Power distribution network fault detection method and system based on multi-scale neural network
By constructing a neural network model that includes a C3K2_EMA module and an SPPF structure, and introducing the NWD loss function, the problems of low fault identification accuracy and insufficient multi-source information collaborative analysis in traditional models are solved. This enables accurate fault location and real-time status assessment in distribution networks, improving the accuracy and efficiency of fault detection.
Patent Information
- Application Number
- CN202610130564.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-30
- Publication Date
- 2026-03-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing technologies, traditional neural network models have limited generalization ability when dealing with high-dimensional nonlinear relationships, resulting in low fault identification accuracy and insufficient ability to perform multi-dimensional information collaborative analysis, making it difficult to achieve accurate fault location and real-time status assessment in distribution networks.
A neural network model containing C3K2_EMA modules and SPPF structures is constructed, and the normalized Wasserstein distance NWD loss function is introduced. Data preprocessing is performed through robust scaling and principal component analysis. Combined with multi-scale feature extraction and fault dynamic tensor analysis, the feature extraction and fault identification capabilities of the model are improved.
It enables precise location and real-time status assessment of distribution network faults, improves fault identification accuracy and multi-source information collaborative analysis capabilities, enhances the robustness and adaptability of the model, and improves the accuracy and efficiency of fault detection.
Smart Images
Figure CN121656750A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power distribution network technology, and in particular relates to a method and system for power distribution network fault detection based on multi-scale neural networks. Background Technology
[0002] Currently, power distribution networks are undergoing rapid evolution towards automation and informatization, leading to increasingly sophisticated requirements for operation and management. As the core terminal unit in overhead lines, feeder units (FTUs) play a crucial role in fault detection, location, and isolation. With the widespread adoption of integrated dual-core systems in the power distribution field, the operational data generated by FTUs is characterized by its large volume, diverse types, and high collection frequency, creating favorable conditions for intelligent analysis. The intelligent development of power distribution networks places higher demands on fault identification accuracy and condition monitoring capabilities, driving the research and application of related technologies.
[0003] In existing technologies, FTU condition monitoring mainly relies on traditional neural network models for data analysis and achieves equipment condition assessment through a combination of manual inspection and single-point data judgment. These methods are typically based on monitoring single parameters such as current and voltage, using preset thresholds or simple models for fault identification. At the same time, some systems attempt to integrate time-series data such as current waveforms and temperature and humidity information to support basic condition analysis.
[0004] However, existing technologies have obvious limitations: traditional neural network models have limited generalization ability when dealing with high-dimensional nonlinear relationships, resulting in unsatisfactory prediction results for complex fault modes; at the same time, existing monitoring methods lack effective collaborative analysis capabilities for diverse heterogeneous information (such as transient current and equipment temperature rise), making it difficult to achieve accurate real-time status assessment and limiting the efficiency of fault identification and handling. Summary of the Invention
[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for detecting distribution network faults based on multi-scale neural networks, so as to solve the problems of low fault identification accuracy and insufficient multi-information collaborative analysis capability in the prior art, thereby realizing accurate location of distribution network faults and real-time status assessment.
[0006] Technical solution: The power distribution network fault detection method of the present invention includes the following steps:
[0007] S1. Collect the operating parameters and measured electrical values of the feeder terminal unit (FTU) through the electrical transformer to form an electrical dataset;
[0008] S2. Preprocess the electrical data collected by S1 and replace abnormal data;
[0009] S3. Divide the preprocessed electrical dataset into training set, validation set and test set according to a preset ratio;
[0010] S4. Construct a neural network model containing C3K2_EMA modules and SPPF structures, and introduce the standardized Wasserstein distance NWD loss function for regression training;
[0011] S5. Use the training set, validation set and test set to train and test the neural network model to obtain the trained FTU fault detection model;
[0012] S6. Use the trained FTU fault detection model to obtain FTU fault detection results.
[0013] This invention constructs an electrical dataset by collecting FTU operating parameters and measuring electrical values from electrical transformers. Abnormal data is replaced through preprocessing to improve data quality, and the dataset is proportionally divided to ensure the effectiveness of model training. A neural network model incorporating C3K2_EMA modules and an SPPF structure is built, and an NWD loss function is introduced to enhance the model's ability to collaboratively analyze multi-source information and improve feature extraction efficiency. The model is trained and tested using training, validation, and test sets, ultimately forming a high-precision FTU fault detection model. This effectively solves the problems of low fault identification accuracy and insufficient collaborative analysis capability of multi-source information in traditional methods, achieving accurate fault location and real-time status assessment in distribution networks.
[0014] Preferably, the electrical data in the electrical dataset described in step S1 includes voltage, current, voltage angle, current angle, line frequency, line active power, line reactive power, line apparent power, power factor, fault current, closing signal, opening signal, no energy storage indicator, longitude, latitude, number of master station communications, remaining capacity of backup power supply, temperature and humidity, CPU load, opening and closing current, battery voltage, and device voltage.
[0015] By collecting diverse data including electrical quantities such as voltage, current, and power, status quantities such as switch position and fault current, as well as device geographical location, communication status, power supply capacity, and internal operating parameters, a comprehensive dataset reflecting the electrical status of the line and the operating status of the terminal itself was constructed. This laid a solid data foundation for subsequent accurate analysis and significantly improved the adaptability of the fault detection model to complex operating conditions and the comprehensiveness of the status assessment.
[0016] Preferably, step S2, which involves preprocessing the electrical data collected in step S1 and replacing abnormal data, includes:
[0017] S21. Robust scaling processing is applied to the collected electrical data. The calculation formula is as follows:
[0018]
[0019] In the formula, X is the original data vector, i.e., all data points of a certain feature; Median(X) represents the median of the feature data; and IQR(X) represents the interquartile range of the feature data, specifically calculated as IQR = Q3 - Q1, where Q3 is the upper quartile and Q1 is the lower quartile. This represents the new data after robust outlier handling;
[0020] S22. Combine multiple feature data that have undergone robust scaling into a standardized data matrix:
[0021]
[0022] In the formula, X1 is an n×d dimensional matrix, where n is the number of samples and d is the number of features. This represents the new data after robust outlier handling;
[0023] S23. Calculate the correlation coefficient matrix between features based on the standardized data matrix:
[0024]
[0025] In the formula, R is the correlation coefficient matrix. It is a data matrix transpose;
[0026] S24. Perform eigenvalue decomposition on the correlation coefficient matrix to extract eigenvalues and corresponding eigenvectors:
[0027]
[0028] In the formula It is the i-th eigenvalue. These are the corresponding eigenvectors, and all eigenvectors together form the feature matrix. ;
[0029] S25. Determine the number of principal components to be retained based on the cumulative contribution rate of the eigenvalues, such that:
[0030]
[0031] In the formula, m is the number of principal components after screening, and u is the screening index;
[0032] S26. Project the original data into a low-dimensional space using the filtered principal components to obtain the dimension-reduced matrix:
[0033]
[0034] In the formula, Y is the projection matrix; Y is the dimension-reduced matrix used to represent the processed historical FTU running data.
[0035] This preprocessing method effectively suppresses outlier interference through robust scaling, extracts the main features inherent in the data using the correlation coefficient matrix and eigenvalue decomposition, and achieves data dimensionality reduction by selecting principal components based on contribution rate. While retaining key information, it eliminates redundancy and noise, thus providing high-quality and highly representative input data for subsequent model training, significantly improving the accuracy and robustness of fault detection.
[0036] Preferably, the ratio of the training set, validation set, and test set in step S3 is 7:2:1.
[0037] By dividing the dataset into training, validation, and test sets in a reasonable proportion, we can ensure that the model has sufficient data to learn and master generalization rules, effectively adjust hyperparameters and prevent overfitting during training using the validation set, and finally evaluate the model performance unbiasedly through the independent test set. This lays a solid foundation for obtaining a reliable and highly generalizable fault detection model.
[0038] Preferably, the implementation process of the C3K2_EMA module in step S4 includes:
[0039] S41. Process the input feature Y sequentially through a convolutional layer, a batch normalization layer, and the SiLU activation function to obtain the output feature:
[0040]
[0041] In the formula, Y represents the processed historical FTU data, Conv represents the convolutional layer, BN represents the batch normalization layer, and SiLU represents the SiLU activation function. This indicates the output characteristics of the module;
[0042] S42. Divide the output feature into two parts along the channel dimension:
[0043]
[0044] In the formula, Proceed to the next branch for calculation. Then it is passed directly to the next output tensor. split means to divide the input tensor into multiple output tensors along the specified dimension. dim means the specified dimension and CHANL means the index of the channel dimension.
[0045] S43, The formula is derived from a sequence of n C3K modules:
[0046]
[0047] In the formula, C3K includes two 2×2 convolution kernels. These are the input features for the previous layer. For output features;
[0048] S44, Process the branch and The data is concatenated along the channel dimension and then fused through a convolutional layer to obtain the output. The formula is:
[0049]
[0050] In the formula, Conv represents a convolutional layer, and Concat represents a concatenation function. This is the output result after passing through the C3K2_EMA module.
[0051] This module extracts basic features through a combination of convolution, normalization, and activation functions, and constructs branch paths using channel segmentation and splicing structures. It combines multi-level sub-modules for deep feature collaborative analysis and finally fuses the information of each path through convolution, effectively enhancing the model's ability to capture complex nonlinear relationships between FTU multivariate operating parameters, thereby improving the richness and robustness of feature representation.
[0052] Preferably, the implementation process of the SPPF structure described in step S4 includes:
[0053] S45. The features output by the C3K2_EMA module are first compressed in the channel dimension through a standard 1×1 convolution to obtain... :
[0054]
[0055] In the formula, Conv represents a 1×1 convolution, X in This indicates the output of the previous module;
[0056] S46, will Multi-scale processing is performed by feeding the data into a max-pooling layer with a shared convolutional kernel. This shared pooling layer uses a fixed pooling kernel and simulates the pooling effect of different receptive fields through cascading, thereby reducing the number of parameters and maintaining computational efficiency.
[0057]
[0058] In the formula, This indicates a max pooling operation with kernel size k = 5, stride s = 1, and padding p = 2, where Y0 equals the output of the previous layer. Y1 represents the result of Y0 after max pooling, Y2 represents the result of Y1 after max pooling, and Y3 represents the result of Y2 after max pooling.
[0059] S47. Combine and fuse features at different scales to merge the original features. The output after triple pooling , , Perform splicing along the channel dimension:
[0060]
[0061] S48. Use a shared convolutional layer to fuse the concatenated multi-scale features and compress their channel count back to the original dimension. This layer uses 1×1 convolutions and applies a uniform linear transformation to all input channels.
[0062]
[0063] In the formula, The characteristics after fusion It is a shared convolution weight matrix. It is the shared weight offset.
[0064] S49. Merged features The features are added to the input features to form the final output feature map. :
[0065] .
[0066] This SPPF structure efficiently extracts multi-scale features through serial max pooling and utilizes concatenation and convolution to fuse information from different receptive fields. This significantly increases the receptive field of the feature map while maintaining computational efficiency, thereby enhancing the model's ability to capture the evolution of fault electrical features at different spatiotemporal scales and providing richer contextual information for accurate localization and state assessment.
[0067] Preferably, the expression for the NWD loss function in step S4 is:
[0068]
[0069] In the formula, Indicates the actual fault type and magnitude. Indicates the predicted fault type and magnitude. To predict the center location coordinates of the fault type, To predict the magnitude of the fault type, The coordinates of the center location of the actual fault type. Where C represents the size of the actual fault type, and C is a user-defined constant. Let represent the square of the l2 norm, and exp represent the exponential function used to map the distance to the range (0,1).
[0070] The NWD loss function, by jointly optimizing the difference between the predicted and actual fault location and size in an exponential form related to a distribution metric, can smoothly handle the regression error between fault features of different scales, effectively enhance the model's ability to distinguish fault location and type, and thus improve the overall accuracy and robustness of fault detection and localization.
[0071] Preferably, step S6, which involves using the trained FTU fault detection model to obtain the FTU fault detection result, includes:
[0072] Construct a fault dynamic tensor based on fault dynamic features;
[0073] The fault dynamic tensor is input into the pre-trained C3k2_EMA module for multi-scale deep feature extraction to obtain an enhanced feature map.
[0074] The enhanced feature map is subjected to channel compression and projection dimensionality reduction to obtain a low-dimensional feature matrix;
[0075] The low-dimensional feature matrix is input into a preset Softmax classifier for classification to obtain an initial classification matrix and its mapped probability matrix; wherein, the elements of the classification matrix represent fault type labels, and the elements of the probability matrix represent the initial confidence of the corresponding fault type.
[0076] Determine whether the confidence level of each element in the probability matrix reaches a preset reliability threshold. If not, mark these low confidence labels and their probability values to generate a matrix to be reviewed.
[0077] When the confidence scores of all labels in the matrix to be verified meet the requirements, the standard classification matrix is obtained.
[0078] The final fault type monitoring report is generated based on the standard classification matrix.
[0079] This step constructs a fault dynamic tensor and extracts multi-scale deep features using a pre-trained model. After dimensionality reduction, the initial results are generated by a classifier. Innovatively, a confidence-based verification mechanism is introduced to label and reprocess low-confidence samples, thereby ensuring that the final output fault type monitoring report has high reliability and accuracy, and realizing accurate identification and stability assessment of the fault status of the distribution network.
[0080] Preferably, the fault type monitoring report includes the following:
[0081] Identified fault types and confidence levels;
[0082] The spatiotemporal distribution information of fault characteristics includes the fault initiation time, fault duration, and fault electrical characteristics;
[0083] Differentiated handling recommendations based on fault type;
[0084] The report metadata and model version information include the timestamp of report generation, the version number of the intelligent diagnostic model used, and the FTU device number from which the data was sourced.
[0085] The fault type monitoring report generated by this technical solution not only provides high-confidence fault identification results and detailed spatiotemporal feature analysis of faults, but also gives differentiated handling suggestions based on the fault type, and includes complete report metadata. This provides maintenance personnel with comprehensive information from accurate diagnosis to decision support, greatly improving the efficiency of fault handling and the level of intelligence in distribution network operation and maintenance.
[0086] Secondly, the power distribution network fault detection system of the present invention includes:
[0087] The data acquisition module is used to collect the operating parameters and measured electrical values of the feeder terminal unit (FTU) through the electrical transformer, forming an electrical dataset;
[0088] The data preprocessing module is used to preprocess the electrical data collected by the data acquisition module and replace abnormal data therein;
[0089] The dataset partitioning module is used to divide the preprocessed electrical dataset into training set, validation set and test set according to a preset ratio;
[0090] The model building and training module is used to build a neural network model containing a C3K2_EMA module and an SPPF structure. The standardized Wasserstein distance (NWD) loss function is introduced for regression training. The neural network model is trained and tested using the training set, validation set, and test set to obtain a trained FTU fault detection model.
[0091] The fault detection execution module is used to obtain FTU fault detection results using the trained FTU fault detection model.
[0092] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the described method for detecting power distribution network faults based on multi-scale neural networks.
[0093] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned power distribution network fault detection method based on a multi-scale neural network.
[0094] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. By constructing a neural network model containing C3K2_EMA modules and an SPPF structure, and introducing the NWD loss function for regression training, this invention effectively improves the accuracy of fault identification and the ability to collaboratively analyze multi-source information, achieving accurate fault location and real-time status assessment in the distribution network; 2. This invention adopts a branched C3K2_EMA module structure, splitting the input features along the channel dimension and processing them in parallel by multiple C3K sub-modules, enhancing the model's ability to extract features from multi-source operating parameters of the FTU, enabling the network to simultaneously learn the complex relationships between electrical quantities, state quantities, and environmental quantities, significantly improving the diversity of fault feature expression; 3. This invention utilizes the serial multi-channel structure in the SPPF structure... The scaled pooling mechanism effectively captures fault types at different time scales, from transient pulses to continuous disturbances, while maintaining computational efficiency, significantly improving the ability to detect weak faults; 4. This invention uses the NWD loss function to replace the traditional bounding box regression loss, and optimizes the regression training process for imbalanced fault samples through Gaussian distribution modeling and Wasserstein distance metric, effectively alleviating the problem of missed detection caused by the imbalance of positive and negative sample ratios, and improving the detection robustness of the model in real complex power distribution scenarios; 5. This invention combines robust scaling and principal component analysis preprocessing, which significantly reduces the data dimensionality while retaining key fault features, improving both model training efficiency and the feasibility of deploying the algorithm on resource-constrained FTU terminal equipment. Attached Figure Description
[0095] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0096] Figure 2 This is a model framework diagram of the present invention;
[0097] Figure 3 This is a schematic diagram of the C3K2_EMA structure of the present invention;
[0098] Figure 4 This is a schematic diagram of the SPPF of the present invention. Detailed Implementation
[0099] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0100] Example 1
[0101] This embodiment provides a distribution network fault detection method based on a multi-scale neural network, such as... Figure 1 As shown, it includes the following steps:
[0102] S1: Acquire electrical data by using the electrical transformers present in the power system to collect the operating parameters and measured electrical values of the FTU;
[0103] Specifically, in this step, various electrical transformers and sensors pre-installed at key nodes of the distribution network are used to collect comprehensive, multi-dimensional data on the operating status and related electrical quantities of the feeder terminal units (FTUs), achieved through current transformers and voltage transformers. These transformers are pre-installed on the power feeders (cable or overhead line sections from substation outgoing lines to user access points) monitored by the FTU, responsible for accurately converting the raw high-voltage, high-current signals in the line into standardized low-level signals that can be processed by the FTU's internal circuitry. Based on these basic electrical quantities, the FTU obtains key parameters characterizing the energy transmission status of the line through internal calculations or direct measurement. These include active power, reactive power, apparent power, and power factor. These data collectively depict the distribution of power flow in the distribution network, which is crucial for determining line load rates, optimizing network operation, and identifying certain abnormal states (such as voltage anomalies caused by reactive power deficits). To ensure the reliable operation of the FTU in complex field environments, it is also necessary to collect data on its operating environment and communication link status.
[0104] S2: Replace outlier values in the collected data. Specific implementation steps include:
[0105] S21: Robust scaling is applied to the acquired abnormal data. The specific formula is as follows:
[0106]
[0107] In the formula, X is the original data vector (all data points for a certain feature), Median(X) represents the median of the feature data, and IQR(X) represents the interquartile range of the feature data, specifically calculated as IQR = Q3 - Q1. Here, Q3 is the upper quartile, and Q1 is the lower quartile. This represents the new data after robust outlier handling.
[0108] S22: Combine multiple robustly processed feature data into a standardized data matrix.
[0109]
[0110] In the formula, X1 is an n×d dimensional matrix, where n is the number of samples and d is the number of features. This represents the new data after robust outlier handling.
[0111] S23: Calculate the correlation coefficient matrix between features based on the standardized data matrix.
[0112]
[0113] In the formula, R is the correlation coefficient matrix. It is a data matrix The transpose of .
[0114] S24: Perform eigenvalue decomposition on the correlation coefficient matrix to extract the main eigenvectors and eigenvalues.
[0115]
[0116] In the formula It is the i-th eigenvalue. These are the corresponding eigenvectors, and all eigenvectors together form the feature matrix.
[0117] S25: Determine the number of principal components to be retained based on the cumulative contribution rate of eigenvalues.
[0118]
[0119] In the formula, m is the number of principal components after screening, and u is the screening index.
[0120] S26: Project the original data into a low-dimensional space using the filtered principal components.
[0121]
[0122] In the formula, Y is the projection matrix; Y is the dimension-reduced matrix used to represent the processed historical FTU running data.
[0123] S3: Divide the processed data into training set, test set and validation set in a ratio of 7:2:1;
[0124] S4: Construct a new neural network model, train it using the training set, and validate it using the validation set;
[0125] Reference Figure 2 The newly constructed model includes the following: First, the C3K2_EMA module is introduced to smooth weight updates, avoiding training instability caused by excessively large learning rates or gradient explosion, thus improving the stability of the training process. Second, the SPPF structure is introduced, which uses a multi-scale pooling mechanism to enhance the network's ability to identify minute faults, improving the model's ability to capture details and effectively reducing the number of parameters and computational cost. Third, the NWD loss function is used instead of the traditional cross-entropy loss function, and the Wasserstein distance metric for fault features is modeled using a Gaussian distribution, optimizing the training process for unbalanced distribution fault samples and effectively mitigating the problem of missed detections caused by a much larger number of normal samples than fault samples.
[0126] like Figure 3 The C3K2_EMA module was designed, and its specific implementation steps include:
[0127] S41: The specific formula for feeding the input feature Y into the CBS module is as follows:
[0128]
[0129] In the formula, Y represents the processed historical FTU data, Conv represents the convolutional layer, BN represents the batch normalization layer, and SiLU represents the SiLU activation function. This indicates the output characteristics of the module.
[0130] S42: The result processed by the CBS module The data is fed into the Splite module, which divides it into two parts along the channel dimension. The specific formula is as follows:
[0131]
[0132] In the formula, Proceed to the next branch for calculation. Then it is passed directly to the next input tensor. `split` means to divide the input tensor into multiple output tensors along the specified dimension. `dim` represents the specified dimension, and `CHANL` represents the index of the channel dimension.
[0133] S43: It will pass through a module sequence consisting of n C3K modules, and the specific formula is:
[0134]
[0135] In the formula, C3K includes two 2×2 convolution kernels. These are the input features for the previous layer. For output features.
[0136] S44: Process the branch and The data is concatenated along the channel dimension and then fused through a convolutional layer to obtain the output M. The specific formula is as follows:
[0137]
[0138] In the formula The output of the C3K2_EMA module
[0139] like Figure 4 The specific steps for introducing the SPPF module include:
[0140] S45: The features output by the C3K2_EMA module are first compressed in the channel dimension through a standard 1×1 convolution, resulting in...
[0141]
[0142] In the formula, Conv represents a 1×1 convolution, X in Indicates the output of the previous module
[0143] S46: Will Multi-scale processing is performed by feeding the data into a max-pooling layer with a shared convolutional kernel. This shared pooling layer uses a fixed pooling kernel and simulates the pooling effect of different receptive fields through a cascaded approach, thereby reducing the number of parameters and maintaining computational efficiency.
[0144]
[0145] In the formula, This indicates a max pooling operation with kernel size k = 5, stride s = 1, and padding p = 2, where Y0 equals the output of the previous layer. Y1 represents the result of Y0 after max pooling, Y2 represents the result of Y1 after max pooling, and Y3 represents the result of Y2 after max pooling.
[0146] S47: This involves stitching and fusing features at different scales, combining the original features... The output after triple pooling , , Then, the components are spliced together along the channel dimension.
[0147]
[0148] S48: An efficient shared convolutional layer is used to fuse the concatenated multi-scale features and compress their channel count back to the original dimension. This layer uses 1×1 convolutions and applies a uniform linear transformation to all input channels, further improving the model's quantization level.
[0149]
[0150] in It is a shared convolution weight matrix. It is the shared weight offset.
[0151] S49: Finally, the fused features The features are added to the input features to form the final output feature map Xout.
[0152]
[0153] Furthermore, this model employs the NWD function, the expression of which is:
[0154]
[0155] In the formula, Indicates the actual fault type and magnitude. Indicates the predicted fault type and magnitude. To predict the center location coordinates of the fault type, To predict the magnitude of the fault type, The coordinates of the center location of the actual fault type. Where C represents the size of the actual fault type, and C is a user-defined constant. Let represent the square of the l2 norm, and exp represent the exponential function used to map the distance to the range (0,1).
[0156] S5: Use the model to dynamically identify the broken wire area, analyze the fault type based on the dynamic characteristics of the fault, and obtain a fault type monitoring report.
[0157] In a specific embodiment, fault type analysis based on fault dynamic features is performed to obtain a fault type monitoring report, including: constructing a fault dynamic tensor based on the fault dynamic features; inputting the fault dynamic tensor into a pre-trained C3k2_EMA module for multi-scale deep feature extraction to obtain an enhanced feature map; performing channel compression and projection dimensionality reduction on the enhanced feature map to obtain a low-dimensional feature matrix; inputting the low-dimensional feature matrix into a preset Softmax classifier for one classification to obtain an initial classification matrix and its mapped probability matrix; wherein, the elements of the classification matrix represent fault type labels, and the elements of the probability matrix represent the initial confidence of the corresponding fault type; determining whether the confidence of each element in the probability matrix reaches a preset reliability threshold; if not, marking these low-confidence labels and their probability values to generate a verification matrix; when the confidence of all labels in the verification matrix meets the requirements, a standard classification matrix is obtained; and generating a final fault type monitoring report based on the standard classification matrix. The report includes: identified fault types and confidence levels, clearly listing the fault types ultimately determined through the standard classification matrix; spatiotemporal distribution information of fault characteristics, including: fault initiation time, fault duration, and fault electrical characteristics; differentiated handling recommendations based on fault type, providing targeted operational guidance; and report metadata and model version information, including the report generation timestamp, the version number of the intelligent diagnostic model used, and the FTU device number from which the data was sourced.
[0158] Example 2
[0159] This embodiment provides a distribution network fault detection system corresponding to the aforementioned distribution network fault detection method, including:
[0160] The data acquisition module is used to collect the operating parameters and measured electrical values of the feeder terminal unit (FTU) through the electrical transformer, forming an electrical dataset;
[0161] The data preprocessing module is used to preprocess the electrical data collected by the data acquisition module and replace abnormal data therein;
[0162] The dataset partitioning module is used to divide the preprocessed electrical dataset into training set, validation set and test set according to a preset ratio;
[0163] The model building and training module is used to build a neural network model containing a C3K2_EMA module and an SPPF structure. The standardized Wasserstein distance (NWD) loss function is introduced for regression training. The neural network model is trained and tested using the training set, validation set, and test set to obtain a trained FTU fault detection model.
[0164] The fault detection execution module is used to obtain FTU fault detection results using the trained FTU fault detection model.
[0165] Example 3
[0166] This embodiment provides an electronic device.
[0167] Specifically, the electronic device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.
[0168] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor executes various functional applications and data processing by running non-transitory software programs, instructions, and modules stored in memory. Memory may include a program storage area and a data storage area. The program storage area may store the control unit and the application program required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, memory may include high-speed random access memory and non-transitory memory. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
[0169] Example 4
[0170] This embodiment provides a computer-readable storage medium.
[0171] Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above method implementation.
[0172] Those skilled in the art will understand that all or part of the processes in the methods described above can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.
Claims
1. A method for fault detection in distribution networks based on multi-scale neural networks, characterized in that, Includes the following steps: S1. Collect the operating parameters and measured electrical values of the feeder terminal unit (FTU) through the electrical transformer to form an electrical dataset; S2. Preprocess the electrical data collected by S1 and replace abnormal data; S3. Divide the preprocessed electrical dataset into training set, validation set and test set according to a preset ratio; S4. Construct a neural network model containing C3K2_EMA modules and SPPF structures, and introduce the standardized Wasserstein distance NWD loss function for regression training; S5. Use the training set, validation set and test set to train and test the neural network model to obtain the trained FTU fault detection model; S6. Use the trained FTU fault detection model to obtain FTU fault detection results.
2. The method according to claim 1, characterized in that, The electrical data in the electrical dataset mentioned in step S1 includes voltage, current, voltage angle, current angle, line frequency, line active power, line reactive power, line apparent power, power factor, fault current, closing signal, opening signal, no energy storage indicator, longitude, latitude, number of master station communications, remaining capacity of backup power supply, temperature and humidity, CPU load, opening and closing current, battery voltage, and device voltage.
3. The method according to claim 1, characterized in that, Step S2, which involves preprocessing the electrical data collected in step S1 and replacing abnormal data, includes: S21. Robust scaling processing is applied to the collected electrical data. The calculation formula is as follows: In the formula, X is the original data vector, i.e., all data points of a certain feature; Median(X) represents the median of the feature data; and IQR(X) represents the interquartile range of the feature data, specifically calculated as IQR = Q3 - Q1, where Q3 is the upper quartile and Q1 is the lower quartile. This represents the new data after robust outlier handling; S22. Combine multiple feature data that have undergone robust scaling into a standardized data matrix: In the formula, X1 is an n×d dimensional matrix, where n is the number of samples and d is the number of features. This represents the new data after robust outlier handling; S23. Calculate the correlation coefficient matrix between features based on the standardized data matrix: In the formula, R is the correlation coefficient matrix. It is a data matrix Transpose of; S24. Perform eigenvalue decomposition on the correlation coefficient matrix to extract eigenvalues and corresponding eigenvectors: In the formula It is the i-th eigenvalue. These are the corresponding eigenvectors, and all eigenvectors together form the feature matrix. ; S25. Determine the number of principal components to be retained based on the cumulative contribution rate of the eigenvalues, such that: In the formula, m is the number of principal components after screening, and u is the screening index; S26. Project the original data into a low-dimensional space using the filtered principal components to obtain the dimension-reduced matrix: In the formula, Y is the projection matrix; Y is the dimension-reduced matrix used to represent the processed historical FTU running data.
4. The method according to claim 1, characterized in that, The ratio of the training set, validation set, and test set mentioned in step S3 is 7:2:
1.
5. The method according to claim 1, characterized in that, The implementation process of the C3K2_EMA module described in step S4 includes: S41. Process the input feature Y sequentially through a convolutional layer, a batch normalization layer, and the SiLU activation function to obtain the output feature: In the formula, Y represents the processed historical FTU data, Conv represents the convolutional layer, BN represents the batch normalization layer, and SiLU represents the SiLU activation function. Indicates the output characteristics of the module; S42. Divide the output feature into two parts along the channel dimension: In the formula, Proceed to the next branch for calculation. Then it is passed directly to the next output tensor. split means to divide the input tensor into multiple output tensors along the specified dimension. dim means the specified dimension and CHANL means the index of the channel dimension. S43, The formula is derived from a sequence of n C3K modules: In the formula, C3K includes two 2×2 convolution kernels. These are the input features for the previous layer. For output features; S44, Process the branch and The data is concatenated along the channel dimension and then fused through a convolutional layer to obtain the output. The formula is: In the formula, Conv represents a convolutional layer, and Concat represents a concatenation function. This is the output result after passing through the C3K2_EMA module.
6. The method according to claim 1, characterized in that, The implementation process of the SPPF structure described in step S4 includes: S45. The features output by the C3K2_EMA module are first compressed in the channel dimension through a standard 1×1 convolution to obtain... : In the formula, Conv represents a 1×1 convolution, X in This indicates the output of the previous module; S46, will Multi-scale processing is performed by feeding the data into a max-pooling layer with a shared convolutional kernel. This shared pooling layer uses a fixed pooling kernel and simulates the pooling effect of different receptive fields through cascading, thereby reducing the number of parameters and maintaining computational efficiency. In the formula, This indicates a max pooling operation with kernel size k = 5, stride s = 1, and padding p = 2, where Y0 equals the output of the previous layer. Y1 represents the result of Y0 after max pooling, Y2 represents the result of Y1 after max pooling, and Y3 represents the result of Y2 after max pooling. S47. Combine and fuse features at different scales, integrating the original features... The output after triple pooling , , Perform splicing along the channel dimension: S48. Use a shared convolutional layer to fuse the concatenated multi-scale features and compress their channel count back to the original dimension. This layer uses 1×1 convolutions and applies a uniform linear transformation to all input channels. In the formula, The characteristics after fusion It is a shared convolution weight matrix. It is the shared weight offset. S49. Merged features The features are added to the input features to form the final output feature map. : 。 7. The method according to claim 1, characterized in that, The expression for the NWD loss function mentioned in step S4 is: In the formula, Indicates the actual fault type and magnitude. Indicates the predicted fault type and magnitude. To predict the center location coordinates of the fault type, To predict the magnitude of the fault type, The coordinates of the center location of the actual fault type. Where C represents the size of the actual fault type, and C is a user-defined constant. Let represent the square of the l2 norm, and exp represent the exponential function used to map the distance to the range (0,1).
8. The method according to claim 1, characterized in that, Step S6, which involves using the trained FTU fault detection model to obtain FTU fault detection results, includes: Construct a fault dynamic tensor based on fault dynamic features; The fault dynamic tensor is input into the pre-trained C3k2_EMA module for multi-scale deep feature extraction to obtain an enhanced feature map. The enhanced feature map is subjected to channel compression and projection dimensionality reduction to obtain a low-dimensional feature matrix; The low-dimensional feature matrix is input into a preset Softmax classifier for classification to obtain an initial classification matrix and its mapped probability matrix; wherein, the elements of the classification matrix represent fault type labels, and the elements of the probability matrix represent the initial confidence of the corresponding fault type. Determine whether the confidence level of each element in the probability matrix reaches a preset reliability threshold. If not, mark these low confidence labels and their probability values to generate a matrix to be reviewed. When the confidence scores of all labels in the matrix to be verified meet the requirements, the standard classification matrix is obtained. The final fault type monitoring report is generated based on the standard classification matrix.
9. The method according to claim 8, characterized in that, The fault type monitoring report includes the following: Identified fault types and confidence levels; The spatiotemporal distribution information of fault characteristics includes the fault initiation time, fault duration, and fault electrical characteristics; Differentiated handling recommendations based on fault type; The report metadata and model version information include the timestamp of report generation, the version number of the intelligent diagnostic model used, and the FTU device number from which the data was sourced.
10. A power distribution network fault detection system based on a multi-scale neural network, characterized in that, include: The data acquisition module is used to collect the operating parameters and measured electrical values of the feeder terminal unit (FTU) through the electrical transformer, forming an electrical dataset; The data preprocessing module is used to preprocess the electrical data collected by the data acquisition module and replace abnormal data therein; The dataset partitioning module is used to divide the preprocessed electrical dataset into training set, validation set and test set according to a preset ratio; The model building and training module is used to build a neural network model containing a C3K2_EMA module and an SPPF structure. The standardized Wasserstein distance (NWD) loss function is introduced for regression training. The neural network model is trained and tested using the training set, validation set, and test set to obtain a trained FTU fault detection model. The fault detection execution module is used to obtain FTU fault detection results using the trained FTU fault detection model.