Power transmission line multi-lightning stroke identification system and method based on multi-source data

Abstract

The invention discloses a multi-source data-based multi-lightning stroke identification system and method for a power transmission line, and the method comprises the steps: carrying out the preprocessing of electromagnetic field data, distributed traveling wave data and fault recording data, generating an electromagnetic field time-frequency spectrogram, a distributed traveling wave time-frequency spectrogram and a fault recording time-frequency spectrogram, and carrying out the feature extraction of the electromagnetic field time-frequency spectrogram, the distributed traveling wave time-frequency spectrogram and the fault recording time-frequency spectrogram; an electromagnetic field one-dimensional feature vector, a distributed traveling wave one-dimensional feature vector and a fault recording one-dimensional feature vector are obtained and combined, and a multi-dimensional feature vector is generated; and finally, performing fusion processing on the multi-dimensional feature vector by adopting a weight adaptive feature fusion method to obtain a fused feature vector, and inputting the fused feature vector into a deep neural network to obtain a power transmission line multiple lightning stroke identification result. Through multi-source data fusion and adaptive weight optimization, the problems of insufficient data utilization and low feature fusion efficiency in the prior art are solved.

Description

Technical Field

[0001] This invention belongs to the field of power system fault detection technology, specifically relating to a transmission line multiple lightning strike identification system and method based on multi-source data. Background Technology

[0002] With the continuous expansion of my country's power grid distribution scale and the long distances of high-altitude power transmission lines, the economic losses caused by lightning strikes to the power system each year are incalculable. Therefore, reducing the damage of lightning strikes to transmission lines is of great significance for ensuring the safe production and stable operation of the power system. Although lightning protection measures have been implemented for some transmission lines, tripping accidents caused by lightning strikes have not been significantly reduced. The reason for this is the difficulty in accurately classifying and identifying lightning waveforms with different return stroke counts, resulting in a lack of targeted lightning protection measures. Therefore, the ability to quickly and accurately identify specific lightning waveforms is of great importance for improving lightning protection and mitigation measures in the power system.

[0003] In the field of lightning waveform recognition, scholars have been conducting research by extracting distinguishable features from different lightning strike types and then selecting appropriate classifiers for classification. Extracting features reflecting the transient current components of different lightning strike types is a key aspect of lightning waveform recognition. Currently, the main methods for extracting features from lightning signals include time-domain methods, frequency-domain methods, S-transform, singular value decomposition, wavelet transform, and Hilbert-Huang transform. To address the incompleteness of single-source information, researchers utilize multi-source information composed of signals from multiple sensors to obtain more complete equipment health status information, constructing multi-source information fusion models to obtain more stable and reliable diagnostic results. Existing methods have certain application value in lightning waveform recognition for transmission lines, but they generally face limitations such as relatively singular information sources and limited applicability, making it difficult to achieve ideal accuracy when facing lightning strikes with different return stroke counts. However, with the rapid development of neural network technology, deep learning has demonstrated superior performance in multiple fields and is gradually replacing some traditional lightning waveform recognition methods. For research on lightning strike waveform identification of transmission lines, deep learning technology can achieve higher accuracy in lightning strike identification and classification through its powerful feature extraction and model recognition capabilities. Summary of the Invention

[0004] This invention provides a system and method for identifying multiple lightning strikes on transmission lines based on multi-source data. By integrating multi-source data and adaptive weight optimization, it solves the problems of insufficient data utilization and low feature fusion efficiency in the prior art.

[0005] A multi-source data-based system for identifying multiple lightning strikes on transmission lines includes a multi-source data preprocessing module, a multi-modal feature extraction module, an adaptive feature fusion module, and a lightning strike classification and identification module. The multi-source data preprocessing module preprocesses electromagnetic field data, distributed traveling wave data, and fault recording data to generate time-frequency spectra of the electromagnetic field, distributed traveling wave, and fault recording data. The multi-modal feature extraction module extracts features from these data to obtain one-dimensional feature vectors for the electromagnetic field, distributed traveling wave, and fault recording data. The adaptive feature fusion module merges these feature vectors to generate a multi-dimensional feature vector. The lightning strike classification and identification module uses a weighted adaptive feature fusion method to fuse the multi-dimensional feature vectors, obtaining a fused feature vector. This fused feature vector is then input into a deep neural network to obtain the result of identifying multiple lightning strikes on the transmission line.

[0006] The steps of the lightning strike classification and recognition module in fusing multi-dimensional feature vectors include: The multidimensional feature vector is subjected to adaptive pooling. The multidimensional feature vector X for X =( x 1, x 2, x 3,..., x J ), x 1, x 2, x 3,..., x J It is a one-dimensional feature vector. J The dimension of a multidimensional feature vector; The adaptive pooling algorithm used in the adaptive pooling process is calculated using the following formula: In the formula, To fuse feature vectors, γ ij for γ Type pooling weight parameters, for Type pooling weight parameters, x ij Multidimensional feature vectors X The Middle i Line number j The feature values ​​are input as column feature values; The Softmax function is chosen as the classifier for converting probability values; the cross-entropy cost function is used.L As a constraint for subsequent iterations, the cross-entropy cost function is calculated using the following formula: In the formula, To predict the output value, This is the actual label value. L Let cost function be M The total number of eigenvectors; Using the cross-entropy cost function L As constraints, the formula for calculating the adaptive pooling algorithm is... Type pooling weight parameters and The partial derivatives of the pooling weight parameters are calculated to obtain the updated... Type pooling weight parameters and updated The pooling weights are updated, and the next iteration is performed using the updated pooling weights until the pooling weights satisfy the cost function. L At this point, the minimum value is obtained. The optimal pooling weight parameters are Type pooling weight parameters , The optimal pooling weight parameters are Type pooling weight parameters This step ensures that the pooling weight parameters can be dynamically updated based on the detected data, accurately preserving feature information; and that the optimal pooling weight parameters are... Type pooling weight parameters and Type pooling weight parameters Substituting the values ​​into the adaptive pooling algorithm formula, the fusion vector is calculated. ; In the formula To fuse feature vectors, To be optimal Type pooling weight parameters, To be optimal Type pooling weight parameters, The input feature vector is used for pooling.

[0007] The update Type pooling weight parameters and The specific method for setting pooling weight parameters is as follows: The updated number l layer Type pooling weight parameters and the updated first l layer Substituting the pooling weight parameters into the adaptive pooling algorithm calculation formula, the first pooling weight is calculated.l The state values ​​of the layer after pooling: In the formula, To fuse feature vectors In the l The state values ​​of the layer after pooling, For the first l layer Type pooling weight parameters, For the first l layer Type pooling weight parameters, For the first l The feature vector of the layer input; Calculate the first l Layer output values Update the formula using backpropagation: In the formula, For the first l The output value of the layer, For the first l The predicted output value of the layer, For the first l Layer nonlinear factors, ,in, The ReLU activation function is used. The derivative of the ReLU activation function. This represents space multiplication operations; Calculate the first l+ Output value of layer 1 The output values ​​of the pooling layers are calculated sequentially from layer N-1 to layer 2, where N is the number of iterations. The calculation formula is: In the formula, For the first l The output value of the layer, For the first l+ The transpose of the weight matrix of layer 1. For the first l+ The output value of layer 1 The derivative of the ReLU activation function. This represents space multiplication operations; Calculate the cost function for the current training data using the formula below. L The partial derivatives; In the formula, Cost functionL right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l The output value of the layer, For the first l- The transpose of a single-layer nonlinear factor matrix; renew Type pooling weight parameters and The formula for calculating the pooling weight parameters is: In the formula, Cost function L right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l+ 1st floor Type pooling weight parameters, For the first l+ 1st floor Pooling weight parameters.

[0008] A method for classifying and identifying lightning strikes on transmission lines based on multi-source data fusion is characterized by the following steps: preprocessing electromagnetic field data, distributed traveling wave data, and fault recording data to generate electromagnetic field time-frequency spectra, distributed traveling wave time-frequency spectra, and fault recording time-frequency spectra; extracting features from the electromagnetic field time-frequency spectra, distributed traveling wave time-frequency spectra, and fault recording time-frequency spectra to obtain one-dimensional feature vectors for the electromagnetic field, distributed traveling wave, and fault recording; merging the one-dimensional feature vectors for the electromagnetic field, distributed traveling wave, and fault recording to generate a multi-dimensional feature vector; fusing the multi-dimensional feature vectors using a weighted adaptive feature fusion method to obtain a fused feature vector, and inputting the fused feature vector into a deep neural network to obtain the identification result of multiple lightning strikes on the transmission line.

[0009] The present invention has the following beneficial effects: (1) Multi-source heterogeneous data fusion mechanism: Through the collaborative processing of electromagnetic, traveling wave and fault recording data, the electromagnetic radiation, traveling wave propagation and transient response characteristics of lightning strikes are covered.

[0010] (2) Weight Adaptive Module: Introduces a learnable feature saliency evaluation function to dynamically optimize the contribution weights of multi-source data and suppress noise interference.

[0011] (3) Time-frequency-space joint modeling: The time-varying characteristics of the lightning strike signal are preserved by using the STFT time-spectrum graph and combined with the local feature extraction capability of CNN to improve the sensitivity of the model. Attached Figure Description

[0012] Figure 1 This is a system structure diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; 1-Multi-source data preprocessing module, 2-Multi-modal feature extraction module, 3-Adaptive feature fusion module, 4-Lightning strike classification and recognition module. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1: A transmission line lightning strike classification and identification system based on multi-source data fusion, such as Figure 1 As shown, it includes a multi-source data preprocessing module 1, a multimodal feature extraction module 2, an adaptive feature fusion module 3, and a lightning strike classification and recognition module 4; The multi-source data preprocessing module 1 preprocesses electromagnetic field data, distributed traveling wave data and fault recording data to generate electromagnetic field time spectrum, distributed traveling wave time spectrum and fault recording time spectrum. The multimodal feature extraction module 2 extracts features from the electromagnetic field time spectrum, the distributed traveling wave time spectrum, and the fault recording time spectrum to obtain a one-dimensional feature vector of the electromagnetic field, a one-dimensional feature vector of the distributed traveling wave, and a one-dimensional feature vector of the fault recording. The adaptive feature fusion module 3 merges the one-dimensional feature vector of the electromagnetic field, the one-dimensional feature vector of the distributed traveling wave, and the one-dimensional feature vector of the fault recording wave to generate a multi-dimensional feature vector. The lightning strike classification and identification module 4 uses a weighted adaptive feature fusion method to fuse multi-dimensional feature vectors to obtain a fused feature vector, and then inputs the fused feature vector into a deep neural network to obtain the multiple lightning strike identification results of the transmission line.

[0014] In the above technical solution, the specific implementation method of the preprocessing method of the multi-source data preprocessing module 1 is as follows: The electromagnetic field data is subjected to a short-time Fourier transform (STFT) to generate a time-frequency spectrum of the electromagnetic field, wherein the time-frequency spectrum of the electromagnetic field has a dimension of 32×32. Perform a short-time Fourier transform (STFT) on the distributed traveling wave data to generate a distributed traveling wave time spectrum, wherein the distributed traveling wave time spectrum has a dimension of 32×32; The fault recording data is extracted to generate the effective waveform segment of the fault recording after the lightning strike. After being converted by short-time Fourier transform (STFT), a fault recording time spectrum is generated. The dimension of the fault recording time spectrum is 32×32.

[0015] In the above technical solution, the specific implementation method of the multimodal feature extraction module 2 is as follows: Convolutional neural networks are constructed to extract features from the electromagnetic field time spectrum, the distributed traveling wave time spectrum, and the fault recording time spectrum, thereby obtaining one-dimensional feature vectors of the electromagnetic field, the distributed traveling wave, and the fault recording. The convolutional neural network includes convolutional layers and fully connected layers. There are 4 convolutional layers with 3×3 kernels and ReLU activation. The fully connected layers include 2 layers. The dimensions of the one-dimensional feature vector of the electromagnetic field, the one-dimensional feature vector of the distributed traveling wave, and the one-dimensional feature vector of the fault recording wave are all 256.

[0016] In the above technical solution, the adaptive feature fusion module 3 is specifically implemented as follows: The one-dimensional feature vector of the electromagnetic field, the one-dimensional feature vector of the distributed traveling wave, and the one-dimensional feature vector of the fault recording wave are connected end to end to generate a multi-dimensional feature vector.

[0017] In the above technical solution, the steps of the lightning strike classification and recognition module 4 in fusing the multi-dimensional feature vector are as follows: Step 41: Perform adaptive pooling processing on the multidimensional feature vector; The multidimensional feature vector X for X =( x 1, x 2, x 3,..., x J ), x 1, x 2, x 3,..., x J It is a one-dimensional feature vector. J The dimension of a multidimensional feature vector; The adaptive pooling algorithm used in the adaptive pooling process is calculated using the following formula: In the formula, To fuse feature vectors, γ ijfor γ Type pooling weight parameters, for Type pooling weight parameters, x ij Multidimensional feature vectors X The Middle i Line number j The feature values ​​are input as column feature values; Step 42: Select the Softmax function as the classifier to convert probability values; use the cross-entropy cost function. L As a constraint for subsequent iterations, the cross-entropy cost function is calculated using the following formula: In the formula, To predict the output value, This is the actual label value. L Let cost function be M The total number of feature vectors; the predicted output value is the probability value after processing by the Softmax function, representing the model's prediction confidence for each category; Step 43, using the cross-entropy cost function L As constraints, the formula for calculating the adaptive pooling algorithm is... Type pooling weight parameters and The partial derivatives of the pooling weight parameters are calculated to obtain the updated... Type pooling weight parameters and updated The pooling weights are updated, and the next iteration is performed using the updated pooling weights until the pooling weights satisfy the cost function. L At this point, the minimum value is obtained. The optimal pooling weight parameters are Type pooling weight parameters , The optimal pooling weight parameters are Type pooling weight parameters This step ensures that the pooling weight parameters can be dynamically updated based on the detected data, accurately preserving feature information; and that the optimal pooling weight parameters are... Type pooling weight parameters and Type pooling weight parameters Substituting the values ​​into the adaptive pooling algorithm formula, the fusion vector is calculated. ; In the formula To fuse feature vectors, To be optimal Type pooling weight parameters, To be optimal Type pooling weight parameters, The input feature vector is used for pooling.

[0018] The The initial value of the pooling weight parameter is The The initial value of the pooling weight parameter is ; In the above technical solution, the update Type pooling weight parameters and The specific method for setting pooling weight parameters is as follows: Step 42.1, update the first l layer Type pooling weight parameters and the updated first l layer Substituting the pooling weight parameters into the adaptive pooling algorithm calculation formula, the first pooling weight is calculated. l The state values ​​of the layer after pooling: In the formula, To fuse feature vectors In the l The state values ​​of the layer after pooling, For the first l layer Type pooling weight parameters, For the first l layer Type pooling weight parameters, For the first l The feature vector of the layer input; Step 42.2, calculate the first... l Layer output values Update the formula using backpropagation: In the formula, For the first l The output value of the layer, For the first l The predicted output value of the layer, For the first l Layer nonlinear factors, ,in, The ReLU activation function is used. The derivative of the ReLU activation function. This represents space multiplication operations; Step 42.3, calculate the first... l+ Output value of layer 1 The output values ​​of the pooling layers are calculated sequentially from layer N-1 to layer 2, where N is the number of iterations. The calculation formula is: In the formula, For the first l The output value of the layer, For the first l+ The transpose of the weight matrix of layer 1. For the first l+ The output value of layer 1 The derivative of the ReLU activation function. This represents space multiplication operations; Step 42.4: Calculate the cost function for the current training data using the formula below. L The partial derivatives; In the formula, Cost function L right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l The output value of the layer, For the first l- The transpose of a single-layer nonlinear factor matrix; Step 42.5, Update Type pooling weight parameters and The formula for calculating the pooling weight parameters is: In the formula, Cost function L right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l+ 1st floor Type pooling weight parameters, For the first l+ 1st floor Type pooling weight parameters; Example 2: A method for classifying and identifying lightning strikes on transmission lines based on multi-source data fusion, such as... Figure 2As shown, it includes the following steps: Preprocessing is performed on electromagnetic field data, distributed traveling wave data, and fault recording data to generate electromagnetic field time spectrum diagrams, distributed traveling wave time spectrum diagrams, and fault recording time spectrum diagrams. Feature extraction was performed on the electromagnetic field time spectrum, distributed traveling wave time spectrum and fault recording time spectrum to obtain one-dimensional feature vectors of electromagnetic field, distributed traveling wave and fault recording. The one-dimensional feature vectors of electromagnetic field, distributed traveling wave, and fault recording wave are merged to generate a multi-dimensional feature vector. A weighted adaptive feature fusion method is used to fuse multidimensional feature vectors to obtain a fused feature vector. The fused feature vector is then input into a deep neural network to obtain the identification results of multiple lightning strikes on transmission lines.

[0019] In the above technical solution, the specific implementation method for preprocessing electromagnetic field data, distributed traveling wave data, and fault recording data is as follows: Perform a short-time Fourier transform (STFT) on the electromagnetic field data to generate a time-frequency spectrum of the electromagnetic field. Perform a short-time Fourier transform (STFT) on the distributed traveling wave data to generate a distributed traveling wave time spectrum. The fault recording data is extracted to generate the effective waveform segment of the fault recording after the lightning strike. After conversion by short-time Fourier transform (STFT), the fault recording time spectrum is generated.

[0020] In the above technical solution, the electromagnetic field time spectrum diagram has a dimension of 32×32, the distributed traveling wave time spectrum diagram has a dimension of 32×32, and the fault recording time spectrum diagram has a dimension of 32×32.

[0021] The specific implementation method for feature extraction from the electromagnetic field time spectrum, distributed traveling wave time spectrum, and fault recording time spectrum in the above technical solution is as follows: A convolutional neural network is constructed to extract features from the electromagnetic field time spectrum, the distributed traveling wave time spectrum, and the fault recording time spectrum, thereby obtaining one-dimensional feature vectors for the electromagnetic field, the distributed traveling wave, and the fault recording.

[0022] In the above technical solution, the convolutional neural network includes convolutional layers and fully connected layers. The convolutional layers have a total of 4 layers with 3×3 kernels and ReLU activation. The fully connected layers include 2 layers. The dimensions of the one-dimensional feature vector of the electromagnetic field, the one-dimensional feature vector of the distributed traveling wave, and the one-dimensional feature vector of the fault recording wave are all 256.

[0023] In the above technical solution, the specific implementation method for generating multidimensional feature vectors is as follows: The one-dimensional feature vector of the electromagnetic field, the one-dimensional feature vector of the distributed traveling wave, and the one-dimensional feature vector of the fault recording wave are connected end to end to generate a multi-dimensional feature vector.

[0024] The steps involved in fusing multi-dimensional feature vectors in the above technical solution include: The multidimensional feature vector is subjected to adaptive pooling. The multidimensional feature vector X for X =( x 1, x 2, x 3,..., x J ), x 1, x 2, x 3,..., x J It is a one-dimensional feature vector. J The dimension of a multidimensional feature vector; The adaptive pooling algorithm used in the adaptive pooling process is calculated using the following formula: In the formula, To fuse feature vectors, γ ij for γ Type pooling weight parameters, for Type pooling weight parameters, x ij Multidimensional feature vectors X The Middle i Line number j The feature values ​​are input as column feature values; The Softmax function is chosen as the classifier for converting probability values; the cross-entropy cost function is used. L As a constraint for subsequent iterations, the cross-entropy cost function is calculated using the following formula: In the formula, To predict the output value, This is the actual label value. L Let cost function be M The total number of eigenvectors; Using the cross-entropy cost function L As constraints, the formula for calculating the adaptive pooling algorithm is... Type pooling weight parameters and The partial derivatives of the pooling weight parameters are calculated to obtain the updated... Type pooling weight parameters and updated The pooling weights are updated, and the next iteration is performed using the updated pooling weights until the pooling weights satisfy the cost function. L At this point, the minimum value is obtained. The optimal pooling weight parameters are Type pooling weight parameters , The optimal pooling weight parameters are Type pooling weight parameters This step ensures that the pooling weight parameters can be dynamically updated based on the detected data, accurately preserving feature information; and that the optimal pooling weight parameters are... Type pooling weight parameters and Type pooling weight parameters Substituting the values ​​into the adaptive pooling algorithm formula, the fusion vector is calculated. ; In the formula To fuse feature vectors, To be optimal Type pooling weight parameters, To be optimal Type pooling weight parameters, The input feature vector is used for pooling.

[0025] In the above technical solution, the update Type pooling weight parameters and The specific method for setting pooling weight parameters is as follows: The updated number l layer Type pooling weight parameters and the updated first l layer Substituting the pooling weight parameters into the adaptive pooling algorithm calculation formula, the first pooling weight is calculated. l The state values ​​of the layer after pooling: In the formula, To fuse feature vectors In the l The state values ​​of the layer after pooling, For the first l layer Type pooling weight parameters, For the first l layer Type pooling weight parameters, For the first l The feature vector of the layer input; Calculate the first l Layer output values Update the formula using backpropagation: In the formula, For the first l The output value of the layer, For the first l The predicted output value of the layer, For the first l Layer nonlinear factors, ,in, The ReLU activation function is used. The derivative of the ReLU activation function. This represents space multiplication operations; Calculate the first l+ Output value of layer 1 The output values ​​of the pooling layers are calculated sequentially from layer N-1 to layer 2, where N is the number of iterations. The calculation formula is: In the formula, For the first l The output value of the layer, For the first l+ The transpose of the weight matrix of layer 1. For the first l+ The output value of layer 1 The derivative of the ReLU activation function. This represents space multiplication operations; Calculate the cost function for the current training data using the formula below. L The partial derivatives; In the formula, Cost function L right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l The output value of the layer, For the first l- The transpose of a single-layer nonlinear factor matrix; renew Type pooling weight parameters and The formula for calculating the pooling weight parameters is: In the formula, Cost function L right The partial derivatives of the pooling weight parameters, Cost function L right The partial derivatives of the pooling weight parameters, For the first l+ 1st floor Type pooling weight parameters, For the first l+ 1st floor Pooling weight parameters.

[0026] Example 3: An electronic device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the steps of the above-described method for classifying and identifying lightning strikes on transmission lines based on multi-source data fusion.

[0027] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for identifying multiple lightning strikes on transmission lines based on multi-source data.

[0028] The contents not described in detail in this specification are prior art known to those skilled in the art. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0029] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0030] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0031] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A lightning stroke classification and identification system for power transmission lines based on multi-source data fusion, characterized by: It comprises a multi-source data preprocessing module (1), a multi-modal feature extraction module (2), an adaptive feature fusion module (3), and a lightning stroke classification and identification module (4). The multi-source data preprocessing module (1) is used for preprocessing electromagnetic field data, distributed traveling wave data, and fault recording data to generate electromagnetic field time-frequency spectrograms, distributed traveling wave time-frequency spectrograms, and fault recording time-frequency spectrograms. The multi-modal feature extraction module (2) is used for extracting features from the electromagnetic field time-frequency spectrograms, distributed traveling wave time-frequency spectrograms, and fault recording time-frequency spectrograms to obtain one-dimensional feature vectors of the electromagnetic field, distributed traveling wave, and fault recording. The adaptive feature fusion module (3) is used for merging the one-dimensional feature vectors of the electromagnetic field, distributed traveling wave, and fault recording to generate a multi-dimensional feature vector. The lightning stroke classification and identification module (4) is used for fusing the multi-dimensional feature vector using a weight adaptive feature fusion method to obtain a fused feature vector, and inputting the fused feature vector into a deep neural network to obtain a multiple lightning stroke identification result of the transmission line.

2. The multi-source data fusion based transmission line lightning classification and identification system of claim 1, wherein: The specific implementation method of the preprocessing method of the multi-source data preprocessing module (1) is as follows: Perform short-time Fourier transform on the electromagnetic field data to generate electromagnetic field time-frequency spectrograms; Perform short-time Fourier transform on the distributed traveling wave data to generate distributed traveling wave time-frequency spectrograms; Intercept the fault recording data to generate a fault recording effective waveform segment after lightning stroke, and convert it into a fault recording time-frequency spectrogram through short-time Fourier transform.

3. The multi-source data fusion based transmission line lightning classification and identification system of claim 2, wherein: The dimension of the electromagnetic field time-frequency spectrogram is 32x32, the dimension of the distributed traveling wave time-frequency spectrogram is 32x32, and the dimension of the fault recording time-frequency spectrogram is 32x32.

4. The multi-source data fusion based transmission line lightning classification and identification system of claim 1, wherein: The specific implementation method of the multi-modal feature extraction module 2 is as follows: Construct a convolutional neural network to extract features from the electromagnetic field time-frequency spectrogram, distributed traveling wave time-frequency spectrogram, and fault recording time-frequency spectrogram to obtain one-dimensional feature vectors of the electromagnetic field, distributed traveling wave, and fault recording.

5. The multi-source data fusion-based transmission line lightning stroke classification and identification system according to claim 4, characterized in that: The convolutional neural network comprises four convolutional layers and two fully connected layers, the convolutional kernel is 3x3, and ReLU activation is selected; The dimensions of the one-dimensional feature vectors of the electromagnetic field, distributed traveling wave, and fault recording are all 256.

6. The multi-source data fusion based transmission line lightning classification and identification system of claim 1, wherein: The specific implementation method of the adaptive feature fusion module (3) is as follows: Connect the one-dimensional feature vectors of the electromagnetic field, distributed traveling wave, and fault recording end to end to generate a multi-dimensional feature vector.

7. The multi-source data fusion based transmission line lightning classification and identification system of claim 1, wherein: The steps of the lightning stroke classification and identification module (4) for fusing the multi-dimensional feature vector include: Perform adaptive pooling processing on the multi-dimensional feature vector; The multi-dimensional feature vector X is X ( x 1, x 2, x 3,..., x J ), x 1, x 2, x 3,..., x J is a one-dimensional feature vector, J denotes the dimension of the multi-dimensional feature vector; The adaptive pooling algorithm calculation formula used in the adaptive pooling processing is as follows: wherein, is a fused feature vector, It comprises the following steps: ij is ​ a pooling weight parameter of type is a pooling weight parameter of type x ij a multi-dimensional feature vector X the i the j the The Softmax function is chosen as the classifier to transform the probability values; the cross-entropy cost function is used L The formula for the cross-entropy cost function for the constraint condition for the subsequent iteration is: wherein is the predicted output value, is the actual label value, L is the cost function, M is the total number of feature vectors; Using the cross-entropy cost function L As constraints, the formula for calculating the adaptive pooling algorithm is... Type pooling weight parameters and The partial derivatives of the pooling weight parameters are calculated to obtain the updated... Type pooling weight parameters and updated The pooling weights are updated, and the next iteration is performed using the updated pooling weights until the pooling weights satisfy the cost function. L At this point, the minimum value is obtained. The optimal pooling weight parameters are Type pooling weight parameters , The optimal pooling weight parameters are Type pooling weight parameters This step ensures that the pooling weight parameters can be dynamically updated based on the detected data, accurately preserving feature information; and that the optimal pooling weight parameters are... Pooling weight parameters and Pooling weight parameters Substituting the values ​​into the adaptive pooling algorithm formula, calculate the fusion vector. ; wherein is a fused feature vector, is an optimal is a type of pooling weight parameter, is an optimal is a type of pooling weight parameter, is a feature vector input to the pooling layer.

8. The multi-source data fusion based transmission line lightning classification and identification system of claim 7, wherein: The update Type pooling weight parameters and The specific method for setting pooling weight parameters is as follows: The updated first l layer type pooling weight parameters and the updated first l layer type pooling weight parameters are substituted into the adaptive pooling algorithm calculation formula to calculate the state value of the first l layer after pooling. wherein is a fused feature vector In a first l layer passed through a pooling state value, is a first l layer type pooling weight parameter, is a first l layer type pooling weight parameter, is a first l layer input feature vector; The output value of the layer is calculated l layer The output value of the layer is calculated The output value of the layer is calculated wherein, is the output value of the l layer, is the predicted output value of the l layer, is the non-linear factor of the l layer, wherein, is the RELU activation function, is the derivative of the RELU activation function, denotes a spatial multiplication operation; The output value of the first layer is calculated l+ 1 layer The output value of the pooling layer is calculated in order from the N-1 layer to the 2 layer, where N is the number of iterations, The calculation formula is: wherein, is the output value of the l layer, is the output value of the l+ 1 layer, is the transpose matrix of the weight matrix of the l+ 1 layer, is the derivative of the RELU activation function, denotes a spatial multiplication operation; The cost function for the current training data is calculated according to the following formula L partial derivative of the cost function with respect to the weight wherein is a cost function L is a cost function is a partial derivative of the pooling weight parameter of the type is a cost function L is a partial derivative of the pooling weight parameter of the type is a partial derivative of the pooling weight parameter of the type is an output value of the l is an output value of the is an output value of the l- is a transpose matrix of the first update pooling weight parameters and The formula for calculating the pooling weight parameters of the type is wherein is a cost function L with respect to is a partial derivative of a pooling weight parameter of the type is a cost function L with respect to is a partial derivative of a pooling weight parameter of the type is a first l+ layer pooling weight parameter of the type is a first l+ layer pooling weight parameter of the type 9. A power transmission line lightning stroke classification and identification method based on multi-source data fusion, characterized in that: ​ The electromagnetic field data, distributed traveling wave data and fault recording data are preprocessed to generate electromagnetic field time-frequency spectrum, distributed traveling wave time-frequency spectrum and fault recording time-frequency spectrum; The electromagnetic field time-frequency spectrum, distributed traveling wave time-frequency spectrum data and fault recording time-frequency spectrum are subjected to feature extraction to obtain electromagnetic field one-dimensional feature vector, distributed traveling wave one-dimensional feature vector and fault recording one-dimensional feature vector; The electromagnetic field one-dimensional feature vector, distributed traveling wave one-dimensional feature vector and fault recording one-dimensional feature vector are combined to generate a multi-dimensional feature vector; The multi-dimensional feature vector is subjected to fusion processing by using a weight adaptive feature fusion method to obtain a fusion feature vector, and the fusion feature vector is input into a deep neural network to obtain a multiple lightning stroke recognition result of the power transmission line.

10. The method for classification and identification of lightning stroke on transmission lines based on multi-source data fusion according to claim 9, characterized in that: The specific implementation method of preprocessing the electromagnetic field data, distributed traveling wave data and fault recording data is: The electromagnetic field data is subjected to short-time Fourier transform to generate electromagnetic field time-frequency spectrum; The distributed traveling wave data is subjected to short-time Fourier transform to generate distributed traveling wave time-frequency spectrum; The fault recording data is intercepted to generate a fault recording effective waveform segment after lightning stroke, which is converted by short-time Fourier transform to generate fault recording time-frequency spectrum.

11. The method for classification and identification of lightning stroke on transmission line based on multi-source data fusion according to claim 10, characterized in that: The dimension of the electromagnetic field time-frequency spectrum is 32x32, the dimension of the distributed traveling wave time-frequency spectrum is 32x32, and the dimension of the fault recording time-frequency spectrum is 32x32.

12. The method for classification and identification of lightning stroke on transmission lines based on multi-source data fusion according to claim 9, characterized in that: The specific implementation method of feature extraction of the electromagnetic field time-frequency spectrum, distributed traveling wave time-frequency spectrum data and fault recording time-frequency spectrum is: A convolutional neural network is constructed to extract features from the electromagnetic field time-frequency spectrum, distributed traveling wave time-frequency spectrum and fault recording time-frequency spectrum to obtain electromagnetic field one-dimensional feature vector, distributed traveling wave one-dimensional feature vector and fault recording one-dimensional feature vector.

13. The lightning stroke classification and recognition method of a power transmission line based on multi-source data fusion according to claim 12, characterized in that: The convolutional neural network includes a convolutional layer and a fully connected layer, the convolutional layer has a total of 4 layers, the convolution kernel is 3x3, and ReLU activation is selected, and the fully connected layer includes 2 layers; The dimensions of the electromagnetic field one-dimensional feature vector, distributed traveling wave one-dimensional feature vector and fault recording one-dimensional feature vector are all 256.

14. The method of claim 9, wherein the method is based on multi-source data fusion for transmission line lightning classification and identification. The specific implementation of generating a multi-dimensional feature vector is: The electromagnetic field one-dimensional feature vector, distributed traveling wave one-dimensional feature vector and fault recording one-dimensional feature vector are connected head to tail to generate a multi-dimensional feature vector.

15. The method for classification and identification of lightning stroke on transmission line based on multi-source data fusion according to claim 9, characterized in that: The steps of fusion processing of the multi-dimensional feature vector include: The multi-dimensional feature vector is subjected to adaptive pooling processing; The multi-dimensional feature vector X is X ( x 1, x 2, x 3,..., x J ), x 1, x 2, x 3,..., x J is a one-dimensional feature vector, J denotes the dimension of the multi-dimensional feature vector; The adaptive pooling algorithm calculation formula used in the adaptive pooling processing is: wherein, is a fused feature vector, The computer program is executed by a processor to implement a lightning stroke classification and recognition method of a power transmission line based on multi-source data fusion according to any one of claims 9-16. ij is ​ a pooling weight parameter of type is a pooling weight parameter of type x ij is a multi-dimensional feature vector X the i row j column feature value input feature value; The Softmax function is chosen as the classifier to transform the probability values; the cross-entropy cost function is used L The formula for the cross-entropy cost function is given by: wherein is the predicted output value, is the actual label value, L is the cost function, M is the total number of feature vectors; Using the cross-entropy cost function L As constraints, the formula for calculating the adaptive pooling algorithm is... Type pooling weight parameters and The partial derivatives of the pooling weight parameters are calculated to obtain the updated... Type pooling weight parameters and updated The pooling weights are updated, and the next iteration is performed using the updated pooling weights until the pooling weights satisfy the cost function. L At this point, the minimum value is obtained. The optimal pooling weight parameters are Type pooling weight parameters , The optimal pooling weight parameters are Type pooling weight parameters This step ensures that the pooling weight parameters can be dynamically updated based on the detected data, accurately preserving feature information; and that the optimal pooling weight parameters are... Pooling weight parameters and Pooling weight parameters Substituting the values ​​into the adaptive pooling algorithm formula, calculate the fusion vector. ; wherein is the fused feature vector, is the optimal is the type of pooling weight parameter, is the optimal is the type of pooling weight parameter, is the input feature vector before pooling.

16. The method of claim 15, wherein the method is based on multi-source data fusion for transmission line lightning classification and identification. The updating a specific method of the type of pooling weight parameters is: a specific method of the type of pooling weight parameters is: The updated first l layer type pooling weight parameters and the updated first l layer type pooling weight parameters are substituted into the adaptive pooling algorithm calculation formula to calculate the state value of the first l layer after pooling. wherein is a fused feature vector In a first l layer passed through a pooling state value, is a first l layer type pooling weight parameter, is a first l layer type pooling weight parameter, is a first l layer input feature vector; The output value of the layer is calculated l layer The output value of the layer is calculated The output value of the layer is calculated wherein, is the output value of the l layer, is the predicted output value of the l layer, is the output value of the l layer nonlinearity, wherein, is the RELU activation function, is the derivative of the RELU activation function, denotes a spatial multiplication operation; The output value of the first layer is calculated l+ 1 layer The output value of the pooling layer is calculated in order from the N-1 layer to the 2 layer, where N is the number of iterations, The calculation formula is: wherein, is the output value of the l layer, is the output value of the l+ 1 layer, is the transpose matrix of the weight matrix of the l+ 1 layer, is the derivative of the RELU activation function, denotes a spatial multiplication operation; The cost function for the current training data is calculated according to the following formula L partial derivative of the cost function with respect to the weight wherein is a cost function L is a cost function is a partial derivative of the pooling weight parameter of the type is a cost function L is a partial derivative of the pooling weight parameter of the type is a partial derivative of the pooling weight parameter of the type is an output value of the i-th layer l is an output value of the i-th layer is an output value of the i-th layer l- is a transpose matrix of the i-th layer nonlinear factor matrix update pooling weight parameters and The formula for calculating the pooling weight parameters of the type is wherein is a cost function L with respect to is a partial derivative of the pooling weight parameter of the type is a cost function L with respect to is a partial derivative of the pooling weight parameter of the type is a cost function l+ 1 layer is a pooling weight parameter of the type is a cost function l+ 1 layer is a pooling weight parameter of the type 17. A computer storage medium, the computer readable storage medium having stored thereon a computer program, characterized in that, ​

Images