Clustering recognition method of GIS partial discharge ultra-high frequency signal based on time domain and frequency domain

Abstract

A clustering and identification method for GIS partial discharge ultra-high frequency (UHF) signals based on time and frequency domains belongs to the field of power equipment condition monitoring technology. The method includes the following steps: Step S1, model training process; Step S2, x-branch encoding of the original electromagnetic wave signal; Step S3, filter mask generation and feature separation; Step S4, independent decoding and weighted fusion; and Step S5, online monitoring and identification process. This solution, capable of online monitoring, receives the original electromagnetic wave signal in real time and generates a single sensor feature map through dual-channel encoding, mask separation, and time-frequency decoding. This map is then input into the trained model and outputs the recognition result.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment status monitoring, and in particular to a GIS partial discharge ultra-high frequency signal clustering recognition method based on time domain and frequency domain. Background Art

[0002] Gas-insulated switchgear (GIS) has become a core component of urban power grids due to its compact structure, small footprint, high operational reliability, long maintenance cycles, and strong environmental adaptability. GIS equipment is prone to insulation defects (such as metal burrs, floating potential bodies, and insulation gaps) during manufacturing, transportation, and on-site installation. Although these defects do not cause breakdown under power frequency voltage, they can induce strong partial discharge (PD) during transient system faults due to electric field distortion, gradually degrading insulation performance and ultimately leading to equipment breakdown.

[0003] Ultra-high frequency (UHF) is the current mainstream method for partial discharge detection. It utilizes the high-frequency electromagnetic wave signals radiated by partial discharge for detection. It offers high sensitivity, strong interference immunity, a large detection range, and online location capabilities. Due to its comprehensive performance advantages, it has become the preferred solution for GIS online monitoring systems, supporting discharge pattern recognition and fault location, significantly improving status warning capabilities.

[0004] GIS equipment may contain multiple insulation defects (such as metal spikes, suspended conductors, and air gaps). The signal received by a single sensor is actually the superimposed waveform of multiple discharge sources. If not separated, the characteristics will be confused. Chinese patent application publication number CN118465446A discloses a method for identifying GIS local multi-source discharge types based on a deep residual network. This method first eliminates external interference and simulates scenarios where different local discharge types overlap. However, it has the following shortcomings:

[0005] Incomplete coverage of discharge types: Only three types of partial discharges, namely tip corona, suspension, and particle, are considered, making it difficult to fully characterize the aliasing effect of multi-source discharges;

[0006] Limited generalization: Existing models struggle to maintain high accuracy and robustness when dealing with scenarios involving the superposition of multiple discharge types or the coexistence of complex interference. For example, simultaneous discharges from multiple discharge sources can cause PRPD spectra of different discharge types to overlap in phase-amplitude space (e.g., the phase intervals of suspended and tip discharges intersect). Summary of the Invention

[0007] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a GIS partial discharge ultra-high frequency signal clustering and identification method based on time domain and frequency domain.

[0008] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0009] The GIS partial discharge ultra-high frequency signal clustering recognition method based on time domain and frequency domain includes the following steps:

[0010] Step S1, model training process: normalize the training set atlas containing spike, particle, air gap, suspension, and surface discharge samples; then use the ImageNet pre-trained VGG16 model, remove its top fully connected layer, and use cross-entropy loss and back-propagation iterative optimization to achieve model training for multiple source discharge types;

[0011] Step S2, original electromagnetic wave signal x is divided into two independent signal streams through a dual-channel separator, and the encoding of the time domain features and the encoding of the frequency domain features are performed respectively to obtain the first time domain feature map M conv With the first time spectrum diagram M spec And fuse the features of the two to obtain the cross-domain feature map H;

[0012] Step S3, filter cover generation and feature separation: A separator is used to pass the cross-domain feature map H through a one-dimensional convolutional neural network and generate a filter cover for each single UHF sensor, thereby separating an independent pure cross-domain feature map for each UHF sensor;

[0013] Step S4, independent decoding and weighted fusion: Separate the pure cross-domain feature map into time domain and frequency domain parts, decode and reconstruct them separately, and fuse them through learnable weights to generate the final signal estimate that takes into account both temporal continuity and spectral resolution capabilities;

[0014] Step S5, online monitoring identification process: normalize and extract features of the final signal estimation value, output 5 types of discharge probabilities, and take the maximum probability as the discharge type identification result of online monitoring.

[0015] Further, step S1 includes:

[0016] Step S101: Input the training set and perform preprocessing:

[0017] Perform pixel normalization on each color channel of the training set atlas;

[0018] Step S102: Adjustment of network structure and parameters:

[0019] The deep convolutional neural network VGG16 model is used and pre-trained on the IMageNet image set. After training, the model parameters are extracted to enable the model to have the ability to extract underlying features.

[0020] Remove the top fully connected layer of the deep convolutional neural network VGG16 model as the CNN model of this solution;

[0021] Create a new fully connected layer with 128 neurons to adapt to small datasets; add a Dropout layer between the fully connected layer and the output layer with a retention probability of p = 0.5;

[0022] The output layer is changed to a Softmax classifier with 5 neurons;

[0023] Step S103, loss calculation and back propagation:

[0024] Input the normalized training set graph, calculate the cross entropy loss between the Softmax output and the true label, update and optimize the model parameters of the fully connected layer and the Softmax layer through the backpropagation algorithm, freeze all model parameters except the newly created fully connected layer and the Softmax layer, and iterate until convergence.

[0025] Further, step S2 includes:

[0026] Step S201: First path, encoding of the first time domain feature map: using a one-dimensional convolutional neural network to extract the time domain features of the input original electromagnetic wave signal, and mapping the original electromagnetic wave signal into the first time domain feature map M in the real number domain conv ; The first time domain feature map M conv Contains T time frames, each time frame contains F conv dimensional frequency band channel;

[0027] Step S202: Second path, encoding of the first time-frequency spectrum: using short-time Fourier transform, extracting the frequency domain features of the input original electromagnetic wave signal to generate the first time-frequency spectrum M in the complex domain. spec ; The first time spectrum diagram M spec Contains T time frames, each time frame contains F spec dimensional frequency band channel;

[0028] Step S203: Feature coding constraints: The two-way coding process uses the same window function and frame shift parameters;

[0029] Step S204: Generation of cross-domain feature maps: Aligning the first time domain feature map M conv and the first time spectrum M spec The time frame dimension T is used to extract the corresponding channel information in the same time frame; a concatenation operation is performed along the channel dimension to integrate the time domain features of the real domain and the spectrum features of the complex domain at the channel level to obtain a cross-domain feature map H; the structure of the cross-domain feature map H is divided into upper and lower layers: the upper layer is the first time domain feature map M conv , the lower layer is the first time spectrum graph M spec.

[0030] Further, step S3 includes:

[0031] Step S301, input preprocessing and initial feature extraction: The cross-domain feature map H is preprocessed by layer normalization to normalize the data distribution and provide stable input for subsequent network layers; then, it flows through a series of point convolution layers to perform preliminary feature extraction operations;

[0032] Step S302, feature extraction and enhancement: the signal after preprocessing and initial feature extraction enters the core feature extraction network composed of 4 stacked recurrent units; the number of channels d of the recurrent unit is 128;

[0033] Step S303, filter mask generation and feature decoupling: the signal after the loop output is generated into parallel filter masks through a parallel point convolution module, and the number of filter masks is the same as the number of pre-decoupled UHF sensors; the obtained filter mask matrix is ​​multiplied element by element with the cross-domain feature map H extracted by the encoder, and finally an independent feature map of each single UHF sensor is obtained.

[0034] Further, step S4 includes:

[0035] Step S401, separation of feature maps: decompose the pure cross-domain feature map of each single UHF sensor into two parts:

[0036] The second time domain feature map 'M conv ; The second time spectrum diagram 'M spec ;

[0037] Step S402: Independently decode and reconstruct the signal: perform specialized decoding processing on the two separated features respectively:

[0038] The second time domain feature map M' conv , through the first-level deconvolution layer, upsampling and feature mapping are performed to restore the time domain waveform details of the signal and obtain the signal estimation value S1 based on time domain reconstruction;

[0039] The second time spectrum M' spec , converted back to the time domain signal through inverse short-time Fourier transform, and the signal estimation value S2 based on spectrum reconstruction is obtained;

[0040] Step S403: weighted fusion to generate a final estimated signal S;

[0041] The two decoded outputs are linearly fused through a learnable weight coefficient α to generate the final single signal estimate S; S = αS1 + (1-α)S2; where α is a freely dynamically adjusted fusion weight, 0≤α≤1.

[0042] Further, step S5 includes:

[0043] Step S501, inputting a single signal estimation value S and performing preprocessing: inputting a single signal estimation value S and performing normalization. At this time, the normalized model parameters are the same as the model parameters in step S1;

[0044] Step S502, feature extraction: use convolutional layers and pooling layers to extract features, and the model parameters for feature extraction are the same as those in step S1;

[0045] Step S503, customizing top-level inference: loading trained fully connected layer and Softmax layer parameters; disabling Dropout;

[0046] Step S504, output classification results: the Softmax layer outputs 5 categories of discharge probabilities; the category with the maximum probability is taken as the final discharge type recognition result.

[0047] This solution uses dual-channel encoding, cross-domain feature fusion, and separation technology to achieve precise decoupling and identification of multi-source discharge signals, improving recognition accuracy in complex scenarios. This solution decomposes the original electromagnetic wave signal and uses a neural network-like model, leveraging its powerful learning capabilities to directly encode and decode the original signal using an end-to-end neural network model, rather than relying on intermediate features extracted by traditional machine learning methods.

[0048] This solution has the ability of online monitoring, receiving the original electromagnetic wave signal in real time, generating a single sensor feature map through dual-channel encoding, mask separation and time-frequency decoding, and inputting the trained model to output the recognition result.

[0049] This solution achieves rapid model iteration and stable operation during recognition through feature reuse and efficient parameter fine-tuning. It freezes the convolutional and pooling layers of the pre-trained model, retaining its general feature extraction capabilities and reusing the pre-trained model's ability to represent time-frequency domain features, improving model generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a structural diagram of the GIS partial discharge UHF signal clustering recognition system;

[0051] Figure 2 is a flow chart of step S1;

[0052] Figure 3 is a flow chart of steps S2 to S4;

[0053] Figure 4 This is a flowchart of step S5. DETAILED DESCRIPTION

[0054] The present invention will be described in further detail below with reference to the accompanying drawings.

[0055] Partial discharge (PD) caused by manufacturing, assembly, or material defects in gas-insulated metal-enclosed switchgear (GIS) is the main cause of insulation degradation. Based on the physical form of insulation defects and discharge mechanisms, GIS PD can be divided into five typical modes:

[0056] Metal spike discharge: discharge caused by burrs or foreign matter on the conductor surface causing electric field distortion;

[0057] Metal particle discharge: intermittent discharge triggered by free metal particles moving in an electric field;

[0058] Suspended electrode discharge: Continuous discharge of suspended potential body formed by poor contact parts;

[0059] Insulator surface discharge: creepage caused by dirt or wet flashover on the insulation surface;

[0060] Air gap discharge: penetrating discharge through bubbles or cracks inside solid insulation.

[0061] Different discharge types reflect significant differences in the insulation degradation mechanisms and fault development speeds. For example, suspended electrode discharge can quickly trigger a flashover to ground, requiring urgent attention. Air gap discharge, while slow to develop, can lead to insulation breakdown over time. Therefore, accurately identifying discharge types is crucial for assessing insulation condition and formulating maintenance strategies.

[0062] Figure 1 This is a structural diagram of the GIS partial discharge UHF signal clustering recognition system; Figure 1 As shown in the figure, the GIS partial discharge UHF signal clustering and identification system consists of three parts: UHF sensor, front-end monitoring unit and background server.

[0063] UHF sensor monitors the ultra-high frequency (UHF) signal of partial discharge during GIS operation and transmits the signal to the front monitoring unit via RF coaxial cable;

[0064] The front-end monitoring unit converts the analog signal of the partial discharge UHF signal into a digital signal, performs active noise suppression, and extracts the original electromagnetic wave signal, PRPD spectrum and PRPS spectrum, and uploads them to the back-end server;

[0065] The backend server automatically identifies the type of partial discharge based on the received signals and graphs.

[0066] Since the convolutional neural network model is translation invariant to the image and the weight sharing property of the convolution kernel can significantly reduce the total number of network parameters, the CNN model is selected to identify the type of partial discharge defects.

[0067] Convolutional Neural Network (CNN) consists of an input layer, a convolution layer, a pooling layer, a fully connected layer, and a Softmax layer connected in series. The functions of each layer are as follows:

[0068] Input layer: normalize the input data;

[0069] Convolutional layer: extract local features and construct feature maps;

[0070] Pooling layer: reduces the dimension of feature maps and enhances translation invariance;

[0071] Fully connected layer: integrates high-level features and outputs feature vectors;

[0072] Softmax layer: converts feature vectors into category probabilities to complete classification.

[0073] The GIS partial discharge ultra-high frequency signal clustering recognition method based on time domain and frequency domain includes the following steps:

[0074] Step S1, model training process. Figure 2 is a flow chart of step S1; Figure 2 shown.

[0075] Step S101: Input the training set and perform preprocessing:

[0076] Perform pixel normalization on each color channel of the training set atlas: eliminate dimensionality differences and accelerate convergence.

[0077] The training set contains training set samples of spike discharge, particle discharge, air gap discharge, suspension discharge and surface discharge.

[0078] Step S102: Adjustment of network structure and parameters:

[0079] The deep convolutional neural network VGG16 model is used and pre-trained with the IMageNet image set. After training, the model parameters are extracted to enable the model to have the ability to extract underlying features.

[0080] Remove the top fully connected layer of the deep convolutional neural network VGG16 model as the CNN model of this solution to simplify the model and improve efficiency;

[0081] A new fully connected layer is created with the number of neurons set to 128 to adapt to small-scale datasets. A Dropout layer is added between the fully connected layer and the output layer with a retention probability of p = 0.5. During training, 50% of the neurons are randomly blocked to suppress overfitting.

[0082] The output layer is changed to a Softmax classifier with 5 neurons;

[0083] The deep convolutional neural network VGG16 model was developed by the Visual GeoMetry Group, a research team at the University of Oxford, and is an open model.

[0084] Step S103, loss calculation and back propagation:

[0085] Input the normalized training set graph, calculate the cross entropy loss between the Softmax output and the true label, update and optimize the model parameters of the fully connected layer and the Softmax layer through the backpropagation algorithm, freeze all model parameters except the newly created fully connected layer and the Softmax layer, and iterate until convergence.

[0086] Step S2: The original electromagnetic wave signal x is encoded in different channels.

[0087] The original electromagnetic wave signal is divided into two independent signal streams through a dual-channel separator, and the encoding of the time domain features and the encoding of the frequency domain features are performed respectively to obtain the first time domain feature map M conv With the first time spectrum diagram M spec And fuse the two features. Figure 3 is a flow chart of steps S2 to S4; Figure 3 The specific process is as follows:

[0088] Step S201: First path, encoding of the first time domain feature map: using a one-dimensional convolutional neural network (1D CNN) to extract the time domain features of the input original electromagnetic wave signal, and map the original electromagnetic wave signal into the first time domain feature map M in the real number domain. conv ; The first time domain feature map M conv Contains T time frames, each time frame contains F conv dimensional frequency band channel. T is the time frame dimension; F conv is the first time domain feature map M conv The frequency band channel dimension.

[0089] Step S202: Second path, encoding of the first time-frequency spectrum: using short-time Fourier transform, extracting the frequency domain features of the input original electromagnetic wave signal to generate the first time-frequency spectrum M in the complex domain. spec ; The first time spectrum diagram M spec Contains T time frames, each time frame contains F spec dimensional frequency band channel. spec The first time spectrum diagram M spec The frequency band channel dimension.

[0090] Step S203: Feature encoding constraints: The two-pass encoding process uses the same specifications of the window function (WindowFunction) and frame shift (HopLength) parameters.

[0091] The window function length is consistent to ensure the synchronization of the local time-frequency analysis windows of the time domain signal;

[0092] The same frame shift parameters ensure that the overlapping areas of adjacent time frames are consistent and maintain the uniformity of the time resolution of the time-frequency analysis.

[0093] Step S204: Generation of cross-domain feature maps.

[0094] Align the first temporal feature map M conv and the first time spectrum M spec The time frame dimension T is used to extract the corresponding channel information in the same time frame; the concatenation operation is performed along the channel dimension to integrate the time domain features of the real domain and the spectrum features of the complex domain at the channel level to obtain a cross-domain feature map H. The structure of the cross-domain feature map H is divided into upper and lower layers: the upper layer is the first time domain feature map M conv , the lower layer is the first time spectrum graph M spec , the upper and lower layers jointly generate cross-domain feature representations containing multi-dimensional information, providing rich feature input for subsequent signal processing tasks.

[0095] Step S3: filter mask generation and feature separation:

[0096] Using a separator, the cross-domain feature map H is passed through a one-dimensional convolutional neural network (1D CNN) and a filter mask is generated for each single UHF sensor, separating the independent and pure cross-domain feature map of each UHF sensor. The specific process is as follows:

[0097] Step S301: Input preprocessing and initial feature extraction. The cross-domain feature map H undergoes layer normalization (LayerNorMalization) preprocessing to normalize the data distribution and provide stable input for subsequent network layers. It then flows through a series of point convolutional layers to perform preliminary feature extraction.

[0098] Step S302: Feature extraction and enhancement. After preprocessing and initial feature extraction, the signal enters a core feature extraction network consisting of four stacked recurrent units. The number of channels d of the recurrent unit is 128 to balance the model's expressive power and computational efficiency.

[0099] Step S303: Filter generation and feature decoupling. The looped output signal is processed through parallel point convolution modules to generate parallel filter masks. The number of filter masks is equal to the number of pre-decoupled UHF sensors. The resulting filter mask matrix is ​​element-wise multiplied with the cross-domain feature map H extracted by the encoder to ultimately obtain an independent, pure cross-domain feature map for each single UHF sensor.

[0100] Step S4: Independent decoding and weighted fusion to achieve signal reconstruction.

[0101] The pure cross-domain feature map is separated into time domain and frequency domain parts, which are decoded and reconstructed separately and then fused through learnable weights to generate the final signal estimation value that takes into account both temporal continuity and spectral resolution capabilities.

[0102] Step S401, separation of feature maps: decompose the pure cross-domain feature map of each single UHF sensor into two parts:

[0103] The second time domain feature map 'M conv : Focus on the time series dynamic characteristics of the signal;

[0104] The second time spectrum diagram 'M spec : Contains the frequency-time joint distribution information of the signal.

[0105] The two parts of features maintain the correspondence between the temporal and spatial dimensions of the original feature map, providing input for subsequent independent decoding.

[0106] Step S402: independently decode and reconstruct the signal.

[0107] The two separated features are decoded separately in a specialized manner:

[0108] The second time domain feature map M' conv , through the first-level deconvolution layer, upsampling and feature mapping are performed to restore the time domain waveform details of the signal and obtain the signal estimation value S1 based on time domain reconstruction;

[0109] The second time spectrum M' spec , through the inverse short-time Fourier transform, it is converted back to the time domain signal to obtain the signal estimation value S2 based on spectrum reconstruction.

[0110] Step S403: weighted fusion to generate a final estimated signal S.

[0111] The two decoded outputs are linearly fused through a learnable weight coefficient α to generate the final single signal estimate S.

[0112] S=αS1+(1-α)S2; where α is the freely dynamically adjusted fusion weight, 0≤α≤1, which balances the contributions of the time domain and frequency domain reconstructed signals, and α defaults to 0.5; S1 is the signal estimation value based on time domain reconstruction, focusing on the detailed reconstruction of the time domain waveform; S2 is the signal estimation value based on spectrum reconstruction, focusing on the integrity restoration of the spectrum structure.

[0113] S combines S1 and S2, taking into account both the temporal continuity and spectrum resolution of the signal.

[0114] Step S5, online monitoring identification process: Figure 4 is a flow chart of step S5, such as Figure 4 shown.

[0115] Step S501: Input a single signal estimation value S and perform preprocessing:

[0116] A single signal estimation value S is input and normalized. At this time, the normalized model parameters are the same as the model parameters in step S1.

[0117] Step S502, feature extraction:

[0118] Convolutional layers and pooling layers are used to extract features. The model parameters for feature extraction are the same as those in step S1.

[0119] Step S503: Customize top-level reasoning:

[0120] Load the trained fully connected layer and softmax layer parameters; turn off Dropout, and all neurons in the fully connected layer participate in the calculation during testing.

[0121] Step S504: Output classification results:

[0122] The Softmax layer outputs five types of discharge probabilities (tip / suspension / particle / air gap / surface discharge); the category with the highest probability is taken as the final discharge type recognition result.

[0123] This solution includes a total of five typical discharge types. The training set is collected through a single fault, so it has no other noise. The number of training set samples for spike discharge, particle discharge, air gap discharge, suspension discharge and surface discharge is 200.

[0124] In actual recognition, there are scenarios where multiple discharge types overlap or complex interference coexists. Therefore, the signal to be identified is refined through steps 2 to 4 to extract the information collected by a single sensor and input it into the trained model for recognition, thereby improving the recognition accuracy.

[0125] In step S5 of this solution, feature reusability is achieved by freezing convolutional and pooling layers to extract common image features, avoiding retraining on small samples. Efficient fine-tuning is achieved by training only the newly created top layers, improving training speed. Engineering reliability is achieved by strictly reusing the normalization parameters and new layer weights from the training path in the test path to ensure deployment consistency.

[0126] It is understandable that for a person skilled in the art who has ordinary skills in the art, the pre-monitoring unit can be equivalently replaced or changed according to the technical solution and inventive concept of the present invention, and all these changes or replacements should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A clustering and identification method for GIS partial discharge ultra-high frequency signals based on time domain and frequency domain, characterized in that: The following steps are involved: Step S1, model training process: normalize the training set atlas containing spike, particle, air gap, suspension, and surface discharge samples; then use the ImageNet pre-trained VGG16 model, remove its top fully connected layer, and use cross-entropy loss and back-propagation iterative optimization to achieve model training for multiple source discharge types; Step S2, original electromagnetic wave signal x is divided into two independent signal streams through a dual-channel separator, and the encoding of the time domain features and the encoding of the frequency domain features are performed respectively to obtain the first time domain feature map M conv With the first time spectrum diagram M spec And fuse the features of the two to obtain the cross-domain feature map H; Step S3, filter cover generation and feature separation: A separator is used to pass the cross-domain feature map H through a one-dimensional convolutional neural network and generate a filter cover for each single UHF sensor, thereby separating an independent pure cross-domain feature map for each UHF sensor; Step S4, independent decoding and weighted fusion: Separate the pure cross-domain feature map into time domain and frequency domain parts, decode and reconstruct them separately, and fuse them through learnable weights to generate the final signal estimate that takes into account both temporal continuity and spectral resolution capabilities; Step S5, online monitoring identification process: normalize and extract features of the final signal estimation value, output 5 types of discharge probabilities, and take the maximum probability as the discharge type identification result of online monitoring.

2. The GIS partial discharge ultra-high frequency signal clustering identification method based on time domain and frequency domain according to claim 1 is characterized in that: Step S1 includes: Step S101: input the training set and perform preprocessing: Perform pixel normalization on each color channel of the training set atlas; Step S102: Adjustment of network structure and parameters: The deep convolutional neural network VGG16 model is used and pre-trained on the IMageNet image set. After training, the model parameters are extracted to enable the model to have the ability to extract underlying features. Remove the top fully connected layer of the deep convolutional neural network VGG16 model as the CNN model of this solution; Create a new fully connected layer with 128 neurons to adapt to small datasets; add a Dropout layer between the fully connected layer and the output layer with a retention probability of p = 0.5; The output layer is changed to a Softmax classifier with 5 neurons; Step S103, loss calculation and back propagation: Input the normalized training set graph, calculate the cross entropy loss between the Softmax output and the true label, update and optimize the model parameters of the fully connected layer and the Softmax layer through the backpropagation algorithm, freeze all model parameters except the newly created fully connected layer and the Softmax layer, and iterate until convergence.

3. The GIS partial discharge ultra-high frequency signal clustering identification method based on time domain and frequency domain according to claim 2 is characterized in that: Step S2 includes: Step S201: First path, encoding of the first time domain feature map: using a one-dimensional convolutional neural network to extract the time domain features of the input original electromagnetic wave signal, and mapping the original electromagnetic wave signal into the first time domain feature map M in the real number domain conv ; The first time domain feature map M conv Contains T time frames, each time frame contains F conv dimensional frequency band channel; Step S202: Second path, encoding of the first time-frequency spectrum: using short-time Fourier transform, extracting the frequency domain features of the input original electromagnetic wave signal to generate the first time-frequency spectrum M in the complex domain. spec ; The first time spectrum diagram M spec Contains T time frames, each time frame contains F spec dimensional frequency band channel; Step S203: Feature coding constraints: The two-way coding process uses the same window function and frame shift parameters; Step S204: Generation of cross-domain feature maps: Aligning the first time domain feature map M conv and the first time spectrum M spec The time frame dimension T is used to extract the corresponding channel information in the same time frame; a concatenation operation is performed along the channel dimension to integrate the time domain features of the real domain and the spectrum features of the complex domain at the channel level to obtain a cross-domain feature map H; the structure of the cross-domain feature map H is divided into upper and lower layers: the upper layer is the first time domain feature map M conv , the lower layer is the first time spectrum graph M spec .

4. The GIS partial discharge ultra-high frequency signal clustering identification method based on time domain and frequency domain according to claim 3 is characterized in that: Step S3 includes: Step S301, input preprocessing and initial feature extraction: The cross-domain feature map H is preprocessed by layer normalization to normalize the data distribution and provide stable input for subsequent network layers; then, it flows through a series of point convolution layers to perform preliminary feature extraction operations; Step S302, feature extraction and enhancement: the signal after preprocessing and initial feature extraction enters the core feature extraction network composed of 4 stacked recurrent units; the number of channels d of the recurrent unit is 128; Step S303, filter mask generation and feature decoupling: the signal after the loop output is generated into parallel filter masks through a parallel point convolution module, and the number of filter masks is the same as the number of pre-decoupled UHF sensors; the obtained filter mask matrix is ​​multiplied element by element with the cross-domain feature map H extracted by the encoder, and finally an independent feature map of each single UHF sensor is obtained.

5. The GIS partial discharge ultra-high frequency signal clustering and identification method based on time domain and frequency domain according to claim 4 is characterized in that: Step S4 includes: Step S401, separation of feature maps: decompose the pure cross-domain feature map of each single UHF sensor into two parts: The second time domain feature map 'M conv ; The second time spectrum diagram 'M spec ; Step S402: Independently decode and reconstruct the signal: perform specialized decoding processing on the two separated features respectively: The second time domain feature map M' conv , through the first-level deconvolution layer, upsampling and feature mapping are performed to restore the time domain waveform details of the signal and obtain the signal estimation value S1 based on time domain reconstruction; The second time spectrum M' spec , converted back to the time domain signal through inverse short-time Fourier transform, and the signal estimation value S2 based on spectrum reconstruction is obtained; Step S403: weighted fusion to generate a final estimated signal S; The two decoded outputs are linearly fused through a learnable weight coefficient α to generate the final single signal estimate S; S = αS1 + (1-α)S2; where α is a freely dynamically adjusted fusion weight, 0≤α≤1.

6. The GIS partial discharge ultra-high frequency signal clustering and identification method based on time domain and frequency domain according to claim 5 is characterized in that: Step S5 includes: Step S501, inputting a single signal estimation value S and performing preprocessing: inputting a single signal estimation value S and performing normalization. At this time, the normalized model parameters are the same as the model parameters in step S1; Step S502, feature extraction: use convolutional layers and pooling layers to extract features, and the model parameters for feature extraction are the same as those in step S1; Step S503, customizing top-level inference: loading trained fully connected layer and Softmax layer parameters; disabling Dropout; Step S504, output classification results: the Softmax layer outputs 5 categories of discharge probabilities; the category with the maximum probability is taken as the final discharge type recognition result.

Images