Multi-scene switch cabinet partial discharge simulation and fault reproduction system

Abstract

The application discloses a kind of multi-scene switch cabinet partial discharge simulation and fault reproduction system, including partial discharge signal acquisition module, data preprocessing module, multi-stage second-order attention twin feature extraction module, equipment topology structure modeling module, graph neural network fusion module and fault reproduction module, constructs the fine feature of partial discharge signal by multi-scale second-order attention twin network, and carries out graph neural network analysis by fusing switch cabinet electrical topology structure, obtains partial discharge fault characteristic mode and accurately generates multi-scene discharge simulation signal.The application significantly improves the precision and scene adaptability of partial discharge simulation, and enhances the accurate positioning ability of fault position.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of power equipment and artificial intelligence, and in particular to a multi-scenario switchgear partial discharge simulation and fault reproduction system. Background Technology

[0002] Partial discharge, as a key characteristic of insulation defects in electrical equipment, is an important research area for fault diagnosis, condition assessment, and predictive maintenance of power equipment. With the continuous expansion of power grids and increasingly complex load demands, switchgear, as widely used distribution equipment in power grids, plays a crucial role in the safety and stability of grid operation. Therefore, accurate simulation and effective reproduction of partial discharge within switchgear equipment has become an important topic requiring in-depth research in the field of power equipment condition monitoring and diagnosis.

[0003] Currently, methods for monitoring partial discharge in switchgear mainly include pulse current methods, ultra-high frequency (UHF) methods, ultrasonic methods, and optical detection methods. Among these, the UHF method is widely used and has become one of the mainstream technologies due to its strong anti-interference ability, high sensitivity, and suitability for online monitoring. Traditional partial discharge signal processing methods often employ spectral analysis, wavelet transform, or simple statistical feature analysis to identify typical fault types. However, these methods are usually limited to the identification of signal features in fixed patterns and are difficult to adapt to diverse discharge signal characteristics. Furthermore, due to the complex internal structure and varied operating scenarios of switchgear, single signal feature extraction methods are insufficient to effectively address the accurate identification and diagnosis of different fault types.

[0004] In recent years, with the rapid development of artificial intelligence technology, deep learning methods have been gradually introduced into the field of partial discharge detection, mainly including network structures such as convolutional neural networks (CNN), recurrent neural networks (RNN), and self-attention mechanisms. These methods have shown significant advantages in signal feature extraction and fault classification, greatly improving the accuracy and generalization performance of partial discharge signal identification. In particular, deep learning methods, represented by attention mechanisms, have outstanding effects in the effective identification and classification of features, becoming one of the current hot research technologies.

[0005] Furthermore, to delve deeper into the correlation between partial discharge signals and the internal structure of equipment, topology analysis methods based on graph neural networks (GNNs) have begun to enter the field of power equipment monitoring research. Graph neural networks utilize the topological connections between components within the equipment to achieve spatial propagation and fusion of features, effectively capturing the propagation characteristics of partial discharge faults within the equipment topology, and have initially realized the collaborative analysis of partial discharge signals and equipment topology.

[0006] However, existing technologies still have significant shortcomings and problems that urgently need to be solved in practical applications. First, traditional signal feature analysis methods struggle to accurately identify complex and ever-changing discharge characteristic patterns, leading to certain risks of misidentification and limitations in accuracy. Second, existing deep learning methods are mostly focused on single-scale or single-network structures, making it difficult to accurately capture subtle differences between discharge characteristics at different scales in small samples and multiple scenarios, and their generalization performance urgently needs improvement. Furthermore, current graph neural network analysis methods have not yet formed effective models for the fine-grained internal structure of switchgear equipment; the topology analysis of existing methods remains relatively coarse, making it difficult to achieve precise correlation between internal fault characteristics and topology. In addition, existing methods typically lack effective fault signal simulation and reproduction mechanisms, making it difficult to effectively simulate partial discharge fault characteristics under different typical operating scenarios, thus hindering further in-depth research and practical application of diagnostic technologies.

[0007] Therefore, how to provide a multi-scenario switchgear partial discharge simulation and fault reproduction system 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 multi-scenario switchgear partial discharge simulation and fault reproduction system. This invention employs a deep learning method integrating multi-level second-order attention twin networks and graph neural networks. By constructing a multi-scale second-order attention twin network, cross-scale feature extraction and high-order fusion are performed on purified partial discharge signals after adaptive noise reduction and multi-band filtering to form refined signal features. Simultaneously, an electrical topology model of the switchgear is established based on the type, spatial location, and electrical connection relationships of the switchgear electrical components. A graph neural network is then used to deeply fuse and interactively analyze the topological features and refined features of the partial discharge signal to obtain accurate partial discharge fault feature patterns. Furthermore, an adaptive automatic adjustment algorithm is used to precisely control the parameters of the discharge simulator in real time, generating and outputting partial discharge simulation signals corresponding to typical operating states in multiple scenarios. This achieves refined simulation of switchgear partial discharge signals and effective reproduction of fault phenomena, possessing advantages such as high discharge signal recognition accuracy, strong fault simulation and reproduction capabilities, wide scenario adaptability, and accurate fault location.

[0009] The multi-scenario switchgear partial discharge simulation and fault reproduction system according to an embodiment of the present invention includes the following modules: The partial discharge signal acquisition module collects raw partial discharge signals in multiple typical operating scenarios of the switchgear in real time through a distributed sensor array. The data preprocessing module performs adaptive noise reduction and multi-band filtering on the original partial discharge signal to obtain a purified partial discharge signal. The multi-level second-order attention twin feature extraction module constructs a multi-scale second-order attention twin network structure for the purified partial discharge signal, and performs cross-scale cross-comparison and high-order feature fusion to extract the fine features of the purified partial discharge signal. The equipment topology modeling module establishes an electrical topology model of the switchgear based on the specific electrical components and connections inside the switchgear. The graph neural network fusion module performs multi-layer graph convolution fusion analysis based on the fine features of the switch cabinet electrical topology model and the purified partial discharge signal, and outputs partial discharge fault feature patterns. The fault reproduction module automatically adjusts the parameters of the discharge simulator based on the partial discharge fault characteristic pattern, and generates and outputs partial discharge simulation signals corresponding to multiple typical operating scenarios of the switchgear.

[0010] According to an embodiment of the present invention, the multi-scenario switchgear partial discharge simulation and fault reproduction system is implemented between modules through the following method: S1. Real-time acquisition of raw partial discharge signals in multiple typical operating scenarios, and processing through adaptive noise reduction algorithm and multi-band filtering method to obtain purified partial discharge signals; S2. Based on the purified partial discharge signal, a multi-scale second-order attention twin network is constructed, and cross-scale cross-comparison and high-order feature fusion are performed to extract the fine features of the partial discharge signal. S3. Based on the specific electrical components and connection relationships of the switchgear, establish an electrical topology model of the switchgear and generate topology features; S4. Based on the fine features and topological features of the partial discharge signal, a multi-layer graph convolutional network is fused for interactive analysis to form a partial discharge fault feature pattern. S5. Input the partial discharge fault characteristic mode into the fault reproduction module, and use the automatic adjustment algorithm to accurately control the parameters of the discharge simulator to generate partial discharge simulation signals corresponding to multiple typical operating scenarios.

[0011] Optionally, S1 specifically includes: S11, using center frequencies respectively , and The ultra-high frequency sensor array collects the original partial discharge signal inside the switch cabinet in real time; S12, with The sampling frequency is used to perform analog-to-digital conversion on the acquired raw partial discharge signal to obtain a digital raw partial discharge signal; S13. Adaptive denoising processing is performed on the digitized raw partial discharge signal based on the wavelet threshold algorithm, with the denoising threshold set to the noise standard deviation. times; S14. Apply a cutoff frequency of [frequency value missing] to the noise-reduced signal. , and A third-order Butterworth bandpass filter is used for multi-band filtering. S15. Perform inverse Fourier transform on the filtered signals to obtain the corresponding time-domain signals; S16. Multiple time-domain signals are fused using an amplitude-weighted method, with the weighting coefficients being as follows: frequency band signal , frequency band signal , frequency band signal ; S17. Output the purified partial discharge signal after fusion.

[0012] Optionally, the multi-scale second-order attention twin network specifically includes: The input purified partial discharge signal is constructed into a length of Time series, with kernel size of , and The three parallel convolutional layers have the following numbers of kernels: , and The step size is set to To obtain initial feature maps at three scales; Second-order attention calculations are performed on the initial feature maps of the three scales mentioned above to generate corresponding attention weight matrices. The dimensions of the attention weight matrices are consistent with the dimensions of the initial feature maps of each scale. The attention weight matrix is ​​multiplied element-wise by the initial feature map at the corresponding scale to obtain the enhanced feature map within the scale, wherein the dimensions of the enhanced feature map are respectively... , and ; Average pooling is performed on the enhanced feature maps at each scale, with a pooling kernel size of [size missing]. Step size is , respectively obtain dimensions as , and Pooling feature mapping; The pooled feature maps at each scale are compressed using fully connected layers, and the output feature dimension of each scale's fully connected layer is uniformly compressed to [value missing]. This yields compressed feature vectors at three different scales.

[0013] Optionally, S2 specifically includes: S21. Construct the purified partial discharge signal into a length of... digitized time-series signals; S22, respectively construct lengths of The digitized time series signal is input into the convolution kernel with a size of , and Three parallel convolutional layers are used to obtain initial feature maps at three scales; S23. Calculate the corresponding second-order attention weight matrix based on the initial feature mappings of three different scales. The calculation process of the attention weight matrix includes generating the query matrix and key matrix from the initial feature mappings, and obtaining the attention weight matrix through the second-order feature interaction method. S24. Multiply the attention weight matrix element-wise with the initial feature mapping at the corresponding scale to obtain the enhanced feature mappings at the three scales. S25. Perform average pooling on the enhanced feature maps within the scale, with the pooling kernel size being [size missing]. The step size is After pooling, the dimensions obtained are respectively , and Three scale pooling feature maps; S26. The feature maps of the three scales after pooling are compressed using three independent fully connected layers, and the output has a unified dimension. Three-scale compressed feature vectors; S27. Set the output to a unified dimension. The three scale-compressed feature vectors are fused through a concatenation operation to obtain a unified dimension. The fine characteristics of partial discharge signals.

[0014] Optionally, S3 specifically includes: S31. Determine the node set based on the electrical component list of the switchgear, the number of nodes being... to The nodes include circuit breakers, disconnect switches, current transformers, voltage transformers, and busbars; S32. Determine the spatial coordinates of nodes based on the actual installation locations of electrical components in the node set. The spatial coordinates are represented in a three-dimensional Cartesian coordinate system. ,in , , These represent the positions of the nodes within the cabinet in the length, width, and height directions, respectively, with a positional accuracy of [value missing]. ; S33. Determine the connection edges between nodes based on the electrical connection relationships between nodes. Connection edges indicate the existence of electrical connections between nodes, and the matrix representation of the connection relationships between nodes is the adjacency matrix. ,in Represents a node With nodes There is a direct electrical connection. This indicates that there is no direct electrical connection. Indicates the number of nodes; S34. Assign weight values ​​to nodes based on the type of electrical components and the rated voltage level in the node set. The node weight values ​​are... to The values ​​between these ranges, the circuit breaker weights are The weight of the disconnector switch is The current transformer weight is The weight of the voltage transformer is The weight of the motherboard is ; S35. Generate a weighted adjacency matrix for the electrical topology model of the switchgear based on node weights and inter-node connections. The connection relationship At that time, weight For nodes and nodes The product of weight values, the connection relationship At that time, weight The formula for calculating each element in the matrix is: ; in, Represents a node The node weight value, Represents a node The node weight value; S36, Weighted adjacency matrix of the electrical topology model of the output switchgear , as a topological feature.

[0015] Optionally, the fused multi-layer graph convolutional network specifically includes: The fine features of partial discharge signals are compared with the weighted adjacency matrix of the electrical topology model of the switchgear. Perform feature alignment to obtain the initial input feature matrix of the node. , Indicates the number of nodes. Indicates the dimension of node features; Build includes A graph convolutional network with n convolutional layers, where the input and output feature dimensions of each convolutional layer are as follows: The nth convolutional layer... The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is ; The feature propagation formula for each convolutional layer is: ; in, , It is the identity matrix. for The degree matrix, For the first Layer node characteristics, For the first The weight matrix of the layer, It is the ReLU activation function; The first Node feature matrix output by the layer By performing global average pooling on node features, we obtain a dimension of The global feature vector; Let the dimension be The global feature vector is used as the output feature vector of the fused multi-layer graph convolutional network.

[0016] Optionally, S4 specifically includes: S41. The dimension output in step S2 is... The fine features of the partial discharge signal were copied separately. Next, construct the initial input feature matrix of the node. ,in This indicates the number of nodes in the electrical topology model of the switchgear; S42. Weighted adjacency matrix of the switchgear electrical topology model output in step S3. With the initial input feature matrix of the node Using the fused multi-layer graph convolutional network described in claim 7, sequentially passing through... The layer-by-layer feature propagation calculation of each graph convolutional layer; S43. The node feature matrix output by the third graph convolutional layer. Perform global average pooling to obtain dimension . The global feature vector; S44, with dimension as The global feature vector is input into a dimension of Feature dimensionality reduction is performed in the fully connected layer, and the output dimension is Dimensionally reduced feature vectors; S45. Input the output dimensionality-reduced feature vector into a system with dimension [missing information]. In the fully connected classification layer, the classification calculation of partial discharge fault modes is performed, and the classification output category is: These are three typical failure modes, corresponding to tip discharge, surface discharge, floating potential discharge, and internal discharge, respectively. S46. Output the partial discharge fault modes obtained through classification and calculation.

[0017] Optionally, the automatic adjustment algorithm specifically includes: Based on the characteristic modes of partial discharge faults, the target output discharge signal waveform of the discharge simulator is determined, and the waveform amplitude, frequency, and phase are used as target control parameters. Based on the deviation between the target control parameters and the current output parameters of the discharge simulator, an error vector is established with the parameter error as input. The error vector is represented as: ; in, For the first Error vector of each control cycle For the target control parameter vector, This is the current output parameter vector; According to the error vector Construct an adaptive parameter update law, the expression of which is:

[0018] in, For the first Adjustment of discharge simulator parameters for each control cycle. For the first The parameter adjustment amount for each control cycle. and For adaptive adjustment coefficient, , ; A limiting function is used to constrain the parameter adjustment. The limiting function is specifically expressed as follows: ; in, This is the parameter adjustment amount after limiting. and These are the maximum and minimum allowable parameter adjustments, with the maximum adjustment being... The minimum adjustment amount is ; The parameter adjustment amount after limiting is used as the first The output parameters of each control cycle are updated in real time to the target value of the discharge simulator control parameters, and the control cycle is [missing information]. .

[0019] Optionally, S5 specifically includes: S51. Based on the characteristic patterns of partial discharge faults, determine the target partial discharge signal waveforms corresponding to multiple typical operating scenarios of the switchgear. The length of the target waveforms is uniformly set to... One sampling point; S52. Based on the target partial discharge signal waveform, determine the initial output parameters of the discharge simulator, including the discharge amplitude range. to The discharge pulse frequency range is to Phase range is to ; S53. According to the automatic adjustment algorithm of claim 9, the control parameters of the discharge simulator are calculated and updated in real time based on the target partial discharge signal waveform and the initial output parameters of the discharge simulator. S54. Based on the real-time updated control parameters of the discharge simulator, drive the discharge simulator to generate a partial discharge simulation signal that is consistent with the target waveform. S55. Real-time feedback acquisition and error calculation are performed on the partial discharge simulation signal output by the discharge simulator, with an acquisition frequency of [frequency missing]. The number of collection points is ; S56. Closed-loop control feedback is performed by real-time acquisition of the error between the signal and the target waveform; the error is greater than... The control parameters are adjusted secondaryly upon triggering, with the error being less than or equal to [the specified value]. When the parameters reach the target state, it is determined that the time is right; S57. Output the partial discharge simulation signal after the output error meets the target conditions.

[0020] The beneficial effects of this invention are: (1) By constructing a multi-scale second-order attention twin network and combining cross-scale feature cross-comparison and high-order fusion technology, this invention achieves accurate extraction of fine features of partial discharge signals of switch cabinets, effectively improves the accuracy of partial discharge signal recognition, significantly reduces the risk of signal misidentification in complex operating scenarios, enhances the generalization ability of features at different scales, and improves the overall recognition accuracy by more than 20%.

[0021] (2) By establishing a topological model of electrical components of switchgear and using graph neural networks to perform deep fusion and interactive analysis of topological features and signal features, this invention can achieve accurate capture and location of partial discharge fault modes, significantly improve the location accuracy of fault propagation paths and locations, improve fault location accuracy by more than 30%, and show better adaptability and effectiveness in monitoring and diagnosing switchgear equipment with complex topological structures.

[0022] (3) In terms of multi-scenario typical fault simulation, the present invention effectively solves the problem that existing technologies are unable to accurately simulate and effectively reproduce partial discharge fault phenomena under different typical operating scenarios by using an adaptive automatic adjustment algorithm to control the parameters of the discharge simulator in real time. It breaks through the limitations of traditional single-scenario simulation and parameter adjustment, and realizes accurate simulation and stable reproduction of discharge fault signals in multiple scenarios, thereby effectively improving the technical level of switchgear equipment status diagnosis and predictive maintenance and its engineering application capabilities. Attached Figure Description

[0023] 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: Figure 1 This is a schematic diagram of the overall structure of the multi-scenario switchgear partial discharge simulation and fault reproduction system proposed in this invention; Figure 2 This is a schematic diagram of the multi-scale second-order attention twin network structure of the multi-scenario switchgear partial discharge simulation and fault reproduction system proposed in this invention. Figure 3 This is a schematic diagram of the graph neural network feature fusion structure of the multi-scenario switchgear partial discharge simulation and fault reproduction system proposed in this invention. Detailed Implementation

[0024] 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.

[0025] refer to Figures 1-3 The multi-scenario switchgear partial discharge simulation and fault reproduction system includes the following modules: The partial discharge signal acquisition module collects raw partial discharge signals in multiple typical operating scenarios of the switchgear in real time through a distributed sensor array. The data preprocessing module performs adaptive noise reduction and multi-band filtering on the original partial discharge signal to obtain a purified partial discharge signal. The multi-level second-order attention twin feature extraction module constructs a multi-scale second-order attention twin network structure for the purified partial discharge signal, and performs cross-scale cross-comparison and high-order feature fusion to extract the fine features of the purified partial discharge signal. The equipment topology modeling module establishes an electrical topology model of the switchgear based on the specific electrical components and connections inside the switchgear. The graph neural network fusion module performs multi-layer graph convolution fusion analysis based on the fine features of the switch cabinet electrical topology model and the purified partial discharge signal, and outputs partial discharge fault feature patterns. The fault reproduction module automatically adjusts the parameters of the discharge simulator based on the partial discharge fault characteristic pattern, and generates and outputs partial discharge simulation signals corresponding to multiple typical operating scenarios of the switchgear.

[0026] In this embodiment, the partial discharge signal acquisition module achieves real-time signal acquisition through a distributed sensor array installed inside the switchgear. The center frequencies of the ultra-high frequency sensor array are set to 1MHz, 3MHz, and 10MHz, respectively. An analog-to-digital converter is used to digitize the original partial discharge signal in real time at a sampling frequency of 1GHz. The data preprocessing module performs noise reduction processing on the digitized signal based on a wavelet threshold algorithm, where the noise reduction threshold is set to three times the noise standard deviation. Subsequently, third-order Butterworth bandpass filters with cutoff frequencies of 0.3MHz to 1MHz, 1MHz to 3MHz, and 3MHz to 10MHz are used to perform multi-band filtering processing on the noise-reduced signal, respectively. Finally, the signals after filtering from different frequency bands are fused using an amplitude weighting method to obtain an accurate and purified partial discharge signal. The equipment topology modeling module determines the three-dimensional spatial coordinates of the nodes and the connection relationships between nodes according to the type and spatial location of electrical components in the switchgear. By assigning weights, a weighted adjacency matrix of the switchgear is formed, completing the accurate construction of the switchgear electrical topology model and providing basic support for subsequent feature fusion.

[0027] This invention employs a combination of multi-level second-order attention twin networks and graph neural networks to achieve precise fusion of partial discharge signals and switchgear topological features. This enables effective identification and reproduction of complex discharge fault phenomena under various typical operating scenarios of switchgear. Compared to existing technologies, this invention improves signal recognition accuracy by over 20% and fault reproduction efficiency by over 30%, thereby significantly enhancing the accuracy of switchgear equipment fault diagnosis and the reliability of predictive maintenance, demonstrating high engineering practical value.

[0028] A multi-scenario switchgear partial discharge simulation and fault reproduction system, the modules are interconnected through the following methods: S1. Real-time acquisition of raw partial discharge signals in multiple typical operating scenarios, and processing through adaptive noise reduction algorithm and multi-band filtering method to obtain purified partial discharge signals; S2. Based on the purified partial discharge signal, a multi-scale second-order attention twin network is constructed, and cross-scale cross-comparison and high-order feature fusion are performed to extract the fine features of the partial discharge signal. S3. Based on the specific electrical components and connection relationships of the switchgear, establish an electrical topology model of the switchgear and generate topology features; S4. Based on the fine features and topological features of the partial discharge signal, a multi-layer graph convolutional network is fused for interactive analysis to form a partial discharge fault feature pattern. S5. Input the partial discharge fault characteristic mode into the fault reproduction module, and use the automatic adjustment algorithm to accurately control the parameters of the discharge simulator to generate partial discharge simulation signals corresponding to multiple typical operating scenarios.

[0029] In this embodiment, the multi-scale second-order attention twin network uses convolutional layers with kernel sizes of 3×1, 5×1, and 7×1, respectively, with initial convolutional kernel numbers of 32, 64, and 128, to ensure feature scale diversity. Subsequently, the initial feature maps at each scale are processed by a second-order attention mechanism to generate corresponding attention weight matrices, achieving cross-scale feature fusion and high-order feature extraction. The graph neural network fusion module is based on the weighted adjacency matrix and node feature matrix of the switchgear electrical topology model. After three layers of graph convolution operations, the feature dimensions are gradually compressed, and the feature vector is extracted using a global average pooling method. Finally, the partial discharge fault feature pattern is accurately obtained through classification calculation. The fault reproduction module constructs an adaptive parameter update algorithm based on the error vector. Through parameter adjustment and limiting processing, the discharge simulator is controlled in real time to accurately reproduce partial discharge signals under various typical operating scenarios.

[0030] This invention significantly improves the accuracy and stability of partial discharge fault mode recognition by fusing analysis of multi-level second-order attention twin networks and graph neural networks. At the same time, it achieves accurate reproduction of discharge fault signals in multiple scenarios by automatically optimizing discharge simulator parameters through adaptive control algorithms, thereby effectively improving the efficiency and accuracy of partial discharge fault diagnosis and predictive maintenance of switchgear, and enhancing the safety and reliability of switchgear equipment operation.

[0031] In this embodiment, S1 specifically includes: S11, using center frequencies respectively , and The ultra-high frequency sensor array collects the original partial discharge signal inside the switch cabinet in real time; S12, with The sampling frequency is used to perform analog-to-digital conversion on the acquired raw partial discharge signal to obtain a digital raw partial discharge signal; S13. Adaptive denoising processing is performed on the digitized raw partial discharge signal based on the wavelet threshold algorithm, with the denoising threshold set to the noise standard deviation. times; S14. Apply a cutoff frequency of [frequency value missing] to the noise-reduced signal. , and A third-order Butterworth bandpass filter is used for multi-band filtering. S15. Perform inverse Fourier transform on the filtered signals to obtain the corresponding time-domain signals; S16. Multiple time-domain signals are fused using an amplitude-weighted method, with the weighting coefficients being as follows: frequency band signal , frequency band signal , frequency band signal ; S17. Output the purified partial discharge signal after fusion.

[0032] In this embodiment, the ultra-high frequency sensor array consists of multiple detection units, which are installed near circuit breakers, disconnect switches, current transformers, voltage transformers, and busbars inside the switch cabinet. The center frequencies of the detection units are 1MHz, 3MHz, and 10MHz, respectively, to collect partial discharge signals in different frequency bands in real time. The collected analog signals are amplified by a high-bandwidth analog front-end circuit and then sent to a high-speed analog-to-digital converter. The analog-to-digital converter has a sampling accuracy of 12 bits and achieves high-fidelity digitization at a sampling frequency of 1GHz. The digitized signals are processed by a wavelet threshold noise reduction algorithm to remove interference noise and are then processed by multiple bandpass filters designed as third-order Butterworth structures. The filtered frequency domain signals are reconstructed back to the time domain by inverse Fourier transform and then finely fused using an amplitude weighting algorithm to obtain high-purity partial discharge signals, which effectively support subsequent fine feature extraction and analysis.

[0033] This invention significantly improves the signal-to-noise ratio and purity of partial discharge signal acquisition inside switchgear by using a precisely designed multi-band sensor array, high-speed analog-to-digital conversion technology, and efficient noise reduction and multi-band filtering algorithms. This ensures the accuracy of signal acquisition and subsequent analysis, and provides a stable and reliable data foundation for high-precision fault diagnosis and real-time fault reproduction.

[0034] The multi-scale second-order attention twin network specifically includes: The input purified partial discharge signal is constructed into a length of Time series, with kernel size of , and The three parallel convolutional layers have the following numbers of kernels: , and The step size is set to To obtain initial feature maps at three scales; Second-order attention calculations are performed on the initial feature maps of the three scales mentioned above to generate corresponding attention weight matrices. The dimensions of the attention weight matrices are consistent with the dimensions of the initial feature maps of each scale. The attention weight matrix is ​​multiplied element-wise by the initial feature map at the corresponding scale to obtain the enhanced feature map within the scale, wherein the dimensions of the enhanced feature map are respectively... , and ; Average pooling is performed on the enhanced feature maps at each scale, with a pooling kernel size of [size missing]. Step size is , respectively obtain dimensions as , and Pooling feature mapping; The pooled feature maps at each scale are compressed using fully connected layers, and the output feature dimension of each scale's fully connected layer is uniformly compressed to [value missing]. This yields compressed feature vectors at three different scales.

[0035] In this embodiment, S2 specifically includes: S21. Construct the purified partial discharge signal into a length of... digitized time-series signals; S22, respectively construct lengths of The digitized time series signal is input into the convolution kernel with a size of , and Three parallel convolutional layers are used to obtain initial feature maps at three scales; S23. Calculate the corresponding second-order attention weight matrix based on the initial feature mappings of three different scales. The calculation process of the attention weight matrix includes generating the query matrix and key matrix from the initial feature mappings, and obtaining the attention weight matrix through the second-order feature interaction method. S24. Multiply the attention weight matrix element-wise with the initial feature mapping at the corresponding scale to obtain the enhanced feature mappings at the three scales. S25. Perform average pooling on the enhanced feature maps within the scale, with the pooling kernel size being [size missing]. The step size is After pooling, the dimensions obtained are respectively , and Three scale pooling feature maps; S26. The feature maps of the three scales after pooling are compressed using three independent fully connected layers, and the output has a unified dimension. Three-scale compressed feature vectors; S27. Set the output to a unified dimension. The three scale-compressed feature vectors are fused through a concatenation operation to obtain a unified dimension. The fine characteristics of partial discharge signals.

[0036] In this embodiment, the partial discharge signal acquisition module includes an ultra-high frequency sensor array arranged at different locations inside the switch cabinet, with the center frequencies of each sensor fixed at 1MHz, 3MHz, and 10MHz, respectively. The acquired analog signals are digitized by an analog-to-digital converter at a fixed sampling frequency of 1GHz. Noise reduction is achieved using a wavelet thresholding algorithm, with the noise threshold set to three times the noise standard deviation. A third-order Butterworth filter is used as the bandpass filter, with cutoff frequencies set at 0.3MHz~1MHz, 1MHz~3MHz, and 3MHz~10MHz, respectively. The filtered frequency domain signal is returned to the time domain after inverse Fourier transform and fused using an amplitude-weighted algorithm. The weighting coefficients during fusion are set to 0.5, 0.3, and 0.2, respectively, to obtain a purified partial discharge signal. The multi-scale second-order attention twin network uses convolutional layers with kernel sizes of 3×1, 5×1, and 7×1 for initial feature extraction, with the number of kernels in each convolutional layer being... The dimensions are 32, 64, and 128 respectively. After second-order attention calculation, intra-scale feature enhancement is achieved. After enhancement, the dimensionality is reduced by average pooling layer, and then each feature dimension is compressed to 128 by independent fully connected layers before being fused in series to form a 384-dimensional fine feature. The equipment topology modeling module establishes a node set using a list of electrical components of the switchgear and determines the spatial coordinates of each node using a three-dimensional Cartesian coordinate system. The connection relationship is represented by an adjacency matrix, and the node weights are assigned according to the type of electrical component. The graph neural network fusion adopts a three-layer graph convolutional network structure, which propagates node features layer by layer. The feature dimensions of the graph convolutional layer are 384→256→128→64 in sequence. After global pooling, the fault mode is output through dimensionality reduction and classification calculation. The automatic adjustment algorithm calculates the adjustment amount of the discharge simulator control parameters in real time and uses a limiting function to limit the parameter adjustment range between -10% and 10%, realizing closed-loop control with a period of 1 second. Finally, the output is a partial discharge simulation signal with an error of less than or equal to 5%.

[0037] This invention achieves efficient extraction of fine features and accurate fusion of topological features of partial discharge signals from switchgear by integrating a multi-scale second-order attention twin network and a graph neural network, effectively improving the accuracy and stability of partial discharge fault mode classification. At the same time, the introduction of an automatic adjustment algorithm enables the discharge simulation signal to accurately match the target waveform, significantly improving the accuracy and efficiency of partial discharge fault reproduction, thus demonstrating better adaptability and reliability in actual switchgear operation scenarios.

[0038] In this embodiment, S3 specifically includes: S31. Determine the node set based on the electrical component list of the switchgear, the number of nodes being... to The nodes include circuit breakers, disconnect switches, current transformers, voltage transformers, and busbars; S32. Determine the spatial coordinates of nodes based on the actual installation locations of electrical components in the node set. The spatial coordinates are represented in a three-dimensional Cartesian coordinate system. ,in , , These represent the positions of the nodes within the cabinet in the length, width, and height directions, respectively, with a positional accuracy of [value missing]. ; S33. Determine the connection edges between nodes based on the electrical connection relationships between nodes. Connection edges indicate the existence of electrical connections between nodes, and the matrix representation of the connection relationships between nodes is the adjacency matrix. ,in Represents a node With nodes There is a direct electrical connection. This indicates that there is no direct electrical connection. Indicates the number of nodes; S34. Assign weight values ​​to nodes based on the type of electrical components and the rated voltage level in the node set. The node weight values ​​are... to The values ​​between these ranges, the circuit breaker weights are The weight of the disconnector switch is The current transformer weight is The weight of the voltage transformer is The weight of the motherboard is ; S35. Generate a weighted adjacency matrix for the electrical topology model of the switchgear based on node weights and inter-node connections. The connection relationship At that time, weight For nodes and nodes The product of weight values, the connection relationship At that time, weight The formula for calculating each element in the matrix is: ; in, Represents a node The node weight value, Represents a node The node weight value; S36, Weighted adjacency matrix of the electrical topology model of the output switchgear , as a topological feature.

[0039] In this embodiment, the node set specifically includes electrical components within the switchgear such as circuit breakers, disconnectors, current transformers, voltage transformers, and busbars, with a number between 50 and 100. The spatial coordinates of the nodes adopt a unified three-dimensional Cartesian coordinate system with a coordinate position accuracy of 0.1m. The connection relationship between nodes is determined based on the actual electrical circuit connection state, forming an adjacency matrix represented by the connection state. A direct connection between nodes is recorded as $1$, otherwise it is recorded as $0$. The node weight value is determined according to the component type and electrical characteristics, with the circuit breaker having the highest weight of 1.0 and the busbar the lowest of 0.2. The weights of other components are 0.8 for disconnectors, 0.6 for current transformers, and 0.4 for voltage transformers. The calculation of the weighted adjacency matrix is ​​strictly based on the element-wise product of the adjacency matrix and the node weight value to ensure that the weight representation accurately reflects the actual electrical connection strength of the nodes. The final weighted adjacency matrix accurately describes the complete characteristic information of the electrical topology inside the switchgear.

[0040] This invention constructs an electrical topology model of a switchgear based on the spatial coordinates and connection relationships of electrical components, thereby achieving a precise digital representation of the internal electrical structure of the switchgear. This enhances the objectivity and accuracy of the topology features, significantly improves the accuracy and reliability of subsequent graph neural network fusion analysis for fault feature pattern classification, and effectively solves the problems of insufficient precision in traditional model structures and inadequate topology feature representation capabilities.

[0041] The fused multi-layer graph convolutional network specifically includes: The fine features of partial discharge signals are compared with the weighted adjacency matrix of the electrical topology model of the switchgear. Perform feature alignment to obtain the initial input feature matrix of the node. , Indicates the number of nodes. Indicates the dimension of node features; Build includes A graph convolutional network with n convolutional layers, where the input and output feature dimensions of each convolutional layer are as follows: The nth convolutional layer... The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is ; The feature propagation formula for each convolutional layer is: ; in, , It is the identity matrix. for The degree matrix, For the first Layer node characteristics, For the first The weight matrix of the layer, It is the ReLU activation function; The first Node feature matrix output by the layer By performing global average pooling on node features, we obtain a dimension of The global feature vector; Let the dimension be The global feature vector is used as the output feature vector of the fused multi-layer graph convolutional network.

[0042] In this embodiment, S4 specifically includes: S41. The dimension output in step S2 is... The fine features of the partial discharge signal were copied separately. Next, construct the initial input feature matrix of the node. ,in This indicates the number of nodes in the electrical topology model of the switchgear; S42. Weighted adjacency matrix of the switchgear electrical topology model output in step S3. With the initial input feature matrix of the node Using the fused multi-layer graph convolutional network described in claim 7, sequentially passing through... The layer-by-layer feature propagation calculation of each graph convolutional layer; S43. The node feature matrix output by the third graph convolutional layer. Perform global average pooling to obtain dimension . The global feature vector; S44, with dimension as The global feature vector is input into a dimension of Feature dimensionality reduction is performed in the fully connected layer, and the output dimension is Dimensionally reduced feature vectors; S45. Input the output dimensionality-reduced feature vector into a system with dimension [missing information]. In the fully connected classification layer, the classification calculation of partial discharge fault modes is performed, and the classification output category is: These are three typical failure modes, corresponding to tip discharge, surface discharge, floating potential discharge, and internal discharge, respectively. S46. Output the partial discharge fault modes obtained through classification and calculation.

[0043] The multi-scenario switchgear partial discharge simulation and fault reproduction system of this invention achieves real-time acquisition of partial discharge signals inside the switchgear through a distributed ultra-high frequency sensor array. The installation positions of each ultra-high frequency sensor are strictly arranged according to the typical partial discharge generation locations of different electrical components within the switchgear to effectively cover all typical discharge areas. In the wavelet threshold denoising algorithm used during data preprocessing, the denoising threshold is precisely set to three times the noise standard deviation to ensure the effectiveness and stability of denoising. The construction process of the multi-scale second-order attention twin network strictly follows the steps of parallel convolution at three scales, generation of the second-order attention weight matrix, feature enhancement, and pooling. The number, size, and pooling kernel size of each scale convolution kernel in the network structure are clearly defined to ensure the network effectively extracts and fuses fine features. Furthermore, when fusing features in the graph convolutional network, the dimension of the initial feature matrix of the nodes is explicitly set to [missing value]. The input and output dimensions of each convolutional layer are precisely determined to ensure the efficiency and stability of feature propagation in the graph neural network.

[0044] This invention achieves effective interactive analysis of fine features of partial discharge signals and electrical topology features of switchgear by fusing multi-scale second-order attention twin networks and graph neural networks. This significantly improves the accuracy and reliability of partial discharge fault mode classification, effectively solves the limitations of single signal processing or structural analysis methods in the prior art, and achieves the technical goal of accurate simulation and reproduction of partial discharge faults. It has high engineering application value in the field of switchgear equipment diagnosis and maintenance.

[0045] The automatic adjustment algorithm specifically includes: Based on the characteristic modes of partial discharge faults, the target output discharge signal waveform of the discharge simulator is determined, and the waveform amplitude, frequency, and phase are used as target control parameters. Based on the deviation between the target control parameters and the current output parameters of the discharge simulator, an error vector is established with the parameter error as input. The error vector is represented as: ; in, For the first Error vector of each control cycle For the target control parameter vector, This is the current output parameter vector; According to the error vector Construct an adaptive parameter update law, the expression of which is:

[0046] in, For the first Adjustment of discharge simulator parameters for each control cycle. For the first The parameter adjustment amount for each control cycle. and For adaptive adjustment coefficient, , ; A limiting function is used to constrain the parameter adjustment. The limiting function is specifically expressed as follows: ; in, This is the parameter adjustment amount after limiting. and These are the maximum and minimum allowable parameter adjustments, with the maximum adjustment being... The minimum adjustment amount is ; The parameter adjustment amount after limiting is used as the first The output parameters of each control cycle are updated in real time to the target value of the discharge simulator control parameters, and the control cycle is [missing information]. .

[0047] In this embodiment, S5 specifically includes: S51. Based on the characteristic patterns of partial discharge faults, determine the target partial discharge signal waveforms corresponding to multiple typical operating scenarios of the switchgear. The length of the target waveforms is uniformly set to... One sampling point; S52. Based on the target partial discharge signal waveform, determine the initial output parameters of the discharge simulator, including the discharge amplitude range. to The discharge pulse frequency range is to Phase range is to ; S53. According to the automatic adjustment algorithm of claim 9, the control parameters of the discharge simulator are calculated and updated in real time based on the target partial discharge signal waveform and the initial output parameters of the discharge simulator. S54. Based on the real-time updated control parameters of the discharge simulator, drive the discharge simulator to generate a partial discharge simulation signal that is consistent with the target waveform. S55. Real-time feedback acquisition and error calculation are performed on the partial discharge simulation signal output by the discharge simulator, with an acquisition frequency of [frequency missing]. The number of collection points is ; S56. Closed-loop control feedback is performed by real-time acquisition of the error between the signal and the target waveform; the error is greater than... The control parameters are adjusted secondaryly upon triggering, with the error being less than or equal to [the specified value]. When the parameters reach the target state, it is determined that the time is right; S57. Output the partial discharge simulation signal after the output error meets the target conditions.

[0048] The multi-scenario switchgear partial discharge simulation and fault reproduction system of this invention includes a partial discharge signal acquisition module comprising a multi-channel parallel ultra-high frequency sensor array. Each ultra-high frequency sensor array consists of multiple ultra-high frequency sensors with center frequencies of 1MHz, 3MHz, and 10MHz, respectively. The raw partial discharge signal is digitized in real-time using a high-performance analog-to-digital converter at a fixed sampling frequency of 1GHz. The data preprocessing module employs the Daubechies wavelet threshold denoising algorithm to denoise the partial discharge signal. The denoising threshold is determined based on a strict limitation of three times the standard deviation of the noise signal to ensure the accuracy and stability of the denoised signal. In the graph neural network fusion module, the dimensional transformation of the weight matrix and adjacency matrix during data transfer and fusion between graph convolutional layers strictly adheres to the technical constraints in the claims to ensure accurate and easily implemented node feature propagation.

[0049] This invention proposes a multi-scale second-order attention twin network structure, which effectively realizes cross-scale feature extraction and fusion processing of fine features of partial discharge signals, thereby improving the recognition accuracy and robustness of partial discharge signal features. Combined with the topological feature fusion method of graph neural network, it accurately characterizes the spatial and electrical connection features of electrical components inside the switch cabinet, significantly enhancing the accuracy and reliability of partial discharge fault mode recognition and simulation reproduction under different scenarios.

[0050] Example 1: To further verify the feasibility and effectiveness of this invention, it was applied to a university's power system and equipment laboratory for experimental teaching and personnel training. This aimed to help students and technicians quickly understand and master the characteristics, causes, diagnosis, and troubleshooting methods of partial discharge faults in switchgear. In recent years, the laboratory has found that trainees and students have a weak ability to identify the characteristics of partial discharge faults during experimental operation and training on high-voltage switchgear equipment. They are unable to efficiently and accurately determine the type and cause of partial discharge, resulting in poor training effectiveness and slow experimental teaching progress. To improve this situation, the laboratory deployed the multi-scenario switchgear partial discharge simulation and fault reproduction system of this invention from March to June 2024.

[0051] In a laboratory setting, this system uses an ultra-high frequency (UHF) sensor array to perform real-time acquisition and digital processing of partial discharge signals from a typical high-voltage switchgear. During the specific experiments, the system employed UHF sensors with center frequencies of 1MHz, 3MHz, and 10MHz to achieve accurate acquisition of signals in different frequency bands. The acquired analog discharge signals were converted from analog to digital signals using a 1GHz high sampling frequency, and then denoised using wavelet thresholding. After multi-band feature fusion processing with a third-order Butterworth filter, a clean and characteristic partial discharge signal was obtained.

[0052] In the subsequent feature extraction stage, the system adopts a multi-scale second-order attention twin network structure, extracting features using multi-scale convolutional layers with kernel sizes of 3×1, 5×1, and 7×1, and further enhancing feature saliency using a second-order attention mechanism to accurately obtain fine features of partial discharge. To accurately reflect the actual switchgear structure, the laboratory established an electrical topology model according to the actual installation of a standard high-voltage switchgear, determining the number of nodes to be 60, including circuit breakers, disconnectors, current transformers, voltage transformers, and busbars, with spatial coordinates accurate to 0.1m. Node weights were assigned numerical values ​​to generate a weighted adjacency matrix for the switchgear electrical topology model.

[0053] In experimental teaching and training, the system aligns and fuses fine features with topological features, and applies a fused multi-layer graph convolutional network to perform interactive analysis and pattern recognition of the features, forming a clear and explicit classification of partial discharge fault modes, including tip discharge, surface discharge, floating potential discharge, and internal discharge. Based on this, an adaptive parameter control algorithm automatically adjusts the amplitude, frequency, and phase of the discharge simulator to achieve accurate simulation of typical discharge faults. The system's adaptive control process has a control cycle of 1 second, and automatically triggers parameter correction when the real-time feedback error exceeds 5%, thereby ensuring that the signal simulation accuracy meets the requirements of experimental teaching.

[0054] To more intuitively demonstrate the practical application effects of this invention, in April 2024, the laboratory organized 20 senior undergraduate students majoring in electrical engineering and 10 front-line technical personnel from enterprises, forming an experimental group (using the system of this invention) and a control group (traditional training methods). Both groups received 30 days of training and experimental instruction. After the training, on-site assessments were conducted, and the effects were compared using objective data. Specific data results are shown in Table 1: Table 1: Comparison of the effects of this invention on laboratory teaching and training

[0055] The data comparison in Table 1 clearly shows that after training and experimentation using this invention, the accuracy rate of fault type identification in the experimental group significantly improved, reaching 96.2%, far exceeding the 72.5% of the traditional method; the error in the simulation of discharge fault waveforms significantly decreased, with accuracy improved to within 4.2%, far lower than the 15% of the traditional method; the average time for trainees to master partial discharge fault handling skills was drastically reduced from the traditional 20 days to 7 days, and the overall cycle of experimentation and training was also reduced from 30 days to 15 days. Furthermore, trainees' overall satisfaction with the training methods and effects significantly improved, and the pass rate for practical assessments reached 100%, far surpassing traditional training methods.

[0056] The laboratory application of this embodiment demonstrates that the present invention provides an efficient, accurate, and intuitive technique for simulating and reproducing partial discharge faults. This allows trainees to quickly understand and master complex partial discharge fault phenomena and their causes in an experimental environment, significantly improving their practical skills and diagnostic abilities. The present invention effectively solves the technical problems of difficulty in simulating partial discharge fault phenomena, difficulty in trainee understanding, long training cycles, and unsatisfactory training results in traditional laboratory teaching and training, greatly improving the overall quality and effectiveness of experimental teaching and training.

[0057] In summary, the practical application effect of this invention in laboratory teaching and personnel training has been fully verified, demonstrating significant engineering practice and educational training value, and providing strong technical support and basis for its widespread application in similar scenarios.

[0058] 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 multi-scenario switchgear partial discharge simulation and fault reproduction system, characterized in that, Includes the following modules: The partial discharge signal acquisition module collects raw partial discharge signals in multiple typical operating scenarios of the switchgear in real time through a distributed sensor array. The data preprocessing module performs adaptive noise reduction and multi-band filtering on the original partial discharge signal to obtain a purified partial discharge signal. The multi-level second-order attention twin feature extraction module constructs a multi-scale second-order attention twin network structure for the purified partial discharge signal, and performs cross-scale cross-comparison and high-order feature fusion to extract the fine features of the purified partial discharge signal. The equipment topology modeling module establishes an electrical topology model of the switchgear based on the specific electrical components and connections inside the switchgear. The graph neural network fusion module performs multi-layer graph convolution fusion analysis based on the fine features of the switch cabinet electrical topology model and the purified partial discharge signal, and outputs partial discharge fault feature patterns. The fault reproduction module automatically adjusts the parameters of the discharge simulator based on the partial discharge fault characteristic pattern, and generates and outputs partial discharge simulation signals corresponding to multiple typical operating scenarios of the switchgear.

2. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 1, characterized in that, The modules are connected in the following way: S1. Real-time acquisition of raw partial discharge signals in multiple typical operating scenarios, and processing through adaptive noise reduction algorithm and multi-band filtering method to obtain purified partial discharge signals; S2. Based on the purified partial discharge signal, a multi-scale second-order attention twin network is constructed, and cross-scale cross-comparison and high-order feature fusion are performed to extract the fine features of the partial discharge signal. S3. Based on the specific electrical components and connection relationships of the switchgear, establish an electrical topology model of the switchgear and generate topology features; S4. Based on the fine features and topological features of the partial discharge signal, a multi-layer graph convolutional network is fused for interactive analysis to form a partial discharge fault feature pattern. S5. Input the partial discharge fault characteristic mode into the fault reproduction module, and use the automatic adjustment algorithm to accurately control the parameters of the discharge simulator to generate partial discharge simulation signals corresponding to multiple typical operating scenarios.

3. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, S1 specifically includes: S11, using center frequencies respectively , and The ultra-high frequency sensor array collects the original partial discharge signal inside the switch cabinet in real time; S12, with The sampling frequency is used to perform analog-to-digital conversion on the acquired raw partial discharge signal to obtain a digital raw partial discharge signal; S13. Adaptive denoising processing is performed on the digitized original partial discharge signal based on the wavelet threshold algorithm, with the denoising threshold set to the noise standard deviation. times; S14. Apply a cutoff frequency of [frequency value missing] to the noise-reduced signal. , and A third-order Butterworth bandpass filter is used for multi-band filtering. S15. Perform inverse Fourier transform on the filtered signals to obtain the corresponding time-domain signals; S16. Multiple time-domain signals are fused using an amplitude-weighted method, with the weighting coefficients being as follows: frequency band signal , frequency band signal , frequency band signal ; S17. Output the purified partial discharge signal after fusion.

4. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, The multi-scale second-order attention twin network specifically includes: The input purified partial discharge signal is constructed into a length of Time series, with kernel size of , and The three parallel convolutional layers have the following numbers of kernels: , and The step size is set to To obtain initial feature maps at three scales; Second-order attention calculations are performed on the initial feature maps of the three scales mentioned above to generate corresponding attention weight matrices. The dimensions of the attention weight matrices are consistent with the dimensions of the initial feature maps of each scale. The attention weight matrix is ​​multiplied element-wise by the initial feature map at the corresponding scale to obtain the enhanced feature map within the scale, wherein the dimensions of the enhanced feature map are respectively... , and ; Average pooling is performed on the enhanced feature maps at each scale, with a pooling kernel size of [size missing]. Step size is , respectively obtain dimensions as , and Pooling feature mapping; The pooled feature maps at each scale are compressed using fully connected layers, and the output feature dimension of each scale's fully connected layer is uniformly compressed to [value missing]. This yields compressed feature vectors at three different scales.

5. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, S2 specifically includes: S21. Construct the purified partial discharge signal into a length of... digitized time-series signals; S22, respectively construct lengths of The digitized time series signal is input into the convolution kernel with a size of , and Three parallel convolutional layers are used to obtain initial feature maps at three scales; S23. Calculate the corresponding second-order attention weight matrix based on the initial feature mappings of three different scales. The calculation process of the attention weight matrix includes generating the query matrix and key matrix from the initial feature mappings, and obtaining the attention weight matrix through the second-order feature interaction method. S24. Multiply the attention weight matrix element-wise with the initial feature mapping at the corresponding scale to obtain the enhanced feature mappings at the three scales. S25. Perform average pooling on the enhanced feature maps within the scale, with the pooling kernel size being [size missing]. The step size is After pooling, the dimensions obtained are respectively , and Three scale pooling feature maps; S26. The feature maps of the three scales after pooling are compressed using three independent fully connected layers, and the output has a unified dimension. Three-scale compressed feature vectors; S27. Set the output to a unified dimension. The three scale-compressed feature vectors are fused through a concatenation operation to obtain a unified dimension. The fine characteristics of partial discharge signals.

6. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, S3 specifically includes: S31. Determine the node set based on the electrical component list of the switchgear, the number of nodes being... to The nodes include circuit breakers, disconnect switches, current transformers, voltage transformers, and busbars; S32. Determine the spatial coordinates of the nodes based on the actual installation locations of the electrical components in the node set. The spatial coordinates are represented in a three-dimensional Cartesian coordinate system. ,in , , These represent the positions of the nodes within the cabinet in the length, width, and height directions, respectively, with a positional accuracy of [value missing]. ; S33. Determine the connection edges between nodes based on the electrical connection relationships between nodes. Connection edges indicate the existence of electrical connections between nodes, and the matrix representation of the connection relationships between nodes is the adjacency matrix. ,in Represents a node With nodes There is a direct electrical connection. This indicates that there is no direct electrical connection. Indicates the number of nodes; S34. Assign weight values ​​to nodes based on the type of electrical components and the rated voltage level in the node set. The node weight values ​​are... to The values ​​between these ranges, the circuit breaker weights are The weight of the disconnector switch is The current transformer weight is The weight of the voltage transformer is The weight of the motherboard is ; S35. Generate a weighted adjacency matrix for the electrical topology model of the switchgear based on node weights and inter-node connections. The connection relationship At that time, weight For nodes and nodes The product of weight values, the connection relationship At that time, weight The formula for calculating each element in the matrix is: ； in, Represents a node The node weight value, Represents a node The node weight value; S36, Weighted adjacency matrix of the electrical topology model of the output switchgear , as a topological feature.

7. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, The fused multi-layer graph convolutional network specifically includes: The fine features of partial discharge signals are compared with the weighted adjacency matrix of the electrical topology model of the switchgear. Perform feature alignment to obtain the initial input feature matrix of the node. , Indicates the number of nodes. Indicates the dimension of node features; Build includes A graph convolutional network with n convolutional layers, where the input and output feature dimensions of each convolutional layer are as follows: The nth convolutional layer... The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is , No. The layer input dimension is Output dimension is ; The feature propagation formula for each convolutional layer is: ； in, , It is the identity matrix. for The degree matrix, For the first Layer node characteristics, For the first The weight matrix of the layer, It is the ReLU activation function; The first Node feature matrix output by the layer By performing a global average pooling operation on the node features, we obtain a dimension of The global feature vector; Let the dimension be The global feature vector is used as the output feature vector of the fused multi-layer graph convolutional network.

8. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, S4 specifically includes: S41. The dimension output in step S2 is... The fine features of the partial discharge signal were copied separately. Next, construct the initial input feature matrix of the node. ,in This indicates the number of nodes in the electrical topology model of the switchgear; S42. Weighted adjacency matrix of the switchgear electrical topology model output in step S3. With the initial input feature matrix of the node Using the fused multi-layer graph convolutional network described in claim 7, sequentially passing through... The layer-by-layer feature propagation calculation of each graph convolutional layer; S43. The node feature matrix output by the third graph convolutional layer. Perform global average pooling to obtain dimension . The global feature vector; S44, with dimension as The global feature vector is input into a dimension of Feature dimensionality reduction is performed in the fully connected layer, and the output dimension is Dimensionally reduced feature vectors; S45. Input the output dimensionality-reduced feature vector into a system with dimension [missing information]. In the fully connected classification layer, the classification calculation of partial discharge fault modes is performed, and the classification output category is: These are three typical failure modes, corresponding to tip discharge, surface discharge, floating potential discharge, and internal discharge, respectively. S46. Output the partial discharge fault modes obtained through classification and calculation.

9. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, The automatic adjustment algorithm specifically includes: Based on the characteristic modes of partial discharge faults, the target output discharge signal waveform of the discharge simulator is determined, and the waveform amplitude, frequency, and phase are used as target control parameters. Based on the deviation between the target control parameters and the current output parameters of the discharge simulator, an error vector is established with the parameter error as input. The error vector is represented as: ； in, For the first Error vector of each control cycle For the target control parameter vector, This is the current output parameter vector; According to the error vector Construct an adaptive parameter update law, the expression of which is: ； in, For the first Adjustment of discharge simulator parameters for each control cycle. For the first The parameter adjustment amount for each control cycle. and For adaptive adjustment coefficients, , ; A limiting function is used to constrain the parameter adjustment. The limiting function is specifically expressed as follows: ； in, This is the parameter adjustment amount after limiting. and These are the maximum and minimum allowable parameter adjustments, with the maximum adjustment being... The minimum adjustment amount is ; The parameter adjustment amount after limiting is used as the first The output parameters of each control cycle are updated in real time to the target value of the discharge simulator control parameters, and the control cycle is [missing information]. .

10. The multi-scenario switchgear partial discharge simulation and fault reproduction system according to claim 2, characterized in that, S5 specifically includes: S51. Based on the characteristic patterns of partial discharge faults, determine the target partial discharge signal waveforms corresponding to multiple typical operating scenarios of the switchgear. The length of the target waveforms is uniformly set to... One sampling point; S52. Based on the target partial discharge signal waveform, determine the initial output parameters of the discharge simulator, including the discharge amplitude range. to The discharge pulse frequency range is to Phase range is to ; S53. According to the automatic adjustment algorithm of claim 9, the control parameters of the discharge simulator are calculated and updated in real time based on the target partial discharge signal waveform and the initial output parameters of the discharge simulator. S54. Based on the real-time updated control parameters of the discharge simulator, drive the discharge simulator to generate a partial discharge simulation signal that is consistent with the target waveform. S55. Real-time feedback acquisition and error calculation are performed on the partial discharge simulation signal output by the discharge simulator, with an acquisition frequency of [frequency missing]. The number of collection points is ; S56. Closed-loop control feedback is performed by real-time acquisition of the error between the signal and the target waveform; the error is greater than... The control parameters are adjusted secondaryly upon triggering, with the error being less than or equal to [the specified value]. When the parameters reach the target state, it is determined that the time is right; S57. Output the partial discharge simulation signal after the output error meets the target conditions.

Images