Partial discharge pattern recognition method and device

Abstract

The application discloses a partial discharge mode recognition method and device, and relates to the field of power equipment fault diagnosis, which comprises the following steps: collecting the partial discharge signals of high-voltage electrical equipment; pre-processing the collected data to obtain single pulse signals; generating the time-frequency graph of single pulse based on the continuous wavelet transform method (CWT) and extracting the spatial features of the generated time-frequency graph by using the multi-scale convolutional neural network (MCNN); inputting the features extracted by the MCNN into the bidirectional long short-term memory neural network (BiLSTM) to obtain the time sequence characteristics of the partial discharge features; using the attention mechanism to focus on the key fragments of the output splicing sequence of the BiLSTM, extracting the global features and improving the accuracy of mode recognition. The application can recognize the mode according to the collected partial discharge signals, judge the type of partial discharge, help relevant personnel to formulate targeted maintenance plans and ensure the long-term reliable operation of electrical equipment.

Description

Technical Field

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

[0002] As a core component of the critical infrastructure of power systems, electrical equipment performs crucial functions such as power conversion, transmission, and distribution. Its operating performance and operating conditions are directly related to the reliability and stability of the overall power system. When the dielectric properties of the equipment's insulation degrade due to factors such as material aging, manufacturing defects, or environmental stress, the electric field distribution becomes distorted. This non-uniform electric field environment can lead to partial discharge (PD) in areas of weak insulation. PD is essentially a repetitive gas ionization process caused by microscopic defects within the insulation.

[0003] From the perspective of insulation degradation mechanisms, partial discharge (PD) exhibits dual characteristics: it serves as a typical indicator of early insulation degradation and, through the combined effects of charged particle bombardment, thermal effects, and chemical corrosion during the discharge process, accelerates the irreversible degradation of insulation properties. Therefore, research on PD pattern recognition has significant theoretical significance and engineering application value for improving the management of power equipment throughout its lifecycle and ensuring the safe and economical operation of power grids.

[0004] The key to partial discharge pattern recognition lies in extracting discriminative feature parameters that can accurately characterize its discharge type. The current mainstream feature extraction methods mainly include phase distribution mode (Phase Resolved Partial Discharge, PRPD) and time distribution mode (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 pattern of discharge in voltage phase. It is suitable for identifying discharge types with obvious phase characteristics. However, this method's reliance on large-scale time series data leads to high computational complexity and low processing efficiency.

[0005] The TRPD method studies single pulses of partial discharge (PD) and classifies them by analyzing their characteristic parameters in the time or frequency domain. This method offers advantages in computational efficiency and is suitable for real-time signal processing. However, due to the inherent non-stationary and highly transient nature of PD signals, their spectral distribution is complex and time-varying. Therefore, single-domain analysis in the time or frequency domain is difficult to fully characterize and can easily result in the loss of important information.

[0006] To this end, researchers introduced time-frequency analysis technology and achieved joint characterization in the time and frequency dimensions by constructing time-frequency graphs of PD signals. This method can effectively reveal the changing patterns of signals in time and frequency, and provide richer feature expressions for the recognition of complex discharge patterns. Based on this, Convolutional Neural Network (CNN), as a typical deep learning structure, has been widely used in automatic feature extraction and classification tasks of time-frequency graphs. CNN can mine high-order spatial features from input images, avoiding the subjectivity and limitations of artificial feature design. However, the convolution kernel size of traditional CNN networks is fixed, making it difficult to capture multi-scale discharge features at the same time, and its modeling ability is limited when processing time-dependent sequence data, resulting in reduced recognition performance.

[0007] Considering the temporal correlation of partial discharge signals, the Long Short-Term Memory (LSTM) neural network was introduced to capture the long-term dependencies in time series, demonstrating excellent performance in modeling non-stationary signals. However, standard LSTM networks only utilize information from the current time step and before, making it difficult to fully exploit backward time series features, limiting their ability to express global temporal context.

[0008] To further enhance the model's comprehensive modeling capabilities for sequence information, a bidirectional long short-term memory (BiLSTM) neural network was proposed. This architecture introduces two LSTM networks, one forward and one backward, to handle the forward and backward dependencies of a time series, respectively. This allows for joint modeling of the contextual information of the entire input sequence, 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 to address the above-mentioned technical problems. The present invention processes the collected partial discharge signal using the continuous wavelet transform method (CWT) to generate a time-frequency diagram of a single pulse, and further processes it using the MCNN-BiLSTM-Attention method to accurately identify the partial discharge type, thereby ensuring the long-term and reliable operation of electrical equipment.

[0010] In one aspect, a partial discharge pattern recognition method includes:

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

[0012] S2, pre-processing the partial discharge signal to obtain a single pulse signal;

[0013] S3, calculating the CWT coefficient of the unit pulse of the partial discharge, and generating a time-frequency diagram of the single pulse based on the CWT coefficient;

[0014] 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;

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

[0016] S6 inputs the spliced ​​sequence into the trained attention mechanism Attention network, focuses on the key segments of the spliced ​​sequence, extracts global semantic features, inputs the global semantic features into the fully connected layer, and outputs the discharge type or fault type used to characterize the partial discharge signal through the classification function.

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

[0018] High-frequency current sensors are used to collect transient signals of various typical partial discharge types within one power frequency cycle.

[0019] Preferably, the S2 specifically includes:

[0020] The partial discharge signal pulse is extracted within a power frequency cycle to obtain a single pulse signal of partial discharge, and the interception width of each pulse is set to 50us.

[0021] Preferably, in S3, the CWT coefficient of the unit pulse is calculated as follows:

[0022] ;

[0023] in, Represents the scale parameter, which represents the expansion and contraction factor of the wavelet basis function on the time axis; It 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; Indicates time.

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

[0025] 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:

[0026] ;

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

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

[0029] Preferably, in S4, the multi-scale convolutional neural network MCNN consists of two parallel convolution channels, and features are merged through an additive fusion layer; the structures of the two convolution channels include a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer and a fully connected layer; the first channel uses convolution kernels of multiple sizes to extract local features; the second channel uses convolution kernels to extract global features; wherein, the size of the convolution kernel of the first channel is smaller than the size of the convolution kernel of the second channel.

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

[0031] The two parallel convolution channels of the multi-scale convolutional neural network (MCNN) extract the local and global features of the input time-frequency graph respectively. After the two channels complete the convolution operation and feature extraction and pass through the fully connected layer, the outputs of the two channels are fused by "element-by-element addition" to generate the final spatial feature representation.

[0032] Preferably, the S5 specifically includes:

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

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

[0035] The one-dimensional feature sequence is input into a bidirectional long short-term memory (BiLSTM) neural network for temporal feature extraction. The BiLSTM neural network consists of two substructures: a forward LSTM network and a reverse LSTM network. The forward LSTM processes the features of each time step in chronological order to capture historical dependency information. The reverse LSTM processes the feature sequence in reverse chronological order to extract the potential impact of future states on current features.

[0036] The output results of the forward and backward LSTM networks are concatenated in the feature dimension to form a temporal feature representation that includes forward and backward semantic information.

[0037] Preferably, the S6 specifically includes:

[0038] The concatenated sequence is introduced as input to the Attention network, which calculates the importance score of each time step output in the sequence and assigns a corresponding weight coefficient based on the importance score;

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

[0040] The semantic representation vector is input into the fully connected layer and finally normalized by the Softmax function to output the corresponding multi-category classification probability distribution, which is used to characterize the discharge type or fault category to which the partial discharge signal belongs.

[0041] In another aspect, a partial discharge pattern recognition device includes:

[0042] A partial discharge signal acquisition module is used to acquire the collected partial discharge signals of high-voltage electrical equipment;

[0043] A single pulse signal acquisition module is used to pre-process the partial discharge signal to obtain a single pulse signal;

[0044] A single pulse time-frequency diagram generation module is used to calculate the CWT coefficient of the unit pulse of partial discharge and generate the time-frequency diagram of the single pulse based on the CWT coefficient;

[0045] A multi-scale convolutional neural network processing module is used to input the time-frequency map 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;

[0046] 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 spliced ​​sequence of partial discharge features; the spliced ​​sequence includes a temporal feature representation of forward and backward semantic information;

[0047] The Attention network processing module is used to input the spliced ​​sequence into the trained Attention network, focus on the key segments of the spliced ​​sequence, extract global semantic features, input the global semantic features into the fully connected layer, and output the discharge type or fault type used to characterize the partial discharge signal through the classification function.

[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 7A t-SNE two-dimensional visualization diagram before final classification of the embodiment of the application after training;

[0058] Figure 8 A schematic diagram of a partial discharge pattern recognition device of the embodiment of the application. DETAILED DESCRIPTION

[0059] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0060] As shown in the figure, the partial discharge pattern recognition method provided by the embodiment comprises the following steps. Figure 1

[0061] S1, acquiring the collected partial discharge signal of the high-voltage electrical equipment.

[0062] Specifically, the transient signal of a plurality of typical partial discharge types in one power frequency cycle is collected by using a high-frequency current sensor. The power frequency cycle is, for example, 20 ms / 50 Hz.

[0063] S2, preprocessing the partial discharge signal to obtain a single pulse signal.

[0064] Specifically, the partial discharge signal pulse in one power frequency cycle is extracted to obtain a single pulse signal of the partial discharge, and the interception width of each pulse can be set to 50 μs.

[0065] S3, calculating the CWT coefficient of the unit pulse of the partial discharge, and generating a time-frequency diagram of the single pulse based on the CWT coefficient.

[0066] Specifically, the calculation method of the CWT coefficient of the unit pulse is as follows:

[0067] ;

[0068] wherein, denotes a scale parameter, and denotes a stretching factor of the wavelet base function on the time axis; denotes a translation position of the wavelet function on the time axis, and is used to analyze the local characteristics of the signal at a specific time; x(t) is the single pulse signal of the partial discharge; denotes time.

[0069] Further, the time-frequency diagram of the single pulse is generated based on the CWT coefficient, and specifically comprises:

[0070] ​The absolute value of the CWT coefficient represents the energy intensity of the signal at different frequencies and time points, and converts the scale parameter a into the actual frequency f, and the specific formula is as follows:

[0071] ;

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

[0073] The horizontal axis is set as the time axis of the time-frequency diagram, covering the complete duration (50 μs) of the single pulse signal; the vertical axis is the frequency axis, covering the main frequency band range of the single pulse signal of partial discharge. The time-frequency diagram uses Jet color spectrum to represent energy intensity, and the more the color deviates from warm color, the stronger the energy. The finally generated time-frequency diagram image size is 64x64x3, where 64x64 represents that the resolution of the image is 64 rows x 64 columns of pixels, and 3 represents three color channels in RGB format.

[0074] 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 spatial feature representation.

[0075] Specifically, the constructed MCNN is composed of two parallel convolution channels, and the features are merged through an addition fusion layer. The structure of the two channels includes a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer and a fully connected layer. The first channel adopts multiple small size convolution kernels (3x3 and 4x4) to enhance the perception ability of local detail features and slight changes, which is suitable for capturing local texture and edge information; the second channel uses a larger size convolution kernel (7x7) to extract structural features with more global perception ability, which is helpful to 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 diagram is used as the input of the constructed MCNN network. The two parallel convolution channels contained in the network are respectively used to extract the local features and global features of the input time-frequency diagram. After the convolution operation, feature extraction and processing through the full connection layer are completed in each channel, the outputs of the two channels are fused in the form of "element-by-element addition" to generate the final spatial feature representation. The fused features have both local detail perception and global structure understanding ability, realizing the collaborative expression of local and global information.

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

[0079] Specifically, the fused feature vector output by the MCNN network is flattened and converted into a one-dimensional feature sequence. Subsequently, the sequence is divided into several time steps, each time step corresponding to a local area or feature block in the original time-frequency graph to ensure the structural consistency of subsequent time series modeling. Next, the feature sequence is input into the BiLSTM for time series feature extraction, where the BiLSTM network contains two substructures: a forward LSTM network and a reverse LSTM network. The forward LSTM processes the features of each time step step by step in chronological order to capture historical dependency information; the reverse LSTM processes the feature sequence in reverse chronological order to extract the potential impact of future states on current features. Finally, the output results of the forward and reverse LSTM networks are spliced ​​in the feature dimension to form a time series feature representation that contains both forward and backward semantic information.

[0080] S6 inputs the spliced ​​sequence into the trained attention mechanism Attention network, focuses on the key segments of the spliced ​​sequence, extracts global semantic features, inputs the global semantic features into the fully connected layer, and outputs the discharge type and fault type used to characterize the partial discharge signal through the classification function.

[0081] Specifically, the concatenated sequence output by the BiLSTM network is introduced as input into the Attention network. The Attention network calculates the importance score for each time step in the sequence and assigns a corresponding weight coefficient based on the score. The feature vectors of all time steps are then weighted and summed according to the weights to produce a semantic representation vector that incorporates global context information. This semantic representation vector is then input into a fully connected layer and finally normalized using a Softmax function to output a corresponding multi-class classification probability distribution, which characterizes the discharge type or fault category to which the input sample belongs.

[0082] The technical solutions of the embodiments of this application will be further explained through specific examples as follows.

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

[0084] Depend on Figure 2It can be seen that a plurality of pulse signals are contained in one power frequency cycle, so the original signals collected are segmented into single pulse signals, as shown in Figure 3 200 single pulse signals are extracted for each of the three fault types, CWT is used to generate the time-frequency diagram of the single pulse of partial discharge, and a data label is added, label 1 represents surface discharge, label 2 represents bubble discharge, and label 3 represents needle-plate discharge. After setting the labels, the training set and the test set are divided, and 160 data of each fault type are selected as the training set and 40 data are selected as the test set.

[0085] As shown in Figure 4 The training set data is input into the MCNN-BiLSTM-Attention model built for training, and cross-entropy is used as the loss function of the model. When the convergence condition is reached, the training is completed, and the model parameters are saved. The test set is input into the trained model, and the diagnostic results of the test set are as shown in Figure 5 The diagnostic accuracy of surface discharge is 85%, the diagnostic accuracy of bubble discharge is 85%, the diagnostic accuracy of needle-plate discharge is 100%, and the average diagnostic accuracy is 90.83%.

[0086] To verify the effectiveness of the feature extraction of the MCNN-BiLSTM-Attention model, t-SNE is used to visualize the features of the partial discharge signal in two dimensions, wherein Figure 6 is the original time-frequency diagram feature before the model has learned, Figure 7 is the feature extracted before the final classification of the model. By comparing the two graphs, it can be seen that before feature extraction, the three types of faults overlap significantly in two-dimensional space, and the clustering is not clear. After the model is trained, the three types of fault samples are obviously clustered, and each type forms an independent and compact cluster, indicating that the model can successfully extract features with strong discriminability and high discrimination after training.

[0087] As shown in Figure 8 The embodiment also discloses a partial discharge pattern recognition device, which comprises:

[0088] A partial discharge signal acquisition module is configured to acquire the partial discharge signal of the high-voltage electrical equipment collected;

[0089] A single pulse signal acquisition module is configured to pre-process the partial discharge signal to obtain a single pulse signal;

[0090] A single pulse time-frequency diagram generation module is configured to calculate the CWT coefficient of the unit pulse of the partial discharge, and generate a time-frequency diagram of the single pulse based on the CWT coefficient;

[0091] A multi-scale convolutional neural network processing module is used to input the time-frequency map 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;

[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 spliced ​​sequence of partial discharge features; the spliced ​​sequence includes a temporal feature representation of forward and backward semantic information;

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

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

[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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A partial discharge pattern recognition method, characterized in that: include: S1, obtaining the collected partial discharge signal of the high-voltage electrical equipment; S2, pre-processing the partial discharge signal to obtain a single pulse signal; S3, calculating the CWT coefficient of the unit pulse of the partial discharge, and generating a time-frequency diagram of the single pulse based on the CWT coefficient; 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; S5, inputting the spatial feature representation into a trained bidirectional long short-term memory neural network BiLSTM to obtain a spliced ​​sequence of partial discharge features; the spliced ​​sequence includes temporal feature representations of forward and backward semantic information; S6, inputs the spliced ​​sequence into the trained attention mechanism Attention network, focuses on the key segments of the spliced ​​sequence, extracts global semantic features, inputs the global semantic features into the fully connected layer, and outputs the discharge type or fault type used to characterize the partial discharge signal through the classification function; In S3, the CWT coefficient of the unit pulse is calculated as follows: Where a is the scale parameter, which represents the expansion factor of the wavelet basis function on the time axis; b is 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; t is time; In S3, generating a time-frequency diagram of a single pulse based on the CWT coefficients specifically includes: 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: Among them, f center is the center frequency of the wavelet basis, Δt is the sampling interval of the partial discharge single pulse signal; The horizontal axis is set as the time axis of the time-frequency diagram, covering the complete duration of a single pulse signal; the vertical axis is the frequency axis, covering the frequency band range of a single pulse signal of partial discharge; the time-frequency diagram uses the Jet color spectrum to represent the energy intensity, and the final image size is 64×64×3, where 64×64 indicates that the image resolution is 64 rows × 64 columns of pixels, and 3 represents the three color channels in RGB format; Said 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, each of which corresponds to a local area or feature block in the original time-frequency graph to ensure the structural consistency of subsequent time series modeling; The one-dimensional feature sequence is input into a bidirectional long short-term memory (BiLSTM) neural network for temporal feature extraction. The BiLSTM neural network consists of two substructures: a forward LSTM network and a reverse LSTM network. The forward LSTM processes the features of each time step in chronological order to capture historical dependency information. The reverse LSTM processes the feature sequence in reverse chronological order to extract the potential impact of future states on current features. The output results of the forward and backward LSTM networks are concatenated in the feature dimension to form a temporal feature representation that includes forward and backward semantic information.

2. The partial discharge pattern recognition method according to claim 1, characterized in that: In S1, the collected partial discharge signal is specifically: High-frequency current sensors are used to collect transient signals of various typical partial discharge types within one power frequency cycle.

3. The partial discharge pattern recognition method according to claim 1, characterized in that: Said S2 specifically includes: The partial discharge signal pulse is extracted within a power frequency cycle to obtain a single pulse signal of partial discharge, and the interception width of each pulse is set to 50us.

4. The partial discharge pattern recognition method according to claim 1, characterized in that: In S4, the multi-scale convolutional neural network MCNN consists of two parallel convolution channels, and features are merged through an additive fusion layer; the structures of the two convolution channels include a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer and a fully connected layer; the first channel uses convolution kernels of multiple sizes to extract local features; the second channel uses convolution kernels to extract global features; wherein, the size of the convolution kernel of the first channel is smaller than the size of the convolution kernel of the second channel.

5. The partial discharge pattern recognition method according to claim 4, characterized in that: In S4, the multi-scale convolutional neural network MCNN performs feature fusion to obtain spatial feature representation, specifically including: The two parallel convolution channels of the multi-scale convolutional neural network (MCNN) extract local and global features of the input time-frequency map respectively. After the two channels complete the convolution operation and feature extraction and pass through the fully connected layer, the outputs of the two channels are fused using the "element-by-element addition" method to generate the final spatial feature representation.

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

7. A partial discharge pattern recognition device, characterized in that: The method for identifying a partial discharge pattern according to any one of claims 1 to 6 comprises: A partial discharge signal acquisition module is used to acquire the collected partial discharge signals of high-voltage electrical equipment; A single pulse signal acquisition module is used to pre-process the partial discharge signal to obtain a single pulse signal; A single pulse time-frequency diagram generation module is used to calculate the CWT coefficient of the unit pulse of partial discharge and generate the time-frequency diagram of the single pulse based on the CWT coefficient; A multi-scale convolutional neural network processing module is used to input the time-frequency map 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; 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 spliced ​​sequence of partial discharge features; the spliced ​​sequence includes a temporal feature representation of forward and backward semantic information; The Attention network processing module is used to input the spliced ​​sequence into the trained Attention network, focus on the key segments of the spliced ​​sequence, extract global semantic features, input the global semantic features into the fully connected layer, and output the discharge type or fault type used to characterize the partial discharge signal through the classification function.

Images