A GIS pulse group fault classification method based on weighted feature attention

Abstract

The application discloses a GIS pulse group fault classification method based on weighted feature attention, comprising the following steps: collecting original time signals of GIS monitoring objects and detecting candidate pulses, calculating the arrival time, peak amplitude, energy and power frequency phase of each pulse; using a sliding time window to construct the candidate pulses into a pulse group to be classified, extracting the dimension correlation features of the pulse group to be classified, including time, amplitude, energy and phase, and constructing a five-dimensional correlation feature vector; constructing a feature-level attention dynamic weighting multi-classification neural network, inputting the five-dimensional correlation feature vector into the neural network for dynamic weighting and score calculation, comprehensively detecting the classification probability in a complete cycle, and finally obtaining the overall fault classification result and classification reliability. The application improves the accuracy, anti-interference ability and engineering application reliability of GIS equipment fault diagnosis in a complex field environment.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault identification and classification technology, specifically to a GIS burst fault classification method and computer equipment based on weighted feature attention. Background Technology

[0002] In actual industrial environments, the raw sampled signals often contain valid fault pulses, complex noise pulses, and various interference pulses simultaneously.

[0003] Existing GIS fault pulse classification methods have the following limitations: First, traditional methods often rely on a single feature threshold or analyze the instantaneous amplitude of a single pulse in isolation, lacking consideration of the overall intensity and evolution of pulse groups within a local time range. Secondly, existing technologies often fail to fully integrate multi-dimensional correlation features such as the time interval between adjacent pulses, the correlation between amplitude and energy, and the power frequency phase distribution when extracting features. Finally, when faced with multi-feature fusion, traditional classification models usually assign fixed weights to each feature, and cannot dynamically weight feature-level attention based on the differences between different pulse group samples. As a result, in complex interference environments, the accuracy of fault type identification and the reliability of classification are difficult to meet the high reliability requirements of actual engineering.

[0004] To address these technical problems, this application proposes a GIS burst fault classification method based on weighted feature attention. Summary of the Invention

[0005] The main objective of this invention is to provide a GIS pulse group fault classification method based on weighted feature attention, which improves the accuracy, anti-interference capability, and engineering application reliability of GIS equipment fault diagnosis in complex field environments, thereby solving the technical problems mentioned in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: A GIS burst fault classification method based on weighted feature attention, comprising the following steps performed by a computer device: Step S1. Collect the raw time signal of the GIS monitoring object and detect candidate pulses, and calculate the arrival time, peak amplitude, energy and power frequency phase of each pulse; Step S2. Using a sliding time window, candidate pulses are constructed into a pulse group to be classified, and the time, amplitude and energy, phase and other dimensional correlation features of the pulse group to be classified are extracted to construct a five-dimensional correlation feature vector. Step S3. Construct a feature-level attention-based dynamically weighted multi-class neural network, and input the five-dimensional associated feature vector into the neural network for dynamic weighting and scoring calculation. By combining the classification probabilities within the complete detection period, the overall fault classification result and classification credibility are finally obtained.

[0007] Preferably, the specific operation procedure for calculating the arrival time, peak amplitude, energy, and power frequency phase of each pulse in step S1 includes: Step S11. Collect raw continuous time signals of GIS monitored objects within the time interval. ,in, , Indicates time The corresponding original signal amplitude, Represents a continuous-time variable. Indicates the total sampling time; Step S12. According to the sampling frequency The original continuous-time signal is discretized to obtain a discrete signal. ,in, , Indicates the first The signal amplitude corresponding to each sampling point Indicates the sampling point number. Let N represent the sampling frequency, and N represent the total number of sampling points. ; Step S13. For discrete signals After performing basic filtering or denoising, the signal is searched for pulse components with local abrupt changes or local peaks that are significantly higher than the background fluctuations, and pulses that meet the preset detection conditions are recorded as candidate pulses. Step S14. Calculate the first one respectively. Arrival time of candidate pulses Peak amplitude ,energy and peak amplitude Finally, the number was obtained. The parameter set corresponding to each candidate pulse is represented as follows: . Preferably, in step S14, for the first... The candidate pulses represent transient pulse signals initially detected in the discrete signal that may correspond to a certain type of fault discharge event. Since the original sampled signal may simultaneously contain valid fault pulses, noise pulses, and interference pulses, the first pulse detected is a candidate pulse, not the final confirmed fault pulse. The calculation process includes: If the detection yields the first The peak positions of the candidate pulses are ,but Indicates the first The peak sampling point positions corresponding to each candidate pulse Indicates the candidate pulse number.

[0008] (1) By the first Peak position of candidate pulses Calculate the arrival time of the candidate pulse. The calculation formula is as follows:

[0009] in, Indicates the first The arrival time of each candidate pulse.

[0010] From the calculated peak position Then calculate the first Peak amplitude of each candidate pulse The calculation formula is:

[0011] in, Indicates the first The peak amplitude of each candidate pulse, peak amplitude Indicates the first The signal amplitude corresponding to the peak position of each candidate pulse is used to characterize the instantaneous intensity of the candidate pulse; Around the first Peak position of candidate pulses Extract a local pulse segment and calculate the first pulse. Energy of each candidate pulse The calculation formula is as follows;

[0012] in, Indicates the first The energy of each candidate pulse This represents the local offset index relative to the peak position. This indicates the number of sampling points intercepted before the peak point. This indicates the number of sampling points truncated after the peak point; The pulse energy It is obtained by summing the squares of the local waveform amplitudes near the peak value, and is used to characterize the overall intensity of the candidate pulse over a local time range, rather than just reflecting the instantaneous amplitude at a single sampling point; Finally, based on the arrival time of the candidate pulses Calculate the first The power frequency phase corresponding to each candidate pulse The calculation formula is as follows:

[0013] in, Indicates the first The power frequency phase corresponding to each candidate pulse. Indicates the fundamental frequency of the power frequency. Indicates to Round down; Therefore, the first Each candidate pulse is represented as

[0014] in, Indicates the first The parameter set corresponding to each candidate pulse.

[0015] Preferably, the specific operation process of step S2, which uses a sliding time window to construct the candidate pulses into a pulse group to be classified, includes: Construct a set of sliding time windows, with the time length of each analysis window being... The sliding time window has a first... The time window is ,in, Indicates the first A time window, Indicates the first The start time of a time window Indicates the time window number; The arrival times of each candidate pulse calculated in step S1 and pulse parameter group The conditions will be met. All candidate pulses constitute the first There are several pulse groups to be classified. ,in, Indicates the first One pulse group to be classified.

[0016] Let the first The number of pulses in the pulse groups to be classified is The minimum number of pulses required to perform population statistical analysis is [threshold value]. ,when At that time, the pulse group to be classified Perform subsequent feature extraction, when At that time, the first Each time window is marked as an invalid window.

[0017] Preferably, the specific operation process for extracting the temporal dimension correlation features of the pulse group to be classified in step S2 includes: From the already constructed first A pulse group to be classified Extract the arrival time of each candidate pulse in the pulse group. And arrange them in ascending order; Calculate the time interval between adjacent pulses from the arrival time sequence of each candidate pulse. The calculation formula is as follows:

[0018] in, Indicates the first The candidate pulse and the first The time interval between candidate pulses; Based on the time interval of each candidate pulse Calculate the first Mean time interval of each pulse group to be classified The calculation formula is as follows:

[0019] in, Indicates the first The average time interval between adjacent pulses within a pulse group to be classified.

[0020] By time interval and the mean of time intervals Calculate the standard deviation of the time interval for the k-th pulse group to be classified. The calculation formula is as follows:

[0021] in, It represents the standard deviation of the time interval between adjacent pulses within the k-th pulse group to be classified.

[0022] Ultimately, it is determined by the average of the time intervals. and the standard deviation of the time interval Calculate the time discrepancy coefficients of the kth pulse group to be classified. The calculation formula is as follows:

[0023] in, ε represents the time discretization coefficient of the k-th pulse group to be classified, and ε represents a small normal number to prevent the denominator from being zero.

[0024] Preferably, the specific operation procedure for extracting the amplitude and energy dimension correlation features of the pulse group to be classified in step S2 includes: For the kth pulse group to be classified that has been constructed The peak amplitude of each candidate pulse Calculate the mean amplitude of the k-th pulse group to be classified. The calculation formula is as follows:

[0025] in, This represents the average peak amplitude of candidate pulses within the k-th pulse group to be classified; The peak amplitude of each candidate pulse and the mean amplitude Calculate the standard deviation of the amplitude of the k-th pulse group to be classified. The calculation formula is as follows:

[0026] in, It represents the standard deviation of the peak amplitude of candidate pulses within the k-th pulse group to be classified.

[0027] From the mean amplitude and amplitude standard deviation Calculate the amplitude dispersion coefficient of the k-th pulse group to be classified. The calculation formula is as follows:

[0028] in, This represents the amplitude dispersion coefficient of the k-th pulse group to be classified.

[0029] Energy of each candidate pulse Calculate the average energy of the k-th pulse group to be classified. The calculation formula is as follows:

[0030] in, This represents the average energy of candidate pulses within the k-th pulse group to be classified.

[0031] Ultimately, the peak amplitude of each candidate pulse is used to determine the outcome. Energy of each candidate pulse Amplitude mean and energy mean Calculate the amplitude-energy correlation coefficient of the k-th pulse group to be classified. The calculation formula is as follows:

[0032] in, denoted as the Pearson correlation coefficient between the peak amplitude and pulse energy in the k-th pulse group to be classified.

[0033] Preferably, the specific operation process for extracting the phase dimension correlation features of the pulse group to be classified in step S2 includes: For the kth pulse group to be classified that has been constructed Extract the power frequency phase corresponding to each candidate pulse in the pulse group to form a phase sequence. ,in, This represents the phase sequence corresponding to the k-th pulse group to be classified; The power frequency phase of each candidate pulse Calculate the phase clustering degree of the k-th pulse group to be classified. The calculation formula is as follows:

[0034] in, denoted by , where j represents the phase clustering degree of the k-th pulse group to be classified.

[0035] power frequency phase interval Divide into B equal-width subintervals, and denote the b-th phase subinterval as . The calculation formula is as follows:

[0036] Where B represents the total number of sub-intervals into which the phase interval is divided. This represents the b-th phase sub-interval, where b represents the phase sub-interval index. The power frequency phase of each candidate pulse The statistics fall within the b-th phase sub-interval Number of pulses The formula is as follows:

[0037] in, This represents the number of pulses in the b-th phase sub-interval. This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise. Number of pulses in phase sub-interval and total number of pulses Calculate the probability corresponding to the b-th phase sub-interval. The calculation formula is as follows:

[0038] in, Let represent the probability value corresponding to the b-th phase sub-interval, and satisfy the condition. ; From the probability distribution Calculate the phase distribution entropy of the k-th pulse group to be classified. The calculation formula is as follows:

[0039] in, Represents the phase distribution entropy of the k-th pulse group to be classified; when When =0, the corresponding term takes the value of 0 according to the limit.

[0040] Preferably, in step S2, the calculated time discrepancy coefficients are used in the process of constructing the five-dimensional correlation feature vector. Calculated amplitude dispersion coefficients And amplitude-energy correlation coefficient and the calculated phase convergence and phase distribution entropy The five-dimensional correlation feature vector constitutes the k-th pulse group to be classified. :

[0041] in, This represents the five-dimensional correlation feature vector corresponding to the k-th pulse group to be classified. This indicates the transpose operation.

[0042] Preferably, the specific calculation process for the dynamic weighting and scoring calculation of the neural network in step S3 includes: Step S31. After normalizing the five-dimensional correlation feature vector, input it into the neural network. Specifically, input the five-dimensional correlation feature vector into the compression mapping layer to obtain the hidden layer vector. ; Step S32. Based on hidden layer vectors The five-dimensional attention score vector is calculated through the extended mapping layer. And construct the corresponding dynamic weight matrix. Used to obtain reconstructed feature vectors Finally, the reconstructed feature vector The fault type score vector corresponding to the k-th pulse group to be classified is obtained through multi-classification output layer calculation. ; Step S33. From the fault type score vector ,through The function calculates the probability vector of the fault type corresponding to the k-th pulse group to be classified. ; Step S34. Train the network using a multi-class cross-entropy loss function, and during the training phase, let the true fault category label of the k-th pulse group sample to be classified be . ,in, , This represents the total number of fault categories to be classified, and the parameter set is also optimized during training using the backpropagation algorithm; Step S35. Based on the fault type probability vector Perform fault type classification on the k-th pulse group to be classified, and take... The category with the highest probability is used as the predicted fault type label for that pulse group. At the same time, a mapping relationship between the fault category index and the fault type name is also established. The output result for the fault type corresponding to the k-th pulse group to be classified is: ; Step S36. Finally, based on the total number of pulse groups to be classified during the complete detection cycle and the predicted fault type label for each pulse group to be classified, The frequency of occurrence of each fault type, and the maximum value of each pulse group to be classified. probability Calculate the overall classification confidence level and compare it with the overall fault type. Output together.

[0043] Preferably, before inputting the five-dimensional correlation feature vector into the neural network in step S3, the constructed five-dimensional correlation feature vector is further processed. Normalization is performed to obtain the normalized feature vector. :

[0044] in, This represents the normalized five-dimensional correlation feature vector of the k-th pulse group to be classified. to These represent the five normalized feature components. This represents a five-dimensional real vector space.

[0045] Preferably, in step S31, the normalized five-dimensional correlation feature vector is input into the compression mapping layer to obtain the hidden layer vector. The process involves the following formula:

[0046] in, This represents the hidden layer vector corresponding to the k-th pulse group to be classified. This represents the weight matrix of the compression mapping layer. Let represent the bias vector of the compression mapping layer, d represent the hidden layer dimension, and satisfy d < 5. This represents a non-linear activation function.

[0047] Preferably, the specific operation process of step S32 includes: From hidden layer vectors The five-dimensional attention score vector corresponding to the k-th pulse group to be classified is calculated through the extended mapping layer. for:

[0048] in, This represents the weight matrix of the extended mapping layer. This represents the bias vector of the extended mapping layer. express Activation function to Let each represent the attention score corresponding to one of the five feature components, and satisfy the following conditions: ; From attention score vector Construct the dynamic weight matrix corresponding to the k-th pulse group to be classified. ,have: , This represents the diagonal matrix construction operator.

[0049] From the dynamic weight matrix and the normalized five-dimensional correlation feature vector Calculate the reconstructed feature vector of the k-th pulse group to be classified after dynamic weighting. The calculation formula is as follows: ,Right now

[0050] Reconstructing the feature vector The fault type score vector corresponding to the k-th pulse group to be classified is obtained through multi-classification output layer calculation. :

[0051] in, This represents the fault type score vector corresponding to the k-th pulse group to be classified. This represents the weight matrix of the multi-class output layer. This represents the bias vector of the multi-class output layer. This indicates that the k-th pulse group to be classified belongs to the k-th pulse group. The original score for the type of fault, Indicates the fault category index, .

[0052] Preferably, in step S33, the fault type score vector ,through The function calculates the probability vector of the fault type. The specific calculation formula is as follows:

[0053]

[0054] in, This represents the probability that the k-th pulse group to be classified belongs to the c-th type of fault. Let represent the probability vector of the fault type corresponding to the k-th pulse group to be classified, and satisfy the following condition: .

[0055] Preferably, in step S34, the training function during training using the multi-class cross-entropy loss function is expressed as follows:

[0056] Where L represents the loss function value, and K represents the total number of pulse group samples to be classified participating in the training. This indicates that the k-th sample belongs to the true class. The predicted probability.

[0057] Preferably, the specific data processing flow of step S35 includes: Pick The category with the highest probability is used as the predicted fault type label for that pulse group. ,have , representing the predicted fault type label for the k-th pulse group to be classified; At this point, there exists a maximum value for the k-th pulse group to be classified in predicting the fault type. The probability is Its formula is: ; Therefore, when it is necessary to output the fault type name, a mapping relationship between the fault category index and the fault type name should be established. The output result for the fault type corresponding to the kth pulse group to be classified is: ,in, This indicates the fault type name corresponding to the c-th type of fault.

[0058] Preferably, the specific operation process for calculating the overall classification confidence level in step S36 includes: The total number of pulse groups involved in the classification process within the complete detection cycle. Fault type labels are predicted from each pulse group to be classified. Count the occurrence frequency of each fault type to obtain the number of pulse groups to be classified as fault type c. The calculation formula is as follows:

[0059] The most frequently occurring fault category is used as the overall fault classification result, using the following formula: ,in This label represents the overall fault type corresponding to the complete testing cycle; The maximum value of each pulse group to be classified probability The overall classification confidence level is calculated using the following formula:

[0060] in, This indicates the overall reliability of the fault classification.

[0061] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0062] As can be seen from the above technical solution, the present invention provides a GIS burst fault classification method based on weighted feature attention. Compared with the prior art, the present invention has the following advantages: 1. This invention constructs a pulse group to be classified within a sliding time window, extracts multi-dimensional correlation features such as time, amplitude, energy, and power frequency phase, comprehensively characterizes the statistical laws of fault discharge, breaks through the limitations of traditional single-feature, single-point pulse analysis methods, and abandons the one-sidedness of relying solely on instantaneous pulse parameters to determine faults. It can accurately capture the differences in the temporal distribution, energy characteristics, and phase distribution of pulse groups corresponding to different GIS fault types, and achieve refined feature characterization of different fault types such as partial discharge, poor contact, and insulation defects, providing a comprehensive and reliable feature basis for subsequent accurate classification.

[0063] 2. This invention introduces a feature-level attention dynamic weighting mechanism to adaptively reconstruct feature weights for different samples, effectively amplifying key identification features and suppressing redundant noise. This weakens the interference of invalid features caused by environmental noise and detection errors, strengthens the core differentiated features of various faults, and significantly improves the model's ability to identify and distinguish weak and similar faults. Ultimately, it solves the problem that traditional fixed-weight feature extraction methods cannot adapt to complex on-site interference scenarios.

[0064] 3. This invention reduces the random errors of single-point instantaneous sampling by combining local waveform energy assessment, and outputs the overall fault classification results and classification credibility by comprehensively analyzing data within the complete detection cycle. This effectively avoids misjudgment and missed judgment caused by instantaneous sampling fluctuations and sudden environmental interference. By optimizing the classification logic through full-cycle data probabilistic fusion, and outputting classification credibility quantification indicators, the fault diagnosis results are more intuitive and reliable, thereby significantly improving the accuracy, anti-interference ability and engineering application reliability of GIS equipment fault diagnosis in complex field environments.

[0065] It should be understood that the descriptions in this section are not intended to identify key or essential features of embodiments of the invention, nor are they intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Of course, implementing any product of the invention does not necessarily require achieving all of the advantages described above simultaneously. Attached Figure Description

[0066] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 A schematic diagram illustrating the construction of the pulse group to be classified using the sliding time window of this invention; Figure 3 This is a schematic diagram of the feature-level attention dynamic weighted multi-class neural network of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] For details in the embodiments, please refer to Figures 1 to 3 .

[0069] like Figure 1 As shown in the figure, the GIS burst fault classification method based on weighted feature attention proposed in this embodiment of the invention includes the following steps: For details in the embodiments, please refer to Figures 1 to 3 .

[0070] like Figure 1 As shown in the figure, the GIS burst fault classification method based on weighted feature attention proposed in this embodiment of the invention includes the following steps: Step L1: Acquire raw GIS signals and detect candidate pulses.

[0071] Step L11: Collect the raw continuous time signal of the GIS monitored object within the time interval, denoted as:

[0072] in, Indicates time The corresponding original signal amplitude, Represents a continuous-time variable. This indicates the total sampling time.

[0073] According to sampling frequency Discretize the original continuous-time signal to obtain a discrete signal:

[0074] in, Indicates the first The signal amplitude corresponding to each sampling point Indicates the sampling point number. Let N represent the sampling frequency, and N represent the total number of sampling points. .

[0075] Step L12: For discrete signals After basic filtering or denoising, the signal is searched for pulse components with local abrupt changes or local peaks significantly higher than background fluctuations. Pulses that meet preset detection conditions are recorded as candidate pulses. Candidate pulses represent transient pulse signals that are initially detected in the discrete signal and may correspond to a certain type of fault discharge event. Since the original sampled signal may simultaneously contain valid fault pulses, noise pulses, and interference pulses, the candidate pulses are detected first in step L, rather than the finally confirmed fault pulses.

[0076] If the detection yields the first The peak positions of the candidate pulses are ,but Indicates the first The peak sampling point positions corresponding to each candidate pulse Indicates the candidate pulse number.

[0077] By the Peak position of candidate pulses Calculate the arrival time of the candidate pulse. :

[0078] in, Indicates the first The arrival time of each candidate pulse.

[0079] From the calculated peak position Then calculate the peak amplitude. :

[0080] in, Specifically represented as the first The signal amplitude corresponding to the peak position of each candidate pulse is used to characterize the instantaneous intensity of the candidate pulse.

[0081] Step L13: Around the peak position Extract a local pulse segment and calculate the first pulse. Energy of each candidate pulse :

[0082] in, This represents the local offset index relative to the peak position. This indicates the number of sampling points intercepted before the peak point. This indicates the number of sampling points extracted after the peak point.

[0083] Pulse energy It is obtained by summing the squares of the local waveform amplitudes near the peak value, and is used to characterize the overall intensity of the candidate pulse over a local time range, rather than just reflecting the instantaneous amplitude at a single sampling point.

[0084] Step L14: Arrival time calculated in step L1 Then calculate the power frequency phase. :

[0085] in, Indicates the first The power frequency phase corresponding to each candidate pulse. Indicates the fundamental frequency of the power frequency. Indicates to Round down to the nearest integer.

[0086] Therefore, the first Each candidate pulse is represented as:

[0087] in, Indicates the first The parameter set corresponding to each candidate pulse.

[0088] Step L2: Construct the pulse group to be classified using a sliding time window.

[0089] Let the length of the sliding time window be... ,in, This indicates the duration of each analysis window.

[0090] Let the first The time window is:

[0091] in, Indicates the first The start time of a time window Indicates the time window number.

[0092] Arrival times of each candidate pulse calculated in step L1 and pulse parameter group ,like Figure 2 As shown, it will satisfy: All candidate pulses constitute the pulse group to be classified as follows:

[0093] in, Indicates the first One pulse group to be classified.

[0094] Let the first The number of pulses in the pulse groups to be classified is ,in, Indicates pulse group The number of candidate pulses contained within.

[0095] Let the minimum pulse number threshold be... ,in, This represents the minimum number of pulses required to perform population statistical analysis.

[0096] when At that time, the pulse group to be classified Perform subsequent feature extraction; when At that time, the first Each time window is marked as an invalid window.

[0097] Step L21: Extract the temporal correlation features of the pulse group to be classified.

[0098] The first one constructed above A pulse group to be classified Extract the arrival time of each candidate pulse in the pulse group. And arrange them in ascending order.

[0099] From the arrival time sequence, the time interval between adjacent pulses is calculated. :

[0100] in, Indicates the first The candidate pulse and the first The time interval between candidate pulses.

[0101] By time interval Calculate the average of the time intervals :

[0102] in, Indicates the first The average time interval between adjacent pulses within a pulse group to be classified.

[0103] By time interval and the mean of time intervals Calculate the standard deviation of the time interval. :

[0104] in, It represents the standard deviation of the time interval between adjacent pulses within the k-th pulse group to be classified.

[0105] From the mean of time intervals and standard deviation of time interval Calculate the time dispersion coefficient :

[0106] in, ε represents the time discretization coefficient of the k-th pulse group to be classified, and ε represents a small normal number to prevent the denominator from being zero.

[0107] Step L22: Extract the amplitude and energy correlation features of the pulse group to be classified.

[0108] The kth pulse group to be classified, constructed in step L2 Extract the peak amplitude of each candidate pulse in the pulse group. and pulse energy .

[0109] The peak amplitude of each candidate pulse calculated in step L1 Calculate the mean amplitude of the pulse group to be classified. :

[0110] in, This represents the average peak amplitude of candidate pulses within the k-th pulse group to be classified.

[0111] The peak amplitude of each candidate pulse calculated in step L1 and the mean amplitude Calculate the standard deviation of the amplitude. :

[0112] in, It represents the standard deviation of the peak amplitude of candidate pulses within the k-th pulse group to be classified.

[0113] From the mean amplitude and amplitude standard deviation Calculate the amplitude dispersion coefficient :

[0114] in, This represents the amplitude dispersion coefficient of the k-th pulse group to be classified.

[0115] The energy of each candidate pulse calculated in step L1 Calculate the average energy :

[0116] in, This represents the average energy of candidate pulses within the k-th pulse group to be classified.

[0117] The peak amplitude of each candidate pulse calculated in step L1 Energy of each candidate pulse Amplitude mean and energy mean Calculate the amplitude-energy correlation coefficient of the pulse group to be classified. :

[0118] in, denoted as the Pearson correlation coefficient between the peak amplitude and pulse energy in the k-th pulse group to be classified.

[0119] Step L23: Extract the phase correlation features of the pulse group to be classified.

[0120] The kth pulse group to be classified, constructed in step L2 Extract the power frequency phase corresponding to each candidate pulse in the pulse group to form a phase sequence:

[0121] in, This represents the phase sequence corresponding to the k-th pulse group to be classified.

[0122] The power frequency phase of each candidate pulse calculated in step L1 Calculate phase aggregation :

[0123] in, denoted by , where j represents the phase clustering degree of the k-th pulse group to be classified.

[0124] power frequency phase interval Divide into B equal-width subintervals, and denote the b-th phase subinterval as . :

[0125] Where B represents the total number of sub-intervals into which the phase interval is divided, and b represents the sequence number of the phase sub-interval.

[0126] The power frequency phase of each candidate pulse calculated in step L1 The statistics fall within the b-th phase sub-interval Number of pulses :

[0127] in, This represents an indicator function, which takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0128] Number of pulses in phase sub-interval and total number of pulses Calculate the probability corresponding to the phase sub-interval :

[0129] in, Let represent the probability value corresponding to the b-th phase sub-interval, and satisfy .

[0130] From probability distribution Calculate the phase distribution entropy of the pulse group to be classified. :

[0131] Among them, when When =0, the corresponding term takes the value of 0 according to the limit.

[0132] Step L24: Construct a five-dimensional associated feature vector.

[0133] The time discrepancy coefficients calculated in step L3 The amplitude dispersion coefficients calculated in step L4 And amplitude-energy correlation coefficient and the phase convergence calculated in step L5 and phase distribution entropy This constitutes a five-dimensional correlation feature vector. :

[0134] in, This represents the five-dimensional correlation feature vector corresponding to the k-th pulse group to be classified. This indicates the transpose operation.

[0135] Step L3: Construct a feature-level attention-based dynamically weighted multi-class neural network, and input the five-dimensional associated feature vector into the neural network for dynamic weighting and scoring calculation. By combining the classification probabilities within the complete detection period, the overall fault classification result and classification credibility are finally obtained.

[0136] Step L31: Construct a feature-level attention-based dynamically weighted multi-class neural network, such as... Figure 3 As shown. In this embodiment, the feature-level attention dynamically weighted multi-class neural network consists of an input layer, a normalization layer, a compression mapping layer, an expansion mapping layer, a feature-level attention weighting layer, a multi-class output layer, and a probability output layer connected sequentially, wherein: The input layer is used to receive the five-dimensional correlation feature vector obtained in step L24. The five-dimensional correlation feature vector includes time dispersion coefficient, amplitude dispersion coefficient, amplitude-energy correlation coefficient, phase clustering degree and phase distribution entropy, which are used to characterize the statistical characteristics of the pulse group to be classified in terms of time distribution, amplitude fluctuation, energy correlation and power frequency phase distribution.

[0137] The normalization layer is used to unify the scale of the five-dimensional correlated feature vectors, avoiding the influence of differences in the numerical range of features with different dimensions on subsequent weight learning. The normalized five-dimensional feature vectors are then fed into the compression mapping layer as the actual input data of the neural network. The compression mapping layer consists of a fully connected map and a nonlinear activation function, used to compress and map the five-dimensional input features to a low-dimensional hidden layer space, thereby extracting the combination relationships and nonlinear correlations between different feature components. The output of this layer is a hidden layer vector, whose dimension is smaller than that of the input features, used to form a compact representation of the overall feature state of the current pulse group.

[0138] The extended mapping layer is connected to the compressed mapping layer to regenerate a five-dimensional attention score vector with the same dimension as the input features based on the hidden layer vector. Each component in this attention score vector corresponds to a feature component in the five-dimensional associated features, representing the relative importance of that feature in the current pulse group fault identification process. Unlike traditional fixed-weight feature fusion methods, the attention score in this embodiment is adaptively calculated from the current input sample, thus different pulse groups to be classified can obtain different feature weight assignments.

[0139] The feature-level attention weighting layer constructs a dynamic weight matrix based on the five-dimensional attention score vector, and applies this dynamic weight matrix to the normalized five-dimensional associated feature vector to obtain a reconstructed feature vector. This reconstructed feature vector retains the physical meaning of the original five-dimensional features, while enhancing feature components that are more sensitive to the current fault type through dynamic weighting, and suppressing redundant feature components that are greatly affected by on-site noise, random impulses, or detection errors. This improves the model's anti-interference ability and classification stability in complex GIS on-site environments.

[0140] A multi-classification output layer is connected to a feature-level attention weighting layer to map the reconstructed feature vector into a fault type score vector. Each element in this fault type score vector corresponds to the original classification score for a preset GIS fault type. Subsequently, the probability output layer uses the Softmax function to normalize the fault type score vector, obtaining the probability distribution of the pulse cluster to be classified belonging to each fault type. Finally, the fault category with the highest probability is taken as the predicted fault type for the pulse cluster to be classified, and the classification results of multiple pulse clusters to be classified within the complete detection period are statistically fused to output the overall fault type and classification confidence.

[0141] thus, Figure 3 The five-dimensional correlation feature vector corresponds to the input layer of the neural network, the normalized feature vector corresponds to the output of the normalized layer, the compression mapping layer is used to generate the hidden layer vector, the expansion mapping layer is used to generate the five-dimensional attention score vector, the dynamic weight matrix and the reconstructed feature vector correspond to the feature-level attention weighting layer, the fault type score vector corresponds to the output of the multi-class output layer, and the fault type probability vector corresponds to the output of the Softmax probability output layer.

[0142] Step L32: The five-dimensional correlation feature vector constructed in step L6 Normalize it to obtain the feature vector. :

[0143] in, This represents the normalized five-dimensional correlation feature vector of the k-th pulse group to be classified. to These represent the five normalized feature components. This represents a five-dimensional real vector space.

[0144] The normalized five-dimensional correlation feature vector is input into the compression mapping layer to obtain the hidden layer vector. :

[0145] in, This represents the hidden layer vector corresponding to the k-th pulse group to be classified. This represents the weight matrix of the compression mapping layer. Let represent the bias vector of the compression mapping layer, d represent the hidden layer dimension, and satisfy d < 5. This represents a non-linear activation function.

[0146] Step L33: From hidden layer vectors The five-dimensional attention score vector is calculated through the extended mapping layer. :

[0147] in, This represents the five-dimensional attention score vector corresponding to the k-th pulse group to be classified. This represents the weight matrix of the extended mapping layer. This represents the bias vector of the extended mapping layer. express Activation function.

[0148] The attention score vector is represented as:

[0149] in, to Let each of the five feature components represent the attention scores, satisfying the following conditions:

[0150] From attention score vector Construct the dynamic weight matrix corresponding to the k-th pulse group to be classified. :

[0151] in, This represents the diagonal matrix construction operator.

[0152] From the dynamic weight matrix and the normalized five-dimensional correlation feature vector Calculate and reconstruct the feature vector :

[0153] in, Let represent the reconstructed feature vector of the k-th pulse group to be classified after dynamic weighting, i.e.:

[0154] Let the total number of fault types be ,in, This represents the total number of fault categories to be classified.

[0155] Reconstructing the feature vector The fault type score vector corresponding to the k-th pulse group to be classified is obtained through multi-classification output layer calculation. :

[0156] in, This represents the weight matrix of the multi-class output layer. This represents the bias vector of the multi-class output layer.

[0157] The fault type score vector is represented as follows:

[0158] in, This represents the raw score of the k-th pulse group to be classified, which belongs to the c-th type of fault, where c represents the fault category index. .

[0159] Step L34: From the fault type score vector ,through The function calculates the probability vector of the fault type. :

[0160]

[0161] in, This represents the probability that the k-th pulse group to be classified belongs to the c-th type of fault. Let represent the probability vector of the fault type corresponding to the k-th pulse group to be classified, and satisfy:

[0162] During the training phase, let the true fault category label of the k-th pulse group sample to be classified be . ,in, .

[0163] Step L35: Train using the multi-class cross-entropy loss function:

[0164] Where L represents the loss function value, and K represents the total number of pulse group samples to be classified participating in the training. This indicates that the k-th sample belongs to the true class. The predicted probability.

[0165] By optimizing the parameter set using the backpropagation algorithm, we have:

[0166] Here, Θ represents the set of all parameters to be optimized in the neural network.

[0167] Step L36: Based on Fault type classification is performed based on probability.

[0168] The fault type probability vector calculated from the above steps Perform fault type classification on the pulse group to be classified, and take... The category with the highest probability is used as the predicted fault type label for that pulse group. :

[0169] in, This represents the predicted fault type label for the k-th pulse group to be classified.

[0170] Let the corresponding maximum class probability be :

[0171] in, This indicates the maximum value of the k-th pulse group to be classified in predicting the fault type. Probability.

[0172] When it is necessary to output the fault type name, establish a mapping relationship between the fault category index and the fault type name. Then the fault type output result corresponding to the kth pulse group to be classified is:

[0173] in, This indicates the fault type name corresponding to the c-th type of fault.

[0174] Step L37: Output the overall fault classification results and reliability.

[0175] Let the total number of pulse groups to be classified during the complete detection cycle be . .

[0176] The fault type labels for each pulse group to be classified, calculated in step L8, are used to predict the fault type labels. Count the number of occurrences of each fault type.

[0177] Let the number of pulse groups to be classified correspond to the c-th type of fault be . ,but:

[0178] The most frequently occurring fault category is used as the overall fault classification result, which includes:

[0179] in, The label indicates the overall fault type corresponding to the complete testing cycle.

[0180] The maximum values ​​of each pulse group to be classified, calculated from the above steps, are as follows: probability To calculate the overall classification confidence, we have:

[0181] in, This indicates the overall reliability of the fault classification.

[0182] Final output: Overall fault type and overall classification credibility .

[0183] In summary, this method constructs the pulse group to be classified by a sliding time window, extracts multi-dimensional correlation features such as time, amplitude, energy and power frequency phase, and comprehensively describes the statistical law of fault discharge. Its core advantage lies in the introduction of a feature-level attention dynamic weighting mechanism, which can adaptively reconstruct feature weights for different samples, effectively amplifying key identification features and suppressing redundant noise.

[0184] Meanwhile, it also incorporates local waveform energy assessment to reduce the random errors of single-point instantaneous sampling, and integrates data from the entire detection cycle to output the overall fault classification results and classification reliability, which significantly improves the accuracy, anti-interference ability and engineering application reliability of GIS equipment fault diagnosis in complex field environments.

[0185] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0186] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0187] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the GIS burst fault classification methods based on weighted feature attention in the above embodiments.

[0188] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above method.

[0189] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. Memory, used to store computer programs; When the processor executes the program stored in memory, it implements the above-mentioned GIS burst fault classification method based on weighted feature attention.

[0190] The communication bus mentioned in the above-mentioned electronic devices can be a standard bus for interconnecting peripheral components or an extended industrial standard structure bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0191] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0192] The memory may include random access memory or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0193] The processors mentioned above can be general-purpose processors, including central processing units, network processors, etc.; they can also be digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0194] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0195] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0196] Furthermore, it should be noted that if any directional indication (such as up, down, left, right, front, back, etc.) is involved in the embodiments of the present invention, the directional indication is only used to explain the relative positional relationship and movement of each component in a specific posture. If the specific posture changes, the directional indication will also change accordingly.

[0197] Furthermore, those skilled in the art should understand that in the actual use of the embodiments of this application, there may be preset thresholds used as the basis for judging the corresponding technical solutions. These thresholds are conventional technical means commonly used in the field to implement functions such as state judgment, condition recognition, and control logic switching. The specific values, setting basis, value selection methods, determination methods, and adjustment rules of the thresholds involved in this technical solution are all conventional technical choices that can be reasonably determined by those skilled in the art based on conventional technical factors such as actual application scenarios, system working states, characteristics of the detection object, hardware performance parameters, and functional requirements, through conventional experiments, calibrations, and debugging. The specific setting and adjustment of the aforementioned thresholds will not cause this technical solution to be unimplementable as a whole, nor will it affect the realization of the core concept and the achievement of the technical effects of this technical solution.

[0198] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, in the embodiments of this invention, "multiple" refers to two or more. Moreover, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

Claims

1. A GIS burst fault classification method based on weighted feature attention, characterized in that, include: Step S1. Collect the raw time signal of the GIS monitoring object and detect candidate pulses, and calculate the arrival time, peak amplitude, energy and power frequency phase of each pulse; Step S2. Using a sliding time window, candidate pulses are constructed into a pulse group to be classified, and the time, amplitude and energy, phase and other dimensional correlation features of the pulse group to be classified are extracted to construct a five-dimensional correlation feature vector. Step S3. Construct a feature-level attention-based dynamically weighted multi-class neural network, and input the five-dimensional associated feature vector into the neural network for dynamic weighting and scoring calculation. By combining the classification probabilities within the complete detection period, the overall fault classification result and classification credibility are finally obtained.

2. The GIS burst fault classification method based on weighted feature attention as described in claim 1, characterized in that, The specific operation procedure for calculating the arrival time, peak amplitude, energy, and power frequency phase of each pulse in step S1 includes: Step S11. Collect raw continuous time signals of GIS monitored objects within the time interval. ,in, , Indicates time The corresponding original signal amplitude, Represents a continuous-time variable. Indicates the total sampling time; Step S12. According to the sampling frequency The original continuous-time signal is discretized to obtain a discrete signal. ,in, , Indicates the first The signal amplitude corresponding to each sampling point Indicates the sampling point number. Let N represent the sampling frequency, and N represent the total number of sampling points. ; Step S13. For discrete signals After performing basic filtering or denoising, the signal is searched for pulse components with local abrupt changes or local peaks that are significantly higher than the background fluctuations, and pulses that meet the preset detection conditions are recorded as candidate pulses. Step S14. Calculate the first one respectively. Arrival time of candidate pulses Peak amplitude ,energy and peak amplitude Finally, the number was obtained. The parameter set corresponding to each candidate pulse is represented as follows: .

3. The GIS burst fault classification method based on weighted feature attention as described in claim 2, characterized in that, In step S14, for the first Each candidate pulse represents a transient pulse signal initially detected in a discrete signal that may correspond to a certain type of fault discharge event. The calculation process includes: If the detection yields the first The peak positions of the candidate pulses are ,but Indicates the first The peak sampling point positions corresponding to each candidate pulse Indicates the candidate pulse number. (1) Calculate the arrival time of the candidate pulse. The calculation formula is as follows: in, Indicates the first The arrival time of each candidate pulse. Calculate the next Peak amplitude of each candidate pulse The calculation formula is as follows: in, Indicates the first The peak amplitude of each candidate pulse; Around the first Peak position of candidate pulses Extract a local pulse segment and calculate the first pulse. Energy of each candidate pulse The calculation formula is as follows; in, Indicates the first The energy of each candidate pulse This represents the local offset index relative to the peak position. This indicates the number of sampling points intercepted before the peak point. This indicates the number of sampling points truncated after the peak point; Finally calculate the... The power frequency phase corresponding to each candidate pulse The calculation formula is as follows: in, Indicates the first The power frequency phase corresponding to each candidate pulse. Indicates the fundamental frequency of the power frequency. Indicates to Round down to the nearest integer.

4. The GIS burst fault classification method based on weighted feature attention as described in claim 2, characterized in that, The specific operation process of step S2, which uses a sliding time window to construct the candidate pulses into a pulse group to be classified, includes: Construct a set of sliding time windows, with the time length of each analysis window being... The sliding time window has a first... The time window is ,in, Indicates the first A time window, Indicates the first The start time of a time window Indicates the time window number; The arrival times of each candidate pulse calculated in step S1 and pulse parameter group The conditions will be met. All candidate pulses constitute the first There are several pulse groups to be classified. ,in, Indicates the first One pulse group to be classified. Let the first The number of pulses in the pulse groups to be classified is The minimum number of pulses required to perform population statistical analysis is [threshold value]. ,when At that time, the pulse group to be classified Perform subsequent feature extraction, when At that time, the first Each time window is marked as an invalid window.

5. The GIS burst fault classification method based on weighted feature attention as described in claim 4, characterized in that, The specific operation process for extracting the temporal dimension correlation features of the pulse group to be classified in step S2 includes: From the already constructed first A pulse group to be classified Extract the arrival time of each candidate pulse in the pulse group. And arrange them in ascending order; Calculate the time interval between adjacent pulses from the arrival time sequence of each candidate pulse. The calculation formula is as follows: in, Indicates the first The candidate pulse and the first The time interval between candidate pulses; Calculate the first Mean time interval of each pulse group to be classified The calculation formula is as follows: in, Indicates the first The average time interval between adjacent pulses within a pulse group to be classified. Calculate the standard deviation of the time interval for the k-th pulse group to be classified. The calculation formula is as follows: in, It represents the standard deviation of the time interval between adjacent pulses within the k-th pulse group to be classified. Finally, the time discrepancy coefficients of the kth pulse group to be classified are calculated. The calculation formula is as follows: in, ε represents the time discretization coefficient of the k-th pulse group to be classified, and ε represents a small normal number to prevent the denominator from being zero.

6. The GIS burst fault classification method based on weighted feature attention as described in claim 4, characterized in that, The specific operation process for extracting the amplitude and energy dimension correlation features of the pulse group to be classified in step S2 includes: For the kth pulse group to be classified that has been constructed The peak amplitude of each candidate pulse Calculate the mean amplitude of the k-th pulse group to be classified. The calculation formula is as follows: in, This represents the average peak amplitude of candidate pulses within the k-th pulse group to be classified; Calculate the standard deviation of the amplitude of the k-th pulse group to be classified. The calculation formula is as follows: in, It represents the standard deviation of the peak amplitude of candidate pulses within the k-th pulse group to be classified. Calculate the amplitude dispersion coefficients of the kth pulse group to be classified. The calculation formula is as follows: in, This represents the amplitude dispersion coefficient of the k-th pulse group to be classified. Calculate the average energy of the k-th pulse group to be classified The calculation formula is as follows: in, This represents the average energy of candidate pulses within the k-th pulse group to be classified. Finally, the amplitude-energy correlation coefficient of the k-th pulse group to be classified is calculated. The calculation formula is as follows: in, denoted as the Pearson correlation coefficient between the peak amplitude and pulse energy in the k-th pulse group to be classified.

7. The GIS burst fault classification method based on weighted feature attention as described in claim 4, characterized in that, The specific operation process for extracting the phase dimension correlation features of the pulse group to be classified in step S2 includes: For the kth pulse group to be classified that has been constructed Extract the power frequency phase corresponding to each candidate pulse in the pulse group to form a phase sequence. ,in, This represents the phase sequence corresponding to the k-th pulse group to be classified; Calculate the phase clustering degree of the k-th pulse group to be classified The calculation formula is as follows: in, denoted by , where j represents the phase clustering degree of the k-th pulse group to be classified. power frequency phase interval Divide into B equal-width subintervals, and denote the b-th phase subinterval as . The calculation formula is as follows: Where B represents the total number of sub-intervals into which the phase interval is divided. This represents the b-th phase sub-interval, where b represents the phase sub-interval index. The statistics fall within the b-th phase sub-interval Number of pulses The formula is as follows: in, This represents the number of pulses in the b-th phase sub-interval. This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise. Calculate the probability corresponding to the b-th phase sub-interval The calculation formula is as follows: in, Let represent the probability value corresponding to the b-th phase sub-interval, and satisfy the condition. ; Calculate the phase distribution entropy of the k-th pulse group to be classified. The calculation formula is as follows: in, Represents the phase distribution entropy of the k-th pulse group to be classified; when When =0, the corresponding term takes the value of 0 according to the limit.

8. The GIS burst fault classification method based on weighted feature attention as described in claim 1, characterized in that, The specific calculation process for the dynamic weighting and scoring calculation of the neural network in step S3 includes: Step S31. After normalizing the five-dimensional correlation feature vector, input it into the neural network. Specifically, input the five-dimensional correlation feature vector into the compression mapping layer to obtain the hidden layer vector. ; Step S32. Based on hidden layer vectors The five-dimensional attention score vector is calculated through the extended mapping layer. And construct the corresponding dynamic weight matrix. Used to obtain reconstructed feature vectors Finally, the reconstructed feature vector The fault type score vector corresponding to the k-th pulse group to be classified is obtained through multi-classification output layer calculation. ; Step S33. From the fault type score vector ,through The function calculates the probability vector of the fault type corresponding to the k-th pulse group to be classified. ; Step S34. Train the network using a multi-class cross-entropy loss function, and during the training phase, let the true fault category label of the k-th pulse group sample to be classified be . ,in, , This represents the total number of fault categories to be classified, and the parameter set is also optimized during training using the backpropagation algorithm; Step S35. Based on the fault type probability vector Perform fault type classification on the k-th pulse group to be classified, and take... The category with the highest probability is used as the predicted fault type label for that pulse group. At the same time, a mapping relationship between the fault category index and the fault type name is also established. The output result for the fault type corresponding to the k-th pulse group to be classified is: ; Step S36. Finally, based on the total number of pulse groups to be classified during the complete detection cycle and the predicted fault type label for each pulse group to be classified, The frequency of occurrence of each fault type, and the maximum value of each pulse group to be classified. probability Calculate the overall classification confidence level and compare it with the overall fault type. Output together.

9. The GIS burst fault classification method based on weighted feature attention as described in claim 8, characterized in that, The specific operation process of step S32 includes: From hidden layer vectors The five-dimensional attention score vector corresponding to the k-th pulse group to be classified is calculated through the extended mapping layer. for: in, This represents the weight matrix of the extended mapping layer. This represents the bias vector of the extended mapping layer. express Activation function to Let each represent the attention score corresponding to one of the five feature components, and satisfy the following conditions: ; From attention score vector Construct the dynamic weight matrix corresponding to the k-th pulse group to be classified. ,have: , This represents the diagonal matrix construction operator. From the dynamic weight matrix and the normalized five-dimensional correlation feature vector Calculate the reconstructed feature vector of the k-th pulse group to be classified after dynamic weighting. The calculation formula is as follows: ,Right now Reconstructing the feature vector The fault type score vector corresponding to the k-th pulse group to be classified is obtained through multi-classification output layer calculation. : in, This represents the fault type score vector corresponding to the k-th pulse group to be classified. This represents the weight matrix of the multi-class output layer. This represents the bias vector of the multi-class output layer. This indicates that the k-th pulse group to be classified belongs to the k-th pulse group. The original score for the type of fault, Indicates the fault category index, .

10. The GIS burst fault classification method based on weighted feature attention as described in claim 8, characterized in that, The specific operational procedure for calculating the overall classification confidence level in step S36 includes: The total number of pulse groups involved in the classification process within the complete detection cycle. Fault type labels are predicted from each pulse group to be classified. Count the occurrence frequency of each fault type to obtain the number of pulse groups to be classified as fault type c. The calculation formula is as follows: The most frequently occurring fault category is used as the overall fault classification result, using the following formula: ,in This label represents the overall fault type corresponding to the complete testing cycle; The maximum value of each pulse group to be classified probability The overall classification confidence level is calculated using the following formula: in, This indicates the overall reliability of the fault classification.

Images