GIS ultrahigh frequency partial discharge signal fault diagnosis and state evaluation method

Abstract

A fault diagnosis and state evaluation method for GIS ultrahigh frequency partial discharge signals is characterized by comprising the following steps: S1, based on a physical model of GIS partial discharge typical defects, adopting a specific test loop to carry out a discharge test, collecting ultrahigh frequency partial discharge signals of multiple power frequency periods of different defects, grouping the signals, and carrying out fault diagnosis and state evaluation on the signals; each group comprises a predetermined number of power frequency periodic signals; s2, calculating a partial discharge statistical characteristic quantity, and performing normalization processing to form a data sample corresponding to each group of partial discharge signals; s3, the data samples are regarded as observation values in clustering, different total clustering numbers k are selected, K-means clustering analysis is carried out on each k value, and in-cluster distance sum of squares (WCSS) under each clustering result is calculated; S4, a relation curve of WCSS-k is drawn, and an elbow rule is adopted to preliminarily judge the optimal clustering number; and S5, calculating absolute values of Pearson's correlation coefficients between the WCSS and k under different k values, and when rho begins to rise to be close to 1, determining an optimal clustering number kopt as final clustering output. The optimal clustering number kopt determined by the method can effectively distinguish and classify GIS partial discharge ultrahigh frequency signals, so that different types of discharge signals are distinguished, and powerful support is provided for effective diagnosis of GIS partial discharge defects.

Description

Technical Field

[0001] The present invention belongs to the field of ultra-high frequency partial discharge signal type diagnosis for gas-insulated metal-enclosed switchgear (GIS), and relates to a fault diagnosis and status evaluation method for GIS ultra-high frequency partial discharge signals. The method can use a theoretical calculation method when classifying different GIS ultra-high frequency partial discharge signals instead of selecting the number of clustering categories based on human experience, thereby achieving better classification effect, facilitating the distinction between interference and useful signals, greatly improving anti-interference capability, providing strong support for equipment status maintenance, improving equipment reliability, and facilitating safe and stable operation of power systems. Background Art

[0002] Gas-insulated metal-enclosed switchgear (GIS) has become an indispensable power transmission and transformation equipment in key places such as urban load center substations due to its many advantages. However, GIS will inevitably have insulation defects in the actual operation process, resulting in a gradually deteriorating partial discharge process, which may eventually lead to insulation breakdown, followed by discharge and accidents. If partial discharge monitoring and diagnosis are performed on the operating equipment during this process, the defect type is determined and handled in a timely manner, serious accidents can be effectively avoided and the safe and stable operation of the power system can be guaranteed.

[0003] The UHF method has high detection sensitivity and wide coverage, and is suitable for various occasions of partial discharge detection. It is also commonly used to detect partial discharge signals of insulation defects in GIS. Different types of discharge defects can be identified through UHF partial discharge signals. In terms of identifying partial discharge signals, although current machine learning algorithms have high accuracy in identifying standard defects in laboratory environments, most of these algorithms rely on laboratory data for training, and their performance in field applications has not yet been verified. The complexity of the field environment, including external electromagnetic interference and discharge defect types that are different from laboratory conditions, poses a challenge to accurate identification. If the total number of categories of all discharge signals can be effectively determined and differentiated and classified without labels, the recognition accuracy of partial discharge signal defect types will be significantly improved.

[0004] Clustering technology is one of the most widely used classification methods. Without knowing the sample labels, samples can be distinguished and classified based on the characteristics of the samples and the differences between different samples. K-means clustering is an unsupervised machine learning algorithm. Its main principle is to automatically divide multiple samples in space into k categories. The classification principle is: the similarity of objects of the same category is large, and the similarity of objects of different categories is small.

[0005] However, there are few studies focusing on the selection of the number of clusters k in K-means clustering. The selection of the number of clusters k has a great impact on the clustering effect. If the k value is not selected appropriately, the clustering effect is often not ideal, and the classification results of the data samples may be very different from the actual results. In practical applications, the number of discharge types contained in the UHF discharge data samples on site is often unknown, which means that for different data samples, it is necessary to determine the optimal number of clusters k before clustering. opt It is necessary. But at present opt The selection of the value is often based on experience, and there is no method to determine the optimal clustering number k through theoretical calculation. opt In many cases, the data samples of different defects to be classified and identified are often composed of multiple (greater than 3) local discharge feature quantities, and the sample points may be relatively scattered. It is difficult to manually determine the total number of categories and thus determine k by drawing a sample distribution diagram. opt Although the elbow rule provides a method to determine the optimal number of clusters in k-means clustering, it lacks theoretical support and is difficult to determine the optimal number of clusters k when the inflection point is not prominent. opt . Summary of the invention

[0006] The present invention aims to solve the shortcomings of the current existing technology and proposes a K-means clustering method for GIS ultra-high frequency partial discharge signals based on the elbow rule and the Pearson correlation coefficient to determine the optimal clustering number, in order to achieve the distinction and classification of GIS partial discharge ultra-high frequency signals and identify different types of signals, which will be more conducive to the effective diagnosis of GIS partial discharge defects in the future.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0008] A method for fault diagnosis and status assessment of ultra-high frequency partial discharge signals of GIS is characterized in that it comprises the following steps:

[0009] Step S1. Based on the physical model of typical GIS partial discharge defects, a specific test circuit is used to perform a discharge test, and ultra-high frequency partial discharge signals of multiple power frequency cycles of different defects are collected and grouped, each group containing a predetermined number of power frequency cycle signals;

[0010] Step S2. Calculate the statistical characteristic quantity of partial discharge, and perform normalization processing to form a data sample corresponding to each group of partial discharge signals;

[0011] Step S3. Treat the data samples as observations in clustering, select different total cluster numbers k, perform K-means clustering analysis for each k value, and calculate the intra-cluster distance sum of squares (WCSS) under each clustering result:

[0012] Step S4. Draw the WCSS-k relationship curve and use the elbow rule to preliminarily determine the optimal number of clusters;

[0013] Step S5. Calculate the absolute value of the Pearson correlation coefficient between WCSS and k under different k values. When |ρ| starts to rise close to 1, determine the optimal number of clusters kopt as the final clustering output.

[0014] Preferably, in step S1, the typical defects of GIS local discharge include corona discharge, surface discharge and internal discharge.

[0015] Preferably, the ultra-high frequency partial discharge signal collected in step S1 is preprocessed to ensure that the signal quality meets the requirements of subsequent analysis.

[0016] Preferably, in step S2, the local discharge statistical characteristics include discharge amplitude, discharge times, discharge phase distribution, discharge energy, discharge rise time and discharge duration.

[0017] Preferably, the calculation of the intra-cluster distance sum of squares (WCSS) under each clustering result in step S3 specifically includes:

[0018] When the number of clusters k is 1, the sum of squares of intra-cluster distances (WCSS) is calculated;

[0019] When k is greater than or equal to 2 and less than or equal to a preset positive integer n, a specific clustering method is used to update the cluster center until the maximum number of iterations is reached or the WCSS change amplitude in two consecutive iterations is less than a preset threshold.

[0020] Preferably, the specific clustering method specifically includes:

[0021] When k is greater than or equal to 2 and less than or equal to n, the clustering method is as follows:

[0022] 1) The first k observations are assigned as k cluster centers respectively.

[0023] 2) Traverse each remaining observation to see if it can replace the current cluster center. The judgment principle is: if the distance between the observation point and its nearest cluster center is greater than the distance between the two nearest cluster centers (cluster A and cluster B), then the observation point will replace the cluster center of cluster A or B, depending on which one is closer to the observation point; otherwise, if the distance between the observation point and the second nearest cluster center is greater than the closest distance between the nearest cluster center and the other cluster centers, then the observation point will replace the nearest cluster center.

[0024] 3) Assign the observation point to the nearest cluster; each observation point is assigned to the cluster with the nearest distance. If the i-th observation point is assigned to the k-th cluster, then the distance d between the observation point and the k-th cluster is ik will be corrected to And the distance d between the observation point and other clusters ij Corrected to

[0025]

[0026] Among them, n k and n j are the number of observations in the kth cluster and the jth cluster, respectively, and d ik is the distance between the ith observation point and cluster k, d ij is the distance between the i-th observation point and cluster j. If the distance between the corrected observation point and the l-th cluster is the smallest, and l≠k, then the observation point will be assigned to the l-th cluster instead of the k-th cluster. Then, each cluster center will be updated to the average value of the observation points in each cluster;

[0027] 4) Using the updated cluster centers, repeat the above process (3), assign each observation point, and update the cluster centers; when the maximum number of iterations is reached or the WCSS change in two consecutive iterations is less than a certain threshold, the iteration will stop; the cluster center updated in the last iteration is called the final cluster center.

[0028] Preferably, the elbow rule refers to observing the WCSS-k curve and identifying the point where the slope of the curve changes significantly as the "elbow", and the k value corresponding to the point is the optimal number of clusters preliminarily estimated.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention proposes a K-means clustering method for ultra-high frequency partial discharge signals of GIS, which combines the elbow rule and the Pearson correlation coefficient to determine the optimal clustering number kopt through theoretical calculation. When classifying ultra-high frequency partial discharge data samples of different defects, the optimal clustering number kopt can be determined in advance through theoretical calculation, thereby achieving more accurate and well-founded K-means clustering and improving the accuracy of defect identification. This not only helps to accurately determine the fault type of GIS equipment, but also provides convenience for the operation, maintenance, inspection and fault diagnosis of the equipment, thereby improving the reliability of the equipment and ensuring the safety of production and life of personnel.

[0031] The optimal clustering number kopt determined by this method can effectively distinguish and classify GIS partial discharge UHF signals, and then identify different types of discharge signals, providing strong support for the effective diagnosis of GIS partial discharge defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Test circuit diagram

[0033] Figure 2 WCSS-k curve when k is 1 to 12

[0034] Figure 3 When q is 1 to 11, the absolute value of the Pearson correlation coefficient between sequence k and sequence WCSS |ρ| DETAILED DESCRIPTION

[0035] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments, but this should not be used to limit the protection scope of the present invention.

[0036] In this embodiment, a K-means clustering method for GIS ultra-high frequency partial discharge signals for determining the optimal number of clusters based on the elbow rule and the Pearson correlation coefficient comprises the following steps:

[0037] Step 1: First, based on the physical model of typical defects of GIS partial discharge, Figure 1 The test circuit shown in the figure is used to conduct a discharge test and collect UHF simulation signals: the prepared defect model is placed in the GIS test tank, and the voltage is applied by the step-by-step voltage step-up method until the defect produces a continuous and stable discharge without overcurrent behavior of the test transformer. Under this voltage, the UHF detection method is used to continuously collect UHF signals of multiple power frequency cycles. For different typical defects, 3000 power frequency cycles of UHF partial discharge data are obtained, which are divided into 60 groups, each containing 50 power frequency cycles of signals.

[0038] Step 2: Based on the discharge data, calculate the local discharge statistical characteristic quantity with clear physical meaning, which can characterize the actual discharge characteristics and is intuitive and clear. There are no special requirements for the local discharge statistical characteristic quantity required to be calculated in the present invention, including but not limited to: pulse basic waveform characteristic quantity (such as pulse rise speed, pulse fall speed, etc.), pulse equivalent time-frequency characteristic quantity (such as pulse time domain equivalent time length, frequency domain equivalent bandwidth, etc.), local discharge statistical characteristic quantity based on amplitude-phase time series data (such as A max : The maximum amplitude of each group of partial discharge signals, reflecting the maximum discharge intensity of the defect; A ave : The average amplitude of each group of partial discharge signals, reflecting the overall discharge intensity; A total : The sum of the discharge amplitudes of each group of partial discharge signals, which characterizes the cumulative discharge intensity of the defect over a period of time, including information characteristics such as the discharge repetition rate; N: The total number of discharges of each group of partial discharge signals, which can reflect the repetition rate of discharge within a certain period of time; r A: The ratio of the sum of the amplitude of the discharge signal strength in the positive half cycle of each group of partial discharge signals to the sum of the amplitude of the discharge signal strength in the negative half axis, which reflects the shape characteristics of the defect. Because the discharge strength of the positive and negative half cycles in a cycle is different for defects of different shapes, it can characterize the shape characteristics of the positive and negative half cycles in the partial discharge PRPS spectrum; std ave : The average value of the standard deviation of the discharge amplitude in each cycle in each group of partial discharge signals, which reflects the volatility of the discharge amplitude and can reflect the shape characteristics of the amplitude change in the PRPS spectrum). The characteristic quantity corresponding to each group of partial discharge signals is taken as a data sample, and 60 data samples can be obtained.

[0039] In order to eliminate the influence of different units between different feature quantities, the statistical feature quantities are normalized to the interval [0,1] before using K-means clustering. The normalization method is to perform a normalization on each feature quantity in all the data of this type of feature quantity. operation, where x i is the i-th feature quantity, x min is the minimum value of the feature quantity of this type among all samples, x max It is the maximum value of this type of feature among all samples.

[0040] Step 3: Treat the above 60 data samples as observations in the cluster:

[0041] When the number of clusters k is 1, all data samples belong to the same category. The average value of all sample observations is used as the cluster center and the within-cluster distance sum (Within Cluster Sum of Square, WCSS) is calculated. WCSS is the sum of the distances between the cluster center of each cluster in the sample space and the sample points within the cluster. The calculation method is:

[0042]

[0043] In the formula, m is the total number of all sample points, n represents the number of features (i.e., dimension) of each sample point, and X j is the jth sample point, μ j is the sample point X j The cluster center of the cluster, x ji Represents the sample point X j The eigenvalue of the i-th dimension, μ ji Represents the cluster center μ j The eigenvalue of the i-th dimension.

[0044] In this method, the jth sample in each cluster is related to the cluster center μ of the cluster to which the sample belongs. j The distance between is Euclidean distance. The calculation method is:

[0045]

[0046] In the formula, d(X j ,μ j ) represents the jth sample X j and the cluster center μ j The Euclidean distance between j represents the jth sample point in the cluster, μ j Represents X j The centroid of the cluster, x ji Represents the sample point X j The eigenvalue of the i-th dimension, n represents the number of features (i.e., dimension) of each sample point, μ ji Represents the cluster center μ j The eigenvalue of the i-th dimension.

[0047] Define n as a positive integer much larger than the actual total number of clusters. When k is greater than or equal to 2 and less than or equal to n, the clustering method is as follows:

[0048] 1) The first k observations are assigned as k cluster centers respectively.

[0049] 2) Traverse each remaining observation to see if it can replace the current cluster center. The judgment principle is: if the distance between the observation point and its nearest cluster center is greater than the distance between the two nearest cluster centers (cluster A and cluster B), then the observation point will replace the cluster center of cluster A or B, depending on which one is closer to the observation point; otherwise, if the distance between the observation point and the second nearest cluster center is greater than the closest distance between the nearest cluster center and the other cluster centers, then the observation point will replace the nearest cluster center.

[0050] 3) Assign the observation point to the nearest cluster. Each observation point is assigned to the cluster with the closest distance. If the i-th observation point is assigned to the k-th cluster, then the distance d between the observation point and the k-th cluster is ik will be corrected to And the distance d between the observation point and other clusters ij Corrected to

[0051]

[0052] Among them, n k and n j are the number of observations in the kth cluster and the jth cluster, respectively, and d ik is the distance between the ith observation point and cluster k, d ijis the distance between the i-th observation and cluster j. If the distance between the corrected observation and the l-th cluster is the smallest and l≠k, then the observation will be assigned to the l-th cluster instead of the k-th cluster. Then, each cluster center will be updated to the average value of the observations in each cluster.

[0053] 4) Using the updated cluster center, repeat the above process (3), assign each observation point, and update the cluster center. When the maximum number of iterations is reached (the default maximum number of iterations in this method is 10, which can be modified as needed) or the WCSS change in two consecutive iterations is less than a certain threshold (which can be modified according to the data), the iteration will stop. The cluster center updated in the last iteration is called the final cluster center.

[0054] In the above clustering process, the centroid of each cluster of sample points is selected as the clustering center of each cluster, and the value of each dimension of the centroid is the average value of each cluster of data points in that dimension.

[0055] Step 4: Plot the WCSS results calculated for each k in the two-dimensional coordinate system with the horizontal axis being the cluster number k and the vertical axis being the sum of squared distances WCSS within the cluster. Table 1 gives an implementation example with 3 defect categories and n=12, and gives the results of the sum of squared distances WCSS within the cluster after clustering with k=1, 2, 3, 4, ..., 12, and plots the sum of squared distances WCSS within the cluster-cluster number k curve, as shown in Table 1. Figure 2 shown.

[0056] Table 1 shows the sum of squares of intra-cluster distances WCSS for different k values

[0057]

[0058]

[0059] from Figure 2 It can be seen that k = 3 is an inflection point, which corresponds to the actual number of clusters in this experiment, that is, the optimal number of clusters k that needs to be determined opt .because Figure 2 The middle curve is shaped like a human arm, and the optimal number of clusters k opt The corresponding inflection point is like the elbow of the arm, so the optimal number of clusters k is found in this way opt This method is also called the "elbow rule".

[0060] In order to more accurately determine and identify the "elbow" of the curve, that is, to find the optimal number of clusters k opt ,This paper adopts a mathematical calculation method.

[0061] In a similar Figure 2In the WCSS-k curve, it is generally believed that when WCSS begins to decrease linearly, the k value corresponding to this point is the optimal clustering number k. opt Therefore, when q=2, 3, 4, ..., n-1, the sequence WCSS={WCSS q ,WCSS q+1 ,……,WCSS n The Pearson correlation coefficient ρ between} and the sequence k = {q, q+1, ..., n}. The Pearson correlation coefficient ρ is a commonly used linear correlation coefficient that can reflect the degree of linear correlation between two variables. It is defined as the quotient of the covariance and standard deviation between the two variables. It is calculated as follows:

[0062]

[0063] In the formula, X and Y are variables, cov(X,Y) is the covariance of X and Y, σ X , σ Y are the standard deviations of X and Y, respectively, n is the total number of values ​​contained in variables X and Y, respectively, i is the i-th value of variable X, Y i , Y j are the i-th and j-th values ​​of variable Y respectively, are the means of variables X and Y respectively.

[0064] The value of Pearson correlation coefficient ρ is [-1,1]. When ρ<0, variables X and Y are negatively correlated; when ρ>0, variables X and Y are positively correlated; when ρ=0, there is no linear correlation between variables X and Y. The closer the absolute value of ρ |ρ| is to 1, the stronger the linear correlation between variables X and Y. Table 2 is used as the standard for evaluating the degree of correlation between variables.

[0065] Table 2 Absolute value of Pearson correlation coefficient |ρ| and correlation between variables

[0066]

[0067]

[0068] Therefore, when q=2, 3, 4, ..., n-1, the sequence k={q, q+1, ..., n} is calculated respectively, and the sequence WCSS={WCSS q ,WCSS q+1 ,……,WCSS n The absolute value of the Pearson correlation coefficient |ρ| between} and sequence k.

[0069] When q starts to take a certain value q optWhen |ρ| rises to a value close to 1, it can be considered that WCSS changes from k=q opt It starts to decrease linearly, and the corresponding k value is the optimal cluster number k opt , that is, k opt =q opt .

[0070] An implementation example with n=12 is given below, as shown in Table 3 and Figure 3 shown.

[0071] Table 3 Absolute value of Pearson correlation coefficient when q is 1 to 11 |ρ|

[0072]

[0073] When q increases from 2 to 3, the absolute value of the Pearson correlation coefficient |ρ| increases rapidly from 0.518818881 to 0.908874263, which is close to 1. This method considers that when |ρ| starts to rise to ≥0.8, the two variables can be regarded as highly linearly correlated, that is, at this time, the sequence k={q,q+1,……,n-1} and the sequence WCSS={WCSS q ,WCSS q+1 ,……,WCSS n}(q=2, 3, 4, ..., n-1) are highly linearly correlated, so it can be considered that q opt =3, which means that the WCSS-k curve is close to linear when k is {3, 4, ..., n}, and WCSS begins to decrease linearly. The optimal cluster number k opt =q opt =3, which is consistent with Figure 2 The result of the preliminary judgment based on the inflection point is consistent with that in the previous section. So far, by using a more rigorous mathematical method, by calculating the absolute value of the Pearson correlation coefficient |ρ| from a small value to the point close to 1, the corresponding q opt , the optimal number of clusters k is determined opt .

[0074] Finally, output k = k opt The clustering results when the number of clusters k is 3 in this example are given below as shown in Table 4.

[0075] Table 4 Clustering results when the number of clusters k is 3

[0076]

[0077] It can be seen from the table that when the number of clusters is 3, the data of each defect type can be accurately classified, the clustering effect is good, and the accuracy rate is 100%, that is, reasonable and accurate K-means clustering of discharge samples with different defects is achieved, which shows that by using this method to find the optimal k value, a better K-means clustering effect can be achieved.

[0078] The present invention proposes a K-means clustering method for GIS ultra-high frequency partial discharge signals based on the elbow rule and the Pearson correlation coefficient to determine the optimal number of clusters. Assuming that the total number of clusters is less than n, the absolute value of the Pearson correlation coefficient |ρ| between the cluster number k sequence {q, q+1, ..., n} corresponding to different q (q is an integer greater than 1) and the intra-cluster distance square sum WCSS sequence is calculated. When |ρ| increases to >0.8, the corresponding k value can be used as the reasonable cluster category number k. opt Theoretical calculation value of the value. This method can reasonably select the optimal clustering number k when classifying the local discharge statistical feature data samples of different defects. opt , and then achieve more accurate K-means clustering. Based on the clustering results, support vector machines, artificial neural networks and other pattern recognition algorithms can be used to more accurately identify and diagnose GIS partial discharge defect types. This method can also accurately classify partial discharge samples with different characteristics before identification, so it can distinguish between noise and defect discharge signals, which can greatly improve the anti-interference ability, provide strong support for the condition maintenance of equipment, improve the reliability of equipment, and facilitate the safe and stable operation of the power system.

[0079] In summary, this technical solution successfully solves the problem of determining the optimal number of clusters in cluster analysis of GIS ultra-high frequency partial discharge signals by combining the elbow rule and the Pearson correlation coefficient, improves the accuracy and reliability of cluster analysis, and provides strong technical support for fault diagnosis and status assessment of GIS equipment.

Claims

1. A method for fault diagnosis and status assessment of GIS ultra-high frequency partial discharge signals, characterized in that: The following steps are involved: Step S1. Based on the physical model of typical GIS partial discharge defects, a specific test circuit is used to perform a discharge test, and ultra-high frequency partial discharge signals of multiple power frequency cycles of different defects are collected and grouped, each group containing a predetermined number of power frequency cycle signals; Step S2. Calculate the statistical characteristic quantity of partial discharge, and perform normalization processing to form a data sample corresponding to each group of partial discharge signals; Step S3. Treat the data samples as observations in clustering, select different total cluster numbers k, perform K-means clustering analysis for each k value, and calculate the intra-cluster distance sum of squares (WCSS) under each clustering result: Step S4. Draw the WCSS-k relationship curve and use the elbow rule to preliminarily determine the optimal number of clusters; Step S5. Calculate the absolute value of the Pearson correlation coefficient between WCSS and k under different k values. When |ρ| starts to rise close to 1, determine the optimal number of clusters kopt as the final clustering output.

2. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 1 is characterized in that: The typical defects of GIS partial discharge in step S1 include corona discharge, surface discharge and internal discharge.

3. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 1 is characterized in that: The ultra-high frequency partial discharge signal collected in step S1 is pre-processed to ensure that the signal quality meets the requirements of subsequent analysis.

4. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 1 is characterized in that: In the step S2, the local discharge statistical characteristics include discharge amplitude, discharge times, discharge phase distribution, discharge energy, discharge rise time and discharge duration.

5. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 1 is characterized in that: The step S3 calculates the intra-cluster distance sum of squares (WCSS) under each clustering result, specifically including: When the number of clusters k is 1, the sum of squares of intra-cluster distances (WCSS) is calculated; When k is greater than or equal to 2 and less than or equal to a preset positive integer n, a specific clustering method is used to update the cluster center until the maximum number of iterations is reached or the WCSS change amplitude in two consecutive iterations is less than a preset threshold.

6. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 4 is characterized in that: The specific clustering method specifically includes: When k is greater than or equal to 2 and less than or equal to n, the clustering method is as follows: 1) The first k observations are assigned as k cluster centers respectively. 2) Traverse each remaining observation to see if it can replace the current cluster center. The judgment principle is: if the distance between the observation point and its nearest cluster center is greater than the distance between the two nearest cluster centers (cluster A and cluster B), then the observation point will replace the cluster center of cluster A or B, depending on which one is closer to the observation point; otherwise, if the distance between the observation point and the second nearest cluster center is greater than the closest distance between the nearest cluster center and the other cluster centers, then the observation point will replace the nearest cluster center. 3) Assign observation points to the nearest cluster; each observation point is assigned to the cluster with the nearest distance. If the i-th observation point is assigned to the k-th cluster, then the distance d between the observation point and the k-th cluster is ik will be corrected to And the distance d between the observation point and other clusters ij Corrected to Among them, n k and n j are the number of observations in the kth cluster and the jth cluster, respectively, and d ik is the distance between the ith observation point and cluster k, d ij is the distance between the i-th observation point and cluster j. If the distance between the corrected observation point and the l-th cluster is the smallest, and l≠k, then the observation point will be assigned to the l-th cluster instead of the k-th cluster. Then, each cluster center will be updated to the average value of the observation points in each cluster; 4) Using the updated cluster centers, repeat the above process (3), assign each observation point, and update the cluster centers; when the maximum number of iterations is reached or the WCSS change in two consecutive iterations is less than a certain threshold, the iteration will stop; the cluster center updated in the last iteration is called the final cluster center.

7. The fault diagnosis and status assessment method of GIS ultra-high frequency partial discharge signal according to claim 1 is characterized in that: The elbow rule refers to observing the WCSS-k curve and identifying the point where the slope of the curve changes significantly as the "elbow". The k value corresponding to this point is the preliminary estimated optimal number of clusters.

Images