Zero sequence voltage sudden change characteristic ground fault identification method and system for ftu

Abstract

This invention belongs to the field of FTU fault identification, and particularly relates to a method and system for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs. The method first obtains the discrete sequence of zero-sequence voltage of the distribution network bus, calculates the absolute value sequence of the difference between its continuous sampling points, and constructs a mutation feature sequence through sliding time window integration; it is then input into a pre-trained capsule neural network, where the primary capsule layer encodes the features of each time window into a primary capsule vector; the connection weights between the primary and higher-level capsules are updated through iterative calculation until convergence, and the fault is determined based on the converged higher-level capsule vector: when the ground fault state capsule magnitude exceeds the first threshold and the angle between it and the fault feature basis vector is less than the second threshold, a fault is determined and a start signal is output; otherwise, the system is determined to be normal, which can improve the accuracy and efficiency of FTU identification.

Description

Technical Field

[0001] This invention belongs to the field of FTU fault identification, and particularly relates to a method and system for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs. Background Technology

[0002] With the widespread integration of distributed photovoltaic, wind power, and other green energy sources into medium- and low-voltage distribution networks via power electronic converters, the power supply structure, power flow distribution, and fault characteristics of distribution networks have undergone profound changes, exhibiting new characteristics such as mixed operation of multiple power sources, flexible and varied network topologies, enhanced asymmetry of ground parameters, and limited and complex fault current levels. In low-current grounding systems of distribution networks, single-phase grounding faults have the highest probability of occurrence. Traditional detection methods based on steady-state power frequency zero-sequence voltage amplitude exceeding limits or harmonic component analysis are ineffective in scenarios with high penetration of distributed power sources, frequent load fluctuations, high grounding resistance, or intermittent arcing. These methods are susceptible to interference from background harmonics, measurement noise, and system imbalances, leading to decreased detection sensitivity, startup delays, and even misjudgments.

[0003] In existing technologies, a common approach relies on the rise in steady-state zero-sequence voltage after a fault for detection. For example, a fixed threshold is set, and a ground fault is identified when the effective value of the zero-sequence voltage exceeds this threshold. This method is simple in principle, but has poor adaptability: in systems with high-resistance grounding or those compensated by arc suppression coils, the rise in steady-state zero-sequence voltage is not significant, making reliable startup difficult; when the system has a large capacitance to ground or there is a neutral point grounding of distributed power sources, zero-sequence voltage deviation may occur during normal operation, easily leading to malfunctions.

[0004] Another approach attempts to utilize the transient characteristics of the fault at the moment of occurrence, such as by detecting the abrupt change rate of zero-sequence voltage, the energy of high-frequency components, or the degree of waveform distortion. This type of method improves the detection speed under high-resistance grounding conditions, but still has significant shortcomings: most methods focus on threshold discrimination of a single transient characteristic of zero-sequence voltage, failing to effectively construct a comprehensive criterion that can highlight the abrupt change characteristics of the fault while suppressing normal operation disturbances; in addition, existing methods rely heavily on experience or fixed parameters for selecting the transient characteristic window and setting the threshold, making it difficult to adapt to different system-to-ground capacitances, different fault initial phase angles, and transient interference caused by distributed power source switching, resulting in insufficient detection reliability and poor anti-interference capabilities under complex operating conditions.

[0005] Therefore, how to accurately extract and quantify the transient change characteristics at the fault initiation moment from the zero-sequence voltage signal, while effectively suppressing transient and steady-state disturbances caused by normal operation interferences such as changes in system-to-ground capacitance, distributed power supply switching, and load fluctuations, is a problem that existing technologies urgently need to solve. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method and system for identifying ground faults caused by zero-sequence voltage mutations in FTUs. The method first obtains a discrete sequence of zero-sequence voltages from the distribution network bus, calculates the absolute difference between consecutive sampling points to obtain a sequence of absolute difference values, integrates this sequence over a sliding time window to construct a sequence of mutation features, and inputs it into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a primary capsule vector, forming an output vector. Iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules representing the system state, updating the connection weights until convergence. The fault is determined based on the converged state of the higher-level capsule vectors. If the vector magnitude of the higher-level capsule in the ground fault state exceeds a first preset threshold, and the angle between it and the preset fault feature basis vector is less than a second preset threshold, a single-phase ground fault is determined, and a fault initiation signal is output. Otherwise, the system is considered to be operating normally. This method improves the accuracy and efficiency of FTU fault identification and adapts to the different operating conditions of the distribution network.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A ground fault identification method for zero-sequence voltage mutation characteristics for FTUs includes: Discrete sequences characterizing the zero-sequence voltage of the distribution network bus are obtained using configured electrical sensors. U 0 ( n )},in n The sampling point number; Based on the discrete sequence { U 0 ( n )}, calculate the absolute difference between its consecutive sampling points to obtain the absolute difference sequence { U 0 ( n )}; For the difference absolute value sequence { U 0 ( n Perform sliding time window integration to construct a sequence of mutation feature quantities, where the feature quantity corresponding to the k-th time window is... , L The preset window length, k For time window sequence number; The mutation feature sequence is input into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a set of primary capsule vectors, forming the output vector of the primary capsule layer. Iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules that represent the system state, updating the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. Fault determination is performed based on the converged vector state of the high-level capsule: if the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault characteristic basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault initiation signal is output. Otherwise, the system is considered to be in normal operating condition.

[0008] Specifically, the primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, including: The standardized mutation feature sequence is divided into multiple consecutive feature subsequences according to a preset non-overlapping partitioning rule; wherein each feature subsequence contains a preset number P consecutive standardized mutation features, where P is an integer greater than 1, and there is no data overlap between adjacent feature subsequences. Multiple consecutive feature subsequences are input into a preset working condition feature perception layer and a pre-configured three-layer parallel convolutional layer is called to obtain the first, second and third feature maps with working condition features. The first, second, and third feature maps with working condition features are concatenated along the channel dimension to generate a fused feature map corresponding to the feature subsequence. Specifically, the primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, and also includes: Based on the fused feature map, the depth control module performs the following operations to generate a depth-encoded feature vector: Extract the original feature subsequence corresponding to the current fused feature map; Calculate the kurtosis and skewness statistical features of the original feature subsequence; Compare the kurtosis statistical feature value with a third preset threshold, and compare the skewness statistical feature value with a fourth preset threshold; If the kurtosis statistical feature value is greater than a third preset threshold and the skewness statistical feature value is greater than a fourth preset threshold, then the fused feature map is input to the deep processing path to obtain a deep encoded feature vector; the deep processing path consists of two concatenated convolutional processing units, each convolutional processing unit including: a one-dimensional convolutional layer configured with 16 convolutional kernels of size 3 and a stride of 1; a batch normalization layer; and a ReLU activation function layer; If the third or fourth preset threshold is not met, the fused feature map is input to the shallow processing path to obtain a deep encoded feature vector. The shallow processing path consists of one of the convolutional processing units.

[0009] Specifically, the primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, and also includes: The deep encoded feature vector is subjected to noise suppression processing to obtain the suppressed feature vector, specifically as follows: Determine the original feature subsequence corresponding to the deep encoded feature vector, and define the original feature subsequence as the current processing subsequence; Based on the division order of the standardized mutation feature sequence, the adjacent feature subsequences that are sequentially preceding the current processing subsequence are obtained as the preceding subsequences, and the adjacent feature subsequences that are sequentially following the current processing subsequence are obtained as the following subsequences. All normalized mutation features contained in the current processing subsequence, the preceding subsequence, and the subsequent subsequence are merged to form a merged sequence.

[0010] Specifically, the primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, and also includes: Calculate the variance of all elements in the merged sequence, and use the value of the variance as the noise evaluation value of the current process; Calculate the median of the noise evaluation values ​​corresponding to all feature subsequences in the current processing batch, and use the median as the baseline noise value; Calculate the dynamic attenuation coefficient α , =min(1, ), where T is the median of the noise evaluation values ​​of all sequence segments processed in the current batch, and V is the noise evaluation value of the current sequence segment; Each element of the deep encoded feature vector is multiplied by the dynamic decay coefficient to obtain the suppressed feature vector.

[0011] Specifically, the primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, and also includes: Converting the suppressed feature vector into a primary capsule vector specifically includes: Obtain the total number of time windows N contained in the standardized mutation feature sequence; The target capsule dimension D is dynamically determined based on the total number of time windows N. The specific rules are as follows: when N is less than or equal to 100, D is set to 8; when N is greater than 100 and less than or equal to 200, D is set to 10; when N is greater than 200, D is set to 12. The suppressed feature vector is input into a fully connected projection layer, which maps the dimension of the input vector to the dimension D of the target capsule. The output vector of the fully connected projection layer is normalized using the L2 norm, and the normalized vector is used as a primary capsule vector. The standardized mutation feature sequence is divided into segments of a preset fixed length, such that each segmented feature subsequence contains 50 consecutive standardized mutation features, and there is no data overlap between any two adjacent feature subsequences. For each feature subsequence obtained by the division, the primary capsule vector acquisition process is executed sequentially to generate a primary capsule vector corresponding to the corresponding feature subsequence; All generated primary capsule vectors are arranged and combined according to the chronological order of their corresponding feature subsequences in the standardized mutation feature sequence to form the output vector of the primary capsule layer.

[0012] Specifically, the multiple consecutive feature subsequences are input into a preset working condition feature perception layer and a pre-configured three-layer parallel convolutional layer is invoked, as follows: The operating condition feature perception layer receives the multiple consecutive feature subsequences and calculates the feature statistics of the multiple consecutive feature subsequences; the feature statistics include mean, standard deviation, and maximum amplitude. F 0max and coefficient of variation C v ; Based on characteristic statistics and by setting a first or second judgment condition, the fault condition type is obtained, specifically: When the first condition is met, that is, A max >0.5% F >0.2 This is determined to be a low-resistance fault basic operating condition. When the second criterion is met, that is, if 0.3 ≤ A max ≤0.5 and 0.1≤ s F ≤0.2, judged as a medium resistance fault basic operating condition; When the third criterion is met, that is, if A max <0.3 and σF <0.1 The condition was determined to be an ultra-high resistance fault. If the first, second, and third judgment conditions are not met, the condition is judged to be normal or transitional. The invocation of the pre-configured three-layer parallel convolutional layer is specifically as follows: Based on the final determined working condition type, the parameters of the pre-configured three-layer parallel convolutional layer are invoked: If the basic operating condition is determined to be a low-resistance fault, the first set of parameters is called: the first convolutional layer uses 32 one-dimensional convolutional kernels with a size of 2, the second convolutional layer uses 64 one-dimensional convolutional kernels with a size of 2, and the third convolutional layer uses 128 one-dimensional convolutional kernels with a size of 2. If the basic operating condition is determined to be a medium resistance fault, the fourth set of parameters is called: the first convolutional layer uses 24 one-dimensional convolutional kernels with a size of 3, the second convolutional layer uses 48 one-dimensional convolutional kernels with a size of 3, and the third convolutional layer uses 96 one-dimensional convolutional kernels with a size of 3. If the basic operating condition is determined to be an ultra-high resistance fault, the second set of parameters is called: the first convolutional layer uses 8 one-dimensional convolutional kernels with a size of 4, the second convolutional layer uses 16 one-dimensional convolutional kernels with a size of 4, and the third convolutional layer uses 32 one-dimensional convolutional kernels with a size of 4. If the condition is determined to be normal or transitional, the third set of parameters is called: the first convolutional layer uses 16 one-dimensional convolutional kernels of size 3, the second convolutional layer uses 32 one-dimensional convolutional kernels of size 3, and the third convolutional layer uses 64 one-dimensional convolutional kernels of size 3.

[0013] Specifically, iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules representing the system state to update the connection weights between the primary capsule vector and each higher-level capsule, including: Based on the output vector of the primary capsule layer, the connection weights between each higher-level capsule and all primary capsule vectors are initialized as follows: Obtain each primary capsule vector contained in the primary capsule layer output vector; Calculate the magnitude of each of the primary capsule vectors; Linearly normalize the magnitudes of all primary capsule vectors so that the sum of the normalized magnitudes equals the total number of higher-level capsules. The magnitude of each primary capsule vector is directly set as the initial connection weight between the corresponding primary capsule vector and each higher-level capsule; Based on the initial connection weights and the output vector of the primary capsule layer, an iterative calculation process is performed. A single iteration includes the following implementation steps: The prediction vector of the high-level capsule is calculated based on the current connection weights. Specifically, each primary capsule vector is mapped through a pre-trained linear transformation matrix to obtain the prediction vector of each high-level capsule for each primary capsule vector. For the same high-level capsule, the prediction vectors of all primary capsule vectors are multiplied by the current connection weights between the corresponding primary capsule vectors and the high-level capsule, and the weighted prediction vectors are summed to obtain the prediction vector of the high-level capsule in the current iteration round.

[0014] Specifically, the iterative calculation process is performed, and a single iteration includes the following implementation steps, as well as: Obtain the vector of each primary capsule and the predicted vector of each higher-level capsule in the current iteration round; Calculate the vector dot product between each primary capsule vector and each high-level capsule prediction vector, and define the result as the first similarity metric. For each high-level capsule, perform an update operation: add the current connection weight between the high-level capsule and each primary capsule vector to the first similarity metric between the primary capsule vector and the high-level capsule to obtain a set of original update values; For all the original updated values ​​belonging to the same high-level capsule, the Softmax function is applied to normalize the calculation and obtain the updated connection weights. For each high-level capsule, obtain the predicted vector for that high-level capsule from all the primary capsule vectors; Each prediction vector is multiplied by the updated connection weight between the corresponding primary capsule vector and the higher-level capsule to obtain a weighted prediction vector. Summing all weighted prediction vectors yields the weighted sum vector of the high-level capsule. Apply a nonlinear compression function to the weighted sum vector: calculate the magnitude of the weighted sum vector, and adjust the values ​​of each component of the vector based on the magnitude using the compression function so that the magnitude of the output vector is no greater than 1; The compressed vector is used as the high-level capsule vector for the current iteration round.

[0015] A ground fault identification system for zero-sequence voltage mutation characteristics for FTUs includes: The acquisition module obtains discrete sequences representing the zero-sequence voltage of the distribution network bus. U 0 ( n )},in n The sampling point number; The difference module, based on the discrete sequence { U 0 ( n )}, calculate the absolute difference between its consecutive sampling points to obtain the absolute difference sequence { D.U. 0 ( n )}; The feature construction module performs a process on the difference absolute value sequence { D.U. 0 ( n Perform sliding time window integration to construct a sequence of mutation feature quantities, where the feature quantity corresponding to the k-th time window is... , L The preset window length, k For time window sequence number; The mapping module inputs the mutation feature sequence into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a set of primary capsule vectors, forming the output vector of the primary capsule layer. The iterative module performs iterative calculations between the output vector of the primary capsule layer and the higher-level capsules that represent the system state, updating the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. The fault detection module determines the fault based on the converged vector state of the high-level capsule: if the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault characteristic basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault start signal is output; otherwise, the system is determined to be in normal operation.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the shortcomings of existing technologies by acquiring the discrete sequence of zero-sequence voltage at the distribution network bus, calculating the absolute value of the difference between consecutive sampling points, and performing sliding time window integration. This allows for the accurate capture of abrupt changes in zero-sequence voltage, effectively filtering sampling noise and irrelevant fluctuations, and solving the problem of traditional methods' difficulty in accurately extracting weak transient features of faults. Furthermore, the abrupt feature sequence is input into a pre-trained capsule neural network. The primary capsule layer specifically encodes the features for each time window. Combined with iterative weight updates between the primary and higher-level capsules, the higher-level capsules can fully learn fault-specific features, significantly improving the accuracy of system state representation and avoiding feature encoding distortion and weight imbalance. The method reduces the judgment bias by using a dual threshold of magnitude and angle of the converged high-level capsule vector to determine faults. Compared with a single judgment standard, this method can effectively distinguish between single-phase grounding faults and normal operating conditions, accurately adapt to different fault conditions such as low resistance and high resistance, and significantly reduce the probability of false or missed faults. The entire process does not require complex manual intervention and can quickly complete fault identification and output a fault start signal that conforms to the FTU standard, providing reliable support for subsequent fault location and isolation. At the same time, it retains the iterated vector and weight to ensure the continuity of judgment, which not only improves the accuracy and efficiency of FTU fault identification, but also enhances the practicality and adaptability of the method, and reduces the cost of handling faults in the distribution network. Attached Figure Description

[0017] Figure 1 This is a flowchart of the zero-sequence voltage mutation characteristic ground fault identification method for FTU in Embodiment 1 of the present invention; Figure 2 This is a block diagram of the zero-sequence voltage mutation characteristic ground fault identification system for FTU in Embodiment 2 of the present invention. Detailed Implementation

[0018] Example 1 Please see Figure 1 The present invention provides an embodiment of a method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs, comprising the following steps: S1. Obtain a discrete sequence characterizing the zero-sequence voltage of the distribution network bus using configured electrical sensors. U 0 ( n )},in n The sampling point number is based on the discrete sequence { U 0 ( n )}, calculate the absolute difference between its consecutive sampling points to obtain the absolute difference sequence { U 0 ( n Furthermore, the electrical sensors in this embodiment are specifically configured by those skilled in the art according to data acquisition requirements.

[0019] S2, for the difference absolute value sequence { U 0 ( n Perform sliding time window integration to construct a sequence of mutation feature quantities, where the feature quantity corresponding to the k-th time window is... , L The preset window length, k For time window sequence number; S3. Input the mutation feature sequence into the pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence in each time window into a set of primary capsule vectors to form the output vector of the primary capsule layer. S4. Using a dynamic routing protocol, iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules representing the system state to update the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. S5. Fault determination based on the converged high-level capsule vector state: If the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault characteristic basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault start signal is output; otherwise, the system is determined to be in normal operation.

[0020] It should be further explained that in this embodiment, the differential absolute value sequence is integrated over a sliding time window to construct a mutation feature sequence, including: Based on the obtained discrete sequence characterizing the zero-sequence voltage of the distribution network bus, which is a discrete sequence of zero-sequence voltage, where each element is the zero-sequence voltage value of the corresponding sampling point, the sampling point number starts from 1 and increases sequentially, the sampling frequency is set to 20000 Hz, and the sampling precision is not less than 16 bits, the zero-sequence voltage differential absolute value sequence is obtained. This zero-sequence voltage differential absolute value sequence is a differential absolute value sequence, where each element is the absolute value of the difference between the zero-sequence voltage values ​​of two consecutive sampling points in the zero-sequence voltage discrete sequence. The calculation method is to take the difference between the zero-sequence voltage value of the nth sampling point and the zero-sequence voltage value of the (n-1)th sampling point in the zero-sequence voltage discrete sequence, and then take the absolute value of the difference to obtain the nth element in the differential absolute value sequence, where n starts from 2 and increases sequentially. The length of the differential absolute value sequence is 1 less than the length of the zero-sequence voltage discrete sequence.

[0021] Based on the obtained difference absolute value sequence, boundary processing is performed to obtain a boundary-processed difference absolute value sequence. The processing method combines forward padding and end windowing. Forward padding involves taking the value of the first element in the difference absolute value sequence and padding it to the beginning of the sequence, making the length of the padded sequence one more than the length of the original difference absolute value sequence. End windowing involves determining whether the number of remaining sampling points at the end of the difference absolute value sequence after forward padding is less than a preset initial window length. The preset initial window length is set to 200 sampling points. If the number of remaining sampling points is less than 200, all remaining sampling points at the end are merged with the corresponding number of sampling points before the remaining sampling points in the sequence, making the number of merged sampling points equal to 200, forming an end window segment. If the number of remaining sampling points is not less than 200, end windowing is not performed. Finally, the boundary-processed difference absolute value sequence is obtained, which can completely cover all elements of the original difference absolute value sequence without missing data.

[0022] Based on the acquired boundary-processed absolute difference value sequence, a dynamic window length calculation is performed on the sequence to obtain a sliding window length adapted to the current working condition. First, a 3-point median filtering algorithm is used to denoise the boundary-processed absolute difference value sequence. For each internal element of the boundary-processed absolute difference value sequence, the median of itself and its two adjacent elements (one before and one after) is taken as the filtered value at that position. For the first boundary element of the boundary-processed absolute difference value sequence, the median of itself, itself, and its next adjacent element (one before and one after) is taken as the filtered value at that position. For the last boundary element of the boundary-processed absolute difference value sequence, the median of its previous adjacent element, itself, and itself (one before and one after) is taken as the filtered value at that position, resulting in a filtered absolute difference value sequence. Then, the mean and standard deviation of the filtered absolute difference value sequence are calculated. The ratio of the standard deviation to the mean is defined as the coefficient of variation. The working condition is determined based on the range of the coefficient of variation, and the corresponding basic window length is matched. Specifically, if the coefficient of variation is greater than 2, the condition is determined... The basic operating condition is defined as a low-resistance fault with strong abrupt change, with a basic window length of 200 sampling points. If the coefficient of variation is greater than 1 and less than or equal to 2, it is determined to be a transitional operating condition with moderate abrupt change, with a basic window length of 300 sampling points. If the coefficient of variation is less than or equal to 1, it is determined to be a normal or high-resistance initial operating condition with weak abrupt change, with a basic window length of 400 sampling points. Subsequently, the maximum value of the discrete sequence representing the zero-sequence voltage of the distribution network bus is extracted. Based on the maximum value, the basic window length is used for ultra-high resistance fault classification adaptation. If the maximum value is greater than or equal to 3 volts and... If the voltage is less than 5 volts, it is determined to be a medium-high resistance fault condition. The basic window length is multiplied by 1.2 to obtain the adaptation window length. If the maximum value is less than 3 volts, it is determined to be an ultra-high resistance fault condition. The basic window length is multiplied by 1.5 to obtain the adaptation window length. If the maximum value is greater than or equal to 5 volts, the basic window length is directly used as the adaptation window length. Finally, the adaptation window length is forcibly rounded to an integer number of sampling points to obtain the sliding window length adapted to the current condition. The value range of the sliding window length adapted to the current condition is from 200 sampling points to 600 sampling points.It should be further explained that the determination of the coefficient of variation threshold and the basic window length in this embodiment is based on the statistical analysis of a large amount of historical fault waveform data. Specifically, no less than 1,000 sets of zero-sequence voltage waveforms under different fault types (low resistance, medium resistance, high resistance, arcing) and normal operating conditions are collected. The difference absolute value sequence of each set of data is calculated according to the S1 method and its coefficient of variation is extracted. The coefficient of variation is divided into three significant clusters through cluster analysis (such as K-means). The boundary value of each cluster is taken as the judgment threshold (for example, the center value of the first cluster is 0.8, the center value of the second cluster is 1.5, and the center value of the third cluster is 2.3, so the thresholds are set to 1.15 and 1.9). At the same time, with the fault detection accuracy (not less than 95%) and response time (not more than 20ms) as optimization objectives, traversal tests are carried out under different window lengths (such as 150 to 500 points). The shortest window length that meets the objectives is selected as the basic window length for the corresponding operating condition to ensure that the matching between the threshold and the window length can be verified.

[0023] In this embodiment, when performing high-resistance fault classification and adaptation based on the maximum zero-sequence voltage for the base window length, the voltage thresholds of 3V and 5V are determined according to typical parameters of the distribution network and statistical results of electromagnetic transient simulation. For a 10kV system grounded via an arc suppression coil, the primary phase voltage is 5.77kV, which is converted to the secondary side by a voltage transformer ratio of 10kV / 100V. Different fault resistances (0Ω, 100Ω, 500Ω, 1000Ω, 2000Ω) are set using PSCAD / EMTDC simulation. The transient peak distribution of the zero-sequence voltage on the secondary side under each resistance is statistically analyzed. It is found that when the fault resistance is ≥1000Ω, the peak value is generally lower than 3V, corresponding to an ultra-high resistance fault; when the resistance is ≥30Ω, the peak value is lower than 3V. When the resistance is between 0Ω and 1000Ω, the peak value is concentrated between 3V and 5V, corresponding to medium-high resistance faults. When the resistance is below 300Ω, the peak value usually exceeds 5V, corresponding to low resistance faults. Thus, 3V and 5V are determined as the dividing points for the classification. The window length adjustment coefficients of 1.2 and 1.5 are determined based on the balance test between detection sensitivity and noise immunity. Taking the ultra-high resistance fault of 2000Ω as an example, the sliding integral is performed at the basic window length of 400. If the signal-to-noise ratio of the fault feature is lower than 3dB, the window length is gradually increased until the signal-to-noise ratio is not lower than 6dB. The window length multiple at this time is recorded as the basis for the value of 1.5. Similarly, the test is performed on the medium-high resistance fault of 500Ω to obtain a window length of 1.2, to ensure the reliability of feature extraction under different fault resistances.

[0024] Based on the obtained sliding window length adapted to the current working condition, the sliding step size of the sliding time window is determined, and the sliding step size parameter is obtained. The sliding step size is calculated by taking half of the sliding window length and rounding down to obtain the integer number of sampling points. If the sliding window length is odd, the sliding step size is half of the sliding window length minus 1, ensuring that the sliding step size is an integer number of sampling points. The value range of the sliding step size is from 100 to 300 sampling points. The overlap ratio of two adjacent sliding time windows is 50% to ensure that the fault change signal does not fall in the gap between two sliding time windows, while balancing the time resolution and computational efficiency of feature extraction.

[0025] Based on the acquired boundary-processed absolute difference sequence, the sliding window length adapted to the current working condition, and the sliding step size parameter, a trapezoidal integral algorithm is used to perform sliding time window integration on the boundary-processed absolute difference sequence to obtain the initial mutation feature quantity sequence. In this embodiment, the integration process starts from the second element of the boundary-processed absolute difference sequence, and sequentially extracts subsequences of length equal to the sliding window length adapted to the current working condition according to the sliding step size as the current sliding time window. Each current sliding time window corresponds to a time window number that increments sequentially from 1. Trapezoidal integral calculation is performed on the subsequence within each current sliding time window, first calculating the average of the absolute differences between the first and last sampling points of the subsequence within the current sliding time window. The value is then summed with the absolute difference of all intermediate sampling points in the subsequence within the current sliding time window, excluding the first and last sampling points. The sum of the average value and the summed value is multiplied by the sampling interval to obtain the initial mutation feature corresponding to the current sliding time window. The sampling interval is the reciprocal of the sampling frequency, the sampling frequency is 20000 Hz, and the corresponding sampling interval is 50 microseconds. The unit of the initial mutation feature is volt. All the initial mutation feature corresponding to the current sliding time window are arranged sequentially according to the time window number to form the initial mutation feature sequence. The length of the initial mutation feature sequence is the length of the absolute difference sequence after boundary processing minus the length of the sliding window adapted to the current working condition, divided by the sliding step size, plus 1, and then rounded down to obtain the integer length.

[0026] Based on the obtained initial mutation feature sequence, outlier removal is performed to obtain the outlier-free mutation feature sequence. The outlier removal uses the 3σ criterion algorithm, which is implemented by calculating the mean and standard deviation of the initial mutation feature sequence, setting the outlier judgment threshold to the mean plus or minus 3 times the standard deviation, and traversing each element in the initial mutation feature sequence. If the value of an element exceeds the range of the mean plus or minus 3 times the standard deviation, the element is determined to be an outlier. The outlier is replaced by the mean of the three adjacent elements before and after the element. If the element is the beginning of the sequence, the mean of the three adjacent elements after the element is used as the replacement. If the element is the end of the sequence, the mean of the three adjacent elements before the element is used as the replacement. After traversal, the outlier-free mutation feature sequence is obtained. This sequence has no isolated outliers and retains the effective features of the initial mutation features.

[0027] Based on the obtained outlier-free mutation feature sequence, the max-min normalization algorithm is used to normalize the sequence, obtaining a standardized mutation feature sequence. This standardized mutation feature sequence is the final mutation feature sequence used as input to the capsule neural network model. The max-min normalization algorithm is implemented by first extracting the features of the normal operating period from the outlier-free mutation feature sequence. The normal operating period is determined by selecting the features corresponding to the first 10 minutes of the outlier-free mutation feature sequence, calculating the minimum and maximum values ​​of the features within this period, and using them as the normalization benchmark. If the difference between the maximum and minimum values ​​is less than 1e- If 6 is true, then all features are normalized to 0. Otherwise, subtract the minimum value from each element in the mutation feature sequence after removing outliers, and then divide by the difference between the maximum and minimum values ​​to obtain the normalized features. At the same time, the normalized features are truncated to the range of 0 to 1. Features outside this range are treated as boundary values, that is, features less than 0 are taken as 0, and features greater than 1 are taken as 1. All normalized features are arranged in order according to the original time window number to form a standardized mutation feature sequence. The value range of each element in this sequence is 0 to 1. Features under different voltage levels and different operating conditions are comparable and can be directly used as input to the capsule neural network model.

[0028] It should be further explained that the reason why this embodiment uses the aforementioned convolutional coding network, which includes multi-scale feature extraction, depth adaptive control, noise suppression, and dynamic dimension adjustment, to construct the primary capsule vector is fundamentally to solve the core problem that a single feature extraction model is difficult to adapt to the complex transient characteristics of single-phase grounding faults in distribution networks. Specifically: existing coding does not adapt to the multi-scale characteristics of fault transient signals, resulting in incomplete feature extraction; fixed convolutional kernel parameters cannot adapt to the feature differences under low resistance, high resistance, and normal operating conditions, resulting in poor coding adaptability; the lack of an effective noise suppression mechanism in the coding process leads to residual noise introducing coding redundancy that interferes with subsequent weight iterations; and fixed capsule vector dimensions cannot match the dynamic changes in input data length, resulting in an imbalance between feature expression capability and computational efficiency. Moreover, these problems are interconnected. Poor operating condition adaptability will exacerbate the problem of incomplete feature extraction; noise redundancy will further interfere with feature expression and weight iteration; and fixed dimensions will prevent the defects caused by the above problems from being compensated for by dynamic adjustment. Therefore, this embodiment generates the primary capsule vector based on a standardized mutation feature sequence through the primary capsule layer of the capsule neural network model. The primary capsule layer is implemented through a preset convolutional coding network, including: S301. Divide the standardized mutation feature sequence into multiple consecutive feature subsequences according to a preset non-overlapping partitioning rule; wherein each feature subsequence contains a preset number P consecutive standardized mutation features, where P is an integer greater than 1, and there is no data overlap between adjacent feature subsequences. S302. Input the multiple consecutive feature sub-sequences into the preset working condition feature perception layer and call the pre-configured three parallel convolutional layers to obtain the first feature map, the second feature map and the third feature map with working condition features. It should be further explained that in this embodiment, the multiple consecutive feature sub-sequences are input into a preset working condition feature perception layer, specifically: In this embodiment, the working condition feature perception layer receives the multiple consecutive feature subsequences and calculates the feature statistics of the multiple consecutive feature subsequences; the feature statistics include mean, standard deviation, and maximum amplitude. F 0max and coefficient of variation C v ; Based on characteristic statistics and by setting a first or second judgment condition, the fault condition type is obtained, specifically: When the first condition is met, that is, A max >0.5% F >0.2 This is determined to be a low-resistance fault basic operating condition. When the second criterion is met, that is, if 0.3 ≤A max ≤0.5 and 0.1≤ s F ≤0.2, judged as a medium resistance fault basic operating condition; When the third criterion is met, that is, if A max <0.3 and σ F <0.1 The condition was determined to be an ultra-high resistance fault. If the first, second, and third judgment conditions are not met, the condition is judged to be normal or transitional. The three parallel convolutional layers pre-configured in Table 1 are invoked as follows: Based on the final determined working condition type, the parameters of the pre-configured three-layer parallel convolutional layer are invoked, as shown in Table 1: If the basic operating condition is determined to be a low-resistance fault, the first set of parameters is called: the first convolutional layer uses 32 one-dimensional convolutional kernels with a size of 2, the second convolutional layer uses 64 one-dimensional convolutional kernels with a size of 2, and the third convolutional layer uses 128 one-dimensional convolutional kernels with a size of 2. If the basic operating condition is determined to be a medium resistance fault, the fourth set of parameters is called: the first convolutional layer uses 24 one-dimensional convolutional kernels with a size of 3, the second convolutional layer uses 48 one-dimensional convolutional kernels with a size of 3, and the third convolutional layer uses 96 one-dimensional convolutional kernels with a size of 3. If the basic operating condition is determined to be an ultra-high resistance fault, the second set of parameters is called: the first convolutional layer uses 8 one-dimensional convolutional kernels with a size of 4, the second convolutional layer uses 16 one-dimensional convolutional kernels with a size of 4, and the third convolutional layer uses 32 one-dimensional convolutional kernels with a size of 4. If the condition is determined to be normal or transitional, the third set of parameters is called: the first convolutional layer uses 16 one-dimensional convolutional kernels of size 3, the second convolutional layer uses 32 one-dimensional convolutional kernels of size 3, and the third convolutional layer uses 64 one-dimensional convolutional kernels of size 3.

[0029] Table 1: Convolutional Layer Parameter Table S303. The first layer feature map, the second layer feature map and the third layer feature map with working condition features are concatenated in the channel dimension to generate a fused feature map corresponding to the feature subsequence. S304. Based on the fused feature map, the depth control module performs the following operations to generate a depth-coded feature vector: Extract the original feature subsequence corresponding to the current fused feature map; Calculate the kurtosis and skewness statistical features of the original feature subsequence; The kurtosis statistical feature value is compared with a third preset threshold of 0.5, and the skewness statistical feature value is compared with a fourth preset threshold of 0.3. If the kurtosis statistical feature value is greater than a third preset threshold and the skewness statistical feature value is greater than a fourth preset threshold, then the fused feature map is input to the deep processing path to obtain a deep encoded feature vector; the deep processing path consists of two concatenated convolutional processing units, each convolutional processing unit including: a one-dimensional convolutional layer configured with 16 convolutional kernels of size 3 and a stride of 1; a batch normalization layer; and a ReLU activation function layer; If the third or fourth preset threshold is not met, the fused feature map is input to the shallow processing path to obtain the deep encoded feature vector. The shallow processing path is composed of one of the convolutional processing units. In this embodiment, the depth control module adaptively determines the depth path for feature extraction based on the statistical characteristics of the original feature subsequences corresponding to the fused feature map. The specific construction logic is as follows: First, the original feature subsequences corresponding to the current fused feature map are extracted and their kurtosis and skewness statistical features are calculated. For example, assuming the kurtosis of a certain feature subsequence in the current processing batch is 0.6 and the skewness is 0.4; then, these two statistical values ​​are compared with preset third thresholds of 0.5 and fourth thresholds of 0.3, respectively. If the conditions of kurtosis greater than 0.5 and skewness greater than 0.3 are met, it is determined that the subsequence contains rich nonlinear mutation information and requires deeper feature mining. At this time, the fused feature map is input to the deep processing path, which consists of two... The system consists of a series of convolutional processing units. Each unit contains a convolutional layer with 16 one-dimensional convolutional kernels of size 3 and a stride of 1, a batch normalization layer, and a ReLU activation function layer. This combination process is repeated twice to obtain a deep-encoded feature vector rich in deep abstract features. Conversely, if the kurtosis is no greater than 0.5 or the skewness is no greater than 0.3, the features are considered relatively flat. In this case, the fused feature map is input to a shallow processing path consisting of only one convolutional processing unit with the same structure to generate a deep-encoded feature vector. Thus, the depth control module achieves hierarchical fine-grained encoding of abrupt features under different operating conditions through a path selection mechanism driven by statistical features, providing a high-quality input feature vector for subsequent noise suppression processing. It should be further explained that in this embodiment, the determination of kurtosis and skewness thresholds in the deep control module is based on the nonlinear feature quantification analysis of a large number of fault waveforms. The specific implementation steps are as follows: First, a sample library is constructed by collecting standardized mutation feature sequences under different operating conditions of the distribution network. For each sample's feature subsequence, the kurtosis statistical feature value and skewness statistical feature value are calculated, and experts label whether the sample needs deep feature extraction. Then, kurtosis and skewness are used as input features, and the labeling results are used as labels to train a logistic regression classifier. The optimal decision boundary is determined through the classifier. In this embodiment, the boundary corresponds to a kurtosis of approximately 0.48 and a skewness of approximately 0.28. To allow for margins and simplify calculations, the third preset threshold of 0.5 and the fourth preset threshold of 0.3 are rounded down. Meanwhile, these thresholds support online learning and updates. If, during the FTU operation, it is found that the current path selection causes the confidence level of subsequent fault determination to be lower than the preset threshold of 0.85, the threshold fine-tuning mechanism is triggered, increasing or decreasing the threshold in steps of 0.01 until the confidence level is restored. This ensures that the depth control module can always adaptively select deep or shallow processing paths based on the statistical characteristics of the feature subsequences.

[0030] S305. Perform noise suppression processing on the deep encoded feature vector to obtain the suppressed feature vector, specifically as follows: Determine the original feature subsequence corresponding to the deep encoded feature vector, and define the original feature subsequence as the current processing subsequence; S306. According to the division order of the standardized mutation feature sequence, obtain the adjacent feature subsequence that is located before the current processing subsequence in time as the preceding subsequence, and obtain the adjacent feature subsequence that is located after the current processing subsequence in time as the following subsequence. S307. Combine all normalized mutation features contained in the current processing subsequence, the preceding subsequence, and the subsequent subsequence to form a merged sequence; S308. Calculate the variance of all elements in the merged sequence, and use the value of the variance as the noise evaluation value of the current processing; S309. Calculate the median of the noise evaluation values ​​corresponding to all feature subsequences in the current processing batch, and use the median as the benchmark noise value; It should be further explained that in this embodiment, the median of the noise evaluation value sequence is calculated and determined as the reference noise value, specifically as follows: Obtain the M feature subsequences contained in the current processing batch, where M is an integer greater than 1; The noise evaluation value corresponding to each of the M feature subsequences is extracted sequentially, resulting in a total of M noise evaluation values, forming a noise evaluation value sequence; Sort all values ​​in the noise assessment value sequence in ascending order; If M is odd, then the value at the (M+1) / 2th position in the sorted sequence is taken as the median; If M is even, the median is the arithmetic mean of the value at the M / 2th position and the value at the M / 2+1th position in the sorted sequence. The median obtained from the final calculation is set as the reference noise value.

[0031] S310, Calculate the dynamic attenuation coefficient α , =min(1, ), where T is the median of the noise evaluation values ​​of all sequence segments processed in the current batch, and V is the noise evaluation value of the current sequence segment; it should be further noted that in this embodiment, the dynamic attenuation coefficient is used in the noise suppression processing. α The design follows the principle that the lower the noise level, the more complete the feature preservation, and the specific calculation method is as follows: αThe minimum value between 1 and T / V is taken, where V is the noise evaluation value (variance) of the current merged sequence, and T is the median of the noise evaluation values ​​of all feature subsequences in the current processing batch. The median is used as the benchmark because it is not sensitive to extreme outliers and can stably represent the typical noise level of the current batch. When V is greater than T, it indicates that the noise level of the current sequence is higher than the typical value and there may be interference. It is necessary to suppress the deep-coded feature vector by the ratio of T / V. When V is less than T, it indicates that the noise level is lower than the typical value and the feature quality is high. Therefore, the coefficient is taken as 1 to completely preserve the features. This formula dynamically adjusts the attenuation intensity by the ratio of the median to the current variance, so that the degree of noise suppression matches the relative noise level of the current segment, thereby suppressing abnormal fluctuation interference while preserving effective mutation features.

[0032] S311. Multiply each vector element in the deep encoded feature vector by the dynamic attenuation coefficient to obtain the suppressed feature vector; S312. Convert the suppressed feature vector into a primary capsule vector, specifically including: Obtain the total number of time windows N contained in the standardized mutation feature sequence; The target capsule dimension D is dynamically determined based on the total number of time windows N. The specific rules are as follows: when N is less than or equal to 100, D is set to 8; when N is greater than 100 and less than or equal to 200, D is set to 10; when N is greater than 200, D is set to 12. The suppressed feature vector is input into a fully connected projection layer, which maps the dimension of the input vector to the dimension D of the target capsule. The output vector of the fully connected projection layer is normalized using the L2 norm, and the normalized vector is used as a primary capsule vector. It should be further explained that in the dynamic adjustment of capsule dimensions in this embodiment, the determination of the target capsule dimension D needs to match the total number of time windows N of the standardized mutation feature sequence. The specific grading rules are based on the principle of balancing feature representation capability and computational efficiency, and have been verified through experiments. First, according to the principles of information theory, the capsule dimension D should be able to encode the main mutation information of the input sequence. Considering that the length P of each feature subsequence is fixed at 50, the theoretical range of D is usually set to 1 / 5 to 1 / 3 of P, corresponding to approximately 10. Based on this, historical data is divided into three groups according to the value of N: N≤100, 100<N≤200, and N>200. Within each group, the fault classification accuracy and single forward propagation time are tested under different D values. The optimal dimension within each group was determined with the goal of balancing classification accuracy and computational cost. Experimental results show that when N≤100, the amount of input data is small, and D=8 is sufficient to meet the feature encoding requirements. Increasing the dimension further has limited effect on improving classification performance and increases computational burden. When 100<N≤200, the amount of input data is moderate, and D=10 can achieve the best balance between feature representation and computational efficiency. When N>200, the amount of data is abundant, and D=12 can fully encode detailed features while the computation time is still within the allowable range of FTU. The selection of the dividing points 100 and 200 is based on the correspondence with the duration of typical fault recordings, ensuring that the dimension transitions smoothly with the length of the input sequence, thereby achieving adaptive feature encoding for input data of different lengths.

[0033] S313. Divide the standardized mutation feature sequence into segments of a preset fixed length, such that each segmented feature subsequence contains 50 consecutive standardized mutation features, and there is no data overlap between any two adjacent feature subsequences. It should be further noted that in this embodiment, during the construction of the primary capsule layer, the determination of the length P of the feature subsequence needs to match the fault transient characteristics and sampling frequency. Specifically, the implementation method is as follows: First, statistically analyze the main frequency component of a typical single-phase ground fault transient process, which is usually in the range of 300Hz to 3000Hz. Taking a main frequency of 1000Hz as an example, its period is 1ms. At a sampling frequency of 20000Hz, a single period corresponds to 20 sampling points. To ensure... To ensure that each feature subsequence can cover at least half a cycle of transient fluctuations, the base length is set to 2 to 3 times the number of points corresponding to the cycle, i.e., 40 to 60 points. The classification accuracy is tested on the validation set with P values ​​of 30, 40, 50, 60, and 70 through ablation experiments. The results show that the classification accuracy reaches 98.2% with moderate computational cost when P equals 50. Therefore, the preset length is determined to be 50 consecutive standardized mutation features. If the sampling frequency or system main frequency changes, the P value is dynamically adjusted according to the following rule: P equals the sampling frequency divided by the expected main frequency, multiplied by 2.5, rounded to the nearest integer, and rounded to a multiple of 10 to ensure that the feature subsequence length always matches the fault transient duration and the network receptive field.

[0034] S314. For each feature subsequence obtained by the division, the primary capsule vector acquisition process is executed sequentially to generate a primary capsule vector corresponding to the corresponding feature subsequence. S315. Arrange and combine all the generated primary capsule vectors according to the chronological order of their corresponding feature subsequences in the standardized mutation feature sequence to form the output vector of the primary capsule layer.

[0035] For example, the specific implementation example of this embodiment in the single-phase grounding fault identification scenario of a 10kV distribution network feeder terminal is as follows: Assuming that the acquired standardized abrupt change feature quantity sequence contains feature data of 150 time windows, it is divided into 3 feature subsequences, each containing 50 consecutive standardized abrupt change feature quantities, according to the non-overlapping rule. After inputting these 3 feature subsequences into the operating condition feature perception layer, their mean, standard deviation, and maximum amplitude are calculated. It is determined that the maximum amplitude is 0.6 and the standard deviation is 0.25, which meets the basic operating condition judgment conditions for low resistance faults. Then, the system is called... The first set of parameters in the three parallel convolutional layers processes the three feature sub-sequences separately, resulting in a three-layer feature map with basic low-resistance fault characteristics. These three feature maps are then concatenated along the channel dimension to generate a fused feature map corresponding to each feature sub-sequence. The original feature sub-sequences corresponding to each fused feature map are extracted. Two of these sub-sequences have a kurtosis of 0.6 and a skewness of 0.4, both exceeding their respective preset thresholds. Their corresponding fused feature maps are then input into the deep processing path to generate a deep-encoded feature vector. The other feature sub-sequence has a kurtosis of 0.4 and a skewness of... The degree is 0.2, which does not meet the threshold requirement. A shallow processing path is input to generate a deep-encoded feature vector. For each deep-encoded feature vector, its corresponding current processing subsequence is determined. The preceding and following subsequences of each subsequence are obtained and merged into a merged sequence. The variance of each merged sequence is calculated as a noise evaluation value. Simultaneously, the median of the three noise evaluation values ​​for the current batch is calculated as the baseline noise value. Based on this, the dynamic attenuation coefficient corresponding to each deep-encoded feature vector is calculated. Each element of each deep-encoded feature vector is multiplied by the corresponding dynamic attenuation coefficient to obtain the suppressed feature vector. Since the total number of time windows is 150, which is between 100 and 200, the target capsule dimension is determined to be 10. The three suppressed feature vectors are input into a fully connected projection layer, mapped to 10-dimensional vectors, and normalized using the L2 norm to obtain three primary capsule vectors. These three primary capsule vectors are arranged and combined according to the temporal order of their corresponding feature subsequences in the standardized mutation feature quantity sequence, ultimately forming the output vector of the primary capsule layer. This provides support for subsequent high-level capsule iterations and accurate single-phase grounding fault determination.

[0036] In this embodiment, during the weight iteration calculation process, the initial connection weights lack consideration for the inherent characteristics of the primary capsule vectors (such as magnitude and dynamic dimension), leading to a deviation in the iteration starting point. This deviation is amplified in subsequent iterations due to the lack of a working condition adaptation mechanism, preventing differentiated correction for different fault characteristics (such as low resistance / high resistance), and further amplified by the lack of a residual noise suppression mechanism. A fixed and unreasonable iteration count causes the deviation correction process to be either insufficient (high resistance condition) or excessive (low resistance condition), exacerbating the contradiction between convergence efficiency and accuracy. Furthermore, a single convergence criterion, not linked to the state representation of higher-level capsules, cannot accurately assess the actual effect of weight correction, easily leading to "pseudo-convergence" or "over-iteration." All these factors ultimately result in a severe imbalance in weight allocation, diluting the weights corresponding to core fault characteristics, and preventing higher-level capsules from accurately mapping the system state. Therefore, it needs to be further explained that this embodiment performs iterative calculations between the output vector of the primary capsule layer and the higher-level capsules representing the system state, updating the connection weights between the primary capsule vector and each higher-level capsule, including: S401. Based on the output vector of the primary capsule layer, initialize the connection weights between each higher-level capsule and all primary capsule vectors, specifically as follows: Obtain each primary capsule vector contained in the primary capsule layer output vector; Calculate the magnitude of each of the primary capsule vectors; Linearly normalize the magnitudes of all primary capsule vectors so that the sum of the normalized magnitudes equals the total number of higher-level capsules. The magnitude of each primary capsule vector is directly set as the initial connection weight between the corresponding primary capsule vector and each higher-level capsule; It should be further explained that in this embodiment, during the weight initialization process, the magnitude of the primary capsule vector is linearly normalized and directly used as its initial connection weight with each higher-level capsule. The rationality of this design lies in the positive correlation between the unbiasedness of the initialization strategy and the saliency of the features: at the beginning of the dynamic routing iteration, the model has no prior information about which higher-level capsule the primary capsule vector should belong to. Therefore, using a uniform distribution method to connect all primary capsule vectors to each higher-level capsule with the same initial weight avoids artificially introduced bias that could lead to local optima in subsequent iterations; simultaneously, the primary capsule vector... The modulus itself represents the probability of existence or significance of the corresponding feature. The larger the modulus, the more significant the feature, and its initial contribution in constructing all high-level capsule representations should be greater. Therefore, using the normalized modulus directly as the weight base is consistent with the intuition that significant features should participate more in the construction of high-level representations. In subsequent iterations, the dynamic routing protocol will gradually concentrate the connection weights on the high-level capsules with higher matching degrees based on the similarity between each primary capsule vector and the high-level capsule prediction vector, thereby correcting the initial uniform distribution and making the final weight distribution accurately reflect the true membership relationship between primary capsules and high-level capsules.

[0037] S402. Based on the initial connection weights and the output vector of the primary capsule layer, perform an iterative calculation process. A single iteration includes the following implementation steps: S4021. Calculate the prediction vector of the high-level capsule based on the current connection weights. Specifically, map each primary capsule vector through a pre-trained linear transformation matrix to obtain the prediction vector of the primary capsule vector for each high-level capsule. It should be further noted that the pre-trained linear transformation matrix in this embodiment is learned through the model optimization algorithm during the overall training phase of the capsule neural network model. Specifically, during the model training process, a training dataset containing a large number of historical fault samples and normal samples is input into the untrained capsule neural network. The model calculates the prediction from the primary capsule vector to the high-level capsule through forward propagation, and based on the classification loss between the model output and the true label, iteratively updates each parameter element in the linear transformation matrix using the backpropagation algorithm until the model loss converges to a preset range. After training, all parameter values ​​of the linear transformation matrix are fixed and used for mapping operations in the subsequent inference process.

[0038] S4022. For the same high-level capsule, multiply the prediction vectors of all primary capsule vectors by the current connection weights between the corresponding primary capsule vectors and the high-level capsule, and sum the weighted prediction vectors to obtain the prediction vector of the high-level capsule in the current iteration round. S4023. Obtain the vector of each primary capsule and the prediction vector of each high-level capsule in the current iteration round; S4024. Calculate the vector dot product between each primary capsule vector and each high-level capsule prediction vector, and define the calculation result as the first similarity metric. S4025. For each high-level capsule, perform an update operation: add the current connection weight between the high-level capsule and each primary capsule vector to the first similarity metric value corresponding to the primary capsule vector and the high-level capsule to obtain a set of original update values. S4026. For all original updated values ​​belonging to the same high-level capsule, apply the Softmax function to perform normalization calculation to obtain the updated connection weights. S4027. For each high-level capsule, obtain the prediction vector of all primary capsule vectors for that high-level capsule; S4028. Multiply each of the prediction vectors by the updated connection weight between the corresponding primary capsule vector and the higher-level capsule to obtain a weighted prediction vector. S4029. Sum all weighted prediction vectors to obtain the weighted sum vector of the high-level capsule; Apply a nonlinear compression function to the weighted sum vector: calculate the magnitude of the weighted sum vector, and adjust the values ​​of each component of the vector based on the magnitude using the compression function so that the magnitude of the output vector is no greater than 1; The compressed vector is used as the high-level capsule vector for the current iteration round.

[0039] S403. After updating the connection weights in each iteration (S4021 to S4026), weight sparsification is performed, specifically as follows: Obtain all connection weights after the current iteration round update. The connection weights form a weight matrix, where each weight element corresponds to the connection strength between a primary capsule vector and a higher-level capsule. Based on the output vector of the primary capsule layer, the magnitude of each primary capsule vector is calculated to obtain a set of magnitude values; Calculate the arithmetic mean of the modulus values ​​of all primary capsule vectors in the current processing batch, and set the arithmetic mean as the modulus threshold; Traverse all primary capsule vectors, compare their modulus values ​​with the modulus threshold, and identify all primary capsule vectors whose modulus values ​​are lower than the modulus threshold; For each primary capsule vector whose modulus value is below a threshold, the connection weight element value between it and all higher-level capsules is uniformly multiplied by a preset attenuation coefficient; the attenuation coefficient is set to 0.1, thereby suppressing the contribution of primary capsule vectors with insignificant features to higher-level capsules. S404. After each iteration of calculation, dynamically determine the convergence state of the connection weights, specifically including: Calculate the change in all connection weights between the current iteration and the previous iteration, which is the absolute value of the difference between the corresponding weight elements, and sum all the absolute values ​​to define the total weight change. If the total weight change is less than the first convergence threshold, then the first convergence condition is satisfied.

[0040] Calculate the magnitude of the vector of each high-level capsule in the current iteration round; Calculate the absolute value of the difference between the magnitudes of all corresponding high-level capsule vectors in the current iteration and the previous iteration, and sum all the absolute values, which is defined as the total magnitude change. If the change in the total modulus is less than the second convergence threshold, then the second convergence condition is satisfied.

[0041] Obtain the high-level capsule vector representing the fault state and the high-level capsule vector representing the normal state in the current iteration round; The difference between the magnitude of the high-level capsule vector in the fault state and the magnitude of the high-level capsule vector in the normal state is calculated and defined as the state magnitude difference. If the difference in state magnitude is greater than the third convergence threshold, then the third convergence condition is satisfied.

[0042] If the first convergence condition, the second convergence condition, and the third convergence condition are all satisfied after one iteration, the connection weights are determined to have converged, and the iteration process is terminated.

[0043] If the number of iterations has reached the preset maximum number of iterations (five) and all convergence conditions are still not met, the iteration process will be forcibly terminated.

[0044] It should be further explained that the convergence threshold settings for each operating condition in this embodiment include: for the low-resistance fault basic operating condition, the first convergence threshold (total weight change) is set to 0.008, the second convergence threshold (total modulus change) is set to 0.004, and the third convergence threshold (state modulus difference) is set to 0.35, in order to strictly constrain rapidly changing transient characteristics; for the medium-resistance fault basic operating condition, because the complexity of its fault characteristics is between low-resistance and high-resistance, the first convergence threshold is set to 0.01, and the second convergence threshold is set to... The first convergence threshold is set to 0.005, and the second convergence threshold is set to 0.3 to balance convergence accuracy and speed. For the basic operating condition of ultra-high resistance faults, the first convergence threshold is set to 0.012, the second convergence threshold is set to 0.006, and the third convergence threshold is set to 0.25, appropriately relaxing the convergence conditions to accommodate weak and gradual changes. For normal or transitional operating conditions, the first convergence threshold is set to 0.01, the second convergence threshold is set to 0.005, and the third convergence threshold is set to 0.3 to ensure that the weight iteration remains stable under normal operating conditions. The above threshold settings ensure that the weight iteration can converge accurately under different operating conditions. This embodiment also adds an "iteration efficiency judgment": if the total weight change decreases by less than 1% in two consecutive iterations, the iteration is terminated early.

[0045] It should be further explained that in this embodiment, the differentiated settings of the first, second, and third convergence thresholds in the weight convergence determination are determined based on a combination of Monte Carlo simulation and field data verification. First, a distribution network simulation model is established, incorporating different fault resistances, initial phase angles, and distributed generation penetration rates, generating a massive amount of training samples covering low-resistance faults, medium-resistance faults, ultra-high-resistance faults, and normal operating conditions. For each operating condition, the weight iteration process is run independently, recording the total weight change, total modulus change, and state modulus difference when correct convergence occurs in each iteration (i.e., the determination result matches the true label). The distribution characteristics of each indicator under this operating condition are statistically analyzed, and the 85th quantile of each indicator's distribution is taken as the basic threshold for that operating condition; for example, the 85th quantile of the total weight change under a low-resistance fault is 0. The threshold is rounded down from 0.0076 to 0.008, the 85th percentile of the total modulus change is 0.0038, rounded down to 0.004, and the 85th percentile of the state modulus difference is 0.34, rounded down to 0.35. For medium-resistance faults, whose characteristics are between low and high resistance, the threshold is set to the middle value. For ultra-high resistance faults, whose characteristics are weak, the 85th percentile of the state modulus difference is approximately 0.24, so a value of 0.25 is used to lower the convergence threshold. Under normal operating conditions, the fault state capsule modulus is extremely small, and the state modulus difference is usually large; a value of 0.3 is used to ensure that the fault state is not mistakenly converged to during normal operation. Finally, the thresholds for each operating condition are verified and fine-tuned using field fault waveform recording data, and an adjustment margin of ±10% is set so that the feeder terminal can adaptively learn and update the thresholds based on local historical data, ensuring the adaptability of the thresholds under different system parameters.

[0046] S405. When the iteration process terminates according to the decision in S404, output the final updated connection weights and the final high-level capsule vector as the basis for fault state determination.

[0047] For example, in the scenario of single-phase grounding fault identification in a 10kV distribution network, the specific implementation example of weighted iterative calculation is as follows: Assume that after primary capsule layer encoding, the output vector contains three primary capsule vectors. These three primary capsule vectors are generated by encoding feature subsequences divided by a standardized mutation feature quantity sequence, corresponding to fault transient or normal operating condition features of the distribution network at different time series. Two primary capsule vectors carry low-resistance fault basic operating condition-related features, and one primary capsule vector carries normal operating condition features. The dimension of all primary capsule vectors is dynamically set to 10 dimensions based on the number of time windows in the standardized mutation feature quantity sequence. The higher-level capsules fixedly contain two core capsules: fault state and normal state, specifically used to characterize the two system states of single-phase grounding fault and normal operation in the distribution network. First, based on these three primary capsule vectors, the magnitude of each primary capsule vector is calculated. The larger the magnitude, the more significant the corresponding feature. The magnitude of the primary capsule vector carrying fault-related features is significantly higher than the magnitude of the primary capsule vector carrying normal operating condition features. Subsequently, the magnitudes of all primary capsule vectors are linearly normalized so that the sum of the normalized magnitudes equals the total number of high-level capsules, which is 2. The normalized magnitudes of each primary capsule vector are then directly used as the initial connection weights between the corresponding primary capsule vector and the two high-level capsules, thus avoiding initial convergence bias caused by random initialization. Iterative computation is then initiated based on these initial connection weights and the output vectors of the primary capsule layer. In each iteration, each primary capsule vector is first mapped to a prediction vector for the two high-level capsules using a pre-trained linear transformation matrix. Primary capsule vectors carrying fault characteristics have a higher similarity to the prediction vectors of the fault-state high-level capsules. Each prediction vector is then multiplied by the current connection weight between the corresponding primary capsule vector and the high-level capsule. Finally, all weighted prediction vectors are summed to obtain the prediction vectors for the two high-level capsules in the current iteration. Next, the vector dot product of each primary capsule vector and its corresponding high-level capsule prediction vector is calculated. This result is used as the first similarity metric, and the current connection weights are updated based on this similarity metric. After the update, the Softmax function is applied to normalize all the original updated values ​​belonging to the same high-level capsule to obtain the updated connection weights. After each connection weight update, weight sparsity processing is immediately implemented. The arithmetic mean of the magnitudes of the three primary capsule vectors in the current batch is calculated, and this arithmetic mean is set as the magnitude threshold. All primary capsule vectors are traversed, and their magnitude values ​​are compared with this magnitude threshold. Primary capsule vectors with magnitude values ​​lower than this threshold are identified. These primary capsule vectors are those carrying normal operating condition features. All connection weight elements between these vectors and the two high-level capsules are multiplied by a preset attenuation coefficient, which is set to 0.1, to suppress the contribution of primary capsule vectors with insignificant features to the high-level capsules.After each iteration, the convergence state of the connection weights is dynamically determined. The change in all connection weights between the current iteration and the previous iteration is calculated, i.e., the absolute value of the difference between the corresponding weight elements. The sum of all absolute values ​​is then obtained to obtain the total weight change. The magnitude of each high-level capsule vector in the current iteration is calculated, and the absolute value of the difference between the magnitudes of all corresponding high-level capsule vectors in the current iteration and the previous iteration is then calculated. The sum of all absolute values ​​is then obtained to obtain the total magnitude change. The high-level capsule vectors representing the fault state and the high-level capsule vectors representing the normal state in the current iteration are obtained, and the difference between the magnitudes of the high-level capsule vectors representing the fault state and the high-level capsule vectors representing the normal state is calculated to obtain the state magnitude difference. If the total weight change is less than 0.01, the total magnitude change is less than 0.005, and the state magnitude difference is greater than 0.3, and these three conditions are met simultaneously, the connection weights are determined to have converged, and the iteration process is terminated. These three thresholds all meet the accuracy requirements for single-phase grounding fault identification in 10kV distribution networks. If the iteration rounds reach the preset maximum of 5 iterations and all convergence conditions are still not met, the iteration process is forcibly terminated. After the iteration process terminates, the final updated connection weights and the final two high-level capsule vectors are output, which serve as the core basis for the accurate determination of single-phase grounding faults in the subsequent 10kV distribution network, ensuring that the basic operating conditions of low-resistance faults, ultra-high-resistance faults, and normal operating conditions can be effectively distinguished.

[0048] It should be further explained that this embodiment performs single-phase grounding fault determination and outputs a fault initiation signal based on the high-level capsule vector state after weight iteration convergence. The specific implementation logic is as follows: S501. Based on the final high-level capsule vector output after the weight iteration terminates, obtain all high-level capsules representing the system state. The high-level capsules include at least high-level capsules representing the ground fault state and high-level capsules representing the normal operation state. The high-level capsules representing the ground fault state are pre-set high-level capsules specifically used to map the characteristics of single-phase ground faults in the distribution network. Their vector characteristics have a fixed correspondence with the standardized mutation characteristic sequence of single-phase ground faults in the distribution network. The final high-level capsule vector is a stable vector obtained after dynamic routing protocol iterative calculation, weight sparsification processing, and convergence determination. Its vector characteristics accurately map the system state characteristics in the primary capsule vector. S502. Extract the final high-level capsule vector corresponding to the high-level capsule representing the grounding fault state, and define it as the fault state capsule vector. Calculate the magnitude of the fault state capsule vector, and define it as the fault capsule magnitude. The calculation of the fault capsule magnitude adopts the L2 norm calculation method, that is, after performing the sum of squares on each dimension component of the fault state capsule vector, the arithmetic square root of the sum of squares is obtained to get the fault capsule magnitude. S503. Retrieve a preset first preset threshold. The first preset threshold is a fixed threshold determined in advance through training with samples of different grounding fault conditions (low-resistance grounding fault, high-resistance grounding fault) in the distribution network. Its value ranges from 0.6 to 0.8. It is used to determine the feature significance of the fault state capsule vector. The value of the first preset threshold must ensure that it can effectively distinguish the difference in magnitude between the fault state capsule vector under fault and normal conditions, and avoid the vector fluctuation under normal conditions from falsely triggering the fault determination. S504. Compare the calculated fault capsule module length with the first preset threshold. If the fault capsule module length does not exceed the first preset threshold, the power distribution network is directly determined to be in normal operation and no fault start signal is output. If the fault capsule module length exceeds the first preset threshold, proceed to the next determination process. S505. Retrieve the preset fault feature basis vector. The fault feature basis vector is a standard vector that has been pre-extracted through training with a large number of single-phase grounding fault samples in the distribution network. Its vector direction accurately corresponds to the core features of the single-phase grounding fault in the distribution network and is consistent with the dimension of the fault state capsule vector. The fault feature basis vector is obtained by collecting single-phase grounding fault samples of the distribution network under different voltage levels and different fault resistances, extracting the standardized mutation feature quantity sequence corresponding to each sample, and obtaining the fault state capsule vector corresponding to each sample after primary capsule layer encoding and weight iterative convergence. The mean of the fault state capsule vectors of all samples is calculated and normalized to obtain the preset fault feature basis vector. S506. Calculate the angle between the fault state capsule vector and the preset fault feature basis vector, which is defined as the fault feature angle. The calculation of the fault feature angle adopts the vector dot product formula, that is, first calculate the dot product of the fault state capsule vector and the fault feature basis vector, then calculate the magnitude of the two vectors respectively, divide the dot product result by the product of the magnitudes of the two vectors to obtain the cosine value of the angle, and then obtain the fault feature angle through the inverse cosine operation. The unit of the fault feature angle is radians, and the value range is 0 to π. S507. Retrieve a preset second preset threshold. The second preset threshold is a fixed threshold determined in advance through training with distribution network fault samples. Its value ranges from 0.2 radians to 0.3 radians. It is used to determine the degree of matching between the fault state capsule vector and the core features of the fault. The value of the second preset threshold must ensure that it can effectively distinguish the vector direction difference between ground fault and non-fault conditions, and avoid interference vectors under non-fault conditions from falsely triggering fault determination. S508. Compare the calculated fault feature angle with the second preset threshold. If the fault feature angle is not less than the second preset threshold, the distribution network is determined to be in normal operation and no fault start signal is output. If the fault feature angle is less than the second preset threshold, the distribution network is determined to have a single-phase ground fault. In addition, this embodiment also adopts another fault determination method, which is as follows: extract the magnitude of the high-rise capsule vector representing the ground fault state and the magnitude of the high-rise capsule vector representing the normal operation state after the weighted iterative convergence. At the same time, calculate the fault feature angle between the high-rise capsule vector of the fault state and the preset fault feature basis vector. Simultaneously, set the timing consistency judgment condition, requiring that the magnitude of the high-rise capsule vector of the fault state is greater than the first preset threshold and the fault feature angle is less than the second preset threshold within three consecutive time windows. The final fault determination rule is that only when the magnitude of the high-rise capsule vector of the fault state is greater than the first preset threshold, the fault feature angle is less than the second preset threshold, and the above timing consistency judgment condition is met, is it determined that a single-phase ground fault has occurred in the distribution network. All other cases are determined to be normal operation. The fault start signal generated after the single-phase ground fault is determined to be a pulse signal with an amplitude of 24V, a frequency of 50Hz, and a duration of 100ms. The interface protocol of the fault start signal conforms to the DL / T634.5101-2002 standard and can be directly transmitted to the distribution network feeder terminal.

[0049] S509. When a single-phase ground fault is detected in the distribution network, a fault initiation signal is immediately generated. The fault initiation signal is an electrical signal in a preset format. Its signal parameters (amplitude and frequency) conform to the signal reception standard of the distribution network feeder terminal and can trigger the fault response mechanism of the feeder terminal. At the same time, the fault initiation signal is output to the fault processing module of the distribution network feeder terminal to provide a triggering basis for subsequent fault location and fault isolation. S510. If the distribution network is determined to be in normal operation after the above steps, no fault start signal will be generated, and the normal operation status identifier will be continuously output. At the same time, the current final high-rise capsule vector and the corresponding connection weight will be retained as the initial parameters for subsequent iterative calculations to achieve the continuity of fault determination.

[0050] For example, in the scenario of single-phase grounding fault identification in a 10kV distribution network, this embodiment combines the core principles of zero-sequence voltage mutation feature encoding and capsule neural network weight iteration. The specific example of fault determination and signal output is as follows: Based on the principle of prior zero-sequence voltage discrete sequence processing, mutation feature encoding, and weight iteration convergence between primary and higher-level capsules, after the weight iteration process terminates, the output contains stable vectors of two higher-level capsules, one for fault state and one for normal state. The higher-level capsule vector for fault state can accurately map the fault characteristics carried by the primary capsule layer. This is the core basis for fault determination. The principle is to allow the higher-level capsules to fully learn the fault-specific characteristics through weight iteration, ensuring the accuracy of vector representation. First, the fault state capsule vector is extracted from the final high-level capsule vector, and its magnitude is calculated to be 0.75. A preset first threshold of 0.7 is then applied. Since the fault capsule magnitude exceeds the threshold, the next step of judgment is initiated. Next, a 10-dimensional fault feature basis vector (with the same dimension as the fault state capsule vector) trained on a large number of low-resistance and ultra-high-resistance fault samples is retrieved. The angle between the two is calculated to be 0.25 radians, which is less than a preset second threshold of 0.3 radians. Based on this, a single-phase ground fault is determined to have occurred. A fault initiation signal conforming to the FTU receiving standard is immediately generated and output to the feeder terminal fault processing module, providing support for subsequent fault location and isolation. If the above two threshold conditions are not met, the system is determined to be normal, while the current vector and weights are retained to ensure the continuity of subsequent judgments. This example fully demonstrates the feature representation advantages of the high-level capsule vector after weight iteration convergence. In practice, it can accurately distinguish between fault and normal operating conditions, improve the FTU fault identification accuracy, effectively avoid misjudgment and missed judgment, and adapt to the different fault conditions required by the distribution network.

[0051] This embodiment accurately captures transient fault characteristics and effectively filters sampling noise and irrelevant fluctuations by performing differential absolute value calculation, boundary processing, dynamic window length adjustment, and standardization on the discrete sequence of zero-sequence voltage of the distribution network bus. It also adapts to different fault conditions, such as low-resistance and ultra-high-resistance faults, solving the problem that traditional fixed window lengths cannot simultaneously address weak fault feature extraction and computational efficiency. By constructing primary capsule vectors through a pre-defined convolutional coding network, and combining condition feature perception, adaptive selection of deep and shallow paths, dynamic noise suppression, and dynamic adjustment of capsule dimensions, feature encoding distortion and redundancy are avoided, significantly improving the feature representation capability of the primary capsule vectors and providing reliable support for subsequent weight iterations. Through initial weight settings based on the normalization of the primary capsule vector magnitude and weight sparsification during the iteration process, the embodiment achieves this. The system employs a triple dynamic convergence criterion to address issues such as initial bias in weight iteration, imbalance in allocation, and pseudo-convergence. This enables the high-level capsule to accurately learn fault-specific characteristics, improving the accuracy of system state representation. Based on a dual-threshold fault determination mechanism using the vector magnitude and included angle of the high-level capsule, it effectively distinguishes between single-phase grounding faults and normal operating conditions, significantly reducing the probability of false or missed fault detections. Simultaneously, it rapidly generates fault initiation signals conforming to FTU standards, providing a reliable triggering basis for subsequent fault location and isolation, while retaining the iterated vectors and weights to ensure continuity of judgment. The entire process requires no complex manual intervention, balancing fault identification accuracy and efficiency, adapting to different fault conditions and data length variations, enhancing the method's practicality and generalization ability, reducing the cost of distribution network fault handling, and effectively improving the intelligence level of FTU fault identification.

[0052] Example 2 Please see Figure 2 Another embodiment of the present invention provides: a ground fault identification system for zero-sequence voltage mutation characteristics for FTU, comprising: The acquisition module obtains a discrete sequence representing the zero-sequence voltage of the distribution network bus. U 0 ( n )},in n The sampling point number; The difference module, based on the discrete sequence { U 0 ( n )}, calculate the absolute difference between its consecutive sampling points to obtain the absolute difference sequence { D.U. 0 ( n )}; The feature construction module performs a process on the difference absolute value sequence { D.U. 0 ( n Perform sliding time window integration to construct a sequence of mutation feature quantities, where the feature quantity corresponding to the k-th time window is... ,L The preset window length, k For time window sequence number; The mapping module inputs the mutation feature sequence into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a set of primary capsule vectors, forming the output vector of the primary capsule layer. The iterative module performs iterative calculations between the output vector of the primary capsule layer and the higher-level capsules that represent the system state, updating the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. The fault detection module determines the fault based on the converged vector state of the high-level capsule: if the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault characteristic basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault start signal is output; otherwise, the system is determined to be in normal operation.

[0053] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs, characterized in that, include: The discrete sequence {U0(n)} characterizing the zero-sequence voltage of the distribution network bus is obtained by the configured electrical sensors, where n is the sampling point number; Based on the discrete sequence {U0(n)}, the absolute difference between its consecutive sampling points is calculated to obtain the absolute difference sequence {...}. U0(n)}; For the difference absolute value sequence { By performing sliding time window integration on U0(n), a sequence of mutation feature quantities is constructed, where the feature quantity corresponding to the k-th time window is... L is the preset window length, and k is the time window number; The mutation feature sequence is input into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a set of primary capsule vectors, forming the output vector of the primary capsule layer. Iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules that represent the system state, updating the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. Fault determination is performed based on the converged vector state of the high-level capsule: if the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault characteristic basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault initiation signal is output. Otherwise, the system is determined to be in normal operating condition; The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, including: The standardized mutation feature sequence is divided into multiple consecutive feature subsequences according to a preset non-overlapping partitioning rule; wherein each feature subsequence contains a preset number P consecutive standardized mutation features, where P is an integer greater than 1, and there is no data overlap between adjacent feature subsequences. Multiple consecutive feature subsequences are input into a preset working condition feature perception layer and a pre-configured three-layer parallel convolutional layer is called to obtain the first, second and third feature maps with working condition features. The first, second, and third feature maps with working condition characteristics are concatenated along the channel dimension to generate a fused feature map corresponding to the feature subsequence.

2. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 1, characterized in that, The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, and is implemented through a pre-defined convolutional coding network. It also includes: Based on the fused feature map, the depth control module performs the following operations to generate a depth-encoded feature vector: Extract the original feature subsequence corresponding to the current fused feature map; Calculate the kurtosis and skewness statistical features of the original feature subsequence; Compare the kurtosis statistical feature value with a third preset threshold, and compare the skewness statistical feature value with a fourth preset threshold; If the kurtosis statistical feature value is greater than a third preset threshold and the skewness statistical feature value is greater than a fourth preset threshold, then the fused feature map is input to the deep processing path to obtain a deep encoded feature vector; the deep processing path consists of two concatenated convolutional processing units, each convolutional processing unit including: a one-dimensional convolutional layer configured with 16 convolutional kernels of size 3 and a stride of 1; a batch normalization layer; and a ReLU activation function layer; If the third or fourth preset threshold is not met, the fused feature map is input to the shallow processing path to obtain a deep encoded feature vector. The shallow processing path consists of one of the convolutional processing units.

3. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 2, characterized in that, The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, and is implemented through a pre-defined convolutional coding network. It also includes: The deep encoded feature vector is subjected to noise suppression processing to obtain the suppressed feature vector, specifically as follows: Determine the original feature subsequence corresponding to the deep encoded feature vector, and define the original feature subsequence as the current processing subsequence; Based on the division order of the standardized mutation feature sequence, the adjacent feature subsequences that are sequentially preceding the current processing subsequence are obtained as the preceding subsequences, and the adjacent feature subsequences that are sequentially following the current processing subsequence are obtained as the following subsequences. All normalized mutation features contained in the current processing subsequence, the preceding subsequence, and the subsequent subsequence are merged to form a merged sequence.

4. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 3, characterized in that, The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, and is implemented through a pre-defined convolutional coding network. It also includes: Calculate the variance of all elements in the merged sequence, and use the value of the variance as the noise evaluation value of the current process; Calculate the median of the noise evaluation values ​​corresponding to all feature subsequences in the current processing batch, and use the median as the baseline noise value; Calculate the dynamic attenuation coefficient α. =min(1, ), where T is the median of the noise evaluation values ​​of all sequence segments processed in the current batch, and V is the noise evaluation value of the current sequence segment; Each element of the deep encoded feature vector is multiplied by the dynamic decay coefficient to obtain the suppressed feature vector.

5. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 4, characterized in that, The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, and is implemented through a pre-defined convolutional coding network. It also includes: Converting the suppressed feature vector into a primary capsule vector specifically includes: Obtain the total number of time windows N contained in the standardized mutation feature sequence; The target capsule dimension D is dynamically determined based on the total number of time windows N. The specific rules are as follows: when N is less than or equal to 100, D is set to 8; when N is greater than 100 and less than or equal to 200, D is set to 10; when N is greater than 200, D is set to 12. The suppressed feature vector is input into a fully connected projection layer, which maps the dimension of the input vector to the dimension D of the target capsule. The output vector of the fully connected projection layer is normalized using the L2 norm, and the normalized vector is used as a primary capsule vector. The standardized mutation feature sequence is divided into segments of a preset fixed length, such that each segmented feature subsequence contains 50 consecutive standardized mutation features, and there is no data overlap between any two adjacent feature subsequences. For each feature subsequence obtained by the division, the primary capsule vector acquisition process is executed sequentially to generate a primary capsule vector corresponding to the corresponding feature subsequence; All generated primary capsule vectors are arranged and combined according to the chronological order of their corresponding feature subsequences in the standardized mutation feature sequence to form the output vector of the primary capsule layer.

6. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 5, characterized in that, The step of inputting multiple consecutive feature sub-sequences into a preset working condition feature perception layer and calling a pre-configured three-layer parallel convolutional layer is as follows: The operating condition feature perception layer receives the multiple consecutive feature subsequences and calculates the feature statistics of the multiple consecutive feature subsequences; the feature statistics include the mean and the standard deviation σ. F Maximum amplitude A max ; Based on characteristic statistics and by setting a first or second judgment condition, the fault condition type is obtained, specifically: When the first condition is met, that is, A max >0.5 and σ F >0.2, which is determined to be a low-resistance fault basic operating condition; When the second criterion is met, that is, if 0.3 ≤ A max ≤0.5 and 0.1≤σ F ≤0.2, judged as a medium resistance fault basic operating condition; When the third criterion is met, that is, if A max <0.3 and σ F <0.1 indicates an ultra-high resistance fault condition. If the first, second, and third judgment conditions are not met, the condition is judged to be normal or transitional. The invocation of the pre-configured three-layer parallel convolutional layer is specifically as follows: Based on the final determined working condition type, the parameters of the pre-configured three-layer parallel convolutional layer are invoked: If the basic operating condition is determined to be a low-resistance fault, the first set of parameters is called: the first convolutional layer uses 32 one-dimensional convolutional kernels with a size of 2, the second convolutional layer uses 64 one-dimensional convolutional kernels with a size of 2, and the third convolutional layer uses 128 one-dimensional convolutional kernels with a size of 2. If the basic operating condition is determined to be a medium resistance fault, the fourth set of parameters is called: the first convolutional layer uses 24 one-dimensional convolutional kernels with a size of 3, the second convolutional layer uses 48 one-dimensional convolutional kernels with a size of 3, and the third convolutional layer uses 96 one-dimensional convolutional kernels with a size of 3. If the basic operating condition is determined to be an ultra-high resistance fault, the second set of parameters is called: the first convolutional layer uses 8 one-dimensional convolutional kernels with a size of 4, the second convolutional layer uses 16 one-dimensional convolutional kernels with a size of 4, and the third convolutional layer uses 32 one-dimensional convolutional kernels with a size of 4. If the condition is determined to be normal or transitional, the third set of parameters is called: the first convolutional layer uses 16 one-dimensional convolutional kernels of size 3, the second convolutional layer uses 32 one-dimensional convolutional kernels of size 3, and the third convolutional layer uses 64 one-dimensional convolutional kernels of size 3.

7. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 6, characterized in that, Iterative calculations are performed between the output vector of the primary capsule layer and the higher-level capsules representing the system state to update the connection weights between the primary capsule vector and each higher-level capsule, including: Based on the output vector of the primary capsule layer, the connection weights between each higher-level capsule and all primary capsule vectors are initialized as follows: Obtain each primary capsule vector contained in the primary capsule layer output vector; Calculate the magnitude of each of the primary capsule vectors; Linearly normalize the magnitudes of all primary capsule vectors so that the sum of the normalized magnitudes equals the total number of higher-level capsules. The magnitude of each primary capsule vector is directly set as the initial connection weight between the corresponding primary capsule vector and each higher-level capsule; Based on the initial connection weights and the output vector of the primary capsule layer, an iterative calculation process is performed. A single iteration includes the following implementation steps: The prediction vector of the high-level capsule is calculated based on the current connection weights. Specifically, each primary capsule vector is mapped through a pre-trained linear transformation matrix to obtain the prediction vector of the primary capsule vector for each high-level capsule. For the same high-level capsule, the prediction vectors of all primary capsule vectors are multiplied by the current connection weights between the corresponding primary capsule vectors and the high-level capsule, and the weighted prediction vectors are summed to obtain the prediction vector of the high-level capsule in the current iteration round.

8. The method for identifying ground faults with zero-sequence voltage mutation characteristics for FTUs as described in claim 7, characterized in that, The execution of the iterative calculation process, a single iteration includes the following implementation steps, and also includes: Obtain the vector of each primary capsule and the predicted vector of each higher-level capsule in the current iteration round; Calculate the vector dot product between each primary capsule vector and each high-level capsule prediction vector, and define the result as the first similarity metric. For each high-level capsule, perform an update operation: add the current connection weight between the high-level capsule and each primary capsule vector to the first similarity metric between the primary capsule vector and the high-level capsule to obtain a set of original update values; For all the original updated values ​​belonging to the same high-level capsule, the Softmax function is applied to normalize the calculation and obtain the updated connection weights. For each high-level capsule, obtain the predicted vector for that high-level capsule from all the primary capsule vectors; Each prediction vector is multiplied by the updated connection weight between the corresponding primary capsule vector and the higher-level capsule to obtain a weighted prediction vector. Summing all weighted prediction vectors yields the weighted sum vector of the high-level capsule. Apply a nonlinear compression function to the weighted sum vector: calculate the magnitude of the weighted sum vector, and adjust the values ​​of each component of the vector based on the magnitude using the compression function so that the magnitude of the output vector is no greater than 1; The compressed vector is used as the high-level capsule vector for the current iteration round.

9. A ground fault identification system for zero-sequence voltage mutation characteristics for FTUs, used to implement the ground fault identification method for zero-sequence voltage mutation characteristics for FTUs according to any one of claims 1-8, characterized in that, include: The acquisition module obtains a discrete sequence {U0(n)} representing the zero-sequence voltage of the distribution network bus, where n is the sampling point number; The difference module calculates the absolute difference between consecutive sampling points based on the discrete sequence {U0(n)}, to obtain the absolute difference sequence {ΔU0(n)}. The feature construction module performs a sliding time window integration on the difference absolute value sequence {ΔU0(n)} to construct a mutation feature quantity sequence, where the feature quantity corresponding to the k-th time window is... L is the preset window length, and k is the time window number; The mapping module inputs the mutation feature sequence into a pre-trained capsule neural network model. The primary capsule layer of the model encodes the feature sequence within each time window into a set of primary capsule vectors, forming the output vector of the primary capsule layer. The iterative module performs iterative calculations between the output vector of the primary capsule layer and the higher-level capsules that represent the system state, updating the connection weights between the primary capsule vector and each higher-level capsule until the weights converge. The fault detection module determines the fault based on the converged vector state of the high-level capsule: if the vector magnitude of the high-level capsule corresponding to the ground fault state exceeds the first preset threshold, and the angle between its vector direction and the preset fault feature basis vector is less than the second preset threshold, then a single-phase ground fault is determined to have occurred, and a fault start signal is output; otherwise, the system is determined to be in normal operation. The primary capsule layer of the capsule neural network model generates the primary capsule vector based on a standardized sequence of mutational features, implemented through a pre-defined convolutional coding network, including: The standardized mutation feature sequence is divided into multiple consecutive feature subsequences according to a preset non-overlapping partitioning rule; wherein each feature subsequence contains a preset number P consecutive standardized mutation features, where P is an integer greater than 1, and there is no data overlap between adjacent feature subsequences. Multiple consecutive feature subsequences are input into a preset working condition feature perception layer and a pre-configured three-layer parallel convolutional layer is called to obtain the first, second and third feature maps with working condition features. The first, second, and third feature maps with working condition characteristics are concatenated along the channel dimension to generate a fused feature map corresponding to the feature subsequence.

Images