A high-voltage transmission line lightning stroke fault intelligent identification method and system

Abstract

The application relates to a lightning stroke fault intelligent identification method and system for a high-voltage transmission line. The method comprises the following steps: acquiring lightning current waveform data, line traveling wave data, insulator leakage current data and meteorological environment data after a lightning stroke fault; performing time-frequency domain joint feature extraction on the lightning current waveform data and the line traveling wave data to obtain lightning transient feature data; performing waveform decomposition and modal analysis on the insulator leakage current data to obtain insulator state feature data; performing multi-factor correlation analysis on the meteorological environment data to obtain environment correction factor data; inputting the lightning transient feature data, the insulator state feature data and the environment correction factor data into a hierarchical fusion identification network for feature fusion to obtain a fault fusion feature vector; inputting the fault fusion feature vector into a fault classification identification model for lightning stroke fault type identification to obtain an identification result. The method can improve the identification robustness of lightning stroke faults under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line fault identification, and in particular relates to an intelligent identification method and system for lightning strike faults in high-voltage power transmission lines. Background Technology

[0002] With the rapid development of smart grids and high-voltage transmission technologies, the identification of lightning strike faults in transmission lines is crucial for the safe and stable operation of power systems. Lightning strike faults can not only cause line tripping but also lead to serious accidents such as insulator flashover and line breakage. Therefore, the rapid and accurate identification of lightning strike fault types has significant engineering implications.

[0003] Traditional lightning fault identification primarily relies on a single data source, such as analyzing lightning current waveform characteristics to determine if a lightning strike has occurred, or using traveling wave ranging to locate the fault point. These methods typically focus only on the time-domain characteristics of the lightning current or traveling wave, employing simple threshold comparisons or waveform matching for discrimination. Furthermore, some methods incorporate the amplitude variation of insulator leakage current to aid in judgment, but these are often analyzed independently without considering the inherent correlations between different physical quantities.

[0004] However, existing technologies have the following problems: First, the transient process of lightning strike faults is complex, and a single data source cannot fully characterize the subtle differences between lightning flashover, non-flashover, and non-lightning strike faults, resulting in low identification accuracy; Second, changes in the state of insulator surfaces such as dirt and moisture, as well as meteorological environmental factors (such as humidity and atmospheric electric field), have a significant impact on the occurrence of lightning strike faults, but traditional methods lack an effective fusion mechanism for these multi-source heterogeneous data; Third, existing feature extraction methods are simple and cannot fully explore the deep time-frequency features in waveform data, making it difficult to adapt to the needs of lightning strike fault identification under complex operating conditions. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and system for intelligent identification of lightning strike faults in high-voltage transmission lines to address the aforementioned technical problems.

[0006] Firstly, this application provides an intelligent identification method for lightning strike faults in high-voltage transmission lines, including:

[0007] S1. Acquire lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line.

[0008] S2. Perform joint time-frequency domain feature extraction on lightning current waveform data and line traveling wave data to obtain lightning transient feature data;

[0009] S3. Perform waveform decomposition and modal analysis on the insulator leakage current data to obtain insulator state characteristic data;

[0010] S4. Perform multi-factor correlation analysis on meteorological and environmental data to obtain environmental correction factor data;

[0011] S5. Input the lightning strike transient characteristic data, insulator state characteristic data and environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain the fault fusion feature vector;

[0012] S6. Input the fault fusion feature vector into the fault classification and recognition model to identify the type of lightning strike fault and obtain the lightning strike fault recognition result.

[0013] In one embodiment, S2 includes:

[0014] S21. Perform empirical mode decomposition on the lightning current waveform data to obtain the intrinsic mode function component data of the lightning current;

[0015] S22. Perform continuous wavelet transform on the traveling wave data of the line to obtain the traveling wave time spectrum data;

[0016] S23. Input the lightning current intrinsic mode function component data and the traveling wave time spectrum data into a dual-channel feature extraction network to extract the time-domain transient features and frequency-domain energy features, thereby obtaining the lightning current time-domain feature vector and the traveling wave frequency-domain feature vector; wherein, the dual-channel feature extraction network includes a first channel and a second channel; the first channel includes a one-dimensional convolutional layer and a max-pooling layer, and the first channel is used to extract the waveform steepness and peak attenuation rate of the lightning current intrinsic mode function components; the second channel includes a two-dimensional convolutional layer and a global average pooling layer, and the second channel is used to extract the frequency band centroid and energy concentration of the traveling wave time spectrum;

[0017] S24. Cross-attention fusion is performed on the lightning current time domain feature vector and the traveling wave frequency domain feature vector to obtain lightning transient feature data.

[0018] In one embodiment, S3 includes:

[0019] S31. Perform sliding window truncation and adaptive filtering on the insulator leakage current data to obtain pure leakage current fragment data;

[0020] S32. Based on the constrained variational objective function, variational mode decomposition is performed on the pure leakage current segment data to obtain multiple intrinsic mode function component data and residual component data; wherein, the expression of the constrained variational objective function is:

[0021]

[0022] in, For time The first time Each intrinsic mode function component For the first The center frequency of each component To preset the number of decomposition layers, For the Dirac function, This represents the convolution operation. Indicates time The partial derivatives, The imaginary unit, For time variables, It is an L2 norm;

[0023] S33. Calculate the peak amplitude, root mean square value and frequency band energy ratio of each intrinsic mode function component data and residual component data to obtain the multidimensional characteristic vector of leakage current.

[0024] S34. Input the multidimensional feature vector of leakage current into the insulator state coding network and perform feature compression to obtain insulator state feature data. The insulator state coding network includes an insulation coding input layer, a fully connected coding layer and an output layer. The insulation coding input layer is used to receive the multidimensional feature vector of leakage current, the fully connected coding layer is used to compress the feature dimension of the multidimensional feature vector of leakage current, and the output layer is used to output the insulator state feature data.

[0025] In one embodiment, S4 includes:

[0026] S41. Impute missing values ​​and standardize the meteorological and environmental data to obtain a standardized meteorological data matrix; the meteorological and environmental data includes meteorological parameters.

[0027] S42. Combining lightning strike fault labels and a standardized meteorological data matrix, calculate the mutual information coefficients between each meteorological parameter and the lightning strike fault label to obtain the initial correlation weight vector; where the expression for the mutual information coefficient is:

[0028]

[0029] in, For the first The mutual information coefficient between meteorological parameters and lightning fault labels, For the first A meteorological parameter, This is a historical lightning strike fault label. A value of 0 indicates a non-lightning strike fault, and a value of 1 indicates a lightning strike fault. Indicates the values ​​of meteorological parameters And the lightning strike fault label is The probability, Indicates the values ​​of meteorological parameters The probability, The value of the lightning strike fault label is indicated. The probability of;

[0030] S43. Input the standardized meteorological data matrix and the initial correlation weight vector into the preset multi-factor attention coding network for nonlinear feature weighting to obtain the meteorological impact feature vector; wherein, the multi-factor attention coding network includes an attention coding input layer, a multi-head self-attention layer, a layer normalization layer and a feedforward network layer. The attention coding input layer is used to receive the standardized meteorological data matrix and the initial correlation weight vector. The multi-head self-attention layer is used to calculate the attention score between each meteorological parameter based on the initial correlation weight vector. The layer normalization layer is used to normalize the attention score. The feedforward network layer is used to output the meteorological impact feature vector.

[0031] S44. Input the meteorological impact feature vector into the spatial mapping network for linear transformation to obtain environmental correction factor data.

[0032] In one embodiment, the hierarchical fusion recognition network includes a normalization layer, a weight calculation layer, a splicing and fusion layer, and a feature mapping layer, wherein S5 includes:

[0033] S51. The transient characteristic data of lightning strike, the state characteristic data of insulator and the environmental correction factor data are normalized by the normalization layer to obtain the first normalized feature vector, the second normalized feature vector and the third normalized feature vector.

[0034] S52. Calculate the dynamic fusion weights of the first normalized feature vector, the second normalized feature vector, and the third normalized feature vector through the weight calculation layer to obtain the weighted first feature vector, the weighted second feature vector, and the weighted third feature vector.

[0035] S53. By using a splicing and fusion layer, the weighted first feature vector, the weighted second feature vector, and the weighted third feature vector are spliced ​​along the feature dimension to obtain a spliced ​​feature tensor.

[0036] S54. By performing nonlinear transformation and cross-layer feature mapping on the spliced ​​feature tensor through the feature mapping layer, the fault fusion feature vector is obtained.

[0037] In one embodiment, the fault classification and identification model includes a feature reconstruction layer, a metric space calculation layer, a radial basis function activation layer, and a decision-making layer, wherein S6 includes:

[0038] S61. The fault fusion feature vector is subjected to orthogonal basis transformation and feature decoupling through the feature reconstruction layer to obtain the orthogonal decoupled feature vector;

[0039] S62. Calculate the Mahalanobis distance between the orthogonal decoupled feature vector and the preset lightning fault category center matrix through the metric space computation layer to obtain the category relative distance vector; where the expression for the Mahalanobis distance is:

[0040]

[0041] in, Represents the orthogonal decoupled eigenvectors and the first The Mahalanobis distance of the center of each lightning strike fault category This represents the orthogonal decoupling eigenvector. The first element in the center matrix of lightning strike fault categories Each lightning strike fault category center vector This represents the matrix transpose operation. Indicates the first The inverse matrix of the covariance matrix of each lightning strike fault category. The serial number indicating the type of lightning strike fault;

[0042] S63. By performing nonlinear probability mapping on the relative distance vector of categories through a radial basis function activation layer, the probability distribution vector of lightning strike failure is obtained.

[0043] S64. The lightning fault identification result is obtained by performing maximum probability matching and category index mapping on the probability distribution vector of lightning faults through the decision-making layer.

[0044] In one embodiment, S63 includes:

[0045] S631. Perform multi-scale mapping on the relative distance vectors of categories based on multi-kernel radial basis functions to obtain the mapped values, and obtain the multi-kernel similarity matrix based on the mapped values; wherein, the expression for the mapped values ​​is:

[0046]

[0047] in, For the first The categories in the Mapping values ​​on the radial basis function kernel For the first Mahalanobis distance for each category For the first Bandwidth parameters per core;

[0048] S632. Perform feature weighting on the multi-kernel similarity matrix to obtain a weighted similarity feature vector;

[0049] S633. Normalize the weighted similarity feature vector to obtain the probability distribution vector of lightning strike failure.

[0050] Secondly, this application also provides an intelligent identification system for lightning strike faults in high-voltage transmission lines, comprising:

[0051] The multi-source data acquisition module is used to acquire lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line.

[0052] The lightning transient feature extraction module is used to perform joint time-frequency domain feature extraction on lightning current waveform data and line traveling wave data to obtain lightning transient feature data.

[0053] The insulator state feature extraction module is used to perform waveform decomposition and modal analysis on insulator leakage current data to obtain insulator state feature data.

[0054] The environmental correction factor analysis module is used to perform multi-factor correlation analysis on meteorological and environmental data to obtain environmental correction factor data.

[0055] The hierarchical feature fusion module is used to input lightning transient feature data, insulator state feature data and environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain fault fusion feature vector;

[0056] The fault classification and identification module is used to input the fault fusion feature vector into the fault classification and identification model to identify the type of lightning strike fault and obtain the lightning strike fault identification result.

[0057] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0058] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0059] The aforementioned intelligent identification method and system for lightning strike faults in high-voltage transmission lines simultaneously acquires four types of multi-source heterogeneous data: lightning current waveform, line traveling wave, insulator leakage current, and meteorological environment data. It performs joint time-frequency domain feature extraction on the lightning current waveform and line traveling wave to fully explore the deep-seated time-domain transient and frequency-domain energy characteristics of the lightning strike transient process. Simultaneously, it performs waveform decomposition and modal analysis on the insulator leakage current to quantify the insulator surface state, and conducts multi-factor correlation analysis on the meteorological environment data to quantify the influence weight of meteorological factors. Then, a hierarchical fusion identification network adaptively weights and fuses the above three types of feature data, finally inputting the data into a fault classification and identification model to determine the type of lightning strike fault. This sampling method can achieve fine differentiation between lightning flashover, lightning-non-flashover, and non-lightning strike faults, improving the accuracy and robustness of lightning strike fault identification under complex operating conditions. It can reduce the risk of misjudgment and missed judgment caused by a single data source and reduce the interference of meteorological environment and insulator state changes on the identification results. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating an intelligent identification method for lightning strike faults in high-voltage transmission lines, as shown in one embodiment.

[0062] Figure 2 This is a schematic diagram of the structure of an intelligent identification system for lightning strike faults on high-voltage transmission lines in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] refer to Figure 1 The document presents a flowchart illustrating an intelligent identification method for lightning strike faults in high-voltage transmission lines, as provided in this application. The method includes the following steps:

[0065] S1. Acquire lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line.

[0066] Optionally, lightning current waveform data can reflect the current change pattern during a lightning strike. It can be collected by a lightning current sensor (Rogowski coil sensor) installed on a transmission line tower. The lightning current waveform data contains key information such as the amplitude, steepness, and duration of the lightning strike current.

[0067] Optionally, the line traveling wave data is the transient traveling wave signal generated when a lightning strike fault occurs. It can be collected by a traveling wave sensor. The line traveling wave data includes transient characteristics related to the fault location and fault type.

[0068] Optionally, insulator leakage current data is a weak current signal generated when partial discharge or flashover occurs on the surface of the insulator. It can be collected by a leakage current monitoring device, and the insulator leakage current data can reflect the insulation status of the insulator.

[0069] Optionally, meteorological environmental data refers to environmental parameters at the time of the lightning strike, which may include atmospheric humidity, ambient temperature, atmospheric electric field strength, wind speed, and precipitation intensity.

[0070] Optionally, the acquisition of the four types of data—lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data—must be synchronized. The acquisition frequency must meet the requirements for transient feature capture, and the acquisition duration must cover 10ms before the lightning strike to 50ms after the fault, ensuring that the data can fully characterize the transient process and influencing factors of the lightning strike fault. To provide a foundation for subsequent feature extraction and fusion analysis, all types of data must be stored in a standard format for easy retrieval in subsequent data processing steps. Specifically, lightning current waveform data and line traveling wave data can be stored in a time series format, insulator leakage current data can be stored in a continuous sampling sequence format, and meteorological environment data can be stored in a multi-dimensional time series matrix format.

[0071] S2. Perform joint time-frequency domain feature extraction on lightning current waveform data and line traveling wave data to obtain lightning transient feature data.

[0072] For example, when a lightning strike occurs, both the lightning current waveform and the line traveling wave exhibit complex transient characteristics. Feature extraction using only one dimension—time domain or frequency domain—cannot fully capture the fault characteristics. Therefore, a joint time-frequency domain feature extraction method can be used to consider both the transient characteristics in the time domain and the energy distribution characteristics in the frequency domain. The core principle of joint time-frequency domain feature extraction is to use different signal processing methods to extract the time-domain and frequency-domain features of the two types of data, and then fuse these features to obtain feature data that can comprehensively characterize the transient process of a lightning strike. Specifically, the time-domain features of the lightning current waveform data mainly reflect the instantaneous change pattern of the current, while the frequency-domain features of the line traveling wave data mainly reflect the energy distribution pattern of the transient signal. Combining these two methods can effectively distinguish the subtle differences between lightning flashover, lightning-without-flashover, and non-lightning strike faults.

[0073] Optionally, a standardization process should be used during feature extraction to eliminate the influence of data units. The standardization expression can be: ,in, These are the standardized eigenvalues. These are the original eigenvalues. This is the minimum value of the feature. The maximum value of this feature is given by the formula, which can map various feature values ​​to the [0,1] interval to avoid the influence of feature amplitude differences on subsequent fusion effects. The final lightning transient feature data is a multi-dimensional feature vector that can comprehensively cover the key information in the time and frequency domains of the lightning transient process.

[0074] S3. Perform waveform decomposition and modal analysis on the insulator leakage current data to obtain the insulator state characteristic data.

[0075] For example, insulator leakage current data can reflect the insulation state of the insulator. When a lightning fault occurs, if flashover or partial discharge occurs on the insulator surface, the leakage current will change significantly. Therefore, by performing waveform decomposition and modal analysis on the insulator leakage current data, characteristic data for characterizing the insulator state can be extracted, providing an insulator state reference for lightning fault identification. The purpose of waveform decomposition is to separate the effective signal from the noise signal in the insulator leakage current data, and the purpose of modal analysis is to mine the inherent modal characteristics of the leakage current waveform, thereby reflecting changes in the insulation state of the insulator. Specifically, the insulator leakage current data is first preprocessed to eliminate noise interference, then the data is decomposed into multiple modal components using waveform decomposition methods, then features are extracted from each modal component, and finally, feature compression is performed through an encoding network to obtain the insulator state characteristic data. The feature extraction process requires calculating the key feature parameters of each modal component and establishing the correlation between the insulator state and the feature parameters. Feature compression is used to reduce the feature dimensionality, avoid the curse of dimensionality, and retain key feature information. The insulator state feature data can effectively distinguish between the normal state, slightly polluted state, severely polluted state and flashover state of the insulator, providing important support for the identification of lightning fault types.

[0076] S4. Perform multi-factor correlation analysis on meteorological and environmental data to obtain environmental correction factor data.

[0077] For example, various meteorological parameters in meteorological environmental data (atmospheric humidity, ambient temperature, atmospheric electric field strength, wind speed, and precipitation intensity) all affect the occurrence and development of lightning strike faults. For instance, when atmospheric humidity increases, the insulation performance of the insulator surface decreases, making flashover faults more likely. When atmospheric electric field strength increases, the probability of lightning strikes significantly increases. Therefore, it is necessary to use multi-factor correlation analysis to explore the correlation between various meteorological parameters and lightning strike faults, obtaining environmental correction factor data. This environmental correction factor data can be used to correct lightning transient characteristic data and insulator state characteristic data, improving the accuracy of fault identification. The purpose of multi-factor correlation analysis is to quantify the degree of correlation between various meteorological parameters and lightning strike faults, and by weighting each meteorological parameter, obtain environmental correction factor data that reflects the degree of influence of the meteorological environment on lightning strike faults.

[0078] S5. Input the lightning strike transient characteristic data, insulator state characteristic data and environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain the fault fusion feature vector.

[0079] For example, transient lightning strike feature data reflects the transient process of a lightning fault, insulator state feature data reflects the insulation state of the insulator, and environmental correction factor data reflects the impact of meteorological environment on lightning faults. These three types of feature data are complementary; using any single type alone cannot achieve accurate identification of lightning faults. Therefore, a hierarchical fusion identification network is needed to fuse features, fully integrating the complementary information of the three types of feature data to obtain a fault fusion feature vector that comprehensively characterizes the features of lightning faults. The hierarchical fusion identification network can adopt a layered fusion approach, first normalizing each type of feature data, then calculating dynamic fusion weights, then concatenating features, and finally obtaining the fault fusion feature vector through feature mapping. This ensures that the fused feature vector has strong discriminative power and can effectively distinguish different types of lightning faults.

[0080] S6. Input the fault fusion feature vector into the fault classification and recognition model to identify the type of lightning strike fault and obtain the lightning strike fault recognition result.

[0081] Optionally, the fault fusion feature vector is input into the fault classification and identification model for lightning fault type identification to obtain the lightning fault identification result. The fault fusion feature vector contains all the key information of lightning transient features, insulator state features, and environmental correction features, which can comprehensively characterize the features of lightning faults. The core function of the fault classification and identification model is to determine the specific type of lightning fault through the analysis and identification of the fault fusion feature vector. The specific type of lightning fault can include lightning flashover faults, lightning non-flashover faults, and non-lightning faults. The specific type of lightning fault can provide an accurate basis for fault handling and maintenance of the power system. The fault classification and identification model adopts a hierarchical identification structure, and through feature reconstruction, metric space calculation, nonlinear probability mapping, and decision judgment, it gradually achieves accurate identification of fault types, ensuring the accuracy and reliability of the identification results.

[0082] The aforementioned intelligent identification method for lightning-induced faults in high-voltage transmission lines simultaneously acquires four types of multi-source heterogeneous data: lightning current waveform, line traveling wave, insulator leakage current, and meteorological environment data. First, it performs joint time-frequency domain feature extraction on the lightning current waveform and line traveling wave using empirical mode decomposition and continuous wavelet transform to fully explore the deep time-frequency characteristics of the lightning transient process. Simultaneously, it performs variational mode decomposition and mode analysis on the insulator leakage current to extract features characterizing the surface contamination and moisture state of the insulator. Then, it calculates the mutual information coefficient of the meteorological environment data and performs multi-factor attention encoding to obtain an environmental correction factor. Finally, it inputs the above three types of features into a hierarchical fusion identification network for dynamic weighted fusion. Finally, it outputs the identification result through a fault classification model based on Mahalanobis distance and multi-kernel radial basis functions. This method can comprehensively characterize the subtle differences between lightning flashover, non-flashover, and non-lightning-induced faults, improving the accuracy and robustness of lightning fault identification, reducing dependence on a single data source, and adapting to engineering application needs under complex meteorological and insulator condition changes.

[0083] In an optional embodiment, S2 includes:

[0084] S21. Perform empirical mode decomposition on the lightning current waveform data to obtain the intrinsic mode function component data of the lightning current.

[0085] For example, Empirical Mode Decomposition (EMD) is an adaptive signal decomposition method that does not require preset basis functions. It can decompose non-stationary and nonlinear lightning current waveform data into multiple physically meaningful intrinsic mode function components and a residual component. Each intrinsic mode function component satisfies two conditions: first, the number of extreme points in the entire data sequence is equal to or differs from the number of zero crossings by no more than 1; second, at any time, the average value of the upper envelope formed by local maxima and the lower envelope formed by local minima is zero.

[0086] Optionally, the specific process of empirical mode decomposition is as follows: First, determine all local maxima and local minima of the lightning current waveform data. Fit the upper and lower envelopes using cubic spline interpolation. Calculate the average value of the upper and lower envelopes. Subtract this average value from the lightning current waveform data to obtain the initial component. Determine whether the initial component satisfies the conditions for intrinsic mode function components. If not, use this initial component as new raw data and repeat the above process until intrinsic mode function components that meet the conditions are obtained. The remaining part is the residual component. A termination condition needs to be set during the decomposition process. Decomposition stops when the amplitude of the residual component is less than a preset threshold or the number of decompositions reaches a preset maximum value. The decomposed intrinsic mode function component data of the lightning current can respectively characterize the transient change features of the lightning current waveform at different time scales, providing a basis for subsequent time-domain feature extraction. Its decomposition formula can be expressed as: ,in, This is the raw data of the lightning current waveform. For the first Data on the intrinsic mode function components of lightning current. For residual component data, This represents the number of intrinsic mode function components.

[0087] S22. Perform continuous wavelet transform on the traveling wave data of the line to obtain the traveling wave time spectrum data.

[0088] Optionally, the Continuous Wavelet Transform (CWT) is a time-frequency analysis method. By convolving the traveling wave data with wavelet basis functions, it can analyze signals at different scales while preserving both time and frequency domain information. It is suitable for extracting time-frequency features of non-stationary transient signals. The core principle of the CWT is to use the scaling and translation of wavelet basis functions to decompose the traveling wave data at multiple scales, obtaining wavelet coefficients at different scales, and then constructing the traveling wave time spectrum, which reflects the energy distribution of the traveling wave signal at different time and frequency points. The expression for the CWT can be: ,in, These are continuous wavelet transform coefficients. The scaling factor is used to control the scaling of the wavelet basis functions. The smaller the scale factor, the narrower the wavelet basis function, the higher the time resolution, and the lower the frequency resolution; the larger the scale factor, the wider the wavelet basis function, the lower the time resolution, and the higher the frequency resolution. This is a translation factor used to control the translation of the wavelet basis function, reflecting the characteristics of the signal at different time points; This is the raw data of the traveling wave of the line. It is the conjugate function of the wavelet basis functions. The time variable is used. During implementation, a suitable wavelet basis function (such as the db4 wavelet) is selected, and the range and step size of the scaling factor are set to ensure that the time-frequency characteristics of the traveling wave signal can be fully captured. The traveling wave time-frequency spectrum data obtained by continuous wavelet transform is a two-dimensional matrix. The rows of the matrix correspond to the time dimension, the columns correspond to the frequency dimension, and the matrix element values ​​correspond to the signal energy at the corresponding time and frequency points, which can clearly reflect the time-frequency distribution law of the traveling wave signal.

[0089] S23. Input the intrinsic mode function component data of lightning current and the time spectrum data of traveling wave into the dual-channel feature extraction network to extract the time-domain transient features and frequency-domain energy features, and obtain the time-domain feature vector of lightning current and the frequency-domain feature vector of traveling wave.

[0090] Optionally, the dual-channel feature extraction network may include a first channel and a second channel. The first channel may include a one-dimensional convolutional layer and a max-pooling layer, and can be used to extract the waveform steepness and peak attenuation rate of the intrinsic mode function components of lightning current. The second channel may include a two-dimensional convolutional layer and a global average pooling layer, and can be used to extract the frequency band centroid and energy concentration of the traveling wave spectrum. The core function of the dual-channel feature extraction network is that it can use an appropriate feature extraction structure for two different types of data to extract their core features separately, avoiding the problem that a single feature extraction structure cannot adapt to different data types.

[0091] Optionally, the one-dimensional convolutional layer in the first channel can be used to extract features from the intrinsic mode function component data of lightning current. The kernel size is set according to the data sampling frequency. The local temporal features of the data are captured through convolution operations. The convolution operation formula can be: ,in, The convolutional layer outputs feature values. For convolution kernel weights, The input is the eigenmode function component data of the lightning current. For bias terms, Where is the kernel size. Max pooling layers can be used to reduce the dimensionality of features output by convolutional layers, preserving key temporal features and reducing the amount of data. The max pooling formula is as follows: ,in, The pooling layer outputs feature values. To determine the pooling window size, the lightning current time-domain feature vector can be obtained through the first channel. The lightning current time-domain feature vector can include key time-domain features such as waveform steepness and peak attenuation rate.

[0092] Optionally, the two-dimensional convolutional layer of the second channel can be used to extract features from the traveling wave time-frequency data, capturing joint time-frequency features. The convolution kernel is a two-dimensional matrix, and the convolution operation formula can be: ,in, The output feature values ​​of the two-dimensional convolutional layer For two-dimensional convolution kernel weights, The input is the time-spectral data of the traveling wave. For bias terms, , These represent the height and width of the convolutional kernel, respectively. The global average pooling layer performs a global average operation on the feature map output by the two-dimensional convolutional layer to obtain frequency domain features of a fixed dimension. The formula for global average pooling is: ,in, The output feature values ​​of the global average pooling layer. , These represent the height and width of the feature map, respectively. The number of feature map channels is represented by the second channel. The traveling wave frequency domain feature vector can be obtained through the second channel. The traveling wave frequency domain feature vector can contain key frequency domain features such as the centroid of the frequency band and the energy concentration.

[0093] S24. Cross-attention fusion is performed on the lightning current time domain feature vector and the traveling wave frequency domain feature vector to obtain lightning transient feature data.

[0094] For example, the core principle of cross-attention fusion is to calculate the attention weight between two types of feature vectors, highlight the features that contribute more to lightning strike fault identification, suppress irrelevant features, achieve efficient fusion of the two types of features, and improve the discriminative ability of features.

[0095] For example, the specific process of cross-attention fusion is as follows: First, the time-domain feature vector and the frequency-domain feature vector of the lightning current are linearly transformed and mapped to the same feature space. The linear transformation formula can be: , , ,in, The query vector is the result of a linear transformation of the lightning current time-domain feature vector. The key vector is the result of a linear transformation of the traveling wave frequency domain eigenvectors. This is the value vector obtained by linear transformation of the frequency domain eigenvectors of the traveling wave. , , These are the linear transformation weight matrices corresponding to the query vector, key vector, and value vector, respectively. This is the time-domain eigenvector of lightning current. Let the key vector be the frequency domain feature vector of the traveling wave. Then, calculate the similarity between the query vector and the key vector to obtain the attention score matrix. The similarity can be calculated using the dot product method. The expression for the attention score matrix can be: ,in, For attention score matrix, The dimension of the key vector, divided by To avoid the vanishing gradient problem caused by excessively large attention scores, the attention score matrix is ​​then subjected to softmax normalization to obtain the attention weight matrix. The normalization expression can be: ,in, This is the normalized attention weight matrix. The first in the attention score matrix Line number Column elements, Let be the dimension of the key vector; finally, the attention weight matrix and the value vector are weighted and summed to obtain the fused feature vector, i.e., the transient feature data of lightning strikes. The weighted summation formula is: ,in, This method, which uses transient feature data of lightning strikes, can fully utilize the complementary information of the two types of features to improve the comprehensiveness and discriminative ability of transient features of lightning strikes.

[0096] In an optional embodiment, S3 includes:

[0097] S31. Perform sliding window truncation and adaptive filtering on the insulator leakage current data to obtain pure leakage current segment data.

[0098] Optionally, insulator leakage current data is affected by environmental noise, electromagnetic interference, and other factors during acquisition, resulting in a large amount of irrelevant noise in the data, affecting the accuracy of subsequent waveform decomposition and modal analysis. Therefore, preprocessing is required through sliding window truncation and adaptive filtering. The purpose of sliding window truncation is to extract effective data segments containing the transient process of lightning strike faults from continuously acquired insulator leakage current data, avoiding interference from invalid data. The size of the sliding window can be set according to the sampling frequency of the leakage current and the duration of the lightning strike fault, and the window sliding step size is set according to the data continuity requirements. During the truncation process, it is necessary to ensure that the extracted segments can completely cover the leakage current change process before and after the lightning strike fault. The specific calculation formula is as follows: ,in, This is the data of the truncated leakage current segment. This is the original insulator leakage current data. For sliding window functions, The window start time is specified. The window function takes a value of 1 within the window and a value of 0 outside the window.

[0099] Optionally, adaptive filtering can be used to eliminate noise interference in the truncated data. An adaptive filtering algorithm (such as the least mean square adaptive filtering algorithm) can be employed. Its core principle is to adaptively adjust the filter coefficients to minimize the error between the filter output and the pure leakage current signal. The update formula for adaptive filtering can be: ,in, For the first The filter coefficient vector of the next iteration. For the first The filter coefficient vector of the next iteration. This is the step size factor, used to control the rate at which coefficients are updated. For the first The filtering error of the next iteration is the difference between the filter output and the desired output. For the first The input leakage current data of the next iteration, after adaptive filtering, can effectively filter out noise signals and obtain clean leakage current segment data, providing a high-quality data foundation for subsequent waveform decomposition.

[0100] S32. Based on the constrained variational objective function, variational mode decomposition is performed on the pure leakage current segment data to obtain multiple intrinsic mode function component data and residual component data.

[0101] Optionally, Variational Mode Decomposition (VMD) is an adaptive signal decomposition method based on variational principles. Compared to Empirical Mode Decomposition (EMD), it has better noise resistance and decomposition stability. It can decompose pure leakage current segment data into multiple eigenmode function components with center frequencies and a residual component. Each eigenmode function component is a narrowband signal, which can accurately reflect the variation characteristics of leakage current in different frequency ranges. The expression for the constrained variational objective function can be:

[0102]

[0103] in, For time The first time Each intrinsic mode function component For the first The center frequency of each component To preset the number of decomposition layers, For the Dirac function, This represents the convolution operation. Indicates time The partial derivatives, The imaginary unit, For time variables, The core of the constrained variational objective function is to minimize the squared L2 norm of the analytic signal after the Hilbert transform of each eigenmode function component, thereby achieving frequency domain separation of the eigenmode function components. The Hilbert transform is achieved through... accomplish, Used to shift the spectrum of each intrinsic mode function component to a position where... The L2 norm is used to measure the energy of an analytic signal within the centered frequency range. By minimizing this objective function, the spectra of each eigenmode function component can be made non-overlapping, achieving efficient decomposition. During implementation, the number of decomposition layers is first preset. The intrinsic mode function components and center frequencies are initialized, and then the constrained variational problem is solved by an alternating iterative optimization algorithm to update the intrinsic mode function components and center frequencies until the convergence condition is met. The convergence condition is that the change in each center frequency in two iterations is less than a preset threshold. Finally, multiple intrinsic mode function component data and residual component data are obtained. The residual component data reflects the trend term of the leakage current, and the intrinsic mode function component data can be used to reflect the transient change characteristics of the leakage current.

[0104] S33. Calculate the peak amplitude, root mean square value and frequency band energy ratio of each intrinsic mode function component data and residual component data to obtain the multidimensional characteristic vector of leakage current.

[0105] Optionally, peak amplitude, root mean square (RMS) value, and frequency band energy ratio are key parameters characterizing the waveform characteristics of leakage current, which can reflect the insulation state of the insulator. Among them, peak amplitude can be used to reflect the maximum intensity of leakage current, RMS value can be used to reflect the average intensity of leakage current, and frequency band energy ratio can be used to reflect the energy distribution of leakage current in different frequency ranges. The combination of the three can comprehensively characterize the waveform characteristics of leakage current.

[0106] For example, the formula for calculating the peak amplitude can be: ,in, For the first Peak amplitude of each intrinsic mode function component data For the first The intrinsic mode function component data. Given the time length of the data, this formula can be used to calculate the maximum absolute value of each intrinsic mode function component and residual component over the entire time range, i.e., the peak amplitude.

[0107] For example, the formula for calculating the root mean square value can be: ,in, For the first The root mean square value of each intrinsic mode function component data. The number of sampling points for the data. For the first The time for each sampling point For the first The intrinsic mode function component data at the th The formula calculates the root mean square value by taking the square root of the average square of the values ​​at each sampling point. The root mean square value can be used to reflect the average intensity of the leakage current.

[0108] For example, the formula for calculating the frequency band energy ratio can be: ,in, For the first The frequency band energy ratio of each intrinsic mode function component data For the first The energy of each intrinsic mode function component data For the first The energy of each component (including all intrinsic mode function components and residual components), energy The calculation formula can be The formula calculates the proportion of energy of each component to the total energy, which can be used to reflect the energy distribution characteristics of the leakage current. After calculating the peak amplitude, root mean square value, and frequency band energy ratio of all intrinsic mode function components and residual components, these parameters are concatenated in sequence to obtain a multidimensional feature vector of the leakage current. The dimension of this vector is 3×(K+1), where K is the preset decomposition level, which can comprehensively cover the waveform characteristics and energy distribution characteristics of the leakage current.

[0109] S34. Input the multidimensional feature vector of leakage current into the insulator state coding network, perform feature compression, and obtain insulator state feature data.

[0110] Optionally, the insulator state coding network may include an insulation coding input layer, a fully connected coding layer, and an output layer. The insulation coding input layer can be used to receive the multi-dimensional feature vector of leakage current, the fully connected coding layer can be used to compress the feature dimension of the multi-dimensional feature vector of leakage current, and the output layer can be used to output the insulator state feature data. The multi-dimensional feature vector of leakage current has a high dimensionality. If it is directly used for subsequent feature fusion, it will increase the computational load and may contain redundant features, affecting the recognition efficiency and accuracy. Therefore, it is necessary to perform feature compression through the insulator state coding network to retain key features and reduce the feature dimension.

[0111] For example, the purpose of the insulation coding input layer is to receive the multidimensional feature vector of the leakage current and standardize it to eliminate the influence of feature dimensions. The standardization process can use the same formula as S2 to ensure that the feature values ​​are mapped to the [0,1] interval. The fully connected coding layer can be composed of multiple fully connected neural networks. Each fully connected layer processes the input features through linear transformation and activation functions to achieve feature dimension compression. The linear transformation formula can be: ,in, For the first The output feature vector of a fully connected layer For the first The weight matrix of the fully connected layer. For the first The output feature vector of a fully connected layer For the first The bias term of the fully connected layer; the activation function can be the ReLU function, and its expression can be: The ReLU function is used to introduce non-linear features, enhancing the feature representation capability of the encoding network. The output layer can be a fully connected layer. A linear transformation maps the output features of the fully connected encoding layer to a preset dimension, obtaining the insulator state feature data. The linear transformation formula is as follows: ,in, This is insulator state characteristic data. The weight matrix of the output layer. This is the output feature vector of the fully connected coding layer. This is the bias term for the output layer. By compressing the features of the insulator state coding network, the feature dimensionality can be reduced while retaining the key features of leakage current, thus obtaining insulator state feature data that accurately characterizes the insulator state.

[0112] In an optional embodiment, S4 includes:

[0113] S41. Impute missing values ​​and standardize the meteorological and environmental data to obtain a standardized meteorological data matrix.

[0114] Optionally, meteorological environmental data may include meteorological parameters. During the acquisition process, meteorological environmental data may contain missing values ​​due to sensor malfunctions, signal transmission interruptions, or other reasons. If these data are directly used for correlation analysis, it will affect the accuracy of the analysis results. Therefore, missing value imputation is required. At the same time, the different units of various meteorological parameters (e.g., atmospheric humidity is expressed as a percentage, and ambient temperature is expressed as a degree Celsius) can lead to significant differences in characteristic amplitudes, affecting the rationality of subsequent weighted calculations. Therefore, standardization processing is necessary.

[0115] Alternatively, missing value imputation can use linear interpolation, which is suitable for scenarios with few missing values ​​and relatively stable data changes. The interpolation formula can be: ,in, The imputation result for missing values, The value of the previous valid data point before the missing value. The value of the next valid data point after the missing value. The time of the previous valid data point before the missing value. The time of the last valid data point after the missing value. The formula represents the time corresponding to the missing value. It can accurately impute missing values ​​based on the values ​​of the valid data points before and after the missing value.

[0116] Optionally, the standardization process can employ the Z-score standardization method, mapping each meteorological parameter to a normal distribution with a mean of 0 and a standard deviation of 1. The standardization formula can be: ,in, These are the standardized meteorological parameter values. These are the original meteorological parameter values. This is the mean value of the meteorological parameter. The standard deviation of this meteorological parameter is used to eliminate the influence of dimensions. After imputing missing values ​​and standardizing all meteorological parameters, a standardized meteorological data matrix is ​​constructed by arranging them by rows. The rows of the matrix correspond to different time points, and the columns correspond to different meteorological parameters. The matrix element values ​​are the standardized meteorological parameter values, providing a standardized data foundation for subsequent correlation analysis.

[0117] S42. Combining the lightning fault labels and the standardized meteorological data matrix, calculate the mutual information coefficients between each meteorological parameter and the lightning fault labels to obtain the initial correlation weight vector.

[0118] Optionally, by calculating the mutual information coefficients between each meteorological parameter and the lightning strike fault label, the influence of each meteorological parameter on the lightning strike fault can be quantified, thereby obtaining the initial association weight vector. The expression for the mutual information coefficient can be: ,in, For the first The mutual information coefficient between meteorological parameters and lightning fault labels, For the first A meteorological parameter, This is a historical lightning strike fault label. A value of 0 indicates a non-lightning strike fault, and a value of 1 indicates a lightning strike fault. Indicates the values ​​of meteorological parameters And the lightning strike fault label is The probability, Indicates the values ​​of meteorological parameters The probability, The value of the lightning strike fault label is indicated. The probability of mutual information coefficients. Mutual information coefficient is an index that quantifies the degree of correlation between two random variables. It ranges from [0,∞). The larger the mutual information coefficient, the stronger the correlation between the two random variables, and vice versa.

[0119] Specifically, firstly, historical lightning strike fault data and corresponding meteorological environmental data are collected to construct a historical dataset, where historical lightning strike fault labels are marked according to the actual fault conditions; then, based on the historical dataset, the probability of each meteorological parameter value is calculated. Probability of the value of the lightning strike fault label and the joint probability of the two The probability calculation can be performed using the frequency estimation method, which involves statistically analyzing the frequency of each value as a probability estimate. Then, the calculated probability value is substituted into the expression for the mutual information coefficient to calculate the mutual information coefficient between each meteorological parameter and the lightning fault label. Finally, the mutual information coefficients of all meteorological parameters are arranged in order to obtain the initial association weight vector. Each element of the vector corresponds to the mutual information coefficient of a meteorological parameter, which can be used to reflect the initial association degree between the corresponding meteorological parameter and the lightning fault.

[0120] S43. Input the standardized meteorological data matrix and the initial correlation weight vector into the preset multi-factor attention coding network for nonlinear feature weighting to obtain the meteorological impact feature vector.

[0121] Optionally, the multi-factor attention coding network may include an attention coding input layer, a multi-head self-attention layer, a layer normalization layer, and a feedforward network layer. The attention coding input layer can be used to receive a standardized meteorological data matrix and an initial correlation weight vector. The multi-head self-attention layer can be used to calculate the attention score between each meteorological parameter based on the initial correlation weight vector. The layer normalization layer can be used to normalize the attention score. The feedforward network layer can be used to output a meteorological impact feature vector. The purpose of the multi-factor attention coding network is to further explore the intrinsic correlation between each meteorological parameter based on the initial correlation weight vector through an attention mechanism, to nonlinearly weight each meteorological parameter, to highlight meteorological parameters with greater influence and to suppress meteorological parameters with less influence, so as to obtain a meteorological impact feature vector that accurately reflects the impact of the meteorological environment on lightning strike faults.

[0122] For example, the attention encoding input layer can concatenate a standardized meteorological data matrix and an initial association weight vector to obtain an input feature matrix. The concatenation method is along the feature dimension, ensuring that the input features simultaneously contain meteorological parameter information and initial association weight information. The multi-head self-attention layer can divide the input feature matrix into multiple heads, calculate the attention score for each head separately, and then concatenate the attention scores of multiple heads to improve the feature capture capability of the attention mechanism. The attention score calculation for each head uses the same cross-attention calculation method as S24, adjusting the weights of each meteorological parameter based on the initial association weight vector. The calculation formula is: ,in, , For the number of heads, , , The first The query, key, and value linear transformation weight matrix of the head. This is the concatenated linear transformation weight matrix. The layer normalization layer can be used to normalize the attention score output by the multi-head self-attention layer, eliminating the gradient vanishing problem. The normalization formula is: ,in, The mean of the input features. The variance of the input features. To prevent the use of tiny constants with a denominator of zero, the feedforward network layer consists of two fully connected layers and a ReLU activation function. It can be used to perform a nonlinear transformation on the normalized attention score and output a meteorological impact feature vector. The calculation method of the fully connected layer is consistent with that of the fully connected encoding layer in S34. Through the feedforward network layer, the nonlinear characteristics of meteorological parameters can be further explored, and the expressive power of the meteorological impact feature vector can be improved.

[0123] S44. Input the meteorological impact feature vector into the spatial mapping network for linear transformation to obtain environmental correction factor data.

[0124] Optionally, the meteorological impact feature vector can reflect the degree of influence of the meteorological environment on lightning strike faults. However, the dimension and value range of this feature vector are inconsistent with the transient feature data of lightning strikes and the insulator state feature data, and cannot be directly used for subsequent feature fusion. Therefore, a spatial mapping network is needed to perform a linear transformation to map the meteorological impact feature vector to the same feature space as the other two types of feature data, obtaining environmental correction factor data to ensure that the three types of feature data can be effectively fused. The spatial mapping network can be a single-layer fully connected network, the purpose of which is to adjust the dimension and value range of the meteorological impact feature vector through linear transformation. The linear transformation formula can be: ,in, For environmental correction factor data, Here is the weight matrix of the spatial mapping network. This is a feature vector representing meteorological impacts. This represents the bias term of the spatial mapping network. During implementation, the weight matrix... and bias terms The training objective, determined through training on historical datasets, is to enable efficient fusion of environmental correction factor data with lightning transient characteristic data and insulator state characteristic data within the same feature space, thereby improving the accuracy of fault identification. Through linear transformation of the spatial mapping network, the resulting environmental correction factor data shares the same dimension and value range as the other two types of feature data, accurately reflecting the corrective effect of meteorological environment on lightning faults and providing suitable feature data for subsequent feature fusion.

[0125] In an optional embodiment, the hierarchical fusion recognition network includes a normalization layer, a weight calculation layer, a splicing and fusion layer, and a feature mapping layer, wherein S5 includes:

[0126] S51. The transient characteristic data of lightning strike, the state characteristic data of insulator and the environmental correction factor data are normalized by the normalization layer to obtain the first normalized feature vector, the second normalized feature vector and the third normalized feature vector.

[0127] Optionally, the transient characteristic data of lightning strikes, the state characteristic data of insulators, and the environmental correction factor data are respectively normalized using a normalization layer to obtain a first normalized feature vector, a second normalized feature vector, and a third normalized feature vector. The value ranges and dimensions of the three types of feature data differ. Direct fusion would result in a fusion result biased towards feature data with larger amplitudes, affecting the fusion effect. Therefore, it is necessary to standardize the various feature data through layer normalization, ensuring that the mean of each type of feature data is 0 and the variance is 1, thus eliminating the influence of differences in dimensions and amplitudes. The formula for calculating layer normalization can be:

[0128]

[0129] in, The normalized feature vectors, This is the original feature data (lightning transient feature data, insulator state feature data, or environmental correction factor data). The mean of the corresponding feature data. The variance of the corresponding feature data. To prevent the use of tiny constants with a denominator of zero, the normalization layer processes the three types of feature data independently. For the lightning strike transient feature data, the mean and variance of all its feature values ​​are calculated and substituted into the above formula to obtain the first normalized feature vector. For the insulator state feature data, the mean and variance of all its feature values ​​are calculated and substituted into the above formula to obtain the second normalized feature vector. For the environmental correction factor data, the mean and variance of all its feature values ​​are calculated and substituted into the formula to obtain the third normalized feature vector. After layer normalization, the three types of feature data have the same value range and statistical characteristics, providing a foundation for subsequent dynamic weight calculation and feature concatenation.

[0130] S52. Calculate the dynamic fusion weights of the first normalized feature vector, the second normalized feature vector, and the third normalized feature vector through the weight calculation layer to obtain the weighted first feature vector, the weighted second feature vector, and the weighted third feature vector.

[0131] Optionally, the dynamic fusion weights of the first, second, and third normalized feature vectors are calculated through a weight calculation layer to obtain weighted first, second, and third feature vectors. In different types of lightning strike faults, the importance of the three types of feature data varies. For example, in lightning flashover faults, the insulator state feature data is more important than the other two types of feature data; in non-lightning strike faults, the lightning transient feature data is less important. Therefore, it is necessary to calculate dynamic fusion weights through a weight calculation layer, assigning different weights according to the importance of the feature data to improve the discriminative ability of the fused features. The calculation of dynamic fusion weights can employ an attention mechanism. The weight calculation layer determines the weights of each type of feature data by calculating the similarity between the three types of normalized feature vectors. The formula for calculating the dynamic fusion weights can be: ,in, For the first Dynamic fusion weights for class-normalized feature vectors These correspond to the first, second, and third normalized feature vectors, respectively. For the first Attention score of class-normalized feature vectors, attention score The similarity between this type of feature vector and the fusion target vector is calculated. The fusion target vector is the mean vector of the three types of normalized feature vectors. The calculation formula is as follows: ,in, For the first Normalized feature vectors To fuse the target vector, The cosine similarity function is used to calculate the similarity between two vectors. The higher the similarity, the higher the attention score and the greater the dynamic fusion weight. After calculating the dynamic fusion weight, the normalized feature vectors of each type are multiplied by their corresponding dynamic fusion weights to obtain a weighted feature vector. Through dynamic weighting, more discriminative feature data in different fault types can be highlighted.

[0132] S53. By using a splicing and fusion layer, the weighted first feature vector, the weighted second feature vector, and the weighted third feature vector are spliced ​​along the feature dimension to obtain a spliced ​​feature tensor.

[0133] Optionally, a splicing and fusion layer is used to splice the weighted first feature vector, weighted second feature vector, and weighted third feature vector along the feature dimensions to obtain a spliced ​​feature tensor. The core function of tensor splicing is to integrate the complementary information of the three types of weighted feature vectors into a single feature tensor, achieving preliminary feature fusion. The splicing method is along the feature dimensions, i.e., keeping the time dimension (if any) unchanged, merging the feature dimensions of the three types of feature vectors. The spliced ​​feature tensor can simultaneously contain all the key information of lightning strike transient features, insulator state features, and environmental correction features. The calculation formula for splicing and fusion is: ,in, To splice the feature tensors, The first weighted eigenvector, The weighted second eigenvector, For the weighted third eigenvector, "" indicates concatenation along the feature dimension. During implementation, it is necessary to ensure that the time dimension (if any) of the three weighted feature vectors is consistent. If there is a discrepancy in dimensions, it needs to be adjusted to be consistent through zero-padding or interpolation before concatenation. The dimension of the concatenated feature tensor is (time dimension × total feature dimension), where the total feature dimension is the sum of the feature dimensions of the three weighted feature vectors. Through concatenation and fusion, the complementary information of the three types of feature data can be fully integrated, providing a comprehensive feature foundation for subsequent feature mapping.

[0134] S54. By performing nonlinear transformation and cross-layer feature mapping on the spliced ​​feature tensor through the feature mapping layer, the fault fusion feature vector is obtained.

[0135] The concatenated feature tensor undergoes nonlinear transformation and cross-layer feature mapping through a feature mapping layer to obtain a fault fusion feature vector. Although the concatenated feature tensor integrates information from three types of feature data, it may contain redundant and invalid features, affecting the accuracy and efficiency of subsequent fault identification. Therefore, a feature mapping layer is needed to perform nonlinear transformation and cross-layer feature mapping to mine deeper features in the concatenated feature tensor, eliminate redundant features, and obtain a fault fusion feature vector with strong discriminative power. The feature mapping layer consists of multiple convolutional neural networks and cross-layer connections. The convolutional neural network can be used to perform nonlinear transformation and extract deep features. The convolution operation formula is consistent with the two-dimensional convolution operation formula in S23, capturing local and global features of the concatenated feature tensor through the convolution kernel. Cross-layer connections are used to fuse shallow and deep features, avoiding feature loss during deep feature extraction and improving the feature representation capability of the feature mapping layer. The nonlinear transformation can use the ReLU activation function, whose expression is: This is used to introduce nonlinear features and enhance the fitting ability of the feature mapping layer. The output of the feature mapping layer is a fault fusion feature vector. The dimension of this vector is lower than the total feature dimension of the concatenated feature tensor. It can retain the key complementary information of the three types of feature data, while eliminating redundant features, providing high-quality feature input for subsequent fault classification and identification.

[0136] In an optional embodiment, the fault classification and identification model includes a feature reconstruction layer, a metric space calculation layer, a radial basis function activation layer, and a decision-making layer, wherein S6 includes:

[0137] S61. The fault fusion feature vector is subjected to orthogonal basis transformation and feature decoupling through the feature reconstruction layer to obtain the orthogonal decoupled feature vector.

[0138] For example, orthogonal basis transformation and feature decoupling are performed on the fault fusion feature vector through a feature reconstruction layer to obtain orthogonally decoupled feature vectors. Feature coupling may exist in the fault fusion feature vector, meaning features of different fault types overlap, affecting the accuracy of fault identification. Therefore, orthogonal basis transformation and feature decoupling are needed through the feature reconstruction layer to decompose the fault fusion feature vector into multiple mutually orthogonal feature components, eliminating feature coupling and improving the discriminative ability of the features. Orthogonal basis transformation can be performed using orthogonal matrix transformation. By constructing an orthogonal basis matrix, the fault fusion feature vector is mapped to an orthogonal feature space. The orthogonal basis matrix is ​​constructed using the Gram-Schmidt orthogonalization method, expressed as: ,in, For the first orthogonal basis vectors The fault fusion feature vector is the first Each feature component The inner product operation ensures that the orthogonal basis matrices constructed using this method satisfy the condition of pairwise orthogonality. Feature decoupling can be achieved through orthogonal basis transformation. Multiplying the fault fusion feature vector by the orthogonal basis matrix yields the orthogonal decoupling feature vector, calculated using the following formula: ,in, These are orthogonal decoupling eigenvectors. It is an orthogonal basis matrix. For the transpose of an orthogonal basis matrix, This is the fault fusion feature vector. Through orthogonal basis transformation and feature decoupling, the feature components in the resulting orthogonal decoupled feature vector are independent of each other, which can clearly distinguish the feature differences of different fault types, providing a foundation for subsequent metric space calculations.

[0139] S62. Calculate the Mahalanobis distance between the orthogonal decoupled feature vector and the preset lightning fault category center matrix through the metric space computing layer to obtain the category relative distance vector.

[0140] Optionally, the Mahalanobis distance between the orthogonal decoupled feature vector and the preset lightning fault category center matrix is ​​calculated through a metric space computation layer to obtain the category relative distance vector; wherein, the expression for the Mahalanobis distance can be: ,in, Represents the orthogonal decoupled eigenvectors and the first The Mahalanobis distance of the center of each lightning strike fault category This represents the orthogonal decoupling eigenvector. The first element in the center matrix of lightning strike fault categories Each lightning strike fault category center vector This represents the matrix transpose operation. Indicates the first The inverse matrix of the covariance matrix of each lightning strike fault category. This represents the index of the lightning strike fault category. Mahalanobis distance is a metric for measuring the similarity between two vectors. Compared to Euclidean distance, Mahalanobis distance can eliminate the correlation and dimensionality effects between features, making it more suitable for distance calculation in fault classification and identification. The preset lightning strike fault category center matrix is ​​trained using a historical fault dataset. The rows of the matrix correspond to different lightning strike fault categories (lightning flashover fault, lightning non-flashover fault, and non-lightning strike fault), and the columns correspond to the feature dimensions. Each row vector is the center vector of the corresponding fault category, i.e., the mean vector of all feature vectors of that fault category. The calculation formula is: ,in, For the first Number of samples of this type of fault For the first Class of faults The orthogonal decoupled feature vectors of each sample. Covariance matrix of each lightning strike fault category The formula used to measure the dispersion of the feature vector of this type of fault is as follows: The inverse of the covariance matrix This is used to eliminate correlations between features. The Mahalanobis distance formula is used to calculate the Mahalanobis distance between the orthogonal decoupled feature vector and the center vector of each fault category, resulting in a category relative distance vector. Each element of the vector corresponds to the Mahalanobis distance between the orthogonal decoupled feature vector and the center of a fault category. The smaller the distance, the higher the similarity between the orthogonal decoupled feature vector and the fault category.

[0141] S63. By performing nonlinear probability mapping on the relative distance vector of categories through a radial basis function activation layer, the probability distribution vector of lightning strike failure is obtained.

[0142] Optionally, the Mahalanobis distance in the relative distance vector between categories can be used to reflect the similarity between the orthogonal decoupling feature vector and each fault category. However, the distance value cannot be directly used as the basis for fault identification. Therefore, a nonlinear probability mapping is needed through a radial basis function activation layer to convert the Mahalanobis distance into probability values ​​for each fault category, resulting in a lightning strike fault probability distribution vector. The larger the probability value, the higher the likelihood of that fault category. The radial basis function activation layer can employ multi-kernel radial basis functions, which can adapt to the feature distribution of different fault categories and improve the accuracy of the probability mapping.

[0143] S64. The lightning fault identification result is obtained by performing maximum probability matching and category index mapping on the probability distribution vector of lightning faults through the decision-making layer.

[0144] Optionally, the core function of the decision-making layer is to determine the final lightning fault type based on the probability values ​​in the lightning fault probability distribution vector. Specifically, firstly, maximum probability matching is performed to find the element with the highest probability value in the lightning fault probability distribution vector; the fault category corresponding to this element is the initially identified fault type. If two or more fault categories have the same probability value, and both are the maximum value, further determination can be made by combining the key feature components of the fault fusion feature vector, prioritizing the fault category with higher similarity to the orthogonal decoupled feature vector to avoid ambiguity. Then, category index mapping is performed to convert the initially identified fault category index into a specific fault type description (such as lightning flashover fault, lightning non-flashover fault, or non-lightning fault), obtaining the final lightning fault identification result. The formula for calculating maximum probability matching can be: ,in The fault category index corresponding to the element with the highest probability value. For the first The probability value of each fault category. Total number of fault categories (e.g.) (Corresponding to three types of faults). The category index mapping is implemented through a preset category index table, which stores a one-to-one correspondence between fault category indexes and specific fault type descriptions. The preset index table is as follows: Index 1 corresponds to lightning flashover faults, Index 2 corresponds to lightning-not-flashover faults, and Index 3 corresponds to non-lightning-flashover faults. Simultaneously, the decision-making layer can output the probability value of each fault category as a reliability reference for the identification results. When the maximum probability value is lower than a preset confidence threshold (e.g., 80%), an identification warning can be issued, prompting further confirmation of the fault type through manual review to ensure the reliability and usability of the fault identification results.

[0145] In an optional embodiment, S63 includes:

[0146] S631. Perform multi-scale mapping on the relative distance vector of categories based on multi-kernel radial basis functions to obtain the mapping value, and obtain the multi-kernel similarity matrix based on the mapping value.

[0147] Optionally, the relative distance vectors of categories are subjected to multi-scale mapping based on multi-kernel radial basis functions to obtain mapping values, and a multi-kernel similarity matrix is ​​obtained based on the mapping values; wherein, the expression for the mapping values ​​can be: ,in, For the first The categories in the Mapping values ​​on the radial basis function kernel For the first Mahalanobis distance for each category For the first The bandwidth parameters of each kernel. Multi-kernel radial basis functions (RBFs) perform multi-scale mapping on the relative distance vectors of categories by setting multiple radial basis function kernels with different bandwidth parameters. RBFs can capture feature similarity at different scales, improving the comprehensiveness of the mapping. The core principle of the radial basis function is to convert Mahalanobis distance into a similarity mapping value using an exponential function. The smaller the Mahalanobis distance, the larger the mapping value, indicating higher similarity. The mapping value ranges from (0,1). Specifically, multiple bandwidth parameters are preset. Each bandwidth parameter corresponds to a radial basis function kernel, which is applied to each Mahalanobis distance in the class-relative distance vector. Substituting the values ​​into the mapping formula of each radial basis function kernel, the mapping values ​​of each category on different kernels are calculated. All mapping values ​​are arranged in rows (rows correspond to fault categories, columns correspond to kernels) to obtain a multi-kernel similarity matrix. The element values ​​of the matrix are the mapping values ​​of the corresponding categories on the corresponding kernels, which can comprehensively reflect the multi-scale similarity between the orthogonal decoupled feature vectors and each fault category.

[0148] S632. Perform feature weighting on the multi-kernel similarity matrix to obtain the weighted similarity feature vector.

[0149] Optionally, the multi-kernel similarity matrix is ​​weighted to obtain a weighted similarity feature vector. Different kernels in the multi-kernel similarity matrix contribute differently to fault identification; therefore, feature weighting is necessary to highlight the mapping values ​​of kernels with higher contributions and suppress those with lower contributions, thereby improving the discriminative power of the similarity features. The weighting process employs an attention weighting mechanism. First, the attention weight of each kernel is calculated. The attention weight is determined based on the variance of the mapping values ​​of each kernel. The larger the variance, the higher the discriminative power of the mapping values ​​of that kernel, and the larger the attention weight. The formula for calculating the attention weight is: ,in For the first Attention weights for each core, For the number of cores, For the first A vector of mapping values ​​for each kernel. This is the variance calculation function. Then, each column of the multi-kernel similarity matrix (corresponding to the mapping value of one kernel) is multiplied by the corresponding attention weight to obtain the weighted mapping value. The calculation formula is: ,in For the first The categories in the The weighted mapping values ​​on each kernel are summed. Finally, all weighted mapping values ​​for each category are summed to obtain the weighted similarity value for each category. The weighted similarity values ​​of all categories are arranged in order to obtain the weighted similarity feature vector, where each element of the vector corresponds to the weighted similarity value of a fault category.

[0150] S633. Normalize the weighted similarity feature vector to obtain the probability distribution vector of lightning strike failure.

[0151] Optionally, the elements in the weighted similarity feature vector are weighted similarity values ​​for each fault category. Since these values ​​have inconsistent ranges and cannot be directly used as probability values, normalization is required. This normalization maps the weighted similarity values ​​to the [0,1] interval, ensuring the sum of all element values ​​is 1, thus obtaining the lightning strike fault probability distribution vector. The normalization process uses the softmax normalization method, calculated using the following formula: ,in, For the first The probability value of each fault category. The first element in the weighted similarity feature vector Each element value The formula represents the number of fault categories. It converts the weighted similarity value into a probability value; a higher probability value indicates a greater likelihood that the orthogonal decoupling feature vector belongs to that fault category. The sum of the elements in the normalized lightning strike fault probability distribution vector is 1, with each element corresponding to the probability of a fault category. This clearly reflects the likelihood of each fault category, providing a basis for subsequent decision-making.

[0152] In the aforementioned intelligent identification method for lightning strike faults in high-voltage transmission lines, four types of multi-source heterogeneous data—lightning current waveform, line traveling wave, insulator leakage current, and meteorological environment—are acquired simultaneously. First, empirical mode decomposition and continuous wavelet transform are performed on the lightning current and traveling wave data to extract joint time-frequency domain features, thereby fully exploring the deep waveform information in the transient process of lightning strikes. Simultaneously, variational mode decomposition is performed on the leakage current to obtain insulator state features, and mutual information analysis and attention encoding are performed on the meteorological data to obtain environmental correction factors. Then, the three types of features are input into a hierarchical fusion identification network for dynamic weighted fusion. Finally, a classification model based on Mahalanobis distance and multi-kernel radial basis functions is used to output the fault type. This method effectively solves the problem of low accuracy in traditional single-data source identification, enabling fine differentiation between lightning flashover, lightning-non-flashover, and non-lightning faults, and improving the robustness of identification under complex operating conditions. By integrating insulator status and meteorological correction information, it effectively reduces the negative impact of environmental interference and changes in insulator surface conditions on the identification results. At the same time, it reduces the subjectivity of manual threshold judgment and can adaptively extract nonlinear time-frequency features, providing reliable technical support for lightning fault diagnosis of high-voltage transmission lines.

[0153] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0154] Based on the same inventive concept, this application also provides a high-voltage transmission line lightning fault intelligent identification system for implementing the above-described intelligent identification method for high-voltage transmission line lightning faults. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the high-voltage transmission line lightning fault intelligent identification system provided below can be found in the limitations of the high-voltage transmission line lightning fault intelligent identification method described above, and will not be repeated here.

[0155] In one exemplary embodiment, such as Figure 2 As shown, a schematic diagram of a high-voltage transmission line lightning strike fault intelligent identification system 10 is provided, including:

[0156] The multi-source data acquisition module 11 is used to acquire lightning current waveform data, line traveling wave data, insulator leakage current data and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line.

[0157] The lightning transient feature extraction module 12 is used to perform joint time-frequency domain feature extraction on lightning current waveform data and line traveling wave data to obtain lightning transient feature data.

[0158] The insulator state feature extraction module 13 is used to perform waveform decomposition and modal analysis on the insulator leakage current data to obtain insulator state feature data.

[0159] The environmental correction factor analysis module 14 is used to perform multi-factor correlation analysis on meteorological and environmental data to obtain environmental correction factor data.

[0160] The hierarchical feature fusion module 15 is used to input lightning transient feature data, insulator state feature data and environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain fault fusion feature vector;

[0161] The fault classification and identification module 16 is used to input the fault fusion feature vector into the fault classification and identification model to identify the type of lightning strike fault and obtain the lightning strike fault identification result.

[0162] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the intelligent identification method for lightning strike faults in high-voltage transmission lines as described above.

[0163] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0164] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0165] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent identification of lightning strike faults in high-voltage transmission lines, characterized in that, The method includes: S1. Acquire lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line. S2. Perform time-frequency domain joint feature extraction on the lightning current waveform data and the line traveling wave data to obtain lightning transient feature data; S3. Perform waveform decomposition and modal analysis on the insulator leakage current data to obtain insulator state characteristic data; S4. Perform multi-factor correlation analysis on the meteorological and environmental data to obtain environmental correction factor data; S5. Input the lightning transient feature data, the insulator state feature data and the environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain the fault fusion feature vector; S6. Input the fault fusion feature vector into the fault classification and recognition model to identify the type of lightning strike fault and obtain the lightning strike fault recognition result.

2. The method according to claim 1, characterized in that, S2 includes: S21. Perform empirical mode decomposition on the lightning current waveform data to obtain the intrinsic mode function component data of the lightning current. S22. Perform continuous wavelet transform on the traveling wave data of the line to obtain the traveling wave time spectrum data; S23. Input the lightning current intrinsic mode function component data and the traveling wave time spectrum data into a dual-channel feature extraction network to extract time-domain transient features and frequency-domain energy features, thereby obtaining lightning current time-domain feature vector and traveling wave frequency-domain feature vector; wherein, the dual-channel feature extraction network includes a first channel and a second channel; the first channel includes a one-dimensional convolutional layer and a max-pooling layer, and the first channel is used to extract the waveform steepness and peak attenuation rate of the lightning current intrinsic mode function components; the second channel includes a two-dimensional convolutional layer and a global average pooling layer, and the second channel is used to extract the frequency band centroid and energy concentration of the traveling wave time spectrum; S24. Perform cross-attention fusion on the lightning current time-domain feature vector and the traveling wave frequency-domain feature vector to obtain the lightning transient feature data.

3. The method according to claim 2, characterized in that, S3 includes: S31. Perform sliding window truncation and adaptive filtering on the insulator leakage current data to obtain pure leakage current fragment data; S32. Based on the constrained variational objective function, perform variational mode decomposition on the pure leakage current segment data to obtain multiple intrinsic mode function component data and residual component data; wherein, the expression of the constrained variational objective function is: in, For time The first time Each intrinsic mode function component For the first The center frequency of each component To preset the number of decomposition layers, For the Dirac function, This represents the convolution operation. Indicates time The partial derivatives, The imaginary unit, For time variables, It is an L2 norm; S33. Calculate the peak amplitude, root mean square value and frequency band energy ratio of each intrinsic mode function component data and the residual component data to obtain the multidimensional feature vector of leakage current. S34. The leakage current multidimensional feature vector is input into the insulator state coding network for feature compression to obtain the insulator state feature data. The insulator state coding network includes an insulation coding input layer, a fully connected coding layer, and an output layer. The insulation coding input layer is used to receive the leakage current multidimensional feature vector, the fully connected coding layer is used to compress the feature dimension of the leakage current multidimensional feature vector, and the output layer is used to output the insulator state feature data.

4. The method according to claim 1, characterized in that, S4 includes: S41. Impute missing values ​​and standardize the meteorological environment data to obtain a standardized meteorological data matrix; the meteorological environment data includes meteorological parameters. S42. Combining the lightning strike fault label and the standardized meteorological data matrix, calculate the mutual information coefficient between each meteorological parameter and the lightning strike fault label to obtain the initial association weight vector; wherein, the expression for the mutual information coefficient is: in, For the first The mutual information coefficient between meteorological parameters and lightning fault labels, For the first The meteorological parameters mentioned above The historical lightning strike fault tag has a value of 0 indicating a non-lightning strike fault and a value of 1 indicating a lightning strike fault. Indicates the value of the meteorological parameter And the lightning strike fault label is The probability, Indicates the value of the meteorological parameter The probability, The value of the lightning strike fault label is indicated. The probability of; S43. The standardized meteorological data matrix and the initial correlation weight vector are input into a preset multi-factor attention coding network for nonlinear feature weighting to obtain a meteorological impact feature vector; wherein, the multi-factor attention coding network includes an attention coding input layer, a multi-head self-attention layer, a layer normalization layer, and a feedforward network layer, the attention coding input layer is used to receive the standardized meteorological data matrix and the initial correlation weight vector, the multi-head self-attention layer is used to calculate the attention score between each meteorological parameter according to the initial correlation weight vector, the layer normalization layer is used to normalize the attention score, and the feedforward network layer is used to output the meteorological impact feature vector; S44. Input the meteorological impact feature vector into a spatial mapping network for linear transformation to obtain the environmental correction factor data.

5. The method according to claim 1, characterized in that, The hierarchical fusion recognition network includes a normalization layer, a weight calculation layer, a splicing and fusion layer, and a feature mapping layer. S5 includes: S51. The lightning transient characteristic data, the insulator state characteristic data and the environmental correction factor data are respectively normalized by the normalization layer to obtain the first normalized feature vector, the second normalized feature vector and the third normalized feature vector. S52. Calculate the dynamic fusion weights of the first normalized feature vector, the second normalized feature vector, and the third normalized feature vector through the weight calculation layer to obtain the weighted first feature vector, the weighted second feature vector, and the weighted third feature vector. S53. The weighted first feature vector, the weighted second feature vector, and the weighted third feature vector are spliced ​​along the feature dimension by a splicing and fusion layer to obtain a spliced ​​feature tensor; S54. The fault fusion feature vector is obtained by performing nonlinear transformation and cross-layer feature mapping on the spliced ​​feature tensor through the feature mapping layer.

6. The method according to claim 1, characterized in that, The fault classification and identification model includes a feature reconstruction layer, a metric space calculation layer, a radial basis function activation layer, and a decision-making layer. S6 includes: S61. The fault fusion feature vector is subjected to orthogonal basis transformation and feature decoupling through the feature reconstruction layer to obtain orthogonal decoupled feature vector; S62. The Mahalanobis distance between the orthogonal decoupled feature vector and the preset lightning fault category center matrix is ​​calculated through the metric space calculation layer to obtain the category relative distance vector; wherein, the expression for the Mahalanobis distance is: in, Represents the orthogonal decoupling eigenvector and the first The Mahalanobis distance of the center of each lightning strike fault category This represents the orthogonal decoupling feature vector. Represents the first in the center matrix of the lightning strike fault categories Each lightning strike fault category center vector This represents the matrix transpose operation. Indicates the first The inverse matrix of the covariance matrix of each lightning strike fault category. The serial number indicating the type of lightning strike fault; S63. The relative distance vector of the categories is nonlinearly probabilistically mapped through the radial basis function activation layer to obtain the lightning strike failure probability distribution vector. S64. The lightning fault identification result is obtained by performing maximum probability matching and category index mapping on the lightning fault probability distribution vector through the decision-making layer.

7. The method according to claim 6, characterized in that, S63 includes: S631. Perform multi-scale mapping based on multi-kernel radial basis functions on the relative distance vector of the categories to obtain mapping values, and obtain a multi-kernel similarity matrix based on the mapping values; wherein, the expression of the mapping values ​​is: in, For the first The categories in the Mapping values ​​on the radial basis function kernel For the first Mahalanobis distance for each category For the first Bandwidth parameters per core; S632. Perform feature weighting on the multi-kernel similarity matrix to obtain a weighted similarity feature vector; S633. Normalize the weighted similarity feature vector to obtain the lightning strike failure probability distribution vector.

8. A smart identification system for lightning strike faults in high-voltage transmission lines, characterized in that, The system includes: The multi-source data acquisition module is used to acquire lightning current waveform data, line traveling wave data, insulator leakage current data, and meteorological environment data after a lightning strike fault occurs on a high-voltage transmission line. The lightning strike transient feature extraction module is used to perform time-frequency domain joint feature extraction on the lightning current waveform data and the line traveling wave data to obtain lightning strike transient feature data. The insulator state feature extraction module is used to perform waveform decomposition and modal analysis on the insulator leakage current data to obtain insulator state feature data. The environmental correction factor analysis module is used to perform multi-factor correlation analysis on the meteorological and environmental data to obtain environmental correction factor data. The hierarchical feature fusion module is used to input the lightning transient feature data, the insulator state feature data and the environmental correction factor data into the hierarchical fusion identification network for feature fusion to obtain the fault fusion feature vector. The fault classification and identification module is used to input the fault fusion feature vector into the fault classification and identification model to identify the type of lightning strike fault and obtain the lightning strike fault identification result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images