Partial discharge mode identification method and device

Abstract

The invention discloses a partial discharge mode identification method and device, and relates to the field of power equipment fault diagnosis, and the method comprises the steps: collecting a partial discharge signal of high-voltage electrical equipment; preprocessing the collected data to obtain a single pulse signal; the method comprises the following steps of: generating a time-frequency graph of a single pulse on the basis of a continuous wavelet transform (CWT) method, and carrying out spatial feature extraction on the generated time-frequency graph by using a multi-scale convolutional neural network (MCNN); the features extracted by the MCNN are input into a bidirectional long short-term memory neural network (BiLSTM), and the time sequence characteristics of the partial discharge features are obtained; an attention mechanism is used to focus key segments of a splicing sequence output by the BiLSTM, global features are extracted, and the accuracy of pattern recognition is improved. According to the invention, mode identification can be carried out according to the collected partial discharge signal, the partial discharge type is judged, related personnel are helped to make a targeted maintenance plan, and long-term reliable operation of electrical equipment is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis of electrical equipment, and particularly to a method and device for partial discharge pattern recognition. Background Art

[0002] As a core component of the critical infrastructure of the power system, electrical equipment undertakes key functions such as power conversion, transmission, and distribution. Its operating performance and working conditions are directly related to the reliability and stability of the overall operation of the power system. When the dielectric properties of the equipment insulation medium deteriorate due to factors such as material aging, manufacturing defects, or environmental stress, it will cause distortion of the electric field distribution state. This non-uniform electric field environment will lead to partial discharge (PD) in the insulation weak area, which is essentially a repetitive gas ionization process caused by microscopic defects inside the insulation medium.

[0003] From the perspective of the insulation degradation mechanism, the partial discharge phenomenon has dual characteristics: it is both a typical characterization parameter of the early degradation of insulation materials and accelerates the irreversible decline of the insulation medium performance through the combined effects of charged particle bombardment, thermal effects, and chemical corrosion during the discharge process. Therefore, the research on partial discharge pattern recognition has significant theoretical significance and engineering application value for improving the management level of the entire life cycle of power equipment and ensuring the safe and economic operation of the power grid.

[0004] The key to partial discharge pattern recognition lies in extracting discriminant feature parameters that can accurately characterize its discharge type. The current mainstream feature extraction methods mainly include the phase distribution pattern (Phase Resolved Partial Discharge, PRPD) and the time distribution pattern (Time Resolved Partial Discharge, TRPD). Among them, the PRPD method collects discharge pulse signals within multiple power frequency cycles to construct a two-dimensional map reflecting the distribution law of discharges in the voltage phase, which is suitable for identifying discharge types with obvious phase characteristics. However, the dependence of this method on large-scale time series data leads to high computational complexity and low processing efficiency.

[0005] The TRPD method takes the single pulse of partial discharge as the research object and classifies it by analyzing its characteristic parameters in the time domain or frequency domain. This method has certain advantages in computational efficiency and is suitable for real-time signal processing. However, due to the non-stationarity and strong transience of partial discharge signals in nature, their spectral distribution is complex and time-varying significantly. Single time domain or frequency domain analysis is difficult to comprehensively characterize their features and is prone to loss of important information.

[0006] Therefore, researchers introduced time-frequency analysis technology. By constructing the time-frequency diagram of PD signals, the joint representation in the time and frequency dimensions was realized. This method can effectively reveal the variation law of signals in time and frequency, and provide a richer feature expression for the recognition of complex discharge patterns. Based on this, the Convolutional Neural Network (CNN), as a typical deep learning structure, has been widely applied to the automatic feature extraction and classification tasks of time-frequency spectrograms. CNN can mine high-order spatial features from input images, avoiding the subjectivity and limitations of manual feature design. However, the convolution kernel size of traditional CNN networks is fixed, making it difficult to capture multi-scale discharge features simultaneously. Moreover, when dealing with sequential data with time dependence, its modeling ability is limited, resulting in a decline in recognition performance.

[0007] Considering the correlation of partial discharge signals in the time dimension, the Long Short-Term Memory (LSTM) neural network was introduced to capture the long-term dependence information in time series, showing good performance in the modeling of non-stationary signals. However, the standard LSTM network can only utilize the information at the current time step and previous steps, making it difficult to fully mine the backward time series features, which limits its ability to express the global time series context.

[0008] To further improve the model's comprehensive modeling ability for sequence information, the Bidirectional Long Short-Term Memory (BiLSTM) neural network was proposed. This structure introduces two LSTM networks, a forward one and a backward one, to process the forward and backward dependencies of the time series respectively, thereby realizing the joint modeling of the context information of the entire input sequence and significantly enhancing the ability to express complex time series features. Summary of the Invention

[0009] The purpose of this application is to propose a partial discharge pattern recognition method for the above-mentioned technical problems. In this invention, the collected partial discharge signals of high-voltage electrical equipment are processed by the continuous wavelet transform method CWT to generate the time-frequency diagram of single pulses, and then accurately identify the partial discharge types after further processing by MCNN-BiLSTM-Attention, ensuring the long-term reliable operation of electrical equipment.

[0010] On the one hand, a partial discharge pattern recognition method includes:

[0011] S1, obtaining the partial discharge signals of high-voltage electrical equipment collected;

[0012] S2, preprocessing the partial discharge signals to obtain single-pulse signals;

[0013] S3. Calculate the CWT coefficients of the unit pulses of partial discharge, and generate a time-frequency diagram of a single pulse based on the CWT coefficients;

[0014] S4. Input the time-frequency diagram into the trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation;

[0015] S5. Input the spatial feature representation into the trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes the temporal feature representations of forward and backward semantic information;

[0016] S6. Input the concatenated sequence into the trained attention mechanism Attention network, focus on the key segments of the concatenated sequence, extract the global semantic features, and after inputting the global semantic features into the fully connected layer, output the discharge type or fault type for characterizing the partial discharge signal through the classification function.

[0017] Preferably, in S1, the collected partial discharge signal is specifically:

[0018] Use a high-frequency current sensor to collect transient signals of multiple typical partial discharge types within one power frequency cycle.

[0019] Preferably, S2 specifically includes:

[0020] Extract the partial discharge signal pulses within one power frequency cycle to obtain the single pulse signal of partial discharge, and set the truncation width of each pulse to 50 us.

[0021] Preferably, in S3, the calculation method of the CWT coefficients of the unit pulses is as follows:

[0022] ;

[0023] Where represents the scale parameter, which is the scaling factor of the wavelet basis function on the time axis; represents the translation position of the wavelet function on the time axis, which is used to analyze the local characteristics of the signal at a specific moment; x(t) is the single pulse signal of partial discharge; represents time.

[0024] Preferably, in S3, generating a time-frequency diagram of a single pulse based on the CWT coefficients specifically includes:

[0025] Extract the absolute value ∣CWT(a,b)∣ of the CWT coefficients to represent the energy intensity of the signal at different frequencies and time points, and convert the scale parameter a into the actual frequency f, and the specific formula is as follows:

[0026] ;

[0027] Among them, f center is the center frequency of the wavelet basis, and is the sampling interval of the single-shot partial discharge pulse signal.

[0028] Set the horizontal axis as the time axis of the time-frequency diagram, covering the complete duration of the single-shot pulse signal; the vertical axis is the frequency axis, covering the frequency band range of the single-shot partial discharge pulse signal; the time-frequency diagram uses the Jet color map to represent the energy intensity, and finally generates a time-frequency diagram with an image size of 64×64×3, where 64×64 represents that the resolution of the image is 64 rows × 64 columns of pixels, and 3 represents three color channels in the RGB format.

[0029] Preferably, in the step S4, the multi-scale convolutional neural network MCNN consists of two parallel convolutional channels and merges features through an addition fusion layer; the structures of both convolutional channels include a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, and a fully connected layer; the first channel uses multiple-sized convolutional kernels to extract local features; the second channel uses convolutional kernels to extract global features; among them, the size of the convolutional kernel in the first channel is smaller than that in the second channel.

[0030] Preferably, in the step S4, the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation, specifically including:

[0031] The two parallel convolutional channels of the multi-scale convolutional neural network MCNN respectively extract the local features and global features of the input time-frequency diagram; after the two channels complete the convolutional operation, feature extraction, and processing through the fully connected layer, the outputs of the two channels are fused in a "per-element addition" manner to generate the final spatial feature representation.

[0032] Preferably, the step S5 specifically includes:

[0033] Flatten the spatial feature representation to convert it into a one-dimensional feature sequence;

[0034] Divide the one-dimensional feature sequence into several time steps, and each time step corresponds to a local area or feature block in the original time-frequency diagram to ensure the structural consistency of subsequent temporal modeling;

[0035] Input the one-dimensional feature sequence into a bidirectional long short-term memory neural network BiLSTM for temporal feature extraction. The bidirectional long short-term memory neural network BiLSTM includes two sub-structures: a forward LSTM network and a backward LSTM network; the forward LSTM processes the features of each time step step by step in chronological order to capture historical dependency information; the backward LSTM processes the feature sequence in reverse chronological order to extract the potential influence of future states on the current features;

[0036] Concatenate the output results of the forward and reverse LSTM networks in the feature dimension to form a temporal feature representation including forward and backward semantic information.

[0037] Preferably, step S6 specifically includes:

[0038] Take the concatenated sequence as input and introduce it into the Attention network. The Attention network calculates the corresponding importance scores for each time step in the sequence and assigns corresponding weight coefficients based on the importance scores;

[0039] Perform a weighted summation operation on the feature vectors of all time steps according to the weight coefficients to obtain a semantic representation vector that integrates global context information;

[0040] Input the semantic representation vector into the fully connected layer, and finally perform normalization processing through the Softmax function to output the corresponding multi-class classification probability distribution, which is used to characterize the discharge type or fault category to which the partial discharge signal belongs.

[0041] On the other hand, a partial discharge pattern recognition device includes:

[0042] A partial discharge signal acquisition module for acquiring the partial discharge signals of the collected high-voltage electrical equipment;

[0043] A single-pulse signal acquisition module for preprocessing the partial discharge signal to obtain a single-pulse signal;

[0044] A single-pulse time-frequency diagram generation module for calculating the CWT coefficients of the unit pulse of the partial discharge and generating a time-frequency diagram of the single pulse based on the CWT coefficients;

[0045] A multi-scale convolutional neural network processing module for inputting the time-frequency diagram into the trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation;

[0046] A bidirectional long short-term memory neural network processing module for inputting the spatial feature representation into the trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes a temporal feature representation of forward and backward semantic information;

[0047] An Attention network processing module for inputting the concatenated sequence into the trained attention mechanism Attention network, focusing on the key segments of the concatenated sequence, extracting global semantic features, inputting the global semantic features into the fully connected layer, and then outputting through the classification function to characterize the discharge type or fault type of the partial discharge signal.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention uses the continuous wavelet transform method CWT to perform time-frequency analysis on the single pulse signal of partial discharge, and constructs a time-frequency spectrum that can simultaneously reflect the time and frequency characteristics; then, a multi-scale convolutional neural network MCNN is introduced to extract spatial features from the time-frequency graph. The MCNN consists of two convolution channels of different scales, which are used to capture local detail features and global structural information respectively. Compared with the traditional single-scale CNN network, it has significant advantages in the comprehensiveness and robustness of spatial feature extraction; further, the spatial feature sequence extracted by MCNN is input into BiLSTM to explore its potential temporal dependency, so as to enhance the modeling ability of the dynamic change law of partial discharge signals; finally, the spatial and temporal feature sequence output by BiLSTM is input into the attention mechanism Attention network, which adaptively allocates weights of different time steps to focus on key signal features, thereby effectively improving the model's recognition accuracy and diagnostic efficiency of fault modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 A schematic flow chart of a partial discharge pattern recognition method according to an embodiment of the present application;

[0052] Figure 2 A partial discharge waveform diagram within one power frequency cycle collected by HFCT according to an embodiment of the present application;

[0053] Figure 3 A single pulse waveform diagram captured from a partial discharge waveform within a power frequency cycle of an embodiment of the present application;

[0054] Figure 4 This is a training flowchart of the CWT-MCNN-BiLSTM-Attention network of an embodiment of the present application;

[0055] Figure 5 This is a confusion matrix diagram of CWT-MCNN-BiLSTM-Attention pattern recognition in an embodiment of the present application;

[0056] Figure 6 This is a t-SNE two-dimensional visualization diagram of the model of the embodiment of the present application before learning;

[0057] Figure 7It is a two-dimensional t-SNE visualization graph before final classification after training for the embodiments of the present application;

[0058] Figure 8 It is a schematic diagram of a partial discharge pattern recognition device for the embodiments of the present application. Specific embodiments

[0059] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0060] As Figure 1 shown, a partial discharge pattern recognition method provided in this embodiment includes the following steps.

[0061] S1. Obtain the partial discharge signals of the high-voltage electrical equipment collected.

[0062] Specifically, a high-frequency current sensor is used to collect the transient signals of various typical partial discharge types within one power frequency cycle. The power frequency cycle is, for example, 20 ms / 50 Hz.

[0063] S2. Preprocess the partial discharge signals to obtain single-pulse signals.

[0064] Specifically, extract the partial discharge signal pulses within one power frequency cycle to obtain the single-pulse signals of the partial discharge. The intercept width of each pulse can be set to 50 μs.

[0065] S3. Calculate the CWT coefficients of the unit pulses of the partial discharge, and generate a time-frequency diagram of the single pulse based on the CWT coefficients.

[0066] Specifically, the calculation method of the CWT coefficients of the unit pulses is as follows:

[0067] ;

[0068] Among them, represents the scale parameter, which is the stretching factor of the wavelet basis function on the time axis; represents the translation position of the wavelet function on the time axis, which is used to analyze the local characteristics of the signal at a specific moment; x(t) is the single-pulse signal of the partial discharge; represents time.

[0069] Furthermore, generating a time-frequency diagram of the single pulse based on the CWT coefficients specifically includes:

[0070] The absolute value of the CWT coefficient |CWT(a,b)| is extracted to represent the energy intensity of the signal at different frequencies and time points, and the scale parameter a is converted into the actual frequency f. The specific formula is as follows:

[0071] ;

[0072] where f center is the center frequency of the wavelet basis, is the sampling interval of the partial discharge single pulse signal.

[0073] The horizontal axis of the time-frequency graph is the time axis, covering the entire duration of a single pulse signal (50 μs); the vertical axis is the frequency axis, covering the main frequency band of a single partial discharge pulse signal. The time-frequency graph uses the Jet color spectrum to represent energy intensity, with warmer colors indicating higher energy. The resulting time-frequency graph has an image size of 64 × 64 × 3, where 64 × 64 represents the image resolution of 64 rows × 64 columns of pixels, and 3 represents the three color channels in RGB format.

[0074] S4, the time-frequency map is input into the trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain spatial feature representation.

[0075] Specifically, the constructed MCNN consists of two parallel convolutional channels, with features merged via an additive fusion layer. Both channels are structured with convolutional layers, batch normalization layers, activation function layers, max pooling layers, and fully connected layers. The first channel uses multiple small convolution kernels (3×3 and 4×4) to enhance the perception of local details and subtle changes, making it suitable for capturing local texture and edge information. The second channel uses larger convolution kernels (7×7) to extract more globally perceptual structural features, which helps understand the overall distribution pattern of the signal.

[0076] The multi-scale convolutional neural network (MCNN) performs feature fusion to obtain spatial feature representation, specifically including:

[0077] The generated time-frequency map serves as the input to the constructed MCNN network. The network's two parallel convolutional channels are used to extract local and global features from the input time-frequency map, respectively. After convolution and feature extraction are completed in each channel, and processed through a fully connected layer, the outputs of the two channels are fused using an element-by-element addition method to generate the final spatial feature representation. This fused feature combines the ability to perceive local details and understand global structure, achieving a coordinated representation of local and global information.

[0078] S5. Input the spatial feature representation into the trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes the temporal feature representations of forward and backward semantic information.

[0079] Specifically, flatten the fused feature vectors output by the MCNN network to convert them into a one-dimensional feature sequence. Subsequently, divide this sequence into several time steps, with each time step corresponding to a local area or feature block in the original time-frequency diagram to ensure the structural consistency of subsequent temporal modeling. Then, input the feature sequence into the BiLSTM for temporal feature extraction, where the BiLSTM network includes two sub-structures: a forward LSTM network and a backward LSTM network. The forward LSTM processes the features of each time step step by step in chronological order to capture historical dependency information; the backward LSTM processes the feature sequence in reverse chronological order to extract the potential influence of future states on the current features. Finally, the output results of the forward and backward LSTM networks are concatenated in the feature dimension to form a temporal feature representation that includes both forward and backward semantic information.

[0080] S6. Input the concatenated sequence into the trained attention mechanism Attention network to focus on the key segments of the concatenated sequence, extract global semantic features, and after inputting the global semantic features into the fully connected layer, output the discharge type and fault type for characterizing the partial discharge signal through a classification function.

[0081] Specifically, take the concatenated sequence output by the BiLSTM network as the input and introduce it into the Attention network. The Attention network can calculate the corresponding importance score for each time step in the sequence and assign the corresponding weight coefficient based on this score. Subsequently, perform a weighted summation operation on the feature vectors of all time steps according to the weights to obtain a semantic representation vector that fuses global context information. This semantic representation vector is input into the fully connected layer and finally normalized through the Softmax function to output the corresponding multi-class classification probability distribution for characterizing the discharge type or fault category to which the input sample belongs.

[0082] The technical solutions of the embodiments of the present application will be further described below through specific embodiments.

[0083] The present invention uses the partial discharge data in transformer oil collected by a high-frequency current sensor in the laboratory, including three fault types: surface discharge, bubble discharge, and needle-plate discharge. 20 groups of data are collected for each discharge type, and the sampling period of each group of data is one power frequency cycle (20 ms). The programming environment is Matlab2024a.

[0084] By Figure 2It can be seen that a power frequency cycle contains multiple pulse signals. Therefore, the collected original signal is segmented into single pulse signals, as Figure 3 shown. Extract 200 single pulse signals for each of the three fault types, generate the time-frequency diagram of partial discharge single pulses using CWT, and add data labels. Label 1 represents surface discharge, label 2 represents bubble discharge, and label 3 represents needle-plate discharge. After setting the labels, divide the training set and the test set. Select 160 data for each type of fault as the training set and 40 data as the test set.

[0085] As Figure 4 shown, input the training set data into the constructed MCNN-BiLSTM-Attention model for training, and use cross-entropy as the loss function of this model. When the convergence condition is reached, complete the training and save the model parameters. Input the test set into the trained model. The diagnostic results of the test set are as Figure 5 shown, where the diagnostic accuracy rate of surface discharge is 85%, the diagnostic accuracy rate of bubble discharge is 85%, the diagnostic accuracy rate of needle-plate discharge is 100%, and the average diagnostic accuracy rate reaches 90.83%.

[0086] To verify the effectiveness of the MCNN-BiLSTM-Attention model in extracting features, use t-SNE to perform two-dimensional visualization on the features of partial discharge signals, where Figure 6 are the original time-frequency diagram features before the model has learned, Figure 7 are the features extracted before the final classification of the model. Comparing the two figures, it can be seen that before feature extraction, the three types of faults overlap significantly in the two-dimensional space and the clustering is not clear; after model training, the three types of fault samples are significantly clustered, and each type forms an independent and compact cluster, indicating that the model can successfully extract features with strong discriminability and high distinguishability after training.

[0087] As Figure 8 shown, this embodiment also discloses a partial discharge pattern recognition device, including:

[0088] A partial discharge signal acquisition module for acquiring the partial discharge signals of high-voltage electrical equipment collected;

[0089] A single pulse signal acquisition module for preprocessing the partial discharge signals to obtain single pulse signals;

[0090] A single pulse time-frequency diagram generation module for calculating the CWT coefficients of the unit pulses of partial discharge and generating the time-frequency diagram of single pulses based on the CWT coefficients;

[0091] A multi-scale convolutional neural network processing module is used to input a time-frequency graph into a trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation;

[0092] A bidirectional long short-term memory neural network processing module is used to input the spatial feature representation into a trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes the temporal feature representations of forward and backward semantic information;

[0093] An Attention network processing module is used to input the concatenated sequence into a trained attention mechanism Attention network, focus on the key segments of the concatenated sequence, extract global semantic features, and after inputting the global semantic features into a fully connected layer, output the discharge type or fault type used to characterize the partial discharge signal through a classification function.

[0094] The specific implementation of each module of a partial discharge pattern recognition device is the same as that of a partial discharge pattern recognition method, and this embodiment will not be repeated here.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for partial discharge pattern recognition, characterized in that, Including: S1. Obtain the partial discharge signals of the high-voltage electrical equipment collected; S2. Preprocess the partial discharge signals to obtain single-pulse signals; S3. Calculate the CWT coefficients of the unit pulses of the partial discharge, and generate a time-frequency diagram of the single pulse based on the CWT coefficients; S4. Input the time-frequency diagram into the trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation; S5. Input the spatial feature representation into the trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes the temporal feature representations of the forward and backward semantic information; S6. Input the concatenated sequence into the trained attention mechanism Attention network, focus on the key segments of the concatenated sequence, extract the global semantic features, and after inputting the global semantic features into the fully connected layer, output the discharge type or fault type for characterizing the partial discharge signal through the classification function.

2. The partial discharge pattern recognition method according to claim 1, wherein In the above S1, the collected partial discharge signals are specifically: Use a high-frequency current sensor to collect the transient signals of multiple typical partial discharge types within one power frequency cycle.

3. The partial discharge pattern recognition method according to claim 1, wherein The above S2 specifically includes: Extract the partial discharge signal pulses within one power frequency cycle to obtain the single-pulse signals of the partial discharge, and set the intercept width of each pulse to 50 us.

4. The partial discharge pattern recognition method according to claim 1, wherein, In the above S3, the calculation method of the CWT coefficients of the unit pulses is as follows: ； Among them, represents the scale parameter, which is the stretching factor of the wavelet basis function on the time axis; represents the translation position of the wavelet function on the time axis, which is used to analyze the local characteristics of the signal at a specific moment; x(t) is the single pulse signal of partial discharge; represents time.

5. The partial discharge pattern recognition method according to claim 1, characterized in that In the above S3, generating a time-frequency diagram of the single pulse based on the CWT coefficients specifically includes: Extract the absolute value |CWT(a,b)| of the CWT coefficients to represent the energy intensity of the signal at different frequencies and time points, and convert the scale parameter a into the actual frequency f. The specific formula is as follows: ； where f center is the center frequency of the wavelet basis, is the sampling interval of the single pulse signal of partial discharge; Set the horizontal axis as the time axis of the time-frequency diagram, covering the complete duration of the single-pulse signal; the vertical axis is the frequency axis, covering the frequency band range of the partial discharge single-pulse signal; the time-frequency diagram uses the Jet chromatogram to represent the energy intensity, and finally generates a time-frequency diagram with an image size of 64×64×3, where 64×64 represents the resolution of the image as 64 rows × 64 columns of pixels, and 3 represents three color channels in the RGB format.

6. The partial discharge pattern recognition method according to claim 1, characterized in that In the above S4, the multi-scale convolutional neural network MCNN consists of two parallel convolutional channels and performs feature merging through an addition fusion layer; the structures of both convolutional channels include a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, and a fully connected layer; the first channel uses multiple-sized convolutional kernels to extract local features; the second channel uses convolutional kernels to extract global features; among them, the size of the convolutional kernels in the first channel is smaller than that in the second channel.

7. The partial discharge pattern recognition method according to claim 6, wherein In the above S4, the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation, specifically including: The two parallel convolutional channels of the multi-scale convolutional neural network MCNN respectively extract the local features and global features of the input time-frequency diagram; after the two channels complete the convolutional operations, feature extraction, and processing through the fully connected layer, the outputs of the two channels are fused in the way of "element-wise addition" to generate the final spatial feature representation.

8. The partial discharge pattern recognition method according to claim 1, wherein The above S5 specifically includes: Flatten the spatial feature representation and convert it into a one-dimensional feature sequence; Divide the one-dimensional feature sequence into several time steps, where each time step corresponds to a local region or feature block in the original time-frequency diagram to ensure the structural consistency of subsequent temporal modeling; Input the one-dimensional feature sequence into a bidirectional long short-term memory neural network BiLSTM for temporal feature extraction. The bidirectional long short-term memory neural network BiLSTM consists of two sub-structures: a forward LSTM network and a backward LSTM network. The forward LSTM processes the features of each time step step by step in chronological order to capture historical dependence information. The backward LSTM processes the feature sequence in reverse chronological order to extract the potential influence of future states on the current features; Concatenate the output results of the forward and backward LSTM networks in the feature dimension to form a temporal feature representation including forward and backward semantic information.

9. The partial discharge pattern recognition method according to claim 1, characterized in that, The S6 specifically includes: Introduce the concatenated sequence as input into the Attention network. The Attention network calculates the corresponding importance score for each time step in the sequence and assigns corresponding weight coefficients based on the importance score; Perform a weighted sum operation on the feature vectors of all time steps according to the weight coefficients to obtain a semantic representation vector that fuses global context information; Input the semantic representation vector into the fully connected layer, and finally perform normalization processing through the Softmax function to output the corresponding multi-class classification probability distribution, which is used to characterize the discharge type or fault category to which the partial discharge signal belongs.

10. A partial discharge pattern recognition device, characterized in that Includes: A partial discharge signal acquisition module for acquiring the partial discharge signals of the collected high-voltage electrical equipment; A single-pulse signal acquisition module for preprocessing the partial discharge signal to obtain a single-pulse signal; A single-pulse time-frequency diagram generation module for calculating the CWT coefficients of the unit pulse of the partial discharge and generating the time-frequency diagram of the single pulse based on the CWT coefficients; A multi-scale convolutional neural network processing module for inputting the time-frequency diagram into the trained multi-scale convolutional neural network MCNN, and the multi-scale convolutional neural network MCNN performs feature fusion to obtain a spatial feature representation; A bidirectional long short-term memory neural network processing module for inputting the spatial feature representation into the trained bidirectional long short-term memory neural network BiLSTM to obtain a concatenated sequence of partial discharge features; the concatenated sequence includes a temporal feature representation of forward and backward semantic information; An Attention network processing module for inputting the concatenated sequence into the trained attention mechanism Attention network, focusing on the key segments of the concatenated sequence, extracting global semantic features, and after inputting the global semantic features into the fully connected layer, outputting through the classification function to characterize the discharge type or fault type of the partial discharge signal.

Images