GIS insulation defect identification method and system, terminal and medium

Abstract

The invention discloses a GIS insulation defect identification method and system, a terminal and a medium, and the method comprises the following steps: employing an adaptive pulse sequence threshold determination method based on a quartile distance, and carrying out the adaptive extraction of a plurality of GIS internal partial discharge pulse signals including HFCT, UHF and optical detection equipment; performing dimensionless energy binarization feature extraction on the HFCT, UHF and AE partial discharge signals within the fixed acquisition time to generate feature vectors; establishing a multi-scale insulation defect partial discharge real-time online incremental database, and automatically adjusting parameters of an incremental support vector machine model by adopting a Bayesian optimization algorithm; and inputting the feature vector into an SVM (Support Vector Machine) model subjected to Bayesian optimization of hyper-parameters, and carrying out multi-parameter partial discharge identification.

Description

Technical Field

[0001] The present invention relates to a field, and more specifically, to a GIS insulation defect identification method, system, terminal and medium. Background Art

[0002] Gas-insulated switchgear (GIS), with its core dielectric being high-performance insulating gases like SF6, offers numerous advantages, including high electrical insulation strength, a compact footprint, and low maintenance requirements, making it widely used in modern power systems. However, throughout its lifecycle, including manufacturing, transportation, installation, and long-term operation, GIS equipment can be subject to numerous unforeseen factors, leading to insulation defects such as metallic foreign matter, insulation surface damage, and floating potentials. These defects, when exposed to strong electric fields, can easily induce partial discharge, a key factor in insulation degradation and a potential safety hazard.

[0003] As an expression of insulation defects, the occurrence and evolution of partial discharge is accompanied by complex microscopic physical processes and energy release, including but not limited to charge migration, radiation of light, electromagnetic waves, and acoustic waves, providing a variety of detection methods for partial discharge monitoring. Although current research has made significant progress in this field, using a variety of measurement methods to achieve separation, feature extraction, and fault identification of discharge signals, and successfully applied to on-site condition monitoring, different detection methods still have limitations in sensitivity, applicability, and effectiveness due to differences in principles. The lack of unified application standards leads to a high false alarm rate. In addition, the asynchrony and structural differences between detection data limit the comprehensive reuse capabilities of the data, making it difficult to fully evaluate the discharge status. Furthermore, the different conclusions generated by independent analysis of each detection method lack an effective comparison and integration mechanism, making it difficult to form a highly reliable diagnostic conclusion.

[0004] Given the limitations of a single detection method in partial discharge diagnosis, especially when the microscopic mechanism of the discharge process is not yet fully understood, evaluating the insulation status based solely on the discharge amplitude or frequency is prone to misjudgment due to the complexity of the physical process and the changes in the energy release mechanism.

[0005] In response to the above problems, a GIS insulation defect identification method, system, terminal and medium are urgently needed. Summary of the Invention

[0006] To address the deficiencies in the prior art, the present invention provides a GIS insulation defect identification method and system, proposes multi-parameter adaptive pulse extraction of partial discharge and multi-physical dimensionless energy binary feature extraction, and automatically adjusts parameters through a Bayesian optimization incremental support vector machine model to ensure the effectiveness and stability of the model under different working conditions, different discharge types, and discharge intensities, thereby enhancing the model's adaptability and generalization capabilities.

[0007] The present invention adopts the following technical solutions.

[0008] A first aspect of the present invention relates to a method for identifying GIS insulation defects, comprising the following steps: adaptively extracting a variety of GIS internal partial discharge pulse signals, including HFCT, UHF, and optical detection equipment, using an interquartile range-based adaptive pulse sequence threshold determination method; performing dimensionless energy binary feature extraction on the HFCT, UHF, and AE partial discharge signals within a fixed acquisition time to generate a feature vector; establishing a multi-scale insulation defect partial discharge real-time online incremental database, and automatically adjusting the parameters of an incremental support vector machine model using a Bayesian optimization algorithm; and inputting the feature vector into an SVM model with Bayesian optimized hyperparameters to perform multi-parameter partial discharge identification.

[0009] Adaptive extraction of various GIS internal partial discharge pulse signals including high-frequency coils, ultra-high frequency and optical detection equipment, including: connecting a protective resistor in series between a high-voltage power supply and the high-voltage port of the GIS device under test, where the high-voltage power supply is composed of the output winding of a transformer; connecting an HFCT between the high-voltage port of the GIS device under test and the ground, and also connecting a voltage divider and a coupling capacitor in parallel between the high-voltage port of the GIS device under test and the ground; UHF and AE are set on the outer surface of the GIS, and SiPM is set outside the GIS observation window; the adjustment port of the voltage divider, the HFCT signal acquisition port, the UHF signal acquisition port, the AE signal acquisition port, and the SiPM signal acquisition port are connected to the partial discharge multi-physics detection host computer through a high-speed acquisition card.

[0010] The method performs dimensionless energy binary feature extraction on HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate a feature vector, including: selecting any three ports from the adjustment port of the voltage divider, the HFCT signal acquisition port, the UHF signal acquisition port, the AE signal acquisition port, and the SiPM signal acquisition port; converting the signals of the three ports into dimensionless energy release power of the corresponding sensors, performing binary feature extraction on the signals; and converting the extracted dimensionless energy release power into a two-dimensional space.

[0011] A real-time online incremental database of partial discharges caused by multi-scale insulation defects was established, and the parameters of the incremental support vector machine model were automatically adjusted using a Bayesian optimization algorithm. This involved taking the dimensionless energy release power in two-dimensional space as input and feeding it into multiple SVM models to obtain partial discharge classification results; and optimizing the SVM hyperparameters using Bayesian optimization.

[0012] The SVM hyperparameters are optimized using Bayesian optimization, including: the SVM hyperparameters include kernel function type, C and gamma.

[0013] Bayesian optimization is used to optimize the SVM hyperparameters, including: For the lth SVM model, the optimization objective is:

[0014]

[0015] Where W l and b l are the training parameters of the SVM model,

[0016] C is the regularization parameter,

[0017] ξ i,l variables added to the optimization process,

[0018] y i,l ∈{-1,1}, is the value of the i-th sample at the l-th encoding position, and N is the number of samples.

[0019] The feature vector is input into an SVM model with Bayesian hyperparameter optimization to perform multi-parameter partial discharge identification, including: constructing an incremental database of partial discharge data of multi-scale insulation defects, conducting online collection of partial discharge multi-physical pulse data under insulation defects of different detection scales; using the optimized SVM model to extract the two-dimensional distribution of the posterior probability of different types of faults, and analyzing the distribution differences of different types of faults.

[0020] The second aspect of the present invention relates to a GIS insulation defect identification system; the system is implemented using a GIS insulation defect identification method described in the first aspect of the present invention; the method includes an extraction module, a feature module, an optimization module, and an identification module; the extraction module is used to use an adaptive pulse sequence threshold determination method based on an interquartile range to adaptively extract a variety of GIS internal partial discharge pulse signals including HFCT, UHF and optical detection equipment; the feature module is used to perform dimensionless energy binary feature extraction on HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate a feature vector; the optimization module is used to establish a real-time online incremental database of multi-scale insulation defect partial discharges, and use a Bayesian optimization algorithm to automatically adjust the parameters of the incremental support vector machine model; the identification module is used to input the feature vector into an SVM model with Bayesian optimized hyperparameters to perform multi-parameter partial discharge identification.

[0021] A third aspect of the present invention relates to a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method according to the first aspect of the present invention.

[0022] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in the first aspect of the present invention.

[0023] The beneficial effect of the present invention is that, compared with the existing technology, the GIS insulation defect identification method and system in the present invention proposes multi-parameter adaptive pulse extraction of local discharge and multi-physical dimensionless energy binary feature extraction, and automatically adjusts the parameters of the incremental support vector machine model through Bayesian optimization to ensure the effectiveness and stability of the model under different working conditions, different discharge types and discharge intensities, thereby enhancing the adaptability and generalization ability of the model.

[0024] The beneficial effects of the present invention also include:

[0025] 1. This invention innovatively proposes a GIS insulation defect identification model based on the fusion of multiple physical parameters. By integrating the advantages of multiple detection technologies and comprehensively utilizing multi-dimensional information such as charge, light, electromagnetics, and sound, this model can fully capture and deeply analyze partial discharge phenomena. Through the Bayesian optimization incremental support vector machine (SVM) algorithm, the model parameters are automatically optimized and adapted to new data, ensuring the high accuracy and strong robustness of the model under complex working conditions. At the same time, the introduction of dynamic learning and uncertainty quantification technology further improves the model's response speed to changes in the discharge process and the diagnostic accuracy, providing a more scientific and reliable solution for insulation status assessment and fault warning of GIS equipment.

[0026] 2. This invention combines online learning with incremental learning mechanisms to establish an incremental database of multi-scale insulation defect partial discharge data. Based on this database, the model is updated instantly, further improving the generalization performance of the recognition model. By introducing uncertainty quantification technology and performing soft output transformation on the model's prediction results, the invention effectively obtains the posterior probability of the recognition results and enhances the confidence level of the recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of incremental acquisition of partial discharge data in a GIS insulation defect identification method according to the present invention;

[0028] Figure 2 A schematic diagram of adaptive extraction of multi-physics pulses of partial discharge in a GIS insulation defect identification method according to the present invention;

[0029] Figure 3 A schematic diagram of binary feature extraction for calculating dimensionless energy release power in a GIS insulation defect identification method according to the present invention;

[0030] Figure 4 Schematic diagram of a Bayesian optimization SVM soft output model in a GIS insulation defect identification method of the present invention;

[0031] Figure 5 A schematic diagram of the two-dimensional distribution of the posterior probability of a partial discharge diagnosis model in a GIS insulation defect identification method according to the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention clearer and more accurate, the technical solutions of the present invention are described in detail below through multiple specific embodiments. It is worth noting that the embodiments adopted in the present invention are only used to explain the present invention and are not intended to limit the content of the present invention.

[0033] Figure 1 Schematic diagram of incremental acquisition of partial discharge data in a GIS insulation defect identification method of the present invention. Figure 1 As shown, the first aspect of the present invention relates to a method for identifying GIS insulation defects, the method comprising:

[0034] In step 1, an adaptive pulse sequence threshold determination method based on the interquartile range is used to adaptively extract the internal partial discharge pulse signals of various GIS including HFCT, UHF and optical detection equipment.

[0035] An adaptive pulse sequence threshold determination method based on interquartile range is used to adaptively extract the partial discharge pulse signals inside various GIS including high frequency coil (HFCT), ultra high frequency (UHF) and optical detection equipment.

[0036] Adaptive extraction of various GIS internal partial discharge pulse signals including high-frequency coils, ultra-high frequency and optical detection equipment, including: connecting a protective resistor in series between a high-voltage power supply and the high-voltage port of the GIS device under test, where the high-voltage power supply is composed of the output winding of a transformer; connecting an HFCT between the high-voltage port of the GIS device under test and the ground, and also connecting a voltage divider and a coupling capacitor in parallel between the high-voltage port of the GIS device under test and the ground; UHF and AE are set on the outer surface of the GIS, and SiPM is set outside the GIS observation window; the adjustment port of the voltage divider, the HFCT signal acquisition port, the UHF signal acquisition port, the AE signal acquisition port, and the SiPM signal acquisition port are connected to the partial discharge multi-physics detection host computer through a high-speed acquisition card.

[0037] Figure 2 Schematic diagram of adaptive extraction of multi-physics pulses of partial discharge in a GIS insulation defect identification method of the present invention. Figure 2 As shown in the figure, under the power frequency AC voltage, the protective resistor, voltage divider, coupling capacitor and transformer in the partial discharge device each play a key role to ensure the safety and accuracy of the test.

[0038] Protective resistors provide current limiting and short-circuit protection. Connected in series between the high-voltage power supply and the device under test, they limit short-circuit current and prevent damage to the device or power supply due to overcurrent. They also damp oscillations, suppressing transient oscillations or voltage reflections in the circuit and reducing the impact of high-frequency interference on test results.

[0039] Voltage dividers are primarily used for voltage measurement, reducing applied high voltages proportionally to a safe range for accurate measurement of the power frequency voltage amplitude by oscilloscopes or digital meters. They are also used to ensure the accuracy of test voltages to determine the inception voltage or extinction voltage of partial discharges.

[0040] The coupling capacitor provides a low-impedance path for the high-frequency pulses of partial discharge (PD), coupling them to the detection circuit while blocking the high power-frequency voltage. Together with the detection impedance (such as an inductor or resistor), it forms a high-pass filter, optimizing PD signal extraction efficiency. The transformer primarily provides a stable high power-frequency voltage, which is applied to the device under test to induce partial discharge. By adjusting the transformation ratio, the amplitude of the test voltage can be precisely controlled to meet diverse testing requirements.

[0041] Synchronous data acquisition of partial discharge signals was performed using a high-frequency Rogowski coil (HFCT), an ultra-high frequency transducer (UHF), and an ultrasonic transducer (AE). The HFCT card was attached to the ground wire connected to the tank and the terminal ground wire. The AE sensor and the UHF transducer were mounted on the surface of the insulating bushing. The SiPM array was fixed to the experimental chamber via a flange, with a quartz light guide mounted on the flange head. All sensor quantities were connected to a high-speed acquisition card via a high-speed USB 3.0 cable and used by a multi-physics detection host for discharge pulse extraction, feature calculation, and discharge identification.

[0042] Step 2: Dimensionless energy binary feature extraction is performed on the HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate feature vectors.

[0043] The method performs dimensionless energy binary feature extraction on HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate a feature vector, including: selecting any three ports from the adjustment port of the voltage divider, the HFCT signal acquisition port, the UHF signal acquisition port, the AE signal acquisition port, and the SiPM signal acquisition port; converting the signals of the three ports into dimensionless energy release power of the corresponding sensors, performing binary feature extraction on the signals; and converting the extracted dimensionless energy release power into a two-dimensional space.

[0044] HFCT utilizes the principle of magnetic coupling, transmitting high-frequency current to the secondary coil via a magnetic field, converting it into a low-frequency current that is easier to measure. The high-frequency current in the primary coil generates a varying magnetic field, which induces current in the secondary coil. Due to the very high coupling coefficient between the primary and secondary coils, a low-frequency current signal proportional to the current in the primary coil is obtained. The high-frequency current sensor uses a PEARSON-8600 Rogowski coil manufactured by Pearson Corporation in the United States, with a detection bandwidth of 460kHz to 120MHz and a sensitivity of 10V / A.

[0045] The present invention uses a calibration pulse generator to conduct discharge capacity calibration tests on a pulse current measurement system in a laboratory, obtaining calibration coefficients for three discharge types. In one embodiment, the pulse amplitude-discharge capacity conversion coefficient for corona discharge is 5.471, for creeping discharge is 3.919, and for suspension discharge is 7.371. The converted amplitudes can be used to construct a ternary characteristic parameter.

[0046] In ultra-high frequency sensors (UHF), when partial discharge occurs, the current I caused by the displacement of positive and negative charges can be represented by a magnetic dipole for a PD with a discharge amount of q within a pulse duration. I is written in vector form as I = jωq, and the differential vector potential dA of any point P in space can be calculated:

[0047]

[0048] Where, represents the wave number in the unbounded medium, μ represents the spatial magnetic permeability, ε is the spatial dielectric constant, and r represents the radial distance between the center of the magnetic dipole and point p. Their geometric relationship is given by the PD-RC type measurement circuit. And:

[0049] dA r =dA z cosθ

[0050] dA θ =-dA z sinθ

[0051] The magnetic field H of a dipole is:

[0052]

[0053] The electric field E is:

[0054]

[0055] in:

[0056]

[0057] Where, is the eigenwave impedance of a uniform plane electromagnetic wave in free space.

[0058] Since the distance between the sensor and the discharge source in the device of the present invention is very close, which is much smaller than the wavelength of the UHF signal, the sensor can be regarded as being located in the near field of the dipole, i.e., e -jβr ≈1.

[0059] From this we can get the electric field and magnetic field as:

[0060]

[0061] Combining the above two equations, the density of the electromagnetic radiation energy flow at the sensor is:

[0062]

[0063] Since the electromagnetic radiation source generated by the PD radiates uniformly in any direction in space, the energy of the electromagnetic wave can be solved by integrating the energy flux density at a point r away from the radiation source. The result is still a quadratic function of the current source I.

[0064]

[0065] Where, α f is the discharge attenuation constant. The amplitude of the pulse voltage Um can be solved from the above formula as follows:

[0066] U m =q / [C c +C a C d / (C a +C d )]

[0067] Then the pulse current can be obtained as:

[0068]

[0069] Where,

[0070]

[0071] Let k = R[C c +C a C d / (C a +C d )],but

[0072] I m =q / k s

[0073] Where, the pulse current Im is the peak current of the radiation current source I, and we can get

[0074]

[0075] According to the principle of UHF antenna receiving electromagnetic waves, the relationship between the induced voltage U on the sensor and the electric field strength E at that location is:

[0076] U=kE

[0077] k is a parameter determined by the antenna's own characteristics and position. In the PRPD statistical mode, the waveform of the partial discharge pulse can be abstracted as a standard Gaussian pulse. Therefore, the quadratic integral of the ultra-high frequency signal during a partial discharge is the square of the discharge pulse amplitude A. U 2 The product of θ and the standard Gaussian pulse integral value δ. In this case, θ, Parameters such as r and β are all constant. Then the optical radiation power P of the partial discharge is UHF for:

[0078]

[0079] For ultrasonic sensors (AE), the ultrasonic signal detected by the acoustic emission sensor on the outer wall of the equipment cavity is generated by the partial discharge source, propagated through the SF6 gas, and transmitted through the cavity wall. When the relative position of the discharge source and the sensor is fixed, the transmission path of the partial discharge ultrasonic signal is fixed. Therefore, the sound power received by the sensor and the sound power released by the discharge source itself have a fixed attenuation multiple. The sound energy received by the acoustic emission sensor is directly proportional to the sum of the squares of the amplitudes:

[0080]

[0081] When monitoring partial discharge (PD), the true values ​​of the various energies released during the discharge process cannot be accurately calculated directly due to the limitations of the sensor's bandwidth, sensitivity, and detection scale. Therefore, the discharge pulse signal obtained by a multi-physics sensing device is essentially only the apparent value of the signal amplitude (or energy). Ignoring the nonlinear attenuation effects that each physical process may encounter along the transmission path, it can be assumed that these measured apparent values ​​maintain a relatively stable proportional relationship. Therefore, based on the sensor's maximum dynamic response value, the energy release power of each physical energy is normalized to a constant.

[0082] Binary feature extraction is performed on the dimensionless energy release power of each sensor, and the three energy release power values ​​are converted into two-dimensional space based on the following formula, namely:

[0083]

[0084] Step 3: Establish a real-time online incremental database of multi-scale insulation defect partial discharge and use the Bayesian optimization algorithm to automatically adjust the parameters of the incremental support vector machine model.

[0085] Figure 3 This is a schematic diagram of binary feature extraction for calculating dimensionless energy release power in a GIS insulation defect identification method of the present invention. Figure 3 , established a real-time online incremental database of multi-scale insulation defect partial discharge, and used the Bayesian optimization algorithm to automatically adjust the parameters of the incremental support vector machine model, including: taking the dimensionless energy release power in two-dimensional space as input and inputting it into multiple SVM models to obtain partial discharge classification results; using Bayesian optimization to optimize the SVM hyperparameters.

[0086] In a conventional binary support vector machine model, the output for a sample classification is either "1" or "-1," indicating whether the sample belongs to the current classification or not. This is called a hard output. Using multiple SVM models can yield a variety of fault classifications. The SVM model input is the result of the calculation in step 2.

[0087] The SVM hyperparameters are optimized using Bayesian optimization, including: the SVM hyperparameters include kernel function type, C and gamma.

[0088] The kernel function type, C value, and gamma value in the SVM model are the main hyperparameters that need to be adjusted. Common kernel functions include the linear kernel function (LinearKernelFunction), the polynomial kernel function (PolynomialKernelFunction), the Gaussian radial basis kernel function (RadialBasisKernelFunction, RBF), and the sigmoid kernel function (SigmoidKernel, Sigmoid).

[0089] The linear kernel function is relatively simple and suitable for linearly separable data sets, but when the data set is linearly inseparable, the linear kernel function performs poorly; the polynomial kernel function is suitable for some nonlinear and complex data sets, but is prone to overfitting; the RBF kernel function can handle nonlinear relationships well and is suitable for most data sets. Among all kernel functions, the RBF kernel function is the most commonly used, but its computational cost is high; the Sigmoid kernel function is suitable for nonlinearly separable data sets, but in practice, the Sigmoid kernel function is rarely used and has a slow training speed.

[0090] The C value is a regularization parameter that controls the misclassification penalty of the classifier. Generally speaking, it can be set to a floating point number between 0 and 10. Smaller C values ​​produce a looser decision boundary, causing more points to be misclassified, while larger C values ​​produce a tighter decision boundary, causing more points to be correctly classified.

[0091] The gamma value is a hyperparameter used to control the influence of the RBF kernel function. It is a positive scalar that represents the degree of influence of a data point on the decision boundary. Generally, it can be set to a floating-point number between 0 and 10. A larger gamma value reduces the influence of each support vector, which may lead to overfitting. Conversely, a smaller gamma value may lead to underfitting. Therefore, it is necessary to adjust these hyperparameters to obtain the optimal combination for the SVM model to achieve optimal classification performance.

[0092] Bayesian optimization mainly includes the following steps:

[0093] 1) Define the objective function: In the context of hyperparameter tuning, the objective function is usually the model evaluation function (such as cross-validation accuracy, loss function, etc.). For the SVM model, the objective function can be the accuracy based on cross-validation.

[0094] 2) Choosing a prior distribution: The prior distribution of a hyperparameter models the probability distribution of its possible values. In Bayesian optimization, the specific form of the prior distribution is usually not specified directly, but is implicitly represented through a surrogate model (such as a Gaussian process).

[0095] 3) Building a Surrogate Model: A surrogate model is used to predict the objective function value for unknown data points based on existing observations (i.e., the objective function values ​​under different hyperparameter configurations). The most commonly used surrogate model is the Gaussian Process (GP). A GP defines a prior distribution by specifying a mean function and a covariance function (kernel function) and then updates the posterior distribution based on the observed data.

[0096] 4) Selecting an Acquisition Function: The acquisition function determines the next hyperparameter configuration to evaluate. Common acquisition functions include expected improvement (EI) and probability of improvement (PI). The goal of the acquisition function is to strike a balance between exploration (exploring untested hyperparameter regions) and exploitation (refining the search within known optimal regions).

[0097] 5) Optimization Iteration: Based on the guidance of the acquisition function, the next hyperparameter configuration point is selected for evaluation, the surrogate model is updated, and this process is repeated until the preset number of evaluations is reached or other stopping conditions are met.

[0098] The present invention uses Bayesian optimization to optimize the SVM hyperparameters. The steps are as follows:

[0099] 1) Define the objective function: In SVM hyperparameter tuning, the objective function can be the accuracy based on cross-validation. Specifically, for each hyperparameter configuration, use cross-validation to evaluate the performance of the SVM model and use the accuracy as the output of the objective function.

[0100] 2) Initialize the surrogate model: Select appropriate mean and covariance functions to initialize the Gaussian process model. The choice of these functions will affect the predictive performance of the surrogate model.

[0101] 3) Optimization Iterations: In each iteration, the acquisition function is used to select the next hyperparameter configuration point for evaluation. The SVM model is trained using the selected hyperparameter configuration and its performance is evaluated through cross-validation. New observations (hyperparameter configurations and corresponding accuracy) are added to the surrogate model and the posterior distribution is updated. The above steps are repeated until the preset number of evaluations is reached or other stopping conditions are met. Select the optimal hyperparameters: Among all the evaluated hyperparameter configurations, the hyperparameter combination that achieves the optimal objective function (accuracy) is selected as the final hyperparameter configuration.

[0102] For the multi-classification task of defect recognition, the number of categories is K, and SVM is used to output the result vector for the input sample x. To this end, the result vector f(x) is converted into a calibrated probability P(y=k|x) through Dirichlet calibration, satisfying the probability normalization:

[0103]

[0104] For the trained SVM classifier, its decision function is:

[0105]

[0106] Where wk is the weight vector of class k and bk is the bias term. The scores of the base classifier f(x) are linearly transformed based on Dirichlet calibration and then the probabilities are generated through softmax:

[0107]

[0108] where α k is the scaling factor, β kis the parameter to be optimized. Then maximizing the likelihood function of the calibration probability is equivalent to minimizing the negative log-likelihood loss (i.e., cross entropy loss):

[0109]

[0110] P(y=k|x i ) and then expand the following formula:

[0111]

[0112] For the optimization parameter {α k ,β k}, we need to calculate the gradient of the loss function with respect to the parameters:

[0113]

[0114] The parameters are iteratively updated by gradient descent to minimize the loss function L:

[0115] 1) Initialization parameters: (usually initialized to unit transform);

[0116] 2) Calculate gradient: use current parameters to calculate and

[0117] 3) Update parameters:

[0118]

[0119] Where η is the learning rate.

[0120] 4) Iteration termination: Stop when the loss function converges or reaches the maximum number of iterations

[0121] For the test sample x', the calibrated recognition probability result is:

[0122]

[0123] Figure 4 This figure shows a schematic diagram of a Bayesian-optimized SVM soft output model used in a GIS insulation defect identification method. By normalizing the posterior probabilities of the outputs of multiple binary SVM models based on a Dirichlet parameter optimization model, the posterior probabilities of the sample in each classification category calculated by the SVM under the multi-classification model are obtained.

[0124] In a computer method, the encoding matrix M∈{-1,0,1} can be introduced N×n , encode the c category labels as Two-class classification problem The zero symbol represents a category that is completely ignored. In addition, by using loss-weighted ternary decoding and a quadratic loss function, the loss values ​​of all categories are kept in the same dynamic range, thereby improving the robustness and reliability of classification.

[0125] For n SVM models, the probability estimate of the jth model is:

[0126]

[0127] In multi-class classification, a K×L encoding matrix M=[m kj ]∈{-1,0,1}, where 1 means k is marked as positive in subproblem l, -1 means negative, and 0 means it does not participate in the training of this subproblem.

[0128] For each column of the encoding matrix, an SVM classifier is trained to distinguish the categories of 1 and -1 in the column (ignoring the category corresponding to 0), thereby forming multiple SVM models of the present invention;

[0129] For the test sample, the prediction results {fl(x)} of all classifiers are collected, and the final category is determined by maximizing the association between the prediction results and the encoded line:

[0130]

[0131] In the ternary error correction coding model, each column of the coding matrix corresponds to an independent support vector machine model. For the SVM in the lth column, the optimization objective is:

[0132]

[0133] Where W l and b l are the training parameters of the SVM model,

[0134] C is the regularization parameter,

[0135] ξ i,l variables added to the optimization process,

[0136] y i,l ∈{-1,1}, is the value of the i-th sample at the l-th encoding position, and N is the number of samples.

[0137] In one embodiment, Corresponding to the quadratic loss function, it replaces the standard SVM's prior loss. In the decoding stage, the prediction results of multiple binary classifiers need to be matched with the category codewords in the encoding matrix. At this time, the quadratic loss function is the aforementioned category loss |M kl -sign(f l (x))|.

[0138] In addition, weights are assigned to different code positions or categories to form a weighted quadratic loss function:

[0139]

[0140] M k,l is the weight of category k at the lth encoding position.

[0141] This loss function can be used to iteratively improve the model.

[0142] Step 4: input the feature vector into the SVM model with Bayesian optimization hyperparameters to perform multi-parameter partial discharge recognition.

[0143] The feature vector is input into an SVM model with Bayesian hyperparameter optimization to perform multi-parameter partial discharge identification, including: constructing an incremental database of partial discharge data of multi-scale insulation defects, conducting online collection of partial discharge multi-physical pulse data under insulation defects of different detection scales; using the optimized SVM model to extract the two-dimensional distribution of the posterior probability of different types of faults, and analyzing the distribution differences of different types of faults.

[0144] Figure 5 A schematic diagram of the two-dimensional distribution of the posterior probability of the partial discharge diagnosis model in the GIS insulation defect identification method of the present invention. Figure 5 As shown in the figure, the proposed model generates a two-dimensional distribution of the posterior probabilities of three typical insulation defects. Feature extraction from field data can be used to visualize the identification results in the figure. Furthermore, a comparison is made with various traditional, unoptimized machine learning classification models, such as Naive Bayesian classification, decision trees, and k-nearest neighbors.

[0145] For the above model, the true positive rate (TPR) (i.e., the correct classification rate of samples) and the positive prediction rate (PPV) (i.e., the correct prediction rate of the results) of the test set are shown in Table 1. The proposed model achieves the best recognition accuracy.

[0146] Table 1 Optimized set of chi-square test for partial discharge characteristics

[0147]

[0148] The second aspect of the present invention relates to a GIS insulation defect identification system; the system is implemented using a GIS insulation defect identification method in the first aspect of the present invention; the method includes an extraction module, a feature module, an optimization module, and an identification module; the extraction module is used to use an adaptive pulse sequence threshold determination method based on the interquartile range to adaptively extract a variety of GIS internal partial discharge pulse signals including HFCT, UHF and optical detection equipment; the feature module is used to perform dimensionless energy binary feature extraction on HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate a feature vector; the optimization module is used to establish a real-time online incremental database of multi-scale insulation defect partial discharge, and use a Bayesian optimization algorithm to automatically adjust the parameters of the incremental support vector machine model; the identification module is used to input the feature vector into an SVM model with Bayesian optimized hyperparameters to perform multi-parameter partial discharge identification.

[0149] A third aspect of the present invention relates to a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; and the processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1-8.

[0150] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in the first aspect of the present invention.

[0151] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art will appreciate that the technical solutions of the present invention still include modifications or equivalent substitutions that may be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are intended to be covered by the claims of the present invention.

Claims

1. A GIS insulation defect identification method, characterized in that: The method comprises the following steps: Adopting an adaptive pulse sequence threshold determination method based on interquartile range, the partial discharge pulse signals inside various GIS including HFCT, UHF and optical detection equipment are adaptively extracted. Dimensionless energy binary feature extraction is performed on HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate feature vectors; Establish a real-time online incremental database of multi-scale insulation defect partial discharge, and use the Bayesian optimization algorithm to automatically adjust the parameters of the incremental support vector machine model; The feature vector is input into a SVM model with Bayesian optimization of hyperparameters to perform multi-parameter partial discharge recognition.

2. A GIS insulation defect identification method according to claim 1, characterized in that: The adaptive extraction of various GIS internal partial discharge pulse signals including high-frequency coils, ultra-high frequency and optical detection equipment includes: Connect the protection resistor in series between the high-voltage power supply and the high-voltage port of the GIS device under test. The high-voltage power supply is composed of the output winding of the transformer. An HFCT is connected between the high-voltage port of the GIS device under test and the ground. A voltage divider and a coupling capacitor are also connected in parallel between the high-voltage port of the GIS device under test and the ground. UHF and AE are installed on the outer surface of GIS, and SiPM is installed outside the GIS observation window; The voltage divider's adjustment port, HFCT signal acquisition port, UHF signal acquisition port, AE signal acquisition port, and SiPM signal acquisition port are connected to the partial discharge multi-physics detection host computer through a high-speed acquisition card.

3. A GIS insulation defect identification method according to claim 2, characterized in that: The dimensionless energy binary feature extraction is performed on the HFCT, UHF and AE partial discharge signals within a fixed acquisition time. Generate feature vectors, including: Any three ports are selected from the adjustment port of the voltage divider, the HFCT signal acquisition port, the UHF signal acquisition port, the AE signal acquisition port, and the SiPM signal acquisition port; The signals of the three ports are converted into dimensionless energy release power of the corresponding sensors, and binary features are extracted from them; The extracted dimensionless energy release power is converted into two-dimensional space.

4. A GIS insulation defect identification method according to claim 3, characterized in that: The method of establishing a real-time online incremental database of multi-scale insulation defect partial discharge and automatically adjusting the parameters of the incremental support vector machine model using a Bayesian optimization algorithm includes: The dimensionless energy release power in two-dimensional space is used as input to multiple SVM models to obtain partial discharge classification results; Bayesian optimization is used to optimize the SVM hyperparameters.

5. A GIS insulation defect identification method according to claim 4, characterized in that: The Bayesian optimization is used to optimize the SVM hyperparameters, including: The SVM hyperparameters include kernel function type, C and gamma.

6. A GIS insulation defect identification method according to claim 5, characterized in that: The Bayesian optimization is used to optimize the SVM hyperparameters, including: For the lth SVM model, the optimization objective is: Where W l and b l are the training parameters of the SVM model, C is the regularization parameter, ξ i,l variables added to the optimization process, y i,l ∈{-1,1}, is the value of the i-th sample at the l-th encoding position, and N is the number of samples.

7. A GIS insulation defect identification method according to claim 6, characterized in that: The step of inputting the feature vector into an SVM model with Bayesian optimization of hyperparameters to perform multi-parameter partial discharge identification includes: Construct an incremental database of partial discharge data for multi-scale insulation defects and conduct online collection of multi-physics pulse data for partial discharge under insulation defects of different detection scales; The optimized SVM model is used to extract the two-dimensional distribution of the posterior probability of different types of faults, and the distribution differences of different types of faults are analyzed.

8. A GIS insulation defect identification system; characterized by: The system is implemented using a GIS insulation defect identification method according to any one of claims 1 to 7; The method includes an extraction module, a feature module, an optimization module, and a recognition module; The extraction module is used to adaptively extract the internal partial discharge pulse signals of various GIS including HFCT, UHF and optical detection equipment using an adaptive pulse sequence threshold determination method based on the interquartile range; The feature module is used to extract dimensionless energy binary features of HFCT, UHF and AE partial discharge signals within a fixed acquisition time to generate feature vectors; The optimization module is used to establish a real-time online incremental database of multi-scale insulation defect partial discharge and automatically adjust the parameters of the incremental support vector machine model using a Bayesian optimization algorithm; The identification module is used to input the feature vector into the SVM model with Bayesian optimization of hyperparameters to perform multi-parameter partial discharge identification.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images