Switch cabinet discharge detection method, system and equipment based on typical characteristics and medium

Abstract

The invention relates to the technical field of discharge point detection, and discloses a typical feature-based switch cabinet discharge detection method, system and device and a medium, and the method comprises the steps: collecting a multi-source signal and an environment parameter of a switch cabinet in real time, and carrying out the first processing; extracting a signal nonlinear feature of the first processed data, obtaining a frequency domain energy feature through a first transformation operation, and constructing a multi-dimensional feature vector through a first fusion analysis operation; constructing an adversarial network model, carrying out adversarial training on the multi-dimensional feature vector and a historical real partial discharge signal, and generating a high-fidelity partial discharge analog signal; classifying the discharge types by using a first classification algorithm, and constructing a fault recognition model in combination with the multi-dimensional feature vector to recognize the discharge fault probability of the switch cabinet; and learning the time sequence characteristics by using the risk prediction model, predicting the discharge risk of the switch cabinet, and generating a detection report. According to the invention, multi-dimensional accurate characterization of the discharge state of the switch cabinet is realized.

Description

Technical Field

[0001] This invention relates to the field of discharge detection technology, and in particular to a method, system, device and medium for discharge detection of switchgear based on typical characteristics. Background Technology

[0002] As a key power device in the smart grid, the real-time monitoring of the insulation status of switchgear is crucial for ensuring the safe and stable operation of the power system. Partial discharge (PD) is one of the important signs reflecting insulation degradation in switchgear, thus partial discharge detection technology has become a core means of condition monitoring. With the advancement of sensing technology and data analysis methods, partial discharge detection has gradually evolved from early single-signal detection to multi-source information fusion detection, significantly improving the comprehensiveness and reliability of detection.

[0003] In recent years, deep learning technology has shown significant advantages in discharge signal analysis and state prediction. For example, generative adversarial networks (GANs) can solve the problem of insufficient partial discharge samples through data augmentation, while long short-term memory networks have performed well in discharge trend prediction due to their excellent time-series modeling capabilities. However, existing detection methods still have obvious limitations: (1) feature extraction is limited in dimension. Traditional methods mostly rely on time-domain statistical features or simple frequency-domain transformations, failing to fully explore the nonlinear dynamic features of the signal and the multi-scale energy distribution in the frequency domain, resulting in insufficient ability to distinguish complex discharge types; (2) environmental interference suppression is insufficient. Environmental parameters such as temperature, humidity, and mechanical vibration at the switchgear site are easily coupled with discharge signals, but existing methods lack effective fusion of these multi-modal data, reducing the robustness of detection; (3) model generalization ability is limited. A single detection algorithm is difficult to adapt to changes in discharge modes under different working conditions, and its ability to process small samples or noisy data is weak.

[0004] Therefore, how to comprehensively utilize the multi-dimensional characteristics of partial discharge signals and adaptively fuse them with environmental parameters to improve detection accuracy and anti-interference capability has become a key technical challenge that urgently needs to be overcome in the field of switchgear insulation condition monitoring. Against this backdrop, researching switchgear discharge detection methods based on typical characteristics is of great significance for promoting condition-based maintenance and fault early warning of smart grid equipment. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for detecting partial discharge in switchgear based on typical features, which solves the problem of insufficient accuracy and distinguishing ability in partial discharge diagnosis caused by single-dimensional feature description.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a switchgear discharge detection method based on typical characteristics, comprising:

[0009] Real-time acquisition of multi-source signals and environmental parameters from the switchgear, followed by initial processing;

[0010] Extract the signal nonlinear features of the first processed data, obtain the frequency domain energy features through the first transformation operation, and construct a multidimensional feature vector through the first fusion analysis operation;

[0011] An adversarial network model is constructed, and the multidimensional feature vectors are trained against historical real partial discharge signals to generate high-fidelity partial discharge simulation signals.

[0012] Based on the high-fidelity partial discharge simulation signal, the discharge type is classified using the first classification algorithm, and a fault identification model is constructed by combining the multi-dimensional feature vector to identify the probability of discharge faults in the switchgear.

[0013] Based on the discharge fault probability of the switchgear, the risk of discharge of the switchgear is predicted by learning time series features using a risk prediction model, and a detection report is generated.

[0014] As a preferred embodiment of the switchgear discharge detection method based on typical features described in this invention, the construction of the multidimensional feature vector includes:

[0015] The phase space of the first processed data is reconstructed according to the first embedding theorem, and the nonlinear characteristics of the multi-source signal are quantified.

[0016] The frequency domain energy characteristics of the multi-source signal are obtained by performing a first transformation operation.

[0017] The nonlinear features, frequency domain energy features, and the first processed environmental parameters are optimized and combined using the first fusion analysis operation to construct a multidimensional feature vector.

[0018] As a preferred embodiment of the switchgear discharge detection method based on typical features described in this invention, the step of classifying the discharge type using a first classification algorithm includes:

[0019] Based on the high-fidelity partial discharge simulation signal, signal features are extracted from the first dimension, the second dimension, and the third dimension respectively to obtain multi-dimensional features;

[0020] The multi-dimensional features are processed by dynamic weight fusion combined with the first dimensionality reduction operation to generate a signal feature vector;

[0021] The signal feature vector is classified using a first classification algorithm to obtain a discharge type label.

[0022] As a preferred embodiment of the switchgear discharge detection method based on typical features described in this invention, the method for identifying the probability of switchgear discharge faults includes:

[0023] Based on logistic regression, the probability of switchgear discharge failure is calculated by combining the discharge type label with the linear combination of the multidimensional feature vectors.

[0024] As a preferred embodiment of the switchgear discharge detection method based on typical characteristics described in this invention, the real-time acquisition of multi-source signals and environmental parameters of the switchgear, and the subsequent first processing, includes:

[0025] The switchgear multi-source signals include ultrasonic signals, ultra-high frequency signals, transient ground voltage, and high frequency current signals; the environmental parameters include temperature data, humidity data, air pressure data, vibration data, and dust concentration.

[0026] The first processing includes multi-source signal denoising, spatiotemporal alignment, environmental parameter correction, and normalization processing.

[0027] As a preferred embodiment of the switchgear discharge detection method based on typical features described in this invention, the generation of high-fidelity partial discharge simulation signals includes:

[0028] Based on the GAN multi-layer neural network architecture, an adversarial network model containing an input layer, a generator, and a discriminator is constructed.

[0029] The generator extracts features and transforms the dimensions of the input multidimensional feature vector through a fully connected layer, and gradually upsamples and reconstructs it through a deconvolution layer to generate a preliminary high-fidelity partial discharge simulation signal.

[0030] The discriminator extracts the time-frequency features of the real discharge signal and the preliminary high-fidelity partial discharge simulation signal through a multi-layer convolutional neural network, and performs binary classification to distinguish the authenticity of the signal based on the time-frequency features.

[0031] By iteratively executing the gradient backpropagation algorithm, the generator and the discriminator are alternately optimized to generate a high-fidelity partial discharge simulation signal.

[0032] The beneficial effects of this preferred technical solution are: by generating high-fidelity partial discharge simulation signals through an adversarial network model, the diversity and realism of the training samples are effectively expanded.

[0033] As a preferred embodiment of the switchgear discharge detection method based on typical features described in this invention, the step of using a risk prediction model to learn time series features and predict the discharge risk of the switchgear includes:

[0034] The construction of the risk prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer;

[0035] The LSTM layer receives the historical switchgear discharge fault probability of the input layer, selectively memorizes and updates the temporal characteristics of the historical fault probability and environmental parameters through a gating mechanism, and generates the hidden state of the current time step;

[0036] The fully connected layer receives the hidden state of the LSTM layer, reconstructs the feature space through the linear transformation of the weight matrix, combines the nonlinear activation function to construct a decision boundary, and generates a high-order feature representation characterizing the fault risk;

[0037] The output layer performs a linear transformation on the high-order feature representation generated by the fully connected layer, and maps it to the switchgear discharge risk probability R of the future time step through the Sigmoid activation function t+Δt ;

[0038] Based on historical data, define a low-risk threshold A1 and a high-risk threshold A2;

[0039] When R t+Δt < A1, it is considered that the switchgear discharge activity is in a normal state;

[0040] When A1 ≤ R t+Δt < A2, it is considered that the switchgear discharge activity exceeds the normal range and needs to be repaired in time;

[0041] When R t+Δt ≥ A2, it is considered that the switchgear discharge activity has serious insulation defects, and power should be cut off immediately and an insulation diagnostic test should be carried out.

[0042] The beneficial effect of this preferred technical solution is: predicting the future risk level, outputting a detection report including specific risk threshold judgments, and providing a comprehensive and reliable decision-making basis for the switchgear state assessment.

[0043] In the second aspect, the present invention provides a switchgear discharge detection system based on typical features, including:

[0044] A data acquisition module for real-time acquisition of multi-source signals and environmental parameters of the switchgear and performing a first processing;

[0045] A feature extraction module for extracting the signal non-linear features of the data after the first processing, obtaining the frequency-domain energy features through a first transformation operation, and constructing a multi-dimensional feature vector through a first fusion analysis operation;

[0046] A discharge signal generation module for constructing an adversarial network model, performing adversarial training on the multi-dimensional feature vector and historical real partial discharge signals, and generating high-fidelity local discharge simulation signals;

[0047] The fault identification module is used to classify the discharge type based on the high-fidelity partial discharge simulation signal using the first classification algorithm, and to construct a fault identification model by combining the multi-dimensional feature vector to identify the probability of discharge faults in the switchgear.

[0048] The risk prediction module is used to predict the discharge risk of the switchgear based on the probability of discharge failure of the switchgear, and to generate a detection report by learning time series features using a risk prediction model.

[0049] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of a switchgear discharge detection method based on typical characteristics.

[0050] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a switchgear discharge detection method based on typical characteristics.

[0051] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method, system, device, and medium for switchgear discharge detection based on typical features. It extracts nonlinear features and frequency domain energy features of signals through chaos theory combined with short-time Fourier transform, and constructs multi-dimensional feature vectors using a dynamic weight fusion strategy, achieving a multi-dimensional and accurate characterization of the switchgear discharge state. Furthermore, it generates high-fidelity partial discharge simulation signals through an adversarial network model, effectively expanding the diversity and realism of the training samples. Based on the multi-dimensional feature vectors and discharge simulation signals, it can accurately identify discharge types and calculate fault probabilities. Finally, it predicts future risk levels through LSTM time-series modeling and outputs a detection report containing specific risk threshold judgments, providing a comprehensive and reliable decision-making basis for switchgear condition assessment. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the overall process logic of a switchgear discharge detection method based on typical features according to an embodiment of the present invention.

[0054] Figure 2 This is a flowchart illustrating the discharge signal generation model structure of a switchgear discharge detection method based on typical features, as described in one embodiment of the present invention.

[0055] Figure 3 This is a flowchart illustrating the risk prediction and report generation process of a switchgear discharge detection method based on typical features, as described in one embodiment of the present invention.

[0056] Figure 4 This is a flowchart of a switchgear discharge detection system based on typical features according to an embodiment of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0058] Example 1, referring to Figure 1 As one embodiment of the present invention, a switchgear discharge detection method based on typical characteristics is provided, such as... Figure 1 The specific steps shown are as follows:

[0059] S100: Real-time acquisition of multi-source signals and environmental parameters from the switchgear, and performs initial processing;

[0060] S200: Extract the signal nonlinear characteristics of the first processed data, obtain the frequency domain energy characteristics through the first transformation operation, and construct a multi-dimensional feature vector through the first fusion analysis operation;

[0061] S300: Construct an adversarial network model, train multi-dimensional feature vectors against historical real partial discharge signals, and generate high-fidelity partial discharge simulation signals.

[0062] S400: Based on high-fidelity partial discharge simulation signals, the first classification algorithm is used to classify the discharge type, and a fault identification model is constructed by combining multi-dimensional feature vectors to identify the probability of discharge faults in the switchgear.

[0063] S500: Based on the probability of switchgear discharge faults, it uses a risk prediction model to learn time series features, predicts the discharge risk of switchgear, and generates a detection report.

[0064] It should be noted that, to address the issue of insufficient accuracy and discriminative power in partial discharge diagnosis caused by single-dimensional feature descriptions, steps S100–S500 extract nonlinear and frequency-domain energy features of the signal using chaos theory combined with short-time Fourier transform. A dynamic weight fusion strategy is then employed to construct a multi-dimensional feature vector, achieving a multi-dimensional and accurate representation of the switchgear's discharge state. Furthermore, a high-fidelity partial discharge simulation signal is generated using an adversarial network model, effectively expanding the diversity and realism of the training samples. Based on the multi-dimensional feature vector and the discharge simulation signal, the discharge type can be accurately identified and the fault probability calculated. Finally, LSTM time-series modeling is used to predict future risk levels, outputting a detection report containing specific risk threshold judgments, providing a comprehensive and reliable decision-making basis for switchgear condition assessment.

[0065] Example 2, refer to Figure 2 and Figure 3 Based on the previous embodiment, this embodiment provides a specific implementation method for the switchgear discharge detection method based on typical characteristics, in order to illustrate the technical means used in this method.

[0066] In this embodiment of the application, the above step S100, which involves real-time acquisition of multi-source signals and environmental parameters from the switchgear and the first processing, includes the following sub-steps A1 and A2:

[0067] In A1: Real-time acquisition of multi-source signals and environmental parameters from the switchgear;

[0068] Specifically, the switchgear's multi-source signals include ultrasonic signals, ultra-high frequency signals, transient ground voltage, and high-frequency current signals, while environmental parameters include temperature data, humidity data, air pressure data, vibration data, and dust concentration.

[0069] In A2: The collected data undergoes a first processing step, which includes multi-source signal denoising, spatiotemporal alignment, environmental parameter correction, and normalization.

[0070] Specifically, in the multi-source signal denoising process, for ultrasonic signals, based on their mechanical vibration characteristics, wavelet basis decomposition is performed and an improved SURE thresholding algorithm is used to eliminate mechanical vibration noise while retaining effective signal components; for ultra-high frequency signals, based on their electromagnetic interference characteristics, a variable step-size LMS adaptive filter is used to dynamically track and eliminate narrowband interference in the communication frequency band; for transient ground voltage, based on its pulse characteristics, a sliding time window moving average filter is used to suppress random pulse interference; for high-frequency current signals, based on the power frequency harmonic characteristics, a Kalman filter based on a state-space model is used to correct sensor baseline drift in real time.

[0071] Specifically, during the spatiotemporal alignment process, a unified time reference for multi-source signals is achieved through precision clock synchronization technology. A master-slave network architecture is established using a standard time synchronization protocol to ensure strict alignment of the acquisition times of each channel. Simultaneously, a three-dimensional coordinate system is constructed using spatial positioning algorithms to complete the spatial position calibration of multiple sensors. Furthermore, a dynamic compensation mechanism is used to correct for time delay changes caused by environmental factors in real time, and an automatic diagnostic function is equipped to continuously monitor the alignment status.

[0072] Specifically, during the environmental parameter correction process, a multi-parameter coupled compensation model is established to systematically correct environmental interference. Specifically, firstly, a temperature compensation module establishes a dynamic compensation curve based on the nonlinear relationship between temperature and signal amplitude; secondly, a humidity correction unit compensates for attenuation based on the impact of relative humidity changes on signal propagation; simultaneously, an air pressure adjustment mechanism optimizes the signal propagation model parameters in real time based on atmospheric pressure changes; furthermore, a vibration suppression algorithm eliminates interference with signal acquisition based on the spectral characteristics of mechanical vibrations; finally, a dust compensation system dynamically corrects for signal scattering effects based on changes in particulate matter concentration. This processing employs an adaptive feedback mechanism, enabling real-time tracking of environmental parameter changes and automatic adjustment of compensation parameters to ensure effective suppression of the impact of various environmental factors on signal acquisition.

[0073] Specifically, the normalization process achieves comparability conversion of multi-source data by establishing a unified standardization system. This includes: signal amplitude normalization processing; employing piecewise linear conversion methods based on the dynamic response characteristics of different signals such as ultrasound and UHF; utilizing statistical distribution-based conversion algorithms based on the physical characteristics of parameters such as temperature and humidity; and establishing an automatically adjusting standardized conversion model based on the overall measurement range through unified processing of multi-source data benchmarks. This process employs an intelligent dynamic adjustment mechanism that automatically optimizes conversion parameters based on the real-time characteristics of the input data, ensuring that all processed signals and parameters have consistent dimensions and comparability.

[0074] Specifically, for ultrasonic signals, amplitude fluctuations are eliminated through dynamic range compression algorithms; for ultra-high frequency signals, energy equalization is achieved through logarithmic transformation; for transient ground voltage signals, DC offset is eliminated through baseline correction algorithms; and for high-frequency current signals, dimensions are unified through proportional scaling. Simultaneously, environmental parameter standardization is applied to monitoring data such as temperature and humidity using Z-score standardization to eliminate dimensional differences. This processing employs an intelligent threshold adaptive adjustment mechanism, which dynamically optimizes transformation parameters based on real-time data characteristics, ensuring that the output data has consistent standardized features.

[0075] In an optional embodiment, the first processing step can also involve using a deep learning model such as a convolutional autoencoder or U-Net to directly perform end-to-end noise suppression and feature enhancement on the original multi-source signal. This approach trains the network with a large amount of real data to automatically learn the distribution differences between signals and noise, eliminating the need for manually designed specific filters. It can simultaneously process multiple signals such as ultrasound, UHF, and transient ground voltage, making it particularly suitable for adaptive denoising in complex noisy environments. For example, the CAE encoder can extract the common time-frequency features of the signal, while the decoder reconstructs the denoised clean signal, and adversarial training can further improve fidelity. In addition, an attention mechanism can be embedded to dynamically weight signal components in different frequency bands, enhancing the ability to preserve key features.

[0076] In another optional embodiment, the first processing step can also involve distributed signal optimization using a federated learning framework. This involves initial processing of sensor data locally on edge computing nodes before uploading it to a central server for fusion. This scheme utilizes graph neural networks to model the spatiotemporal correlations between multiple sensors. For example, it constructs a graph structure showing the spatial location of ultrasonic sensors and the propagation path of UHF signals, achieving cross-modal information complementarity through message passing mechanisms. Simultaneously, federated learning protects data privacy, allowing switchgear data from different substations to collaboratively train a global model, improving preprocessing generalization.

[0077] In this embodiment of the application, the above step S200, which extracts the signal nonlinear characteristics of the first processed data, obtains the frequency domain energy characteristics through the first transformation operation, and constructs a multidimensional feature vector through the first fusion analysis operation, includes the following sub-steps B1 to B3:

[0078] In B1: The phase space of the first processed data is reconstructed according to the first embedding theorem, and the nonlinear characteristics of the multi-source signal are quantified;

[0079] Specifically, the phase space of the preprocessed multi-source signal is reconstructed using the Takens embedding theorem, also known as the first embedding theorem. First, the optimal delay parameter and embedding dimension are determined, and the time-domain signal is mapped to a geometric trajectory in a high-dimensional phase space. Then, the Lyapunov exponent is calculated, and the chaotic characteristics are characterized by the exponential divergence rate of adjacent orbits. At the same time, the fractal dimension is estimated to evaluate the complexity of the phase space attractor from a geometric perspective.

[0080] In an optional embodiment, the first embedding theorem can also be phase space reconstruction based on singular spectral analysis, which decomposes the signal trajectory matrix through singular value decomposition and extracts the dominant modes to reconstruct the phase space. This method does not require preset delay parameters, but automatically separates noise and effective components using data-driven eigenorthogonal bases, making it particularly suitable for non-stationary signals.

[0081] In another alternative embodiment, the first embedding theorem can also be based on recursive graphs and recursive quantitative analysis. The time-domain signal is converted into a recursive graph, the recursive characteristics of the phase space orbit are visualized through thresholding, and the nonlinear dynamic characteristics are directly quantified using the RQA index. This avoids the parameter selection problem of traditional phase space reconstruction and is suitable for short data sequence analysis.

[0082] In B2: Frequency domain analysis is performed on the multi-source signal through the first transform operation to obtain the frequency domain energy characteristics; the specific steps include:

[0083] Based on the Hanning window function and adaptive window length selection algorithm, multi-source signals are converted into time-frequency distribution maps;

[0084] The energy distribution parameters of each characteristic frequency band are calculated through spectral energy analysis.

[0085] By integrating time-varying characteristics, an energy distribution feature including indicators such as instantaneous frequency and spectral centroid is constructed.

[0086] In an optional embodiment, the first transformation operation can also be to adaptively decompose the signal into several intrinsic mode functions (IMFs) using a VMD algorithm, with each IMF corresponding to a narrowband component with a defined center frequency. For example, the number of IMFs can be optimized using a kurtosis criterion, the Hilbert energy spectrum of each component can be extracted, and a multi-scale energy feature vector can be constructed based on the frequency band energy ratio, while preserving the nonlinear coupling characteristics between modes.

[0087] In another alternative embodiment, the first transformation operation can also employ high-resolution WVD time-frequency analysis, using an adaptive kernel function to suppress cross-term interference. For example, by combining a smooth pseudo-Wigner-Ville distribution to extract time-frequency ridges, calculating instantaneous spectral entropy and marginal spectral energy moments, an interference-resistant joint time-frequency feature set can be constructed, suitable for fine analysis of pulsed discharge signals.

[0088] In B3: The first fusion analysis operation is used to optimize and combine nonlinear features, frequency domain energy features, and the first-processed environmental parameters to construct a multi-dimensional feature vector; the specific steps include:

[0089] Based on feature importance and correlation analysis, a dynamic weight allocation model is established to weight and fuse nonlinear features, energy features and environmental parameters.

[0090] Principal component analysis is used to extract key principal components based on the contribution rate of feature variance. The fused high-dimensional features are then projected onto a low-dimensional orthogonal space to construct a multi-dimensional feature vector.

[0091] It should be noted that, through an adaptive weight adjustment mechanism and orthogonal transformation technology, an optimized feature vector was constructed that retains the core features of the original data while reducing dimensionality, which significantly improves the recognition accuracy and computational efficiency of the subsequent fault diagnosis model.

[0092] In an optional embodiment, the first fusion analysis operation can also be achieved by automatically learning the correlation weights between different features through a multi-head attention network, thereby realizing the intelligent fusion of nonlinear features, frequency domain features, and environmental parameters.

[0093] In another alternative embodiment, the first fusion analysis operation can also be to construct a heterogeneous graph structure from multi-source features, where nodes represent various types of features and edges represent physical or statistical relationships between features. Node information is aggregated through a graph attention network to achieve higher-order interactions while preserving feature topological relationships.

[0094] It should be noted that the above step S200 significantly improves the characterization capability of the features, enabling more accurate differentiation between weak discharge signals and mixed modes under complex interference environments, providing richer discrimination criteria for subsequent discharge type identification and fault probability calculation.

[0095] In this embodiment, step S300, which constructs an adversarial network model and performs adversarial training on the multidimensional feature vectors and historical real partial discharge signals to generate high-fidelity partial discharge simulation signals, includes the following sub-steps C1 to C4:

[0096] In C1: Based on the GAN multi-layer neural network architecture, an adversarial network model is constructed, which includes an input layer, a generator, and a discriminator.

[0097] Specifically, the network topology is designed based on the dimensionality characteristics of the multidimensional feature vectors: the input layer uses a fully connected network to receive the dimensionality-reduced multidimensional feature vectors; the generator consists of a 5-layer deconvolutional neural network, with each layer configured with a ReLU activation function and batch normalization processing, progressively upsampling the latent space vectors into complete time-frequency signals; the discriminator uses a 4-layer convolutional neural network architecture, achieving signal authenticity discrimination through stride convolutional layers and spectral normalization processing. This adversarial network model, through adversarial training between the generator and the discriminator, ultimately achieves a nonlinear mapping from the feature space to the signal space, generating physically realistic partial discharge simulation signals.

[0098] In C2: The generator extracts features and transforms the dimensions of the input multidimensional feature vector through a fully connected layer, and then gradually upsamples and reconstructs it through a deconvolutional layer to generate a preliminary high-fidelity partial discharge simulation signal.

[0099] Specifically, the generator extracts features and transforms the dimensions of the input multidimensional feature vector through a fully connected layer, and establishes a nonlinear mapping from the low-dimensional feature space to the high-dimensional latent space based on the statistical properties of the feature vector. It then progressively upsamples and reconstructs the signal through deconvolution layers. Based on the time-frequency characteristics of the partial discharge signal, it uses cascaded deconvolution operations to transform abstract features into specific waveforms. Each deconvolution layer is followed by batch normalization and LeakyReLU activation function to enhance feature representation. Finally, the generator outputs a preliminary high-fidelity partial discharge simulation signal that conforms to the dynamic range of the real signal through the activation function.

[0100] It should be noted that step C2 achieves end-to-end generation from abstract features to specific signals through a deep neural network. The generated simulated signals are highly consistent with real partial discharge signals in terms of time-domain waveform, frequency-domain energy distribution, and nonlinear characteristics.

[0101] In C3: The discriminator extracts the time-frequency features of the real discharge signal and the preliminary high-fidelity partial discharge simulation signal through a multi-layer convolutional neural network, and performs binary classification to distinguish the authenticity of the signal based on the time-frequency features;

[0102] Specifically, the deep convolutional neural network consists of 5 convolutional layers. Each convolutional layer is followed by a ReLU activation function and Dropout regularization, specifically designed to analyze the distribution characteristics of signals in the joint time-frequency domain. Based on the extracted time-frequency features, the discriminator constructs a discriminant function containing a Sigmoid activation function, outputting a probability assessment of the signal's authenticity.

[0103] In C4: High-fidelity partial discharge simulation signals are generated by iteratively executing the gradient backpropagation algorithm, alternately optimizing the generator and discriminator.

[0104] Specifically, the difference in distribution between the generated signal and the real signal is calculated based on the loss function constructed using the Wasserstein distance; through an adversarial training mechanism, the network parameters of the generator are dynamically adjusted using the Adam optimizer based on the feedback information from the discriminator, so that the generated signal approximates the distribution of the real signal in dimensions such as time-frequency characteristics and nonlinear dynamic characteristics; through a Nash equilibrium optimization strategy, the model converges based on the game process between the generator and the discriminator, and finally outputs a high-fidelity partial discharge simulation signal with physical realism.

[0105] It should be noted that step S300 effectively alleviates the problems of insufficient sample size and uneven distribution in actual engineering. By generating diverse discharge signal data, not only is the coverage of training samples expanded, but the model's ability to identify rare discharge patterns is also improved, thereby enhancing the generalization and robustness of the entire detection system and enabling it to maintain high diagnostic accuracy under complex working conditions.

[0106] In this embodiment, step S400, based on a high-fidelity partial discharge simulation signal, classifies the discharge type using a first classification algorithm and constructs a fault identification model by combining multi-dimensional feature vectors, identifying the probability of switchgear discharge faults, includes the following sub-steps D1 to D4:

[0107] In D1: Based on the high-fidelity partial discharge simulation signal, signal features are extracted from the first dimension, the second dimension, and the third dimension respectively to obtain multi-dimensional features;

[0108] In this embodiment, the first dimension is to extract the energy spectrum characteristics of the discharge simulation signal through frequency domain FFT analysis, and calculate key frequency domain indicators such as center frequency and bandwidth based on the peak distribution of the spectrum and the energy ratio of the frequency band.

[0109] In this embodiment, the second dimension involves analyzing the time-varying characteristics of the discharge simulation signal through time-frequency wavelet transform, constructing a time-frequency joint distribution based on wavelet basis functions, and extracting the statistical features of the time-frequency matrix and the instantaneous frequency change law.

[0110] In this embodiment of the application, the third dimension is to quantify the complexity and randomness of the signal by nonlinear entropy feature analysis based on indicators such as sample entropy and permutation entropy.

[0111] It should be noted that the analysis processes of these three dimensions are independent of each other but complementary, and together they generate a multi-dimensional feature set that includes frequency domain features, time-frequency features and nonlinear features.

[0112] In an alternative embodiment, signal feature extraction can also be achieved by designing a hybrid architecture of a one-dimensional convolutional neural network and a long short-term memory network, which automatically learns multi-level feature representations directly from the original signal. The 1D-CNN layer is responsible for extracting local temporal and frequency domain features, while the LSTM layer captures the long-term temporal dependencies of the signal. Finally, the importance of each feature dimension is dynamically weighted through an attention mechanism.

[0113] In another alternative embodiment, signal feature extraction can also be achieved by combining prior physical knowledge with a data-driven approach: first, a feature template library is constructed based on the physical mechanism of discharge; then, these physical constraints are embedded into the feature extraction process of the neural network using differentiable programming techniques. Simultaneously, a graph neural network is used to model the physical relationships between different feature nodes, forming a hybrid feature representation that combines physical interpretability and data adaptability.

[0114] In D2: Multi-dimensional features are processed by dynamic weight fusion combined with the first dimensionality reduction operation to generate signal feature vectors;

[0115] In this embodiment, the first dimensionality reduction operation is achieved through t-SNE nonlinear dimensionality reduction. Through a dynamic weighted fusion mechanism, multi-dimensional features are optimized and combined according to the energy distribution characteristics of frequency domain features, the transient change law of time-frequency features, and the dynamic behavior of nonlinear features to obtain the fused feature representation. Subsequently, through the manifold learning method, the high-dimensional features are nonlinearly reduced according to the principle of preserving the topology of the feature space to obtain a discriminative low-dimensional feature vector.

[0116] In an optional embodiment, the first dimensionality reduction operation can also construct a deep generative model through a variational autoencoder, which achieves nonlinear feature compression in the encoder part while using KL divergence to constrain the probability distribution of the latent space.

[0117] In another alternative embodiment, the first dimensionality reduction operation can also employ a unified manifold approximation and projection algorithm to optimize the local and global structure preservation of the feature space using Riemannian geometry methods, and dynamically adjust the contribution weights of different feature dimensions by combining an attention mechanism.

[0118] In D3: The first classification algorithm is used to classify the signal feature vector to obtain the discharge type label; the specific steps include:

[0119] A high-dimensional feature space mapping is constructed using radial basis kernel functions, and the optimal penalty parameters and kernel parameters are determined through grid search and cross-validation.

[0120] A multi-class SVM model is constructed based on a one-to-one strategy, and binary classifiers are trained for each discharge type.

[0121] The probability estimation algorithm outputs the probability distribution of each sample belonging to different discharge types, and the category corresponding to the highest probability is taken as the final discharge type label.

[0122] In an optional embodiment, the first classification algorithm can also be an end-to-end classification based on a deep residual network, designing a 1D-CNN architecture with residual connections, automatically extracting multi-level features by stacking residual blocks, and introducing an attention mechanism at the end of the network to strengthen key features.

[0123] In another alternative embodiment, the first classification algorithm can also be a topology-aware classification based on graph neural networks, constructing the signal feature vectors as graph-structured data, where nodes represent feature dimensions and edges represent the correlation strength between features. Neighborhood information is aggregated through a graph attention network, and a multi-head attention mechanism is used to learn the nonlinear interaction patterns between features.

[0124] In D4: Based on logistic regression, the probability of switchgear discharge faults is calculated by combining a linear combination of multidimensional feature vectors with discharge type labels. The formula is expressed as:

[0125]

[0126] Where y is the discharge type label, X is the multidimensional feature vector, W1 is the feature transformation weight matrix, W2 is the output layer weight vector, b1 is the feature transformation layer bias term, b2 is the output layer bias term, λ is the regularization coefficient, w is the regularization term, and P(y=1|X) is the discharge fault probability of the switchgear.

[0127] The specific operating steps include:

[0128] The input multidimensional feature vector is subjected to a two-stage linear transformation. The first-stage transformation maps the original feature space to the latent feature space through the weight matrix W2 and the bias term b1. The second-stage transformation converts the latent features into linear scores through the weight vector W2 and the bias term b2.

[0129] The linear score is converted into a probability value by using the Sigmoid function. The non-linear nature of the Sigmoid function ensures that the output value is strictly limited to between 0 and 1, which meets the definition of probability.

[0130] Through regularization term λ‖w‖ 2 Constraints are imposed on the weight parameters;

[0131] The output switchgear discharge fault probability value P(y=1|X) intuitively reflects the possibility of the switchgear experiencing a discharge fault under given characteristic conditions.

[0132] It should be noted that step S400 significantly improves the accuracy of discharge mode classification by integrating multi-dimensional features and simulated data, especially demonstrating stronger distinguishing ability for discharge types with similar characteristics. Simultaneously, the quantitative output of fault probability provides maintenance personnel with an intuitive basis for risk assessment, making maintenance decisions more scientific and refined.

[0133] In this embodiment of the application, step S500 above, based on the probability of switchgear discharge faults, uses a risk prediction model to learn time series features to predict the discharge risk of the switchgear and generates a detection report, including the following sub-steps E1 to E2:

[0134] In E1: Based on the probability of switchgear discharge failure, the risk prediction model is used to learn time series features to predict the discharge risk of switchgear.

[0135] In this embodiment of the application, the risk prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer;

[0136] It should be noted that, based on the time-series characteristics of the discharge fault probability of the switchgear, a deep time-series prediction architecture is designed: the input layer receives time-series data containing discharge fault probability, multi-dimensional feature vectors, and environmental parameters, and organizes the input data using a sliding time window; the LSTM layer contains 128 memory units, captures temporal dependencies through a gating mechanism, and outputs the hidden state at each time step; the fully connected layer integrates the features of the output of the last time step of the LSTM and uses the ReLU activation function to enhance the nonlinear expression capability; the output layer maps the predicted value to the risk probability in the [0,1] interval through the Sigmoid activation function.

[0137] Specifically, the LSTM layer receives the historical switchgear discharge fault probability from the input layer, selectively memorizes and updates the temporal characteristics of the historical fault probability and environmental parameters through a gating mechanism, and generates the hidden state of the current time step.

[0138] It should be noted that the LSTM layer employs a triple gating mechanism of forget gate, input gate, and output gate to dynamically extract features from historical switchgear discharge fault probabilities and environmental parameters. The forget gate comprehensively evaluates the correlation between current monitoring parameters and historical operating states, using a sigmoid function to calculate information retention weights, thus updating the cell state memory. The input gate analyzes the relationship between current input features and the hidden state from the previous time step, identifying key features such as sudden changes in discharge probability and abnormal environmental parameters, and updating the cell state content. The output gate adjusts the output ratio of the current cell state, combined with tanh activation processing, to ultimately generate the hidden state for the current time step. This hidden state not only contains important temporal features filtered by the gating mechanism but also integrates the dynamic evolution of fault development, providing a time-dimensional state representation for subsequent risk prediction.

[0139] Specifically, the fully connected layer receives the hidden state of the LSTM layer, reconstructs the feature space through linear transformation of the weight matrix, and constructs the decision boundary by combining a nonlinear activation function to generate a high-order feature representation that characterizes the fault risk.

[0140] It should be noted that a high-order feature representation is generated based on the temporal hidden state output by the LSTM layer through a combination of linear transformation and nonlinear activation. Specifically, this layer first reconstructs and spatially maps the feature dimensions through linear transformation of the weight matrix, adjusting the transformation parameters according to the distribution characteristics of the input features; then, it dynamically adjusts the feature activation level according to the intensity of the input signal through the ReLU activation function and nonlinear transformation; finally, a high-order feature representation is generated.

[0141] Specifically, the output layer performs a linear transformation on the high-order feature representation generated by the fully connected layer, and maps it to the switchgear discharge risk probability R at a future time step through a Sigmoid activation function. t+Δt The formula is:

[0142]

[0143] Among them, R t+Δt is the predicted value of the discharge risk probability at the future time t + Δt, and R t-i is the predicted risk value at the historical moment t - i. m1 is the memory weight coefficient of the historical risk value, and G t is the temporal gradient feature at time t. E is the environmental parameter at the current time, H is the non - linear transformation function, H(E) is the environmental feature mapping function, α1 is the current feature weight, α2 is the temporal gradient weight, β is the historical memory term, γ is the adjustment intensity of the environmental impact, ε is the regularization coefficient, m2 is the constraint term to prevent over - fitting of the memory weight, k is the number of historical time steps, and i is the time step index.

[0144] It should be noted that first, a multi - source feature linear combination is performed, and the high - order features output by the fully - connected layer are weighted and fused with the temporal gradient features reflecting the risk change trend, the historical risk sequence capturing periodic patterns, and the environmental feature mapping quantifying external influencing factors. The contribution degrees of different features are balanced by adjusting the weight coefficients of each dimension. Second, through the regularization constraint mechanism, a quadratic penalty term of the weight parameter is added to the loss function, and the gradient decay effect is used to automatically control the model complexity, effectively improving the generalization performance. Finally, a probabilistic transformation is performed, and the linear combination result is mapped to a standardized probability value through a sigmoid function with non - linear characteristics. This transformation process can accurately describe the gradual change characteristics of the risk state. When the input features are significantly enhanced, the output approaches a high - risk warning, and when the features weaken, it corresponds to a low - risk state, showing a smooth transition in the critical region.

[0145] In E2: Generate a detection report; the specific steps include:

[0146] Based on historical data, define the low - risk threshold A1 and the high - risk threshold A2;

[0147] When R t+Δt < A1, it is considered that the discharge activity of the open cabinet is in a normal state;

[0148] When A1 ≤ R t+Δt < A2, it is considered that the discharge activity of the open cabinet exceeds the normal range and needs to be repaired in time;

[0149] When R t+Δt ≥ A2, it is considered that the open cabinet discharge activity has serious insulation defects, and power should be cut off immediately and an insulation diagnosis test should be carried out.

[0150] It should be noted that when the predicted risk value R t+Δt is less than the low - risk threshold A1, for example, R t+ΔtWhen A1 is 0.03 and A1 is 0.3, the switchgear is considered to be in normal operation. At this time, the discharge activity conforms to the historical normal operating conditions of the equipment, characterized by: the discharge signal amplitude remaining stable within the reference range, the discharge frequency meeting the expected level for the equipment's service life, and a reasonable correlation with changes in environmental parameters. Maintenance personnel can perform status monitoring according to the regular inspection cycle without initiating special handling procedures.

[0151] It should be noted that when predicting the risk probability R t+Δt When R is greater than or equal to the low-risk threshold A1 and less than the high-risk threshold A2, for example, R t+Δt When A1 is 0.4, A2 is 0.3, and A2 is 0.7, the open cabinet is determined to be in a warning state. At this time, the discharge activity exhibits the following typical characteristics: the discharge signal amplitude is significantly increased compared to the reference level but does not reach the danger threshold; the discharge frequency shows an abnormally increasing trend; and the correlation with environmental parameters deviates. This state indicates that the equipment insulation performance has begun to deteriorate, and there may be a risk of partial discharge or early insulation defects.

[0152] It should be noted that the predicted risk probability R t+Δt When the high-risk threshold A2 is reached or exceeded, the equipment is determined to be in an emergency fault state, for example, R. t+Δt When the amplitude is 0.7 and A2 is 0.6, the discharge activity exhibits the following typical characteristics: the discharge signal amplitude exceeds the safety limit, the discharge frequency shows an exponential growth trend, and the discharge phase distribution shows significant distortion. This state indicates that the equipment has a serious insulation degradation problem, which may be accompanied by the following specific fault modes: internal air gap discharge of solid insulation materials, development of surface creepage marks, or intermittent arc discharge caused by poor contact of conductive parts. The maintenance team must arrive at the site within 2 hours of receiving the alarm, prioritize using an infrared thermal imager to identify hotspot locations, and use a high-frequency current transformer to accurately measure the discharge quantity. Based on the diagnostic results, the insulation components confirmed to have dendritic discharge or carbonization channels must be replaced, and preventive inspections should be carried out on adjacent equipment.

[0153] It should be noted that step S500 not only provides early warning of potential insulation faults and avoids sudden accidents, but also reveals the trend of discharge development through time-series modeling, providing forward-looking guidance for condition-based maintenance. The final generated inspection report integrates multi-dimensional analysis results and predictive data, providing comprehensive and reliable decision support for the intelligent operation and maintenance of switchgear.

[0154] Example 3, referring to Figure 4 This embodiment provides a switchgear discharge detection system based on typical characteristics, including:

[0155] The data acquisition module is used to acquire multi-source signals and environmental parameters of the switchgear in real time and perform initial processing.

[0156] The feature extraction module is used to extract the signal nonlinear features of the first processed data, obtain the frequency domain energy features through the first transformation operation, and construct a multi-dimensional feature vector through the first fusion analysis operation.

[0157] The discharge signal generation module is used to build an adversarial network model, which performs adversarial training on multidimensional feature vectors and historical real partial discharge signals to generate high-fidelity partial discharge simulation signals.

[0158] The fault identification module is used to classify the discharge type based on high-fidelity partial discharge simulation signals using the first classification algorithm, and to construct a fault identification model by combining multi-dimensional feature vectors to identify the probability of discharge faults in the switchgear.

[0159] The risk prediction module is used to predict the discharge risk of switchgear based on the probability of switchgear discharge failure, and to generate a detection report by learning time series features using a risk prediction model.

[0160] It should be noted that the technical solution of the switchgear discharge detection system based on typical features is based on the same concept as the technical solution of the switchgear discharge detection method based on typical features described above. For details not described in detail in the technical solution of the switchgear discharge detection system based on typical features in this embodiment, please refer to the description of the technical solution of the switchgear discharge detection method based on typical features described above.

[0161] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0162] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a switchgear discharge detection method based on typical characteristics. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0163] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0164] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0165] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting discharge in switchgear based on typical characteristics, characterized in that, include: Real-time acquisition of multi-source signals and environmental parameters from the switchgear, followed by initial processing; Extract the signal nonlinear features of the first processed data, obtain the frequency domain energy features through the first transformation operation, and construct a multidimensional feature vector through the first fusion analysis operation; An adversarial network model is constructed, and the multidimensional feature vectors are trained against historical real partial discharge signals to generate high-fidelity partial discharge simulation signals. Based on the high-fidelity partial discharge simulation signal, the discharge type is classified using the first classification algorithm, and a fault identification model is constructed by combining the multi-dimensional feature vector to identify the probability of discharge faults in the switchgear. Based on the discharge fault probability of the switchgear, the risk of discharge of the switchgear is predicted by learning time series features using a risk prediction model, and a detection report is generated.

2. The switchgear discharge detection method based on typical characteristics as described in claim 1, characterized in that, The construction of the multidimensional feature vector includes: The phase space of the first processed data is reconstructed according to the first embedding theorem, and the nonlinear characteristics of the multi-source signal are quantified. The frequency domain energy characteristics of the multi-source signal are obtained by performing a first transformation operation. The nonlinear features, frequency domain energy features, and the first processed environmental parameters are optimized and combined using the first fusion analysis operation to construct a multidimensional feature vector.

3. The switchgear discharge detection method based on typical characteristics as described in claim 2, characterized in that, The classification of discharge types using the first classification algorithm includes: Based on the high-fidelity partial discharge simulation signal, signal features are extracted from the first dimension, the second dimension, and the third dimension respectively to obtain multi-dimensional features; The multi-dimensional features are processed by dynamic weight fusion combined with the first dimensionality reduction operation to generate a signal feature vector; The signal feature vector is classified using a first classification algorithm to obtain a discharge type label.

4. The switchgear discharge detection method based on typical characteristics as described in claim 3, characterized in that, The probability of identifying a switchgear discharge fault includes: Based on logistic regression, the probability of switchgear discharge failure is calculated by combining the discharge type label with the linear combination of the multidimensional feature vectors.

5. The switchgear discharge detection method based on typical characteristics as described in claim 1, characterized in that, The real-time acquisition of multi-source signals and environmental parameters from the switchgear, and the subsequent first processing, include: The switchgear multi-source signals include ultrasonic signals, ultra-high frequency signals, transient ground voltage, and high frequency current signals; the environmental parameters include temperature data, humidity data, air pressure data, vibration data, and dust concentration. The first processing includes multi-source signal denoising, spatiotemporal alignment, environmental parameter correction, and normalization processing.

6. The switchgear discharge detection method based on typical characteristics as described in claim 5, characterized in that, The generation of the high-fidelity partial discharge simulation signal includes: Based on the GAN multi-layer neural network architecture, an adversarial network model containing an input layer, a generator, and a discriminator is constructed. The generator extracts features and transforms the dimensions of the input multidimensional feature vector through a fully connected layer, and then gradually upsamples and reconstructs it through a deconvolution layer to generate a preliminary high-fidelity partial discharge simulation signal. The discriminator extracts the time-frequency features of the real discharge signal and the preliminary high-fidelity partial discharge simulation signal through a multi-layer convolutional neural network, and performs binary classification to distinguish the authenticity of the signal based on the time-frequency features. By iteratively executing the gradient backpropagation algorithm, the generator and the discriminator are alternately optimized to generate a high-fidelity partial discharge simulation signal.

7. The switchgear discharge detection method based on typical characteristics as described in claim 6, characterized in that, The method of using a risk prediction model to learn time series features to predict the discharge risk of switchgear includes: The risk prediction model consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. The LSTM layer receives the historical switchgear discharge fault probability from the input layer, and selectively memorizes and updates the temporal characteristics of the historical fault probability and environmental parameters through a gating mechanism to generate the hidden state of the current time step. The fully connected layer receives the hidden state of the LSTM layer, reconstructs the feature space through linear transformation of the weight matrix, and constructs the decision boundary by combining a nonlinear activation function to generate a high-order feature representation that characterizes the fault risk. The output layer performs a linear transformation on the high-order feature representation generated by the fully connected layer, and maps it to the switchgear discharge risk probability R at a future time step through a Sigmoid activation function. t+Δt ; Based on historical data, a low-risk threshold A1 and a high-risk threshold A2 are defined. When R t+Δt <is less than A1, it is considered that the discharge activity of the open cabinet is in a normal state; When A1 ≤ R t+Δt <When A2, it is considered that the discharge activity of the open cabinet exceeds the normal range and needs to be repaired in time; When R t+Δt If the value is ≥A2, it is considered that there is a serious insulation defect in the open cabinet discharge activity, and the power should be cut off immediately and an insulation diagnostic test should be carried out.

8. A switchgear discharge detection system based on typical characteristics, employing the switchgear discharge detection method based on typical characteristics as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire multi-source signals and environmental parameters of the switchgear in real time and perform initial processing. The feature extraction module is used to extract the signal nonlinear features of the first processed data, obtain the frequency domain energy features through the first transformation operation, and construct a multi-dimensional feature vector through the first fusion analysis operation. The discharge signal generation module is used to construct an adversarial network model, perform adversarial training on the multidimensional feature vector and historical real partial discharge signals, and generate high-fidelity partial discharge simulation signals. The fault identification module is used to classify the discharge type based on the high-fidelity partial discharge simulation signal using the first classification algorithm, and to construct a fault identification model by combining the multi-dimensional feature vector to identify the probability of discharge faults in the switchgear. The risk prediction module is used to predict the discharge risk of the switchgear based on the probability of discharge failure of the switchgear, and to generate a detection report by learning time series features using a risk prediction model.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the switchgear discharge detection method based on typical features as described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the switchgear discharge detection method based on typical features as described in any one of claims 1 to 7.

Images