Small current grounding system fault cause analysis method and system
By acquiring topology and equipment waveforms in a low-current grounding system, and using feature modules and classifiers for fault cause analysis, the problem of high hardware dependence and insufficient real-time performance in existing technologies is solved, achieving low-cost and rapid fault detection and isolation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING DAN HUA HAO BO POWER SCI & TECH CO LTD
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for fault analysis in low-current grounding systems suffer from high hardware dependence, high cost, difficulty in large-scale deployment, limited sensing dimensions, inability to detect latent faults, and difficulty in meeting real-time requirements for electrical quantity detection methods.
A fault cause analysis method for low-current grounding systems is adopted. By acquiring the topology, equipment waveforms, and zero-sequence voltage and current for cycle interception, and using feature modules and classifiers to perform fault cause analysis, rapid diagnosis and isolation can be achieved.
It enables full coverage deployment in existing power grid systems, low-cost detection of both explicit and implicit faults, and meets the real-time requirements of modern distribution networks for rapid fault diagnosis and isolation.
Smart Images
Figure CN122020290A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of low-current grounding systems, and particularly relates to a method and system for analyzing the causes of faults in low-current grounding systems. Background Technology
[0002] A low-current grounding system refers to a three-phase system where the neutral point is ungrounded, grounded through an arc suppression coil, or grounded with high impedance; it is also known as a neutral-point indirect grounding system. When a ground fault occurs in one phase, because a short-circuit loop cannot be formed, the ground fault current is often much smaller than the load current, hence the name low-current grounding system.
[0003] Common faults include single-phase grounding, open circuits, and short circuits, each with different characteristics and causes. Single-phase grounding faults manifest as a decrease in voltage in the faulty phase rather than an increase in voltage in the non-faulty phase, usually caused by foreign objects or tree obstructions. Open circuit faults lead to abnormal voltage and load imbalance, often caused by external forces or lightning strikes. Short circuit faults are accompanied by a sudden increase in current and protection tripping, mainly caused by insulation breakdown or foreign objects. Fault Cause Analysis: Analyzing the causes of faults after they occur can prevent similar faults from recurring and guide maintenance personnel to take targeted measures (such as strengthening insulation monitoring, optimizing lightning protection, or tightening connection points), thereby improving system reliability and reducing power outage losses. In-depth and systematic cause analysis of faults occurring in low-current grounding systems is far more than a routine post-incident accountability procedure; it is a strategic core element in improving the intelligent operation and maintenance level of the power grid and ensuring power supply reliability.
[0004] Existing methods for analyzing and detecting faults in low-current grounding systems can be divided into two categories: non-electrical quantity-based detection and electrical quantity-based detection.
[0005] Based on non-electrical quantity detection: This method mainly uses information from non-electrical quantities for comprehensive judgment, such as... Reference 1 (Early warning and optimization of tree-line conflict risk in rural power distribution networks considering severe convective weather, Yao Fuxing et al., *Journal of Electrical Engineering*) uses historical meteorological data and fault records to establish a predictive model to determine which areas are prone to faults under specific weather conditions. Reference 2 (Design of a tree obstacle detection method for power line inspection based on lidar, Li Junpeng, *Electronic Design Engineering*) installs multispectral lidar on equipment to collect the distance between trees and power lines to determine whether tree-line grounding has occurred.
[0006] Based on electrical quantity detection: This method mainly uses electrical quantity information to build a database. When a new fault is detected, it is compared with existing data to obtain the distance. The closer the distance, the higher the similarity between the current fault and similar or historical faults. For example: Reference 3 (Fault Diagnosis Algorithm Based on Case Reasoning, Kong Qin et al., Computer Systems Applications) proposes a case matching algorithm based on vector computation, which is based on a case reasoning-based fault diagnosis algorithm. This algorithm abstracts the fault information and performs similarity calculation with the cases in the system's case library to match the fault information with the case information one by one.
[0007] However, existing technologies have some defects and shortcomings: 1. The hardware dependence of non-electrical quantity detection methods is high, costly, and difficult to deploy on a large scale: These methods heavily rely on external precision detection equipment such as lidar, satellite remote sensing, high-definition cameras, and weather sensors. Not only are these devices expensive to install, calibrate, maintain, and handle subsequent data transmission and storage, but they also require continuous and substantial investment. This makes it difficult to achieve full coverage deployment in distribution networks (especially in rural or old urban power grids with numerous branches and complex structures), resulting in numerous monitoring blind spots.
[0008] 2. Fault identification based on non-electrical quantity detection methods is limited in types and has a single perception dimension: The core of this method is to discover external hidden dangers through "seeing" and "measuring." It cannot effectively detect latent faults: such as insulation aging, deterioration, internal moisture, short circuits between winding turns, and increased contact resistance due to loose conductive connections, which are "slow-changing" or "latent" faults occurring inside the equipment or electrical connection points, are almost powerless by non-electrical quantity detection methods. These faults do not show obvious changes in external physical or chemical characteristics in the early stages, but they are important root causes of serious failures.
[0009] 3. The electrical quantity detection method relies on high-quality and complete fault electrical quantity data: This method presupposes the ability to acquire complete and accurate fault waveform data. However, in actual distribution networks, especially when a fault occurs, data acquisition devices may experience missing, distorted, or erroneous records due to overvoltage, electromagnetic interference, or other reasons, leading to case matching failures.
[0010] 4. As the case database grows larger, the number of similarity calculations required for each fault diagnosis increases, and the retrieval time increases linearly or even exponentially, making it difficult to meet the real-time requirements of modern power distribution networks for rapid fault diagnosis and isolation. Summary of the Invention
[0011] To address the shortcomings of existing technologies, this invention provides a method and system for analyzing fault causes in low-current grounding systems. When a low-current grounding fault occurs, the method involves: acquiring the topology of the system; acquiring the waveform recording of the device closest to the fault point; extracting the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle from the waveform recording; truncating the waveform recording based on the length of the single cycle and a set step size; processing the truncated cycles; inputting the zero-sequence voltage, zero-sequence current, and processed cycles into several discriminators targeting different fault causes to obtain classification probabilities for different types of fault causes; wherein the discriminator includes a feature module and a classifier; and selecting the final judgment result based on the classification probabilities of different types of fault causes according to actual needs. This invention relies on feature modules and classifiers for judgment, which can meet the real-time requirements of modern power distribution networks for rapid fault diagnosis and isolation.
[0012] The present invention adopts the following technical solution.
[0013] This invention proposes a method for analyzing the causes of faults in low-current grounding systems, including: When a low-current ground fault occurs, obtain the topology of the low-current ground fault system. Locate the fault point in the topology and obtain the device waveform of the device that is physically closest to the fault point; Extract the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the device waveform recording; truncate the device waveform recording based on the length of the single cycle and a set step size; process the truncated cycles. The zero-sequence voltage, the zero-sequence current, and the processed cycle are input into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; Based on the classification probability of different types of fault causes, the final judgment result is selected according to actual needs.
[0014] More preferably, the processing of the intercepted frequency specifically includes: The extracted cycles are resampled and normalized. The resampled and normalized cycles are then decomposed using FFT to obtain the zero-sequence voltage FFT result and the zero-sequence current FFT result.
[0015] More preferably, the discriminator includes a feature module and a classifier, and its specific structure includes: The feature module includes an input layer, a feature extraction part, a feature fusion part, and an output layer; the input layer reads the zero-sequence voltage, the zero-sequence current, and the processed cycle. The discriminators share the same network parameters for the same feature extraction part, while different discriminators have feature fusion parts and classifiers with different network parameters.
[0016] More preferably, the input format of the input layer is: The format is [m,n,4,256], where m represents the number of input samples, n represents the number of cycles of the input samples, 4 represents four channels, and 256 is the default number of nodes in one cycle; The four channels are input sequentially: zero-sequence voltage, zero-sequence voltage FFT result, zero-sequence current, and zero-sequence current FFT result, respectively. The output format of the output layer is [m, 128], where m represents the number of input samples and 128 is the dimension of the extracted abstract features.
[0017] More preferably, the structure of the feature extraction part is as follows: Part 1: One-dimensional convolutional layer, batch normalization layer, ReLU activation function layer, one-dimensional max pooling layer; the output of Part 1 serves as the input of Part 2. Part Two: Flattening layer, linear transformation layer, Dropout layer, ReLU activation function layer.
[0018] More preferably, the specific structure of the feature fusion portion is as follows: The output of the feature extraction part is multiplied element-wise by the result of the linear transformation layer and the Softmax activation function layer, and the result of the multiplication is then passed through the linear transformation layer and the ReLU activation function layer to obtain the output.
[0019] More preferably, the specific structure of the classifier includes: The input layer receives the output from the feature module. The input format is [m, 128], where m represents the number of input samples and 128 is the feature dimension. The input layer's results are sequentially passed through a one-dimensional convolutional layer, a linear transformation layer, and a sigmoid function layer to obtain the probability of belonging to the corresponding category.
[0020] This invention also proposes a fault cause analysis system for low-current grounding systems, including a topology acquisition module, a device waveform acquisition module, a waveform processing module, a discriminator discrimination module, and an analysis module: The topology acquisition module acquires the topology of the system when a low-current ground fault occurs. The device waveform acquisition module locates the fault point in the topology and acquires the device waveform of the device that is physically closest to the fault point. The waveform processing module extracts the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the waveform recorded by the device; performs cycle truncation on the waveform recorded by the device based on the length of the single cycle and a set step size; and processes the truncated cycles. The discriminator discrimination module inputs the zero-sequence voltage, the zero-sequence current, and the processed cycle into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; The analysis module, based on the classification probability of different types of fault causes, selects the final analysis result according to actual needs.
[0021] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.
[0022] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention does not require additional hardware equipment. It only uses faults or DTU / FTU measurements and can be fully deployed in the existing power grid system by simply fusing algorithms. The cost is very low and the deployment is quick and convenient.
[0024] 2. This invention can detect not only explicit fault causes, but also "slow-change" or "latent" faults such as insulation aging.
[0025] 3. The present invention requires only one type of data for judgment, without any other data, and it is almost an essential element for judging faults. It is easy to obtain and can quickly accumulate a large number of cases.
[0026] 4. The present invention relies on feature modules and classifiers for judgment. As the case library increases, it will only increase the model training time, not the computation time during deployment. This can meet the real-time requirements of modern power distribution networks for rapid fault diagnosis and isolation. Attached Figure Description
[0027] Figure 1 This is a flowchart of a method for analyzing the causes of faults in a low-current grounding system according to the present invention; Figure 2 This is an overall flowchart of a fault cause analysis method for a low-current grounding system according to Embodiment 1 of the present invention; Figure 3 This is a structural diagram of the feature module of Embodiment 1 of the present invention; Figure 4This is a structural diagram of the classifier in Embodiment 1 of the present invention. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.
[0029] This invention provides the following technical solution: like Figure 1 As shown, this invention proposes a method for analyzing the causes of faults in low-current grounding systems, including: When a low-current ground fault occurs, obtain the topology of the low-current ground fault system. Locate the fault point in the topology and obtain the device waveform of the device that is physically closest to the fault point; Extract the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the device waveform recording; truncate the device waveform recording based on the length of the single cycle and a set step size; process the truncated cycles. The processing of the intercepted frequency specifically includes: The extracted cycles are resampled and normalized. The resampled and normalized cycles are then decomposed using FFT to obtain the zero-sequence voltage FFT result and the zero-sequence current FFT result.
[0030] The zero-sequence voltage, the zero-sequence current, and the processed cycle are input into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; The discriminator includes a feature module and a classifier, and its specific structure includes: The feature module includes an input layer, a feature extraction part, a feature fusion part, and an output layer; the input layer reads the zero-sequence voltage, the zero-sequence current, and the processed cycle. The discriminators share the same network parameters for the same feature extraction part, while different discriminators have feature fusion parts and classifiers with different network parameters.
[0031] The input format of the input layer is: The format is [m,n,4,256], where m represents the number of input samples, n represents the number of cycles of the input samples, 4 represents four channels, and 256 is the default number of nodes in one cycle; The four channels are input sequentially: zero-sequence voltage, zero-sequence voltage FFT result, zero-sequence current, and zero-sequence current FFT result, respectively. The output format of the output layer is [m, 128], where m represents the number of input samples and 128 is the dimension of the extracted abstract features.
[0032] The structure of the feature extraction part is as follows: Part 1: One-dimensional convolutional layer, batch normalization layer, ReLU activation function layer, one-dimensional max pooling layer; the output of Part 1 serves as the input of Part 2. Part Two: Flattening layer, linear transformation layer, Dropout layer, ReLU activation function layer.
[0033] The specific structure of the feature fusion part is as follows: The output of the feature extraction part is multiplied element-wise by the result of the linear transformation layer and the Softmax activation function layer, and the result of the multiplication is then passed through the linear transformation layer and the ReLU activation function layer to obtain the output.
[0034] The specific structure of the classifier includes: The input layer receives the output from the feature module. The input format is [m, 128], where m represents the number of input samples and 128 is the feature dimension. The input layer's results are sequentially passed through a one-dimensional convolutional layer, a linear transformation layer, and a sigmoid function layer to obtain the probability of belonging to the corresponding category.
[0035] Based on the classification probability of different types of fault causes, the final judgment result is selected according to actual needs.
[0036] This invention also proposes a fault cause analysis system for low-current grounding systems, including a topology acquisition module, a device waveform acquisition module, a waveform processing module, a discriminator discrimination module, and an analysis module: The topology acquisition module acquires the topology of the system when a low-current ground fault occurs. The device waveform acquisition module locates the fault point in the topology and acquires the device waveform of the device that is physically closest to the fault point. The waveform processing module extracts the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the waveform recorded by the device; performs cycle truncation on the waveform recorded by the device based on the length of the single cycle and a set step size; and processes the truncated cycles. The discriminator discrimination module inputs the zero-sequence voltage, the zero-sequence current, and the processed cycle into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; The analysis module, based on the classification probability of different types of fault causes, selects the final analysis result according to actual needs.
[0037] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.
[0038] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0039] Example 1 This invention proposes a method for analyzing the causes of faults in low-current grounding systems, such as... Figure 2 As shown, the zero-sequence voltage is comprehensively utilized. and zero-sequence current Calculate using a single cycle as the unit. The features are then fed into the feature module to obtain abstract features, which are then fed into the classifier to obtain the judgment result. The specific steps are as follows: 1. Construct the discriminator, including the feature module and the classifier, and initialize the discriminator and its parameters.
[0040] Specifically, the network structure adopted by the feature module is as follows: Figure 3 The input format is [m,n,4,256], where m represents the number of input samples, n represents the number of cycles of the input samples, 4 represents four channels (channel order: zero-sequence voltage, zero-sequence voltage FFT result, zero-sequence current, zero-sequence current FFT result), 256 is the default number of nodes in one cycle (if the original number of nodes is inconsistent, resampling is used for preprocessing), Conv1d is one-dimensional convolution, BN1d is batch normalization, ReLU is the activation function, MaxPool1d is one-dimensional max pooling, Linear is the linear transformation in the fully connected layer, Dropout layer achieves regularization by randomly discarding neurons, and softmax is the activation function.
[0041] The network is divided into four main parts: input, feature extraction, feature fusion, and output.
[0042] The feature extraction part consists of two parts: a convolutional layer and a linear transformation. The convolution is used to extract features, and the linear transformation transforms the features into one dimension, while regularization prevents overfitting.
[0043] The feature fusion section introduces an attention mechanism. An attention score is calculated by performing a linear transformation on the feature extraction result using a Linear function. This score is then multiplied by the extracted feature score and summed to amplify important features and scale ineffective features. Finally, the features are fused using a Linear and ReLU activation function, outputting an abstract feature set in the range [m, 128].
[0044] Furthermore, the network structure used by the classifier is as follows: Figure 4 The input format is [m, 128], where m represents the number of input samples and 128 is the extracted abstract feature dimension. The input is first normalized using a Batch Normalization (BN) layer, followed by a linear transformation. Finally, the sigmoid function locks the output to a number between 0 and 1, representing the probability of a sample belonging to that class. If the probability exceeds a set threshold (usually 0.5), the sample is considered to belong to that class. The output format is [m, 1], where m represents the number of input samples, 1 is the feature dimension, and the output channel is 1.
[0045] The training process of the feature module and classifier, including the design of the loss function, the selection of the optimizer, the learning rate, the number of iterations, the batch size, etc., is determined by the actual experimental performance.
[0046] 2. Read the topology of the low-current grounding fault system to locate the fault point and record waveforms of the nearest device. The location of the fault point is based on mature existing technologies and will not be elaborated upon here.
[0047] 3. Read the waveform recorded by the device and the length of a single cycle.
[0048] 4. Extract cycles according to the single-cycle length and the set step size (the default is the same as the single-cycle length).
[0049] 5. Resample and normalize the extracted cycles, and then perform FFT decomposition.
[0050] Specifically, after the truncation is completed, the sample is resampled to the set value (default 256), then normalized to unify the voltage and current. The normalized domain value is [0,1]. Finally, FFT is used to extract features of different frequencies.
[0051] The resampling process involves two steps: To achieve a unified frequency, data points are supplemented using linear interpolation. For each pair of adjacent points, N points are added for upsampling, and for every N points, one point is taken for downsampling (N is the value of the expected frequency divided by the actual frequency and rounded off).
[0052] Even after unifying the frequency, there may still be inconsistencies in the number of dots, specifically referring to situations where the actual frequency and the default frequency (12800Hz) are not integer multiples of each other. If there are too many dots, simply delete them from the end; if there are too few dots, add 0 to the end. (This is because a 0 value will output 0 in the convolution operation if there is no bias, which is equivalent to not extracting the abstract features.)
[0053] 6. Arrange the data according to ]arrangement.
[0054] 7. Input the data into multiple feature modules and output their respective abstract features.
[0055] 8. Input the abstract features into the corresponding classifier and output the probability that the sample belongs to each class.
[0056] 9. The analysis results are obtained by combining the analysis strategy.
[0057] Specifically, several discriminators are trained for each category. Each discriminator includes a feature module and a classifier. The feature extraction part of the feature module is shared across different discriminators, while the feature fusion part and the classifier are trained separately according to the cause category. During a sample test, the input is given to all discriminators to obtain the probability of it being identified as a certain fault cause. The output can be adjusted according to actual needs. For example, if only the most likely cause needs to be output, the category corresponding to the largest probability value greater than a threshold can be selected. If multiple causes can be pushed, the categories that meet the threshold can be sorted and output in descending order of probability. Furthermore, when a fault occurs, the cause of a single fault is determined by the device closest to the fault point.
[0058] The reason why feature extraction is common is that the input and output must remain consistent during training, which is equivalent to representing the original data (i.e., abstract low-dimensional features) with a smaller amount of data. The reason why feature fusion and classifier each correspond to a category is that each category has a different degree of emphasis on abstract features, and also for flexible combination and pushing unknown reasons.
[0059] The feature module and classifier structure proposed in this invention are relatively simple because of insufficient data volume, and the fact that the model is not necessarily better the larger and more complex it is. With a small amount of data, it is sufficient as long as it is suitable for the task volume. A huge model may overfit and is not conducive to deployment on mobile devices.
[0060] In this invention, the classification categories include {0: 'Unknown cause', 1: 'Transformer burnout', 2: 'External force', 3: 'Wire detachment', 4: 'Foreign object contact', 5: 'Porcelain insulator detachment', 6: 'Cable fault', 7: 'Surge arrester breakdown', 8: 'Needle insulator fault', 9: 'Lightning strike'}.
[0061] In this invention, each classifier only determines whether a product belongs to that category, rather than performing multi-classification directly. The main reason is that the classification task of fault causes will change due to actual needs, i.e., customer needs. For example, customer A is concerned about causes 1 and 2, and other causes do not need to be analyzed, while customer B is concerned about causes 2 and 3. In this way, each category can be flexibly matched with its own classifier. At the same time, when faced with unseen samples, they will be classified as unknown causes rather than being forcibly classified into a certain set category.
[0062] 10. Output results.
[0063] Example 2 This invention also proposes a fault cause analysis system for low-current grounding systems, including a topology acquisition module, a device waveform acquisition module, a waveform processing module, a discriminator discrimination module, and an analysis module: The topology acquisition module acquires the topology of the system when a low-current ground fault occurs. The device waveform acquisition module locates the fault point in the topology and acquires the device waveform of the device that is physically closest to the fault point. The waveform processing module extracts the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the waveform recorded by the device; performs cycle truncation on the waveform recorded by the device based on the length of the single cycle and a set step size; and processes the truncated cycles. The discriminator discrimination module inputs the zero-sequence voltage, the zero-sequence current, and the processed cycle into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; The analysis module, based on the classification probability of different types of fault causes, selects the final analysis result according to actual needs.
[0064] Example 3 The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps according to the method described above.
[0065] Example 4 The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.
Claims
1. A method for analyzing the causes of faults in low-current grounding systems, characterized in that, include: When a low-current ground fault occurs, obtain the topology of the low-current ground fault system. Locate the fault point in the topology and obtain the device waveform of the device that is physically closest to the fault point; Extract the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the device waveform recording; truncate the device waveform recording based on the length of the single cycle and a set step size; process the truncated cycles. The zero-sequence voltage, the zero-sequence current, and the processed cycle are input into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; Based on the classification probability of different types of fault causes, the final judgment result is selected according to actual needs.
2. The fault cause analysis method for low-current grounding systems according to claim 1, characterized in that: The processing of the intercepted frequency specifically includes: The extracted cycles are resampled and normalized. The resampled and normalized cycles are then decomposed using FFT to obtain the zero-sequence voltage FFT result and the zero-sequence current FFT result.
3. The fault cause analysis method for low-current grounding systems according to claim 1, characterized in that: The discriminator includes a feature module and a classifier, and its specific structure includes: The feature module includes an input layer, a feature extraction part, a feature fusion part, and an output layer; the input layer reads the zero-sequence voltage, the zero-sequence current, and the processed cycle. The discriminators share the same network parameters for the same feature extraction part, while different discriminators have feature fusion parts and classifiers with different network parameters.
4. The fault cause analysis method for low-current grounding systems according to claim 3, characterized in that: The input format of the input layer is: The format is [m,n,4,256], where m represents the number of input samples, n represents the number of cycles of the input samples, 4 represents four channels, and 256 is the default number of nodes in one cycle; The four channels are input sequentially: zero-sequence voltage, zero-sequence voltage FFT result, zero-sequence current, and zero-sequence current FFT result, respectively. The output format of the output layer is [m, 128], where m represents the number of input samples and 128 is the dimension of the extracted abstract features.
5. The fault cause analysis method for low-current grounding systems according to claim 3, characterized in that: The structure of the feature extraction part is as follows: Part 1: One-dimensional convolutional layer, batch normalization layer, ReLU activation function layer, one-dimensional max pooling layer; the output of Part 1 serves as the input of Part 2. Part Two: Flattening layer, linear transformation layer, Dropout layer, ReLU activation function layer.
6. The fault cause analysis method for low-current grounding systems according to claim 3, characterized in that: The specific structure of the feature fusion part is as follows: The output of the feature extraction part is multiplied element-wise by the result of the linear transformation layer and the Softmax activation function layer, and the result of the multiplication is then passed through the linear transformation layer and the ReLU activation function layer to obtain the output.
7. The fault cause analysis method for low-current grounding systems according to claim 3, characterized in that: The specific structure of the classifier includes: The input layer receives the output from the feature module. The input format is [m, 128], where m represents the number of input samples and 128 is the feature dimension. The input layer's results are sequentially passed through a BatchNorm one-dimensional convolutional layer, a linear transformation layer, and a Sigmoid function layer to obtain the probability of belonging to the corresponding category.
8. A fault cause analysis system for a low-current grounding system using the method described in any one of claims 1-7, comprising a topology acquisition module, a device waveform acquisition module, a waveform processing module, a discriminator discrimination module, and an analysis module, characterized in that: The topology acquisition module acquires the topology of the system when a low-current ground fault occurs. The device waveform acquisition module locates the fault point in the topology and acquires the device waveform of the device that is physically closest to the fault point. The waveform processing module extracts the zero-sequence voltage and zero-sequence current, as well as the length of a single cycle, from the waveform recorded by the device; performs cycle truncation on the waveform recorded by the device based on the length of the single cycle and a set step size; and processes the truncated cycles. The discriminator discrimination module inputs the zero-sequence voltage, the zero-sequence current, and the processed cycle into several discriminators targeting different fault causes to obtain the classification probability of different types of fault causes; wherein, the discriminator includes a feature module and a classifier; The analysis module, based on the classification probability of different types of fault causes, selects the final analysis result according to actual needs.
9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.