Circuit protection method and device based on current data

Abstract

The invention relates to the field of circuit protection analysis, and discloses a circuit protection method and device based on current data, and the method specifically comprises the steps: carrying out the layering noise estimation of an original current signal, obtaining a layering threshold value, carrying out the dynamic iteration noise reduction, avoiding the noise estimation deviation, carrying out the dynamic optimization of the layering threshold value through combining kurtosis and skewness, and obtaining a current signal; analyzing the optimized noise reduction current signal after optimization to make a protection decision; current feature extraction and analysis are carried out on the optimized noise reduction current signal, a bidirectional LSTM is combined with an attention mechanism to establish a circuit fault model and train the circuit fault model, a current threshold value is corrected in real time based on the circuit fault model, the current threshold value is compared with a real-time current value output by the optimized noise reduction current signal, and a protection decision is generated through analysis in combination with current features.

Description

Technical Field

[0001] The present invention relates to the technical field of circuit protection analysis, and in particular to a circuit protection method and device based on current data. Background Art

[0002] In the field of circuit protection, accurate analysis of current signals is crucial. With the increasing complexity of power systems and the increasing requirements for circuit reliability, traditional circuit protection methods face many challenges. The original current signal will inevitably be interfered by various noises during the actual acquisition process, making the effective processing of current signals a key link in ensuring the accuracy of circuit protection. Traditional signal noise reduction methods mostly use global threshold processing. In complex circuit environments, it is difficult to accurately capture the characteristics of noise in different frequency bands. The global threshold cannot perform differentiated processing for noise in each frequency band. In terms of building circuit fault models, early models lack full utilization of current signal timing information and key information. Since the occurrence and evolution of circuit faults are dynamic, it is difficult to accurately describe and predict fault states by relying on simple models. Traditional models are highly dependent on specific scenario data during training and lack versatility. When the application scenario changes, the performance and accuracy of the model will drop significantly. In the prior art, traditional noise reduction methods are mostly based on the Gaussian noise assumption. The noise in the actual circuit contains a large number of outliers and outliers. The global threshold processing method relied on by the traditional method cannot adapt to the complex noise environment and is extremely sensitive to non-Gaussian noise and outliers. It is easy to cause large deviations in noise estimation, which in turn affects the noise reduction effect and subsequent analysis of the signal. It fails to fully consider the timing information in the current signal. The occurrence of circuit faults is a dynamic process. There is a correlation between the current signals at the previous and next moments. Ignoring the timing information will cause the model to be unable to accurately capture the law of fault evolution, reducing the ability to predict and diagnose faults. The traditional model has no effective The mechanism is used to highlight the key information in the current signal. When processing large-scale data, it is difficult to focus on the data that is important for fault judgment, resulting in poor adaptability of the model to complex fault scenarios. It is unable to accurately distinguish different types of faults, judge the severity of faults, and make timely decisions. Traditional models rely heavily on a large amount of data from specific scenarios during training. It is necessary to collect and annotate a large amount of data for training for different circuits and application scenarios, which consumes a lot of manpower, material resources, and time. When faced with new and insufficiently trained scenarios, the model's generalization ability is insufficient, making it difficult to accurately judge faults and unable to effectively and timely protect circuits. In view of this, it is necessary to provide a circuit protection method and device based on current data. Summary of the invention

[0003] The object of the present invention is to provide a circuit protection method and device based on current data. To solve the above-mentioned prior art problems, the present invention is implemented through the following technical solutions: In a first aspect, an embodiment of the present invention provides a circuit protection method based on current data, which specifically includes the following steps: The original current signal is subjected to hierarchical noise estimation to obtain hierarchical thresholds and dynamically iterate noise reduction to avoid noise estimation deviation. Dynamically optimize the stratified threshold by combining kurtosis and skewness, and make protection decisions by analyzing the optimized noise-reduced current signal; The optimized noise-reduced current signal is used to extract and analyze the current features, and a bidirectional LSTM combined with an attention mechanism is used to establish and train a circuit fault model. The current threshold is corrected in real time based on the circuit fault model, compared with the real-time current value output by the optimized noise reduction current signal, and analyzed in combination with the current characteristics to generate a protection decision.

[0004] In a second aspect, an embodiment of the present invention provides a circuit protection system based on current data, which specifically includes the following units: Signal preprocessing unit: Use wavelet transform to decompose the original current signal into layers, calculate the high-frequency coefficients of each layer and analyze them to obtain the layer threshold, soft-threshold process the high-frequency coefficients of each layer to mark and exclude outliers, perform secondary noise estimation and threshold correction through signal residuals to dynamically iterate noise reduction; calculate the kurtosis and skewness of each layer of current data based on the high-frequency coefficients, compare with the preset standard value to determine whether there is non-Gaussian noise and outliers, calculate the dynamically optimized layer threshold and dynamically iterate noise reduction based on it to obtain the optimized noise-reduced current signal, and output the real-time current value; Feature extraction unit: extract time domain, frequency domain, time-frequency domain and nonlinear features to generate multi-dimensional feature vectors; Fault model unit: uses bidirectional LSTM combined with attention mechanism to build a circuit fault model, combined with Kirchhoff's law constraints, and outputs the fault probability; Protection decision unit: Based on the fault output probability output by the circuit fault model, the current threshold is corrected in real time using a preset formula, the real-time current value output by the optimized noise reduction current signal is combined with the corrected current threshold, and a comprehensive analysis is performed with reference to the multi-dimensional characteristics of the circuit current to generate a protection decision.

[0005] Beneficial effects of the present invention: 1. Wavelet transform is used for multi-resolution decomposition to achieve independent noise estimation of different frequency components. Compared with the traditional global threshold method, layered processing can more accurately capture the noise characteristics of different frequency bands, estimate the standard deviation based on the median absolute deviation, and improve the robustness to non-Gaussian noise and outliers through conversion coefficient correction, avoiding the sensitivity of traditional methods to outliers. Through residual signal reconstruction and secondary noise estimation, iterative correction of thresholds is achieved; the problem of initial noise estimation deviation in traditional single noise reduction is solved, which is suitable for non-stationary noise environments, and kurtosis and skewness are introduced as statistical criteria to break through the limitation of traditional noise reduction methods that rely on Gaussian noise assumptions. When non-Gaussian noise or outliers are detected, the threshold is adaptively increased by adjusting the coefficient to achieve a dynamic balance between noise suppression and signal fidelity. 2. By extracting current features, the characteristics of the current signal are comprehensively described from different angles, and the operating status of the circuit is accurately reflected. The circuit fault model of the bidirectional LSTM combined with the attention mechanism makes full use of the timing information and key information of the current signal, and is pre-trained on a large-scale IEEE power fault data set, so that the model can learn common power fault modes and characteristics. By fine-tuning to adapt to specific circuit fault protection scenarios, the dependence on a large amount of specific scenario data is reduced. The method of real-time correction of the current threshold can dynamically adjust the current threshold according to the failure probability of the circuit, and a comprehensive analysis is performed based on the real-time current value, the corrected current threshold and the current characteristics to generate a protection decision, so as to more comprehensively evaluate the operating status of the circuit and protect the circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0007] Figure 1 is a flow chart of the steps of a circuit protection method based on current data provided by Embodiment 1 and Embodiment 2 of the present invention; Figure 2 It is a structural schematic diagram of a circuit protection system based on current data provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0008] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0009] Embodiment 1 An embodiment of the present invention provides a circuit protection method based on current data, such as Figure 1 As shown, the specific steps include: Step 1: Perform hierarchical noise estimation on the original current signal to obtain hierarchical thresholds and dynamically iterate noise reduction to avoid noise estimation deviation; In a specific embodiment, the original current signal The specific method for layered noise estimation is: The original current signal is transformed by wavelet transform conduct The decomposition transformation of the layer is obtained High frequency coefficients of the layer ,in is the number of layers into which the original current signal is decomposed and transformed; For each layer of high frequency coefficients , calculate the high frequency coefficients The median of the absolute values ​​of Get the median absolute deviation , divide the obtained median absolute deviation by the conversion coefficient to obtain the Estimated standard deviation of layer noise ; It should be noted that the conversion coefficient represents the conversion relationship coefficient between the median and standard deviation of the high-frequency coefficient, and the value is preset to 0.6745. For the standard normal distribution, 67.45% of the data are within the range of ±0.6745 times the standard deviation of the mean; Based on the estimated standard deviation The noise of each layer is calculated by the formula Perform threshold calculation to obtain the stratified threshold ,in, It means the The length of the high-frequency coefficients of the layer; In a specific embodiment, the specific method of dynamic iterative noise reduction to avoid noise estimation deviation is: Apply layered thresholding to high-frequency coefficients at each layer , perform soft threshold processing, through the formula Calculate the iterative processing coefficient , based on the obtained processing coefficients, analyze and judge, and mark the abnormal values ​​of high-frequency coefficients; Specifically, if , it means that the high-frequency coefficients are abnormal. The high-frequency coefficients marked as outliers are excluded to obtain the reconstructed preliminary noise-reduced signal. ; Preliminary noise reduction signal based on reconstruction , the original current signal Subtract the signal from the initial noise reduction signal to get the signal residual ; Perform SWT decomposition on the signal residual to obtain the residual high-frequency coefficient ; For each layer of residual high frequency coefficients , calculate the residual high frequency coefficients The median of the absolute values ​​of Get the first corrected median absolute deviation , divide the obtained median absolute deviation by the conversion coefficient to obtain the The first corrected estimate standard deviation of the layer noise ; Based on the obtained first revised estimated standard deviation The noise of each layer is calculated by the formula Perform threshold calculation to obtain the first modified stratification threshold ,in, It means the The length of the high-frequency coefficients of the layer; The first modified stratification threshold Perform soft threshold processing to obtain the second iterative processing coefficient Based on the obtained iterative processing coefficient, the original current signal is dynamically iterated to reduce noise and obtain the first Iteration processing coefficient ,in, is the maximum number of iterations; Step 2: Dynamically optimize the stratified threshold by combining kurtosis and skewness, and make protection decisions by analyzing the optimized noise-reduced current signal; In a specific embodiment, the specific method for dynamically optimizing the stratification threshold by combining kurtosis and skewness is: Based on the obtained high-frequency coefficients, the kurtosis of each layer of current data is calculated by the formula Calculate the kurtosis of the data , based on the obtained high-frequency coefficient, the skewness of each layer of current data is calculated by the formula Calculate the data skewness ,in, It means the The mean of the high-frequency coefficients of the layer, It means the Standard deviation of layer high frequency coefficients; The calculated kurtosis is compared with the preset kurtosis standard value Compare and compare the calculated skewness with the preset skewness standard value Compare and analyze whether non-Gaussian noise and outliers appear in the stratification, and increase the stratification threshold; Specifically, if the kurtosis is greater than the preset kurtosis standard value and / or the skewness is greater than the preset skewness standard value, it indicates that non-Gaussian noise and outliers exist in the stratification. Calculate the optimized stratification threshold after dynamic optimization ,in, It indicates the preset adjustment coefficient; Based on the optimized layered threshold after dynamic optimization, dynamic iterative noise reduction is performed to obtain the optimized noise reduction current signal and output the real-time current value ; The optimized noise reduction current signal output by dynamic iterative noise reduction is analyzed with the preset current threshold value, and the circuit is powered off for protection when the real-time current value is greater than the preset current threshold value; The technical solution of the present invention is as follows: multi-resolution decomposition is performed using wavelet transform to achieve independent noise estimation of different frequency components; compared with the traditional global threshold method, layered processing can more accurately capture the noise characteristics of different frequency bands, estimate the standard deviation based on the median absolute deviation, and improve the robustness to non-Gaussian noise and outliers through conversion coefficient correction, avoiding the sensitivity of traditional methods to outliers, and realizing iterative correction of thresholds through residual signal reconstruction and secondary noise estimation; solving the problem of initial noise estimation deviation in traditional single denoising, being suitable for non-stationary noise environments, introducing kurtosis and skewness as statistical criteria, breaking through the limitation of traditional denoising methods relying on the Gaussian noise assumption, and when non-Gaussian noise or outliers are detected, the threshold is adaptively raised by adjusting the coefficient to achieve a dynamic balance between noise suppression and signal fidelity.

[0010] Embodiment 2 like Figure 1 As shown, a circuit protection method based on current data provided by an embodiment of the present invention specifically includes the following steps: Step 3: Extract and analyze the current features of the optimized noise-reduced current signal, and use a bidirectional LSTM combined with an attention mechanism to establish and train a circuit fault model; In a specific embodiment, the specific method of extracting current features from the optimized noise reduction current signal is: Acquire current characteristics of the optimized noise reduction current signal based on the optimized noise reduction current signal, the current characteristics include: time domain characteristics, frequency domain characteristics, time-frequency domain characteristics and nonlinear characteristics; It should be noted that the time domain characteristics of the optimized noise reduction current signal include but are not limited to: mean, variance, kurtosis and waveform factor; the frequency domain characteristics of the optimized noise reduction current signal include but are not limited to: total harmonic distortion rate, fundamental energy proportion and specific frequency component proportion; the time and frequency domain characteristics of the optimized noise reduction current signal include but are not limited to: wavelet entropy and local mean decomposition energy distribution; the nonlinear characteristics of the optimized noise reduction current signal include but are not limited to: permutation entropy and sample entropy; Specifically, based on optimizing the noise reduction current signal setting sampling points, get The current sampling value of the sampled current signal , the mean value of the sampling current signal is calculated by the mean value formula based on the current sampling value obtained; The variance of the sampled current signal is calculated by using the mean value of the sampled current signal and the current sampling value through a variance formula; The mean value of the sampled current signal is calculated using the formula Calculate the waveform factor of the sampled current signal ,in, It represents the number of sampled current signals. It means the A sampled current signal, It represents the mean value of the sampled current signal; Get the harmonic effective value of the current signal, and use the calculation formula Get the total harmonic distortion rate of the current signal ,in, It indicates the specific harmonic order. It means the The effective value of the subharmonics, It represents the effective value of the fundamental wave; Obtain the energy of the fundamental wave of the current signal and divide it by the total harmonic energy to obtain the proportion of the energy of the fundamental wave of the current signal in the total harmonic energy; Obtain the energy proportion of harmonics of a specific order, identify specific circuit faults, and calculate the formula Get the proportion of specific frequency components ,in, It indicates the specific harmonic order. It means the The energy of subharmonics; After decomposing the signal through wavelet transform, the information entropy of the energy distribution of each frequency band is calculated to obtain the wavelet entropy, which measures the complexity of the signal; Decompose the signal into local characteristic scale components (Product Function components), calculate the energy proportion of each component to obtain the local mean decomposition energy distribution, which is used to analyze the transient components in non-stationary signals; It should be noted that the local characteristic scale component represents a representation form used to describe the local characteristics of signal data at different scales; The permutation entropy is calculated by the permutation pattern of the time series. The larger the permutation entropy, the stronger the randomness. Obtain the sample entropy of the current signal to measure the probability of pattern repetition in the time series. The smaller the sample entropy, the stronger the regularity. In a specific embodiment, the specific method of using bidirectional LSTM combined with attention mechanism to establish and train the circuit fault model is: A circuit fault model is established using a bidirectional long short-term memory network (BidirectionalLSTM) combined with an attention mechanism; The circuit fault model includes: current feature input layer, bidirectional LSTM layer, attention mechanism layer and fully connected layer; It should be noted that the bidirectional long short-term memory network is an extension based on the long short-term memory network LSTM, which consists of two LSTMs in opposite directions. A forward LSTM processes input data from the beginning to the end of the sequence, and the other reverse LSTM processes input data from the end to the beginning of the sequence. The bidirectional LSTM captures forward and reverse timing information and processes the dynamic characteristics of circuit fault evolution. The attention mechanism is a technology widely used in the field of deep learning and artificial intelligence. It simulates the way attention is allocated in the human visual system, so that the model focuses on important information in the input data and ignores irrelevant or minor information, thereby improving the performance and efficiency of the model, dynamically focusing on the key parts of the input sequence and improving the performance of the model; The calculated multi-dimensional current features are input into the bidirectional LSTM layer. Kirchhoff's law of the circuit equation is introduced as a regularization term in the bidirectional LSTM to constrain the model to learn features that conform to the physical laws of the circuit. The forward and reverse input current feature sequences are processed through the bidirectional LSTM layer. The output of the bidirectional LSTM layer is weighted by the attention mechanism, and the output of the attention mechanism layer is mapped to the output of the fault probability to obtain the fault output probability. ; Specifically, the current circuit fault type is mapped to a query vector through a fully connected layer to obtain a query vector , for each time step the hidden state The linear transformation is done by the formula Get the key vector ,in, It means that the hidden state The linear transformation weight matrix mapped to the key vector space determines the contribution weight of different features to the attention calculation. It means adding a learnable offset to the key vector; The time step is obtained by weight normalization through the attention score formula Importance weight for fault judgment ; Hidden State The linear transformation is calculated by the formula to obtain the value vector , weighted fusion Get the center vector ; Through the fully connected layer after nonlinear transformation Mapping output failure output probability ,in, represents the weight matrix of the fully connected layer, It represents the offset of the fully connected layer. represents the fault output probability weight matrix, It represents the offset of the fault output probability. ; Pre-training is performed on the IEEE power fault dataset, and the pre-trained model is fine-tuned to the circuit fault protection scenario. The model is trained using the training dataset, and the L2 regularization loss function is used to prevent overfitting. Step 4: Based on the circuit fault model, the current threshold is corrected in real time, compared with the real-time current value output by the optimized noise reduction current signal, and analyzed in combination with the current characteristics to generate a protection decision; In a specific embodiment, the specific method of correcting the current threshold in real time based on the circuit fault model and making a protection decision for the circuit in combination with current characteristic analysis is: The calculated fault output probability is used to correct the current threshold in real time, through the formula Get the corrected current threshold ,in, It indicates the preset current threshold. The figure shows the preset sensitivity coefficient, which is 0.3; The real-time current value obtained by optimizing the noise reduction current signal output is compared with the modified current threshold and analyzed in combination with the current characteristics to generate a protection decision; Exemplarily, a circuit preset current threshold is set For 10A, the sensitivity coefficient is 0.3; when the model outputs the probability of failure output is 0.1, and the corrected current threshold is calculated by the formula 9.7A, optimized real-time current value of noise reduction current signal output is 8A, and through analysis At the same time, combined with the analysis of the current characteristics, the circuit is judged to be in normal operation and no protection operation is required; when the fault output probability of the model output is is 0.5, and the corrected current threshold is calculated by the formula 8.5A, optimize the real-time current value of the noise reduction current signal output is 8.2A, obtained through analysis At the same time, combined with the analysis of the sudden increase in the total harmonic distortion rate in the current characteristics, the circuit has a potential fault risk, and an early warning signal is issued to remind the operation and maintenance personnel to check; when the fault output probability of the model output is 0.9, and the corrected current threshold is calculated by the formula 7.3A, optimize the real-time current value of the noise reduction current signal output is 8A, and through analysis At the same time, combined with the analysis of the abnormal waveform factor in the current characteristics, it is determined that the circuit has a fault and the circuit is immediately powered off for protection to prevent the fault from further expanding; The technical solution of the embodiment of the present invention is: based on optimizing the noise reduction current signal, obtaining the current characteristics in the time domain, frequency domain, time-frequency domain and nonlinearity, using the bidirectional LSTM combined with the attention mechanism to establish a circuit fault model, the bidirectional LSTM is composed of two LSTMs in opposite directions, capturing the forward and reverse timing information, introducing the circuit equation Kirchhoff's law as a regularization term in the bidirectional LSTM, constraining the model to learn features that conform to the physical laws of the circuit, and following the basic physical principles of the circuit, the attention mechanism performs attention weighting on the output of the bidirectional LSTM layer, dynamically focusing on the key parts of the input sequence, so that the model focuses on important information and ignores irrelevant or minor information, the model is pre-trained on the IEEE power fault data set, fine-tuned to the circuit fault protection scenario, trained using the training data set, using the L2 regularization loss function to prevent overfitting, improve the generalization ability and stability of the model, and the calculated fault input The probability of output is corrected in real time through a specific formula, and the real-time current value of the optimized noise reduction current signal output is combined with the corrected current threshold for analysis in combination with the current characteristics, and corresponding protection decisions are generated according to different situations; by extracting current characteristics, the characteristics of the current signal are comprehensively described from different angles, and the operating status of the circuit is accurately reflected. The circuit fault model of the bidirectional LSTM combined with the attention mechanism makes full use of the timing information and key information of the current signal, and is pre-trained on a large-scale IEEE power fault data set, so that the model learns common power fault modes and characteristics, and adapts to specific circuit fault protection scenarios through fine-tuning, reducing dependence on a large amount of specific scenario data. The method of real-time correction of the current threshold can dynamically adjust the current threshold according to the fault probability of the circuit, and combines the real-time current value, the corrected current threshold and the current characteristics for comprehensive analysis to generate protection decisions, more comprehensively evaluate the operating status of the circuit and protect the circuit.

[0011] Embodiment 3 like Figure 2 As shown, a circuit protection system based on current data provided by an embodiment of the present invention specifically includes the following units: Signal preprocessing unit: Use wavelet transform to decompose the original current signal into layers, calculate the high-frequency coefficients of each layer and analyze them to obtain the layer threshold, soft-threshold process the high-frequency coefficients of each layer to mark and exclude outliers, perform secondary noise estimation and threshold correction through signal residuals to dynamically iterate noise reduction; calculate the kurtosis and skewness of each layer of current data based on the high-frequency coefficients, compare with the preset standard value to determine whether there is non-Gaussian noise and outliers, calculate the dynamically optimized layer threshold and dynamically iterate noise reduction based on it to obtain the optimized noise-reduced current signal, and output the real-time current value; Feature extraction unit: extract time domain, frequency domain, time-frequency domain and nonlinear features to generate multi-dimensional feature vectors; Fault model unit: uses bidirectional LSTM combined with attention mechanism to build circuit fault model, combined with Kirchhoff's law constraints, and outputs fault probability; Protection decision unit: Based on the fault output probability output by the circuit fault model, the current threshold is corrected in real time using a preset formula, the real-time current value output by the optimized noise reduction current signal is combined with the corrected current threshold, and a comprehensive analysis is performed with reference to the multi-dimensional characteristics of the circuit current to generate a protection decision.

[0012] An embodiment of the present invention is described in detail above, but the content described is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention; the above formulas are all dimensionless and numerical calculations, and the formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions and historical experience, and can be adjusted according to actual conditions; the above description is only a preferred embodiment of the present invention and is not used to limit the present invention. All equal changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A circuit protection method based on current data, characterized in that: The following steps are involved: The original current signal is subjected to hierarchical noise estimation to obtain hierarchical thresholds and dynamically iterate noise reduction to avoid noise estimation deviation. Dynamically optimize the stratified threshold by combining kurtosis and skewness, and make protection decisions by analyzing the optimized noise-reduced current signal; The optimized noise-reduced current signal is used to extract and analyze the current features, and a bidirectional LSTM combined with an attention mechanism is used to establish a circuit fault model and perform model training. The current threshold is corrected in real time based on the circuit fault model, compared with the real-time current value output by the optimized noise reduction current signal, and analyzed in combination with the current characteristics to generate a protection decision.

2. A circuit protection method based on current data according to claim 1, characterized in that: The method for performing hierarchical noise estimation on the original current signal is: The original current signal is transformed by wavelet transform Perform hierarchical decomposition transformation to obtain the high-frequency coefficients of each layer ; For each layer of high-frequency coefficients, the median of the absolute values ​​of the high-frequency coefficients is calculated to obtain the median absolute deviation, and the median absolute deviation is divided by the conversion coefficient to obtain the estimated standard deviation of the noise.

3. The circuit protection method based on current data according to claim 1, characterized in that: The method for obtaining the stratification threshold is: Based on the estimated standard deviation The noise of each layer is calculated to obtain the layer threshold ,in, It means the The length of the high frequency coefficients of the layer.

4. The circuit protection method based on current data according to claim 1, characterized in that: The specific method of dynamic iterative noise reduction is: Apply layered thresholds to the high-frequency coefficients of each layer, perform soft threshold processing, and calculate the iterative processing coefficients through the formula , based on the obtained processing coefficients, analysis and judgment are performed, abnormal values ​​of high-frequency coefficients are marked, and the original current signal is subtracted from the preliminary noise reduction signal to obtain the signal residual; Perform SWT decomposition on the signal residual to obtain the residual high-frequency coefficient ; For each layer of residual high frequency coefficients , calculate the residual high frequency coefficients The median of the absolute values ​​of the first corrected median absolute deviation is calculated , divide the obtained median absolute deviation by the conversion coefficient to obtain the The first corrected estimate standard deviation of the layer noise ; Based on the obtained first revised estimated standard deviation The noise of each layer is calculated by first correcting the layer threshold ,in, It means the The length of the high-frequency coefficients of the layer; The first modified stratification threshold Perform soft threshold processing to obtain the second iterative processing coefficient Based on the obtained iterative processing coefficient, the original current signal is dynamically iterated to reduce noise and obtain the first Iteration processing coefficient ,in, is the maximum number of iterations.

5. The circuit protection method based on current data according to claim 1, characterized in that: The method for dynamically optimizing the stratification threshold by combining kurtosis and skewness is: The kurtosis of each layer of current data is calculated based on the obtained high-frequency coefficients, and the data kurtosis is calculated by the formula , the skewness of each layer of current data is calculated based on the obtained high-frequency coefficient, and the data skewness is calculated by the formula ; If the kurtosis is greater than the preset kurtosis standard value or the skewness is greater than the preset skewness standard value, it indicates that there are non-Gaussian noise and outliers in the stratification. The optimized stratification threshold after dynamic optimization is calculated by the formula .

6. The circuit protection method based on current data according to claim 1, characterized in that: The method for establishing a circuit fault model by using a bidirectional LSTM combined with an attention mechanism is as follows: the multi-dimensional current features obtained by calculation are input into a bidirectional LSTM layer, Kirchhoff's law of the circuit equation is introduced into the bidirectional LSTM as a regularization term, the model is constrained to learn features that conform to the physical laws of the circuit, and the forward and reverse input current feature sequences are processed by the bidirectional LSTM layer; The output of the bidirectional LSTM layer is weighted by the attention mechanism, and the output of the attention mechanism layer is mapped to the output of the fault probability to obtain the fault output probability. .

7. The circuit protection method based on current data according to claim 6, characterized in that: The specific method for obtaining the fault output probability is: The query vector is obtained by mapping the current circuit fault type to the query vector through the fully connected layer. , for each time step the hidden state The linear transformation obtains the key vector through the formula ; The hidden state linear transformation is calculated by the formula to obtain the value vector , weighted fusion to obtain the concentrated vector ; The fault output probability is output through the nonlinear transformation mapping through the fully connected layer , fault output probability .

8. The circuit protection method based on current data according to claim 1, characterized in that: The model training method is: Pre-training is performed on the IEEE power fault dataset, and the pre-trained model is fine-tuned to the circuit fault protection scenario. The model is trained using the training dataset, and the L2 regularized loss function is adopted.

9. The circuit protection method based on current data according to claim 1, characterized in that: The method of comparing the real-time corrected current threshold with the real-time current value output by the optimized noise reduction current signal and analyzing and generating a protection decision in combination with the current characteristics is as follows: The calculated fault output probability is used to correct the current threshold in real time, through the formula Get the corrected current threshold ,in, It indicates the preset current threshold. It represents the preset sensitivity coefficient, which is 0.3; The optimized noise-reduced current signal outputs the real-time current value and the modified current threshold, combined with the current characteristics to analyze and generate a protection decision.

10. A circuit protection system based on current data, the system being used to execute the circuit protection method according to any one of claims 1 to 9, characterized in that: The following units are included: Signal preprocessing unit: Perform layered noise estimation on the original current signal to obtain the layered threshold and dynamically iterate the noise reduction to avoid noise estimation deviation. Combine the kurtosis and skewness to dynamically optimize the layered threshold to obtain the optimized noise-reduced current signal. Feature extraction unit: extract and analyze the current features of the optimized noise reduction current signal; Fault model unit: Use bidirectional LSTM combined with attention mechanism to establish and train circuit fault model; Protection decision unit: Based on the circuit fault model, the current threshold is corrected in real time, compared with the real-time current value output by the optimized noise reduction current signal, and analyzed in combination with the current characteristics to generate a protection decision.

Images