GIS partial discharge fault simulation experiment system based on real defect

Abstract

The application discloses a GIS partial discharge fault simulation experiment system based on real defects, comprising a real defect signal acquisition module, a signal adaptive feature extraction module, a multi-dimensional feature modulation and coding module, a conditional generative adversarial network precise simulation module, a joint discrimination and defect self-supervised classification module and a real-time feedback dynamic optimization module, through constructing a real defect feature database, a feature adaptive modulation and coding structure and a conditional generative adversarial network precise simulation structure, accurate mapping and dynamic optimization of the partial discharge simulation signal and the real defect type are realized. The application improves the fitting precision of the simulation signal to the real defect discharge characteristics in the GIS partial discharge experiment, and is suitable for the development and verification of the GIS equipment partial discharge detection method and detection equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis and intelligent detection technology, and in particular to a GIS partial discharge fault simulation experimental system based on real defects. Background Technology

[0002] In the field of power equipment fault detection, accurately detecting partial discharge defects inside GIS (Gas Insulated Metal Enclosed Switchgear) has always been a key issue in ensuring the safe and stable operation of the power grid. With the increasing demands for equipment reliability and safety in power systems, existing partial discharge defect detection methods based on traditional simulation and experimentation are gradually failing to meet the needs of detecting and evaluating the complex and diverse types of real-world defects in GIS equipment. To more accurately analyze and model partial discharge defect signals, academia and industry have recently introduced artificial intelligence technologies based on deep learning, especially generative adversarial networks, to generate simulated signals that resemble the actual distribution of defect discharge signals, thus promoting the development of intelligent fault diagnosis technology for GIS equipment.

[0003] Existing methods for partial discharge detection and simulation in GIS generally employ traditional neural networks or conventional GANs to perform simple feature extraction and simulation of partial discharge signals, thereby assisting detection systems or devices in verifying the effectiveness of defect detection. For example, existing technologies commonly use a single neural network structure, relying on manually defined static signal features for defect signal simulation. However, such methods are often limited to general signal feature processing and struggle to effectively incorporate the unique multidimensional time-frequency characteristics specific to real GIS defect types. This results in significant discrepancies in feature distribution between the generated simulated signals and real defect signals, limiting the model's generalization ability and the accuracy of the detection system.

[0004] In terms of simulation data generation, existing technologies often neglect the deep integration and utilization of real defect type information, lack effective feature modulation and condition control mechanisms, and struggle to achieve accurate mapping of specific defect types. Specifically, real defect types in GIS equipment, such as conductive particles, tip discharge, and insulator defects, exhibit significantly different time-frequency characteristics of partial discharge. Existing methods fail to finely encode defect features, resulting in simulation signals that cannot reflect the differentiated characteristic information between defect types. Furthermore, existing methods lack specificity in dynamic optimization of the feature space, failing to form an efficient feedback closed-loop adjustment process, leading to poor stability and robustness of the generated simulation signals.

[0005] In terms of model structure, existing generative adversarial network (GAN) methods typically employ a general GAN ​​architecture. However, they fail to establish a sufficiently organic and dynamic interaction between signal feature input, simulation output, and the authenticity verification process, and fail to construct an effective joint learning mechanism between the generator and discriminator networks. Furthermore, existing techniques often fail to effectively integrate the simulation signal authenticity verification task with the defect type classification task, resulting in a simplistic model structure that cannot accurately determine the precision of the generated signal in terms of defect type, thus limiting the generation quality and classification accuracy of the simulation model.

[0006] Regarding feedback optimization, current partial discharge simulation models generally lack real-time dynamic parameter adaptive adjustment mechanisms. Once the simulation data differs from the actual defect signal, existing methods typically struggle to identify the source of error and proactively update network parameters in real time. This leads to the model's inability to adapt to simulation deviations caused by changes in the actual operating environment of GIS equipment during field operation, severely impacting the long-term reliability of the simulation and detection system.

[0007] Therefore, how to provide a GIS partial discharge fault simulation experimental system based on real defects is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a GIS partial discharge fault simulation experimental system based on real defects. This invention integrates real defect type feature information and partial discharge signal data, constructs a feature adaptive modulation structure and an improved conditional generative adversarial network architecture, and achieves accurate mapping of defect types and dynamic optimization processing of simulation signals through feature attention mechanism and joint discriminant feedback method. It has the advantages of high simulation accuracy, accurate defect feature fitting and adaptive dynamic optimization of parameters.

[0009] The GIS partial discharge fault simulation experimental system based on real defects according to an embodiment of the present invention includes the following modules:

[0010] The real defect signal acquisition module is used to acquire and output the original partial discharge signals that correspond one by one to various real defect types under the operating conditions of GIS equipment;

[0011] The signal adaptive feature extraction module is used to receive the original partial discharge signal and perform adaptive time-frequency feature decomposition, output high-dimensional features corresponding to each real defect type, and construct a real defect feature database.

[0012] The multidimensional feature modulation and coding module is used to extract high-dimensional features corresponding to the real defect type from the real defect feature database, and to perform feature modulation on the high-dimensional features based on the attention mechanism, and output the modulated defect features and the real defect type code.

[0013] The Conditional Generative Adversarial Network (CGAN) Precision Simulation Module is used to receive the modulated defect features and the actual defect type encoding, and generate a partial discharge simulation signal that is consistent with the time-frequency domain characteristic distribution of the original partial discharge signal based on the improved CGAN architecture.

[0014] The joint discrimination and defect self-supervised classification module is used to simultaneously perform authenticity identification and self-supervised classification on the partial discharge simulation signal and the original partial discharge signal, and output the discrimination error of the partial discharge simulation signal and the defect classification result.

[0015] The real-time feedback dynamic optimization module is used to receive the discrimination error and defect classification results, dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, and achieve adaptive fitting between the partial discharge simulation signal and the original partial discharge signal distribution.

[0016] Optionally, modules can be integrated using the following methods:

[0017] S1. Collect the original partial discharge signals of GIS equipment under various real defect types, and establish data channels for the original partial discharge signals according to the defect type.

[0018] S2. Based on the adaptive time-frequency analysis algorithm, high-dimensional features are extracted from each data channel in the original partial discharge signal, and a real defect feature database is generated through feature decomposition and reconstruction.

[0019] S3. Based on the real defect feature database, the high-dimensional features of each real defect type are adaptively modulated using an attention mechanism, and encoded in combination with the corresponding real defect type information in the real defect feature database to form modulated defect features and real defect type codes.

[0020] S4. Based on the modulated defect features and the encoding of the real defect type, a feature attention modulation mechanism is embedded to construct an improved conditional generative adversarial network architecture. Through adversarial training, a partial discharge simulation signal with the same time-frequency characteristic distribution as the original partial discharge signal is generated.

[0021] S5. Based on the partial discharge simulation signal and the original partial discharge signal, perform self-supervised classification of simulation signal authenticity and defect type, calculate and output the discrimination error and classification result;

[0022] S6. Based on the discrimination error and classification result, a real-time feedback dynamic optimization model is constructed to adaptively and dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, so that the partial discharge simulation signal is adaptively fitted to the feature distribution of the original partial discharge signal.

[0023] Optionally, S1 specifically includes:

[0024] S11. Based on the structural and internal insulation design requirements of the GIS equipment, determine the structural dimensions, material, and placement parameters of the actual defective components;

[0025] S12. Based on the actual defective component's structural dimensions, material, and placement parameters, fabricate actual defective components such as conductive microparticles, metal particles, insulator tip burrs, and gas gap defects.

[0026] S13. Install the actual defective component in the predetermined position in the GIS equipment, and use a three-dimensional positioning device to calibrate the spatial coordinate position of the defective component relative to the insulator and conductor inside the GIS equipment.

[0027] S14. Set the experimental voltage and SF6 gas pressure parameters according to the rated operating conditions of the GIS equipment, start the GIS equipment, and use a high-frequency sensor array to synchronously collect the original partial discharge signals generated by the discharge of real defective components from multiple angles.

[0028] S15. The original partial discharge signals synchronously acquired from multiple angles by the high-frequency sensor array are digitally converted and labeled with the corresponding real defect types. Independent data channels are established according to the labeled real defect types to form data channels that correspond one-to-one with each real defect type.

[0029] Optionally, S2 specifically includes:

[0030] S21. Input the data channels of the original partial discharge signal into the time-frequency analysis algorithm based on adaptive short-time Fourier transform respectively;

[0031] S22. Perform adaptive short-time Fourier transform on the original partial discharge signal in each data channel to obtain time-frequency domain feature parameters that can accurately characterize the corresponding real defect type, and form a high-dimensional feature set.

[0032] S23. Construct a feature decomposition and reconstruction mapping structure based on the high-dimensional feature set, set a shared representation of the mapping relationship between the input channel of the high-dimensional feature set and the corresponding real defect type identification channel, and each output channel is composed of a nonlinear mapping structure of the time-frequency domain feature parameters, and is connected to the corresponding real defect type identification node at the output end respectively;

[0033] S24. Two optimization processing steps are set for the high-dimensional feature set: feature correlation analysis and feature redundancy removal. By removing redundant feature parameters and enhancing core feature parameters, an optimized high-dimensional feature set is formed.

[0034] S25. Based on the optimized high-dimensional feature set, establish the data structure of the real defect feature database, and map the optimized high-dimensional feature set to the corresponding real defect type identifier nodes one by one and store them in the real defect feature database.

[0035] Optionally, S3 specifically includes:

[0036] S31. Based on the mapping relationship between the optimized high-dimensional feature set stored in the real defect feature database and the corresponding real defect type identifier node, construct the feature adaptive modulation structure of the attention mechanism, wherein the real defect type identifier node is the basis for calculating the feature modulation weight.

[0037] S32. In the feature adaptive modulation structure, a feature attention weight calculation unit is established to calculate the attention weight of each feature dimension based on the mapping relationship between the input optimized high-dimensional feature set and the corresponding real defect type identification node.

[0038] S33. Adaptively weighted modulation is applied to each feature dimension in the optimized high-dimensional feature set using the calculated attention weights to generate modulated defect features.

[0039] S34. Based on the real defect type identifier node corresponding to the real defect feature database, construct a real defect type digital encoding unit to obtain a real defect type digital encoding that corresponds one-to-one with the modulated defect feature.

[0040] S35. Establish a feature fusion mapping structure to fuse the modulated defect features with the corresponding real defect types in digital encoding to form a unified modulated defect feature and real defect type encoding.

[0041] Optionally, the feature attention weight calculation unit includes:

[0042] The system receives the optimized high-dimensional feature set and the corresponding real defect type identifier node, and performs linear mapping processing on the input high-dimensional feature set to form the corresponding query feature vector, key feature vector and value feature vector.

[0043] The importance coefficients of each feature dimension are calculated based on the degree of correlation between the query feature vector and the key feature vector, thus obtaining the feature correlation coefficients that can reflect the intrinsic feature differences of the high-dimensional feature set.

[0044] The feature correlation coefficient is used to perform a dimension-by-dimensional weighted fusion of the feature dimensions in the value feature vector to generate feature attention weights;

[0045] Based on the calculation deviation between the actual defect type identification node and the feature attention weight, the network parameters in the linear mapping process are dynamically adjusted using the error backpropagation method to optimize the calculation accuracy of the feature attention weight in real time.

[0046] Optionally, S4 specifically includes:

[0047] S41. The modulated defect features and the real defect type encoding are used as conditional input sources, and the real defect type encoding is used as conditional control information.

[0048] S42. The modulated defect features are input to the input of the generator network of the improved conditional generative adversarial network. Through the feature attention modulation mechanism, the modulated defect features are adaptively modulated with feature weights in each dimension to form a generated feature vector with time-frequency characteristic distribution information.

[0049] S43. Establish a corresponding real defect type condition constraint structure based on the real defect type encoding, and generate a corresponding initial partial discharge simulation signal based on the generated feature vector and the real defect type condition constraint structure.

[0050] S44. Input the initial partial discharge simulation signal to the input terminal of the discriminant network of the improved conditional generative adversarial network, and at the same time introduce the corresponding original partial discharge signal from the data channel to perform a true / false adversarial identification between the initial partial discharge simulation signal and the original partial discharge signal.

[0051] S45. Based on the discrimination result of the discrimination network, determine the difference in time-frequency characteristic distribution between the initial partial discharge simulation signal and the original partial discharge signal, calculate the difference loss function value, and update the network parameters of the adversarial network by reversing the improvement conditions through the difference loss function value.

[0052] S46. Repeat S42 to S45 until the difference loss function value reaches the preset convergence threshold, and output a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal.

[0053] Optionally, the improved conditional generative adversarial network specifically includes:

[0054] A feature attention modulation generative network is constructed. The modulated defect features are fused with the corresponding real defect type encoding and used as the network input. The feature weights are dynamically allocated dimension by dimension using a multi-head self-attention mechanism to represent the time-frequency domain distribution features corresponding to the real defect type.

[0055] In the generative network, a feature dimension expansion layer and a deep convolution-deconvolution network are constructed to progressively transform the feature vectors modulated by feature attention into a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal.

[0056] A dual-task joint discrimination network structure is constructed. The bottom layer adopts a shared feature extraction structure, while the upper layer sets up independent authenticity discrimination structure and defect type classification structure, which respectively output the authenticity judgment probability value and defect type prediction value of the partial discharge simulation signal.

[0057] A feature interaction attention mechanism is embedded between the shared feature extraction structure and the defect type classification structure to explicitly interact the feature association relationship between the partial discharge simulation signal and the real defect type;

[0058] Based on the authenticity probability value and defect type prediction value output by the discrimination network, the simulation error is calculated and fed back to the generation network in real time, and the parameters of the generation network are dynamically updated using the gradient backpropagation method.

[0059] Optionally, S5 specifically includes:

[0060] S51. Receive the partial discharge simulation signal generated by the improved conditional generative adversarial network, and receive the corresponding original partial discharge signal to form a paired input structure of simulation signal and original signal.

[0061] S52. Input the paired input structure into the authenticity discrimination structure, perform feature extraction and difference analysis on the partial discharge simulation signal and the original partial discharge signal respectively, and obtain the authenticity probability value of the partial discharge simulation signal relative to the original partial discharge signal.

[0062] S53. Based on the information of the corresponding real defect type identifier node in the real defect feature database, construct a self-supervised classification structure for defect types, input the partial discharge simulation signal into the self-supervised classification structure for defect types, and determine the predicted value of the defect type corresponding to the partial discharge simulation signal.

[0063] S54. The true / false probability value and the defect type prediction value are compared with the corresponding original partial discharge signal and the real defect type identification node information, respectively. The discrimination error that can characterize the degree of difference between the true and false partial discharge simulation signal and the classification error that can characterize the degree of difference between the defect type prediction are obtained by calculation.

[0064] S55. Output the discrimination error and classification error as feedback for the improved conditional generative adversarial network to update network parameters in real time.

[0065] Optionally, S6 specifically includes:

[0066] S61. Receive the discrimination error calculated by comparing the partial discharge simulation signal with the original partial discharge signal, and the classification error calculated by comparing the defect type prediction value of the partial discharge simulation signal with the corresponding real defect type identifier node information.

[0067] S62. Construct a real-time feedback dynamic optimization objective function based on the discrimination error and classification error, and use the objective function to characterize the degree of difference between the network parameters of the improved conditional generative adversarial network and the original partial discharge signal;

[0068] S63. The optimization direction of the network parameters is determined by the real-time feedback dynamic optimization objective function, and the update gradient of the network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network is calculated based on the gradient descent algorithm.

[0069] S64. The network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network are dynamically adjusted according to the calculated update gradient to form an adaptive update process of the parameters of the partial discharge simulation signal.

[0070] S65. Repeat the above S61 to S64 processes until the value of the real-time feedback dynamic optimization objective function reaches the preset convergence threshold, thus completing the dynamic optimization of the network parameters of the improved conditional generative adversarial network.

[0071] The beneficial effects of this invention are:

[0072] (1) This invention constructs a real defect feature database and introduces a feature adaptive modulation structure. Through the attention mechanism, it accurately captures the multidimensional feature differences of different defect types in GIS partial discharge signals, effectively realizes the accurate mapping between feature information and real defect types, improves the accuracy and distinguishability of partial discharge simulation signals, breaks through the limitation of traditional methods that rely on static features and cannot accurately distinguish multiple types of defect signals, and significantly improves the accuracy of partial discharge defect detection and diagnosis.

[0073] (2) The present invention constructs an improved conditional generative adversarial network architecture, integrates a dual-task joint discrimination structure of authenticity identification and defect type classification, and embeds a feature interaction attention mechanism, which effectively improves the fitting degree between the simulation signal and the actual defect discharge signal, realizes the accurate generation and stable output of the simulation signal in the time-frequency characteristic distribution, overcomes the shortcomings of insufficient generalization ability and poor signal quality of the traditional single-task GAN architecture, and significantly enhances the quality and stability of the model-generated signal.

[0074] (3) By constructing a real-time feedback dynamic optimization model, the present invention uses the discrimination error and classification error to dynamically adjust the parameters of the generation network and the discrimination network in real time, which significantly improves the real-time dynamic adaptability and long-term stability of the model, solves the technical defect of the traditional simulation method lacking a real-time parameter adaptive adjustment mechanism, realizes accurate adaptive fitting of partial discharge simulation signals under complex and variable actual working conditions, and significantly enhances the robustness and generalization ability of the model to changes in on-site working conditions. Attached Figure Description

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0076] Figure 1 This is a schematic diagram of the overall system structure of the GIS partial discharge fault simulation experimental system based on real defects proposed in this invention;

[0077] Figure 2 This is a schematic diagram of the improved conditional generative adversarial network structure of the GIS partial discharge fault simulation experimental system based on real defects proposed in this invention.

[0078] Figure 3 This is a flowchart of the real-time feedback dynamic optimization model of the GIS partial discharge fault simulation experimental system based on real defects proposed in this invention. Detailed Implementation

[0079] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0080] refer to Figures 1-3 The GIS partial discharge fault simulation experimental system based on real defects includes the following modules:

[0081] The real defect signal acquisition module is used to acquire and output the original partial discharge signals that correspond one by one to various real defect types under the operating conditions of GIS equipment;

[0082] The signal adaptive feature extraction module is used to receive the original partial discharge signal and perform adaptive time-frequency feature decomposition, output high-dimensional features corresponding to each real defect type, and construct a real defect feature database.

[0083] The multidimensional feature modulation and coding module is used to extract high-dimensional features corresponding to the real defect type from the real defect feature database, and to perform feature modulation on the high-dimensional features based on the attention mechanism, and output the modulated defect features and the real defect type code.

[0084] The Conditional Generative Adversarial Network (CGAN) Precision Simulation Module is used to receive the modulated defect features and the actual defect type encoding, and generate a partial discharge simulation signal that is consistent with the time-frequency domain characteristic distribution of the original partial discharge signal based on the improved CGAN architecture.

[0085] The joint discrimination and defect self-supervised classification module is used to simultaneously perform authenticity identification and self-supervised classification on the partial discharge simulation signal and the original partial discharge signal, and output the discrimination error of the partial discharge simulation signal and the defect classification result.

[0086] The real-time feedback dynamic optimization module is used to receive the discrimination error and defect classification results, dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, and achieve adaptive fitting between the partial discharge simulation signal and the original partial discharge signal distribution.

[0087] The operating conditions of the GIS equipment for the real defect signal acquisition module described in this invention include a rated operating voltage of 110kV to 550kV, an SF6 gas pressure of 0.4MPa to 0.6MPa, and an ambient operating temperature of 20℃ to 40℃. Real defect types include conductive particle defects with a diameter range of 0.2mm to 2mm, insulator tip burr defects with a tip curvature radius of less than 0.1mm, gas gap defects with a gap distance of 0.1mm to 3mm, and metal particle defects with a diameter range of 0.5mm to 3mm. The real defect elements are manufactured using conductive metal materials such as copper and aluminum through precision lathe machining and are precisely installed on the conductor surface or insulator fixed position inside the GIS equipment using a three-dimensional positioning device, with installation positioning accuracy controlled within ±0.1mm. The original partial discharge signal is acquired using a four-frequency sensor array installed at different locations on the GIS equipment housing. The sensor frequency response range is 300MHz to 1500MHz, and the digital sampling frequency is set to 10GHz to ensure that the acquired original signal has complete time-frequency domain characteristics without distortion. The adaptive time-frequency analysis algorithm described in this invention employs adaptive short-time Fourier transform (ASTFT), uses the Hamming window function for adaptive time window adjustment, and sets the dynamic adjustment range of the window length within 100ns to 500ns. It uses signal energy peak, instantaneous frequency, and spectral energy distribution as feature extraction criteria. After feature redundancy removal processing, the feature information corresponding to each real defect type is stored in the real defect feature database in matrix form.

[0088] The present invention adopts the above technical solution, which can accurately realize the generation of simulation signals of partial discharge defect characteristics of GIS equipment, significantly improve the accuracy and stability of the generated simulation signals corresponding to the real defect types, and effectively improve the reliability and generalization ability of the partial discharge defect detection and verification system.

[0089] In this embodiment, the modules are interconnected using the following method:

[0090] S1. Collect the original partial discharge signals of GIS equipment under various real defect types, and establish data channels for the original partial discharge signals according to the defect type.

[0091] S2. Based on the adaptive time-frequency analysis algorithm, high-dimensional features are extracted from each data channel in the original partial discharge signal, and a real defect feature database is generated through feature decomposition and reconstruction.

[0092] S3. Based on the real defect feature database, the high-dimensional features of each real defect type are adaptively modulated using an attention mechanism, and encoded in combination with the corresponding real defect type information in the real defect feature database to form modulated defect features and real defect type codes.

[0093] S4. Based on the modulated defect features and the encoding of the real defect type, a feature attention modulation mechanism is embedded to construct an improved conditional generative adversarial network architecture. Through adversarial training, a partial discharge simulation signal with the same time-frequency characteristic distribution as the original partial discharge signal is generated.

[0094] S5. Based on the partial discharge simulation signal and the original partial discharge signal, perform self-supervised classification of simulation signal authenticity and defect type, calculate and output the discrimination error and classification result;

[0095] S6. Based on the discrimination error and classification result, a real-time feedback dynamic optimization model is constructed to adaptively and dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, so that the partial discharge simulation signal is adaptively fitted to the feature distribution of the original partial discharge signal.

[0096] The data channels described in this invention employ independent data storage paths, specifically including: storing the original partial discharge signals acquired from four real defect types—conductive particle defects, insulator tip burr defects, gas gap defects, and metal particle defects—in separate digital channels. The sampling frequency of each channel is uniformly set to 10 GHz, and the data sampling bit depth is set to 16 bits to ensure the accuracy and consistency of the original signal data. This invention uses adaptive short-time Fourier transform to extract time-frequency features from the original signals stored in each data channel. A Hamming window is selected as the window function, with a fixed window length of 300 ns and a window overlap rate of 50%. The extracted time-frequency features specifically include signal energy accumulation features, spectral peak features, and instantaneous frequency features, ensuring that the feature parameter dimensions for each defect type are uniformly set to 256 dimensions. A feature redundancy removal method is used to filter the feature parameters of each dimension, retaining only the 128 feature dimensions with the highest correlation to defect type identification. Finally, these are stored in a unified, standardized feature matrix format in the real defect feature database, and each feature matrix is ​​labeled with the corresponding digital encoding information of the real defect type, ensuring sufficient database retrieval efficiency and accuracy.

[0097] Through the above-described implementation scheme, this invention achieves high precision and standardization in the extraction and storage of original partial discharge signal features, effectively ensuring the mapping relationship between feature data and real defect types, improving the consistency between simulated signal generation and real defect signals, and effectively enhancing the reliability and accuracy of defect diagnosis and verification processes.

[0098] In this embodiment, S1 specifically includes:

[0099] S11. Based on the structural and internal insulation design requirements of the GIS equipment, determine the structural dimensions, material, and placement parameters of the actual defective components;

[0100] S12. Based on the actual defective component's structural dimensions, material, and placement parameters, fabricate actual defective components such as conductive microparticles, metal particles, insulator tip burrs, and gas gap defects.

[0101] S13. Install the actual defective component in the predetermined position in the GIS equipment, and use a three-dimensional positioning device to calibrate the spatial coordinate position of the defective component relative to the insulator and conductor inside the GIS equipment.

[0102] S14. Set the experimental voltage and SF6 gas pressure parameters according to the rated operating conditions of the GIS equipment, start the GIS equipment, and use a high-frequency sensor array to synchronously collect the original partial discharge signals generated by the discharge of real defective components from multiple angles.

[0103] S15. The original partial discharge signals synchronously acquired from multiple angles by the high-frequency sensor array are digitally converted and labeled with the corresponding real defect types. Independent data channels are established according to the labeled real defect types to form data channels that correspond one-to-one with each real defect type.

[0104] The specific structural dimensions of the actual defect elements in this invention are as follows: the diameters of the conductive microparticles are 0.2mm, 0.5mm, and 1mm, and the material is copper microparticles; the metal particles are aluminum particles with diameters of 0.5mm, 1mm, and 3mm; the tip curvature radius of the insulator tip burr is 0.05mm; the gap distances of the gas gap defects are set to 0.5mm, 1mm, and 2mm, and the gap structure is formed using standard metal electrode pairs. The actual defect elements are installed on the surface of the conductor inside the GIS equipment and on the surface of the supporting insulator. Installation uses a high-precision three-dimensional positioning device, with positioning accuracy limited to ±0.05mm. The high-frequency sensor array adopts a spatial layout, including four high-frequency sensors installed in four mutually perpendicular positions on the GIS equipment housing. The frequency response range of the sensors is fixed at 300MHz~1500MHz. The digital conversion of the acquired signals uses a high-speed data acquisition device with a sampling rate of 10GHz and a bit depth of 16bit. Each data acquisition is labeled with a digital code indicating the defect type to ensure the integrity, accuracy, and uniqueness of the acquired data. Separate and independent data channels are established for storage and management.

[0105] This embodiment, through the above technical solution, defines the size, material, installation location and parameters of the actual defective component, as well as the specific specifications for signal acquisition and data channel construction. It significantly improves the standardization and accuracy of partial discharge signal acquisition and storage, ensures the high reliability of subsequent feature extraction and simulation analysis, and provides an accurate data foundation for subsequent precise simulation of partial discharge signals and defect diagnosis.

[0106] In this embodiment, S2 specifically includes:

[0107] S21. Input the data channels of the original partial discharge signal into the time-frequency analysis algorithm based on adaptive short-time Fourier transform respectively;

[0108] S22. Perform adaptive short-time Fourier transform on the original partial discharge signal in each data channel to obtain time-frequency domain feature parameters that can accurately characterize the corresponding real defect type, and form a high-dimensional feature set.

[0109] S23. Construct a feature decomposition and reconstruction mapping structure based on the high-dimensional feature set, set a shared representation of the mapping relationship between the input channel of the high-dimensional feature set and the corresponding real defect type identification channel, and each output channel is composed of a nonlinear mapping structure of the time-frequency domain feature parameters, and is connected to the corresponding real defect type identification node at the output end respectively;

[0110] S24. Two optimization processing steps are set for the high-dimensional feature set: feature correlation analysis and feature redundancy removal. By removing redundant feature parameters and enhancing core feature parameters, an optimized high-dimensional feature set is formed.

[0111] S25. Based on the optimized high-dimensional feature set, establish the data structure of the real defect feature database, and map the optimized high-dimensional feature set to the corresponding real defect type identifier nodes one by one and store them in the real defect feature database.

[0112] This invention proposes an Adaptive Short-Time Fourier Transform (ASTFT) algorithm, where the adaptive window function is a Hamming window with a fixed length of 300 ns and a fixed window overlap rate of 50%, ensuring the integrity and consistency of signal features during time-frequency analysis. Specific time-frequency domain feature parameters include the signal's spectral energy distribution parameters, peak energy parameters, and instantaneous frequency parameters. The time-frequency domain feature matrix output by each data channel has a fixed dimension of 256. The feature decomposition and reconstruction mapping structure is implemented through a nonlinear fully connected network with two hidden layers. Each hidden layer uses the ReLU activation function, and the number of nodes in the input and output layers is set to 256. Each output channel is connected to a unique and fixed real defect type identifier node. Feature correlation analysis uses the Pearson correlation coefficient calculation method, and redundancy removal uses analysis of variance, retaining the 128 feature dimensions with the highest correlation and largest variance. Finally, the optimized feature matrix is ​​stored in a real defect feature database using a standardized storage structure, ensuring a clear and unique mapping relationship between the feature set and defect type in the database.

[0113] This embodiment achieves high-precision feature extraction and standardized storage of the original partial discharge signal through the above technical solution, ensuring that the feature distribution of the subsequently generated simulation signal is accurately mapped to the real defect type, which greatly improves the accuracy and stability of partial discharge defect simulation and classification, and provides a reliable technical foundation for the accurate detection of partial discharge defects in GIS equipment.

[0114] In this embodiment, S3 specifically includes:

[0115] S31. Based on the mapping relationship between the optimized high-dimensional feature set stored in the real defect feature database and the corresponding real defect type identifier node, construct the feature adaptive modulation structure of the attention mechanism, wherein the real defect type identifier node is the basis for calculating the feature modulation weight.

[0116] S32. In the feature adaptive modulation structure, a feature attention weight calculation unit is established to calculate the attention weight of each feature dimension based on the mapping relationship between the input optimized high-dimensional feature set and the corresponding real defect type identification node.

[0117] S33. Adaptively weighted modulation is applied to each feature dimension in the optimized high-dimensional feature set using the calculated attention weights to generate modulated defect features.

[0118] S34. Based on the real defect type identifier node corresponding to the real defect feature database, construct a real defect type digital encoding unit to obtain a real defect type digital encoding that corresponds one-to-one with the modulated defect feature.

[0119] S35. Establish a feature fusion mapping structure to fuse the modulated defect features with the corresponding real defect types in digital encoding to form a unified modulated defect feature and real defect type encoding.

[0120] The feature attention weight calculation unit includes:

[0121] The system receives the optimized high-dimensional feature set and the corresponding real defect type identifier node, and performs linear mapping processing on the input high-dimensional feature set to form the corresponding query feature vector, key feature vector and value feature vector.

[0122] The importance coefficients of each feature dimension are calculated based on the degree of correlation between the query feature vector and the key feature vector, thus obtaining the feature correlation coefficients that can reflect the intrinsic feature differences of the high-dimensional feature set.

[0123] The feature correlation coefficient is used to perform a dimension-by-dimensional weighted fusion of the feature dimensions in the value feature vector to generate feature attention weights;

[0124] Based on the calculation deviation between the actual defect type identification node and the feature attention weight, the network parameters in the linear mapping process are dynamically adjusted using the error backpropagation method to optimize the calculation accuracy of the feature attention weight in real time.

[0125] In this embodiment, the adaptive modulation structure of the attention mechanism of the feature attention weight calculation unit specifically includes an input layer, a feature mapping layer, an attention calculation layer, and an output layer. The input layer receives high-dimensional feature set data and generates query feature vectors, key feature vectors, and value feature vectors through independent linear mapping operations. The feature mapping layer contains a fully connected network with 256 input nodes and 256 output nodes. Each node forms a feature vector through matrix multiplication. The attention calculation layer calculates the feature correlation coefficient through the inner product between the query feature vector and the key feature vector output by the feature mapping layer. The correlation coefficient is then normalized by the Softmax function to form the attention weight for each feature dimension. The output layer uses the attention weight of the feature dimension to perform a weighted summation of the value feature vector dimension by dimension to obtain the modulated defect features, ensuring that there is a unique and clear mapping relationship between the final output modulated feature set and the real defect type identification node. The real defect type digital encoding unit encodes the real defect type identification node using a fixed digital encoding method. Each real defect type identification node corresponds to a unique and fixed 32-bit binary code. The code is then spliced ​​and fused with the modulated defect features through a feature fusion mapping structure, ultimately forming a unified digital feature encoding structure that can be used as input for conditional generative adversarial networks.

[0126] This embodiment achieves accurate and stable mapping and modulation between defect types and high-dimensional features through the above technical solution, which greatly improves the distinguishability of defect signal features, significantly enhances the accuracy of features in the subsequent simulation generation process, and effectively improves the accuracy and reliability of the partial discharge signal simulation corresponding to the real defect type.

[0127] In this embodiment, S4 specifically includes:

[0128] S41. The modulated defect features and the real defect type encoding are used as conditional input sources, and the real defect type encoding is used as conditional control information.

[0129] S42. The modulated defect features are input to the input of the generator network of the improved conditional generative adversarial network. Through the feature attention modulation mechanism, the modulated defect features are adaptively modulated with feature weights in each dimension to form a generated feature vector with time-frequency characteristic distribution information.

[0130] S43. Establish a corresponding real defect type condition constraint structure based on the real defect type encoding, and generate a corresponding initial partial discharge simulation signal based on the generated feature vector and the real defect type condition constraint structure.

[0131] S44. Input the initial partial discharge simulation signal to the input terminal of the discriminant network of the improved conditional generative adversarial network, and at the same time introduce the corresponding original partial discharge signal from the data channel to perform a true / false adversarial identification between the initial partial discharge simulation signal and the original partial discharge signal.

[0132] S45. Based on the discrimination result of the discrimination network, determine the difference in time-frequency characteristic distribution between the initial partial discharge simulation signal and the original partial discharge signal, calculate the difference loss function value, and update the network parameters of the adversarial network by reversing the improvement conditions through the difference loss function value.

[0133] S46. Repeat S42 to S45 until the difference loss function value reaches the preset convergence threshold, and output a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal.

[0134] The improved conditional generative adversarial network specifically includes:

[0135] A feature attention modulation generative network is constructed. The modulated defect features are fused with the corresponding real defect type encoding and used as the network input. The feature weights are dynamically allocated dimension by dimension using a multi-head self-attention mechanism to represent the time-frequency domain distribution features corresponding to the real defect type.

[0136] In the generative network, a feature dimension expansion layer and a deep convolution-deconvolution network are constructed to progressively transform the feature vectors modulated by feature attention into a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal.

[0137] A dual-task joint discrimination network structure is constructed. The bottom layer adopts a shared feature extraction structure, while the upper layer sets up independent authenticity discrimination structure and defect type classification structure, which respectively output the authenticity judgment probability value and defect type prediction value of the partial discharge simulation signal.

[0138] A feature interaction attention mechanism is embedded between the shared feature extraction structure and the defect type classification structure to explicitly interact the feature association relationship between the partial discharge simulation signal and the real defect type;

[0139] Based on the authenticity probability value and defect type prediction value output by the discrimination network, the simulation error is calculated and fed back to the generation network in real time, and the parameters of the generation network are dynamically updated using the gradient backpropagation method.

[0140] In this embodiment, the modulated defect feature dimension input to the generator network of the improved conditional generative adversarial network is 128-dimensional, and the real defect type encoding dimension is 32-dimensional. The two are concatenated into a unified 160-dimensional input. The feature attention modulation mechanism specifically adopts four parallel self-attention substructures. In each substructure, the query, key, and value feature vector dimensions are all set to 40-dimensional. The self-attention structure adaptively calculates the dimension weights and fuses them into a unified generated feature vector. The feature dimension expansion layer of the generator network includes a fully connected network layer that expands to 512-dimensional features and is input into a 5-layer convolutional-deconvolutional network. The kernel size of each convolutional layer is fixed at 3×3. The features are progressively mapped to a time-domain signal length of 1024 points. The output partial discharge simulation signal has a sampling rate of 10GHz and a bit depth of 16bit. The shared feature extraction structure of the discriminant network contains three layers of one-dimensional convolutions. The system consists of layers with a fixed kernel size of 5. Each layer outputs 256 dimensions. The true / false discrimination structure and the defect type classification structure each employ a two-layer fully connected network for task classification. The discriminant probability output layer uses the Sigmoid activation function, while the defect classification output layer uses the Softmax activation function. The classification dimensions correspond to the four types of real defect types. The feature interaction attention mechanism interacts through two independent self-attention sub-networks, dynamically adjusting the output weights of the shared feature extraction structure. Real-time feedback dynamic optimization calculates the true / false judgment probability and the deviation between the predicted defect type and the true label using the mean squared error function. The Adam optimization algorithm dynamically updates the weight parameters of the generated network. The learning rate for network parameter updates is fixed at 0.001, and the convergence threshold is set when the error function value is below 0.001.

[0141] This embodiment achieves refined design and real-time dynamic feedback optimization of conditional generative adversarial networks through the above technical solution. The generated partial discharge simulation signal is highly consistent with the real original partial discharge signal in terms of time-frequency characteristic distribution, significantly improving the simulation accuracy and enhancing the reliability and accuracy of real defect detection and fault analysis of GIS equipment.

[0142] In this embodiment, S5 specifically includes:

[0143] S51. Receive the partial discharge simulation signal generated by the improved conditional generative adversarial network, and receive the corresponding original partial discharge signal to form a paired input structure of simulation signal and original signal.

[0144] S52. Input the paired input structure into the authenticity discrimination structure, perform feature extraction and difference analysis on the partial discharge simulation signal and the original partial discharge signal respectively, and obtain the authenticity probability value of the partial discharge simulation signal relative to the original partial discharge signal.

[0145] S53. Based on the information of the corresponding real defect type identifier node in the real defect feature database, construct a self-supervised classification structure for defect types, input the partial discharge simulation signal into the self-supervised classification structure for defect types, and determine the predicted value of the defect type corresponding to the partial discharge simulation signal.

[0146] S54. The true / false probability value and the defect type prediction value are compared with the corresponding original partial discharge signal and the real defect type identification node information, respectively. The discrimination error that can characterize the degree of difference between the true and false partial discharge simulation signal and the classification error that can characterize the degree of difference between the defect type prediction are obtained by calculation.

[0147] S55. Output the discrimination error and classification error as feedback for the improved conditional generative adversarial network to update network parameters in real time.

[0148] In this embodiment, the authenticity discrimination structure specifically includes a feature extraction network and an authenticity classification network. The feature extraction network uses a three-layer one-dimensional convolutional network with a fixed kernel size of 5 and a fixed output feature dimension of 256 for each layer. The authenticity classification network uses a two-layer fully connected network with 128 nodes and 1 node per layer, and the last layer uses the Sigmoid activation function to output a authenticity probability value between 0 and 1. The defect type self-supervised classification structure specifically includes a feature mapping network and a classification output network. The feature mapping network uses a two-layer one-dimensional convolutional network with a fixed kernel size of 5 and a 256-dimensional output dimension for each layer. The classification output network consists of two fully connected layers with 128 nodes and 4 nodes per layer, and the output layer uses the Softmax activation function to achieve classification and prediction of four real defect types: conductive particle defects, insulator tip burr defects, gas gap defects, and metal particle defects. The discrimination error and classification error are calculated using the mean squared error function. By calculating the error between the true and false probability values ​​and the true label, as well as the error between the classification prediction probability value and the true type label, a quantified error result with a clear numerical value is obtained.

[0149] This embodiment, through the above technical solution, achieves accurate identification of the authenticity of partial discharge simulation signals and precise classification of defect types. The generated error information can provide accurate real-time feedback to the generation network, greatly improving the consistency and accuracy between partial discharge simulation signals and actual defect signals, and significantly enhancing the reliability of defect simulation analysis.

[0150] In this embodiment, S6 specifically includes:

[0151] S61. Receive the discrimination error calculated by comparing the partial discharge simulation signal with the original partial discharge signal, and the classification error calculated by comparing the defect type prediction value of the partial discharge simulation signal with the corresponding real defect type identifier node information.

[0152] S62. Construct a real-time feedback dynamic optimization objective function based on the discrimination error and classification error, and use the objective function to characterize the degree of difference between the network parameters of the improved conditional generative adversarial network and the original partial discharge signal;

[0153] S63. The optimization direction of the network parameters is determined by the real-time feedback dynamic optimization objective function, and the update gradient of the network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network is calculated based on the gradient descent algorithm.

[0154] S64. The network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network are dynamically adjusted according to the calculated update gradient to form an adaptive update process of the parameters of the partial discharge simulation signal.

[0155] S65. Repeat the above S61 to S64 processes until the value of the real-time feedback dynamic optimization objective function reaches the preset convergence threshold, thus completing the dynamic optimization of the network parameters of the improved conditional generative adversarial network.

[0156] In this embodiment, the real-time feedback dynamic optimization objective function is specifically a weighted combination of discrimination error and classification error. The discrimination error is calculated using the mean square error between the simulated partial discharge signal and the original partial discharge signal, while the classification error is calculated using the cross-entropy error between the predicted defect type value and the actual defect type identifier node. The weight ratio of the two errors is fixed at 1:1. The network parameters are updated using the Adam optimization algorithm, with the initial learning rate of the gradient descent algorithm fixed at 0.001. During each iteration, the update gradients of all trainable network parameters for both the generator and discriminator networks are calculated and updated accordingly. The update gradients are obtained through backpropagation, and the gradient update process iterates using a batch size of 32 data points until the objective function value drops below 0.001, at which point the optimization process terminates. The initial values ​​of the network parameters are initialized using a random normal distribution. The updated network parameters automatically replace the network parameters from the previous iteration and are used in the next iteration for further optimization.

[0157] This embodiment achieves precise and dynamic adjustment of the parameters of the generation network and the discrimination network through the above technical solution, ensuring that the time-frequency characteristics of the simulated signal are highly matched with the actual defect signal, greatly improving the simulation accuracy and generalization ability of partial discharge defect signal, and enabling the system to have high reliability and stability.

[0158] Example 1:

[0159] To verify the feasibility and practical application effect of this invention, it was applied in an electrical engineering laboratory at a university for conducting simulation experiments on partial discharge faults in GIS equipment and for training professionals. Since traditional GIS equipment fault diagnosis methods generally rely on manual experience and offline data analysis, they are difficult to accurately simulate and identify partial discharge phenomena caused by real internal defects in equipment in real time. Therefore, the laboratory urgently needs an intelligent experimental system that can highly fit the characteristics of partial discharge caused by various defects inside real GIS equipment to effectively improve experimental accuracy and training effectiveness.

[0160] The laboratory is equipped with a typical 220kV GIS experimental platform, including switch units, busbar connection devices, and supporting SF6 gas insulation structures. During implementation, based on the GIS equipment's structural dimensions, internal insulation design requirements, and actual operating conditions, four typical real-world defect components—conductive particles, metal particles, insulator tip burrs, and gas gaps—were selected and fabricated. These components were placed in specific insulating spatial locations within the GIS equipment, and precise positioning and calibration were performed using a high-precision three-dimensional positioning device to ensure that the positional relationship between the defect components and the GIS equipment's insulators and conductors conformed to actual operating conditions.

[0161] During the experiment, the GIS equipment was set to its rated voltage (220kV), and the SF6 gas pressure was controlled at 0.5MPa. A high-frequency sensor array was used to synchronously acquire partial discharge signals from multiple angles for each type of real defect, obtaining a large number of raw partial discharge signals. Data channels corresponding to each real defect type were then established. These data channels were then input into an analysis algorithm based on adaptive short-time Fourier transform, and a database of real defect features was generated through refined feature extraction techniques.

[0162] Based on a database of real defect features, an attention mechanism was used to adaptively modulate and digitally encode the features. A feature fusion mapping structure was then used to form modulated defect features and corresponding real defect type codes. These features and codes were subsequently input into an improved conditional generative adversarial network (GAN). The network, utilizing a deep convolutional-deconvolutional structure and a multi-head self-attention mechanism, generated simulated signals highly similar to the discharge phenomena observed in actual GIS equipment. A joint discriminant network structure performed authenticity verification and self-supervised defect type classification on the generated simulated signals and the actual acquired raw signals. The discrimination error and classification error were calculated, and the network parameters were dynamically updated through real-time feedback to achieve accurate fitting between the simulated partial discharge signal and the original signal.

[0163] To quantify the experimental results and compare the performance improvement of the system of the present invention, the experimental data of partial discharge fault diagnosis before and after the implementation of the system are compared in the following table:

[0164] Table 1: Performance Comparison of GIS Partial Discharge Fault Simulation Experiments

[0165]

[0166] As shown in Table 1 above, after adopting the GIS partial discharge fault simulation experimental system based on real defects of the present invention, the accuracy rate of partial discharge defect type identification in the laboratory GIS increased from 78.6% to 98.5%, and the fitting degree between the simulated partial discharge signal and the original signal also significantly improved from 75.2% to 97.8%. Simultaneously, the system of the present invention shortened the network training convergence time by approximately 5.5 hours compared to the traditional system (from 7 hours to 1.5 hours), and the defect diagnosis response time also decreased significantly from 25 minutes in the traditional system to 2 minutes. Furthermore, the system of the present invention significantly improved the utilization rate of experimental data and the repeatability and stability of the experiment, making the experimental results more reliable and valuable for reference.

[0167] This invention significantly improves laboratory personnel's understanding and mastery of real-world defect characteristics by accurately simulating partial discharge phenomena under actual operating conditions of GIS equipment. Simultaneously, the high-precision, high-efficiency, and rapid-response partial discharge fault diagnosis training tool provided by the experimental system of this invention significantly improves the effectiveness of experimental training, substantially enhances the overall fault diagnosis capabilities of the laboratory and the technical level of personnel, and ensures the high reliability of experimental data and the effectiveness of experimental training results.

[0168] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A GIS partial discharge fault simulation experimental system based on real defects, characterized in that, Includes the following modules: The real defect signal acquisition module is used to acquire and output the original partial discharge signals that correspond one by one to various real defect types under the operating conditions of GIS equipment; The signal adaptive feature extraction module is used to receive the original partial discharge signal and perform adaptive time-frequency feature decomposition, output high-dimensional features corresponding to each real defect type, and construct a real defect feature database. The multidimensional feature modulation and coding module is used to extract high-dimensional features corresponding to the real defect type from the real defect feature database, and to perform feature modulation on the high-dimensional features based on the attention mechanism, and output the modulated defect features and the real defect type code. The Conditional Generative Adversarial Network (CGAN) Precision Simulation Module is used to receive the modulated defect features and the actual defect type encoding, and generate a partial discharge simulation signal that is consistent with the time-frequency domain characteristic distribution of the original partial discharge signal based on the improved CGAN architecture. The joint discrimination and defect self-supervised classification module is used to simultaneously perform authenticity identification and self-supervised classification on the partial discharge simulation signal and the original partial discharge signal, and output the discrimination error of the partial discharge simulation signal and the defect classification result. The real-time feedback dynamic optimization module is used to receive the discrimination error and defect classification results, dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, and achieve adaptive fitting between the partial discharge simulation signal and the original partial discharge signal distribution.

2. The GIS partial discharge fault simulation experimental system based on real defects according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect the original partial discharge signals of GIS equipment under various real defect types, and establish data channels for the original partial discharge signals according to the defect type. S2. Based on the adaptive time-frequency analysis algorithm, high-dimensional features are extracted from each data channel in the original partial discharge signal, and a real defect feature database is generated through feature decomposition and reconstruction. S3. Based on the real defect feature database, the high-dimensional features of each real defect type are adaptively modulated using an attention mechanism, and encoded in combination with the corresponding real defect type information in the real defect feature database to form modulated defect features and real defect type codes. S4. Based on the modulated defect features and the encoding of the real defect type, a feature attention modulation mechanism is embedded to construct an improved conditional generative adversarial network architecture. Through adversarial training, a partial discharge simulation signal with the same time-frequency characteristic distribution as the original partial discharge signal is generated. S5. Based on the partial discharge simulation signal and the original partial discharge signal, perform self-supervised classification of simulation signal authenticity and defect type, calculate and output the discrimination error and classification result; S6. Based on the discrimination error and classification result, a real-time feedback dynamic optimization model is constructed to adaptively and dynamically optimize the network parameters of the improved conditional generative adversarial network architecture, so that the partial discharge simulation signal is adaptively fitted to the feature distribution of the original partial discharge signal.

3. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S1 specifically includes: S11. Based on the structural and internal insulation design requirements of the GIS equipment, determine the structural dimensions, material, and placement parameters of the actual defective components; S12. Based on the actual defective component's structural dimensions, material, and placement parameters, fabricate actual defective components such as conductive microparticles, metal particles, insulator tip burrs, and gas gap defects. S13. Install the actual defective component in the predetermined position in the GIS equipment, and use a three-dimensional positioning device to calibrate the spatial coordinate position of the defective component relative to the insulator and conductor inside the GIS equipment. S14. Set the experimental voltage and SF6 gas pressure parameters according to the rated operating conditions of the GIS equipment, start the GIS equipment, and use a high-frequency sensor array to synchronously collect the original partial discharge signals generated by the discharge of real defective components from multiple angles. S15. The original partial discharge signals synchronously acquired from multiple angles by the high-frequency sensor array are digitally converted and labeled with the corresponding real defect types. Independent data channels are established according to the labeled real defect types to form data channels that correspond one-to-one with each real defect type.

4. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S2 specifically includes: S21. Input the data channels of the original partial discharge signal into the time-frequency analysis algorithm based on adaptive short-time Fourier transform respectively; S22. Perform adaptive short-time Fourier transform on the original partial discharge signal in each data channel to obtain time-frequency domain feature parameters that can accurately characterize the corresponding real defect type, and form a high-dimensional feature set. S23. Construct a feature decomposition and reconstruction mapping structure based on the high-dimensional feature set, set a shared representation of the mapping relationship between the input channel of the high-dimensional feature set and the corresponding real defect type identification channel, and each output channel is composed of a nonlinear mapping structure of the time-frequency domain feature parameters, and is connected to the corresponding real defect type identification node at the output end respectively; S24. Two optimization processing steps are set for the high-dimensional feature set: feature correlation analysis and feature redundancy removal. By removing redundant feature parameters and enhancing core feature parameters, an optimized high-dimensional feature set is formed. S25. Based on the optimized high-dimensional feature set, establish the data structure of the real defect feature database, and map the optimized high-dimensional feature set to the corresponding real defect type identifier nodes one by one and store them in the real defect feature database.

5. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S3 specifically includes: S31. Based on the mapping relationship between the optimized high-dimensional feature set stored in the real defect feature database and the corresponding real defect type identifier node, construct the feature adaptive modulation structure of the attention mechanism, wherein the real defect type identifier node is the basis for calculating the feature modulation weight. S32. In the feature adaptive modulation structure, a feature attention weight calculation unit is established to calculate the attention weight of each feature dimension based on the mapping relationship between the input optimized high-dimensional feature set and the corresponding real defect type identification node. S33. Adaptively weighted modulation is applied to each feature dimension in the optimized high-dimensional feature set using the calculated attention weights to generate modulated defect features. S34. Based on the real defect type identifier node corresponding to the real defect feature database, construct a real defect type digital encoding unit to obtain a real defect type digital encoding that corresponds one-to-one with the modulated defect feature. S35. Establish a feature fusion mapping structure to fuse the modulated defect features with the corresponding real defect types in digital encoding to form a unified modulated defect feature and real defect type encoding.

6. The GIS partial discharge fault simulation experimental system based on real defects according to claim 5, characterized in that, The feature attention weight calculation unit includes: The system receives the optimized high-dimensional feature set and the corresponding real defect type identifier node, and performs linear mapping processing on the input high-dimensional feature set to form the corresponding query feature vector, key feature vector and value feature vector. The importance coefficients of each feature dimension are calculated based on the degree of correlation between the query feature vector and the key feature vector, thus obtaining the feature correlation coefficients that can reflect the intrinsic feature differences of the high-dimensional feature set. The feature correlation coefficient is used to perform a dimension-by-dimensional weighted fusion of the feature dimensions in the value feature vector to generate feature attention weights; Based on the calculation deviation between the actual defect type identification node and the feature attention weight, the network parameters in the linear mapping process are dynamically adjusted using the error backpropagation method to optimize the calculation accuracy of the feature attention weight in real time.

7. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S4 specifically includes: S41. The modulated defect features and the real defect type encoding are used as conditional input sources, and the real defect type encoding is used as conditional control information. S42. The modulated defect features are input to the input of the generator network of the improved conditional generative adversarial network. Through the feature attention modulation mechanism, the modulated defect features are adaptively modulated with feature weights in each dimension to form a generated feature vector with time-frequency characteristic distribution information. S43. Establish a corresponding real defect type condition constraint structure based on the real defect type encoding, and generate a corresponding initial partial discharge simulation signal based on the generated feature vector and the real defect type condition constraint structure. S44. Input the initial partial discharge simulation signal to the input terminal of the discriminant network of the improved conditional generative adversarial network, and at the same time introduce the corresponding original partial discharge signal from the data channel to perform a true / false adversarial identification between the initial partial discharge simulation signal and the original partial discharge signal. S45. Based on the discrimination result of the discrimination network, determine the difference in time-frequency characteristic distribution between the initial partial discharge simulation signal and the original partial discharge signal, calculate the difference loss function value, and update the network parameters of the adversarial network by reversing the improvement conditions through the difference loss function value. S46. Repeat S42 to S45 until the difference loss function value reaches the preset convergence threshold, and output a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal.

8. The GIS partial discharge fault simulation experimental system based on real defects according to claim 7, characterized in that, The improved conditional generative adversarial network specifically includes: A feature attention modulation generative network is constructed. The modulated defect features are fused with the corresponding real defect type encoding and used as the network input. The feature weights are dynamically allocated dimension by dimension using a multi-head self-attention mechanism to represent the time-frequency domain distribution features corresponding to the real defect type. In the generative network, a feature dimension expansion layer and a deep convolution-deconvolution network are constructed to progressively transform the feature vectors modulated by feature attention into a partial discharge simulation signal that is consistent with the time-frequency characteristic distribution of the original partial discharge signal. A dual-task joint discrimination network structure is constructed. The bottom layer adopts a shared feature extraction structure, while the upper layer sets up independent authenticity discrimination structure and defect type classification structure, which respectively output the authenticity judgment probability value and defect type prediction value of the partial discharge simulation signal. A feature interaction attention mechanism is embedded between the shared feature extraction structure and the defect type classification structure to explicitly interact the feature association relationship between the partial discharge simulation signal and the real defect type; Based on the authenticity probability value and defect type prediction value output by the discrimination network, the simulation error is calculated and fed back to the generation network in real time, and the parameters of the generation network are dynamically updated using the gradient backpropagation method.

9. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S5 specifically includes: S51. Receive the partial discharge simulation signal generated by the improved conditional generative adversarial network, and receive the corresponding original partial discharge signal to form a paired input structure of simulation signal and original signal. S52. Input the paired input structure into the authenticity discrimination structure, perform feature extraction and difference analysis on the partial discharge simulation signal and the original partial discharge signal respectively, and obtain the authenticity probability value of the partial discharge simulation signal relative to the original partial discharge signal. S53. Based on the information of the corresponding real defect type identifier node in the real defect feature database, construct a self-supervised classification structure for defect types, input the partial discharge simulation signal into the self-supervised classification structure for defect types, and determine the predicted value of the defect type corresponding to the partial discharge simulation signal. S54. The true / false probability value and the defect type prediction value are compared with the corresponding original partial discharge signal and the real defect type identification node information, respectively. The discrimination error that can characterize the degree of difference between the true and false partial discharge simulation signal and the classification error that can characterize the degree of difference between the defect type prediction are obtained by calculation. S55. Output the discrimination error and classification error as feedback for the improved conditional generative adversarial network to update network parameters in real time.

10. The GIS partial discharge fault simulation experimental system based on real defects according to claim 2, characterized in that, S6 specifically includes: S61. Receive the discrimination error calculated by comparing the partial discharge simulation signal with the original partial discharge signal, and the classification error calculated by comparing the defect type prediction value of the partial discharge simulation signal with the corresponding real defect type identifier node information. S62. Construct a real-time feedback dynamic optimization objective function based on the discrimination error and classification error, and use the objective function to characterize the degree of difference between the network parameters of the improved conditional generative adversarial network and the original partial discharge signal; S63. The optimization direction of the network parameters is determined by the real-time feedback dynamic optimization objective function, and the update gradient of the network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network is calculated based on the gradient descent algorithm. S64. The network parameters of the generator network and the discriminator network in the improved conditional generative adversarial network are dynamically adjusted according to the calculated update gradient to form an adaptive update process of the parameters of the partial discharge simulation signal. S65. Repeat the above S61 to S64 processes until the value of the real-time feedback dynamic optimization objective function reaches the preset convergence threshold, thus completing the dynamic optimization of the network parameters of the improved conditional generative adversarial network.

Images