GIS partial discharge ultrahigh frequency signal clustering identification method based on time domain and frequency domain

Abstract

A GIS partial discharge ultrahigh frequency signal clustering identification method based on a time domain and a frequency domain belongs to the technical field of power equipment state monitoring, and comprises the following steps: step S1, a model training process; s2, the original electromagnetic wave signal x is subjected to shunt coding; s3, filter casing generation and feature separation; s4, independent decoding and weighted fusion are carried out; and S5, an identification process of online monitoring. According to the scheme, the online monitoring capability is achieved, original electromagnetic wave signals are received in real time, a single sensor characteristic pattern is generated through dual-channel coding, mask separation and time-frequency decoding, and a trained model is input to output a recognition result.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment condition monitoring, and particularly relates to a clustering and recognition method for UHF signals of GIS partial discharge based on time domain and frequency domain. Background Art

[0002] Gas Insulated Switchgear (GIS) has become the core equipment of urban power grids due to its advantages such as compact structure, small floor area, high operation reliability, long maintenance cycle, and strong environmental adaptability. During the manufacturing, transportation, and on-site installation processes of GIS equipment, insulation defects (such as metal burrs, floating potential bodies, insulation air gaps, etc.) are likely to be introduced. Although such defects do not break down under power frequency voltage, during the transient process of system faults, they will induce strong partial discharges (PD) due to electric field distortion, gradually damage the insulation performance, and ultimately lead to equipment breakdown accidents.

[0003] The Ultra-High Frequency (UHF) method is the current mainstream partial discharge detection method: it uses the high-frequency electromagnetic wave signals radiated by partial discharges for detection, and has the advantages of high sensitivity, strong anti-interference ability, large detection range, and on-line positioning ability. Due to its comprehensive performance advantages, it has become the preferred solution for GIS on-line monitoring systems, supporting discharge pattern recognition and fault location, and significantly improving the state warning ability.

[0004] There may be multiple insulation defect points inside GIS equipment (such as the coexistence of metal spikes, floating conductors, and air gaps). The signal received by a single sensor is actually the superimposed waveform of multiple discharge sources. If not separated, it will lead to feature confusion. Chinese patent application with publication number CN118465446A discloses a method for identifying multi-source discharge types of GIS local based on a deep residual network. This method first excludes external interference and simulates the scenario of superposition of different partial discharge types. However, it has the following deficiencies: Incomplete coverage of discharge types: Only considering three partial discharge types, namely tip corona, floating, and particle, it is difficult to comprehensively characterize the aliasing effect of multi-source discharges; Limited generalization ability: Existing models are difficult to maintain high accuracy and robustness when dealing with scenarios of superposition of multiple discharge types or coexistence of complex interferences. For example, when multiple discharge sources discharge simultaneously, the PRPD patterns of different discharge types will overlap in the phase-amplitude space (such as the phase intervals of floating discharge and tip discharge crossing). Summary of the Invention

[0005] In view of the above deficiencies of the prior art, the purpose of the present invention is to provide a clustering and recognition method for UHF signals of GIS partial discharge based on time domain and frequency domain.

[0006] To achieve the above object, the present invention adopts the following technical solutions.

[0007] A method for clustering and identifying UHF signals of partial discharge in GIS based on time domain and frequency domain, comprising the following steps: Step S1, model training process: Normalize the training set spectrograms containing spike, particle, air gap, suspension and surface discharge samples; Subsequently, use the pre-trained VGG16 model of ImageNet, remove its top fully connected layer, and use cross-entropy loss and backpropagation iteration optimization to realize the model training of multi-source discharge types; Step S2, split-channel encoding of the original electromagnetic wave signal x: Divide the original electromagnetic wave signal into two independent signal streams through a two-channel separator, and perform encoding processing of time-domain features and frequency-domain features respectively to obtain the first time-domain feature map M conv and the first time-frequency spectrogram M spec and fuse the two features to obtain a cross-domain feature map H; Step S3, filter mask generation and feature separation: Use a separator to pass the cross-domain feature map H through a one-dimensional convolutional neural network-like structure to generate a filter mask for each single UHF sensor, and separate the independent and pure cross-domain feature maps of 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 respectively, and fuse them through learnable weights to generate a final signal estimate value that takes into account both time continuity and spectral analysis ability; Step S5, online monitoring identification process: Normalize and extract features from the final signal estimate value, output the discharge probabilities of 5 types, and take the maximum probability as the identification result of the discharge type for online monitoring.

[0008] Further, step S1 includes: Step S101, input the training set and perform preprocessing: Perform pixel normalization on each color channel of the training set spectrogram; Step S102, adjustment of network structure and parameters: Adopt the deep convolutional neural network VGG16 model, and pre-train it with the IMageNet image set. After training, extract the model parameters 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, set the number of neurons to 128 to adapt to small-scale data sets; Add a Dropout layer inserted between the fully connected layer and the output layer, with a retention probability p = 0.5; Change the output layer to a 5-neuron Softmax classifier; Step S103, Loss Calculation and Backpropagation: Input the normalized training set spectrogram, 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.

[0009] Furthermore, step S2 includes: Step S201: First path, encoding of the first time-domain feature map: Use a one-dimensional convolutional neural network to extract time-domain features from the input original electromagnetic wave signal, and map the original electromagnetic wave signal to 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, and each time frame contains F conv frequency band channels; Step S202: Second path, encoding of the first time-frequency spectrogram: Use the short-time Fourier transform to extract frequency-domain features from the input original electromagnetic wave signal, and generate the first time-frequency spectrogram M in the complex number domain spec ; The first time-frequency spectrogram M spec contains T time frames, and each time frame contains F spec frequency band channels; Step S203: Feature encoding constraint condition: The same specifications of window functions and frame shift parameters are used in the two encoding processes; Step S204: Generation of the cross-domain feature map: Align the time frame dimension T of the first time-domain feature map M conv and the first time-frequency spectrogram M spec ; Extract the corresponding channel information in the same time frame; Perform a concatenation operation along the channel dimension to integrate the time-domain features in the real number domain and the spectral features in the complex number 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 , and the lower layer is the first time-frequency spectrogram M spec .

[0010] Furthermore, 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 a stable input for the subsequent network layers; Subsequently, it flows through a series of point convolutional 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, decoupling of filter mask generation and features: The signal after cyclic output generates parallel filter masks through a parallel point convolution module. The number of filter masks is the same as the number of pre-decoupled UHF sensors. The obtained filter mask matrix is multiplied element-wise with the cross-domain feature map H extracted by the encoder, and finally, the independent feature map of each single UHF sensor is obtained.

[0011] Further, step S4 includes: Step S401, separation of feature maps: The pure cross-domain feature map of each single UHF sensor is decomposed into two parts: Second time-domain feature map ’M conv ; Second time-frequency spectrum map ’M spec ; Step S402, independent decoding and reconstruction of signals: Professional decoding processing is performed on the two separated types of features respectively: Second time-domain feature map M’ conv , through a first-level transposed convolution layer, perform upsampling and feature mapping to restore the time-domain waveform details of the signal, and obtain the signal estimation value S1 based on time-domain reconstruction; Second time-frequency spectrum map M’ spec , through the inverse short-time Fourier transform, convert it back to the time-domain signal, and obtain the signal estimation value S2 based on spectrum reconstruction; Step S403, weighted fusion to generate the final estimated signal S; The two decoding outputs are linearly fused through a learnable weight coefficient α to generate the final single signal estimation value S; S = αS1+(1 - α)S2; where α is a fusion weight that can be freely and dynamically adjusted, and 0 ≤ α ≤ 1.

[0012] Further, step S5 includes: Step S501, input the single signal estimation value S and perform preprocessing: Input the single signal estimation value S and perform normalization. At this time, the normalization model parameters are the same as those in step S1; Step S502, feature extraction: Use a convolution layer and a pooling layer to extract features. The model parameters for feature extraction are the same as those in step S1; Step S503, custom top-level inference: Load the parameters of the trained fully connected layer and Softmax layer; turn off Dropout; Step S504, output classification result: The Softmax layer outputs the discharge probabilities of 5 categories; Take the category with the maximum probability as the recognition result of the final discharge type.

[0013] This solution realizes the precise decoupling and identification of multi-source discharge signals and improves the identification accuracy in complex scenarios through dual-channel coding, cross-domain feature fusion, and separation technologies. This solution disassembles the original electromagnetic wave signals, uses a neural network-like approach, and takes advantage of its powerful learning ability to directly encode and decode the original signals using an end-to-end neural network model, rather than relying on intermediate features extracted by traditional machine learning methods.

[0014] This solution has the ability of online monitoring, receives the original electromagnetic wave signals in real time, generates a single sensor feature map through dual-channel coding, mask separation, and time-frequency decoding, and inputs it into the trained model to output the identification result.

[0015] In this solution, during identification, the model realizes the rapid iteration and stable operation of the model through feature reuse and efficient parameter fine-tuning. Freeze the convolutional layer and pooling layer of the pre-trained model, retain its general feature extraction ability, reuse the representation ability of the pre-trained model for time-frequency domain features, and improve the generalization of the model. Description of the Drawings

[0016] Figure 1 It is a schematic structural diagram of a GIS partial discharge UHF signal clustering and identification system; Figure 2 It is a flowchart of step S1; Figure 3 It is a flowchart of steps S2 - S4; Figure 4 It is a flowchart of step S5. Detailed Implementation Manner

[0017] The present invention will be further described in detail below with reference to the accompanying drawings.

[0018] Partial discharge (PD) caused by manufacturing, assembly, or material defects in gas-insulated metal-enclosed switchgear (GIS) is the main cause of insulation deterioration. According to the physical form and discharge mechanism of insulation defects, GIS partial discharge can be divided into five typical modes: Metal spike discharge: Discharge caused by burrs or foreign objects on the conductor surface that cause electric field distortion; Metal particle discharge: Intermittent discharge triggered by the movement of free metal particles in an electric field; Floating electrode discharge: Continuous discharge of a floating potential body formed by a poorly contacted component; Insulator surface discharge: Creepage phenomenon caused by contamination or wet flash on the insulation surface; Air gap discharge: Penetrating discharge of bubbles or cracks inside the solid insulation.

[0019] There are significant differences in the insulation deterioration mechanism and fault development speed reflected by different discharge types. For example, floating electrode discharge can quickly trigger a ground flashover and requires emergency treatment; although air gap discharge develops slowly, long-term accumulation will lead to insulation breakdown. Therefore, accurately identifying the discharge type is of decisive significance for evaluating the insulation status and formulating maintenance strategies.

[0020] Figure 1 is the structural schematic diagram of the UHF signal clustering recognition system for partial discharge in GIS; as Figure 1 shown, the UHF signal clustering recognition system for partial discharge in GIS consists of three parts: UHF sensors, pre-monitoring units, and background servers.

[0021] UHF sensors monitor the UHF signals of partial discharge during the operation of GIS equipment and transmit the signals to the pre-monitoring unit through RF coaxial cables; The pre-monitoring unit converts the analog signals of the UHF signals of partial discharge into digital signals, performs active noise suppression, and extracts the original electromagnetic wave signals, PRPD patterns, and PRPS patterns, and uploads them to the background server; The background server automatically identifies the partial discharge type based on the received signals and patterns.

[0022] Since the convolutional neural network model has translational invariance for images, and the weight sharing characteristic of the convolutional kernel can significantly reduce the total number of network parameters, the CNN model is selected to identify the types of partial discharge defects.

[0023] The convolutional neural network (CNN) is composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a Softmax layer connected in series. The functions of each layer are as follows: Input layer: Normalize the input data; Convolutional layer: Extract local features and construct feature maps; Pooling layer: Reduce the dimension of the feature map and enhance translational invariance;

[0024] Fully connected layer: Integrate high-level features and output feature vectors;

[0025] Softmax layer: Convert the feature vector into a class probability to complete classification.

[0026] The UHF signal clustering recognition method for partial discharge in GIS based on time domain and frequency domain includes the following steps: Step S1, the model training process. Figure 2 is the flowchart of step S1; as Figure 2 shown.

[0027] Step S101, input the training set and perform preprocessing: Perform pixel normalization on each color channel of the training set spectrogram: eliminate the dimensional difference and accelerate convergence.

[0028] The training set contains training set samples of spike discharge, granular discharge, air gap discharge, floating discharge and surface discharge.

[0029] Step S102, adjustment of network structure and parameters: Adopt the deep convolutional neural network VGG16 model, which is pre-trained on the ImageNet image set. After training, extract the model parameters to enable the model to have the ability of underlying feature extraction.

[0030] Remove the top fully connected layer of the deep convolutional neural network VGG16 model as the CNN model of this scheme to simplify the model and improve efficiency; Create a new fully connected layer with the number of neurons set to 128 to adapt to small-scale data sets; add a Dropout layer inserted between the fully connected layer and the output layer, with a retention probability p = 0.5, and randomly mask 50% of the neurons during training to inhibit overfitting; Change the output layer to a 5-neuron Softmax classifier; The deep convolutional neural network VGG16 model was developed by the research team Visual Geometry Group at the University of Oxford and is an open model.

[0031] Step S103, loss calculation and backpropagation: Input the normalized training set spectrogram, 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.

[0032] Step S2: Split and encode the original electromagnetic wave signal x.

[0033] Divide the original electromagnetic wave signal into two independent signal streams through a two-channel separator. After separately performing the encoding of time-domain features and the encoding of frequency-domain features, obtain the first time-domain feature map M conv And the first time-frequency spectrogram M spec And fuse the two features. Figure 3 is the flowchart of steps S2~S4; as Figure 3 shown. The specific process is as follows: Step S201: For the first path, encoding of the first time-domain feature map: Use a one-dimensional convolutional neural network (1D 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 to the first time-domain feature map M in the real number domainconv ; The first time-domain feature map M conv contains T time frames, and each time frame contains F conv dimensional frequency band channels. T is the time frame dimension; F conv is the frequency band channel dimension of the first time-domain feature map M conv .

[0034] Step S202: Second path, encoding of the first time-frequency spectrum map: Using the short-time Fourier transform, perform frequency-domain feature extraction on the input original electromagnetic wave signal to generate the first time-frequency spectrum map M in the complex domain spec ; The first time-frequency spectrum map M spec contains T time frames, and each time frame contains F spec dimensional frequency band channels. F spec is the frequency band channel dimension of the first time-frequency spectrum map M spec .

[0035] Step S203: Feature encoding constraint conditions: The same specifications of window function (WindowFunction) and hop length parameters are used in the two encoding processes

[0036] The window function lengths are the same to ensure the synchronization of the local time-frequency analysis window of the time-domain signal The hop length parameters are the same to ensure that the overlapping regions of adjacent time frames are the same and maintain the unity of the time resolution of time-frequency analysis

[0037] Step S204: Generation of the cross-domain feature map

[0038] Align the time frame dimension T of the first time-domain feature map M conv and the first time-frequency spectrum map M spec , extract the corresponding channel information under the same time frame; perform a concatenation operation along the channel dimension to integrate the time-domain features in the real domain and the frequency spectrum features in 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 , and the lower layer is the first time-frequency spectrum map M spec . The upper and lower layers jointly generate a cross-domain feature representation containing multi-dimensional information, providing rich feature inputs for subsequent signal processing tasks

[0039] Step S3: Mask generation and feature separation Use a separator to pass the cross-domain feature map H through a one-dimensional convolutional neural network (1D CNN) and generate masks for each single UHF sensor, separating the independent and pure cross-domain feature maps of each UHF sensor. The specific process is as follows 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 a stable input for subsequent network layers; then, it flows through a series of pointwise convolutional layers to perform preliminary feature extraction operations.

[0040] 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 to balance the model's expressive ability and computational efficiency.

[0041] Step S303: Mask generation and feature decoupling. The signal after recurrent output generates parallel masks through parallel pointwise convolutional modules, and the number of masks is the same as the number of pre-decoupled UHF sensors; the obtained mask matrix is multiplied element-wise with the cross-domain feature map H extracted by the encoder, and finally, an independent and pure cross-domain feature map of each single UHF sensor is obtained.

[0042] Step S4: Independent decoding and weighted fusion to achieve signal reconstruction.

[0043] Separate the pure cross-domain feature map into time-domain and frequency-domain parts, decode and reconstruct them respectively, and then fuse them through learnable weights to generate the final signal estimate that takes into account both time continuity and spectral analysis ability.

[0044] Step S401, Separation of feature maps: Separate the pure cross-domain feature map of each single UHF sensor into two parts: The second time-domain feature map 'M conv : Focus on the time-series dynamic characteristics of the signal; The second time-frequency spectrogram 'M spec : Contain the frequency-domain - time joint distribution information of the signal.

[0045] The two parts of the features maintain the spatio-temporal dimensional correspondence of the original feature map and provide input for subsequent independent decoding.

[0046] Step S402, Independent decoding to reconstruct the signal.

[0047] Perform specialized decoding processing on the two separated types of features respectively: The second time-domain feature map M' conv , through a first-level transposed convolutional layer, perform upsampling and feature mapping to restore the time-domain waveform details of the signal and obtain the signal estimate S1 based on time-domain reconstruction; The second time-frequency spectrogram M' spec , through the inverse short-time Fourier transform, convert it back to the time-domain signal to obtain the signal estimate S2 based on spectral reconstruction.

[0048] Step S403, weighted fusion to generate the final estimated signal S.

[0049] Linearly fuse the two decoding outputs with a learnable weight coefficient α to generate the final single signal estimated value S.

[0050] S = αS1 + (1 - α)S2; where α is a fusion weight that can be dynamically adjusted freely, 0 ≤ α ≤ 1, balancing the contributions of the time-domain and frequency-domain reconstructed signals, and α is defaulted to 0.5; S1 is the signal estimated value based on time-domain reconstruction, focusing on the detail reconstruction of the time-domain waveform; S2 is the signal estimated value based on spectrum reconstruction, focusing on the integrity recovery of the spectrum structure.

[0051] S combines S1 and S2, taking into account both the time continuity and spectrum analysis ability of the signal.

[0052] Step S5, the identification process of on-line monitoring: Figure 4 is the flow chart of step S5, as Figure 4 shown.

[0053] Step S501, input the single signal estimated value S and perform preprocessing: Input the single signal estimated value S and perform normalization. At this time, the normalization model parameters are the same as those in step S1.

[0054] Step S502, feature extraction: Use the convolutional layer and pooling layer to extract features. The model parameters for feature extraction are the same as those in step S1.

[0055] Step S503, custom top-level inference: Load the parameters of the trained fully connected layer and Softmax layer; turn off Dropout, and all neurons in the fully connected layer participate in the calculation during testing.

[0056] Step S504, output the classification result: The Softmax layer outputs the probabilities of 5 types of discharges (tip / suspension / particle / gap / surface discharge); take the category with the maximum probability as the recognition result of the final discharge type.

[0057] In this scheme, there are a total of 5 types of typical discharge types. The training set is collected through single faults, so there is no other noise. The number of training set samples for spike discharge, particle discharge, gap discharge, suspension discharge, and surface discharge is 200 each.

[0058] In actual identification, there are scenarios where multiple discharge types are superimposed or complex interferences coexist. Therefore, the signal to be identified goes through steps 2 to 4 to refine the information collected by a single sensor and input it into the trained model for identification, thereby improving the identification accuracy.

[0059] In step S5 of this solution, for feature reusability: the frozen convolutional / pooling layers extract general image features to avoid retraining with small samples. For the efficient fine-tuning mechanism: only the newly built top layer is trained to improve the training speed. For engineering reliability: the normalization parameters of the test path strictly reuse those of the training path and the weights of the new layers to ensure deployment consistency.

[0060] It can be understood that for those of ordinary skill in the art, equivalent replacements or changes can be made according to the technical solution of the present invention and its inventive concept, and all such changes or replacements should fall within the protection scope of the claims appended to the present invention.

Claims

1. A method for clustering and identifying UHF signals of partial discharge in GIS based on time domain and frequency domain, characterized in that, It includes the following steps: Step S1, model training process: Normalize the training set spectrograms containing spike, particle, air gap, suspension, and surface discharge samples; subsequently, use the VGG16 model pre-trained on ImageNet, remove its top fully connected layer, and utilize cross-entropy loss and backpropagation for iterative optimization to achieve model training for multi-source discharge types. Step S2, split encoding of the original electromagnetic wave signal x: 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 a first time domain feature map M conv With the first time spectrum M spec And fuse the features of the two to obtain the cross-domain feature map H; Step S3, filter mask generation and feature separation: Use a separator to pass the cross-domain feature map H through a one-dimensional convolutional neural network-like structure to generate a filter mask for each single UHF sensor, and separate the independent and pure cross-domain feature maps of 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 respectively, and then fuse them through learnable weights to generate the final signal estimate value that takes into account both time continuity and spectral analysis ability. Step S5, online monitoring recognition process: Normalize and extract features from the final signal estimate value, output the probabilities of 5 discharge types, and take the maximum probability as the recognition result of the discharge type for online monitoring.

2. The method for cluster recognition of UHF signals of partial discharge in GIS based on time domain and frequency domain according to claim 1, 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 spectrograms. Step S102, adjustment of network structure and parameters: Adopt the VGG16 model of deep convolutional neural network, pre-train it on the ImageNet image set, and extract the model parameters after training to enable the model to have the ability to extract underlying features. Remove the top fully connected layer of the VGG16 model of deep convolutional neural network as the CNN model of this solution. Create a new fully connected layer with the number of neurons set to 128 to adapt to small-scale data sets; add a Dropout layer inserted between the fully connected layer and the output layer with a retention probability p = 0.

5. Change the output layer to a Softmax classifier with 5 neurons. Step S103, loss calculation and backpropagation: Input the normalized training set spectrograms, calculate the cross-entropy loss between the Softmax output and the true labels, 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 method for clustering and identifying UHF signals of GIS partial discharge based on time domain and frequency domain according to claim 2, wherein, Step S2 includes: Step S201: First path, encoding of the first time-domain feature map: Using a one-dimensional convolutional neural network, perform time-domain feature extraction on the input original electromagnetic wave signal, and map the original electromagnetic wave signal to 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, and each time frame contains F conv frequency band channels of dimension; Step S202: Second path, encoding of the first time-frequency spectrogram: Using short-time Fourier transform, frequency-domain features of the input original electromagnetic wave signal are extracted to generate the first time-frequency spectrogram M in the complex domain spec ; The first time-frequency spectrogram M spec contains T time frames, and each time frame contains F spec dimensional frequency band channels; Step S203: Feature encoding constraint conditions: Use the same specifications of window functions and frame shift parameters for the two encoding processes. Step S204: Generation of cross-domain feature map: Align the first time-domain feature map M conv and the first time-frequency spectrum map M spec in terms of the time frame dimension T, extract the corresponding channel information in the same time frame; perform a concatenation operation along the channel dimension to integrate the time-domain features in the real number domain and the frequency spectrum features in the complex number domain at the channel level, obtaining a cross-domain feature map H; the structure of the cross-domain feature map H is hierarchical: the upper layer is the first time-domain feature map M conv , and the lower layer is the first time-frequency spectrum map M spec .

4. The method for clustering and identifying UHF signals of partial discharge in GIS based on time domain and frequency domain according to claim 3, 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 a stable input for subsequent network layers; subsequently, it flows through a series of point convolutional 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, decoupling the filter mask generation and features: The signal after the loop output generates parallel filter masks through a parallel point convolution module. The number of filter masks is the same as the number of pre-decoupled UHF sensors. The obtained filter mask matrix is multiplied element-wise with the cross-domain feature map H extracted by the encoder, and finally, the independent feature map of each single UHF sensor is obtained.

5. The method for clustering and identifying UHF signals of GIS partial discharge based on time domain and frequency domain according to claim 4, characterized in that, Step S4 includes: Step S401, separation of the feature map: The pure cross-domain feature map of each single UHF sensor is decomposed into two parts: Second time-domain feature map 'M' conv ; Second time-frequency spectrum map 'M' spec ; Step S402, independently decoding and reconstructing the signal: The two types of separated features are respectively subjected to specialized decoding processing: Second time-domain feature map M' conv , through a first-level transposed convolutional layer, perform upsampling and feature mapping to restore the time-domain waveform details of the signal, and obtain a signal estimate value S1 based on time-domain reconstruction; Second time-frequency spectrum diagram M' spec , through inverse short-time Fourier transform, it is converted back to the time-domain signal to obtain the signal estimated value S2 based on spectrum reconstruction; Step S403, weighted fusion to generate the final estimated signal S; The two decoding outputs are linearly fused through a learnable weight coefficient α to generate the final single signal estimated value S; S = αS1+(1 - α)S2; where α is a fusion weight that can be dynamically adjusted freely, and 0 ≤ α ≤ 1.

6. The method for clustering and identifying UHF signals of GIS partial discharge based on time domain and frequency domain according to claim 5, characterized in that, Step S5 includes: Step S501, input the single signal estimated value S and perform preprocessing: Input the single signal estimated value S and perform normalization. At this time, the normalization model parameters are the same as those in Step S1; Step S502, feature extraction: Use convolutional layers and pooling layers to extract features. The model parameters for feature extraction are the same as those in Step S1; Step S503, custom top-level inference: Load the parameters of the trained fully connected layer and Softmax layer; turn off Dropout; Step S504, output the classification result: The Softmax layer outputs the probabilities of 5 types of discharges; the category with the maximum probability is taken as the recognition result of the final discharge type.

Images