCAN cannot meet this requirement. Therefore, an efficient adaptive online data stream clustering algorithm is proposed.

The efficiency of querying neighbors is improved by creating a KD tree, traversing the entire data set, starting from the root node, and accessing the KD tree recursively. Find the K-nearest neighbors and inverse K-nearest neighbors of each data point. NaN method definition: Given a set of data points ρ ₁ , ρ ₂ , ρ ₃ , ..., ρ _N the concept of similarity s _ij between all data points ρ _i and ρ _j , the purpose is to find the similarity of these points in the data set Natural neighborhoods, computed and stored in a distance matrix, one of the most popular choices for measuring this distance is the Euclidean distance. The condition for reaching a natural steady state is that the number of data points in the dataset with zero neighbors no longer changes or that all objects have inverse neighbors. For a data point, if point ρ _i is considered as ρ _j and point ρ _i as ρ _j , then ρ _i is point ρ _j , then ρ _i is one of the natural neighbors of point ρ _j . The natural stable structure of the data points is formulated in the following way:

The multi-source partial discharge signal separation method is an adaptive fully online clustering method, which consists of two stages. In the first stage, the natural neighborhood algorithm is introduced, and the data set composed of the first n data points is processed by the natural neighborhood algorithm to obtain the natural eigenvalue λ, and the minimum threshold D in the algorithm is set, and then the relationship between each data point and each data point is calculated. After the distance of the minimum density D, it is then averaged. Set the neighborhood radius of the algorithm γ(ε), by introducing the natural neighborhood algorithm, the required minimum threshold and neighborhood radius, so that there is no need to initialize the parameter values, and the neighborhood radius and the minimum threshold are adaptively updated by continuously entering data points. The calculation formula of the neighborhood radius of the algorithm is as follows:

The second stage is the intersecting microclusters. The microclusters are divided into shell regions and core regions. By grouping the microclusters by considering the core region that intersects the shell of the microclusters, the edge microclusters can be automatically determined. Microclusters that do not have a minimum threshold have outliers, and each microcluster contains a graph that demonstrates the intersection of the microclusters. The computation required to separate the microclusters when they break up or eventually die can be minimized by applying a graph structure. The clustering result is obtained by updating the graph structure in real time. When the data point arrives, the reachability of several connected microclusters around the modified microcluster is calculated, and the rest of the points do not need to be modified, which can ensure the division of the microclusters. effectiveness. The cluster formation process and results are shown in Figure 2. Microclusters with different colors represent different categories.

Step 1 Parameter selection: The natural neighborhood algorithm is used to obtain the natural eigenvalue λ for the first n data, which is set as the density value D in the algorithm. Then calculate the sum of the distances of the neighbors of each data point D, and calculate the average value. The obtained value is set as the neighborhood radius R of the algorithm, so that the neighborhood radius and area density can be obtained adaptively.

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

Step 3 Allocate core microclusters: When a new data point is reached, determine whether the data sample belongs to any of the current microclusters. If not, create a new microcluster. If the data is in the current microcluster, further check that the data is within the core radius or shell radius of the microcluster. If it is judged that the data point falls within the shell radius area, the center position of the microcluster is updated.

Step 4 Remove microclusters that decay to a minimum threshold: This part reduces the lifetime of microclusters, removing them when their lifetime falls below zero. All microcluster lifetimes are reduced to the decay amount, the microcluster is removed, and the edges connected to it are deleted.

Step 5: Update the cluster: the cluster map update occurs when the center point of the existing micro-cluster has changed; the micro-cluster moves or generates a new micro-cluster; the life of the micro-cluster decays to a set threshold.

In the above case, the edge list of the algorithm can be changed, and the number of microclusters needs to be updated. First, the appropriate neighborhood radius and area density can be calculated through the natural neighborhood algorithm. Microclusters that have been modified to have recently reached the threshold or moved their center position. At this time, the graph edges of the micro-clusters have been modified, and the number of the generated micro-clusters has also changed.

3. On-site measurement of clustering and separation of various partial discharge source signals

In order to verify the feasibility and effectiveness of clustering separation of partial discharge signal time-domain features, this paper establishes a partial discharge monitoring system with a school-enterprise cooperation project, and conducts systematic tests on a variety of mixed partial discharge types in the laboratory. The physical diagram of the UHF defect discharge analog signal generator is shown in Figure 3, and the collected signals are input into the hardware system board as shown in Figure 4, which includes a signal acquisition unit and a signal processing unit, and the clustering algorithm is transplanted to the hardware system In the board, the data after real-time online processing is uploaded to the upper computer for display. In this paper, a two _- dimensional map composed of the skewness _Sk and the steepness Ku in the statistical parameters of the PRPD map is selected to realize the separation of various PD signals. A variety of PD signals were collected in the laboratory to form a data set. The comparison of the effects of the three clustering algorithms is shown in Figure 5. The categories are judged by color. The clusters of the same color represent one category. It can be seen from Figure 5 that there are five categories of signals. , Denstream and SE-Stream algorithms have no effect on the separation effect of the first and second types in the figure. The first type of signal and the second type of signal are close to each other, and it is easy to judge them as a type of signal. The colors of the first type signal and the second type signal shown in 5(bc) are determined by the algorithm as one color. The algorithm in this paper can accurately separate the two signals.

After the real-time data is processed in the hardware system board, the marker bits for different types of signals are provided and uploaded to the host computer. The data includes three parameters: phase, discharge amplitude and marker bits. The results of the PRPD map display are shown in Figure 6. The abscissa represents the phase, the ordinate represents the discharge assignment, and the number of discharges is mapped to the color space. Each distribution position can be superimposed, so that color identification can be achieved. 6 color levels. In the experiment, two kinds of signal mixed input are used, which are the discharge signal of the floating electrode and the discharge signal of the tip. Since there is no fixed phase in the generation of the discharge signal from the suspended electrode in the laboratory, the phase shifts over time, resulting in a green line. The tip discharge signal is generated by the laboratory medium and high voltage test transformer and the UHF defect discharge analog signal generator. Due to the fixed phase, the color of the middle part will always become darker with time accumulation, and finally tends to be stable.

Claims (5)

1. a method for clustering and separating based on GIS multi-source partial discharge signal data stream, is characterized in that, comprises: Step 1: Feature extraction, this paper extracts the skewness Sk, steepness Ku, phase Φ and discharge quantity Q and other feature quantities in the signal features. Step 2: Natural neighborhood algorithm, adaptive radius neighborhood and area density, initialize natural neighborhood, put data into KD tree for nearest neighbor search, and find its K-nearest neighbor and inverse K-neighbor for each data point. Neighbors, and the number of neighbors of each data point, determine whether a stable termination equilibrium condition is reached. Step 3: Parameter selection: It is divided into two stages. In the first stage, the natural neighborhood algorithm is introduced, and the data set composed of the first n data points is processed by the natural neighborhood algorithm to obtain the natural eigenvalue λ and the minimum threshold M of microclusters. Then after calculating the distance of each data point from the minimum preset, then take the average. Set the neighborhood radius of the algorithm γ(ε), by introducing the natural neighborhood algorithm, the required area density and neighborhood radius are obtained, so that there is no need to initialize the parameter values, and the data points are continuously entered, and M and γ(ε) are adaptively updated. . Step 4: Cluster separation: assign core microclusters, when a new data point arrives, determine whether the data sample belongs to the current microcluster, and if not, create a new microcluster. If the data is in the current microcluster, further check that the data is within the core radius or shell radius of the microcluster. If it is judged that the data point falls within the shell radius area, the center position of the microcluster is updated. Delete microclusters that decay to the minimum threshold: When all microclusters have reduced their lifespans to the decay amount, remove the microcluster and delete the edges connected to it. Update cluster: Cluster map update occurs when the position of the existing micro-cluster center point has changed; the micro-cluster moves or generates a new micro-cluster; the life of the micro-cluster decays to a set threshold. 2. The GIS-based multi-source partial discharge signal data stream clustering and separation method according to claim 1, to realize the separation of different discharge fault signals, firstly, the feature quantity that must be extracted can reflect the time domain characteristics of the partial discharge signal. Therefore, the partial discharge signal is represented by the eigenvalues extracted by the UHF. Various partial discharge signals will show different differences in eigenvalues due to different discharge mechanisms, discharge defect locations, and discharge signal propagation paths. The signal is separated, and the PD feature selection method for type identification mainly focuses on the PD phase analysis mode, and the statistical feature parameters are used for the feature description of PRPD. In this paper, the skewness Sk, the steepness Ku, the power frequency phase Φ, and the discharge quantity Q are extracted from the signal features. 3. the method for clustering and separating data stream based on GIS multi-source partial discharge signal according to claim 2, is characterized in that: initialize natural neighborhood, put data into KD tree and carry out nearest neighbor search, put data into The nearest neighbor search is performed on the KD tree to determine whether the stable termination equilibrium condition is reached. 4. The method for clustering and separating data streams based on GIS multi-source partial discharge signal according to claim 3, characterized in that: the algorithm processes the data set formed by the first n data points through the natural neighborhood algorithm to obtain natural characteristic values λ, set the minimum threshold D in the algorithm, and then calculate the distance between each data point and the minimum density D, and then calculate the average value. Set the neighborhood radius γ(ε) of the algorithm, and by introducing the natural neighborhood algorithm, the required minimum threshold and neighborhood radius are obtained, so that there is no need to initialize parameter values, and the neighborhood radius and minimum threshold are adaptively updated through continuous data point entry. . The calculation formula of the neighborhood radius of the algorithm is as follows:

where d _λ (i) represents the distance between the λ neighbors of data point i, and n represents the number of data points used to obtain the minimum threshold. In view of the natural law of data set distribution, for areas with sparse or dense distribution of data points, the search radius can be weighted by a Weighted Gaussian Kernel with Multiple Widths (WGKMW) to dynamically set the neighbors of microclusters. Domain radius. When the density value D is larger, the neighborhood radius is larger, otherwise, the neighborhood radius is smaller. The weighting formula is as follows:

Where n represents the number of data points in the microcluster, M is the minimum threshold of the microcluster, and m Gaussian kernels with different widths are linearly combined into a weighted multi-width Gaussian kernel. ^Rms are the weight coefficients on m Gaussian kernels of different widths. The constant factor ^Rm enlarges the linear translation of the distance between the data points in the microcluster, and enlarges the difference of the samples within its characteristic interval, which can better realize the clustering between the weakly different microclusters. Therefore, the radius and density of microclusters can be adaptively changed through the natural neighborhood algorithm. The microclusters are then tuned, the algorithm employs a simple linear aging determination method that shortens the lifespan of the microclusters, which disappear completely as unused microclusters, allowing not only alternative aging techniques to be applied, but also The microcluster can also be activated by adding more data points, thereby updating the map of the microcluster. Microclusters gradually disappear when no usable data is received. When this phenomenon is widespread, the lifetime of the microclusters will gradually reach zero, thereby eliminating these microclusters. 5. The method for clustering and separating data streams of GIS-based multi-source partial discharge signals according to claim 4, is an intersecting micro-cluster, dividing the micro-cluster into a shell area and a core area, by considering the intersection with the micro-cluster shell. The core region is used to group microclusters, and the edge microclusters can be automatically determined. Microclusters that do not have a minimum threshold have outliers, and each microcluster contains a graph that demonstrates the intersection of the microclusters. The computation required to separate the microclusters when they break up or eventually die can be minimized by applying a graph structure. The clustering result is obtained by updating the graph structure in real time. When the data point arrives, the reachability of several connected microclusters around the modified microcluster is calculated, and the rest of the points do not need to be modified, which can ensure the division of the microclusters. effectiveness. In the above case, the edge list of the algorithm can be changed, and the number of microclusters needs to be updated. First, the appropriate neighborhood radius and area density can be calculated through the natural neighborhood algorithm. Microclusters that have been modified to have recently reached the threshold or moved their center position. At this time, the graph edges of the micro-clusters have been modified, and the number of the generated micro-clusters has also changed.

Images