A Method and System for Fault Location of High and Low Voltage Switchgear Feeders Based on Transient Traveling Waves

Abstract

This invention discloses a method and system for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves, belonging to the field of power system fault detection and location technology. It includes: synchronously acquiring electrical and acoustic multimodal signals to generate a weighted transient synchronous feature matrix; performing time-frequency transformation and feedback optimization on the matrix to output a multi-scale time-frequency feature set; analyzing the feature set using an ensemble learning network to output a hierarchical preliminary location result; iteratively calibrating to converge to the precise fault segment; and finally fusing multi-model calculations to output a comprehensive fault location report. This invention adopts a technical approach combining multimodal signal fusion and physical models, enabling phased convergence from fuzzy region segmentation to precise location, significantly improving the accuracy, speed, and anti-interference capability of switchgear feeder fault location.

Description

Technical Field

[0001] This invention relates to the field of power system fault detection and location technology, and in particular to a method and system for locating faults in high and low voltage switchgear feeders based on transient traveling waves. Background Technology

[0002] High- and low-voltage switchgear are key hub equipment connecting the power grid and users in a power system, and their feeder lines are an important component of the distribution network. When a feeder experiences a short circuit, grounding fault, or other fault, it generates a transient traveling wave signal containing rich fault information. This is an electromagnetic wave that propagates along the line at near the speed of light. Using traveling waves for fault location is an important electrical measurement technique. It determines the physical location of the fault by detecting the arrival time of the traveling wave signal at the monitoring point or analyzing its characteristics, which is of great significance for ensuring the safe and stable operation of the power grid and quickly restoring power supply.

[0003] In existing technologies, feeder fault location methods based on transient traveling waves mainly rely on the analysis of voltage or current traveling wave signals. Some methods measure the distance by installing monitoring devices at both ends or multiple points of the feeder and utilizing the time difference of the traveling wave reaching different devices. Other methods utilize single-ended measured traveling wave signals and locate the fault by analyzing the time interval between the initial wavefront and the secondary wavefront reflected back from the fault point. At the signal processing level, time-frequency analysis tools such as wavelet transform are typically used to identify the traveling wavefront and calculate the fault distance.

[0004] However, the aforementioned existing technologies have significant technical shortcomings in practical applications. Inside and near high- and low-voltage switchgear, the electromagnetic environment is complex, with numerous metal structure reflections and coupling interferences. This makes it extremely easy for single current or voltage traveling wave signals to be distorted, with wavefront characteristics being submerged or false wavefronts generated, resulting in low location accuracy. Furthermore, when facing non-ideal fault conditions such as high-resistance grounding and multi-point faults, the traveling wave characteristics are weak and atypical. Traditional methods based on fixed models and algorithms have poor adaptability, leading to unstable or even invalid location results, making it difficult to meet the high requirements of modern distribution networks for rapid and accurate fault location. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves. By employing a technical approach that combines multi-modal signal fusion with a physical model, it can achieve convergence from fuzzy region segmentation to precise location in stages, significantly improving the accuracy, speed, and anti-interference capability of switchgear feeder fault location.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for fault location of high- and low-voltage switchgear feeders based on transient traveling waves includes synchronously acquiring broadband voltage and current traveling wave signals of the faulty feeder, as well as acoustic emission and ultrasonic signals from inside the switchgear; constructing a signal cross-matrix; performing signal correlation analysis on the signal cross-matrix to adjust the mode weights; generating a weighted transient synchronization feature matrix; performing a two-dimensional time-frequency transformation on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram; performing real-time feature scoring on the transient spectrum time series diagram; adjusting the weighted transient synchronization feature matrix based on the real-time feature scores; and outputting a multi-scale time-frequency feature set; inputting the multi-scale time-frequency feature set into a preset ensemble learning network to obtain... The preliminary location results are obtained and divided into a reliable region and a region to be verified through confidence distribution analysis. For each segment of the region to be verified, waveform edge detection and data-driven detection are performed to calibrate the traveling wave front times. The calibration results are fed back to the ensemble learning network to update its internal parameters. The process is iterated until the location results converge to the fault segment, and precise wavefront time pairs are output. Based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints, multi-dimensional location results are calculated, and consistency checks are performed. The verification results are fed back to update the parameter configuration of the ensemble learning network, and a comprehensive fault location report is output.

[0008] Optionally, generating the weighted transient synchronization feature matrix includes: taking the broadband voltage and current traveling wave signal as a first signal group and the acoustic emission and ultrasonic signals as a second signal group; performing correlation analysis on the first signal group and the second signal group to calculate the canonical correlation coefficient; iteratively correcting the modal weights based on the canonical correlation coefficient and monitoring the changes in the cross-correlation of the signal cross matrix to generate an iterative weight adjustment sequence; and applying the iterative weight adjustment sequence to perform weighted fusion on the signal cross matrix to output the weighted transient synchronization feature matrix.

[0009] Optionally, the output multi-scale time-frequency feature set includes: adaptively adjusting the time-frequency resolution of the weighted transient synchronization feature matrix, optimizing the selection of transformation parameters through a dynamic window function, and performing two-dimensional time-frequency transformation processing to generate a transient spectrum time series diagram; calculating real-time feature scores by evaluating information entropy for the transient spectrum time series diagram, quantifying the spectrum distribution uniformity and time series consistency, and outputting a feature score vector; and based on the feature score vector, performing parameter feedback correction on the weighted transient synchronization feature matrix, and fusing the transient spectrum time series diagram to generate a multi-scale time-frequency feature set.

[0010] Optionally, the output feature scoring vector includes: performing multi-scale entropy spectrum analysis on the transient spectral time series graph, calculating the local information entropy distribution and global spectral complexity, filtering out noise interference, and generating a preliminary entropy spectrum quantization set; calculating the spectral distribution uniformity and temporal consistency of the preliminary entropy spectrum quantization set, and generating a real-time feature scoring sequence through weighted aggregation; performing vector normalization processing on the real-time feature scoring sequence, optimizing the scoring distribution by statistical confidence interval correction, and outputting the feature scoring vector.

[0011] Optionally, the method further includes: performing multi-level association mapping between the iterative weight adjustment sequence and the feature scoring vector to generate a weight scoring joint matrix; calculating a dynamic optimization factor based on the weight scoring joint matrix, and performing vector-level enhancement by analyzing signal modes and time-frequency features to output an enhanced weight vector set; injecting the enhanced weight vector set into the parameters of the ensemble learning network, optimizing the internal weights and bias terms through gradient guidance, and generating a fusion feature optimization map.

[0012] Optionally, dividing the preliminary localization result into a reliable region and a region to be verified includes: inputting the multi-scale time-frequency feature set into the ensemble learning network to calculate the preliminary localization result; analyzing the confidence probability distribution of the preliminary localization result, performing distribution entropy calculation to quantify the uncertainty level, and outputting a confidence distribution map; and based on the confidence distribution map, assigning dynamic verification priorities to each region and dividing the region into a reliable region and a region to be verified.

[0013] Optionally, the calculation to obtain the preliminary location result includes: acquiring historical fault recording data of high and low voltage switchgear and historical power grid physical simulation data to generate a historical fault dataset; using multi-scale time-frequency features as input and actual location as output, using the historical fault dataset to establish and train a neural network model to obtain an ensemble learning network; inputting the multi-scale time-frequency feature set into the ensemble learning network to output the preliminary location result.

[0014] Optionally, the output of precise wavefront time pairs includes: performing dual-path preliminary calibration for each segment of the region to be verified, executing traveling wave edge detection and data-driven detection in parallel, and cross-validating the calibration results of the two paths to output initial wavefront times; calculating intermediate localization results for the initial wavefront times and evaluating the confidence gradient of the ensemble learning network for the intermediate localization results; using the confidence gradient to update the internal parameters of the ensemble learning network; iteratively executing the preliminary localization and the confidence gradient evaluation until the drift of the localization result and the confidence gain of the ensemble learning network are both lower than a preset convergence threshold, and outputting the wavefront times of the localization results in the converged state as precise wavefront time pairs.

[0015] Optionally, the output comprehensive fault location report includes: performing multi-end ranging and energy distribution analysis based on the precise wavefront time pair, incorporating the fusion feature optimization map for propagation path constraint correction, and generating preliminary multi-dimensional location results; applying cross-consistency checks to the preliminary multi-dimensional location results, and feeding the check results back to the ensemble learning network to update parameter configuration, and outputting an optimized location result set; performing weighted fusion on the location results in the optimized location result set, calculating the fault location coordinates and posterior probability, and encapsulating the fault location coordinates and the posterior probability into the comprehensive fault location report for output.

[0016] Based on the same inventive concept, this invention also provides a high- and low-voltage switchgear feeder fault location system based on transient traveling waves. The system includes: an adaptive signal fusion module, used to synchronously acquire broadband voltage and current traveling wave signals of the faulty feeder, acoustic emission and ultrasonic signals from inside the switchgear, construct a signal cross-matrix, perform signal correlation analysis on the signal cross-matrix to adjust mode weights, and generate a weighted transient synchronization feature matrix; a time-frequency feature optimization module, used to perform a two-dimensional time-frequency transformation on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram, and perform real-time feature scoring on the transient spectrum time series diagram, adjusting the weighted transient synchronization feature matrix according to the real-time feature score, and outputting a multi-scale time-frequency feature set; and an intelligent pre-location and region division module, used to... A time-frequency feature set is input into a preset ensemble learning network to obtain preliminary positioning results. Through confidence distribution analysis, the preliminary positioning results are divided into a reliable region and a region to be verified. An iterative convergence positioning module performs waveform edge detection and data-driven detection on each segment of the region to be verified, calibrates the traveling wave front time, and feeds the calibration results back to the ensemble learning network to update its internal parameters. This process iterates until the positioning results converge to the fault segment, outputting precise wavefront time pairs. A multi-model fusion and verification module calculates multi-dimensional positioning results based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints. It performs result consistency checks, feeds back the verification results to update the parameter configuration of the ensemble learning network, and outputs a comprehensive fault location report.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. This invention effectively overcomes the limitations of single signal sources being susceptible to interference and lacking information dimensions in complex electromagnetic environments by dynamically and collaboratively fusing multi-source signals and adaptively adjusting modal weights based on the inherent correlation between signals. It improves the quality and signal-to-noise ratio of input features from the source, laying a solid and reliable data foundation for subsequent accurate positioning.

[0019] 2. This invention constructs a hierarchical processing framework from intelligent prediction to iterative precise localization. By intelligently dividing regions through confidence analysis, high-cost, fine-grained computing resources are focused on the most uncertain segments. Through closed-loop feedback iteration between model and physical calibration results, high-precision and high-efficiency localization of complex and difficult faults is achieved, improving the robustness and response speed of the method.

[0020] 3. This invention establishes a multi-level, end-to-end closed-loop optimization system encompassing signal processing, model parameters, physical computation, and model configuration. By feeding back the results of feature quality assessment and cross-consistency testing to update the core intelligent network, it possesses the ability to continuously learn and self-evolve, enabling it to continuously optimize its performance during use and maintain a high level of positioning accuracy over the long term.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the high- and low-voltage switchgear feeder fault location method based on transient traveling waves, according to an embodiment of the present invention.

[0024] Figure 2 This is a cross-correlation and distribution matrix diagram of an embodiment of the present invention.

[0025] Figure 3 This is a confidence distribution and regional division cloud and rain map according to an embodiment of the present invention.

[0026] Figure 4 This is a composite time series diagram of the closed-loop intervention and model adaptive optimization process in an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram of the structure of a high- and low-voltage switchgear feeder fault location system based on transient traveling waves according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 One embodiment of the present invention proposes a fault location method for high and low voltage switchgear feeders based on transient traveling waves. It adopts a technical path that combines multi-mode signal fusion with physical models, which can achieve convergence from fuzzy region division to precise location in stages, significantly improving the accuracy, speed and anti-interference capability of switchgear feeder fault location.

[0030] The method described in this embodiment specifically includes:

[0031] The broadband voltage and current traveling wave signals of the faulty feeder and the acoustic emission and ultrasonic signals inside the switch cabinet are collected synchronously to construct a signal cross matrix. The signal cross matrix is ​​then subjected to signal correlation analysis to adjust the mode weights and generate a weighted transient synchronization feature matrix.

[0032] A two-dimensional time-frequency transformation is performed on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram, and a real-time feature score is performed on the transient spectrum time series diagram. The weighted transient synchronization feature matrix is ​​adjusted according to the real-time feature score, and a multi-scale time-frequency feature set is output.

[0033] The multi-scale time-frequency feature set is input into a preset ensemble learning network to obtain preliminary localization results. The preliminary localization results are then divided into reliable regions and regions to be verified through confidence distribution analysis.

[0034] For each segment of the region to be verified, waveform edge detection and data-driven detection are performed, the traveling wave head time is calibrated, and the calibration result is fed back to the integrated learning network to update the internal parameters. The process is iterated until the positioning result converges to the fault segment, and the accurate wave head time pair is output.

[0035] Based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints, multi-dimensional positioning results are calculated, and consistency checks are performed. The verification results are then fed back to update the parameter configuration of the integrated learning network, and a comprehensive fault location report is output.

[0036] By adopting a technical approach that combines multimodal signal fusion with physical models, it is possible to achieve convergence from fuzzy region segmentation to precise location in stages, which significantly improves the accuracy, speed and anti-interference capability of switchgear feeder fault location.

[0037] Optionally, the generation of the weighted transient synchronization feature matrix includes:

[0038] The broadband voltage and current traveling wave signal is used as the first signal group, and the acoustic emission and ultrasonic signal are used as the second signal group. Correlation analysis is performed on the first signal group and the second signal group to calculate the canonical correlation coefficient.

[0039] Specifically, this step aims to quantify the intrinsic correlation strength between two heterogeneous signal groups. The synchronously acquired broadband voltage and current traveling wave signals are classified into the first signal group, and the acoustic emission and ultrasonic signals are classified into the second signal group. To maximize the correlation between the two sets of variables, canonical correlation analysis (CCA) is performed on the first and second signal groups. This analysis maximizes the correlation coefficient between the two signal groups by finding a linear combination of the two signal groups, i.e., the canonical variables; this maximized correlation coefficient is the first canonical correlation coefficient.

[0040] Based on the canonical correlation coefficient, the modal weights are iteratively corrected, and the cross-correlation changes of the signal cross matrix are monitored to generate an iterative weight adjustment sequence.

[0041] Specifically, the canonical correlation coefficient directly reflects the degree of consistency between the fault information represented by the electrical signal group and the acoustic signal group under the current weight allocation. An iterative optimization process is initiated, with maximizing the canonical correlation coefficient as the objective function. Gradient ascent or similar optimization algorithms are used to adjust the weights of each signal mode in each iteration. Simultaneously, changes in the correlation within the signal cross-matrix are monitored in real time until the canonical correlation coefficient reaches stability or convergence. The weight change trajectory during this iteration is recorded, forming an iterative weight adjustment sequence.

[0042] The iterative weight adjustment sequence is applied to the signal cross matrix to perform weighted fusion, and a weighted transient synchronization feature matrix is ​​output.

[0043] Specifically, the optimized weights that finally converge in the iterative weight adjustment sequence are applied to perform weighted fusion on the original signal cross matrix. This weighted fusion process can be represented by the following matrix operation formula:

[0044] ,

[0045] in, The weighted transient synchronization feature matrix represents the final output; It is a signal cross matrix composed of the original multimodal signals; It is a weight vector whose elements are the final optimized weights of each signal mode derived from the iterative weight adjustment sequence; This is an operator that transforms a weight vector into a diagonal matrix. This fusion operation is achieved by multiplying each column of the signal cross matrix by its corresponding final optimized weight, thereby highlighting components most relevant to the fault and jointly verified by multiple signal sources, while suppressing noise or artifacts present only in a single signal source, such as... Figure 2 As shown in the figure, the diagonal lines represent the probability density distribution of the three core signal modes: broadband traveling wave, acoustic emission, and ultrasonic signals; the off-diagonal scatter plot reveals the correlation between any two signal modes. This figure serves as an important basis for signal correlation analysis and mode weight adjustment before generating the weighted transient synchronization feature matrix.

[0046] Optionally, the output multi-scale time-frequency feature set includes:

[0047] The weighted transient synchronization feature matrix is ​​adaptively adjusted for time and frequency resolution. The transformation parameters are optimized by dynamic window function, and two-dimensional time and frequency transformation is performed to generate a transient spectrum time series diagram.

[0048] Specifically, this step aims to dynamically optimize the time-frequency transform parameters based on the instantaneous characteristics of the weighted transient synchronization feature matrix. For example, when performing a Short-Time Fourier Transform (STFT), the transform parameters are optimized using a dynamic window function. A shorter analysis window is used for the rapidly changing transient portion of the signal to achieve high time resolution, while a longer analysis window is used for the more gradual changes to achieve high frequency resolution. After determining the optimized transform parameters, a two-dimensional time-frequency transform is performed on the weighted transient synchronization feature matrix, converting it from a one-dimensional time-domain signal to a two-dimensional time-frequency domain representation, thereby generating a transient spectral time series diagram.

[0049] Information entropy is evaluated and real-time feature scores are calculated for the transient spectrum time series diagram. The spectrum distribution uniformity and temporal consistency are quantified, and a feature score vector is output.

[0050] Specifically, to quantify the quality and information content of transient spectral time series plots, information entropy is evaluated to calculate real-time feature scores. This step aims to assess the clarity and reliability of time-frequency features. A real-time feature score for a local region of a transient spectral time series plot is provided. It can be calculated using a composite formula that includes multiple metrics:

[0051] ,

[0052] in, It is the information entropy calculated and normalized from the energy distribution of the region. A region with concentrated energy and clear characteristics corresponds to a lower information entropy, thus making... The value is relatively high; It is an index of the uniformity of spectral distribution obtained by calculating the Gini coefficient of spectral energy. The more concentrated the energy, the higher the value of this index. It is a time series consistency index obtained by calculating the autocorrelation coefficient of feature vectors within adjacent time windows. The more stable the features, the higher the index value. , , These are weight coefficients derived from a feature evaluation rule base, used to balance the contributions of information entropy, uniformity, and consistency to the overall score. This calculation ultimately outputs a feature score vector, where each element corresponds to the feature quality score for a different region or time step on the transient spectral time series plot.

[0053] Based on the feature scoring vector, the weighted transient synchronization feature matrix is ​​corrected by parameter feedback, and the transient spectrum time series diagram is fused to generate a multi-scale time-frequency feature set.

[0054] Specifically, this step uses the feature scoring vector generated in the previous step to perform parameter feedback correction on the weighted transient synchronization feature matrix. For example, for time periods with lower scores, the corresponding data points in the original weighted transient synchronization feature matrix are suppressed or filtered. Finally, the weighted transient synchronization feature matrix after parameter feedback correction is fused with the transient spectrum time series plot that clearly shows the time-frequency dynamics, forming a multi-scale time-frequency feature set with richer information dimensions and more reliable features, which is then output to the subsequent processing module.

[0055] Optionally, the output feature scoring vector includes:

[0056] Multi-scale entropy spectrum analysis is performed on the transient spectrum time series diagram to calculate the local information entropy distribution and global spectrum complexity, and noise interference is filtered out to generate a preliminary entropy spectrum quantization set.

[0057] Specifically, this step aims to comprehensively evaluate the uncertainty of the signal at different time resolutions. Multi-scale entropy spectrum analysis is performed on the transient spectral time series graph, for example using a sample entropy algorithm, to calculate its information entropy at different time scales, i.e., after coarsening the signal to different degrees. Through this analysis, the local information entropy distribution of the transient spectral time series graph can be calculated, i.e., the degree of disorder in each small region of the graph, and the global spectral complexity reflecting the overall signal complexity. High-entropy regions typically correspond to irregular noise, while low-entropy regions correspond to structured fault characteristics. Next, based on the results of the entropy spectrum analysis, an entropy threshold is set, and regions with entropy values ​​higher than this threshold are identified as noise and suppressed, thereby filtering and eliminating noise interference and generating a preliminary entropy spectrum quantization set.

[0058] For the preliminary entropy spectrum quantization set, the spectral distribution uniformity and temporal consistency are calculated, and a real-time feature scoring sequence is generated through weighted aggregation;

[0059] Specifically, after obtaining the initial entropy spectrum quantization set cleaned by the entropy dimension, its spectral distribution uniformity and temporal consistency are further calculated. Spectral distribution uniformity measures the distribution pattern of energy in the frequency domain and can be quantified by calculating the Gini coefficient or spectral flatness. Temporal consistency is measured by evaluating the autocorrelation or stability of features in adjacent time slices. Subsequently, the three dimensions of entropy, uniformity, and consistency are fused, with weighting coefficients derived from a feature evaluation rule base, thereby generating a real-time feature scoring sequence. Each value in this sequence represents the comprehensive feature quality of the transient spectral time series graph at the corresponding time point.

[0060] Based on the real-time feature scoring sequence, vector normalization is performed, statistical confidence interval correction is used to optimize the scoring distribution, and feature scoring vectors are output.

[0061] Specifically, to facilitate subsequent processing, the real-time feature score sequence is first subjected to vector normalization, for example, by using min-max normalization to scale its numerical range to between 0 and 1. Finally, to enhance the robustness of the scores, robust statistical methods, such as calculating the interquartile range, are used to construct confidence intervals, and these intervals are used to correct extreme score values ​​that deviate from the main distribution. Through this process of optimizing the score distribution, a stable feature score vector that accurately reflects the quality of the features is ultimately output.

[0062] Optionally, the method further includes:

[0063] The iterative weight adjustment sequence is correlated with the feature score vector through a multi-level association mapping to generate a weight score joint matrix.

[0064] Specifically, this step aims to establish a deep connection between front-end signal processing and mid-stage feature evaluation. The iterative weight adjustment sequence reflects the evolution of the importance of different signal modes during the fusion process, while the feature score vector quantifies the quality of the extracted time-frequency features at different times. A two-dimensional weight-score joint matrix is ​​constructed, with rows corresponding to signal modes and columns corresponding to time or feature dimensions. The matrix elements are calculated using a nonlinear function from the weight values ​​in the iterative weight adjustment sequence and the score values ​​in the feature score vector. This matrix reveals which signal modes contribute the most at which high-quality feature moments.

[0065] The dynamic optimization factor is calculated based on the weight scoring joint matrix, and the signal mode and time-frequency characteristics are analyzed to perform vector-level enhancement, outputting an enhanced weight vector set;

[0066] Specifically, this step aims to extract key optimization guidance information from the joint weight-score matrix. First, by applying Principal Component Analysis (PCA) or similar dimensionality reduction techniques to the joint weight-score matrix, its main trends are extracted, and a dynamic optimization factor is calculated. Then, this dynamic optimization factor is used to perform vector-level enhancement on the original iterative weight adjustment sequence. An enhanced weight vector is then generated. It can be calculated using the following formula:

[0067] ,

[0068] in, It is the original weight vector derived from the iterative weight adjustment sequence; This represents the element-wise product. It is a hyperbolic tangent function used for nonlinear mapping, which smoothly limits the enhancement effect to the range of -1 to 1; It is a hyperparameter used to control the magnitude of enhancement; It is the dynamic optimization factor obtained from the aforementioned calculation; It is the original weight vector and the feature score derived from the feature score vector. The mutual information between the two is used to quantify the nonlinear correlation between them. This formula achieves intelligent reinforcement of those weight components that are both important and reliable by fusing dynamic optimization factors and mutual information, and finally outputs an enhanced weight vector set.

[0069] The enhanced weight vector set is injected into the parameters of the ensemble learning network, and the internal weights and bias terms are optimized through gradient guidance to generate a fusion feature optimization map.

[0070] Specifically, this step applies the enhanced weight vector set generated in the previous step to optimize the internal parameters of the ensemble learning network. This process is not a simple replacement of the original parameters, but rather performed through gradient guidance. For example, the enhanced weight vector set is used as an attention mask, and element-wise multiplication is performed on the feature maps of specific layers of the ensemble learning network. During backpropagation in network training or fine-tuning, this operation naturally modulates the gradients, making the network parameter updates more biased towards the important features identified by the enhanced weight vector set. The final product of this series of operations is a fused feature optimization map, which is both a graphical representation of the entire optimization process and represents the new state of the ensemble learning network after this round of optimization.

[0071] Optionally, dividing the preliminary localization result into a trusted region and a region to be verified includes:

[0072] The multi-scale time-frequency feature set is input into the ensemble learning network to calculate the preliminary localization result;

[0073] Specifically, this step involves feeding the information-rich multi-scale time-frequency feature set generated in the previous step as input to a trained ensemble learning network. The ensemble learning network, through its complex internal nonlinear mapping relationships, performs in-depth analysis and reasoning on the input features, ultimately calculating and outputting a preliminary localization result. This preliminary localization result may consist of one or more potential fault segments and their corresponding initial confidence probabilities.

[0074] The confidence probability distribution of the preliminary positioning results is analyzed, the distribution entropy is calculated to quantify the uncertainty level, and a confidence distribution map is output.

[0075] Specifically, this step aims to quantify the uncertainty of the model regarding its own predictions. First, the confidence probability distribution in the preliminary location results is analyzed. Then, a generalized information entropy, the Rényi entropy, is used to calculate the distribution entropy. A probability distribution... Reni entropy It can be calculated using the following formula:

[0076] ,

[0077] in, It is the first in the preliminary positioning results The confidence probability of a possible faulty section; This indicates summing over all possible segments; It is an order parameter greater than or equal to 0 and not equal to 1. By adjusting this order parameter, the sensitivity to different parts of the probability distribution can be changed, thereby achieving multi-dimensional quantification of uncertainty. For example, when As the value approaches 1, the formula converges to the traditional Shannon entropy; when... When the value is large, more attention is paid to the peak portion of the probability distribution. This calculation outputs a confidence distribution map, which visually shows the positioning uncertainty level of each segment along the feeder.

[0078] Based on the confidence distribution map, a dynamic verification priority is assigned to each region, and a trusted region and a region to be verified are divided.

[0079] Specifically, based on the confidence distribution map generated in the previous step, a dynamic verification priority is assigned to each potential fault region. Regions with high entropy or low confidence will be given higher verification priority. Finally, according to this dynamic verification priority, a confidence threshold or entropy threshold is set. Regions with priority above this threshold are classified as regions to be verified, which require further computational resources for detailed verification; while regions with priority below this threshold are classified as reliable regions, and the location results of these regions are considered preliminarily reliable. Figure 3 As shown in the figure, the confidence distribution of four potential faulty sections is illustrated. The shape of the violin reflects the probability density of the confidence, the box plots inside indicate a statistical summary, and the scatter points represent the original confidence scores from multiple assessments. For example, the violin shapes of "section A" and "section C" are tall and slender with concentrated data points, indicating that the model is very certain about their location results and belongs to the "confidence region"; while the shapes of "section B" and "section D" are short and wide with scattered data, indicating high uncertainty and belonging to the "region to be verified".

[0080] Optionally, the calculation to obtain preliminary positioning results includes:

[0081] Acquire historical fault recording data of high and low voltage switchgear and historical power grid physical simulation data to generate a historical fault dataset.

[0082] Specifically, this step aims to provide sufficient and diverse samples for training the ensemble learning network. Historical fault waveform data comes from waveform files recorded by field equipment during real fault events. Historical power grid physical simulation data is augmented data generated by simulating various boundary conditions, fault types, and fault locations in electromagnetic transient simulation software based on the actual topology and parameters of the switchgear. After integrating, cleaning, and uniformly formatting these two types of data, a historical fault dataset containing rich fault modes is generated.

[0083] Using multi-scale time-frequency features as input and actual location as output, a neural network model is established and trained using the historical fault dataset to obtain an ensemble learning network.

[0084] Specifically, this step involves constructing the core intelligent localization model. A neural network model is built, employing a stacked generalization strategy from ensemble learning. This strategy comprises two layers. The first layer is the base learner, which consists of a Convolutional Neural Network (CNN) unit for extracting spatial features from a multi-scale time-frequency feature set, and a Long Short-Term Memory (LSTM) unit for analyzing its temporal evolution pattern. The CNN unit may contain three convolutional layers, each followed by a Rectified Linear Unit (ReLU) activation function and a max-pooling layer. The LSTM unit may contain two stacked LSTM layers, each with 128 hidden units. The second layer is the meta-learner, which concatenates the output feature vectors from the CNN and LSTM units in the first layer and uses this as input to a Gradient Boosting Decision Tree (GBDT) model. The GBDT model then outputs the final preliminary localization result. The training process of this ensemble learning network is as follows: A multi-scale time-frequency feature set extracted from a historical fault dataset is used as the model input, and the actual fault location labels and corresponding confidence distributions are used as the output. The Adam optimizer is used during training, and the learning rate can be dynamically adjusted during training, with its initial value set in the range of 1e-4 to 1e-3. To enhance the model's generalization ability, the historical fault dataset is augmented by adding Gaussian white noise with different signal-to-noise ratios, random time offsets, and amplitude scaling. To prevent overfitting, Dropout and L2 regularization are used in the training of the CNN and LSTM units. The training objective of the model is to optimize a composite loss function. To achieve this:

[0085] ,

[0086] in, It is the localization loss term, for example, using the mean squared error (MSE) to calculate the deviation between the model's predicted location and the actual location label; It is the confidence loss term, which penalizes predictions with high uncertainty by calculating the information entropy of the confidence distribution output by the model, and guides the model to produce more confident positioning. It is a regularization term used to prevent the model from overfitting; and These are two hyperparameters derived from the model training configuration, used to balance the contributions of the confidence loss and the regularization term to the total loss, respectively. By minimizing this composite loss function using optimization algorithms such as gradient descent, a convergent and stable ensemble learning network is finally trained.

[0087] The multi-scale time-frequency feature set is input into the ensemble learning network, and preliminary localization results are output.

[0088] Specifically, upon receiving a multi-scale time-frequency feature set generated by a new fault event, it is input into the ensemble learning network trained in the previous step. The network performs forward propagation calculations through its internal weights and activation functions, conducting in-depth analysis and inference on the input features, and finally calculating and outputting a preliminary localization result. This preliminary localization result includes predictions of one or more potential fault segments, as well as the model's confidence probability distribution for each predicted segment.

[0089] Optionally, the output precise wavefront time pair includes:

[0090] For each segment of the region to be verified, a dual-path preliminary calibration is performed, and traveling wave edge detection and data-driven detection are executed in parallel. The calibration results of the two paths are cross-validated, and the initial wavefront time is output.

[0091] Specifically, this step aims to perform high-precision wavefront calibration on the uncertain region to be verified by the ensemble learning network. Two independent detection paths are initiated in parallel: the first path is physical model-based traveling wave edge detection, such as applying an image-based line detection algorithm to find steep edges in the two-dimensional representation of the original traveling wave signal; the second path is data-driven detection, such as applying a pre-trained temporal pattern recognition network to find patterns similar to historical fault wavefronts. The calibration results output from the two paths are cross-validated, and only when the two results are consistent within the time tolerance range are they confirmed as a valid initial wavefront moment.

[0092] For the initial wavefront time, the intermediate localization result is calculated, and the confidence gradient of the ensemble learning network for the intermediate localization result is evaluated;

[0093] Specifically, the initial wavefront time output from the previous step is substituted into a two-end traveling wave ranging algorithm to calculate an intermediate localization result. This intermediate localization result is then used as a query and input back into the ensemble learning network. Through one forward and backward propagation calculation, the model's confidence gradient at this localization result is evaluated. The confidence gradient is a vector that indicates the direction in which the model's internal parameters should be adjusted to maximize the model's confidence in the current localization result.

[0094] The confidence gradient is used to update the internal parameters of the ensemble learning network.

[0095] Specifically, this step involves using the confidence gradient calculated in the previous step to perform a feedback update on the internal parameters of the ensemble learning network. This update process can employ gradient ascent or similar optimization algorithms to fine-tune the model, enabling it to generate higher confidence in the wavefront time calibrated and the calculated localization results in the current iteration.

[0096] The initial localization and confidence gradient evaluation are performed iteratively until the drift of the localization result and the confidence gain of the ensemble learning network are both lower than the preset convergence threshold. The wavefront time of the localization result in the converged state is then output as an accurate wavefront time pair.

[0097] Specifically, the aforementioned preliminary localization, confidence gradient evaluation, and parameter update steps are performed iteratively. At the end of each iteration, a convergence criterion function is used. To determine whether the iteration has terminated:

[0098] ,

[0099] in, and They are the first and The positioning result vector calculated in each iteration; It is the first The increase in the model's confidence in the localization result after each iteration is the confidence gain of the ensemble learning network. It is a weighting coefficient derived from the configuration and used to balance the importance of result stability and confidence gain; It is the convergence threshold, derived from the calibration experiment, when The iteration terminates when the value is less than the threshold. The wavefront time corresponding to the positioning result in the final convergence state is output as the precise wavefront time pair.

[0100] Optionally, the output comprehensive fault location report includes:

[0101] Based on the precise wavefront time pairs, multi-end ranging and energy distribution analysis are performed, and the propagation path constraint correction is performed by incorporating the fusion feature optimization map to generate preliminary multi-dimensional positioning results.

[0102] Specifically, this step aims to obtain a set of independent location solutions through parallel computation using multiple physical principles. Based on the precise wavefront time pairs output from the previous step, at least two location algorithms are initiated in parallel. The first is the multi-end traveling wave ranging method, which directly utilizes the time difference for calculation. The second is the transient energy distribution analysis method, which infers the fault point by analyzing the distribution gradient of fault energy in each section of the feeder. During the calculation process, information from the fused feature optimization map is used as a constraint condition for the propagation path to ensure that the calculation conforms to the actual credibility of the signal characteristics, thereby generating a preliminary multi-dimensional location result containing multiple independent physical location solutions.

[0103] A cross-consistency test is applied to the preliminary multi-dimensional localization results, and the test results are fed back to the ensemble learning network to update the parameter configuration and output an optimized localization result set.

[0104] Specifically, this step aims to validate and optimize the results obtained from parallel computation. Cross-consistency checks are performed on the individual physical location solutions in the initial multi-dimensional positioning results, for example, by calculating the Euclidean distance or physical deviations in the Geographic Information System (GIS) between the solutions. The check results, i.e., the deviation between solutions, are quantified and fed back to the ensemble learning network to fine-tune the parameter configuration of its output layer or confidence evaluation module. This feedback process aims to allow the ensemble learning network to learn and understand the performance differences of different physical positioning algorithms under specific conditions. After the parameter updates are completed, positioning calculations are performed again, outputting a pre-optimized and more consistent set of optimized positioning results.

[0105] The positioning results in the optimized positioning result set are weighted and fused to calculate the fault location coordinates and posterior probability. The fault location coordinates and posterior probability are then encapsulated into the comprehensive fault location report and output.

[0106] Specifically, to obtain a unique and most reliable final location conclusion, this step performs a weighted fusion of multiple location results from the optimized location result set. This fusion process employs the Bayesian Model Averaging (BMA) method. This method does not perform a simple arithmetic average; instead, it treats each location algorithm as an independent expert model and calculates the posterior probability of each location result based on its historical performance under the current operating conditions and the consistency of its current output. The final fault location coordinates are the weighted average of all location results, with the posterior probability as the weight. This method not only calculates an accurate fault location coordinate but also provides a statistically significant confidence evaluation, i.e., the posterior probability. Finally, the coordinates and the posterior probability are packaged together into a comprehensive fault location report for output, such as... Figure 4As shown in the figure, with time as the horizontal axis, this graph illustrates the dynamic evolution of robust risk indicators and their uncertainties. When a risk indicator reaches a dynamically adjusted safety threshold, the system triggers a closed-loop intervention command. Each successful intervention not only effectively controls the risk, but the experience gained is also learned by the system, manifested as a step-like increase in the amount of knowledge in the reusable intelligent template library corresponding to the Y-axis on the right. This reflects the complete intelligent closed loop of the entire system's decision-making, intervention, learning, and evolution.

[0107] To verify the feasibility of this invention in practice, it was applied to fault location of a switchgear feeder in an underground 10kV substation at a core urban transportation hub. This switchgear supplies power to critical loads and operates in a complex environment with strong electromagnetic interference and a compact internal structure. Traditional fault location methods struggle to quickly and accurately identify and locate early, high-resistance grounding faults caused by slow insulation degradation and subtle characteristics. To improve maintenance efficiency and prevent fault escalation, the project team adopted the method of this invention to achieve proactive and precise location of feeder faults within the switchgear.

[0108] In this embodiment, the project manager first inputs the feeder topology, equipment model, and historical operating data of all high- and low-voltage switchgear in the underground substation into the diagnostic process corresponding to the method of this invention. Through analysis, corresponding models and parameters are prepared for each potential fault diagnosis task. When a fault occurs, the method of this invention can perform multi-modal signal acquisition and fusion, extraction and optimization of time-frequency features, preliminary localization of the integrated learning network, and precise calibration of iterative convergence, until a comprehensive fault location report containing location and reliability is finally output.

[0109] To verify the effectiveness of the present invention, the diagnostic process of a real early high-resistance grounding fault that occurred on the night of July 10, 2024, is selected for illustration.

[0110] At 23:15 on July 10, 2024, a weak transient disturbance was detected in feeder A12 of switchgear #2 in the station. The method of this invention was immediately activated, and the broadband voltage and current traveling wave signals at the feeder port, as well as the acoustic emission and ultrasonic signals inside the switchgear were synchronously acquired.

[0111] The collected electrical signals were used as the first signal group, and the acoustic signals as the second signal group. Canonical correlation analysis was performed, and the canonical correlation coefficient was calculated to be 0.78. Through three rounds of iterative correction, the optimal iterative weight adjustment sequence was generated, and this sequence was applied to perform weighted fusion of the signals, outputting a weighted transient synchronization feature matrix with optimized signal-to-noise ratio.

[0112] An adaptive time-frequency transformation is performed on the weighted transient synchronization feature matrix, and a feature score vector is calculated using indicators such as information entropy. Based on this score vector, it is found that the feature quality of the ultrasonic signal in the initial stage of the fault is low. Therefore, the weighting process is corrected by feedback, and the weight of the ultrasonic signal is appropriately reduced, ultimately outputting a high-quality multi-scale time-frequency feature set.

[0113] The optimized multi-scale time-frequency feature set is input into a pre-defined ensemble learning network. The initial localization result output by the network points to a section approximately 1.5 meters near the A12 feeder cable termination. However, the confidence distribution map shows that the uncertainty level of this result is high. Therefore, this section is classified as a "region to be verified," and the next step of the precise calibration process is initiated.

[0114] For the 1.5-meter verification area, dual-path wavefront calibration was initiated, and traveling wave edge detection and data-driven detection were performed in parallel, with cross-validation to output the initial wavefront time. Based on the intermediate localization results calculated at this time and the confidence gradient of the ensemble learning network, four rounds of iterative feedback were conducted. During the iteration process, the drift of the localization results converged from the initial 0.5 meters to 0.05 meters, and the confidence gain also tended to stabilize, ultimately outputting a pair of accurate wavefront time pairs.

[0115] Based on precise wavefront time pairs, a parallel dual-end ranging algorithm and energy distribution analysis were applied to obtain two independent location results. Then, a weighted fusion feature-optimized map was incorporated, and a Bayesian consistency check was initiated. The final calculated fault location coordinates were 0.73 meters from the cable termination, with a posterior probability of 98.5%. The entire process took 1.2 seconds, generating and outputting a comprehensive fault location report. Based on this report, maintenance personnel accurately detected early insulation degradation at the cable termination during a power outage maintenance operation the following morning, successfully eliminating a major potential hazard.

[0116] Table 1 Adaptive Fusion Data Table for Multimodal Signals

[0117]

[0118] Table 2 Data on AI Pre-positioning and Iterative Convergence Process

[0119]

[0120] Table 3 Final localization results and model update data

[0121]

[0122] Tables 1-3 above record the practical application data of this invention in switchgear feeder fault location. Table 1 shows that, through canonical correlation analysis, adaptive weighting of multimodal signals was completed within tens of milliseconds, highlighting its intelligent fusion capability. Table 2 clearly records the convergence process from AI fuzzy pre-positioning to iterative precise positioning, proving the efficiency of its phased, step-by-step approximation strategy. Table 3 quantifies the final positioning accuracy and speed, and demonstrates the self-learning capability, using the experience feedback from this successful positioning to update and optimize the background model. This embodiment fully demonstrates the significant technical advantages of this invention in solving the problem of precise fault location in complex environments.

[0123] Based on the same inventive concept, this invention also provides a high- and low-voltage switchgear feeder fault location system based on transient traveling waves, such as... Figure 5 As shown, the system includes:

[0124] An adaptive signal fusion module is used to synchronously acquire broadband voltage and current traveling wave signals from faulty feeders, acoustic emission and ultrasonic signals from inside the switch cabinet, construct a signal cross matrix, perform signal correlation analysis on the signal cross matrix to adjust the mode weights, and generate a weighted transient synchronization feature matrix.

[0125] The time-frequency feature optimization module is used to perform a two-dimensional time-frequency transformation on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram, perform real-time feature scoring on the transient spectrum time series diagram, adjust the weighted transient synchronization feature matrix according to the real-time feature score, and output a multi-scale time-frequency feature set.

[0126] The intelligent pre-positioning and region division module is used to input the multi-scale time-frequency feature set into a preset ensemble learning network to obtain preliminary positioning results, and to divide the preliminary positioning results into reliable regions and regions to be verified through confidence distribution analysis.

[0127] The iterative convergence localization module is used to perform waveform edge detection and data-driven detection for each segment of the region to be verified, calibrate the traveling wave head time, and feed the calibration result back to the integrated learning network to update the internal parameters. The module iterates until the localization result converges to the fault segment and outputs the accurate wave head time pair.

[0128] The multi-model fusion and verification module is used to calculate multi-dimensional positioning results based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints, and to perform result consistency verification. The verification results are then fed back to update the parameter configuration of the integrated learning network, and a comprehensive fault location report is output.

[0129] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for locating faults in feeders of high- and low-voltage switchgear based on transient traveling waves, characterized in that, The method includes: The broadband voltage and current traveling wave signals of the faulty feeder and the acoustic emission and ultrasonic signals inside the switch cabinet are collected synchronously to construct a signal cross matrix. The signal cross matrix is ​​then subjected to signal correlation analysis to adjust the mode weights and generate a weighted transient synchronization feature matrix. A two-dimensional time-frequency transformation is performed on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram, and a real-time feature score is performed on the transient spectrum time series diagram. The weighted transient synchronization feature matrix is ​​adjusted according to the real-time feature score, and a multi-scale time-frequency feature set is output. The multi-scale time-frequency feature set is input into a preset ensemble learning network to obtain preliminary localization results. The preliminary localization results are then divided into reliable regions and regions to be verified through confidence distribution analysis. For each segment of the region to be verified, waveform edge detection and data-driven detection are performed, the traveling wave head time is calibrated, and the calibration result is fed back to the integrated learning network to update the internal parameters. The process is iterated until the positioning result converges to the fault segment, and the accurate wave head time pair is output. Based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints, multi-dimensional positioning results are calculated, and consistency checks are performed. The verification results are then fed back to update the parameter configuration of the integrated learning network, and a comprehensive fault location report is output.

2. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 1, characterized in that, The generated weighted transient synchronization feature matrix includes: The broadband voltage and current traveling wave signal is used as the first signal group, and the acoustic emission and ultrasonic signal are used as the second signal group. Correlation analysis is performed on the first signal group and the second signal group to calculate the canonical correlation coefficient. Based on the canonical correlation coefficient, the modal weights are iteratively corrected, and the cross-correlation changes of the signal cross matrix are monitored to generate an iterative weight adjustment sequence. The iterative weight adjustment sequence is applied to the signal cross matrix to perform weighted fusion, and a weighted transient synchronization feature matrix is ​​output.

3. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 2, characterized in that, The output multi-scale time-frequency feature set includes: The weighted transient synchronization feature matrix is ​​adaptively adjusted for time and frequency resolution. The transformation parameters are optimized by dynamic window function, and two-dimensional time and frequency transformation is performed to generate a transient spectrum time series diagram. Information entropy is evaluated and real-time feature scores are calculated for the transient spectrum time series diagram. The spectrum distribution uniformity and temporal consistency are quantified, and a feature score vector is output. Based on the feature scoring vector, the weighted transient synchronization feature matrix is ​​corrected by parameter feedback, and the transient spectrum time series diagram is fused to generate a multi-scale time-frequency feature set.

4. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 3, characterized in that, The output feature scoring vector includes: Multi-scale entropy spectrum analysis is performed on the transient spectrum time series diagram to calculate the local information entropy distribution and global spectrum complexity, and noise interference is filtered out to generate a preliminary entropy spectrum quantization set. For the preliminary entropy spectrum quantization set, the spectral distribution uniformity and temporal consistency are calculated, and a real-time feature scoring sequence is generated through weighted aggregation; Based on the real-time feature scoring sequence, vector normalization is performed, statistical confidence interval correction is used to optimize the scoring distribution, and feature scoring vectors are output.

5. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 3, characterized in that, The method further includes: The iterative weight adjustment sequence is correlated with the feature score vector through a multi-level association mapping to generate a weight score joint matrix. The dynamic optimization factor is calculated based on the weight scoring joint matrix, and the signal mode and time-frequency characteristics are analyzed to perform vector-level enhancement, outputting an enhanced weight vector set; The enhanced weight vector set is injected into the parameters of the ensemble learning network, and the internal weights and bias terms are optimized through gradient guidance to generate a fusion feature optimization map.

6. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 1, characterized in that, The step of dividing the preliminary localization result into a reliable region and a region to be verified includes: The multi-scale time-frequency feature set is input into the ensemble learning network to calculate the preliminary localization result; The confidence probability distribution of the preliminary positioning results is analyzed, the distribution entropy is calculated to quantify the uncertainty level, and a confidence distribution map is output. Based on the confidence distribution map, a dynamic verification priority is assigned to each region, and a trusted region and a region to be verified are divided.

7. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 1, characterized in that, The calculation yields the following preliminary positioning results: Acquire historical fault recording data of high and low voltage switchgear and historical power grid physical simulation data to generate a historical fault dataset. Using multi-scale time-frequency features as input and actual location as output, a neural network model is established and trained using the historical fault dataset to obtain an ensemble learning network. The multi-scale time-frequency feature set is input into the ensemble learning network, and preliminary localization results are output.

8. The method for locating faults in high- and low-voltage switchgear feeders based on transient traveling waves according to claim 1, characterized in that, The output precise wavefront time pair includes: For each segment of the region to be verified, a dual-path preliminary calibration is performed, and traveling wave edge detection and data-driven detection are executed in parallel. The calibration results of the two paths are cross-validated, and the initial wavefront time is output. For the initial wavefront time, the intermediate localization result is calculated, and the confidence gradient of the ensemble learning network for the intermediate localization result is evaluated; The confidence gradient is used to update the internal parameters of the ensemble learning network. The initial localization and confidence gradient evaluation are performed iteratively until the drift of the localization result and the confidence gain of the ensemble learning network are both lower than the preset convergence threshold. The wavefront time of the localization result in the converged state is then output as an accurate wavefront time pair.

9. The method for fault location of high and low voltage switchgear feeders based on transient traveling waves according to claim 5, characterized in that, The output comprehensive fault location report includes: Based on the precise wavefront time pairs, multi-end ranging and energy distribution analysis are performed, and the propagation path constraint correction is performed by incorporating the fusion feature optimization map to generate preliminary multi-dimensional positioning results. A cross-consistency test is applied to the preliminary multi-dimensional localization results, and the test results are fed back to the ensemble learning network to update the parameter configuration and output an optimized localization result set. The positioning results in the optimized positioning result set are weighted and fused to calculate the fault location coordinates and posterior probability. The fault location coordinates and posterior probability are then encapsulated into the comprehensive fault location report and output.

10. A high- and low-voltage switchgear feeder fault location system based on transient traveling waves, applied to the high- and low-voltage switchgear feeder fault location method based on transient traveling waves as described in any one of claims 1-9, characterized in that, The system includes: An adaptive signal fusion module is used to synchronously acquire broadband voltage and current traveling wave signals from faulty feeders, acoustic emission and ultrasonic signals from inside the switch cabinet, construct a signal cross matrix, perform signal correlation analysis on the signal cross matrix to adjust the mode weights, and generate a weighted transient synchronization feature matrix. The time-frequency feature optimization module is used to perform a two-dimensional time-frequency transformation on the weighted transient synchronization feature matrix to obtain a transient spectrum time series diagram, perform real-time feature scoring on the transient spectrum time series diagram, adjust the weighted transient synchronization feature matrix according to the real-time feature score, and output a multi-scale time-frequency feature set. The intelligent pre-positioning and region division module is used to input the multi-scale time-frequency feature set into a preset ensemble learning network to obtain preliminary positioning results, and to divide the preliminary positioning results into reliable regions and regions to be verified through confidence distribution analysis. The iterative convergence localization module is used to perform waveform edge detection and data-driven detection for each segment of the region to be verified, calibrate the traveling wave head time, and feed the calibration result back to the integrated learning network to update the internal parameters. The module iterates until the localization result converges to the fault segment and outputs the accurate wave head time pair. The multi-model fusion and verification module is used to calculate multi-dimensional positioning results based on the precise wavefront time pairs, combined with energy distribution analysis and fault propagation constraints, and to perform result consistency verification. The verification results are then fed back to update the parameter configuration of the integrated learning network, and a comprehensive fault location report is output.

Images