Railway power distribution network fault information automatic diagnosis and positioning method and system

Abstract

The application discloses a railway power distribution network fault information automatic diagnosis and positioning method and system, belongs to the technical field of power distribution fault diagnosis and positioning, and comprises the following steps: generating an auxiliary classification generative adversarial network, obtaining a standardized training data set balanced in categories and performing preprocessing; mapping the preprocessed data in the standardized training data set into a snowflake-shaped SDP image in a two-dimensional polar coordinate system, inputting the SDP image into a fault diagnosis model based on a deep convolutional neural network, and outputting a fault section and a preliminary fault distance; constructing a physical circuit model of a railway power distribution network distribution parameter, and calculating a theoretical simulation voltage and a simulation current by using a transmission line equation; starting iteration correction based on the Pearson correlation coefficient of each data, and outputting a diagnosis result. Through the fusion of data enhancement, feature transformation, intelligent diagnosis, physical inversion checking and topological adaptive adjustment, the application realizes high-precision, high-reliability and self-adaptive fault information automatic diagnosis and positioning.

Description

Technical Field

[0001] This invention belongs to the field of power distribution fault diagnosis and location technology, specifically relating to an automatic diagnosis and location method and system for railway power distribution network fault information. Background Technology

[0002] The railway power distribution system is a dedicated non-traction load power supply network for railways, independent of the traction power supply system. The railway power distribution network consists of high-voltage through power lines laid parallel to the railway line, automatic block power lines, and box-type substations distributed by section. It is responsible for providing continuous power to signal interlocking devices, communication transmission equipment, train dispatching and command systems, and station power lighting facilities along the railway line. In terms of physical topology, the railway power distribution network exhibits the structural characteristics of multiple power supply points, long and narrow power supply lines, and chain-cascaded load nodes. The railway power distribution network adopts a dual-power-source dual-circuit power supply mode and a neutral point ungrounded operation mode to meet the strict requirements of the railway traffic command and control system for power supply reliability.

[0003] To address the issue of low fault diagnosis accuracy in railway power distribution networks under conditions of scarce fault samples and complex operation, existing technologies employ signal processing methods based on traveling wave principles and deep neural network fault diagnosis models driven by single data. However, these approaches still face challenges such as insufficient model training due to the extremely limited number of fault samples in actual operation, weak nonlinear transient characteristics of intermittent arcing and high-resistance grounding faults that are easily masked by noise, poor interpretability of output results due to the lack of physical mechanism support in purely data-driven models, and the failure of original fixed-parameter models when the power distribution network's operating topology changes. Consequently, these issues lead to insufficient accuracy in fault diagnosis and location in railway power distribution networks, high false alarm rates, and an inability to meet the high reliability maintenance requirements of the power supply system for railway traffic safety. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the present invention aims to provide an automatic diagnosis and location method and system for fault information in railway power distribution networks. By integrating data augmentation, feature transformation, intelligent diagnosis, physical inversion verification and topology adaptive adjustment, it achieves high-precision, high-reliability and adaptive automatic diagnosis and location of fault information.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] An automatic diagnosis and location method for fault information in railway power distribution networks includes the following steps:

[0007] S1. Collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. Use auxiliary classification generative adversarial network to generate virtual fault samples with fault category labels. Mix them with the original sample set to obtain a class-balanced standardized training dataset. Then, preprocess the standardized training dataset.

[0008] S2. Using the symmetric point pattern transformation technique, the preprocessed data in the standardized training dataset is mapped into a snowflake-shaped SDP image in a two-dimensional polar coordinate system.

[0009] S3. Input the snowflake-shaped SDP image into the fault diagnosis model based on a deep convolutional neural network, and output the fault section and the preliminary fault distance.

[0010] S4. Construct a physical circuit model of the railway power distribution network distribution parameters, determine the corresponding line impedance parameters and topology path according to the fault section, and substitute the preliminary fault distance into the physical circuit model in reverse, and use the transmission line equation to deduce the theoretical simulation voltage and simulation current.

[0011] S5. Calculate the Pearson correlation coefficients of the simulated voltage and transient zero-sequence voltage waveform data, and the simulated current and transient zero-sequence current waveform data, respectively. If the Pearson correlation coefficient is lower than the preset confidence threshold, start the iterative correction program to correct the preliminary fault distance until the error requirement is met and then output the diagnostic results.

[0012] As a preferred embodiment of the present invention, in step S1, the process of obtaining a class-balanced standardized training dataset and preprocessing the standardized training dataset includes:

[0013] Fault recording and monitoring devices are deployed at the busbar outgoing terminals of substations and on the high-voltage side of box-type substations along the railway power distribution network.

[0014] The analog input terminal of the fault recording monitoring device is connected to the secondary side of the voltage transformer and the secondary side of the current transformer in the power distribution network line via a cable.

[0015] Voltage transformers acquire the transient zero-sequence voltage analog signal of the line, and current transformers acquire the transient zero-sequence current analog signal of the line.

[0016] The analog-to-digital converter inside the fault recording and monitoring device converts the transient zero-sequence voltage analog signal and the transient zero-sequence current analog signal into a digital sequence;

[0017] When the transient zero-sequence voltage amplitude or transient zero-sequence current amplitude exceeds the corresponding preset voltage start threshold or current start threshold, the fault recording monitoring device extracts the digital sequence from one cycle before the start time to five cycles after the start time to form transient zero-sequence voltage recording data and transient zero-sequence current recording data.

[0018] The transient zero-sequence voltage and transient zero-sequence current waveforms collected at historical fault moments are compiled as the original sample set.

[0019] Construct an auxiliary classification generative adversarial network that includes a generator model and a discriminator model;

[0020] The generator model receives a random noise vector that follows a standard normal distribution generated by a random number generation algorithm and a fault category label vector selected from a preset fault category set. It then uses a deconvolutional neural network layer to upsample and feature map the random noise vector and the fault category label vector, mapping the low-dimensional noise distribution to a high-dimensional data space with the same dimension as the transient zero-sequence voltage waveform data and the transient zero-sequence current waveform data, and outputs a virtual fault sample with fault category labels.

[0021] The discriminator model receives input samples, which are either real fault samples from the original sample set or virtual fault samples generated by the generator model. The discriminator model uses a convolutional neural network layer to extract features from the input samples and outputs the probability of the input sample being a real or fake fault sample and the probability of the input sample belonging to a specific fault category.

[0022] During the training of the auxiliary classification generative adversarial network, the generator model parameters and the discriminator model parameters are updated by optimizing a composite objective function that includes data source discrimination loss and fault category classification loss.

[0023] After the auxiliary classification generative adversarial network is trained, the fault category label vectors of the original sample set that do not reach the preset threshold are input into the generator model. The generator model outputs virtual fault samples with the corresponding number of padded samples. The virtual fault samples are merged with the real fault samples in the original sample set so that the number of samples of each fault category is consistent after merging, forming a standardized training dataset with class balance.

[0024] Z-score normalization is performed on each transient zero-sequence voltage and transient zero-sequence current waveform in the normalized training dataset.

[0025] As a preferred embodiment of the present invention, the process of mapping the preprocessed data in the standardized training dataset to a snowflake-shaped SDP image in a two-dimensional polar coordinate system in S2 includes:

[0026] Set the mapping parameters required for the symmetric point mode transformation, including the time lag coefficient, the initial deflection angle of the polar coordinate system, and the angle mapping coefficient;

[0027] Using the symmetric point mode transformation technique, the preprocessed transient zero-sequence voltage waveform data and transient zero-sequence current waveform data are regarded as time series signals, and the polar radius and polar angle corresponding to each data point in the time series signal in the two-dimensional polar coordinate system are calculated.

[0028] Using minimum-maximum normalization logic, the data values ​​of the time series signal are mapped to a preset radius range, and the polar radius is calculated for the data points in the time series signal.

[0029] The data values ​​at the time lag points are mapped to angular offsets, and the initial deflection angle is added to calculate the polar angle for the data points in the time series signal.

[0030] Mapping points determined by polar radius and polar angle are plotted in a two-dimensional polar coordinate system. The mapping points corresponding to the data points in the time series signal are plotted in the same two-dimensional polar coordinate plane. Different initial deflection angles are set for transient zero-sequence voltage waveform data and transient zero-sequence current waveform data, so that the mapping points of different data sources are centrally symmetrically distributed in the two-dimensional polar coordinate system, forming a graphic with a symmetrical structure, and generating a snowflake-shaped SDP image that reflects the transient characteristics of the fault.

[0031] As a preferred embodiment of the present invention, the process of outputting the fault section and preliminary fault distance in S3 includes:

[0032] A fault diagnosis model based on a deep convolutional neural network is constructed. The structure of a deep convolutional neural network includes an input layer, a feature extraction layer, a fully connected layer, and a multi-task output layer.

[0033] The feature extraction layer consists of alternating convolutional layers and pooling layers. The convolutional layers use convolutional kernels to perform convolution operations on the input snowflake-shaped SDP image and combine them with non-linear activation functions to extract local spatial features of the image and generate feature maps. The pooling layers use max pooling logic to downsample the feature maps output by the convolution, retaining the maximum values ​​in local regions of the feature maps.

[0034] The multi-task output layer contains two parallel branch structures: a fault segment classification branch and a fault distance regression branch. The fault segment classification branch receives the output of the fully connected layer, uses a normalized exponential function to calculate the probability distribution of the input sample belonging to each fault segment, and selects the fault segment with the highest probability as the output fault segment. The fault distance regression branch receives the output of the fully connected layer, uses a linear activation function to map the input features into continuous values, and uses this as the initial fault distance for the output.

[0035] The deep convolutional neural network was trained using snowflake-shaped SDP images from the standardized training dataset as model input and the corresponding real fault segment labels and real fault distance labels as supervision signals.

[0036] During training, a composite loss function is calculated, which includes classification loss and regression loss.

[0037] The classification loss is constructed using cross-entropy calculation logic. By calculating the negative logarithm of the predicted probability corresponding to the actual fault segment label, the difference between the fault diagnosis model's predicted fault segment distribution and the actual distribution is measured.

[0038] The regression loss is constructed using the mean squared error calculation logic. The accuracy of the regression prediction is measured by calculating the square of the difference between the preliminary fault distance output by the fault diagnosis model and the normalized true fault distance.

[0039] The composite loss function is constructed as a weighted sum of classification loss and regression loss, and the proportion of classification task and regression task in the total loss is adjusted by introducing a balancing hyperparameter;

[0040] The gradient of the composite loss function with respect to the weight and bias parameters of each layer in the deep convolutional neural network is calculated using the backpropagation algorithm. The weight and bias parameters are then updated along the direction of gradient descent using the stochastic gradient descent optimization algorithm until the composite loss function converges, thus obtaining the trained fault diagnosis model.

[0041] During online diagnosis, the real-time generated snowflake-shaped SDP image is input into the trained fault diagnosis model. The deep convolutional neural network performs forward propagation calculation on the snowflake-shaped SDP image, extracts local spatial features of the image through the feature extraction layer, and outputs the classification results of the fault section and the regression value of the preliminary fault distance through the multi-task output layer.

[0042] As a preferred embodiment of the present invention, the process of deriving the theoretical simulation voltage and simulation current using the transmission line equation in S4 includes:

[0043] Based on the physical structure of high-voltage through power lines and automatic block power lines in railway distribution networks, a physical circuit model of distributed parameters in railway distribution networks is constructed. The physical circuit model equates the railway distribution network lines to a continuous distributed parameter circuit system composed of unit length resistance, unit length inductance, unit length conductance, and unit length capacitance.

[0044] Using the fault sections output by the fault diagnosis model, the physical attribute information corresponding to the fault sections is indexed in the geographic information database and line parameter database of the railway power distribution network;

[0045] The physical attribute information includes the conductor type, laying method, and topological connection relationship of the fault section relative to the outgoing line of the substation bus.

[0046] Based on physical property information, the zero-sequence unit resistance, zero-sequence unit inductance, zero-sequence unit conductance and zero-sequence unit capacitance of the fault section are extracted, and the line topology path length from the installation location of the fault recording monitoring device to the starting point of the fault section is determined.

[0047] By combining the line topology path length and the preliminary fault distance output by the fault diagnosis model, the relative position of the fault point in the physical circuit model is determined. Using the fault point as the boundary condition, the transmission line equation is used to describe the spatiotemporal distribution of transient zero-sequence voltage and transient zero-sequence current in the physical circuit model.

[0048] In the physical circuit model, the propagation process of transient zero-sequence voltage and transient zero-sequence current along the railway distribution network line satisfies a set of partial differential equations. The set of partial differential equations is used to characterize the spatial rate of change of transient zero-sequence voltage along the line as equal to the negative sum of the resistive voltage drop generated by the zero-sequence unit resistance and the inductive voltage drop generated by the zero-sequence unit inductance, and to characterize the spatial rate of change of transient zero-sequence current along the line as equal to the negative sum of the leakage current generated by the zero-sequence unit conductance and the displacement current generated by the zero-sequence unit capacitance.

[0049] The partial differential equations are discretized and solved using the finite difference method. The initial fault distance output by the fault diagnosis model is set as the boundary position parameter of the partial differential equations. Combined with the power supply side boundary conditions and load side boundary conditions of the railway power distribution network system, the numerical sequence of the partial differential equations at the installation location of the fault recording and monitoring device is solved over time, and the theoretical simulated voltage and simulated current are obtained respectively.

[0050] As a preferred embodiment of the present invention, the process of calculating the Pearson correlation coefficients of the simulated voltage and transient zero-sequence voltage waveform data, and the simulated current and transient zero-sequence current waveform data in S5 includes:

[0051] By calculating the Pearson correlation coefficient, the waveform similarity between simulated voltage and transient zero-sequence voltage recording data, and the waveform similarity between simulated current and transient zero-sequence current recording data are quantified respectively. The Pearson correlation coefficient includes the voltage Pearson correlation coefficient and the current Pearson correlation coefficient.

[0052] Calculate the voltage covariance between the simulated voltage and the transient zero-sequence voltage waveform data, and calculate the voltage standard deviation between the simulated voltage and the transient zero-sequence voltage waveform data respectively. Divide the voltage covariance by the product of the two voltage standard deviations to obtain the voltage Pearson correlation coefficient.

[0053] Calculate the current covariance between the simulated current and the transient zero-sequence current waveform data, and calculate the current standard deviation between the simulated current and the transient zero-sequence current waveform data respectively. Divide the current covariance by the product of the two current standard deviations to obtain the current Pearson correlation coefficient.

[0054] As a preferred embodiment of the present invention, the process of starting the iterative correction procedure to correct the initial fault distance until the error requirement is met and then outputting the diagnostic result includes:

[0055] Set a confidence threshold, compare the voltage Pearson correlation coefficient and the current Pearson correlation coefficient with the confidence threshold. When both the voltage Pearson correlation coefficient and the current Pearson correlation coefficient are greater than or equal to the confidence threshold, the preliminary fault distance output by the fault diagnosis model is deemed reliable, and the preliminary fault distance is output as the final diagnosis result.

[0056] When either the voltage Pearson correlation coefficient or the current Pearson correlation coefficient is below the confidence threshold, it is determined that the preliminary fault distance output by the fault diagnosis model has a deviation, and an iterative correction procedure is initiated. The iterative correction procedure includes:

[0057] Using the double-ended traveling wave method, the propagation speed of the traveling wave signal in the railway power distribution network and the time difference between the arrival of the traveling wave front at both ends of the line are calculated to estimate the theoretical location range of the fault point. Combined with the preliminary fault distance output by the fault diagnosis model, and with the preliminary fault distance as the center, and combined with the error margin determined by the theoretical location range, a search interval containing the actual location of the fault is constructed.

[0058] Within the search interval, a fine-tuning step size is set, and different corrected fault distances are selected by discretization within the search interval according to the fine-tuning step size. Each corrected fault distance is substituted into the physical circuit model, and the corresponding simulation voltage and simulation current are recalculated using the transmission line equation. The corresponding voltage Pearson correlation coefficient and current Pearson correlation coefficient are recalculated, and the weighted sum of the voltage Pearson correlation coefficient and current Pearson correlation coefficient is calculated as a comprehensive similarity index.

[0059] Iterate through the comprehensive similarity index corresponding to each corrected fault distance within the search interval, select the corrected fault distance corresponding to the maximum value of the comprehensive similarity index, and output it as the corrected preliminary fault distance.

[0060] As a preferred embodiment of the present invention, after mapping the preprocessed data in the standardized training dataset to a snowflake-shaped SDP image in a two-dimensional polar coordinate system, the method further includes a step of feature extraction and fusion of the snowflake-shaped SDP image, specifically:

[0061] Calculate the gray-level co-occurrence matrix of the snowflake-shaped SDP image and extract the image texture feature vector including energy, contrast, correlation, and entropy;

[0062] Multi-scale decomposition is performed on the preprocessed data in the standardized training dataset, multi-scale weighted permutation entropy is calculated, and entropy feature vector is constructed.

[0063] A channel attention mechanism is configured for the fault diagnosis model to adaptively weight and fuse image texture feature vectors, entropy feature vectors, and depth features of snowflake-shaped SDP images extracted by deep convolutional neural networks to generate fused feature vectors for diagnosis.

[0064] As a preferred embodiment of the present invention, after outputting the diagnostic results, topology adaptive adjustment is performed according to the changes in the power grid topology, specifically as follows:

[0065] Real-time acquisition of the switching status of circuit breakers and disconnectors in the railway power distribution network, and construction of topology state vector;

[0066] The fault diagnosis model is adaptively adjusted based on the topology state vector.

[0067] Topology adaptive adjustment includes at least one of the following methods:

[0068] One approach involves using the topological state vector as an auxiliary feature, concatenating it with either the fusion feature vector or the deep feature, and then inputting the result into the fully connected layer of the fault diagnosis model.

[0069] Method 2 involves using a transfer learning framework to fine-tune the fault diagnosis model based on the differences in data distribution under different topological state vectors, and updating the model parameters to adapt to the current operating topology.

[0070] An automatic fault diagnosis and location system for railway power distribution networks, used for the above-mentioned methods, includes:

[0071] The sample construction module is used to collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. It uses auxiliary classification generative adversarial network to generate virtual fault samples with fault category labels, mixes them with the original sample set, constructs a class-balanced standardized training dataset, and performs preprocessing.

[0072] The image transformation module is used to map the preprocessed data in the standardized training dataset into a snowflake-shaped SDP image in a two-dimensional polar coordinate system using the symmetric point pattern transformation technique.

[0073] The intelligent diagnostic module is used to input snowflake-shaped SDP images into a fault diagnosis model based on a deep convolutional neural network, and output the fault section and preliminary fault distance.

[0074] The physical inversion module is used to construct a physical circuit model of the distributed parameters of the railway power distribution network. It determines the corresponding line impedance parameters and topology path based on the fault section, and substitutes the preliminary fault distance into the physical circuit model in reverse. It then uses the transmission line equation to deduce the theoretical simulation voltage and simulation current.

[0075] The closed-loop verification module is used to calculate the Pearson correlation coefficient between the simulated voltage and the transient zero-sequence voltage waveform data, and between the simulated current and the transient zero-sequence current waveform data. If the Pearson correlation coefficient is lower than the preset confidence threshold, the iterative correction program is started to correct the preliminary fault distance until the error requirement is met and the diagnostic result is output.

[0076] The advantages of this invention are:

[0077] This invention expands and enhances transient zero-sequence voltage and current waveform data of railway power distribution networks by introducing an auxiliary classification generative adversarial network. It can construct a standardized training dataset with class balance, solving the problem of insufficient training of deep learning models due to the scarcity of fault samples in actual operation scenarios. At the same time, it uses the symmetric point mode transformation technology to map one-dimensional time-domain signals into two-dimensional snowflake-shaped SDP images, and combines multi-scale weighted permutation entropy, image texture feature vectors extracted from gray-level co-occurrence matrix, and channel attention mechanism to perform adaptive weighted fusion of multi-modal features. This effectively enhances the expressive ability of nonlinear transient features generated by intermittent arc and high-resistance grounding faults, significantly suppresses the interference of background harmonic noise on fault feature extraction, and improves the model's identification accuracy for complex fault types.

[0078] This invention constructs a physical circuit model of distributed parameters in railway power distribution networks. It uses transmission line equations to reverse-engineer the preliminary fault distance output by a deep convolutional neural network. By calculating the Pearson correlation coefficients of simulated voltage, simulated current, and transient zero-sequence voltage and current waveforms, closed-loop verification and iterative correction are implemented. This achieves deep integration of data-driven algorithms and physical mechanism models, overcoming the shortcomings of single artificial intelligence models, such as lack of physical interpretability, unreliable output results, and susceptibility to false alarms. It ensures that the final fault diagnosis and location results conform to the physical operation laws of power systems and meet the high reliability requirements of railway power supply systems for fault handling.

[0079] This invention constructs a topology state vector by acquiring the status of circuit breakers and disconnectors in real time, and adopts an adaptive adjustment strategy of feature splicing and transfer learning fine-tuning. This enables the fault diagnosis model to dynamically perceive and adapt to the ever-changing operating network structure of the railway distribution network, solving the problem of the failure of the original fixed parameter model caused by changes in the operating topology of the distribution network. It ensures that the system can maintain stable fault diagnosis and location performance under dual power supply, switching operations and different operating modes. Attached Figure Description

[0080] Figure 1 This is a schematic flowchart of the method of the present invention;

[0081] Figure 2 This is a schematic diagram of the modules of the system of the present invention;

[0082] Figure 3 This is a schematic diagram of the sample distribution for constructing a class-balanced standardized training dataset using the auxiliary classification generative adversarial network in Embodiment 3 of the present invention;

[0083] Figure 4 This is a time-domain waveform diagram of the transient zero-sequence current recording data collected in Embodiment 3 of the present invention;

[0084] Figure 5 This is a snowflake-shaped SDP image generated based on the symmetric point mode transformation technique in Embodiment 3 of the present invention;

[0085] Figure 6 This is a comparison diagram of closed-loop verification between the physical inversion simulation current and the transient zero-sequence current recording data in Embodiment 3 of the present invention;

[0086] Figure 7 This is a graph showing the change of Pearson correlation coefficient with the number of iterations during the iterative correction process in Embodiment 3 of the present invention.

[0087] Figure 8 This is a graph showing the change in fault diagnosis accuracy during the topology adaptive adjustment process in Embodiment 3 of the present invention. Detailed Implementation

[0088] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0089] Example 1: As Figure 1 As shown, the automatic diagnosis and location method for railway power distribution network fault information includes the following steps:

[0090] S1. Collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. Use auxiliary classification generative adversarial network to generate virtual fault samples with fault category labels. Mix them with the original sample set to obtain a class-balanced standardized training dataset. Then, preprocess the standardized training dataset.

[0091] S2. Using the symmetric point pattern transformation technique, the preprocessed data in the standardized training dataset is mapped into a snowflake-shaped SDP image in a two-dimensional polar coordinate system.

[0092] S3. Input the snowflake-shaped SDP image into the fault diagnosis model based on a deep convolutional neural network, and output the fault section and the preliminary fault distance.

[0093] S4. Construct a physical circuit model of the railway power distribution network distribution parameters, determine the corresponding line impedance parameters and topology path according to the fault section, and substitute the preliminary fault distance into the physical circuit model in reverse, and use the transmission line equation to deduce the theoretical simulation voltage and simulation current.

[0094] S5. Calculate the Pearson correlation coefficients of the simulated voltage and transient zero-sequence voltage waveform data, and the simulated current and transient zero-sequence current waveform data, respectively. If the Pearson correlation coefficient is lower than the preset confidence threshold, start the iterative correction program to correct the preliminary fault distance until the error requirement is met and then output the diagnostic results.

[0095] In S1, a class-balanced standardized training dataset is obtained, and the standardized training dataset is preprocessed as follows:

[0096] Fault recording and monitoring devices are deployed at the busbar outgoing terminals of substations and on the high-voltage side of box-type substations along the railway power distribution network.

[0097] The analog input terminal of the fault recording monitoring device is connected to the secondary side of the voltage transformer and the secondary side of the current transformer in the power distribution network line via a cable.

[0098] Voltage transformers acquire the transient zero-sequence voltage analog signal of the line, and current transformers acquire the transient zero-sequence current analog signal of the line.

[0099] The analog-to-digital converter inside the fault recording and monitoring device converts the transient zero-sequence voltage analog signal and the transient zero-sequence current analog signal into a digital sequence;

[0100] When the transient zero-sequence voltage amplitude or transient zero-sequence current amplitude exceeds the corresponding preset voltage start threshold or current start threshold, the fault recording monitoring device extracts the digital sequence from one cycle before the start time to five cycles after the start time to form transient zero-sequence voltage recording data and transient zero-sequence current recording data.

[0101] The transient zero-sequence voltage and transient zero-sequence current waveforms collected at historical fault moments are compiled into a raw sample set.

[0102] Construct an auxiliary classification generative adversarial network that includes a generator model and a discriminator model;

[0103] The generator model receives a random noise vector that follows a standard normal distribution generated by a random number generation algorithm and a fault category label vector selected from a preset fault category set. It then uses a deconvolutional neural network layer to upsample and feature map the random noise vector and the fault category label vector, mapping the low-dimensional noise distribution to a high-dimensional data space with the same dimension as the transient zero-sequence voltage waveform data and the transient zero-sequence current waveform data, and outputs a virtual fault sample with fault category labels.

[0104] The discriminator model receives input samples, which are either real fault samples from the original sample set or virtual fault samples generated by the generator model. The discriminator model uses a convolutional neural network layer to extract features from the input samples and outputs the probability of the input sample being a real or fake fault sample and the probability of the input sample belonging to a specific fault category.

[0105] During the training of the auxiliary classification generative adversarial network, the generator model parameters and the discriminator model parameters are updated by optimizing a composite objective function that includes data source discrimination loss and fault category classification loss.

[0106] The data source discrimination loss is constructed by summing the expected value of the log-likelihood probability that the discriminator model correctly identifies a real fault sample as real and the log-likelihood probability that the discriminator model correctly identifies a virtual fault sample as false. It is used to measure the discriminator model's ability to distinguish between real and virtual fault samples.

[0107] The fault category classification loss is constructed based on the mathematical expectation of the log-likelihood probability of the discriminator model correctly classifying the input sample as the corresponding fault category label. It is used to measure the ability of the discriminator model to correctly classify the sample into the corresponding fault category.

[0108] The discriminator model updates its parameters by maximizing the sum of the data source discrimination loss and the fault category classification loss to improve its ability to distinguish between real and fake samples and identify fault categories. The generator model updates its parameters by maximizing the discriminator's category classification loss for generated samples and minimizing the data source discrimination loss to generate realistic virtual fault samples with clear fault category characteristics, until the discriminator model can no longer distinguish between virtual fault samples and real fault samples.

[0109] After the auxiliary classification generative adversarial network is trained, the generator model is fed with fault category label vectors whose number of samples in the original sample set does not reach the preset threshold (the preset threshold for the number of samples in a single fault category can be set to 200~500). The generator model outputs virtual fault samples with the corresponding number of padded samples. The virtual fault samples are then merged with the real fault samples in the original sample set to ensure that the number of samples for each fault category is consistent after merging, thus forming a standardized training dataset with class balance.

[0110] Z-score standardization is performed on each transient zero-sequence voltage and current waveform data in the standardized training dataset. This is achieved by calculating the arithmetic mean and standard deviation of each transient zero-sequence voltage and current waveform data, and then subtracting the corresponding arithmetic mean from each original data value and dividing by the corresponding standard deviation. This yields preprocessed transient zero-sequence voltage and current waveform data that satisfy the distribution characteristics of a mean of 0 and a standard deviation of 1, thus eliminating the dimensional differences between the two datasets.

[0111] In S2, the preprocessed data in the standardized training dataset is mapped to a snowflake-shaped SDP image in a two-dimensional polar coordinate system. The process is as follows:

[0112] Define the mapping parameters required for the symmetric point mode transformation. These parameters include the time lag coefficient L and the initial deflection angle of the polar coordinate system. and angle mapping coefficient ;

[0113] Using the symmetric point mode transformation technique, the preprocessed transient zero-sequence voltage and current waveforms are treated as time series signals, and the polar radius of each data point (the i-th data point) in the time series signal is calculated in a two-dimensional polar coordinate system. and polar angle ;

[0114] The minimum-maximum normalization logic is used to map the data values ​​of the time series signal to a preset radius range. Calculate the polar radius for the i-th data point (N is the number of data points). In the formula, Let be the value of the i-th data point in the time series signal. , These represent the maximum and minimum values ​​in the time series signal, respectively.

[0115] The data values ​​at the time lag point are mapped to angular offsets, and the initial deflection angle is added. For time series signals Calculate the polar angle for the i-th data point. In the formula, Let be the value of the (i+L)th data point in the time series signal;

[0116] Depicting the polar radius in a two-dimensional polar coordinate system and polar angle The determined mapping points are plotted on the same two-dimensional polar coordinate plane, corresponding to the data points in the time series signal. Different initial deflection angles are set for transient zero-sequence voltage waveform data and transient zero-sequence current waveform data, so that the mapping points of different data sources are centrally symmetrically distributed in the two-dimensional polar coordinate system, forming a graphic with a symmetrical structure, and generating a snowflake-shaped SDP image that reflects the transient characteristics of the fault.

[0117] In S3, the fault section and preliminary fault distance are output. The process is as follows:

[0118] A fault diagnosis model based on a deep convolutional neural network is constructed. The structure of a deep convolutional neural network includes an input layer, a feature extraction layer, a fully connected layer, and a multi-task output layer.

[0119] The feature extraction layer consists of alternating convolutional and pooling layers. The convolutional layers use convolutional kernels to perform convolution operations on the input snowflake-shaped SDP image and combine them with non-linear activation functions to extract local spatial features of the image and generate feature maps. The pooling layers use max pooling logic to downsample the feature maps output by the convolution, retaining the maximum values ​​in local regions of the feature maps to reduce feature dimensionality and maintain the translation invariance of features.

[0120] The multi-task output layer contains two parallel branch structures: a fault segment classification branch and a fault distance regression branch. The fault segment classification branch receives the output of the fully connected layer, uses a normalized exponential function to calculate the probability distribution of the input sample belonging to each fault segment, and selects the fault segment with the highest probability as the output fault segment. The fault distance regression branch receives the output of the fully connected layer, uses a linear activation function to map the input features into continuous values, and uses this as the initial fault distance for the output.

[0121] The deep convolutional neural network is trained using snowflake-shaped SDP images from the standardized training dataset as model input and the corresponding real fault segment labels and real fault distance labels as supervision signals.

[0122] During training, a composite loss function is calculated, which includes classification loss and regression loss.

[0123] The classification loss is constructed using cross-entropy calculation logic. By calculating the negative logarithm of the predicted probability corresponding to the actual fault segment label, the difference between the fault diagnosis model's predicted fault segment distribution and the actual distribution is measured.

[0124] The regression loss is constructed using the mean squared error calculation logic. The accuracy of the regression prediction is measured by calculating the square of the difference between the preliminary fault distance output by the fault diagnosis model and the normalized true fault distance.

[0125] The composite loss function is constructed as a weighted sum of classification loss and regression loss, and the proportion of classification task and regression task in the total loss is adjusted by introducing a balancing hyperparameter.

[0126] The gradient of the composite loss function with respect to the weight and bias parameters of each layer in the deep convolutional neural network is calculated using the backpropagation algorithm. The weight and bias parameters are then updated along the direction of gradient descent using the stochastic gradient descent optimization algorithm until the composite loss function converges, thus obtaining the trained fault diagnosis model.

[0127] During online diagnosis, the real-time generated snowflake-shaped SDP image is input into the trained fault diagnosis model. The deep convolutional neural network performs forward propagation calculation on the snowflake-shaped SDP image, extracts local spatial features of the image through the feature extraction layer, and outputs the classification results of the fault section and the regression value of the preliminary fault distance through the multi-task output layer.

[0128] In S4, the theoretical simulated voltage and simulated current are derived using the transmission line equations. The process is as follows:

[0129] Based on the physical structure of high-voltage through power lines and automatic block power lines in railway distribution networks, a physical circuit model of distributed parameters in railway distribution networks is constructed. The physical circuit model equates the railway distribution network lines to a continuous distributed parameter circuit system composed of unit length resistance, unit length inductance, unit length conductance, and unit length capacitance.

[0130] Using the fault sections output by the fault diagnosis model, the physical attribute information corresponding to the fault sections is indexed in the geographic information database and line parameter database of the railway power distribution network.

[0131] The physical attribute information includes the conductor type, laying method, and topological connection relationship of the fault section relative to the outgoing line of the substation bus.

[0132] Based on physical property information, the zero-sequence unit resistance of the fault section is extracted. Zero-sequence unit inductance Zero-sequence unit conductance and zero-sequence unit capacitance And determine the length of the line topology path from the installation location of the fault recording monitoring device to the starting point of the fault section;

[0133] By combining the line topology path length and the preliminary fault distance output by the fault diagnosis model, the relative position of the fault point in the physical circuit model is determined. Using the fault point as the boundary condition, the transmission line equation is used to describe the spatiotemporal distribution of transient zero-sequence voltage and transient zero-sequence current in the physical circuit model.

[0134] In the physical circuit model, the propagation process of transient zero-sequence voltage and transient zero-sequence current along the railway distribution network line satisfies the following set of partial differential equations. These partial differential equations characterize the spatial rate of change of the transient zero-sequence voltage along the line as equal to the zero-sequence unit resistance of the line. The resulting resistive voltage drop and zero-sequence unit inductance The negative sum of the resulting inductive voltage drops, and the value used to characterize the spatial rate of change of the transient zero-sequence current along the line, which is equal to the zero-sequence unit conductance of the line. The resulting leakage current and zero-sequence unit capacitance The negative of the sum of the generated displacement currents:

[0135] ;

[0136] ;

[0137] In the formula, x represents the distance coordinate from the installation location of the fault recording monitoring device to the fault point along the railway power distribution network line, and t is the time variable. Let x be the transient zero-sequence voltage at time t at coordinate x. Let x be the transient zero-sequence current at time t at coordinate x. The zero-sequence unit resistance of the circuit is given. The zero-sequence unit inductance of the circuit. The zero-sequence unit conductance of the circuit. This is the zero-sequence unit capacitance of the circuit.

[0138] The partial differential equations are discretized and solved using the finite difference method. The initial fault distance output by the fault diagnosis model is set as the boundary position parameter of the partial differential equations. Combined with the power supply side boundary conditions and load side boundary conditions of the railway power distribution network system, the numerical sequence of the partial differential equations at the installation location of the fault recording and monitoring device is solved over time, and the theoretical simulated voltage and simulated current are obtained respectively.

[0139] The specific implementation process of S5 is as follows:

[0140] By calculating the Pearson correlation coefficient, the waveform similarity between simulated voltage and transient zero-sequence voltage recording data, and the waveform similarity between simulated current and transient zero-sequence current recording data are quantified respectively. The Pearson correlation coefficient includes the voltage Pearson correlation coefficient and the current Pearson correlation coefficient.

[0141] Calculate the voltage covariance between the simulated voltage and the transient zero-sequence voltage waveform data, and calculate the voltage standard deviation between the simulated voltage and the transient zero-sequence voltage waveform data respectively. Divide the voltage covariance by the product of the two voltage standard deviations to obtain the voltage Pearson correlation coefficient.

[0142] Calculate the current covariance between the simulated current and the transient zero-sequence current waveform data, and calculate the current standard deviation between the simulated current and the transient zero-sequence current waveform data respectively. Divide the current covariance by the product of the two current standard deviations to obtain the current Pearson correlation coefficient.

[0143] Set a confidence threshold, compare the voltage Pearson correlation coefficient and the current Pearson correlation coefficient with the confidence threshold. When both the voltage Pearson correlation coefficient and the current Pearson correlation coefficient are greater than or equal to the confidence threshold, the preliminary fault distance output by the fault diagnosis model is deemed reliable, and the preliminary fault distance is output as the final diagnosis result.

[0144] When either the voltage Pearson correlation coefficient or the current Pearson correlation coefficient is below the confidence threshold, it is determined that the preliminary fault distance output by the fault diagnosis model has a deviation, and an iterative correction procedure is initiated. The iterative correction procedure includes:

[0145] Using the double-ended traveling wave method, the propagation speed of the traveling wave signal in the railway power distribution network and the time difference between the arrival of the traveling wave front at both ends of the line are calculated to estimate the theoretical location range of the fault point. Combined with the preliminary fault distance output by the fault diagnosis model, a search interval containing the actual fault location is constructed with the preliminary fault distance as the center and the error margin determined by the theoretical location range.

[0146] Within the search interval, a fine-tuning step size is set, and different corrected fault distances are selected by discretization within the search interval according to the fine-tuning step size. Each corrected fault distance is substituted into the physical circuit model, and the corresponding simulation voltage and simulation current are recalculated using the transmission line equation. The corresponding voltage Pearson correlation coefficient and current Pearson correlation coefficient are recalculated, and the weighted sum of the voltage Pearson correlation coefficient and current Pearson correlation coefficient is calculated as a comprehensive similarity index.

[0147] Iterate through the comprehensive similarity index corresponding to each corrected fault distance within the search interval, select the corrected fault distance corresponding to the maximum value of the comprehensive similarity index, and output it as the corrected preliminary fault distance.

[0148] After mapping the preprocessed data in the standardized training dataset to a snowflake-shaped SDP image in a two-dimensional polar coordinate system, the process also includes feature extraction and fusion of the snowflake-shaped SDP image, specifically:

[0149] Calculate the gray-level co-occurrence matrix of the snowflake-shaped SDP image and extract the image texture feature vector including energy, contrast, correlation, and entropy;

[0150] Multi-scale decomposition is performed on the preprocessed data in the standardized training dataset, multi-scale weighted permutation entropy is calculated, and entropy feature vector is constructed.

[0151] A channel attention mechanism is configured for the fault diagnosis model to adaptively weight and fuse image texture feature vectors, entropy feature vectors, and depth features of snowflake-shaped SDP images extracted by deep convolutional neural networks to generate fused feature vectors for diagnosis.

[0152] By adding a feature extraction and fusion step after the snowflake-shaped SDP image mapping stage, key fault-sensitive features can be accurately extracted and aggregated from high-dimensional image data, effectively suppressing noise interference. At the same time, through deep fusion of multi-dimensional features, the limitations of the single image feature representation capability are made up for, and the recognition accuracy and generalization ability of the subsequent deep neural network for complex fault modes are significantly improved.

[0153] Based on the obtained fused feature vector, after outputting the diagnostic results, topology adaptive adjustment is performed according to the changes in the power grid topology, specifically as follows:

[0154] Real-time acquisition of the switching status of circuit breakers and disconnectors in the railway power distribution network, and construction of topology state vector;

[0155] The fault diagnosis model is adaptively adjusted based on the topology state vector.

[0156] Topology adaptive adjustment includes at least one of the following methods:

[0157] One approach involves using the topological state vector as an auxiliary feature, concatenating it with either the fusion feature vector or the deep feature, and then inputting the result into the fully connected layer of the fault diagnosis model.

[0158] Method 2 involves using a transfer learning framework to fine-tune the fault diagnosis model based on the differences in data distribution under different topological state vectors, and updating the model parameters to adapt to the current operating topology.

[0159] After outputting the final diagnostic results, a topology adaptive adjustment step is introduced. This step can automatically correct the diagnostic model parameters or adapt to the new topology environment based on the dynamic changes in the topology structure during the actual operation of the railway power distribution network. This effectively solves the problem of misdiagnosis caused by the fixed model's inability to adapt to topology switching. As a closed-loop optimization link after the diagnostic results are output, this step ensures the long-term effectiveness and robustness of the diagnostic positioning results under various operating conditions.

[0160] Example 2: Figure 2 As shown, the automatic diagnosis and location system for railway power distribution network fault information is used to implement the method in Example 1. It includes a sample construction module, an image transformation module, an intelligent diagnosis module, a physical inversion module, and a closed-loop verification module, with electrical signal connections between the modules.

[0161] The sample construction module is used to collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. Using auxiliary classification generative adversarial network, virtual fault samples with fault category labels are generated, mixed with the original sample set, and a class-balanced standardized training dataset is constructed and preprocessed.

[0162] The image transformation module is used to map the preprocessed data in the standardized training dataset into a snowflake-shaped SDP image in a two-dimensional polar coordinate system using the symmetric point pattern transformation technique.

[0163] The intelligent diagnostic module is used to input snowflake-shaped SDP images into a fault diagnosis model based on a deep convolutional neural network, and output the fault section and preliminary fault distance.

[0164] The physical inversion module is used to construct a physical circuit model of the distributed parameters of the railway power distribution network. It determines the corresponding line impedance parameters and topology path based on the fault section, and substitutes the preliminary fault distance into the physical circuit model in reverse. It then uses the transmission line equation to deduce the theoretical simulation voltage and simulation current.

[0165] The closed-loop verification module is used to calculate the Pearson correlation coefficient between the simulated voltage and the transient zero-sequence voltage waveform data, and between the simulated current and the transient zero-sequence current waveform data. If the Pearson correlation coefficient is lower than the preset confidence threshold, the iterative correction program is started to correct the preliminary fault distance until the error requirement is met and the diagnostic result is output.

[0166] The implementation details of each module are the same as in Example 1.

[0167] Example 3: This example is a verification example based on the method of Example 1. Through simulation experiments, it verifies the ability of the method of Example 1 to achieve high-precision fault diagnosis and adaptive location under high-resistance grounding faults and changes in the operating topology of the distribution network.

[0168] Based on the parameters of a railway power distribution network, a simulation model was developed using a high-resistance grounding fault (grounding resistance 1000 ohms) superimposed with 30dB Gaussian white noise as the test scenario. The simulation results are shown in the figure below. Figures 3-8 As shown.

[0169] Figure 3 The solid square markers represent real fault samples in the original sample set, while the solid circle markers represent virtual fault samples generated by the auxiliary classification generative adversarial network. It can be seen that the original sample set is sparsely distributed and has gaps in the feature space, while a large number of virtual fault samples emerge in batches and evenly fill the blank areas around the real fault samples, thus constructing a standardized training dataset with class balance, which effectively solves the problem of insufficient model training caused by the scarcity of fault samples.

[0170] Figure 4 The acquired transient zero-sequence current waveform data is subject to interference from high-impedance grounding and noise, resulting in chaotic waveforms and weak characteristics. Figure 5 The snowflake-shaped SDP image in the two-dimensional polar coordinate system generated by the symmetric point mode transformation technique is presented as a snowflake shape with three symmetrical petals. The comparison shows that the transformation can transform weak signals that are difficult to identify in the one-dimensional time domain into snowflake-shaped SDP images with clear texture and regular structure in the two-dimensional polar coordinate system, which significantly enhances the identification of nonlinear fault features.

[0171] Figure 6 The comparison between the simulated current obtained from the physical inversion calculation and the transient zero-sequence current waveform data is shown. The solid line represents the transient zero-sequence current waveform data, and the dashed line represents the simulated current derived from the physical circuit model with distributed parameters using the transmission line equation. It can be seen that when the Pearson correlation coefficient exceeds the preset confidence threshold, the dashed line and the solid line almost perfectly coincide, which verifies the physical consistency of the positioning results. Figure 7 The curves showing the changes in the comprehensive similarity index during the iterative correction process are presented. The curve marked with a diamond continuously rises with the increase of the number of iterations and eventually breaks through the preset confidence threshold, indicating that by using the double-ended traveling wave method to construct the search interval and performing physical inversion calculations for fine-tuning, the deviation of the initial fault distance was successfully eliminated.

[0172] Figure 8 The diagram shows the topology adaptive adjustment stage, with the curve representing the change in the diagnostic accuracy of the fault diagnosis model over time. Figure 8 As shown, when the distribution network operating topology undergoes a sudden change, the diagnostic accuracy drops significantly. Subsequently, by constructing a topology state vector and initiating a transfer learning framework, the fault diagnosis model is fine-tuned, which enables the diagnostic accuracy to quickly recover and stabilize at a high level, verifying the adaptive capability of this invention in response to changes in operating mode.

[0173] This verification example demonstrates that by combining the data generation capability of auxiliary classification generative adversarial networks with the feature extraction capability of symmetric point pattern transformation technology, this invention effectively solves the problem of difficult fault feature extraction in small sample and high noise environments. Through a closed-loop verification mechanism driven by both physical and data, the physical interpretability and high reliability of the fault location results are guaranteed.

[0174] In summary, the automatic diagnosis and location method and system for railway distribution network fault information proposed in this invention demonstrates excellent automatic diagnosis and accurate location capabilities under complex operating environments and weak fault characteristics. At the same time, through topology adaptive adjustment, an intelligent diagnostic architecture that dynamically adapts to changes in the distribution network structure is constructed, which has extremely high engineering application value and promotion prospects.

[0175] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatic diagnosis and location of fault information in railway power distribution networks, characterized in that, Includes the following steps: S1. Collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. Use auxiliary classification generative adversarial network to generate virtual fault samples with fault category labels. Mix them with the original sample set to obtain a class-balanced standardized training dataset. Then, preprocess the standardized training dataset. S2. Using the symmetric point pattern transformation technique, the preprocessed data in the standardized training dataset is mapped into a snowflake-shaped SDP image in a two-dimensional polar coordinate system. S3. Input the snowflake-shaped SDP image into the fault diagnosis model based on a deep convolutional neural network, and output the fault section and the preliminary fault distance. S4. Construct a physical circuit model of the railway power distribution network distribution parameters, determine the corresponding line impedance parameters and topology path according to the fault section, and substitute the preliminary fault distance into the physical circuit model in reverse, and use the transmission line equation to deduce the theoretical simulation voltage and simulation current. S5. Calculate the Pearson correlation coefficients between the simulated voltage and the transient zero-sequence voltage waveform data, and between the simulated current and the transient zero-sequence current waveform data, respectively. If the Pearson correlation coefficient is lower than the preset confidence threshold, start the iterative correction program to correct the preliminary fault distance until the error requirement is met, and then output the diagnostic results. The process includes: Set a confidence threshold, compare the voltage Pearson correlation coefficient and the current Pearson correlation coefficient with the confidence threshold. When both the voltage Pearson correlation coefficient and the current Pearson correlation coefficient are greater than or equal to the confidence threshold, the preliminary fault distance output by the fault diagnosis model is deemed reliable, and the preliminary fault distance is output as the final diagnosis result. When either the voltage Pearson correlation coefficient or the current Pearson correlation coefficient is below the confidence threshold, it is determined that the preliminary fault distance output by the fault diagnosis model has a deviation, and an iterative correction procedure is initiated. The iterative correction procedure includes: Using the double-ended traveling wave method, the propagation speed of the traveling wave signal in the railway power distribution network and the time difference between the arrival of the traveling wave front at both ends of the line are calculated to estimate the theoretical location range of the fault point. Combined with the preliminary fault distance output by the fault diagnosis model, and with the preliminary fault distance as the center, and combined with the error margin determined by the theoretical location range, a search interval containing the actual location of the fault is constructed. Within the search interval, a fine-tuning step size is set, and different corrected fault distances are selected by discretization within the search interval according to the fine-tuning step size. Each corrected fault distance is substituted into the physical circuit model, and the corresponding simulation voltage and simulation current are recalculated using the transmission line equation. The corresponding voltage Pearson correlation coefficient and current Pearson correlation coefficient are recalculated, and the weighted sum of the voltage Pearson correlation coefficient and current Pearson correlation coefficient is calculated as a comprehensive similarity index. Iterate through the comprehensive similarity index corresponding to each corrected fault distance within the search interval, select the corrected fault distance corresponding to the maximum value of the comprehensive similarity index, and output it as the corrected preliminary fault distance.

2. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, In S1, the process of obtaining a class-balanced standardized training dataset and preprocessing the standardized training dataset includes: Fault recording and monitoring devices are deployed at the busbar outgoing terminals of substations and on the high-voltage side of box-type substations along the railway power distribution network. The analog input terminal of the fault recording monitoring device is connected to the secondary side of the voltage transformer and the secondary side of the current transformer in the power distribution network line via a cable. Voltage transformers acquire the transient zero-sequence voltage analog signal of the line, and current transformers acquire the transient zero-sequence current analog signal of the line. The analog-to-digital converter inside the fault recording and monitoring device converts the transient zero-sequence voltage analog signal and the transient zero-sequence current analog signal into a digital sequence; When the transient zero-sequence voltage amplitude or transient zero-sequence current amplitude exceeds the corresponding preset voltage start threshold or current start threshold, the fault recording monitoring device extracts the digital sequence from one cycle before the start time to five cycles after the start time to form transient zero-sequence voltage recording data and transient zero-sequence current recording data. The transient zero-sequence voltage and transient zero-sequence current waveforms collected at historical fault moments are compiled as the original sample set. Construct an auxiliary classification generative adversarial network that includes a generator model and a discriminator model; The generator model receives a random noise vector that follows a standard normal distribution generated by a random number generation algorithm and a fault category label vector selected from a preset fault category set. It then uses a deconvolutional neural network layer to upsample and feature map the random noise vector and the fault category label vector, mapping the low-dimensional noise distribution to a high-dimensional data space with the same dimension as the transient zero-sequence voltage waveform data and the transient zero-sequence current waveform data, and outputs a virtual fault sample with fault category labels. The discriminator model receives input samples, which are either real fault samples from the original sample set or virtual fault samples generated by the generator model. The discriminator model uses a convolutional neural network layer to extract features from the input samples and outputs the probability of the input sample being a real or fake fault sample and the probability of the input sample belonging to a specific fault category. During the training of the auxiliary classification generative adversarial network, the generator model parameters and the discriminator model parameters are updated by optimizing a composite objective function that includes data source discrimination loss and fault category classification loss. After the auxiliary classification generative adversarial network is trained, the fault category label vectors of the original sample set that do not reach the preset threshold are input into the generator model. The generator model outputs virtual fault samples with the corresponding number of padded samples. The virtual fault samples are merged with the real fault samples in the original sample set so that the number of samples of each fault category is consistent after merging, forming a standardized training dataset with class balance. Z-score normalization is performed on each transient zero-sequence voltage and transient zero-sequence current waveform in the normalized training dataset.

3. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, In S2, the process of mapping the preprocessed data in the standardized training dataset to a snowflake-shaped SDP image in two-dimensional polar coordinates includes: Set the mapping parameters required for the symmetric point mode transformation, including the time lag coefficient, the initial deflection angle of the polar coordinate system, and the angle mapping coefficient; Using the symmetric point mode transformation technique, the preprocessed transient zero-sequence voltage waveform data and transient zero-sequence current waveform data are regarded as time series signals, and the polar radius and polar angle corresponding to each data point in the time series signal in the two-dimensional polar coordinate system are calculated. Using minimum-maximum normalization logic, the data values ​​of the time series signal are mapped to a preset radius range, and the polar radius is calculated for the data points in the time series signal. The data values ​​at the time lag points are mapped to angular offsets, and the initial deflection angle is added to calculate the polar angle for the data points in the time series signal. Mapping points determined by polar radius and polar angle are plotted in a two-dimensional polar coordinate system. The mapping points corresponding to the data points in the time series signal are plotted in the same two-dimensional polar coordinate plane. Different initial deflection angles are set for transient zero-sequence voltage waveform data and transient zero-sequence current waveform data, so that the mapping points of different data sources are centrally symmetrically distributed in the two-dimensional polar coordinate system, forming a graphic with a symmetrical structure, and generating a snowflake-shaped SDP image that reflects the transient characteristics of the fault.

4. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, In S3, the process of outputting the fault section and preliminary fault distance includes: A fault diagnosis model based on a deep convolutional neural network is constructed. The structure of a deep convolutional neural network includes an input layer, a feature extraction layer, a fully connected layer, and a multi-task output layer. The feature extraction layer consists of alternating convolutional layers and pooling layers. The convolutional layers use convolutional kernels to perform convolution operations on the input snowflake-shaped SDP image and combine them with non-linear activation functions to extract local spatial features of the image and generate feature maps. The pooling layers use max pooling logic to downsample the feature maps output by the convolution, retaining the maximum values ​​in local regions of the feature maps. The multi-task output layer contains two parallel branch structures: a fault segment classification branch and a fault distance regression branch. The fault segment classification branch receives the output of the fully connected layer, uses a normalized exponential function to calculate the probability distribution of the input sample belonging to each fault segment, and selects the fault segment with the highest probability as the output fault segment. The fault distance regression branch receives the output of the fully connected layer, uses a linear activation function to map the input features into continuous values, and uses this as the initial fault distance for the output. The deep convolutional neural network was trained using snowflake-shaped SDP images from the standardized training dataset as model input and the corresponding real fault segment labels and real fault distance labels as supervision signals. During training, a composite loss function is calculated, which includes classification loss and regression loss. The classification loss is constructed using cross-entropy calculation logic. By calculating the negative logarithm of the predicted probability corresponding to the actual fault segment label, the difference between the fault diagnosis model's predicted fault segment distribution and the actual distribution is measured. The regression loss is constructed using the mean squared error calculation logic. The accuracy of the regression prediction is measured by calculating the square of the difference between the preliminary fault distance output by the fault diagnosis model and the normalized true fault distance. The composite loss function is constructed as a weighted sum of classification loss and regression loss, and the proportion of classification task and regression task in the total loss is adjusted by introducing a balancing hyperparameter; The gradient of the composite loss function with respect to the weight and bias parameters of each layer in the deep convolutional neural network is calculated using the backpropagation algorithm. The weight and bias parameters are then updated along the direction of gradient descent using the stochastic gradient descent optimization algorithm until the composite loss function converges, thus obtaining the trained fault diagnosis model. During online diagnosis, the real-time generated snowflake-shaped SDP image is input into the trained fault diagnosis model. The deep convolutional neural network performs forward propagation calculation on the snowflake-shaped SDP image, extracts local spatial features of the image through the feature extraction layer, and outputs the classification results of the fault section and the regression value of the preliminary fault distance through the multi-task output layer.

5. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, In S4, the process of deriving the theoretical simulation voltage and simulation current using the transmission line equation includes: Based on the physical structure of high-voltage through power lines and automatic block power lines in railway distribution networks, a physical circuit model of distributed parameters in railway distribution networks is constructed. The physical circuit model equates the railway distribution network lines to a continuous distributed parameter circuit system composed of unit length resistance, unit length inductance, unit length conductance, and unit length capacitance. Using the fault sections output by the fault diagnosis model, the physical attribute information corresponding to the fault sections is indexed in the geographic information database and line parameter database of the railway power distribution network; The physical attribute information includes the conductor type, laying method, and topological connection relationship of the fault section relative to the outgoing line of the substation bus. Based on physical property information, the zero-sequence unit resistance, zero-sequence unit inductance, zero-sequence unit conductance and zero-sequence unit capacitance of the fault section are extracted, and the line topology path length from the installation location of the fault recording monitoring device to the starting point of the fault section is determined. By combining the line topology path length and the preliminary fault distance output by the fault diagnosis model, the relative position of the fault point in the physical circuit model is determined. Using the fault point as the boundary condition, the transmission line equation is used to describe the spatiotemporal distribution of transient zero-sequence voltage and transient zero-sequence current in the physical circuit model. In the physical circuit model, the propagation process of transient zero-sequence voltage and transient zero-sequence current along the railway distribution network line satisfies a set of partial differential equations. The set of partial differential equations is used to characterize the spatial rate of change of transient zero-sequence voltage along the line as equal to the negative sum of the resistive voltage drop generated by the zero-sequence unit resistance and the inductive voltage drop generated by the zero-sequence unit inductance, and to characterize the spatial rate of change of transient zero-sequence current along the line as equal to the negative sum of the leakage current generated by the zero-sequence unit conductance and the displacement current generated by the zero-sequence unit capacitance. The partial differential equations are discretized and solved using the finite difference method. The initial fault distance output by the fault diagnosis model is set as the boundary position parameter of the partial differential equations. Combined with the power supply side boundary conditions and load side boundary conditions of the railway power distribution network system, the numerical sequence of the partial differential equations at the installation location of the fault recording and monitoring device is solved over time, and the theoretical simulated voltage and simulated current are obtained respectively.

6. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, In S5, the process of calculating the Pearson correlation coefficients between the simulated voltage and the transient zero-sequence voltage waveform data, and between the simulated current and the transient zero-sequence current waveform data, includes: By calculating the Pearson correlation coefficient, the waveform similarity between simulated voltage and transient zero-sequence voltage recording data, and the waveform similarity between simulated current and transient zero-sequence current recording data are quantified respectively. The Pearson correlation coefficient includes the voltage Pearson correlation coefficient and the current Pearson correlation coefficient. Calculate the voltage covariance between the simulated voltage and the transient zero-sequence voltage waveform data, and calculate the voltage standard deviation between the simulated voltage and the transient zero-sequence voltage waveform data respectively. Divide the voltage covariance by the product of the two voltage standard deviations to obtain the voltage Pearson correlation coefficient. Calculate the current covariance between the simulated current and the transient zero-sequence current waveform data, and calculate the current standard deviation between the simulated current and the transient zero-sequence current waveform data respectively. Divide the current covariance by the product of the two current standard deviations to obtain the current Pearson correlation coefficient.

7. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 1, characterized in that, After mapping the preprocessed data in the standardized training dataset to a snowflake-shaped SDP image in a two-dimensional polar coordinate system, the process also includes feature extraction and fusion of the snowflake-shaped SDP image, specifically: Calculate the gray-level co-occurrence matrix of the snowflake-shaped SDP image and extract the image texture feature vector including energy, contrast, correlation, and entropy; Multi-scale decomposition is performed on the preprocessed data in the standardized training dataset, multi-scale weighted permutation entropy is calculated, and entropy feature vector is constructed. A channel attention mechanism is configured for the fault diagnosis model to adaptively weight and fuse image texture feature vectors, entropy feature vectors, and depth features of snowflake-shaped SDP images extracted by deep convolutional neural networks to generate fused feature vectors for diagnosis.

8. The method for automatic diagnosis and location of fault information in railway power distribution networks according to claim 7, characterized in that, After outputting the diagnostic results, topology adaptive adjustment is performed based on changes in the power grid topology, specifically as follows: Real-time acquisition of the switching status of circuit breakers and disconnectors in the railway power distribution network, and construction of topology state vector; The fault diagnosis model is adaptively adjusted based on the topology state vector. Topology adaptive adjustment includes at least one of the following methods: One approach involves using the topological state vector as an auxiliary feature, concatenating it with either the fusion feature vector or the deep feature, and then inputting the result into the fully connected layer of the fault diagnosis model. Method 2 involves using a transfer learning framework to fine-tune the fault diagnosis model based on the differences in data distribution under different topological state vectors, and updating the model parameters to adapt to the current operating topology.

9. An automatic diagnosis and location system for railway power distribution network fault information, used to implement the automatic diagnosis and location method for railway power distribution network fault information as described in any one of claims 1-8, characterized in that, include: The sample construction module is used to collect transient zero-sequence voltage waveform data and transient zero-sequence current waveform data of railway power distribution network as the original sample set. It uses auxiliary classification generative adversarial network to generate virtual fault samples with fault category labels, mixes them with the original sample set, constructs a class-balanced standardized training dataset, and performs preprocessing. The image transformation module is used to map the preprocessed data in the standardized training dataset into a snowflake-shaped SDP image in a two-dimensional polar coordinate system using the symmetric point pattern transformation technique. The intelligent diagnostic module is used to input snowflake-shaped SDP images into a fault diagnosis model based on a deep convolutional neural network, and output the fault section and preliminary fault distance. The physical inversion module is used to construct a physical circuit model of the distributed parameters of the railway power distribution network. It determines the corresponding line impedance parameters and topology path based on the fault section, and substitutes the preliminary fault distance into the physical circuit model in reverse. It then uses the transmission line equation to deduce the theoretical simulation voltage and simulation current. The closed-loop verification module is used to calculate the Pearson correlation coefficient between the simulated voltage and the transient zero-sequence voltage waveform data, and between the simulated current and the transient zero-sequence current waveform data. If the Pearson correlation coefficient is lower than the preset confidence threshold, the iterative correction program is started to correct the preliminary fault distance until the error requirement is met and the diagnostic result is output.

Images