A partial discharge pattern recognition method based on combination of Kmeans and DBSCAN

Abstract

The application discloses a partial discharge mode recognition method based on combination of Kmeans and DBSCAN, and comprises the following steps: data acquisition, continuously acquiring time domain data of a partial discharge sensor, performing real-time processing on the acquired time domain data, and generating a PRPD spectrum diagram; performing envelope waveform discharge / noise calculation on the PRPD spectrum diagram; performing white noise filtering on the PRPD spectrum diagram based on the Kmeans method for the PRPD spectrum diagram judged as discharge; performing PD cluster clustering on the filtered PRPD spectrum diagram based on a hierarchical density clustering method, performing feature extraction on the clustered clusters; establishing a model, training the model, and obtaining a model credibility; inputting actually measured data of a partial discharge fault into the trained model for recognition, and obtaining a final recognition result. The application performs filtering and feature extraction based on the PRPD spectrum diagram, reduces communication pressure of data transmission of an online monitoring system, and only performs feature extraction on PD clusters in the spectrum diagram, so that the accuracy of feature extraction is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment insulation state monitoring, and particularly relates to a partial discharge mode recognition method based on combination of Kmeans and DBSCAN. BACKGROUND

[0002] A power distribution system is a complex power supply network, which is in the form of a power grid composed of a large number of power assets such as switching devices, transformers, power cables, etc. These devices were installed several decades ago and their service life is close to the end, so their condition needs to be monitored and possibly improved to avoid major damage. According to statistics, 40% of power equipment failures are caused by insulation faults or potential defects, such as cracks in insulators, poor electrical contact, contamination of insulating bushings or dust penetration, etc. These insulation defects can excite partial discharge under the action of an electric field, which poses a potential hazard to the equipment. Therefore, monitoring and diagnosing partial discharge can help to discover potential safety hazards of the equipment and avoid serious damage to the equipment.

[0003] Common partial discharge fault diagnosis is mostly based on analysis of PRPD spectrum characteristics of partial discharge, but when diagnosing complex systems or equipment faults, the extracted features are very sensitive, such as noise and phase shift, which can affect the consistency / generality of the extracted features. SUMMARY

[0004] To solve the above problems, the application provides a partial discharge mode recognition method based on combination of Kmeans and DBSCAN, which extracts features only from specific regions of a PRPD graph indicating the presence of PD, so as to improve the accuracy of PD features and further improve the reliability and accuracy of partial discharge fault diagnosis.

[0005] To achieve the above purpose, the specific technical solutions are as follows:

[0006] A partial discharge mode recognition method based on combination of Kmeans and DBSCAN includes the following steps:

[0007] Data acquisition: continuously acquire time domain data of a partial discharge sensor, and perform real-time processing on the acquired time domain data to generate a PRPD spectrum graph;

[0008] Perform envelope waveform discharge / noise calculation on the PRPD spectrum graph;

[0009] Perform white noise filtering on the PRPD spectrum graph judged as discharge based on the Kmeans method;

[0010] Perform PD cluster clustering on the filtered PRPD spectrum graph based on the hierarchical density clustering method, and extract features from the clustered clusters;

[0011] Modeling, model training and model credibility identification are obtained;

[0012] The data of the partial discharge fault are input into the trained model for identification to obtain the final identification result.

[0013] Further, the data collection comprises continuously collecting time domain data 1s of the partial discharge sensor.

[0014] Further, the real-time processing of the collected time domain data comprises dividing every 20ms data into 360 intervals, calculating the maximum amplitude of each interval, and generating 360*50 data.

[0015] Further, the PRPD spectrum graph judged as discharge is subjected to white noise filtering based on the Kmeans method.

[0016] The Kmeans is used for clustering the amplitudes of the PRPD spectrum graph phase intervals, and k is set to 2 during clustering.

[0017] For each phase interval, threshold selection is performed based on the clustering result to remove low-amplitude white noise, and PRPD white noise filtering processing is completed.

[0018] Further, the PRPD spectrum graph is composed of 360 degrees of phase on the horizontal axis, and the 360 degrees are divided into 36 regions, each region ranging from 0 to 35, and each region is provided with 10 phase angle data points.

[0019] Further, the PRPD spectrum graph after filtering is subjected to PD cluster clustering based on the hierarchical density clustering method, and the clusters after clustering are subjected to feature extraction.

[0020] The hierarchical density clustering method is used for clustering the PRPD spectrum graph after filtering, and the number of clusters is not greater than 4.

[0021] Each cluster after clustering is subjected to feature extraction, and 4 features are extracted for each cluster, namely the length of the cluster, the height of the cluster, the gradient from the upper right to the lower left of the cluster, and the gradient from the upper left to the lower right of the cluster.

[0022] Compared with the prior art, the beneficial effects are:

[0023] The present application filters and extracts features based on the PRPD spectrum graph, not only reduces the data transmission communication pressure of the online monitoring system, but also only extracts features from the PD cluster in the spectrum graph, improves the accuracy of feature extraction, provides theoretical guidance for improving the accuracy of power equipment insulation defect monitoring, and is applicable to power equipment fault detection in different environments. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a partial discharge pattern recognition method based on the combination of Kmeans and DBSCAN according to the present invention.

[0026] Figure 2 This is a schematic diagram of the PRPD filtering method for partial discharge pattern recognition based on a combination of K-means and DBSCAN according to the present invention. Figure 1 ;

[0027] Figure 3 This is a schematic diagram of the PRPD filtering method for partial discharge pattern recognition based on a combination of K-means and DBSCAN according to the present invention. Figure 2 . Detailed Implementation

[0028] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0029] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0030] like Figure 1 As shown, a partial discharge pattern recognition method based on a combination of K-means and DBSCAN includes the following steps:

[0031] S1: Data acquisition, continuously acquires time-domain data from the partial discharge sensor, processes the acquired time-domain data in real time, and generates PRPD spectrum;

[0032] S2: Calculate the envelope waveform discharge / noise of the PRPD spectrum;

[0033] S3: For the PRPD spectrum judged to be discharge, white noise filtering is performed on the PRPD spectrum based on the Kmeans method;

[0034] S4: Based on the hierarchical density clustering method, PD cluster clustering is performed on the filtered PRPD spectrum, and feature extraction is performed on the clustered clusters;

[0035] S5: A model is established, and model training and identification model credibility are obtained;

[0036] S6: The data of the partial discharge fault actually measured are input into the trained model for identification, and the final identification result is obtained.

[0037] In S1, time domain data of the partial discharge sensor is continuously collected for 1s, 20ms data is divided into 360 intervals, the maximum value of the amplitude is calculated in each interval, 360*50 data is generated, and a PRPD spectrum is generated.

[0038] In S2, envelope waveform calculation is performed on the PRPD spectrum, and it is judged whether the waveform envelope is noise or discharge based on the fractal nature of the envelope waveform. The specific steps include:

[0039] Based on the PRPD spectrum, the phase frequency diagram is obtained, and the envelope waveform is calculated;

[0040] Based on the envelope waveform, the Hurst index is calculated, the waveform is divided into 36 intervals, the mean M(n) and the variance S(n) of each interval are calculated, the deviation X and the range R are calculated,

[0041] R(n) = maxX(k) - minX(k);

[0042] In the formula, i is the number of points in the interval, M(n) is the average value of the interval, and k is the interval number; based on the ratio of the range and the standard deviation: R s = R / S;

[0043] In the formula, R s is the waveform index calculated based on this, when R s is around 0.5, it indicates that the waveform sequence is not related at this time, and the data at this time is judged to be noise, otherwise it is discharge, and further data processing is performed.

[0044] In S3, for the PRPD spectrum judged to be discharge, white noise filtering is performed on the PRPD spectrum based on the Kmeans method, as shown in Figures 2-3 , which is a schematic diagram before white noise filtering, Figure Two , which is a schematic diagram after white noise filtering, and the specific steps include: Figure Three

[0045] ​The PRPD spectrum is composed of 360 degrees of phase on the horizontal axis, and 360 degrees is divided into 36 regions, each region ranging from 0 to 35, and each region is provided with 10 phase angle data points.

[0046] The Kmeans is used for clustering the amplitude of the phase interval of the PRPD spectrum, and k is set to 2 during clustering. When there is only noise in the PRPD, the clustering result is that the distances of the two clusters are relatively close. When there is a PD cluster, the clustering result is that the distances of the two clusters are relatively far.

[0047] For each phase interval, the threshold value is selected based on the clustering result to remove the white noise with low amplitude, and the PRPD white noise filtering processing is completed. The specific method for selecting the threshold value for each interval is as follows: for the phase interval, when the distance between the two clusters is far, the maximum amplitude in the cluster with low amplitude is selected as the noise filtering threshold value of the interval; if the distance between the two clusters is close, the centroid of the higher cluster is selected as the noise threshold value of the interval. In the embodiment, the distance threshold value between clusters is 0.4 in the embodiment, that is, when the distance between the centroids of two clusters is less than 0.4, it is considered that the distance is relatively close, and when it is greater than 0.4, it is considered that the distance is relatively far. The threshold value can be adjusted according to the specific situation in actual use.

[0048] Finally, the maximum value in the noise threshold values of the 36 intervals is selected as the noise threshold value of the overall filtering of the PRPD, and the underlying noise signal is filtered out.

[0049] In the embodiment, in S4, the PD cluster is clustered based on the hierarchical density clustering method for the filtered PRPD spectrum, and the features of the clustered clusters are extracted, which specifically includes:

[0050] The filtered PRPD spectrum is clustered based on the hierarchical density clustering method, and the number of clusters is not greater than 4.

[0051] The features of each cluster after clustering are extracted, and four features are extracted for each cluster, which are: the length L of the cluster, the height H of the cluster, the gradient T1 from the upper right to the lower left of the cluster, and the gradient T2 from the upper left to the lower right of the cluster.

[0052] The length of the cluster is the distance from the leftmost point to the rightmost point of the cluster, that is, the horizontal coordinate value of the PRPD spectrum, L=x max -x min .

[0053] The height of the cluster is the value corresponding to the uppermost point of the cluster minus the value corresponding to the lowermost point of the cluster, that is, the vertical coordinate value of the PRPD spectrum, H=Y max -Y min .

[0054] The gradient from the upper right to the lower left of the cluster is calculated from the uppermost and rightmost point of the cluster to the lowermost and leftmost point, that is, T1=(y Rmax-y Lmin ) / (x Rmax -x Lmin )。

[0055] The gradient from the upper left to the lower right of the cluster is calculated from the uppermost left point to the lowermost right point, T2=(y Lmax -y Rmin ) / (x Lmax -x Rmin )。

[0056] In S5, the acquisition data is calculated based on the typical defect type, the sample effective feature is extracted in the method of steps S2, S3 and S4, the extracted feature is input into the radial basis neural network recognition model training, and the model is obtained.

[0057] The application filters and extracts features based on the PRPD spectrum, reduces the data transmission communication pressure of the online monitoring system, extracts features from the PD cluster in the spectrum only, improves the accuracy of feature extraction, provides theoretical guidance for improving the accuracy of power equipment insulation defect monitoring, and is suitable for power equipment fault detection in different environments.

[0058] The above only illustrates the embodiments of the application and is not used to limit the application, any modification, equivalent replacement, improvement, etc. made by the skilled in the art within the spirit and principle of the application without creative labor should be included in the protection scope of the application.

Claims

1. A partial discharge pattern recognition method based on a combination of K-means and DBSCAN, characterized in that, The steps include the following: Data acquisition involves continuously acquiring time-domain data from the partial discharge sensor, processing the acquired time-domain data in real time, and generating PRPD spectra. Calculate the envelope waveform discharge / noise from the PRPD spectrum; For PRPD spectra identified as discharge, white noise filtering is performed on the PRPD spectra based on the Kmeans method; The filtered PRPD spectrum is clustered into PD clusters using hierarchical density clustering, and features are extracted from the clusters. Build a model, train the model, and identify the model's credibility. The measured data of partial discharge faults are input into the trained model for identification, and the final identification result is obtained. The process of performing white noise filtering on the PRPD spectrum determined to be discharged, based on the K-means method, includes: The amplitude of the PRPD spectrum phase interval was clustered using K-means with k set to 2. When only noise exists in the PRPD, the clustering result is that the two clusters are relatively close. When PD clusters exist, the clustering result is that the two clusters are relatively far apart. For each phase interval, a threshold is selected based on the clustering results to remove white noise with low amplitude, thus completing the PRPD white noise filtering process. The PRPD spectrum consists of 360-degree phases on the horizontal axis, which is divided into 36 regions. Each region ranges from 0 to 35, and each region has 10 data points for phase angles. The process of clustering PD clusters in the filtered PRPD spectrum using hierarchical density clustering, followed by feature extraction of the clustered clusters, includes: Clustering of the filtered PRPD spectrum based on hierarchical density clustering method, with no more than 4 clusters; For each cluster after clustering, feature extraction is performed. Four features are extracted for each cluster: cluster length, cluster height, gradient from the top right to the bottom left of the cluster, and gradient from the top left to the bottom right of the cluster.

2. The partial discharge pattern recognition method based on the combination of K-means and DBSCAN according to claim 1, characterized in that, The data acquisition includes continuously acquiring time-domain data from the partial discharge sensor for 1 second.

3. The partial discharge pattern recognition method based on the combination of K-means and DBSCAN according to claim 2, characterized in that, The real-time processing of the acquired time-domain data includes: dividing the data into 360 intervals every 20ms, calculating the maximum amplitude of each interval, and generating 360*50 data points.

Images