A GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization

Abstract

The application discloses a GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization, and belongs to the field of large power grid on-site power equipment defect diagnosis and recognition. Firstly, aiming at the signal missing phenomenon of photoelectric partial discharge signals, a non-subsampled contourlet transform is used to fuse and process photoelectric PRPD graphs, so that photoelectric fusion graphs are obtained. Then, feature points and scale vectors of the fusion PRPD graphs are extracted based on a KAZE algorithm. Next, the feature points are dispersed to polar coordinate expressions according to coordinates and scale vectors. Distribution entropy of each region divided on the polar coordinates is calculated to form a feature vector. Finally, the feature vector is loaded into a long short-term memory network optimized by an Adaboost algorithm to verify the accuracy of partial discharge pattern recognition. Compared with a statistical parameter method and a KAZE feature extraction method, the algorithm can better extract a PRPD graph feature vector, and the accuracy of partial discharge defect pattern recognition is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of defect diagnosis and identification technology for power equipment in large power grids, and more specifically to a GIS partial discharge pattern identification method based on KAZE and polar coordinate distribution entropy optimization. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in power systems due to its broad applicability, excellent insulation performance, and reliable stability. However, due to the unique internal structure of GIS, certain insulation defects may occur during production, transportation, and assembly. Insulation aging caused by long-term operation may also lead to insulation faults in GIS, resulting in partial discharge or even insulation breakdown, thus endangering power grid safety. Different insulation defects cause different partial discharge modes, with significantly varying degrees of threat to power grid safety. Therefore, the efficiency of partial discharge pattern recognition in GIS is crucial for equipment condition monitoring and fault diagnosis.

[0003] In recent years, image feature extraction technology has rapidly developed in the field of remote sensing registration. Significant features in images can be spontaneously extracted, and numerous related algorithms, such as scale-invariant feature transform (SIFT), speed-up robust features (SURF), and binary robust invariant salable keypoints (BRISK), have been proposed and applied to feature descriptor extraction. However, remote sensing images, especially optical and SAR images, exhibit significant nonlinear radiometric differences, weakening the feature extraction capabilities of these algorithms. To address this issue, a multi-scale two-dimensional feature registration algorithm based on nonlinear scale space (KAZE) has been proposed and applied. Some literature uses an accelerated KAZE algorithm for feature extraction from aerial images, solving the problem of poor real-time performance and improving image registration accuracy. Other literature first extracts more information-rich parts from high-resolution, large-scale images by setting a threshold, and then performs accelerated KAZE feature extraction, ultimately achieving registration accuracy rates exceeding 90%. However, in GIS partial discharge pattern recognition, the feature vector obtained by the KAZE algorithm has a huge dimension, which is very easy to cause data redundancy and overfitting. Moreover, the registration results are usually given by Hanning distance between a small number of images, making it difficult to apply to partial discharge pattern recognition with a large number of images.

[0004] Therefore, proposing a GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization, which can better extract PRPD map feature vectors and improve pattern recognition accuracy, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a GIS partial emission pattern recognition method based on polar coordinate distribution entropy optimization, which improves the pattern recognition accuracy and has better robustness and more stable recognition results.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A GIS partial-effect pattern recognition method based on polar coordinate distribution entropy optimization includes the following steps:

[0008] S1: Simulate the partial discharge signal of GIS through experimental simulation, collect it using an optical integrated sensor, and convert it into a photoelectric PRPD spectrum;

[0009] S2: The photoelectric PRPD map is fused using the NSCT image fusion algorithm to obtain the photoelectric fused PRPD map;

[0010] S3: Use the KAZE algorithm to extract feature points from the optoelectronic fusion PRPD map, and obtain the coordinate information and description vector of the feature points. The description vector contains the scale information of the feature points.

[0011] S4: The feature points containing coordinate and scale information are dispersed to polar coordinates. The distribution entropy of each sub-region divided on the polar coordinates is calculated to form a feature vector, and the optimal feature vector is selected.

[0012] S5: Input the optimal feature vector into the Adaboost-LSTM network model for partial discharge pattern recognition.

[0013] Preferably, the photoelectric PRPD spectrum includes an optical PRPD spectrum and an ultra-high frequency PRPD spectrum;

[0014] The optical integrated sensor integrates an optical sensor and an ultra-high frequency sensor, which respectively convert the GIS partial discharge signal into the optical PRPD spectrum and the ultra-high frequency PRPD spectrum.

[0015] Preferably, S2 includes:

[0016] S21: Perform multi-scale decomposition on the optical PRPD spectrum and the ultra-high frequency PRPD spectrum respectively to obtain high-frequency sub-band maps and low-frequency sub-band maps at each level; perform multi-directional decomposition on the high-frequency sub-band map to finally obtain the low-frequency sub-band coefficients and high-frequency sub-band coefficients in multiple different directions of the optical PRPD spectrum and the ultra-high frequency PRPD spectrum respectively.

[0017] S22: Based on the fusion rule of averaging the low-frequency subband coefficients and taking the larger high-frequency subband coefficients, the photoelectric PRPD spectrum and the ultra-high frequency PRPD spectrum are fused at each scale and direction to obtain a photoelectric fused sub-map.

[0018] S23: Perform NSCT inverse transform on the photoelectric fusion subgraph to obtain the photoelectric fusion PRPD map.

[0019] Preferably, S3 includes:

[0020] S31: Based on the original image, a nonlinear scale space is constructed using an additive operator splitting algorithm and a variable propagation-diffusion method;

[0021] S32: At different scales in the nonlinear scale space, any point is compared with each point in its neighborhood to find the maximum point of the normalized Hessian matrix as the feature point; the position coordinates of the feature point are solved by Taylor expansion fitting.

[0022] S33: Determine the main direction of the feature point based on its neighborhood, and construct a description vector for each feature point using the M-SURF algorithm.

[0023] Preferably, in step S33, the M-SURF algorithm is used to construct a description vector for each feature point, including:

[0024] S331: For scale s i Feature points, with them as the center, select a size of 24s i ×24s i The matrix neighborhood is divided into 4×4 sub-regions, with an overlap of 2s between each pair of sub-regions. i ;

[0025] S332: Perform Gaussian weighting on each sub-region to obtain a 4-dimensional feature point description vector d. v as follows:

[0026] d v =(∑L x ,∑L y ,∑|L x |,∑|L y |);

[0027] Among them, L x and L y Let x and y be the differentials at x and y, respectively;

[0028] S333: Using a Gaussian kernel for the 4D description vector d of each sub-region v Weighted summation and normalization are performed to obtain a 64-dimensional description vector for each feature point.

[0029] Preferably, S4 includes:

[0030] S41: Using the origin of the photoelectric fusion PRPD map as the pole, construct a polar coordinate system, initialize the number of polar angle segments p and the number of polar radius segments q, divide the polar angle into p segments and the polar radius into q segments with the pole as the center, and divide the polar coordinate system into several sub-regions;

[0031] S42: The extracted feature points are dispersed into polar coordinates according to coordinate and scale information, and the distribution entropy of each sub-region is calculated to form a feature vector;

[0032] S43: Calculate the intra-class scatter and inter-class scatter of the feature vector and determine whether it is optimal; if yes, obtain the optimal feature vector; if no, update the values ​​of the number of polar angle segments p and the number of polar diameter segments q, and return to S41.

[0033] Preferably, the distribution entropy H c The calculation formula is as follows:

[0034] H c =-P j log c P j ;

[0035] Among them, P j The probability that a feature point falls within region j is calculated using the following formula:

[0036]

[0037] n j Let P be the number of feature points falling within region j, and M be the number of subregions, satisfying P1 + P2 + ... + P M =1.

[0038] Preferably, the intra-class scatter and inter-class scatter of the feature vector are calculated as follows:

[0039]

[0040]

[0041] Where ε1 is the intra-class dispersion of the same type of discharge defect, ε2 is the inter-class dispersion of different types of discharge defect, and Hkj Let E(H) be the j-th dimension of the feature vector of the k-th sample. j Let E(H) be the expectation of the j-th feature. aj ), E(H bj ) are the expected values ​​of the j-th dimension features of the feature vectors of discharge defects of types a and b, respectively, and K is the total number of samples.

[0042] As can be seen from the above technical solution, compared with the prior art, this invention discloses a GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization. Addressing the problem of insufficient feature extraction accuracy in partial discharge defect pattern recognition, this invention spontaneously extracts image feature point information using the KAZE algorithm, then diverges these feature points to polar coordinates based on coordinate and scale information, extracts the sub-region distribution entropy as feature information, and finally uses Adaboost-LSTM classification. This method effectively avoids the problem of feature parameter selection relying on human experience, improves pattern recognition accuracy, and has better robustness and more stable recognition results. It helps staff to understand the equipment's operating status in a timely manner, ensuring the stability and safety of the equipment. Attached Figure Description

[0043] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 The flowchart of the GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization is provided in the embodiment of the present invention.

[0045] Figure 2 The partial discharge spectra before and after fusion are provided in the embodiments of the present invention; (a) represents the photoelectric PRPD spectrum before fusion, (b) represents the ultra-high frequency PRPD spectrum before fusion, and (c) represents the photoelectric fusion PRPD spectrum after fusion.

[0046] Figure 3 The following are feature point distribution maps extracted based on the KAZE algorithm provided for embodiments of the present invention: (a) represents the feature point distribution map of suspended discharge defects; (b) represents the feature point distribution map of free particle discharge defects; and (c) represents the feature point distribution map of tip discharge defects.

[0047] Figure 4The following are polar coordinate distribution diagrams of feature points provided in the embodiments of the present invention: (a) polar coordinate distribution diagram of feature points of suspended discharge defects; (b) polar coordinate distribution diagram of feature points of free particle discharge defects; (c) polar coordinate distribution diagram of feature points of tip discharge defects; and (d) polar coordinate distribution comparison diagram of feature points of three defect types.

[0048] Figure 5 The intraclass dispersion values ​​of various partial discharge signals provided in the embodiments of the present invention.

[0049] Figure 6 The average value of inter-class dispersion is provided in the embodiments of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This invention discloses a GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization, the flowchart of which is shown below. Figure 1 As shown. The specific steps are as follows:

[0052] S1: Simulate the partial discharge signal of GIS through experimental simulation, collect it using an optical integrated sensor, and convert it into a photoelectric PRPD spectrum;

[0053] The partial discharge signal in GIS here includes three defect modes: tip discharge defect, suspension discharge defect, and free particle discharge defect.

[0054] Specifically, photoelectric PRPD maps include optical PRPD maps and ultra-high frequency PRPD maps;

[0055] The optical integrated sensor integrates an optical sensor and an ultra-high frequency sensor, which respectively convert the GIS partial discharge signal into an optical PRPD spectrum and an ultra-high frequency PRPD spectrum.

[0056] S2: The NSCT image fusion algorithm is used to fuse the photoelectric PRPD maps, resulting in a photoelectric fused PRPD map. The maps before and after fusion are shown below. Figure 2 As shown. The NSCT image fusion algorithm comprises two processes: decomposition and fusion. The optical PRPD spectrum in the fused optoelectronic PRPD spectrum effectively compensates for the lack of a positive half-cycle in the UHF PRPD spectrum, specifically including:

[0057] S21: NSCT decomposition. Three-level scale decomposition was performed on both the optical PRPD and UHF PRPD maps, resulting in one low-frequency sub-band map and three high-frequency sub-band maps. Then, the high-frequency sub-band maps were decomposed using level 1, 2, and 3 multi-directional decomposition, dividing them into 2, 4, and 8 directional sub-band maps respectively. That is, each input map was decomposed into 15 sub-band maps of the same size as the source map, effectively preserving the image's detailed information in all directions.

[0058] S22: Based on the fusion rule of averaging the low-frequency subband coefficients and taking the larger high-frequency subband coefficients, the photoelectric PRPD spectrum and the ultra-high frequency PRPD spectrum are fused at each scale and direction to obtain the photoelectric fused sub-map.

[0059] S23: Perform NSCT inverse transform on the optoelectronic fusion subgraph to obtain the optoelectronic fusion PRPD map.

[0060] S3: Use the KAZE algorithm to extract feature points from the photoelectric fusion PRPD map, obtain the coordinate information and description vector of the feature points, and the description vector contains the scale information of the feature points.

[0061] The KAZE algorithm constructs a scale space through nonlinear diffusion filtering. The high-scale space describes the image contour information, while the low-scale space describes the image texture, edges, and other details. It also uses M-SURF to generate the image feature descriptor. Therefore, this algorithm can better preserve image information and reduce the impact of noise.

[0062] S31: The KAZE algorithm is based on the original graph operation and uses the Additive Operators Splitting (AOS) algorithm and variable propagation diffusion to construct a nonlinear scale space, as shown in the following equation:

[0063]

[0064] Where L is the Gaussian filtered image, I is the identity matrix, t is time, and A l (L i Let be the transmission matrix of image L in dimension i.

[0065] S32: At different scales, compare any point with each point in its neighborhood to find the maximum point of the normalized Hessian matrix, and then use Taylor expansion to fit and approximate the position coordinates of the feature point.

[0066] The formula for calculating the Hessian matrix is ​​as follows:

[0067]

[0068] Where σ is the scale parameter σ i Integer value, Lxx L yy L xy All are second-order derivatives of L.

[0069] The location of the maximum point can be determined using Taylor expansion, as shown in the following equation:

[0070]

[0071] Where x is the position coordinate of the feature point, and the sub-pixel coordinates of the feature point are as follows:

[0072]

[0073] To improve the search speed of the KAZE algorithm, the search range for the maximum point of the Hessian matrix is ​​fixed at a 3×3 matrix. That is, each selected pixel is compared with 18 pixels at its neighboring scale and 8 pixels at its adjacent scale to obtain the maximum point within the matrix range. Taylor expansion is then used to accurately locate the feature point within this scale. Regarding the number of feature points, different thresholds can be set to change the number of feature points extracted. Too few feature points will result in incomplete PRPD map feature extraction, with some edge and texture information being ignored. Too many feature points will lead to data redundancy, with a large amount of unimportant data being considered, reducing recognition accuracy. In this embodiment, the mean of the 64-dimensional description vector is used as the scale information of the feature points. The feature vectors of the KAZE algorithm are constructed by arranging the feature points in descending order of importance. Six cases were selected: 50, 100, 150, 200, 250, and 300 feature points. Finally, 200 feature points were selected.

[0074] S33: After determining the main orientation of a feature point based on its neighborhood, the M-SURF algorithm is used to construct a description vector for each feature point. For a scale of s... i Feature points, with them as the center, select a size of 24s i ×24s i The matrix neighborhood is divided into 4×4 sub-regions, with an overlap of 2s between each pair of sub-regions. i Finally, Gaussian weighting is applied to each sub-region to obtain a 4-dimensional feature point description vector d. v .

[0075] d v =(∑L x ,∑L y ,∑|L x |,∑|L y |);

[0076] Among them, L x and L yLet x and y be the differentials, respectively. Then, use a Gaussian kernel to evaluate the 4D description vector d of each of the 4×4 sub-regions divided in the matrix neighborhood. v After weighted summation and normalization, a 64-dimensional description vector for each feature point is obtained.

[0077] S4: Distribute the feature points containing coordinate and scale information to polar coordinates, calculate the distribution entropy for each sub-region divided in polar coordinates, construct a feature vector, and select the optimal feature vector. This includes:

[0078] S41: Construct a polar coordinate system with the origin of the optoelectronic fusion PRPD map as the pole, initialize the number of polar angle segments p and the number of polar radius segments q, and divide the polar angle into p segments and the polar radius into q segments with the pole as the center. In this embodiment, p = 6 and q = 3 (the optimal values ​​are obtained below based on the intra-class and inter-class dispersion). Divide the polar coordinate system into 60 sub-regions, and set the spacing of each grid in the polar radius direction to 0.2 after normalization.

[0079] S42: The 200 feature points extracted from each map are distributed into polar coordinates according to their coordinates and scale. Specifically, the feature points containing feature information and scale information are distributed into rectangular coordinates (x, y, z). i y i The vectors are diverged to polar coordinates (ρ, θ) by a rotation factor ξ, and the mean of the vectors describing each feature point is taken as the polar radius. The polar coordinates (ρ, θ) are obtained by the following formula:

[0080]

[0081]

[0082] Where N is the dimension of the feature point description vector, S i is the description vector of feature point i, where ρ and θ are the polar radius and polar angle, respectively, and the rotation factor ξ is set to 4 to ensure that the feature point diverges across the entire polar coordinate system.

[0083] like Figure 4 As shown, the feature points of the three defect types are distributed in a ring around the pole in polar coordinates. However, the distribution radius of the feature points of the suspended discharge, particle discharge, and tip discharge defects increases sequentially, and there are obvious differences in distribution. The number of feature points in each sub-region is counted, and the distribution probability and distribution entropy of each sub-region are calculated according to the following formula. The feature vector H of the PRPD map is composed of the distribution entropy of each sub-region. f = (H1, H2, ..., H M ):

[0084]

[0085] Among them, P jn is the probability that a feature point falls within region j. j Let P be the number of feature points falling within region j, and M be the number of subregions, satisfying P1 + P2 + ... + P M =1. Define the distribution entropy within each sub-region as H. c ,

[0086] H c =-P j log c P j .

[0087] S43: Determine the p and q values ​​by using the optimal intra- and inter-class dispersion of the feature vectors. The greater the inter-class dispersion of each type of sample, the higher the discrimination; the smoother the fluctuation of the intra-class dispersion curve, the smaller the impact. Based on this, determine the optimal intra- and inter-class dispersion.

[0088] To determine the optimal values ​​of p and q and obtain the optimal feature vector, this embodiment uses the feature vectors H of three defect types. f To observe the target, the values ​​of p and q were changed, and the intra-class and inter-class dispersion of the feature vectors for the three types of discharge defects were calculated as the basis for the values. The calculation formula is as follows:

[0089]

[0090]

[0091] Where ε1 is the intra-class dispersion of the same type of discharge defect, ε2 is the inter-class dispersion of different types of discharge defect, and H kj Let E(H) be the j-th dimension of the feature vector of the k-th sample. j Let E(H) be the expectation of the j-th feature. aj ), E(H bj ) are the expected values ​​of the j-th dimension features of the feature vectors of discharge defects of types a and b, respectively, and K is the total number of samples.

[0092] In this embodiment, the number of polar angle segments p is set to 6, 9, 12, and 15, and the number of polar radius segments q is set to 3, 4, 5, and 6. The intra-class and inter-class scatter of the feature vectors obtained under 16 combinations are calculated as the basis for the values ​​of p and q. The 16 cases are then classified and numbered as shown in the table below:

[0093] Table 1 Sample numbers for p and q parameters

[0094]

[0095]

[0096] Since the present invention classifies discharge defect types into three categories, the inter-class dispersion is taken as the mean of pairwise dispersions, and the intra-class dispersions are sorted by size, such as... Figure 5 As shown, the inter-class dispersion is as follows Figure 6 As shown, observations reveal that the intra-class dispersion trends for the three discharge defect types are roughly the same, increasing gradually with the increase of the number of polar angle segments p. When the number of polar diameter segments q is 5, the distribution regularity of feature points in the polar diameter direction is stronger, and the inter-class dispersion is smaller. The inter-class dispersion is greater when the number of polar angle segments p is larger, reaching a peak at sample 10. At this point, the inter-class dispersion of each sample is the largest, and the distinguishing ability is the highest. The intra-class dispersion curve fluctuates gently and has little impact. Therefore, the values ​​of p and q are 18 and 4, respectively.

[0097] S5: Input the optimal feature vector into the Adaboost-LSTM network model for partial discharge pattern recognition.

[0098] The experimental data consisted of 390 samples, with 130 samples for each of the three partial discharge defect types. PRPD (partial discharge defect) maps were fused from these samples to obtain 390 sets of optoelectronic fused PRPD maps. Based on the KAZE feature extraction algorithm proposed above, which optimizes polar coordinate distribution entropy, features were extracted from the UHF PRPD map, optical PRPD map, and optoelectronic fused PRPD map, respectively, and feature vectors were constructed. The necessity of map fusion was then verified. The Adaboost cascade classifier, as a representative of ensemble learning techniques, can improve the classifier's recognition ability through the cascading of sub-classifiers and online assignment of weights. The optimal feature vectors were then substituted into the Adaboost-LSTM for pattern recognition. The experimental data sample statistics are shown in Table 2.

[0099] Table 2. Statistical Table of Experimental Samples

[0100]

[0101] To verify the effectiveness of the algorithm under different sample sizes, this embodiment conducted three experiments. The experiments tested the pattern recognition accuracy of photoelectric fusion maps, UHF maps, and optical maps under different training set sample ratios. The classifier selected was Adaboost-LSTM. In the experiments, the total number of samples was 390 groups. The training set sample ratios were changed to 70%, 50%, and 30% in the three experiments, respectively. The recognition accuracy of optical maps, UHF maps, and photoelectric fusion maps was compared. The experimental results are shown in Tables 3, 4, and 5.

[0102] Table 3. Pattern Recognition Results of Experiment 1

[0103]

[0104] Table 4. Pattern Recognition Results of Experiment 2

[0105]

[0106] Table 5. Pattern Recognition Results of Experiment 3

[0107]

[0108] It can be observed that, under different training set sample proportions, the optoelectronic fusion spectrum can achieve an accuracy of about 80% in identifying different defects, with a maximum of 89.7%, which is significantly higher than the accuracy of optical spectrum and UHF spectrum, thus verifying the effectiveness of spectrum fusion.

[0109] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0110] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A GIS partial emission pattern recognition method based on polar coordinate distribution entropy optimization, characterized in that, Includes the following steps: S1: The partial discharge signal of GIS is simulated by experimental simulation, collected by optical integrated sensor, and converted into photoelectric PRPD spectrum; the photoelectric PRPD spectrum includes optical PRPD spectrum and ultra-high frequency PRPD spectrum; S2: The photoelectric PRPD map is fused using the NSCT image fusion algorithm to obtain the photoelectric fused PRPD map; S3: Use the KAZE algorithm to extract feature points from the optoelectronic fusion PRPD map, and obtain the coordinate information and description vector of the feature points. The description vector contains the scale information of the feature points. S4: Distribute the feature points containing coordinate and scale information to polar coordinates, calculate the distribution entropy for each sub-region divided on the polar coordinates, construct a feature vector, and select the optimal feature vector, including: S41: Using the origin of the photoelectric fusion PRPD map as the pole, construct a polar coordinate system, initialize the number of polar angle segments p and the number of polar radius segments q, and divide the polar angle into p segments and the polar radius into q segments, with the pole as the center, and divide the polar coordinate system into several sub-regions. S42: The extracted feature points are dispersed into polar coordinates according to coordinate and scale information, and the distribution entropy of each sub-region is calculated to form a feature vector; The distribution entropy The calculation formula is as follows: ； in, P j For feature points to fall within the region j The probability within is calculated using the following formula: ； n j For falling in the area j The number of feature points, M, is the number of sub-regions, satisfying P1+P2+…+P M =1; S43: Calculate the intra-class scatter and inter-class scatter of the feature vector and determine whether it is optimal; if yes, obtain the optimal feature vector; if no, update the values ​​of the number of polar angle segments p and the number of polar radius segments q, and return to S41; The intra-class and inter-class scatter of the feature vectors are calculated as follows: ； ； in, For the intra-class dispersion of the same type of discharge defect, H represents the inter-class dispersion of different types of discharge defects. kj Let E(H) be the j-th dimension of the feature vector of the k-th sample. j Let E(H) be the expectation of the j-th feature. aj ), E(H bj ) are the expected values ​​of the j-th dimension features of the feature vectors of discharge defects of types a and b, respectively, and K is the total number of samples; S5: Input the optimal feature vector into the Adaboost-LSTM network model for partial discharge pattern recognition.

2. The GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization according to claim 1, characterized in that, The optical integrated sensor integrates an optical sensor and an ultra-high frequency sensor, which respectively convert the GIS partial discharge signal into the optical PRPD spectrum and the ultra-high frequency PRPD spectrum.

3. The GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization according to claim 2, characterized in that, S2 includes: S21: Perform multi-scale decomposition on the optical PRPD spectrum and the ultra-high frequency PRPD spectrum respectively to obtain high-frequency sub-band maps and low-frequency sub-band maps at each level; perform multi-directional decomposition on the high-frequency sub-band map to finally obtain the low-frequency sub-band coefficients and high-frequency sub-band coefficients in multiple different directions of the optical PRPD spectrum and the ultra-high frequency PRPD spectrum respectively. S22: Based on the fusion rule of averaging the low-frequency subband coefficients and taking the larger high-frequency subband coefficients, the optical PRPD map and the ultra-high frequency PRPD map are fused at each scale and direction to obtain the optoelectronic fusion sub-map. S23: Perform NSCT inverse transform on the photoelectric fusion subgraph to obtain the photoelectric fusion PRPD map.

4. The GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization according to claim 1, characterized in that, S3 includes: S31: Based on the original image, a nonlinear scale space is constructed using an additive operator splitting algorithm and a variable propagation-diffusion method; S32: At different scales in the nonlinear scale space, any point is compared with each point in its neighborhood to find the maximum point of the normalized Hessian matrix as the feature point; the position coordinates of the feature point are solved by Taylor expansion fitting. S33: Determine the main direction of the feature point based on its neighborhood, and construct a description vector for each feature point using the M-SURF algorithm.

5. The GIS partial discharge pattern recognition method based on polar coordinate distribution entropy optimization according to claim 4, characterized in that, In step S33, the M-SURF algorithm is used to construct a description vector for each feature point, including: S331: For scale s i Feature points, with them as the center, select a size of 24s i ×24s i The matrix neighborhood is divided into 4×4 sub-regions, with an overlap of 2s between each pair of sub-regions. i ; S332: Perform Gaussian weighting on each sub-region to obtain a 4-dimensional feature point description vector d. v as follows: ； Among them, L x and L y Let x and y be the differentials at x and y, respectively; S333: Using a Gaussian kernel for the 4D description vector d of each sub-region v Weighted summation and normalization are performed to obtain a 64-dimensional description vector for each feature point.

Images