GIL partial discharge mode identification method and system based on visual word bag model

Abstract

The invention relates to the technical field of image recognition, and discloses a GIL partial discharge mode recognition method and system based on a visual word bag model, and the method comprises the steps: carrying out the statistics of typical partial discharge defects of a GIL, and obtaining GIL partial discharge original pulse data under different typical defect conditions; reading the discharge capacity, discharge phase and discharge polarity information of the obtained pulse data from a database; feature extraction is carried out on the basis of a PRPD atlas of phase-resolved partial discharge, and unsupervised clustering is carried out on high-dimensional features of the PRPD through a K-means method; and constructing an XgBoost algorithm model, taking the obtained visual word bag features of the PRPD as model input of an XgBoost algorithm, outputting a recognition result, and performing evaluation according to the feature evaluation index set. According to the method, the self-adaptive extraction of the discharge features is realized, the computing resources are reduced, and the self-adaptive extraction of the features and the reduction of the risk of redundant features are realized. And the GIL defect with relatively high discharge similarity can be effectively distinguished and identified.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a GIL partial discharge pattern recognition method and system based on a visual word bag model. Background Art

[0002] Partial discharge (PD) refers to the incomplete discharge phenomenon that occurs inside or on the surface of an insulator. It is a common fault phenomenon in the operation of a gas insulated transmission line (GIL). The presence of partial discharge can lead to a decrease in insulation performance. In severe cases, it may cause insulation damage and affect the safe operation of the GIL. Therefore, the identification and classification of partial discharge patterns are of great significance.

[0003] Existing GIL partial discharge pattern recognition technologies are mostly based on the statistical feature extraction of phase resolved partial discharge (PRPD) spectra, including parameters such as kurtosis, skewness, discharge factor, mutual correlation coefficient and discharge asymmetry. These statistical features can characterize the characteristics of partial discharge signals, and these features can be identified and classified through machine learning algorithms to achieve the recognition of partial discharge patterns.

[0004] Existing GIL partial discharge recognition models rely on manually selected features, which not only increases the computational cost, but also may generate redundant features, affecting the recognition accuracy of the model. Manual feature selection requires prior knowledge and experience, and is difficult to automate. The manually constructed feature set may contain redundant features, resulting in increased model complexity and reduced accuracy of the recognition results. These redundant features not only increase the computational cost, but may also have a negative impact on the recognition results. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that the feature extraction of the existing GIL partial discharge recognition model relies on manual feature selection, which not only increases the computational cost, but also may generate redundant features, resulting in a decrease in the recognition accuracy of the model. The model complexity and accuracy of the traditional machine learning algorithm are low, and the recognition performance is average when faced with the recognition problem of highly similar discharge defect types.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a GIL partial discharge pattern recognition method based on a visual word bag model, comprising:

[0008] Count typical partial discharge defects of GIL and obtain the original pulse data of GIL partial discharge under different typical defect conditions;

[0009] Read the discharge amount, discharge phase and discharge polarity information of the acquired pulse data from the database;

[0010] Feature extraction is performed based on the PRPD spectrum of phase-resolved partial discharge, and the high-dimensional features of PRPD are clustered unsupervisedly using the K-means method;

[0011] The XgBoost algorithm model is constructed, and the obtained visual word bag features of PRPD are used as the model input of the XgBoost algorithm, and the recognition results are evaluated based on the evaluation parameters.

[0012] As a preferred solution of the GIL partial discharge pattern recognition method based on the visual word bag model described in the present invention, the partial discharge defects include: carrying out a long-term withstand voltage test, obtaining the GIL partial discharge raw pulse data under different typical defect conditions, and storing the discharge pulse data through a MySQL database; the four typical defects include metal contamination on the insulation surface, air gap inside the insulation, floating electrodes, and metal spikes;

[0013] The discharge amount, discharge phase and discharge polarity information of the acquired pulse data are read from the database, the discharge amount data within a power frequency cycle is normalized according to the maximum value, the discharge data of different defects are superimposed, and each original pulse discharge data is filled in the spectrum grid according to the discharge amount and phase information. The number of discharges in the grid is represented by different colors, and the original data is plotted as a phase-resolved partial discharge PRPD spectrum.

[0014] As a preferred solution of the GIL partial discharge pattern recognition method based on the visual word bag model described in the present invention, the feature extraction includes characterizing the local features of the obtained PRPD spectrum through the SIFT feature extraction algorithm, constructing the scale space of each PRPD through the Gaussian difference pyramid theory, and each layer of the difference pyramid is obtained based on the Gaussian blur and downsampling of the previous layer; the PRPD scale space is represented as a series of Gaussian blurred images L (x, y, σ), and the formula is expressed as:

[0015] L(x,y,σ)=G(x,y,σ)·I(x,y)

[0016] Among them, I(x,y) represents the original image, G(x,y,σ) represents the scale-variable Gaussian function, and the calculation formula is expressed as:

[0017]

[0018] Based on the obtained Gaussian difference pyramids of different PRPD scale spaces, the grayscale images in different scale spaces are compared, the extreme point threshold is set, and based on the set threshold, the grayscale value of each pixel point is compared with that in the same scale and adjacent scales to obtain candidate key extreme points;

[0019] The position and scale of the key points are determined by fitting a three-dimensional quadratic function, and unstable edge response points and key points with low contrast are removed from the candidate key extreme points. According to the key SIFT feature extreme points determined after screening, the gradient direction amplitude of the neighboring pixels adjacent to the key extreme points is given to the key SIFT feature extreme points. The calculation formula of the modulus and angle of the direction parameter is expressed as:

[0020]

[0021] In the 4×4 window of the feature key point, the gradient information in 8 directions of each window is calculated based on the above formula to obtain a 128-dimensional SIFT feature vector; a SIFT feature descriptor is generated to obtain a high-dimensional feature matrix that represents the spectrum.

[0022] As a preferred solution of the GIL partial discharge pattern recognition method based on the visual word bag model described in the present invention, wherein: the unsupervised clustering by the K-means method includes performing unsupervised clustering of the high-dimensional features of all PRPDs by the K-means method, setting the number of cluster center clusters,

[0023] According to the preset number of visual words N, the number of K-means cluster centers is set, and N points are randomly selected as the initial cluster centers; all SIFT feature points of all PRPDs are brought into the calculation, and for each data point, it is assigned to the data cluster where the nearest cluster center is located according to its spatial distance from the cluster center; for each data cluster, the centroid of all points in the cluster is calculated and used as the new cluster center; the steps are repeated until the sum of squared errors in the cluster is locally minimized. The calculation formula is expressed as:

[0024]

[0025] Where K represents the number of clusters, n represents the total number of data points, and x j represents the jth data point in the data set, μ i Represents the cluster center of the i-th cluster.

[0026] As a preferred solution of the GIL partial discharge pattern recognition method based on the bag-of-visual-words model described in the present invention, the bag-of-visual-words features of the PRPD obtained include: taking all the acquired cluster centers as words of the bag-of-visual-words model, calculating the Euclidean distance from each SIFT feature point in each PRPD to all the visual words, and arranging the distance calculation results; according to the distance arrangement results, selecting the visual word with the shortest distance to each SIFT feature point, using a hard-coding method, setting the additional weight of the word to 1, and setting the additional weights of the remaining visual words to 0, calculating all the feature descriptors in the image in turn and accumulating the results, and generating the bag-of-visual-words features of the PRPD described by visual words.

[0027] As a preferred solution of the GIL partial discharge pattern recognition method based on the visual word bag model described in the present invention, the model of the XgBoost algorithm includes taking the obtained visual word bag features of each PRPD as the model input of the XgBoost algorithm, setting the discharge type as the objective function of XgBoost, and the objective function formula is expressed as:

[0028]

[0029] in, represents the target loss function, y i Indicates the actual value of the discharge type, represents the prediction value of the discharge type by the first t decision trees, Ω(f t ) represents the model complexity of the first t decision trees, which is used to reduce the overfitting of the model;

[0030] Traverse all the XgBoost model feature partitioning points, iterate each residual tree according to the feature partitioning, build a residual tree based on the optimal prediction result of the previous tree each time, and gradually approach the objective function. The iterative formula of XgBoost is expressed as:

[0031]

[0032] in, represents the result of the model prediction in the tth round, represents the result of the model prediction in the t-1th round, f t (x i ) represents the new function added in round t.

[0033] As a preferred solution of the GIL partial discharge pattern recognition method based on the visual word bag model described in the present invention, wherein: the method includes calculating the Accuracy, Precision, Recall, and F1 evaluation parameters of the recognition result based on the XgBoost model, and outputting the calculation results of the feature evaluation index set;

[0034] The ratio of the number of predicted samples to the total number of samples is Accuracy, which is expressed as:

[0035]

[0036] Among them, TP represents the number of samples that the algorithm makes a positive judgment and the judgment is correct, TN represents the number of samples that the algorithm makes a negative judgment and the judgment is correct, FN represents the number of samples that the algorithm makes a negative judgment but fails to judge, and FP represents the number of samples that the algorithm makes a positive judgment but fails to judge;

[0037] The number of samples that are predicted to be positive is Precision, which is actually correct. The formula is:

[0038]

[0039] How many positive samples are successfully retrieved? The formula is:

[0040]

[0041] The formula for the harmonic mean F1 of precision and recall is:

[0042]

[0043] Among them, P represents Precision and R represents Recall value.

[0044] A GIL partial discharge pattern recognition system based on a bag-of-visual-words model using any method described in the present invention, wherein:

[0045] Data acquisition and preprocessing: collect partial discharge data of typical defects of GIL and perform preprocessing;

[0046] Feature extraction and visual word bag model: extract image features based on PRPD graph, build visual word bag model, perform K-means clustering analysis and generate visual word bag features;

[0047] Model training and recognition, using the XgBoost algorithm to train and recognize patterns on the extracted features, conduct model evaluation and optimization, and perform hyperparameter adjustment to optimize model performance;

[0048] Result output and application,output the recognition results and apply them to GIL defect detection and classification.

[0049] A computer device comprises: a memory and a processor; the memory stores a computer program, comprising: the steps of implementing any one of the methods of the present invention when the processor executes the computer program.

[0050] A computer-readable storage medium stores a computer program, comprising: when the computer program is executed by a processor, the steps of implementing any one of the methods of the present invention are implemented.

[0051] The beneficial effects of the present invention are as follows: 1. Adaptive extraction of discharge features. The machine learning algorithm needs to obtain discharge features as the input of the recognition model. However, the input features of the model rely on feature extraction. There are two problems with the artificially constructed discharge feature set. First, the extraction of features requires manual computing resources, and the feature construction work is relatively complicated. Second, the constructed features may contain redundant features, which will have a negative impact on the recognition accuracy. The PRPD graph greatly retains the characteristic information of the discharge. The visual word bag model is based on PRPD and can directly realize the high-dimensional feature extraction of the discharge, which not only reduces the computing resources, but also realizes the adaptive extraction of features and the reduction of the risk of redundant features. 2. XgBoost based on the gradient boosting decision tree can effectively improve the robustness of the model and reduce the risk of overfitting. For GIL defects with high discharge similarity, it can effectively distinguish and identify them. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0053] Figure 1 An overall flow chart of a GIL partial discharge pattern recognition method based on a visual word bag model provided by the first embodiment of the present invention;

[0054] Figure 2 A comparison diagram of typical partial discharge defects of a GIL partial discharge pattern recognition method based on a bag-of-visual-words model provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0056] Example 1, reference Figure 1 , as an embodiment of the present invention, provides a GIL partial discharge pattern recognition method based on a visual word bag model, comprising:

[0057] S1: Count typical partial discharge defects of GIL and obtain the original pulse data of GIL partial discharge under different typical defect conditions.

[0058] Furthermore, the partial discharge defects include conducting long-term voltage withstand tests, obtaining the original pulse data of GIL partial discharge under different typical defect conditions, and storing the discharge pulse data through the MySQL database; the four typical defects include metal contamination on the insulation surface, air gaps inside the insulation, suspended electrodes, and metal spikes.

[0059] Furthermore, the discharge amount, discharge phase and discharge polarity information of the acquired pulse data are read from the database, the discharge amount data within a power frequency cycle is normalized according to the maximum value, the discharge data of different defects are superimposed, and each original pulse discharge data is filled in the spectrum grid according to the discharge amount and phase information. The number of discharges in the grid is represented by different colors, and the original data is plotted as a phase-resolved partial discharge PRPD spectrum.

[0060] It should be noted that the feature extraction of the existing GIL partial discharge recognition model relies on manual feature selection, which not only increases the computational cost, but also may generate redundant features, resulting in a decrease in the recognition accuracy of the model. The model complexity and accuracy of traditional machine learning algorithms (such as support vector machines, KNN, etc.) are relatively low, and the recognition performance is average when faced with the recognition problem of highly similar discharge defect types. To solve the above problems, this paper proposes a GIL typical defect partial discharge pattern recognition method based on a visual word bag model and XgBoost. This feature extraction method is based on the PRPD spectrum of phase-resolved partial discharge, and realizes the adaptive extraction of partial discharge features. At the same time, the advanced machine learning algorithm XgBoost is used to achieve further improvement on the model side.

[0061] S2: Read the discharge amount, discharge phase and discharge polarity information of the acquired pulse data from the database.

[0062] Furthermore, partial discharge is an important phenomenon in GIL insulation characterization. Existing GIL partial discharge pattern recognition technology extracts the characteristics of phase resolved partial discharge (PRPD) statistical feature set (including kurtosis, skewness, discharge factor, mutual correlation coefficient and discharge asymmetry parameter) of discharge signals. The calculation formula of each feature is as follows: Skewness (S_k), used to describe the skewness of the partial discharge spectrum:

[0063]

[0064] Wherein, n represents the number of windows in the power frequency cycle of the PRPD spectrum. In this patent, a 360° signal of a power frequency cycle is divided into 360 windows with equal phases of 1°. represents the phase value of the i-th phase window in the PRPD spectrum, represents the phase window width, μ is the mean value of the i-th phase window in the PRPD spectrum, and p i is the probability density of the i-th phase window in the PRPD spectrum, and σ is the variance of the i-th phase window in the PRPD spectrum.

[0065] Furthermore, the kurtosis (K_u) is used to describe the protrusion of the partial discharge spectrum:

[0066]

[0067] Furthermore, the discharge factor Q p , used to describe the difference in the discharge amount of partial discharge in the positive and negative half periods of the power frequency:

[0068]

[0069] in, Represents the total discharge amount in the positive cycle of the PRPD spectrum. Represents the total discharge amount in the negative cycle of the PRPD spectrum. Represents the total number of discharges in the positive cycle of the PRPD spectrum. The sum of the discharge times in the negative cycle of the PRPD spectrum. The cross-correlation coefficient (CC) is used to describe the similarity of the discharge profiles of the positive and negative half-cycles of the PRPD.

[0070]

[0071] Furthermore, the discharge asymmetry (p) is used to describe the difference between the starting phases of the PRPD spectrum in the positive and negative half cycles:

[0072]

[0073] in, is the starting phase of the negative half cycle of discharge, The starting phase of the positive half cycle of discharge.

[0074] It should be noted that the GL partial discharge statistical feature parameter set is constructed according to the above formula, and then the partial discharge is identified and classified based on the statistical feature set and the discharge training set label through the support vector machine (SVM) machine learning algorithm. This technology relies on the effectiveness of the proposed features. Redundant features will affect the recognition results. Feature extraction requires manual prior and knowledge judgment. At the same time, the algorithm complexity of the traditional machine learning algorithm is relatively low, and the recognition accuracy of some similar defects is not good. Therefore, there are deficiencies in the existing technology.

[0075] S3: Feature extraction is performed based on the PRPD spectrum of phase-resolved partial discharge, and the high-dimensional features of PRPD are clustered unsupervisedly using the K-means method.

[0076] Furthermore, the feature extraction includes characterizing the local features of the obtained PRPD spectrum through the SIFT feature extraction algorithm, constructing the scale space of each PRPD through the Gaussian difference pyramid theory, and each layer of the difference pyramid is obtained based on the Gaussian blur and downsampling of the previous layer; the PRPD scale space is represented as a series of Gaussian blurred images L (x, y, σ), and the formula is expressed as:

[0077] L(x,y,σ)=G(x,y,σ)·I(x,y)

[0078] Among them, I(x,y) represents the original image, G(x,y,σ) represents the scale-variable Gaussian function, and the calculation formula is expressed as:

[0079]

[0080] Furthermore, based on the obtained Gaussian difference pyramids of different PRPD scale spaces, the grayscale images in different scale spaces are compared, the extreme point threshold is set, and based on the set threshold, the grayscale value of each pixel point is compared with that in the same scale and adjacent scales to obtain candidate key extreme points.

[0081] Furthermore, the position and scale of the key points are determined by fitting a three-dimensional quadratic function, and the unstable edge response points and key points with low contrast are removed from the candidate key extreme points; according to the key SIFT feature extreme points determined after screening, the gradient direction amplitude of the neighboring pixels adjacent to the key extreme points is given to the key SIFT feature extreme points, and the calculation formula of the modulus and angle of the direction parameter is expressed as:

[0082]

[0083] Furthermore, the gradient information of 8 directions of each window is calculated based on the above formula in the 4×4 window of the feature key point to obtain a 128-dimensional SIFT feature vector; a SIFT feature descriptor is generated to obtain a high-dimensional feature matrix representing the spectrum.

[0084] Furthermore, the unsupervised clustering by the K-means method includes performing unsupervised clustering of all high-dimensional features of PRPD by the K-means method and setting the number of cluster center clusters.

[0085] Furthermore, according to the preset number of visual words N, the number of K-means cluster centers is set, and N points are randomly selected as the initial cluster centers; all SIFT feature points of all PRPDs are brought into the calculation, and for each data point, it is assigned to the data cluster where the nearest cluster center is located according to its spatial distance from the cluster center; for each data cluster, the centroid of all points in the cluster is calculated and used as the new cluster center; the steps are repeated until the sum of squared errors in the cluster is locally minimized, and the calculation formula is expressed as:

[0086]

[0087] Where K represents the number of clusters, n represents the total number of data points, and x j represents the jth data point in the data set, μ i Represents the cluster center of the i-th cluster.

[0088] It should be noted that the obtained bag-of-visual-words features of the PRPD include taking all the obtained cluster centers as words of the bag-of-visual-words model, calculating the Euclidean distance from each SIFT feature point in each PRPD to all visual words, and arranging the distance calculation results; according to the distance arrangement results, selecting the visual word with the shortest distance to each SIFT feature point, using a hard-coding method, setting the additional weight of the word to 1, and setting the additional weights of the remaining visual words to 0, calculating all feature descriptors in the image in turn and accumulating the results to generate the bag-of-visual-words features of the PRPD described by visual words.

[0089] S4: Construct the XgBoost algorithm model, use the obtained visual word bag features of PRPD as the model input of the XgBoost algorithm, and evaluate the recognition results based on the evaluation parameters.

[0090] The model of the XgBoost algorithm includes taking the obtained visual word bag features of each PRPD as the model input of the XgBoost algorithm, setting the discharge type as the objective function of XgBoost, and the objective function formula is expressed as:

[0091]

[0092] in, represents the target loss function, y i Indicates the actual value of the discharge type, represents the prediction value of the discharge type by the first t decision trees, Ω(f t ) represents the model complexity of the first t decision trees, which is used to reduce the overfitting of the model;

[0093] Traverse all the XgBoost model feature partitioning points, iterate each residual tree according to the feature partitioning, build a residual tree based on the optimal prediction result of the previous tree each time, and gradually approach the objective function. The iterative formula of XgBoost is expressed as:

[0094]

[0095] in, represents the result of the model prediction in the tth round, represents the result of the model prediction in the t-1th round, f t (x i ) represents the new function added in round t.

[0096] The method includes calculating the Accuracy, Precision, Recall, and F1 evaluation parameters of the recognition results of the XgBoost model, and outputting the calculation results of the feature evaluation index set;

[0097] The ratio of the number of predicted samples to the total number of samples is Accuracy, which is expressed as:

[0098]

[0099] Among them, TP represents the number of samples that the algorithm makes a positive judgment and the judgment is correct, TN represents the number of samples that the algorithm makes a negative judgment and the judgment is correct, FN represents the number of samples that the algorithm makes a negative judgment but fails to judge, and FP represents the number of samples that the algorithm makes a positive judgment but fails to judge;

[0100] The number of samples that are predicted to be positive is Precision, which is actually correct. The formula is:

[0101]

[0102] How many positive samples are successfully retrieved? The formula is:

[0103]

[0104] The formula for the harmonic mean F1 of precision and recall is:

[0105]

[0106] Among them, P represents Precision and R represents Recall value.

[0107] It should be noted that the XgBoost algorithm's GIL typical defect partial discharge pattern recognition method aims to detect and classify typical defects in GIL through image feature extraction technology and advanced machine learning algorithms, thereby improving the operational safety and reliability of GIL.

[0108] On the other hand, this embodiment also provides a GIL partial discharge pattern recognition system based on a visual word bag model, which includes:

[0109] Data acquisition and preprocessing, collect partial discharge data of typical defects of GIL and perform preprocessing.

[0110] Feature extraction and visual word bag model, extract image features based on PRPD graph, build a visual word bag model, perform K-means clustering analysis and visual word bag feature generation.

[0111] Model training and recognition, using XgBoost algorithm to train and recognize patterns on the extracted features, conduct model evaluation and optimization, and perform hyperparameter adjustment to optimize model performance.

[0112] Result output and application,output the recognition results and apply them to GIL defect detection and classification.

[0113] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0114] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0115] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0116] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0117] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides a GIL partial discharge pattern recognition method based on a visual word bag model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0118] Based on the typical partial discharge defects commonly seen in engineering GILs, a variety of defective short samples containing discharge defects were designed. The original pulse data of GIL partial discharge under different typical defect conditions were obtained based on the test platform. The discharge data of every 100 groups of power frequency AC cycles (20ms) of different defects were superimposed and drawn as PRPD maps. The visual word bag model was used to generate PRPD visual word bag features described by visual words. The obtained PRPD visual word bag features were input into the XgBoost classification model for training, the pattern recognition results of the model were obtained, the model results were saved, and the evaluation index set containing the evaluation parameters of Accuracy, Precision, Recall, and F1 after the model training was output.

[0119] Typical partial discharge defects commonly seen in statistical engineering GILs, such as Figure 2 As shown in the figure, based on the IEC 60270 partial discharge test standard, an AC high-voltage GIL partial discharge test platform is built, and according to the common GIL partial discharge defects in engineering statistics, a variety of defect short samples with corresponding defects are designed, and long-term withstand voltage tests are carried out to obtain the GIL partial discharge raw pulse data under different typical defect conditions, and the discharge pulse data is stored in the MySQL database. The production process of the four typical defects is as follows:

[0120] (1) Control group, normal sample: The sample is made of a cylindrical epoxy resin with a bottom diameter of 80 mm and a height of 60 mm sandwiched between two circular plate electrodes with a diameter of 120 mm and a thickness of 10 mm. It has good insulation performance and is used to simulate the guide rod and metal shell of the GIL under normal operating conditions.

[0121] (2) Metal contamination on the insulation surface: Based on the normal sample, copper chips with a length of 50 mm and a width of 10 mm were glued to the surface of the epoxy resin to simulate the adhesion of metal debris on the surface of the insulator, with a distance of 5 mm between the upper and lower plates.

[0122] (3) Insulation internal air gap: This defective sample was made by stacking three thin epoxy resin sheets of 5 mm, 2 mm, and 5 mm between two circular electrode plates, and an air gap of 5 mm in diameter was left in the middle of the second epoxy resin layer.

[0123] (4) Suspended electrode: The defective sample is made of an epoxy resin ring with a diameter of 50 mm and a thickness of 5 mm, with two pointed electrodes. The lower electrode is connected to a cylindrical epoxy resin with a diameter of 80 mm and a height of 30 mm and an electrode plate. The gap between the electrodes is 2 mm.

[0124] (5) Metal spikes: This defective sample is also composed of two plates, and a steel needle with a length of 10 mm and a diameter of 0.8 mm is embedded in the middle of the upper plate. The data obtained are shown in Table 1.

[0125] Table 1 Experimental results comparison table

[0126]

[0127] The discharge amount and discharge phase analysis show that the discharge amount and discharge phase of normal samples are relatively low, while the discharge amount and discharge phase of defective samples are higher, especially the discharge amount of metal spike samples, which is the highest, reaching 90pC. The discharge phases of different defect types are also significantly different. For example, the discharge phase of metal contamination on the insulating surface is 90°, while the discharge phase of metal spikes is 225°.

[0128] The PRPD feature shows the difference in feature vectors of different defect samples. The PRPD feature of normal samples is relatively low, while the PRPD feature of defect samples is higher, especially the PRPD feature value of metal spike samples is the highest, indicating that its discharge feature is the most significant. Through K-means clustering analysis, the PRPD features of different defect types can be effectively classified. The clustering results show that the PRPD features of different defect types have obvious clustering centers.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A GIL partial discharge pattern recognition method based on a bag-of-visual-words model, characterized in that: include: Count typical partial discharge defects of GIL and obtain the original pulse data of GIL partial discharge under different typical defect conditions; Read the discharge amount, discharge phase and discharge polarity information of the acquired pulse data from the database; Feature extraction is performed based on the PRPD spectrum of phase-resolved partial discharge, and the high-dimensional features of PRPD are clustered unsupervisedly using the K-means method; The XgBoost algorithm model is constructed, the obtained visual bag-of-words features of PRPD are used as the model input of the XgBoost algorithm, the recognition results are output, and they are evaluated according to the feature evaluation index set.

2. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model according to claim 1, characterized in that: The partial discharge defect includes conducting a long-term withstand voltage test, obtaining the original pulse data of the GIL partial discharge under different typical defect conditions, and storing the discharge pulse data in a MySQL database; Four typical defects include metal contamination on the insulation surface, air gaps inside the insulation, floating electrodes, and metal spikes; The discharge amount, discharge phase and discharge polarity information of the acquired pulse data are read from the database, the discharge amount data within a power frequency cycle is normalized according to the maximum value, the discharge data of different defects are superimposed, and each original pulse discharge data is filled in the spectrum grid according to the discharge amount and phase information. The number of discharges in the grid is represented by different colors, and the original data is plotted as a phase-resolved partial discharge PRPD spectrum.

3. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model according to claim 2, characterized in that: The feature extraction includes characterizing the local features of the obtained PRPD spectrum through the SIFT feature extraction algorithm, constructing the scale space of each PRPD through the Gaussian difference pyramid theory, and each layer of the difference pyramid is obtained based on the Gaussian blur and downsampling of the previous layer; the PRPD scale space is represented as a series of Gaussian blurred images L (x, y, σ), and the formula is expressed as: L(x,y,σ)=G(x,y,σ)·I(x,y) Among them, I(x,y) represents the original image, G(x,y,σ) represents the scale-variable Gaussian function, and the calculation formula is expressed as: Based on the obtained Gaussian difference pyramids of different PRPD scale spaces, the grayscale images in different scale spaces are compared, the extreme point threshold is set, and based on the set threshold, the grayscale value of each pixel point is compared with that in the same scale and adjacent scales to obtain candidate key extreme points; The position and scale of the key points are determined by fitting a three-dimensional quadratic function, and unstable edge response points and key points with low contrast are removed from the candidate key extreme points. According to the key SIFT feature extreme points determined after screening, the gradient direction amplitude of the neighboring pixels adjacent to the key extreme points is given to the key SIFT feature extreme points. The calculation formula of the modulus and angle of the direction parameter is expressed as: In the 4×4 window of the feature key point, the gradient information in 8 directions of each window is calculated based on the above formula to obtain a 128-dimensional SIFT feature vector; a SIFT feature descriptor is generated to obtain a high-dimensional feature matrix that represents the spectrum.

4. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model as claimed in claim 3, characterized in that: The unsupervised clustering by the K-means method includes performing unsupervised clustering of all high-dimensional features of PRPD by the K-means method, setting the number of cluster centers, According to the preset number of visual words N, the number of K-means cluster centers is set, and N points are randomly selected as the initial cluster centers; all SIFT feature points of all PRPDs are brought into the calculation, and for each data point, it is assigned to the data cluster where the nearest cluster center is located according to its spatial distance from the cluster center; for each data cluster, the centroid of all points in the cluster is calculated and used as the new cluster center; the steps are repeated until the sum of squared errors in the cluster is locally minimized. The calculation formula is expressed as: Where K represents the number of clusters, n represents the total number of data points, and x j represents the jth data point in the data set, μ i Represents the cluster center of the i-th cluster.

5. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model according to claim 4, characterized in that: The obtained bag-of-visual-words features of the PRPD include: taking all the obtained cluster centers as words of the bag-of-visual-words model, calculating the Euclidean distance from each SIFT feature point in each PRPD to all the visual words, and arranging the distance calculation results; according to the distance arrangement results, selecting the visual word with the shortest distance to each SIFT feature point, using a hard-coding method, setting the additional weight of the word to 1, and setting the additional weights of the remaining visual words to 0, calculating all the feature descriptors in the image in turn and accumulating the results, and generating the bag-of-visual-words features of the PRPD described by visual words.

6. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model according to claim 5, characterized in that: The model of the XgBoost algorithm includes taking the obtained visual word bag features of each PRPD as the model input of the XgBoost algorithm, setting the discharge type as the objective function of XgBoost, and the objective function formula is expressed as: in, represents the target loss function, y i Indicates the actual value of the discharge type, represents the prediction value of the discharge type by the first t decision trees, Ω(f t ) represents the model complexity of the first t decision trees, which is used to reduce the overfitting of the model; Traverse all the XgBoost model feature partitioning points, iterate each residual tree according to the feature partitioning, build a residual tree based on the optimal prediction result of the previous tree each time, and gradually approach the objective function. The iterative formula of XgBoost is expressed as: in, represents the result of the model prediction in the tth round, represents the result of the model prediction in the t-1th round, f t (x i ) represents the new function added in round t.

7. The GIL partial discharge pattern recognition method based on the bag-of-visual-words model according to claim 6, characterized in that: The evaluation according to the feature evaluation index set includes calculating the recognition result Accuracy, Precision, Recall, and F1 evaluation parameters based on the XgBoost model, and outputting the calculation result of the feature evaluation index set; The ratio of the number of predicted samples to the total number of samples is Accuracy, which is expressed as: Among them, TP represents the number of samples that the algorithm makes a positive judgment and the judgment is correct, TN represents the number of samples that the algorithm makes a negative judgment and the judgment is correct, FN represents the number of samples that the algorithm makes a negative judgment but fails to judge, and FP represents the number of samples that the algorithm makes a positive judgment but fails to judge; The number of samples that are predicted to be positive is Precision, which is actually correct. The formula is: How many positive samples are successfully retrieved? The formula is: The formula for the harmonic mean F1 of precision and recall is: Among them, P represents Precision and R represents Recall value.

8. A GIL partial discharge pattern recognition system based on a bag-of-visual-words model using the method according to any one of claims 1 to 7, characterized in that: Data acquisition and preprocessing: collect partial discharge data of typical defects of GIL and perform preprocessing; Feature extraction and visual word bag model: extract image features based on PRPD graph, build visual word bag model, perform K-means clustering analysis and generate visual word bag features; Model training and recognition, using the XgBoost algorithm to train and recognize patterns on the extracted features, conduct model evaluation and optimization, and perform hyperparameter adjustment to optimize model performance; Result output and application,output the recognition results and apply them to GIL defect detection and classification.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of any one of the GIL partial discharge pattern recognition methods based on the visual word bag model according to claims 1-7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of any one of the GIL partial discharge pattern recognition methods based on the visual word bag model according to claims 1-7 are implemented.

Images