Partial discharge type identification method and system based on multi-feature extraction and fusion

Abstract

The present invention discloses a method and system for identifying local discharge types based on multi-feature extraction and fusion, the method comprising: extracting graph structure features from local discharge signal data to obtain graph structure features corresponding to the local discharge signal data, fusing period features and graph structure features, and performing layer normalization processing on the fused fusion features to obtain target fusion features; adding position coding to the target fusion features, and extracting time series features from the added features to be input to obtain a final feature vector; inputting the final feature vector into a pre-built local discharge recognition model, and the local discharge recognition model outputs the local discharge type corresponding to the final feature vector. The attention weight can be adaptively adjusted to enhance the model's responsiveness to changes in the input signal, thereby improving the accuracy of feature extraction and its adaptability to dynamic changes.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system monitoring and diagnosis, and in particular relates to a method and system for identifying partial discharge types based on multi-feature extraction and fusion. Background Art

[0002] In high-voltage cable systems, partial discharge (PD) is a localized electrical discharge caused by insulation defects, and is an important precursor to cable insulation degradation and failure. Therefore, the identification and diagnosis of PD is crucial to ensure the safe operation of power systems.

[0003] In the existing technology, the machine learning-based method uses traditional machine learning algorithms such as support vector machines (SVM) and decision trees to extract the features of partial discharge signals for classification. These methods have improved the recognition accuracy to a certain extent, but they are highly dependent on feature extraction and are difficult to adapt to complex and changing actual environments; the deep learning-based method uses deep learning models such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) to automatically extract features from partial discharge signals for recognition. Deep learning methods perform well in processing complex signals, but require a large amount of labeled data for training, the model has poor interpretability, and consumes a lot of computing resources. Summary of the invention

[0004] The present invention provides a partial discharge type recognition method and system based on multi-feature extraction and fusion, which are used to solve the technical problems of requiring a large amount of labeled data for training, poor model interpretability, and high consumption of computing resources.

[0005] In a first aspect, the present invention provides a method for identifying partial discharge types based on multi-feature extraction and fusion, comprising:

[0006] Obtain partial discharge signal data, and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis ;

[0007] Extracting graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ;

[0008] Add position encoding to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector ;

[0009] The final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0010] In a second aspect, the present invention provides a partial discharge type identification system based on multi-feature extraction and fusion, comprising:

[0011] An acquisition module is configured to acquire partial discharge signal data and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis. ;

[0012] The first extraction module is configured to extract graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data. , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ;

[0013] A second extraction module is configured to add the position code to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector ;

[0014] An output module is configured to convert the final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0015] In a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the local discharge type identification method based on multi-feature extraction and fusion of any embodiment of the present invention.

[0016] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor executes the steps of the method for identifying local discharge types based on multi-feature extraction and fusion according to any embodiment of the present invention.

[0017] The local discharge type identification method and system based on multi-feature extraction and fusion of the present application proposes a comprehensive importance measurement method combining node degree and node feature mean, through which the importance of each node is calculated. On this basis, an adaptive pooling strategy is adopted to prioritize the features of important nodes, thereby significantly improving the effectiveness and computational efficiency of feature extraction; in the process of temporal feature extraction, a dynamic perception attention mechanism is introduced. This mechanism can enhance the model's responsiveness to changes in input signals by adaptively adjusting attention weights, thereby improving the accuracy of feature extraction and its adaptability to dynamic changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 A flowchart of a method for identifying partial discharge types based on multi-feature extraction and fusion provided by an embodiment of the present invention;

[0020] Figure 2 A structural block diagram of a partial discharge type identification system based on multi-feature extraction and fusion provided by an embodiment of the present invention;

[0021] Figure 3 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1 , which shows a flow chart of a local discharge type identification method based on multi-feature extraction and fusion of the present application.

[0024] like Figure 1 As shown, the local discharge type recognition method based on multi-feature extraction and fusion specifically includes the following steps:

[0025] Step S101, obtaining partial discharge signal data, and extracting periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis .

[0026] In this step, the partial discharge signal is collected in real time by the sensor installed at the cable joint. The collected data is a time series signal, including information such as amplitude, period, phase, etc.

[0027] The collected signal is denoised using a continuous adaptive wavelet threshold denoising algorithm. The specific steps include performing wavelet transform on the signal to obtain wavelet coefficients of different scales and positions.

[0028] Wavelet transform: Perform wavelet transform on the signal to obtain wavelet coefficients of different scales and positions.

[0029] Adaptive threshold calculation: According to the statistical characteristics of noise, the threshold of each level of wavelet coefficient is adaptively calculated. The formula is as follows:

[0030] ,

[0031] In the formula, is the noise standard deviation, is the number of wavelet coefficients. By continuously and adaptively adjusting the threshold, dynamic denoising can be achieved at all scales.

[0032] Threshold processing: Apply the threshold to perform soft threshold processing on the wavelet coefficients, set the coefficients less than the threshold to zero, and subtract the threshold from the remaining coefficients. The processing formula is:

[0033] ,

[0034] in, is the threshold value, To adjust the parameters.

[0035] Inverse wavelet transform: Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the denoised signal.

[0036] It should be noted that the periodic feature extraction is:

[0037] Fast Fourier Transform: Perform fast Fourier transform on partial discharge signal data to extract frequency domain features , the frequency domain characteristics The expression is:

[0038] ,

[0039] In the formula, is the signal after denoising;

[0040] Spectral Analysis: Frequency Domain Characteristics Each frequency Corresponding to an amplitude, calculate each frequency The amplitude , and select the frequency with amplitude greater than the preset threshold as the target frequency , and the target frequency The corresponding amplitude Save as a periodic feature, the expression is:

[0041] ,

[0042] ,

[0043] In the formula, is the main frequency component set, which represents the set of frequencies whose amplitudes are greater than the preset threshold. For frequency The corresponding absolute value of the spectrum amplitude is is the amplitude corresponding to the main frequency component, is a frequency in the set of frequency components;

[0044] Empirical mode decomposition: Perform empirical mode decomposition on the partial discharge signal data and decompose it into several intrinsic mode functions and residuals , to obtain the characteristic components on different time scales, the expression is:

[0045] ,

[0046] In the formula, is the denoised signal, is the number of intrinsic mode functions (IMFs), For the Solid mode functions;

[0047] Adaptive signal decomposition: further decompose the IMFS to enhance the accuracy of period detection. The solid mode functions are analyzed and the periodic components are combined to form the overall periodic characteristics. The expression for analyzing the solid mode function is:

[0048] ,

[0049] In the formula, is the number of components of the intrinsic mode function, For the The jth component of the intrinsic mode function;

[0050] The expression of the overall periodic characteristics is:

[0051] ,

[0052] In the formula, is the periodic feature set after adaptive signal decomposition, is the frequency component, is the amplitude corresponding to the frequency component, For the The set of main frequencies corresponding to the jth component in the intrinsic mode function, For frequency The corresponding spectrum amplitude;

[0053] Short-time Fourier transform: Perform short-time Fourier transform on the partial discharge signal data to analyze the time-frequency characteristics of the partial discharge signal data and identify the change of frequency over time. The expression of short-time Fourier transform is:

[0054] ,

[0055] In the formula, is the time domain feature, indicating the moment and frequency The spectrum value under is the denoised signal, is a window function, is the time variable, is a complex exponential function, representing the kernel function of Fourier transform;

[0056] ,

[0057] In the formula, is the amplitude of the time-frequency feature;

[0058] ,

[0059] In the formula, is the time-frequency feature matrix, indicating that at each moment and frequency The power spectrum value under For the A point in time, For the frequency points;

[0060] Autocorrelation function: Use the autocorrelation function to analyze the partial discharge signal data. The expression of the autocorrelation function is:

[0061] ,

[0062] In the formula, is the autocorrelation function, which indicates that the signal is delayed in time. The correlation of time, is the denoised signal at time The value of is the denoised signal at time The value of

[0063] Period detection: Combine frequency domain analysis and autocorrelation function to detect the period of partial discharge signal data , the expression is:

[0064] ,

[0065] ,

[0066] In the formula, is the main frequency component, is the peak position of the autocorrelation function;

[0067] The periodic characteristics are obtained as follows:

[0068] ,

[0069] In the formula, is a set of periodic features, is the set of main frequency components, is the amplitude set corresponding to the main frequency components, For the The intrinsic mode functions, is the frequency corresponding to the intrinsic mode function, is the main period of the signal.

[0070] Step S102: extracting graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data. , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features .

[0071] In this step, the graph structure features are extracted as follows:

[0072] Signal representation as a graph structure: Assume that the signal in the partial discharge signal data is ,in, is the number of signal points, each signal point is regarded as a node in the graph, and each node Corresponding signal point ;

[0073] Defining a dynamic adjacency matrix : Use Euclidean distance to determine whether there is a connection, specifically: ,in, is the distance threshold;

[0074] Degree Matrix: The degree matrix is ​​a diagonal matrix where the diagonal elements Representation Node In time degree, that is, the degree of the node The number of connected nodes, the matrix size is , the diagonal elements The expression is:

[0075] ;

[0076] Each node The initial features are directly set to the corresponding signal values , then the initial feature matrix H ( 0 ) = [ s ′ 0 , s ′ 1 , … , s ′ N − 1 ] T The size is ;

[0077] Graph convolution: In graph convolution, feature matrix and dynamic adjacency matrix Combined with the degree matrix, it can capture the relationship between signals and obtain graph structure features. , the expression is:

[0078] ,

[0079] In the formula, is the activation function, for The spatial convolution operation, is the weight matrix of the initial convolutional layer, is the degree matrix.

[0080] It should be noted that the periodic characteristics and graph structural features The expression for fusion is: ,

[0081] In the formula, is the fusion weight of graph structure features, is the fusion weight of the periodic feature. Specifically, the coefficient is initialized as , use cross-validation to conduct multiple experiments on the training set, and select the best one based on the performance of the validation set and value.

[0082] Furthermore, an adaptive pooling method is introduced after the graph convolution operation and residual connection. The pooling strategy is adaptively selected according to the importance of the node, including:

[0083] Node importance measurement calculation: The node importance measurement calculation is performed by combining the node degree and the node feature mean. The expression is:

[0084] ,

[0085] In the formula, For Node The importance measure, For Node The degree of is the number of nodes connected to it. For Node The characteristic average value of is the weight of the node measurement, is the weight of the average value of node features; and The initial value is set to 1. During the experimental training process, the grid search optimization algorithm is used to select the optimal and value.

[0086] ,

[0087] In the formula, is the degree of node i, that is, the number of nodes connected to it, For Node In the The characteristic value of the dimension;

[0088] ,

[0089] In the formula, is the adjacency matrix element, representing the node and nodes Whether a connection exists;

[0090] Node selection: select nodes with a value greater than a threshold based on the comprehensive importance metric The node expression is:

[0091] ,

[0092] ,

[0093] In the formula, is the mean of the importance measure of the current node, is the standard deviation of the current node importance measure, To adjust the parameters, used to control the ratio of retained nodes; The range is generally between 0.5 and 2. Initialized to 1, by changing different The data set is divided into a training set and a validation set. The impact of the value on the model performance, choose the best value.

[0094] Feature retention and pooling: When performing feature retention, weighted pooling is performed on the selected node features. The expression of weighted pooling is:

[0095] ,

[0096] Weighted pooled feature: features after weighted pooling;

[0097] In the formula, is the set of important nodes selected, For Node The weighting coefficient of For Node The feature matrix of

[0098] Multi-scale feature extraction: Use convolution kernels of different scales to perform convolution operations on the input features, and the multi-scale features are represented as follows:

[0099] ,

[0100] In the formula, is the fused feature matrix, For the The weight coefficients of each scale are initialized using five-fold cross validation. Different weight coefficient combinations are set. Five-fold validation is performed on each weight coefficient combination. The performance of each combination on the validation set is calculated. The weight coefficient combination with the best performance is selected. For the Adaptive pooling results at each scale;

[0101] Adaptive graph convolution pooling: Adaptive graph convolution is introduced in the pooling process to obtain the feature matrix after adaptive graph convolution, which is expressed as:

[0102] ,

[0103] In the formula, is the feature matrix after adaptive graph convolution, and are the adaptive adjacency matrix and degree matrix respectively, is the activation function, is the weight matrix.

[0104] In this embodiment, the periodic features are fused with the graph structure features before adaptive graph pooling, which has the following effects compared to simply putting the periodic features alone for the final fusion: Deep optimization of feature fusion: Through multi-level processing of fused features, higher-level and more discerning features can be extracted, thereby improving the overall performance of the model; early fusion allows the periodic features and graph structure features to complement and enhance each other during the convolution process; early fusion can process all features under a unified framework to ensure the consistency and unity of the model; multi-layer graph convolution and residual connection can effectively prevent feature degradation and ensure the stability of the fused features in subsequent processing, thereby improving the robustness and generalization ability of the model. If the periodic features are retained alone, lacking this multi-level optimization and enhancement, the robustness and generalization ability of the model may be limited.

[0105] Step S103, adding the position code to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector .

[0106] In this step, the position encoding is added to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector include:

[0107] Adaptive position coding: Adaptive position coding is added to the feature sequence, generated using a neural network. The input is the time step and signal features, and the output is the adaptive position coding. Finally, a feature matrix containing the position coding is generated. , a feature matrix containing positional encodings The expression is:

[0108] ,

[0109] ,

[0110] ,

[0111] In the formula, is the adaptive position encoding, is the original time series data, A neural network is used to generate positional encodings, For the original data The function to extract time step information from is the time step information;

[0112] Multi-scale feature extraction: Before extracting temporal features, the input features are processed by a multi-scale convolution layer. A multi-scale convolution layer is added to extract multi-scale features through convolution kernels of different sizes. The expression of multi-scale features is:

[0113] ,

[0114] In the formula, For splicing operation, is the multi-scale feature matrix, , , are convolution operations of different scales, is the encoded feature matrix;

[0115] Dynamically aware attention mechanism, which includes generating queries, keys and values: For multi-scale features , generate the corresponding query matrix, key matrix and value matrix, the expression is:

[0116] ,

[0117] ,

[0118] ,

[0119] In the formula, is the query matrix, is the weight matrix of the query matrix, is the bias term of the key matrix, is the key matrix, is the weight matrix of the key matrix, is the bias term of the key matrix, is the value matrix, is the weight matrix of the value matrix, is the bias term of the value matrix;

[0120] Initialization of weight matrix and bias term: Xaier initialization is used to initialize the weight matrix and bias term. The initialization expression is:

[0121] ,

[0122] In the formula, is any one of the weight matrix of the query matrix, the weight matrix of the key matrix, and the weight matrix of the value matrix, For uniform distribution, is the input dimension, is the output dimension;

[0123] Feature space projection: Through linear transformation, the input features are projected into different feature spaces. The query matrix is ​​projected into one feature space to ask questions, the key matrix is ​​projected into another feature space to provide clues to the answer, and the value matrix contains the actual answer information.

[0124] Dynamic perception: input features Perform a linear transformation to generate a preliminary bias term , is the weight matrix of the linear transformation, is the bias term of the linear transformation;

[0125] Perform nonlinear activation on the result after linear transformation to generate dynamic bias terms ;

[0126] The dynamic bias term Introduced into the calculation of attention weights for adaptive adjustment;

[0127] Calculate the attention weights: Calculate the dot product of the query matrix Q and the key matrix K and scale it by dividing by the square root of the key dimension: in, Represents the dot product of the query matrix and the key matrix, divided by the square root of the key dimension To scale;

[0128] The dynamic bias term Introduced into the dot product result, the formula is: ;

[0129] The attention score is normalized through the softmax function to generate the final attention weight A;

[0130] Calculate the attention output: The calculated final attention weight A is applied to the value matrix V to obtain the final attention output , the attention output calculation formula is:

[0131] ,

[0132] In the formula, For attention output, is the adjacency matrix element, representing the node and nodes Is there a connection? is the j-th value vector.

[0133] Step S104: The final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0134] In this step, the attention output generated by the dynamic-aware attention mechanism can be passed as input to the multi-layer Transformer encoder for further processing and extracting long-term dependencies. The Transformer encoder is a model based on the multi-head attention mechanism that can effectively capture complex dependencies in sequence data.

[0135] Input layer: The output of the dynamic perception attention mechanism AttentionOutput is used as the input of the Transformer encoder : ;

[0136] Transformer encoder layer: The Transformer encoder is composed of multiple layers stacked together, each of which contains a multi-head attention mechanism and a feed-forward neural network (FFN). The specific formula is as follows:

[0137] ,

[0138] in, is the Transformer encoder The output feature matrix of the layer, is the Transformer encoder The input feature matrix of the layer, It is a single-layer structure in the Transformer encoder, which is used to transform the Transformer encoder The input feature matrix of the layer is processed by the multi-head attention mechanism and the feedforward neural network, and the output Transformer encoder is The output feature matrix of the layer, , that is, the output of the dynamic perception attention mechanism is used as the first layer The input of the layer. in, Including multi-head attention mechanism and feedforward neural network , , are the weight matrices of the first and second layers in the feedforward neural network, , are the bias items of the first and second layers respectively. The L-th layer output feature matrix of the encoder ,in yes The Lth layer of the encoder outputs a feature matrix.

[0139] The final feature vector extracted Input into the model for classification and recognition.

[0140] Classification Models and Optimization

[0141] Eigenvector generation: The final eigenvector matrix Perform a flatten operation to generate the eigenvector z: Through the flatten operation, the feature matrix is ​​converted into the input feature vector z for subsequent classification and regression tasks.

[0142] Self-attention mechanism: Use the self-attention mechanism to weight the feature vector, enhance the feature selection ability, and generate the attention feature vector .

[0143] Self-Attention Mechanism: Compute Query (Query, Key and Value):

[0144] , where z is the input feature vector, in, and are the first Row and OK, Is the dimension of the key. Apply the Softmax function: Weighted sum: in, is the attention output of the i-th query, is the attention score between the i-th query and the j-th key, is the attention score between the i-th query and the k-th key, is the attention weight between the i-th query and the j-th key, is the value matrix Row. Output attention vector: .

[0145] Classification model application: Use Fully Connected Neural Network (FCNN) to focus on feature vectors Make classification predictions.

[0146] Input layer: attention vector .

[0147] Hidden layer: contains 3 fully connected layers, the first fully connected layer , the second fully connected layer , the third fully connected layer .

[0148] Output layer: Use the softmax activation function to output the classification results:

[0149] Adversarial training (performed during model training): Adversarial training is introduced to improve the robustness and generalization ability of the model. Adversarial training generates adversarial samples: in is the eigenvector, is the amplitude of the disturbance, indicating the size of the disturbance, is the sign function that determines the direction of the gradient. is the classification loss function relative to the feature vector gradient.

[0150] Adversarial Sample Training: , and are the weights and biases of the classification model. Adversarial loss function: ,in, It is against loss, is the value of the cth class of the true label.

[0151] Total loss function: ,in, is the weight against the loss, is a regularization term used to prevent overfitting. is the regularization strength.

[0152] Model training and prediction: using the total loss function Perform model training and update model parameters and To minimize the loss. Apply the trained model on the test data to generate the final classification prediction results .

[0153] In this embodiment, the partial discharge type identification method based on multi-feature extraction and fusion has the following effects:

[0154] Improve recognition accuracy: Through the adaptive graph pooling layer, the importance of nodes is calculated by comprehensively utilizing the node degree and the mean value of node features, and the features of important nodes are retained first, which effectively improves the accuracy of feature extraction and thus significantly improves the accuracy of partial discharge type recognition.

[0155] Enhanced noise resistance: The dynamic perception attention mechanism can flexibly respond to changes in input signals by adaptively adjusting attention weights, effectively suppress environmental noise and electromagnetic interference, and improve the noise resistance of the recognition system.

[0156] Improve real-time performance: The optimized adaptive graph pooling layer and dynamic perception attention mechanism excel in computational efficiency of feature extraction and signal processing, meeting the real-time and high-efficiency requirements of the power system, enabling the recognition system to maintain efficient operation in large-scale real-time data processing.

[0157] Enhanced adaptability: The adaptive graph pooling layer and dynamic perception attention mechanism enable the recognition system to adapt to different power equipment and complex environments, with good versatility and adaptability, ensuring stable and reliable operation in various practical scenarios.

[0158] Dynamic perception capability: The dynamic perception attention mechanism performs well in processing dynamically changing signals. It can flexibly adjust the attention weight according to the changes in the signal, thereby improving the processing and response capabilities for complex signals.

[0159] Enhanced robustness and generalization capabilities: The adversarial training mechanism is introduced during the model training process to improve the robustness and generalization capabilities of the model, ensuring stable performance in different data sets and practical applications.

[0160] Effectiveness and computational efficiency: The adaptive graph pooling layer ensures the effectiveness of feature retention by adaptively selecting important nodes for pooling, while reducing feature dimensions and improving computational efficiency, allowing the system to operate efficiently in an environment with limited resources.

[0161] High-order feature extraction: The combination of multi-layer graph convolutional networks and residual connections makes the feature extraction level deeper, which can capture the high-order features in the signal and further improve the performance of the recognition system.

[0162] In summary, the method of this application proposes a comprehensive importance measurement method that combines node degree and node feature mean, and calculates the importance of each node through this method. On this basis, an adaptive pooling strategy is adopted to prioritize the features of important nodes, thereby significantly improving the effectiveness and computational efficiency of feature extraction; in the process of temporal feature extraction, a dynamic perception attention mechanism is introduced. This mechanism can enhance the model's responsiveness to changes in input signals by adaptively adjusting attention weights, thereby improving the accuracy of feature extraction and its adaptability to dynamic changes.

[0163] See also Figure 2 , which shows a structural block diagram of a local discharge type identification system based on multi-feature extraction and fusion of the present application.

[0164] like Figure 2 As shown, the partial discharge type identification system 200 includes an acquisition module 210 , a first extraction module 220 , a second extraction module 230 and an output module 240 .

[0165] The acquisition module 210 is configured to acquire partial discharge signal data and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis. ;

[0166] The first extraction module 220 is configured to extract the graph structure feature of the partial discharge signal data to obtain the graph structure feature corresponding to the partial discharge signal data. , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ;

[0167] The second extraction module 230 is configured to add the position code to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector ;

[0168] Output module 240 is configured to output the final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0169] It should be understood that Figure 2 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 2 The modules in it will not be described in detail here.

[0170] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the partial discharge type identification method based on multi-feature extraction and fusion in any of the above method embodiments;

[0171] As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0172] Acquire partial discharge signal data, and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis ;

[0173] Extracting graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ;

[0174] Add position encoding to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector ;

[0175] The final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0176] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the partial discharge type identification system based on multi-feature extraction and fusion, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the partial discharge type identification system based on multi-feature extraction and fusion via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0177] Figure 3 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 3In the example, the bus connection is used. The memory 320 is the computer-readable storage medium mentioned above. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the partial discharge type identification method based on multi-feature extraction and fusion of the above method embodiment is implemented. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the partial discharge type identification system based on multi-feature extraction and fusion. The output device 340 may include display devices such as display screens.

[0178] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0179] As an implementation mode, the electronic device is applied to a partial discharge type identification system based on multi-feature extraction and fusion, and is used for a client, and includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can:

[0180] Obtain partial discharge signal data, and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis ;

[0181] Extracting graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ;

[0182] Add position encoding to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector ;

[0183] The final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

[0184] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0185] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying partial discharge types based on multi-feature extraction and fusion, characterized in that: include: Acquire partial discharge signal data, and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis ; Extracting graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ; Add position encoding to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector , wherein the position encoding is added to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector include: Adaptive position coding: Adaptive position coding is added to the feature sequence, generated using a neural network, with the input being the time step and signal features, the output being the adaptive position coding, and finally generating a feature matrix containing the position coding , a feature matrix containing positional encodings The expression is: ， ， ， In the formula, is the adaptive position encoding, is the original time series data, A neural network is used to generate positional encodings, is a function used to extract time step information from the raw data, is the time step information; Multi-scale feature extraction: Before extracting temporal features, the input features are processed by a multi-scale convolution layer. A multi-scale convolution layer is added to extract multi-scale features through convolution kernels of different sizes. The expression of multi-scale features is: ， In the formula, For splicing operation, is the multi-scale feature matrix, , , are convolution operations of different scales, is the encoded feature matrix; Dynamically aware attention mechanism, which includes generating queries, keys and values: For multi-scale features , generate the corresponding query matrix, key matrix and value matrix, the expression is: ， ， ， In the formula, is the query matrix, is the weight matrix of the query matrix, is the bias term of the bond matrix, is the key matrix, is the weight matrix of the key matrix, is the bias term of the bond matrix, is the value matrix, is the weight matrix of the value matrix, is the bias term of the value matrix; Initialization of weight matrix and bias term: Xaier initialization is used to initialize the weight matrix and bias term. The initialization expression is: ， In the formula, is any one of the weight matrix of the query matrix, the weight matrix of the key matrix, and the weight matrix of the value matrix, For uniform distribution, is the input dimension, is the output dimension; Feature space projection: Through linear transformation, the input features are projected into different feature spaces. The query matrix is ​​projected into one feature space to ask questions, the key matrix is ​​projected into another feature space to provide clues to the answer, and the value matrix contains the actual answer information. Dynamic perception: input features Perform a linear transformation to generate a preliminary bias term , is the weight matrix of the linear transformation, is the bias term of the linear transformation; Perform nonlinear activation on the result after linear transformation to generate dynamic bias terms ; The dynamic bias term Introduced into the calculation of attention weights for adaptive adjustment; Calculate the attention weights: Calculate the dot product of the query matrix Q and the key matrix K and scale it by dividing by the square root of the key dimension: in, Represents the dot product of the query matrix and the key matrix, divided by the square root of the key dimension To scale; The dynamic bias term Introduced into the dot product result, the formula is: ; The attention score is normalized through the softmax function to generate the final attention weight A; Calculate the attention output: The calculated final attention weight A is applied to the value matrix V to obtain the final attention output , the attention output calculation formula is: ， In the formula, For attention output, is an element of the adjacency matrix, indicating whether there is a connection between nodes. is the j-th value vector; The final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

2. The method for identifying partial discharge types based on multi-feature extraction and fusion according to claim 1, characterized in that: The extracting of periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis comprises: Fast Fourier Transform: Perform fast Fourier transform on partial discharge signal data to extract frequency domain features , the frequency domain characteristics The expression is: ， In the formula, is the signal after denoising; Spectral Analysis: Frequency Domain Characteristics Each frequency Corresponding to an amplitude, calculate each frequency The amplitude , and select the frequency with amplitude greater than the preset threshold as the target frequency , and the target frequency The corresponding amplitude Save as a periodic feature, the expression is: ， ， In the formula, is the main frequency component set, which represents the set of frequencies whose amplitudes are greater than the preset threshold. For frequency The corresponding absolute value of the spectrum amplitude is is the amplitude corresponding to the main frequency component, is a frequency in the set of frequency components; Empirical mode decomposition: Perform empirical mode decomposition on the partial discharge signal data and decompose it into several intrinsic mode functions and residuals , to obtain the characteristic components on different time scales, the expression is: ， In the formula, is the denoised signal, is the number of intrinsic mode functions (IMFs), For the Solid mode functions; Adaptive signal decomposition: The solid mode functions are analyzed and the periodic components are combined to form the overall periodic characteristics. The expression for analyzing the solid mode function is: ， In the formula, is the number of components of the intrinsic mode function, For the The jth component of the intrinsic mode function; The expression of the overall periodic characteristics is: ， In the formula, is the periodic feature set after adaptive signal decomposition, is the frequency component, is the amplitude corresponding to the frequency component, For the The set of main frequencies corresponding to the jth component in the intrinsic mode function, For frequency The corresponding spectrum amplitude; Short-time Fourier transform: Perform short-time Fourier transform on the partial discharge signal data to analyze the time-frequency characteristics of the partial discharge signal data and identify the change of frequency over time. The expression of short-time Fourier transform is: ， In the formula, is the time domain feature, which represents the spectrum value at time and frequency. is the denoised signal, is the window function, is the time variable, is a complex exponential function, representing the kernel function of Fourier transform; ， In the formula, is the amplitude of the time-frequency feature; ， In the formula, is the time-frequency feature matrix, indicating that at each moment and frequency The power spectrum value under For the first time point, is the th frequency point; Autocorrelation function: Use the autocorrelation function to analyze the partial discharge signal data. The expression of the autocorrelation function is: ， In the formula, is the autocorrelation function, which indicates that the signal is delayed in time. The correlation of time, is the value of the denoised signal at time, is the denoised signal at time The value of Period detection: Combine frequency domain analysis and autocorrelation function to detect the period of partial discharge signal data , the expression is: ， ， In the formula, is the main frequency component, is the peak position of the autocorrelation function; The periodic characteristics are obtained as follows: ， In the formula, is a set of periodic features, is the set of main frequency components, is the amplitude set corresponding to the main frequency components, For the The intrinsic mode functions, is the frequency corresponding to the intrinsic mode function, is the main period of the signal.

3. The method for identifying partial discharge types based on multi-feature extraction and fusion according to claim 1, characterized in that: The graph structure feature extraction is performed on the partial discharge signal data to obtain a graph structure feature corresponding to the partial discharge signal data. include: Signal representation as a graph structure: Assume that the signal in the partial discharge signal data is , where is the number of signal points. Each signal point is regarded as a node in the graph. Corresponding signal point ; Defining a dynamic adjacency matrix : Use Euclidean distance to determine whether there is a connection, specifically: ,in, is the distance threshold; Degree Matrix: The degree matrix is ​​a diagonal matrix where the diagonal elements Representation Node In time degree, that is, the degree of the node The number of connected nodes, the matrix size is , the diagonal elements The expression is: ； Each node The initial features are directly set to the corresponding signal values , then the size of the initial feature matrix is ; Graph convolution: In graph convolution, feature matrix and dynamic adjacency matrix Combined with the degree matrix, it can capture the relationship between signals and obtain graph structure features. , the expression is: ， In the formula, is the activation function, for The spatial convolution operation, is the weight matrix of the initial convolutional layer, is the degree matrix.

4. The method for identifying partial discharge types based on multi-feature extraction and fusion according to claim 1, characterized in that: The periodic characteristics and graph structural features The expression for fusion is: ， In the formula, is the fusion weight of graph structure features, is the fusion weight of periodic features.

5. The method for identifying partial discharge types based on multi-feature extraction and fusion according to claim 1, characterized in that: The fusion features after fusion Perform layer normalization to obtain target fusion features include: Adaptive graph pooling: Adaptive pooling is introduced after graph convolution and residual connection, and the pooling strategy is adaptively selected according to the importance metric of the nodes. Layer normalization: Perform layer normalization on the generated feature matrix to obtain the target fusion feature , where the target fusion feature The expression is: ， In the formula, Normalize the layer. is the feature matrix after adaptive graph convolution.

6. The method for identifying partial discharge types based on multi-feature extraction and fusion according to claim 5, characterized in that: The adaptive pooling method is introduced after the graph convolution operation and the residual connection, and the pooling strategy is adaptively selected according to the importance of the node, including: Node importance measurement calculation: The node importance measurement calculation is performed by combining the node degree and the node feature mean. The expression is: ， In the formula, is the importance measure of the node, is the degree of a node, i.e. the number of nodes connected to it, is the feature average of the node, is the weight of the node measurement, is the weight of the average value of node features; ， In the formula, is the degree of node i, that is, the number of nodes connected to it, is the eigenvalue of the node in the th dimension; ， In the formula, It is an element of the adjacency matrix, indicating whether there is a connection between nodes; Node selection: select nodes with a value greater than a threshold based on the comprehensive importance metric The node expression is: ， ， In the formula, is the mean of the importance measure of the current node, is the standard deviation of the current node importance measure, To adjust the parameters, used to control the ratio of retained nodes; Feature retention and pooling: When performing feature retention, weighted pooling is performed on the selected node features. The expression of weighted pooling is: ， In the formula, is the set of important nodes selected, is the weight coefficient of the node, is the feature matrix of the node; Multi-scale feature extraction: Use convolution kernels of different scales to perform convolution operations on the input features, and the multi-scale features are represented as follows: ， In the formula, is the fused feature matrix, For the The weight coefficient of each scale, For the Adaptive pooling results at each scale; Adaptive graph convolution pooling: Adaptive graph convolution is introduced in the pooling process to obtain the feature matrix after adaptive graph convolution, which is expressed as: ， In the formula, is the feature matrix after adaptive graph convolution, and are the adaptive adjacency matrix and degree matrix respectively, is the activation function, is the weight matrix.

7. A partial discharge type recognition system based on multi-feature extraction and fusion, characterized in that: include: An acquisition module is configured to acquire partial discharge signal data and extract periodic features in the partial discharge signal data by combining frequency domain analysis and time domain analysis. ; The first extraction module is configured to extract graph structure features from the partial discharge signal data to obtain graph structure features corresponding to the partial discharge signal data. , and the periodic characteristics and graph structural features Fusion is performed and the fused features are Perform layer normalization to obtain target fusion features ; A second extraction module is configured to add the position code to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector , wherein the position encoding is added to the target fusion feature , and the added features to be input Perform time series feature extraction to obtain the final feature vector include: Adaptive position coding: Adaptive position coding is added to the feature sequence, generated using a neural network, with the input being the time step and signal features, the output being the adaptive position coding, and finally generating a feature matrix containing the position coding , a feature matrix containing positional encodings The expression is: ， ， ， In the formula, is the adaptive position encoding, is the original time series data, A neural network is used to generate positional encodings, is a function used to extract time step information from the raw data, is the time step information; Multi-scale feature extraction: Before extracting temporal features, the input features are processed by a multi-scale convolution layer. A multi-scale convolution layer is added to extract multi-scale features through convolution kernels of different sizes. The expression of multi-scale features is: ， In the formula, For splicing operation, is the multi-scale feature matrix, , , are convolution operations of different scales, is the encoded feature matrix; Dynamically aware attention mechanism, the dynamically aware attention mechanism includes generating queries, keys and values: for multi-scale features , generate the corresponding query matrix, key matrix and value matrix, the expression is: ， ， ， In the formula, is the query matrix, is the weight matrix of the query matrix, is the bias term of the bond matrix, is the key matrix, is the weight matrix of the key matrix, is the bias term of the bond matrix, is the value matrix, is the weight matrix of the value matrix, is the bias term of the value matrix; Initialization of weight matrix and bias term: Xaier initialization is used to initialize the weight matrix and bias term. The initialization expression is: ， In the formula, is any one of the weight matrix of the query matrix, the weight matrix of the key matrix, and the weight matrix of the value matrix, For uniform distribution, is the input dimension, is the output dimension; Feature space projection: Through linear transformation, the input features are projected into different feature spaces. The query matrix is ​​projected into one feature space to ask questions, the key matrix is ​​projected into another feature space to provide clues to the answer, and the value matrix contains the actual answer information. Dynamic perception: input features Perform a linear transformation to generate a preliminary bias term , is the weight matrix of the linear transformation, is the bias term of the linear transformation; Perform nonlinear activation on the result after linear transformation to generate dynamic bias terms ; The dynamic bias term Introduced into the calculation of attention weights for adaptive adjustment; Calculate the attention weights: Calculate the dot product of the query matrix Q and the key matrix K and scale it by dividing by the square root of the key dimension: in, Represents the dot product of the query matrix and the key matrix, divided by the square root of the key dimension To scale; The dynamic bias term Introduced into the dot product result, the formula is: ; The attention score is normalized through the softmax function to generate the final attention weight A; Calculate the attention output: The calculated final attention weight A is applied to the value matrix V to obtain the final attention output , the attention output calculation formula is: ， In the formula, For attention output, is an element of the adjacency matrix, indicating whether there is a connection between nodes. is the j-th value vector; An output module is configured to convert the final feature vector The input is into a pre-built partial discharge recognition model, and the partial discharge recognition model output is consistent with the final feature vector Corresponding partial discharge type.

8. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images