Withstand voltage flashover ultrasonic signal determination method and related device

Abstract

The invention provides a voltage-withstanding flashover ultrasonic signal determination method and a related device. The method comprises the following steps: acquiring a discharge data set; first feature information is determined based on the discharge data set and a first model, second feature information is determined based on the discharge data set and a second model, the first model is an unsupervised dimensionality reduction feature extraction model and is used for analyzing principal components of the discharge data set, and the second model is a supervised dimensionality reduction feature extraction model and is used for analyzing principal components of the discharge data set. The feature classification module is used for performing feature classification; determining a double-branch prediction probability based on the first feature information, the second feature information, and an accurate clustering sub-model and a fuzzy clustering sub-model of a pre-constructed clustering model; and determining the withstand voltage flashover ultrasonic signal type of the target network based on the double-branch prediction probability. In this way, high-precision and high-robustness recognition of multi-type flashover and noise in the GIS is achieved, the feature utilization rate can be increased, the model generalization ability can be improved, and the calculation efficiency can be improved.

Description

Technical Field

[0001] This invention relates to the field of electrical technology, and in particular to a method and apparatus for determining withstand voltage flashover ultrasonic signals. Background Technology

[0002] In power equipment such as gas-insulated switchgear (GIS), insulation defects can lead to GIS flashover, which in turn causes equipment failure. Ultra-high frequency (UHF) technology is an important means of detecting GIS flashover discharge; by performing pattern recognition on flashover signals, the type of insulation defect can be diagnosed.

[0003] Current methods still suffer from the following technical problems: insufficient integration of multi-dimensional features leads to redundant or missing feature information; chaotic feature extraction requires high-dimensional phase space reconstruction, resulting in high computational costs; in field detection, voltage phase shifts and grid noise interference cause phase mismatch in the spectrum, and current classifiers have low accuracy in recognizing phase-shifted data. Single neural fuzzy classifiers or single radial basis function neural networks struggle to handle complex flashover discharge data distributions and are prone to getting trapped in local optima; adding input vectors or fuzzy rules leads to a surge in model complexity and poor field applicability; while deep learning methods offer high recognition accuracy, their deep structure results in long training times, failing to meet the demands of real-time field detection. In summary, current methods also suffer from low feature utilization, poor phase shift robustness, insufficient model generalization ability, and low computational efficiency. Summary of the Invention

[0004] This application provides a method and related apparatus for determining withstand voltage flashover ultrasonic signals, which can improve feature utilization, phase shift robustness, model generalization ability, and computational efficiency.

[0005] In a first aspect, embodiments of this application provide a method for determining a withstand voltage flashover ultrasonic signal, including: Obtain a discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network; First feature information is determined based on the discharge dataset and the first model, and second feature information is determined based on the discharge dataset and the second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization. The bi-branch prediction probability is determined based on the first feature information, the second feature information, the precise clustering sub-model and the fuzzy clustering sub-model of the pre-constructed clustering model; The type of withstand voltage flashover ultrasonic signal of the target network is determined based on the bi-branch prediction probability.

[0006] In one possible embodiment, the first model includes a principal component analysis model, and the determination of first feature information based on the discharge dataset and the first model includes: Determine the covariance matrix based on the discharge dataset; The eigenvectors and eigenvalues ​​are determined based on the covariance matrix; A first feature space is determined based on the feature vector and the feature value, and the first feature space is the first feature information.

[0007] In one possible embodiment, the discharge dataset includes multiple types of discharge data, which correspond to multiple types of fault conditions. The second model includes a linear discriminant analysis model. Determining the second feature information based on the discharge dataset and the second model includes: A statistical feature space is determined based on the discharge dataset, and the statistical feature space is used to characterize the waveform features of the multiple types of discharge data. Based on the discharge dataset, an intra-class scatter matrix and an inter-class scatter matrix are determined. The intra-class scatter matrix represents the distribution of data points within the same class of discharge data, and the inter-class scatter matrix represents the distribution of data points within different classes of discharge data. The projection matrix is ​​determined based on the intra-class scatter matrix and the inter-class scatter matrix; The second feature space is determined based on the projection matrix; A third feature space is determined by feature fusion based on the second feature space and at least one statistical feature space, wherein the third feature space is the second feature information and the statistical feature space.

[0008] In one possible embodiment, determining the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-built precise clustering sub-model and fuzzy clustering sub-model of the clustering model includes: The first feature information and the second feature information are input into the precise clustering sub-model, and the first feature information and the second feature information are input into the fuzzy clustering sub-model; The first branch result of the output of the precise clustering sub-model is determined based on the precise clustering sub-model, and the second branch result of the output of the fuzzy clustering sub-model is determined based on the fuzzy clustering sub-model; The dual-branch prediction probability is determined by performing multi-feature fusion based on the results of the first branch and the second branch.

[0009] In one possible embodiment, the precise clustering sub-model includes a hard C-means clustering model, and determining the first branch result output by the precise clustering sub-model based on the precise clustering sub-model includes: Multiple first activation intensities are determined based on the first feature information, the second feature information, and the first membership function; The normalized activation intensity is determined based on the plurality of first activation intensities; The first consequent function is determined based on the normalized activation intensity; The first branch result is determined based on the first consequent function, the normalized activation strength, and the preset weight.

[0010] In one possible embodiment, the fuzzy clustering sub-model includes a fuzzy C-means clustering model, and determining the second branch result output by the fuzzy clustering sub-model based on the fuzzy clustering sub-model includes: Multiple second activation intensities are determined based on the first feature information, the second feature information, and the second membership function; The second consequent function is determined based on the plurality of second activation intensities; The second branch result is determined based on the second consequent function and the preset weight.

[0011] In one possible embodiment, the step of performing multi-feature fusion based on the first branch result and the second branch result to determine the dual-branch prediction probability includes: Based on the results of the first branch, determine the first loss function of the precise clustering sub-model; The second loss function of the fuzzy clustering sub-model is determined based on the results of the second branch. Based on the first loss function, the second loss function, and the influence factor, multi-feature fusion is performed to determine the dual-branch prediction probability.

[0012] Secondly, embodiments of this application provide a device for determining withstand voltage flashover ultrasonic signals, comprising: The data determination module is used to acquire a discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network; The feature extraction module is used to determine first feature information based on the discharge dataset and the first model, and to determine second feature information based on the discharge dataset and the second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization. The model analysis module is used to determine the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-built precise clustering sub-model and fuzzy clustering sub-model of the clustering model; The result determination module is used to determine the type of withstand voltage flashover ultrasonic signal of the target network based on the bi-branch prediction probability.

[0013] Thirdly, embodiments of this application provide a computer-readable storage medium storing a withstand voltage flashover ultrasonic signal determination program. The withstand voltage flashover ultrasonic signal determination program includes execution instructions. When a processor executes the execution instructions stored in the memory, the processor performs some or all of the steps described in the first aspect.

[0014] Fourthly, embodiments of this application provide an electronic device, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and when the processor executes the one or more programs, the processor executes some or all of the instructions of the steps described in the first aspect of the embodiments of this application.

[0015] Fifthly, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the first aspect of embodiments of this application. The computer program product may be a software installation package.

[0016] By implementing the embodiments of this application, a discharge dataset is obtained, which consists of discharge data from multiple experimental models with offset voltage phases, used to simulate a target network. First feature information is determined based on the discharge dataset and a first model, and second feature information is determined based on the discharge dataset and a second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used for feature categorization. A bi-branch prediction probability is determined based on the first feature information, the second feature information, and pre-constructed precise and fuzzy clustering sub-models of a clustering model. The bi-branch prediction probability is then used to determine the withstand voltage flashover ultrasonic signal type of the target network. This achieves high-precision and robust identification of multiple types of flashovers and noise in GIS, improving feature utilization, model generalization ability, and computational efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the accompanying drawings used in the embodiments of the present invention or the background art will be described below.

[0018] Figure 1 This is a schematic diagram of the architecture of a withstand voltage flashover ultrasonic signal determination system provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the first method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment; Figure 3 This is a flowchart illustrating the second method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment; Figure 4 This is a flowchart illustrating the third method for determining withstand voltage flashover ultrasonic signals provided in the embodiments of this application; Figure 5 This is a schematic diagram of the algorithm flow for a method for determining a withstand voltage flashover ultrasonic signal proposed in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a withstand voltage flashover ultrasonic signal determination device provided in an embodiment of this application; Figure 7 This is a schematic diagram of another pressure-resistant flashover ultrasonic signal determination device provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0020] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or electronic device that includes a series of steps or units is not limited to the listed steps or units, but in an alternative example also includes steps or units not listed, or in an alternative example also includes other steps or units inherent to these processes, methods, products, or electronic devices.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] In power equipment such as gas-insulated switchgear (GIS), insulation defects can lead to GIS flashover, which in turn causes equipment failure. Flashover refers to the breakdown discharge phenomenon that occurs along the surface of insulating materials or insulators under high voltage. It is not a breakdown inside the insulator, but rather surface creepage. In GIS, flashover is a serious insulation fault that releases enormous energy and endangers equipment safety. Ultra-high frequency (UHF) technology is an important means of detecting flashover discharge in GIS. By performing pattern recognition on flashover signals, the type of insulation defect can be diagnosed. In existing technologies, partial discharge pattern recognition methods mainly include statistical classifiers (such as Bayesian classifiers), artificial neural networks (ANNs), fuzzy logic classifiers, and deep learning methods (such as convolutional neural networks CNNs). Current technical solutions still have the following technical problems: limitations in feature extraction; traditional methods mostly rely on single feature extraction techniques and do not fully integrate multi-dimensional features, resulting in redundant or missing feature information; chaotic feature extraction requires high-dimensional phase space reconstruction, which has high computational costs. The model suffers from insufficient robustness. Voltage phase shifts and grid noise interference during field detection can lead to phase mismatch in the data spectrum, significantly reducing the accuracy of existing classifiers in recognizing phase-shifted data. The network structure is also limited. Single neural fuzzy classifiers or radial basis function neural networks (RBFNNs) struggle to handle complex flashover discharge data distributions and are prone to getting trapped in local optima. Increasing the input vector or fuzzy rules leads to a surge in model complexity, resulting in poor field applicability. Furthermore, computational efficiency is a concern. While deep learning methods (such as CNNs) offer high recognition accuracy, their deep structures result in long training times (typically exceeding 100 seconds), failing to meet the demands of real-time field detection.

[0023] To address the aforementioned issues, this application provides a method and related apparatus for determining withstand voltage flashover ultrasonic signals, which can achieve high-precision and robust identification of various types of flashovers and noise in GIS, improve feature utilization, enhance model generalization ability, and increase computational efficiency.

[0024] The method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment can be applied to, for example... Figure 1 In the withstand voltage flashover ultrasonic signal determination system shown, please refer to... Figure 1 , Figure 1 This is a schematic diagram of the architecture of a withstand voltage flashover ultrasonic signal determination system provided in an embodiment of this application. The withstand voltage flashover ultrasonic signal determination system 100 includes a server 110 and a network under test 120. The network under test 120 can transmit information to the server 110 through a network or wired connection.

[0025] In this scheme, the network under test 120 is the circuit under test that may generate GIS data. Server 110 refers to a remote computer used for processing large amounts of computational tasks and storing data. In this scheme, a pre-built clustering model is deployed on server 110. This pre-built clustering model includes precise clustering sub-models and fuzzy clustering sub-models, used for clustering analysis based on feature information. Server 110 can also be used to collect data during model usage, facilitating subsequent optimization of the pre-built clustering model.

[0026] Based on this, this application provides a method and related apparatus for determining withstand voltage flashover ultrasonic signals, which will be described in detail below with reference to the accompanying drawings.

[0027] Please see Figure 2 , Figure 2 This is a flowchart illustrating the first method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment, as shown below. Figure 2 As shown, the method includes the following steps: S210, Obtain the discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network.

[0028] Among them, multiple experimental models are used to simulate artificial defect models of multiple flashover types, and discharge data are obtained through experiments based on multiple experimental models.

[0029] Specifically, the various experimental models include at least one of the following: a flashover model of solid insulation materials, a flashover model of conductor burrs or scratches, and a flashover and noise model of foreign object protrusions or metal particles. The aforementioned flashover model of solid insulation materials is used to simulate surface discharge on the surface of the supporting insulator (pot-type insulator) in GIS caused by pollution, aging, moisture, etc. The flashover model of conductor burrs or scratches is used to simulate the sharp burrs or scratches on the high-voltage conductor caused by manufacturing, installation, or electrodynamic vibration. These sharp points can cause electric field concentration, which can easily induce partial discharge and may develop into flashover. The flashover and noise model of foreign object protrusions or metal particles is used to simulate metal debris, screws, or foreign objects that are accidentally dropped into the GIS cavity. These foreign objects will move and jump under the action of the electric field and may form bridging on the surface of the insulator, eventually causing flashover.

[0030] Optionally, discharge data generated by the model can be collected using sensors or other methods. Preferably, ultrasonic sensors can be used to collect discharge data because when partial discharge or flashover occurs, a tiny explosion is generated instantaneously, and energy propagates in the form of sound waves and ultrasonic waves. In some cases, in addition to ultrasonic sensors, other monitoring methods can be used, such as ultra-high frequency (UHF) sensors to capture the electromagnetic waves radiated by the discharge; pulsed current to directly measure the amount of charge discharged; and high-speed cameras to observe the physical process of flashover. Other methods can also be used, and are not limited here. In some cases, multiple technologies can be used for joint detection and mutual verification to obtain more comprehensive information.

[0031] The discharge dataset is a set of characteristic data obtained by processing discharge data generated from multiple experimental models. It includes discharge signals under normal conditions and various typical defects (such as flashover of solid insulation materials, conductor burrs, metal particles, etc.) after preprocessing. Since there may be a phase difference between the experimental model and the actual field voltage, in order to simulate the field voltage phase situation, it is necessary to convert the flashover discharge data detected in the laboratory into discharge data with offset voltage phase.

[0032] Optionally, the process of converting laboratory-detected discharge data into discharge data with an offset voltage phase may include: determining zero-crossing point data by performing voltage zero-crossing detection based on experimental model discharge data; determining the discharge event time based on the zero-crossing point data; determining the offset phase difference based on the discharge time; and determining discharge data with an offset voltage phase based on the offset phase difference and the experimental model discharge data.

[0033] Optionally, the zero-crossing detection based on experimental model discharge data includes: zero-crossing detection, extracting the zero-crossing point from the acquired power supply voltage signal as the phase reference point (0° phase); and period segmentation, dividing one power frequency cycle (20ms, corresponding to 360°) into several phase windows (e.g., one window per 1°) to establish a phase reference coordinate system.

[0034] Optionally, determining the discharge event time based on the zero-crossing data includes: pulse detection, performing pulse detection on the preprocessed discharge signal, and extracting the trigger time of each discharge pulse. Peak correction: Based on the pulse waveform, correct the timing corresponding to the pulse peak. ,in This represents the time offset of the pulse peak relative to the trigger moment.

[0035] Optionally, determining discharge data with an offset voltage phase based on the offset phase difference and the experimental model discharge data includes: time difference calculation, calculating the time difference of the i-th discharge pulse relative to the first pulse: ; Phase difference conversion converts time difference into phase difference (one power frequency cycle is 20ms, corresponding to 360°). ; Normalization is performed by taking the phase difference modulo 360° and reducing it to the range of 0–360°: (mod 360°); Experimental model discharge data includes the discharge amount corresponding to the discharge pulse. The discharge amount of each discharge pulse Its corresponding offset phase Combined, discharge data points with offset voltage phase are formed. These data points can be used to plot... The phase-discharge distribution map is used for flashover discharge mode identification and fault location.

[0036] S220, determine first feature information based on the discharge dataset and the first model, and determine second feature information based on the discharge dataset and the second model, wherein the first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization.

[0037] The first model can include unsupervised models such as principal component analysis (PCA), factor analysis, independent component analysis, nonnegative matrix factorization, t-SNE, UMAP, autoencoders, and random projection. The second model can include supervised models such as linear discriminant analysis (LDA), quadratic discriminant analysis, large-margin nearest neighbor algorithm, supervised t-SNE, supervised autoencoders, random forest feature importance, and gradient boosting machines. The combination of the first and second models can be selected from combinations such as PCA + LDA, UMAP + t-SNE, and autoencoders + supervised learning.

[0038] In one possible embodiment, the first model includes a principal component analysis model, and the step of determining the first feature information based on the discharge dataset and the first model includes: determining a covariance matrix based on the discharge dataset; determining eigenvectors and eigenvalues ​​based on the covariance matrix; and determining a first feature space based on the eigenvectors and eigenvalues, wherein the first feature space is the first feature information.

[0039] Specifically, based on the discharge dataset, its covariance matrix is ​​calculated. This covariance matrix characterizes the linear correlation and variance distribution among the feature parameters of each dimension in the discharge dataset. Eigenvalue decomposition is performed on the covariance matrix to obtain a set of eigenvectors and corresponding eigenvalues. Each eigenvector represents a major direction of change in the data within the original feature space, and the eigenvalue characterizes the variance carried by that direction of change. Based on the eigenvectors and eigenvalues, according to a preset variance contribution rate threshold or a specified number of principal components, the eigenvector corresponding to the largest eigenvalue is selected to form a new linear subspace.

[0040] Specifically, determining the covariance matrix based on the discharge dataset includes: standardizing the data and calculating the covariance matrix C: ; Where N represents the number of data points, and M represents the mean vector of the entire class. This represents the mean vector for each class, where X is the input sample. This represents the input data vector. Determining the eigenvectors and eigenvalues ​​based on the covariance matrix includes: calculating the eigenvectors of the covariance matrix. and eigenvalues : , Determining the first feature space based on the feature vector and the feature value includes: transforming the data X into a feature space using a transformation matrix W. : .

[0041] As can be seen, in this embodiment, the principal component analysis model can reduce the dimensionality of the original discharge dataset into low-dimensional data containing the main feature vectors. It can complete the data dimensionality reduction and feature extraction while retaining the key information of the original data, effectively eliminating data redundancy and noise interference.

[0042] In one possible embodiment, the discharge dataset includes multiple types of discharge data, which correspond to multiple types of fault conditions. The second model includes a linear discriminant analysis model. Determining the second feature information based on the discharge dataset and the second model includes: determining a statistical feature space based on the discharge dataset, wherein the statistical feature space is used to characterize the waveform features of the multiple types of discharge data; determining an intra-class scatter matrix and an inter-class scatter matrix based on the discharge dataset, wherein the intra-class scatter matrix represents the distribution of data points within the same type of discharge data, and the inter-class scatter matrix represents the distribution of data points within discharge data from different classes; determining a projection matrix based on the intra-class scatter matrix and the inter-class scatter matrix; determining a second feature space based on the projection matrix; and performing feature fusion based on the second feature space and at least one statistical feature space to determine a third feature space, wherein the third feature space is the second feature information, and the statistical feature space is the statistical feature space.

[0043] Based on the discharge dataset, statistical features of its waveforms in the time and frequency domains are extracted to construct a statistical feature space. This statistical feature space is used to quantify the differences in shape, distribution, and spectral components of the waveforms of the various types of discharge data. Based on the discharge dataset and its corresponding fault category labels, intra-class scatter matrices and inter-class scatter matrices are calculated. The intra-class scatter matrix represents the density of all discharge data sample points around their category center within the same fault category; the inter-class scatter matrix represents the dispersion of the category center points around the overall sample center between different fault categories. Based on the intra-class and inter-class scatter matrices, an optimization objective function is constructed to find a projection direction that maximizes the ratio of inter-class divergence to intra-class divergence after projection transformation. This optimization problem is solved to obtain the projection matrix. The projection matrix is ​​applied to the original discharge dataset or its feature representation, projecting it to a new low-dimensional subspace, defined as the second feature space, where the features have the strongest class discrimination ability. The feature representation of the second feature space is fused with the feature representation of at least one statistical feature space. This fusion operation includes, but is not limited to, vector concatenation, weighted superposition, or nonlinear combination based on kernel functions. Through this fusion process, a third feature space with both strong discriminative power and rich original statistical information is generated.

[0044] Specifically, the LDA feature extraction process is as follows: Calculate the within-class scatter matrix. Scatter matrix between classes : ; ; Where cs represents the number of classes, N represents the total data, and N represents the data for each class. This represents the mean vector for each class. Each class corresponds to the flashover type in the aforementioned flashover model, such as flashover of solid insulation materials, flashover of conductor burrs or scratches, flashover of foreign object protrusions or metal particles, and noise. The entire class refers to the global mean vector calculated after treating all samples (regardless of their defect category) as a single dataset. Then, the discriminant projection matrix is ​​calculated. : ; in This represents the eigenvector. The eigenvalues ​​are sorted in descending order based on their corresponding eigenvectors. ; ; However, during the transformation process, the feature dimension must be less than or equal to the number of classes minus 1. To address this issue, statistical features are used as complementary features.

[0045] Feature concatenation involves combining the dimensionality-reduced feature vector from LDA with multiple statistical features to obtain a new feature vector. For example, these statistical features might include skewness SK (negative half / positive half / full period), salience K, etc. u At least one of the following: (negative half / positive half / full cycle), correlation degree cc, asymmetry degree Q, and eigenvector X P2 The concatenation process includes feature extraction, feature concatenation, and inputting it into one branch of the EFRBFNN as input to the other branch. Refer to the following formula, where X... LDA+SC The concatenated vector: , If the LDA output is 4-dimensional, , , , These are vectors representing four of the dimensions, with the remainder being the statistical feature vectors mentioned above. The total dimensions are: 4 (LDA) + 8 (statistics) = 12 dimensions.

[0046] Among these, skewness (three characteristics) measures the direction and degree of skewness in the distribution of statistical data, and is a numerical characteristic of the degree of asymmetry in the distribution of statistical data. Kurtosis (three characteristics), also known as the kurtosis coefficient, is a characteristic number that characterizes the peak value of the probability density distribution curve at the mean. Intuitively, kurtosis reflects the sharpness of the peak. The kurtosis of a random variable is calculated as the ratio of the fourth central moment of the random variable to the square of the variance.

[0047] ; Among these, the aforementioned correlation (a feature) measures the "discharge amplitude". "and "number of discharges" "The degree of linear correlation over the entire period, ranging from [-1, 1], is calculated using the following formula:" ; The aforementioned asymmetry (one characteristic) describes the difference in total discharge during the positive and negative half-cycles, and is calculated using the following formula: .

[0048] As can be seen, in this embodiment, by constructing a third feature space (second feature information) that combines discriminative and statistical information, the complementarity and enhancement of feature information are achieved. This fused feature not only retains the main statistical characteristics of the original waveform, but also strengthens the separability between different fault categories, thereby significantly improving the accuracy and robustness of subsequent fault diagnosis models (such as classifiers).

[0049] In one possible embodiment, other dimensionality reduction algorithms such as Independent Component Analysis (ICA) can be used to replace PCA or LDA for feature extraction, but it is necessary to verify their effectiveness in extracting flashover discharge features and their complementarity with statistical features.

[0050] S230, determine the bi-branch prediction probability based on the first feature information, the second feature information, the precise clustering sub-model and the fuzzy clustering sub-model of the pre-constructed clustering model.

[0051] The pre-built clustering model can be a neural network model. Specifically, in this embodiment, it can be an ensemble fuzzy-radial basis function neural network (EFRBFNN) model, which includes a precise clustering sub-model and a fuzzy clustering sub-model. The precise clustering sub-model and the fuzzy clustering sub-model are two complementary clustering sub-models. The precise clustering sub-model is based on the hard partitioning principle, deterministically assigning each data sample to a specific cluster, and its output is a clear and unique membership relationship of the sample to a certain category. The fuzzy clustering sub-model is based on fuzzy set theory, assigning each data sample to all clusters with a certain membership degree, and its output is a membership degree vector of the sample relative to each category. This vector quantifies the probability or degree that the sample belongs to different categories.

[0052] Specifically, the first feature information and the second feature information are input into the clustering model separately or together, and processed in parallel by its two sub-models; or, the first feature information and the second feature information are input into the clustering model separately and then merged, and the merged feature information is input into the first branch and the second branch for parallel processing; or, the first feature information and the second feature information are first merged to obtain merged feature information, and then the merged feature information is input into the clustering model for model processing.

[0053] In the first branch, cluster labels for samples are obtained based on the output of the precise clustering sub-model. These hard labels are then converted into a probabilistic representation, for example, by calculating the relative distance from the sample to its cluster center, or by using a Softmax function to map the distance information into a normalized probability distribution, thus generating the first probability prediction vector (precise branch probability). The second branch utilizes the output of the fuzzy clustering sub-model, namely the membership vector of the sample to all clusters. This vector itself constitutes a probability distribution and is directly used as the second probability prediction vector (fuzzy branch probability). The bi-branch prediction probability is a comprehensive probability output composed of the first and second probability prediction vectors, used to characterize the likelihood of a data sample belonging to a category in both deterministic and fuzzy dimensions.

[0054] Optionally, the aforementioned bi-branch prediction probabilities can be further input into a decision facilitator for fusion to generate the final comprehensive category determination.

[0055] In one possible embodiment, determining the bi-branch prediction probability based on the first feature information, the second feature information, and a pre-built precise clustering sub-model and a fuzzy clustering sub-model of the clustering model includes: inputting the first feature information and the second feature information into the precise clustering sub-model, and inputting the first feature information and the second feature information into the fuzzy clustering sub-model; determining the first branch result output by the precise clustering sub-model based on the precise clustering sub-model, and determining the second branch result output by the fuzzy clustering sub-model based on the fuzzy clustering sub-model; and performing multi-feature fusion based on the first branch result and the second branch result to determine the bi-branch prediction probability.

[0056] The process involves inputting the first and second feature information as input vectors to both a fuzzy clustering sub-model and a precise clustering sub-model. Based on the precise clustering sub-model, the first branch output is obtained, representing a set of discrete hard-classified facial expressions, with each data sample assigned to a specific cluster. Based on the fuzzy clustering sub-model, the second branch output is obtained, representing a membership matrix where each sample data point has a continuous membership value across all clusters, indicating the probability that the sample belongs to a particular cluster. Finally, the first and second branch outputs are fused using multi-feature analysis to produce a dual-branch prediction probability.

[0057] Optionally, the multi-feature fusion process can involve transforming the first branch result into a first probability distribution vector, the second branch result into a second probability distribution vector, and combining the first and second distribution vectors to obtain a dual-branch prediction probability vector. Optionally, the vector fusion of the first and second branch results can be performed using methods such as weighted linear fusion, nonlinear confidence fusion, or decision-level fusion; other methods are also possible and are not limited here. Specifically, the aforementioned weighted linear fusion process can be P... 总 =αP e +βP f α and β are the fusion weight coefficients, and α + β = 1, P e Let P be the first probability distribution vector. f This represents the second probability distribution vector. Specifically, the nonlinear fusion of the confidence scores described above can be achieved by calculating the uncertainty measures (such as information entropy) of the first and second branch results, assigning higher weights to branches with lower uncertainty, thereby realizing adaptive probability fusion. Specifically, the decision-level fusion process described above can use the first and second branch results as input features of a high-level classifier (such as a support vector machine or neural network) for deeper pattern recognition and decision-making.

[0058] As can be seen, in this embodiment, by using precise clustering and fuzzy clustering in parallel, both deterministic and uncertain information in data classification are captured. The multi-feature fusion mechanism effectively avoids misjudgments that may occur due to noise interference, data distribution shift, or parameter mismatch in a single model.

[0059] In one possible embodiment, please refer to Figure 3 , Figure 3 This is a flowchart illustrating the third method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment, as follows: Figure 3 As shown, the precise clustering sub-model includes a hard C-means clustering model, and determining the first branch result output by the precise clustering sub-model based on the precise clustering sub-model includes: S2311, determine multiple first activation intensities based on the first feature information, the second feature information, and the first membership function; S2312, determine a normalized activation intensity based on the plurality of first activation intensities; determine a first consequent function based on the normalized activation intensity; S2313, the first branch result is determined based on the first consequent function, the normalized activation strength, and the preset weight.

[0060] Among them, the precise clustering sub-model includes the hard C-means clustering model. The hard C-means clustering model determines the center and width of the Gaussian radial basis function. The output of the hidden layer nodes is the Gaussian function value, and the output layer uses a linear function. The connection weights are trained through the backpropagation algorithm.

[0061] The first membership function is configured to characterize the degree of association between the input feature vector and each cluster prototype (i.e., cluster center). For each cluster, a first activation intensity is calculated. The multiple first activation intensities together constitute a preliminary membership vector, which characterizes the original association strength between the input sample and each cluster.

[0062] The normalization operation can be performed by calculating the proportion of each first activation intensity in the total sum of all first activation intensities, converting the multiple first activation intensities into a vector of probability distribution, i.e., the normalized activation intensities. Each element of this vector represents the normalized preliminary probability that the input sample belongs to the corresponding cluster.

[0063] The first consequent function is configured as a function mapping, with the input being the normalized activation intensity vector and the output being an inference result directly related to the cluster label.

[0064] Specifically, the output of the first consequent function for each cluster is multiplied by its corresponding normalized activation strength and preset weight, or a weighted sum is performed. This synthesizes the contributions of all clusters to generate a comprehensive discriminant vector. The index, or probability distribution, corresponding to the maximum value of this vector is the result of the first branch.

[0065] Specifically, firstly, in the hidden layer, the i-th hidden layer node pairs with the input... activation intensity Calculated using Gaussian membership functions: ; in, Let i be the i-th cluster center determined by hard C-means clustering. The width parameter of this Gaussian function is usually related to the clustering distribution. Let be the Euclidean distance. Then normalize the activation strengths so that their sum is 1: ; This can help improve the stability of the model. Each hidden layer node corresponds to a consequent function. , Among them, and These are the weight parameters to be learned. Then, the original output of the branch is obtained based on the following formula. : , The original output of the above branch It is a weighted sum of the normalized activation strength and the output of the consequent function. Then, the original output of the branch is... Convert to a probability distribution. The original output above is a vector, with each element corresponding to a category. The process of converting the original output to a probability distribution can be achieved using the softmax function: , in Indicates sample The probability of belonging to class j.

[0066] As can be seen, in this embodiment, the clustering transformation by using the hard C-means clustering model and introducing normalized activation intensity can enhance the model's ability to identify specific fault modes and its robustness.

[0067] In one possible embodiment, please refer to Figure 4 , Figure 4 This is a flowchart illustrating the third method for determining withstand voltage flashover ultrasonic signals provided in this application embodiment, as follows: Figure 4 As shown, the fuzzy clustering sub-model includes a fuzzy C-means clustering model, and determining the second branch result output by the fuzzy clustering sub-model based on the fuzzy clustering sub-model includes: S2321, determine multiple second activation intensities based on the first feature information, the second feature information, and the second membership function; S2322, determine the second consequent function based on the plurality of second activation intensities; S2323, determine the second branch result based on the second consequent function and the preset weight.

[0068] The second membership function is configured to characterize the degree of soft association between the input feature vector and each fuzzy cluster prototype (i.e., the fuzzy cluster center). For each fuzzy cluster, a second activation intensity is calculated, and the multiple second activation intensities together constitute a fuzzy membership vector, which characterizes the original, unnormalized membership strength of the input sample relative to each fuzzy cluster.

[0069] The second consequent function is configured as a function mapping, whose input is the plurality of second activation intensity vectors. Optionally, the second consequent function is a linear function that maps fuzzy membership information to an output with higher decision-making power.

[0070] Specifically, the output of the second consequent function of each fuzzy cluster is weighted and combined with its corresponding preset weights, for example by weighted summation or weighted averaging, to integrate the contributions of all fuzzy clusters and generate a comprehensive fuzzy discriminant vector. The fuzzy discriminant vector reflects the comprehensive probability assessment of various fault modes derived from fuzzy clustering and consequent inference of the input sample, and is the result of the second branch.

[0071] Specifically, the membership function obtained from fuzzy C-means clustering is used as the activation strength, eliminating the need for an additional activation function. The output layer is also trained using a linear function and a backpropagation function. The second branch result of the output of the fuzzy clustering sub-model can be determined based on the following process: First, the hidden layer calculates the activation intensity of the input data based on the following formula: ; in, Let m be the i-th cluster center determined by FCM clustering, and m be the fuzzification coefficient, controlling the degree of fuzziness in the clustering; activation intensity For the i-th hidden layer node pair input The membership degree is obtained by calculating the activation intensity. Then, similar to the hard C-means clustering model branch, each hidden layer node corresponds to a consequent function, which is also a linear consequent function: , This uses the original fuzzy C-means clustering model membership. Normalization is usually not performed because FCM itself already satisfies the following conditions: Among them, and These are the weight parameters to be learned. Then, the original output of the branch is obtained based on the following formula. Similarly, the original output is converted using the softmax function. Convert to probability distribution .

[0072] As can be seen, in this embodiment, the fuzzy C-means clustering model effectively handles the inherent uncertainty and fuzziness in the data. By using the second activation intensity, the association information between the input sample and all clusters is preserved, avoiding information loss in the boundary region of the feature space, thus describing the transition state between discharge modes more delicately.

[0073] In one possible implementation, other clustering algorithms such as K-means clustering can be used to replace HCM or FCM to construct different RBFNN branches, but the network parameters may need to be readjusted to ensure recognition performance.

[0074] In one possible embodiment, the step of performing multi-feature fusion based on the first branch result and the second branch result to determine the dual-branch prediction probability includes: determining a first loss function of the precise clustering sub-model based on the first branch result; determining a second loss function of the fuzzy clustering sub-model based on the second branch result; and performing multi-feature fusion based on the first loss function, the second loss function, and the influencing factor to determine the dual-branch prediction probability.

[0075] The first loss function is used to quantify the degree of deviation between the result of the first branch and the actual situation or the expected target. In one embodiment, the first loss function may be a cross-entropy loss function, a mean squared error loss function, or a custom loss function based on a distance metric. The second loss function is used to quantify the degree of deviation between the result of the second branch and the actual situation or the expected target. In one embodiment, the second loss function may be a loss function based on fuzzy set theory, such as fuzzy cross-entropy loss. Based on the values ​​of the first loss function, the second loss function, and a preset influence factor, multi-feature fusion is performed to determine the final bi-branch prediction probability.

[0076] Optionally, the fusion process specifically includes: using the first loss function value and the second loss function value as inverse indicators to evaluate the confidence or reliability of the precise clustering sub-model and the fuzzy clustering sub-model on the current input sample, respectively. That is, the lower the loss function value, the more reliable the prediction result of that branch. Using the influence factor, the reliability evaluation of the two branches is weighted and adjusted, and the original outputs of the two branches (i.e., the first branch result and the second branch result) are fused accordingly to generate the dual-branch prediction probability.

[0077] Specifically, the predicted probabilities of the two branches are aggregated using cross-entropy based on softmax: , in, For the first loss function, The second loss function is β, which is an influence factor. Introducing the influence factor β adjusts the contribution of different branches to the final classification, thereby enhancing the defect identification capability.

[0078] Optionally, the weights can be adaptively adjusted using adaptive weight fusion strategies, gating mechanism fusion strategies, or learner fusion strategies. Specifically, an adaptive weight fusion strategy can involve the following steps: determining the fusion weights based on the first loss function and the second loss function. and ,in, , ; λ is a temperature coefficient greater than 0, used to control the sensitivity of the loss function to weight allocation; the bi-branch prediction probability P final Determined by the following formula: P final =ω1×P e +ω2×P f ; Among them, P e and P f These are the standardized probability vectors of the first branch result and the second branch result, respectively.

[0079] As can be seen, in this embodiment, a loss function is introduced as a confidence evaluation index, which enables the fusion process to adaptively adjust according to the performance of each specific sample on the two branches.

[0080] In one possible embodiment, the training process of the pre-built clustering model can be as follows: Discharge data can be divided into training and testing subsets using five-fold cross-validation, followed by multiple rounds of testing, and the clustering model can be trained using an error backpropagation algorithm. The backpropagation algorithm calculates the network output through forward propagation and backpropagates the output error, iteratively optimizing the connection weights, biases, and radial basis function center and width parameters in the network according to the loss function. The goal of the optimization process is to minimize the loss function, which is typically configured as a mean squared error loss or a cross-entropy loss function. Through this optimization process, the difference between the network output and the expected output is continuously reduced. On each round of the test set, the classification rate is used as the core metric to evaluate the performance of the trained EFRBFNN network. The classification rate is defined as the percentage of samples correctly classified by the model on the test set out of the total number of samples in the test set. This metric intuitively reflects the model's accuracy in classifying unknown data. The original values ​​of the network output layer are transformed using the Softmax function into a normalized probability distribution vector. Each element of this vector represents the predicted probability that the input sample belongs to the corresponding fault category. The final classification result is determined by the category label corresponding to the highest probability value in the probability distribution vector.

[0081] As can be seen, the use of five-fold cross-validation in this embodiment ensures the statistical rigor of the model performance evaluation. This method ensures that each data sample is used for one test, greatly reducing the dependence of the evaluation results on the randomness of data partitioning.

[0082] S240, determine the type of withstand voltage flashover ultrasonic signal of the target network based on the bi-branch prediction probability.

[0083] This can be achieved by defining a predefined mapping relationship between the output dimension of the bi-branch prediction probability and specific withstand voltage flashover ultrasonic signal types. Each signal type corresponds to a specific internal insulation state or fault mode of the GIS (Gas Insulation System). For example: Type A signal corresponds to "normal operating condition no-discharge ultrasonic signal"; Type B signal corresponds to "surface flashover ultrasonic signal of solid insulation material"; Type C signal corresponds to "discharge ultrasonic signal caused by conductor burrs or scratches"; and Type D signal corresponds to "discharge ultrasonic signal of free metal particles".

[0084] Specifically, the final signal type is determined based on the dual-branch prediction probability vector. Specifically, the index corresponding to the maximum probability value in the probability vector is read, and according to the predefined mapping relationship, the corresponding withstand voltage flashover ultrasonic signal type is output as the final diagnostic result of the target network.

[0085] For an example, please refer to Figure 5 , Figure 5 This is a schematic diagram of the algorithm flow for a method for determining withstand voltage flashover ultrasonic signals provided in an embodiment of this application, as shown below. Figure 5As shown, three artificial defect models were constructed: flashover of solid insulation materials, flashover of conductor burrs or scratches, flashover of foreign object protrusions or metal particles, and a noise model. GIS flashover discharge data were collected in a laboratory environment using an ultrasonic sensor. To simulate field voltage phase conditions, the laboratory-detected flashover discharge data was converted into data with offset voltage phase as a discharge dataset to test the classifier's field applicability. PCA feature extraction was employed. Principal Component Analysis (PCA) was used to reduce the dimensionality of the input space and extract features, removing redundant information and retaining the main features to obtain the first feature information. This extracted feature was used as input to a branch of a pre-built clustering model (EFRBFNN). LDA feature extraction and statistical feature supplementation were performed. Linear Discriminant Analysis (LDA) was used to extract features with class discriminative power. Since the output dimension of LDA is limited by the number of classes, statistical features (including skewness, kurtosis, correlation, and asymmetry, etc.) were used as supplementary features to obtain the second feature information LDA+SC. This second feature information, along with the first feature information, was used as input. EFRBFNN consists of two branches based on fuzzy rules, employing Hard C-means (HCM) and Fuzzy C-means (FCM) clustering algorithms respectively. The HCM clustering RBFNN branch determines the center and width of a Gaussian RBF cluster using HCM clustering. Hidden layer node outputs are Gaussian function values, and the output layer uses a linear function, with connection weights trained using backpropagation (BP). The FCM clustering RBFNN branch uses the membership function obtained from FCM clustering as the activation intensity, requiring no additional activation function. The output layer also uses a linear function and is trained using BP. A multi-feature fusion strategy is designed to aggregate the predicted probabilities from the two branches to obtain an evaluation classification rate.

[0086] As can be seen, by acquiring a discharge dataset, which is discharge data from multiple experimental models with offset voltage phases, and these multiple experimental models are used to simulate the target network; determining first feature information based on the discharge dataset and a first model, and determining second feature information based on the discharge dataset and a second model, where the first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used for feature categorization; determining the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-constructed precise clustering sub-model and fuzzy clustering sub-model of the clustering model; and determining the withstand voltage flashover ultrasonic signal type of the target network based on the bi-branch prediction probability. Thus, high-precision and robust identification of multiple types of flashovers and noise in GIS is achieved, improving feature utilization, model generalization ability, and computational efficiency.

[0087] Please see Figure 6 , Figure 6This is a schematic diagram of a withstand voltage flashover ultrasonic signal determination device 600 according to an embodiment of this application. The withstand voltage flashover ultrasonic signal determination device 600 includes: a data determination module 610, a feature extraction module 620, a model analysis module 630, and a result determination module 640, wherein... The data determination module 610 is used to acquire a discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network. The feature extraction module 620 is used to determine first feature information based on the discharge dataset and the first model, and to determine second feature information based on the discharge dataset and the second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization. The model analysis module 630 is used to determine the bi-branch prediction probability based on the first feature information, the second feature information, the precise clustering sub-model and the fuzzy clustering sub-model of the pre-constructed clustering model; The result determination module 640 is used to determine the type of withstand flashover ultrasonic signal of the target network based on the bi-branch prediction probability.

[0088] In one possible embodiment, the first model includes a principal component analysis model, and the feature extraction module 620, in determining the first feature information based on the discharge dataset and the first model, is specifically used for: Determine the covariance matrix based on the discharge dataset; The eigenvectors and eigenvalues ​​are determined based on the covariance matrix; A first feature space is determined based on the feature vector and the feature value, and the first feature space is the first feature information.

[0089] In one possible embodiment, the discharge dataset includes multiple types of discharge data, which correspond to multiple types of fault conditions. The second model includes a linear discriminant analysis model. The feature extraction module 620, in determining the second feature information based on the discharge dataset and the second model, is specifically used for: A statistical feature space is determined based on the discharge dataset, and the statistical feature space is used to characterize the waveform features of the multiple types of discharge data. Based on the discharge dataset, an intra-class scatter matrix and an inter-class scatter matrix are determined. The intra-class scatter matrix represents the distribution of data points within the same class of discharge data, and the inter-class scatter matrix represents the distribution of data points within different classes of discharge data. The projection matrix is ​​determined based on the intra-class scatter matrix and the inter-class scatter matrix; The second feature space is determined based on the projection matrix; A third feature space is determined by feature fusion based on the second feature space and at least one statistical feature space, wherein the third feature space is the second feature information and the statistical feature space.

[0090] In one possible embodiment, the model analysis module 630, in determining the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-built clustering model's precise clustering sub-model and fuzzy clustering sub-model, is specifically used for: The first feature information and the second feature information are input into the precise clustering sub-model, and the first feature information and the second feature information are input into the fuzzy clustering sub-model; The first branch result of the output of the precise clustering sub-model is determined based on the precise clustering sub-model, and the second branch result of the output of the fuzzy clustering sub-model is determined based on the fuzzy clustering sub-model; The dual-branch prediction probability is determined by performing multi-feature fusion based on the results of the first branch and the second branch.

[0091] In one possible embodiment, the precise clustering sub-model includes a hard C-means clustering model, and the model analysis module 630 is specifically used for determining the first branch result output by the precise clustering sub-model based on the precise clustering sub-model: Multiple first activation intensities are determined based on the first feature information, the second feature information, and the first membership function; The normalized activation intensity is determined based on the plurality of first activation intensities; The first consequent function is determined based on the normalized activation intensity; The first branch result is determined based on the first consequent function, the normalized activation strength, and the preset weight.

[0092] In one possible embodiment, the fuzzy clustering sub-model includes a fuzzy C-means clustering model, and the model analysis module 630 is specifically used for determining the second branch result output by the fuzzy clustering sub-model based on the fuzzy clustering sub-model as follows: Multiple second activation intensities are determined based on the first feature information, the second feature information, and the second membership function; The second consequent function is determined based on the plurality of second activation intensities; The second branch result is determined based on the second consequent function and the preset weight.

[0093] In one possible embodiment, the model analysis module 630, in determining the dual-branch prediction probability by performing multi-feature fusion based on the first branch result and the second branch result, is specifically used for: Based on the results of the first branch, determine the first loss function of the precise clustering sub-model; The second loss function of the fuzzy clustering sub-model is determined based on the results of the second branch. Based on the first loss function, the second loss function, and the influence factor, multi-feature fusion is performed to determine the dual-branch prediction probability.

[0094] It is worth noting that the specific functional implementation of the withstand voltage flashover ultrasonic signal determination device is described above. Figures 2-5 The description of the withstand voltage flashover ultrasonic signal determination method illustrates that, for example, the data determination module 610 is used to implement the relevant content of S210, the feature extraction module 620 is used to implement the relevant content of S220, the model analysis module 630 is used to implement the relevant content of S230, and the result determination module 640 is used to implement the relevant content of S240. Each unit or module in the withstand voltage flashover ultrasonic signal determination device 600 can be individually or entirely merged into one or more other units or modules, or some of the units or modules can be further divided into multiple functionally smaller units or modules. This achieves the same operation without affecting the technical effect of the embodiments of the present invention. The above-mentioned units or modules are based on logical function division. In practical applications, the function of one unit (or module) is implemented by multiple units (or modules), or the function of multiple units (or modules) is implemented by one unit (or module).

[0095] In the case of using integrated units, please refer to Figure 7 , Figure 7 This is a schematic diagram of another withstand voltage flashover ultrasonic signal determination device provided in the embodiments of this application, as shown below. Figure 7 As shown, the withstand voltage flashover ultrasonic signal determination device 600 includes a processing module 602 and a communication module 601. The processing module 602 controls and manages the operation of the withstand voltage flashover ultrasonic signal determination device 600, for example, executing the steps of the data determination module 610, feature extraction module 620, model analysis module 630, and result determination module 640, and / or performing other processes of the technology described herein. The communication module 601 is used for interaction between the withstand voltage flashover ultrasonic signal determination device 600 and other devices. Figure 7 As shown, the withstand voltage flashover ultrasonic signal determination device 600 may further include a storage module 603, which is used to store the program code and data of the withstand voltage flashover ultrasonic signal determination device 600.

[0096] The processing module 602 can be a processor or controller, such as a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication module 601 can be a transceiver, RF circuitry, or a communication interface, etc. The storage module 603 can be a memory.

[0097] All relevant content in each scenario involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here. The above-mentioned withstand voltage flashover ultrasonic signal determining device 600 can perform the above-mentioned... Figure 2 The method for determining the withstand voltage flashover ultrasonic signal is shown.

[0098] As can be seen, the withstand voltage flashover ultrasonic signal determination device described in this application embodiment acquires a discharge dataset, which is discharge data from multiple experimental models with offset voltage phases. These multiple experimental models are used to simulate a target network. Based on the discharge dataset and a first model, it determines first feature information; and based on the discharge dataset and a second model, it determines second feature information. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used for feature categorization. Based on the first feature information, the second feature information, and pre-constructed precise clustering sub-models and fuzzy clustering sub-models of a clustering model, it determines a bi-branch prediction probability. Based on the bi-branch prediction probability, it determines the withstand voltage flashover ultrasonic signal type of the target network. Thus, it achieves high-precision and robust identification of multiple types of flashovers and noise in GIS, improving feature utilization, model generalization ability, and computational efficiency.

[0099] Please see Figure 8 , Figure 8This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application. As shown in the figure, the electronic device 800 includes a processor 810, a memory 820, a communication interface 830, and one or more programs 821. The one or more programs 821 are stored in the memory 820 and are configured to be executed by the processor 810.

[0100] The processor 810, memory 820, and communication interface 830 are interconnected and perform communication with each other. The memory 820 can be a volatile memory such as dynamic random access memory (DRAM) or a non-volatile memory such as a hard disk drive (HDD). The memory 820 stores a set of executable program code, and the processor 810 calls one or more programs 821 stored in the memory 820 to execute any part or all of the steps described in the embodiments of the withstand voltage flashover ultrasonic signal determination method.

[0101] Among them, electronic devices 800 may include smartphones (such as Android phones, iOS phones, Windows Phones, etc.), tablet computers, handheld computers, dashcams, in-vehicle electronic devices, servers, laptops, mobile internet electronic devices (MIDs) or wearable electronic devices (such as smartwatches, Bluetooth headsets), etc. The above are just examples and not an exhaustive list, including but not limited to the above electronic devices.

[0102] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes an electronic device.

[0103] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include an electronic device.

[0104] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0105] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0107] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0109] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer electronic device (which may be a personal computer, electronic device, or network electronic device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0110] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0111] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for determining withstand voltage flashover ultrasonic signals, characterized in that, include: Obtain a discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network; First feature information is determined based on the discharge dataset and the first model, and second feature information is determined based on the discharge dataset and the second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization. The bi-branch prediction probability is determined based on the first feature information, the second feature information, the precise clustering sub-model and the fuzzy clustering sub-model of the pre-constructed clustering model; The type of withstand voltage flashover ultrasonic signal of the target network is determined based on the bi-branch prediction probability.

2. The method according to claim 1, characterized in that, The first model includes a principal component analysis model, and the step of determining the first feature information based on the discharge dataset and the first model includes: Determine the covariance matrix based on the discharge dataset; The eigenvectors and eigenvalues ​​are determined based on the covariance matrix; A first feature space is determined based on the feature vector and the feature value, and the first feature space is the first feature information.

3. The method according to claim 1, characterized in that, The discharge dataset includes multiple types of discharge data, which correspond to multiple types of fault conditions. The second model includes a linear discriminant analysis model. Determining the second feature information based on the discharge dataset and the second model includes: A statistical feature space is determined based on the discharge dataset, and the statistical feature space is used to characterize the waveform features of the multiple types of discharge data. Based on the discharge dataset, an intra-class scatter matrix and an inter-class scatter matrix are determined. The intra-class scatter matrix represents the distribution of data points within the same class of discharge data, and the inter-class scatter matrix represents the distribution of data points within different classes of discharge data. The projection matrix is ​​determined based on the intra-class scatter matrix and the inter-class scatter matrix; The second feature space is determined based on the projection matrix; A third feature space is determined by feature fusion based on the second feature space and at least one statistical feature space, wherein the third feature space is the second feature information and the statistical feature space.

4. The method according to any one of claims 1-3, characterized in that, The determination of the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-constructed precise clustering sub-model and fuzzy clustering sub-model of the clustering model includes: The first feature information and the second feature information are input into the precise clustering sub-model, and the first feature information and the second feature information are input into the fuzzy clustering sub-model; The first branch result of the output of the precise clustering sub-model is determined based on the precise clustering sub-model, and the second branch result of the output of the fuzzy clustering sub-model is determined based on the fuzzy clustering sub-model; The dual-branch prediction probability is determined by performing multi-feature fusion based on the results of the first branch and the second branch.

5. The method according to claim 4, characterized in that, The precise clustering sub-model includes a hard C-means clustering model, and the step of determining the first branch result output by the precise clustering sub-model includes: Multiple first activation intensities are determined based on the first feature information, the second feature information, and the first membership function; The normalized activation intensity is determined based on the plurality of first activation intensities; The first consequent function is determined based on the normalized activation intensity; The first branch result is determined based on the first consequent function, the normalized activation strength, and the preset weight.

6. The method according to claim 4, characterized in that, The fuzzy clustering sub-model includes a fuzzy C-means clustering model, and the step of determining the second branch result output by the fuzzy clustering sub-model includes: Multiple second activation intensities are determined based on the first feature information, the second feature information, and the second membership function; The second consequent function is determined based on the plurality of second activation intensities; The second branch result is determined based on the second consequent function and the preset weight.

7. The method according to claim 4, characterized in that, The step of performing multi-feature fusion based on the first branch result and the second branch result to determine the dual-branch prediction probability includes: Based on the results of the first branch, determine the first loss function of the precise clustering sub-model; The second loss function of the fuzzy clustering sub-model is determined based on the results of the second branch. Based on the first loss function, the second loss function, and the influence factor, multi-feature fusion is performed to determine the dual-branch prediction probability.

8. A device for determining withstand voltage flashover ultrasonic signals, characterized in that, include: The data determination module is used to acquire a discharge dataset, which is discharge data of multiple experimental models with offset voltage phase, and the multiple experimental models are used to simulate the target network; The feature extraction module is used to determine first feature information based on the discharge dataset and the first model, and to determine second feature information based on the discharge dataset and the second model. The first model is an unsupervised dimensionality reduction feature extraction model used to analyze the principal components of the discharge dataset, and the second model is a supervised dimensionality reduction feature extraction model used to perform feature categorization. The model analysis module is used to determine the bi-branch prediction probability based on the first feature information, the second feature information, and the pre-built precise clustering sub-model and fuzzy clustering sub-model of the clustering model; The result determination module is used to determine the type of withstand voltage flashover ultrasonic signal of the target network based on the bi-branch prediction probability.

9. A computer-readable storage medium, characterized in that, The device stores a withstand voltage flashover ultrasonic signal determination program, including execution instructions, which, when executed by a processor of an electronic device, perform the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes a processor, memory, a communication interface, and one or more programs, which are stored in the memory and configured to be executed by the processor; When the processor executes the one or more programs stored in the memory, the processor performs the method as described in any one of claims 1 to 7.

Images