Distribution line fault intelligent section positioning method

Abstract

The invention belongs to the field of power distribution network fault diagnosis, and particularly discloses a power distribution line fault intelligent section positioning method, which does not depend on real-time waveform information in a traditional power supply state, does not adopt an accurate distance measurement method depending on a traveling wave propagation path, and does not depend on a high-frequency injection signal in combination with synchronous multi-point response. A fault feature matrix is constructed, and the AI classification model obtained through training is utilized to realize rapid and accurate identification of a fault section; and meanwhile, a cloud model continuous learning mechanism is designed, so that the system has the intelligent characteristics of self-evolution and self-adaption, and the technical bottleneck of insufficient consideration of intelligence, practicability and robustness in a power-off working condition of a traditional method is broken through.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network fault diagnosis, and specifically relates to a method for intelligent section location of power distribution line faults. Background Technology

[0002] With the increasing electrification of cities and the growing complexity of power distribution systems, 10kV overhead distribution networks play a core role in my country's power system. As the backbone of medium-voltage power supply, they not only bear the responsibility of providing stable power to residential and industrial users, but also undertake multiple functions such as distributed energy access, electric vehicle charging, and new load regulation. However, due to the susceptibility of overhead lines to external environmental factors, such as tree obstructions, animal contact with lines, and insulation aging, single-phase grounding faults occur frequently, seriously threatening the safety and reliability of the power distribution system.

[0003] If a grounding or open-circuit fault occurs and the faulty section cannot be accurately located within a short time, it will lead to a large-scale power outage, increase repair costs and time, and even cause secondary accidents. Currently, most traditional distribution network fault location methods rely on the normal operation of communication and voltage monitoring networks. Once the distribution network loses power due to a fault, these methods that rely on real-time power supply and communication become unusable, thus falling into the "last mile" location dilemma.

[0005] To address the problem of fault location in distribution networks, numerous studies have proposed a series of methods, such as methods based on one-dimensional electrical quantity feature extraction and machine learning classification, methods using graph neural networks to model and identify current waveforms, and methods converting electrical signals into two-dimensional images and then performing deep learning recognition. Specifically, Shi Yuxuan et al. proposed a fault section location method for distribution networks based on GAF-ResNet50 in March 2025. This method targets single-phase grounding faults in non-effectively grounded systems, proposing a section location method based on time-series signal image transformation combined with deep neural networks. It utilizes image representation combined with convolutional neural networks (CNNs) to achieve autonomous feature extraction. Although this method enhances feature extraction capabilities through image modeling, the GAF mapping used is sensitive to amplitude and easily affected by noise interference. Ding Xindi et al. proposed a fault location method for distribution networks based on improved Naive Bayes in April 2025. This method constructs typical feature quantities based on the waveform change characteristics of the current at the end of the line after a single-phase ground fault, and uses an improved Naive Bayes classification model for segment identification. However, this method is entirely based on the statistical features of the fault current waveform and does not introduce any high-frequency excitation mechanism, resulting in a high misclassification rate under noise interference. The Naive Bayes model has a simple structure but limited classification ability and is difficult to handle complex scenarios. Ling Feng et al. proposed a distributed fault location method for active distribution networks based on Empirical Wavelet Transform (EWT) in May 2025. This scheme proposes to use EWT to decompose the voltage signal and extract instantaneous energy features, and combine it with distributed information to achieve segment location. However, the EWT decomposition method is sensitive to signal boundary ambiguity and has high processing complexity. Patent application number 2024118459275 discloses a method for locating single-phase grounding fault sections in distribution networks based on multi-source data fusion. The waveform analysis method used is based on first-half-wave correlation clustering, emphasizing data comparison and rule-based judgment from the cloud master station, but it heavily relies on the accuracy and synchronization of data acquisition. Patent application number 2025102124810 proposes a scheme for identifying fault sections in distribution networks using the differences in three-phase current signals and a clustering algorithm. Its core is "differential elimination of load current + multi-dimensional feature extraction + clustering location"; it emphasizes multi-dimensional time-domain statistical feature extraction, suitable for handling fault types with significant signal changes; however, it does not involve spectral domain analysis, resulting in lower accuracy in locating high-frequency, small-disturbance faults.

[0006] It is evident that while the methods described above have achieved certain results under specific conditions, they still suffer from the following problems:

[0007] 1. Reliance on data acquisition during operation: Most methods are only applicable to the condition that the distribution network is still energized. However, in real-world scenarios, the distribution network cannot be located using conventional voltage or communication signals after a power outage.

[0008] 2. Lagging feature extraction methods and weak model generalization ability: Some methods use simple statistical features or traditional signal processing techniques (such as short-time Fourier transform and wavelet transform) for feature extraction. When faced with complex and ever-changing real-world conditions, the models often cannot adapt and lack robustness.

[0009] 3. Lack of closed-loop optimization mechanism: Most existing methods are one-time model training, lacking subsequent feedback from real-world data and the ability to continuously evolve the model, and cannot adaptively optimize the identification strategy according to new working conditions and fault types.

[0010] Therefore, in order to address the need for fault location in power distribution networks during power outages, there is a current demand for an intelligent solution that can acquire fault responses through active excitation signals in offline mode, combine advanced feature extraction technology, and possess cloud learning capabilities. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention aims to provide an intelligent fault location method for power distribution lines. This method does not rely on real-time waveform information under traditional power supply conditions, nor does it employ precise ranging methods that depend on traveling wave propagation paths. Instead, it constructs a fault feature matrix by combining high-frequency injection signals with synchronous multi-point responses, and utilizes a trained AI classification model to achieve rapid and accurate identification of faulty sections. Simultaneously, a cloud-based model continuous learning mechanism is designed, enabling the system to possess "self-evolutionary and adaptive" intelligent characteristics, overcoming the technical bottleneck of traditional methods in failing to balance intelligence, practicality, and robustness under power outage conditions.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A method for intelligent section location of faults in power distribution lines, which is implemented based on the following steps:

[0014] S1. Training Phase: Simulation Modeling and Feature Pre-extraction

[0015] Simulation tools were used to construct models of various typical operating and fault conditions of a 10kV power distribution system, generating a large number of voltage and current waveform samples. Key fault characteristic parameters were extracted through time domain, frequency domain, and time-frequency joint analysis methods, and multi-dimensional feature vectors were constructed to form a training sample library, providing high-quality data support for AI models.

[0016] S2. Core Positioning Stage: Initial AI-Powered Segment Assessment:

[0017] After a fault occurs in the distribution network and power is lost, the system actively applies a high-frequency voltage pulse signal to the faulty line to simulate and trigger a fault response. The synchronous sampling device deployed at each measuring point collects and uploads the response waveform in real time. After being processed by a pre-trained AI classification model, the system identifies the response differences between each measuring point and achieves rapid and intelligent identification of the faulty section through the constructed fault feature matrix.

[0018] S3. Continuous Optimization Phase: Cloud-based Model Evolution and Self-Learning

[0019] After fault location is completed, all on-site data and model recognition results are uploaded to the cloud platform for unified management; real operation samples are continuously accumulated, the training feature library is updated, online updates and migration optimization of the model are realized, and the adaptability and robustness of the model to unknown working conditions are continuously improved.

[0020] Furthermore, in step S1, during the early stages when the power distribution network is out of service and the actual waveform cannot be obtained, a simulation model of the power distribution network is built on the MATLAB / Simulink platform based on the actual characteristics.

[0021] Furthermore, during the simulation modeling in step S1, each fault condition is configured by randomly specifying the initial phase angle, grounding resistance, and fault location parameters during the data generation process. Under each fault scenario, the three-phase current and zero-sequence current waveforms of multiple measurement points are collected sequentially.

[0022] To ensure the representativeness and uniqueness of the dataset, the data labeling strategy is as follows: For the nth set of fault data, the waveform sample label collected at the mth measurement point is set to 10×m+n, ensuring that each data point can be uniquely mapped to the corresponding measurement point and fault condition.

[0023] Furthermore, in step S1, during feature selection, in order to comprehensively characterize the waveform change features under different fault conditions, three types of typical features are extracted: time-domain features, frequency-domain features, and time-frequency-domain features.

[0024] The time-domain features include instantaneous voltage amplitude, mean, peak value, kurtosis, skewness, margin factor, impulse factor, and waveform factor;

[0025] The frequency domain features include spectral energy distribution, dominant frequency components, frequency shift, and power spectral entropy;

[0026] The time-frequency domain features are obtained by using the energy distribution map obtained by S-transform, and the instantaneous energy, center frequency, sample entropy, and wavelet singular entropy are extracted.

[0027] Furthermore, step S1 requires effective screening of fault characteristics before modeling, specifically including:

[0028] The XGBoost algorithm is used to sort and filter the features by importance, and select the most discriminative feature subset as the core input for the subsequent construction of the fault feature matrix and intelligent recognition model.

[0029] The goal of the XGBoost algorithm is to minimize the overall loss function. It consists of a prediction error term and a regularization term:

[0030] (1)

[0031] In the formula, This represents the true label of the i-th sample; This represents the prediction value of XGBoost in the t-th iteration; Represents the loss function; This represents the regularization term for the k-th tree, which controls complexity and prevents overfitting. For the newly added trees in this round, The number of leaves, Leaf weights and The coefficient of the regularization term, It is a regularization term;

[0032] Each iteration adds a new regression tree, fits the residual from the previous iteration, and accumulates the gain from feature splitting to obtain a feature importance score; the total gain for a given feature... Defined as:

[0033] (2)

[0034] in, Indicates the total gain. Representation of features The total number of nodes used for splitting. This represents the decrease in the loss function caused by the split at the s-th node; finally, we obtain the gain value corresponding to each feature, where a larger value indicates a greater contribution of that feature to the model's discriminative ability.

[0035] Furthermore, the initial AI intelligent segment judgment in step S2 specifically includes:

[0036] S2.1. Actively apply a high-frequency voltage pulse signal with controllable frequency and amplitude to the target faulty line, and when it encounters the fault point, it will excite a series of electrical responses with fault characteristics, and comprehensively capture and collect the excitation signal;

[0037] S2.2. Based on the electrical quantity signals acquired synchronously from multiple points, a topological matrix expressing the response differences between measurement points is formed through feature extraction, normalization processing, and distance modeling to determine the fault range.

[0038] Furthermore, in step S2.1, in order to fully capture the excitation signal and ensure high precision and high synchronization of data acquisition, multi-point synchronous sampling signal collection devices are deployed at multiple key nodes. Each sampling node's signal collection device has a built-in high-precision GPS module to provide a unified clock source and ensure strict alignment of the collected data on the time axis.

[0039] Furthermore, step S2.2 specifically includes:

[0040] S2.2.1 Simultaneous sampling and feature extraction at multiple measurement points:

[0041] Suppose there are M measurement points in the entire system, and the time-domain waveform acquired by each measurement point is of length N. For each record, feature extraction is performed to extract a K-dimensional feature vector. Then, the feature vector of the m-th measurement point under the nth sampling is expressed as:

[0042] , (3)

[0043] Where m=1,…,M; n=1,…,N; This represents the k-th feature extracted from the m-th measurement point in the n-th sampling; fault features are filtered by the XGBoost algorithm;

[0044] S2.2.2 Feature Normalization and Matrix Construction:

[0045] To eliminate the influence of dimensions, each feature is normalized:

[0046] (4)

[0047] Define the normalized feature matrix for the nth failure event as:

[0048] (5);

[0049] in, This indicates that the matrix has M rows and K columns, and each element is a real number;

[0050] S2.2.3 Feature distance calculation and topological matrix generation:

[0051] To quantify the response differences between different measurement points, Euclidean distance is introduced to construct the difference matrix D between the measurement points. (n) Then we have:

[0052] (6)

[0053] in, This represents the sum of variances among the optimal fault characteristics at different measurement points; and For the matrix in formula (5) The set of values ​​represents several optimal fault characteristics of a measurement point, which is the difference measure between measurement point i and measurement point j. It is an element in the difference matrix.

[0054] The final generated symmetric matrix This refers to the topological structure of the measurement point response under this fault, which has the following form:

[0055] (7)

[0056] Where, d ij This indicates the degree of difference between the characteristic responses of measurement point i and measurement point j;

[0057] S2.2.4 Fault Section Identification and Pattern Analysis:

[0058] After completing the synchronous feature extraction at multiple measurement points, the system constructs a high-dimensional symmetric feature Euclidean distance matrix to measure the degree of difference in fault response among the measurement points; each element in this matrix... This reflects the similarity between the extracted features of the i-th and j-th measurement points. The larger the value, the more significant the difference. Under normal circumstances, if all measurement points are in the same segment and the feature changes are stable, the elements in the Euclidean distance matrix will be evenly distributed and there will be no obvious anomalies. However, once a certain part of the measurement points are in the fault-affected area, the feature difference between them and other measurement points will increase significantly, and a prominent "strip-like jump" area will appear in the matrix.

[0059] To accurately quantify the location of structural jumps and further infer the upper and lower boundaries of the faulty segment, a sliding window-based abrupt change detection algorithm is introduced. This algorithm detects the changing trend of each row or column of the matrix. The specific steps include:

[0060] ① Construction of the difference change:

[0061] For each row or column of the matrix Calculate the Euclidean difference with the next row:

[0062] (8)

[0063] in, This represents the distance vector between the i-th measuring point and other measuring points. G represents the distance vector between the (i+1)th measuring point and other measuring points; i It represents the degree of jump between the i-th and i+1-th measurement points in the "global feature difference distribution", which is an indicator of whether these two points belong to different fault sections;

[0064] ②Mutation threshold setting:

[0065] Because different fault conditions, loads, and line structures can lead to diverse differential amplitude distributions, a fixed threshold cannot be used to determine sudden changes. Therefore, an adaptive threshold strategy is introduced. :

[0066] (9)

[0067] in, Represents all g i The mean reflects the level of normal fluctuation; For all g i The standard deviation of the difference represents the overall fluctuation of the difference. These parameters are used to adjust sensitivity.

[0068] ③ Fault boundary determination:

[0069] when When the point is determined to be a sudden change point, it means that there may be a significant change in structural features between measurement point i and i+1, which is one of the boundaries of the fault section.

[0070] By statistically analyzing all those that satisfy The location set is used to identify the start and end boundary intervals of the faulty section. , An index for the first occurrence of a significant mutation location. This serves as an index for adjacent locations after the last significant mutation, thus dividing the original line into "potentially affected areas" and "unaffected areas," significantly reducing the search space for subsequent traveling wave precision localization algorithms.

[0071] Furthermore, the cloud-based model evolution and self-learning in step S3 specifically includes:

[0072] S3.1. After completing the initial identification of the fault section and the precise location by traveling wave ranging, all field measurement data and identification results, including injected signal characteristics, fault response waveforms, and traveling wave propagation time, will be automatically uploaded to the remote cloud platform to form a unified cloud data warehouse.

[0073] S3.2. Perform structured processing on the uploaded data, including data cleaning, feature extraction and standardization, and build a unified fault feature database and a corresponding tag library including fault type, specific location and waveform features;

[0074] S3.3. Utilize uploaded data to continuously update the fault diagnosis and location AI model in the cloud: Through transfer learning and incremental learning processes that are initiated regularly or in real time, the model actively adapts to new fault characteristics and changes in operating conditions;

[0075] S3.4. After the new model completes training and self-optimization in the cloud, it will first undergo automated performance evaluation. Only when the key performance indicators of accuracy and recall meet the set thresholds will the model be sent to the field device in a differential update manner to achieve rapid deployment and low-cost upgrade.

[0076] The beneficial effects of this invention are as follows:

[0077] 1. This invention can be used when the power is off: After the power is off a 10 kV overhead line, a high-frequency voltage pulse is injected into the line to simulate the power-on state and trigger a fault response.

[0078] 2. The present invention adopts a multi-point synchronous sampling framework: it has at least two small sampling terminals, which can be expanded to N, and has a built-in low-cost GPS / BeiDou time synchronization or wireless time synchronization module to ensure that the timestamps of the sampled data are aligned at the millisecond level.

[0079] 3. The present invention adopts an S-transform-driven fault feature matrix construction method: performing an S-transform on the injection response waveform of each measuring point to extract a high-resolution time-frequency energy vector; and splicing the vectors of each measuring point into a two-dimensional "synchronization feature matrix" in a preset order to directly characterize the differences between sections.

[0080] 4. This invention establishes an AI segment classification model based on preferred features: First, XGBoost is used to score 15 classic candidate features (including time-domain and time-frequency domain) to select K (1-6) core features with the highest discriminative power; then, the synchronous feature matrix composed of these features is used as input, and finally, the fault matrix is ​​analyzed by the mutation detection algorithm to obtain the fault interval.

[0081] 5. This invention constructs a cloud-based continuous learning closed-loop mechanism: on-site sampling data and segment judgment results are automatically uploaded to the cloud platform; after incremental training and transfer learning are implemented in the cloud, the updated model difference is sent to the terminal to realize model self-evolution and self-adaptation.

[0082] 6. This invention adopts a low-cost deployment scheme of single-end injection + flexible point placement: only one injection device and several portable sampling boxes are needed to complete the segment positioning; the communication link only uploads compressed features or results, with extremely low bandwidth requirements, making it suitable for rural and mountainous areas with weak communication. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0084] Figure 2 This is a simulation model of a 10kV three-phase four-wire overhead power distribution network for this invention;

[0085] Figure 3 This is a time-domain diagram of the signal of the present invention;

[0086] Figure 4 This is the frequency domain diagram of the signal of the present invention;

[0087] Figure 5 This is the time-frequency domain matrix diagram of the present invention;

[0088] Figure 6 This is a three-dimensional time-frequency domain diagram of the present invention;

[0089] Figure 7 This is an importance graph of the fault features before screening in this invention;

[0090] Figure 8 This is an importance graph after fault feature screening in this invention;

[0091] Figure 9 This is a simulation experiment verification diagram of the present invention. Detailed Implementation

[0092] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention.

[0093] like Figure 1 As shown, this invention addresses a typical 10kV overhead distribution network fault diagnosis scenario after a power outage. It integrates AI intelligent identification methods and an injection signal excitation mechanism to propose an intelligent fault section location method for distribution lines based on injection excitation and a synchronous feature matrix. This method does not rely on real-time waveform information under traditional power supply conditions, nor does it employ precise ranging methods dependent on traveling wave propagation paths. Instead, it constructs a fault feature matrix by combining high-frequency injection signals with synchronous multi-point responses and utilizes a trained AI classification model to achieve rapid and accurate identification of fault sections. Simultaneously, a cloud-based model continuous learning mechanism is designed, enabling the system to possess "self-evolutionary and adaptive" intelligent characteristics, overcoming the technical bottleneck of traditional methods' insufficient balance of intelligence, practicality, and robustness under power outage conditions. This effectively solves the technical challenges of "inability to perceive faults in real time" and "difficulty in quickly determining fault location" after a fault occurs. The specific implementation steps of this method include:

[0094] Step S1. Simulation modeling + feature pre-extraction (training phase):

[0095] In situations where obtaining realistic waveform data is difficult in the early stages of a power distribution network, simulation tools such as MATLAB are used to construct models of various typical operating and fault conditions of a 10kV power distribution system, generating a large number of voltage and current waveform samples. Through time-domain, frequency-domain, and time-frequency joint analysis methods, key fault characteristic parameters are extracted, and multi-dimensional feature vectors are constructed to form a training sample library. This provides high-quality data support for the AI ​​model, laying the foundation for its generalization ability. Specifically, this includes:

[0096] S1.1. Simulation Modeling:

[0097] In the early stages of a power outage in a distribution network where actual waveforms cannot be obtained, this invention aims to extract the optimal fault characteristics required for segment location technology using portable fault location devices. Based on real-world characteristics, this invention constructs a simulated distribution network model on the MATLAB / Simulink platform. This invention constructs a representative 10 kV three-phase four-wire overhead distribution network simulation model for illustration and demonstration. Its structure is as follows: Figure 2 As shown in the figure, this model uses a 10kV power supply to replace the injection device and simulates the typical "T" structure of the main line plus three branches (a total of three layers of branches). It has typical characteristics such as a large number of nodes, wide distribution of the end points, and deep path hierarchy, which fits the actual power distribution network form in rural and suburban areas.

[0098] In the model, all load nodes are configured as 220V residential loads, primarily considering the portability requirements of positioning technology deployment in low-voltage power scenarios, while also reflecting the realistic characteristics of distributed and complex scenarios. Typical load types, such as resistive-inductive residential loads, are selected to enhance the real representativeness of the data.

[0099] To study the changes in fault characteristics under various possibilities, this invention sets three typical single-phase grounding fault points in the figure, located at: the end node of the main line (phase A single-phase grounding); the end node of the middle branch line (phase B single-phase grounding); and the end node of the farthest branch line (phase C single-phase grounding).

[0100] To achieve comprehensive acquisition and analysis of waveform data from various measurement points under typical power distribution network fault scenarios, five measurement points and two controllable fault injection locations were set in the model, which can flexibly simulate single-phase grounding fault conditions under different distances and fault parameters.

[0101] During the data generation process, each fault condition is configured by randomly specifying key parameters such as initial phase angle, grounding resistance, and fault location. Under each fault scenario, the three-phase current (Ig) at 5 measurement points is collected sequentially. a , I b , I c The waveforms of the zero-sequence current (I0) and the zero-sequence current (I0) were collected, totaling 1024 sampling points. To ensure the representativeness and uniqueness of the dataset, the following data labeling strategy was designed:

[0102] For the nth set of fault data, the waveform sample label collected at the mth measurement point is set to 10 × m + n, thus ensuring that each data point can be uniquely mapped to the corresponding measurement point and fault condition. For example, the five samples in the first set are labeled as 11, 21, 31, 41, 51, respectively, and the second set is labeled as 12, 22, 32, 42, 52, and so on.

[0103] Ultimately, this patent generated 400 sets of fault condition data, covering various electrical initial conditions and fault characteristics, totaling 2000 waveform samples. Each sample contains four channels of data: three-phase current and zero-sequence current, as well as a unique identification label, providing a rich and well-structured high-quality data foundation for subsequent fault feature extraction and multi-measurement point identification model training.

[0104] S1.2. Feature Selection:

[0105] To comprehensively characterize waveform changes under different fault conditions, this invention extracts waveforms by combining the following three types of typical features:

[0106] Time-domain characteristics include voltage instantaneous amplitude, mean, peak value, kurtosis, skewness, margin factor, impulse factor, waveform factor, and other indicators.

[0107] Frequency domain characteristics include spectral energy distribution, dominant frequency components, frequency shift, and power spectral entropy.

[0108] Time-frequency domain features: Energy distribution map obtained by S-transform, and indicators such as instantaneous energy, center frequency, sample entropy, and wavelet singular entropy are extracted.

[0109] A total of 15 types of features were extracted. This paper takes a lightning-induced single-phase ground fault as an example, and examines the time-domain, frequency-domain, and time-frequency-domain related features.

[0110] Time-frequency domain features cannot be directly observed and studied by collecting signals; they require transformation of the original time-domain signal. Among the methods for converting ordinary time-domain waveforms into time-frequency matrices, the S-transform (Stockwell transform), as an emerging time-frequency analysis method, has received widespread attention in recent years. Compared to traditional Short-Time Fourier Transform (STFT) and Wavelet Transform, the S-transform maintains the time-frequency localization characteristics while also possessing the global phase-preserving capability of the Fourier transform. This allows it to exhibit higher analytical accuracy and information retention in processing non-stationary signals with rapidly changing frequencies.

[0111] Specifically, the S-transform has high frequency resolution in the low-frequency region and better time resolution in the high-frequency region, making it naturally suitable for signal analysis tasks such as power system fault signals, which have significant time-varying characteristics and drastic spectral changes. Its adaptive window function design combines the multi-scale concept of wavelet transform with the stable structure of Fourier transform, enabling it to obtain clear and accurate time-frequency distributions in a variety of complex scenarios.

[0112] Therefore, the S-transform is considered a more advanced time-frequency transformation method than the STFT and wavelet transform, and is more suitable for the analysis of non-stationary signals. In particular, it has shown better performance and stronger adaptability in power system applications such as distribution network fault identification and pulse traveling wave analysis.

[0113] Through comprehensive analysis of waveforms in the time, frequency, and time-frequency domains, the system extracts multi-dimensional fault features, providing more comprehensive, accurate, and reliable data support for subsequent AI identification. The time-domain, frequency-domain, and time-frequency domain matrices and three-dimensional graphs are shown below. Figure 3 As shown in 4, 5, and 6.

[0114] S1.3. AI algorithm selects optimal features:

[0115] To improve the training efficiency and generalization ability of the distribution network fault identification model, it is essential to effectively screen fault features before modeling. Since the original extracted features have high dimensionality and contain a lot of redundant and noisy information, directly using them for model training not only leads to high computational costs but may also affect the final identification accuracy. Therefore, this FAMING algorithm employs XGBoost (ExtremeGradient Boosting) to prioritize and filter features, selecting the most discriminative subset as the core input for subsequently constructing the fault feature matrix and the intelligent identification model.

[0116] XGBoost is an enhanced version of Gradient Boosting Tree (GBDT) model widely used in machine learning. Compared with traditional methods, it has the following advantages: Strong feature scoring capability: It can automatically evaluate the degree of improvement of each feature on model performance through the gain function; Strong robustness: It is not sensitive to noisy features and missing values, and is suitable for non-ideal data in real-world engineering; Strong nonlinear modeling capability: It can uncover the intrinsic relationship between complex nonlinear features and failure modes; Built-in regularization mechanism: It effectively avoids overfitting and improves the model's generalization ability; High efficiency: It supports parallel optimization and distributed training, and is suitable for large-scale feature selection. XGBoost is the preferred algorithm for implementing an efficient, stable, and highly interpretable feature selection mechanism in this scheme.

[0117] The goal of the XGBoost algorithm is to minimize the overall loss function. It consists of a prediction error term and a regularization term:

[0118] (1)

[0119] In the formula, This represents the true label of the i-th sample; This represents the prediction value of XGBoost in the t-th iteration; Represent the loss function (often chosen as squared error or logarithmic loss); This represents the regularization term for the k-th tree, which controls complexity and prevents overfitting. For the newly added trees in this round, The number of leaves, Leaf weights and The coefficient of the regularization term, It is a regularization term;

[0120] Each iteration adds a new regression tree, fits the residual from the previous iteration, and accumulates the gain from feature splitting to obtain a feature importance score; the total gain for a given feature... Defined as:

[0121] (2)

[0122] in, Indicates the total gain. Representation of features The total number of nodes used for splitting. This represents the decrease in the loss function caused by the split at the s-th node; finally, we obtain the gain value corresponding to each feature, where a larger value indicates a greater contribution of that feature to the model's discriminative ability.

[0123] After using XGBoost to score and rank the original extracted features in the time domain, frequency domain, and time-frequency domain, the top k features with the highest gain were selected. These key features serve the following purposes:

[0124] Constructing a fault feature matrix: compressing the original feature space while retaining the most sensitive indicators for fault identification;

[0125] Improve subsequent model performance: increase model training speed and stability, and enhance accuracy and generalization ability;

[0126] Enhancing engineering interpretability: Helping to understand which physical features best characterize the type and location of faults;

[0127] Provide support for the two-layer intelligent positioning system: Select the best features as input for the subsequent "coarse location of the fault section" to ensure the overall performance of the system;

[0128] Based on the fault data collected from the most representative 10kV three-phase four-wire overhead distribution network simulation model built according to this invention, the XGBoost algorithm was used to score and rank 15 features extracted in the original time domain, frequency domain, and time-frequency domain. The top four feature variables with the highest gain were then selected: transient energy, first peak percentage, wavelet singular entropy, and frequency band energy entropy. The figures before and after the selection of fault feature importance are shown below. Figure 7 As shown in Figure 8.

[0129] Step S2. AI Intelligent Segment Preliminary Judgment (Core Localization Stage):

[0130] When a power distribution network experiences a fault and power outage, the system actively applies a high-frequency voltage pulse signal (e.g., 10kV / 100A level) to the faulty line to simulate a fault response. Synchronous sampling devices deployed at each measuring point collect and upload the response waveform in real time. After processing by a pre-trained AI classification model, the response differences between measuring points can be identified. A constructed fault feature matrix enables rapid and intelligent identification of the faulty section. This method can accurately determine the approximate location of the fault without depending on the power supply status, providing effective guidance for subsequent maintenance. Specifically, it includes:

[0131] S2.1. Fault waveform excitation and signal collection

[0132] When a fault occurs in the distribution network, leading to a loss of power to the lines, traditional fault diagnosis and location methods that rely on normal power supply conditions become ineffective. To address the challenge of fault identification under power outage conditions, this method proposes high-frequency voltage pulse injection, which can actively apply a high-frequency voltage pulse signal with controllable frequency and amplitude to the target faulty line without an external power source. This high-frequency signal propagates along the line and, upon encountering the fault point, excites a series of electrical responses with fault characteristics, such as reflection, attenuation, and distortion, thereby simulating the dynamic electromagnetic behavior during a fault occurrence.

[0133] To comprehensively capture these excitation signals and ensure high precision and synchronization in data acquisition, the system deploys multi-point synchronous sampling signal collection devices at several key nodes, enabling real-time, high-speed sampling of the injected excitation response waveform. Each sampling node has a built-in high-precision GPS module to provide a unified clock source, ensuring strict alignment of the acquired data on the time axis. This provides a solid foundation for lateral comparative analysis of waveform characteristics and time-series modeling of signal propagation paths.

[0134] Waveform data acquired through multi-point synchronous acquisition can be used to construct a multi-dimensional spatiotemporal response matrix, which reflects the response pattern of the high-frequency excitation signal at each sampling point. Combined with a previously trained fault identification model or differential analysis method, it is possible to further infer the abnormal intervals of the excitation signal propagation, enabling rapid localization of the approximate fault location and providing accurate interval boundaries and a narrowed search space for subsequent high-precision traveling wave ranging.

[0135] S2.2. Fault Feature Matrix Construction to Determine Fault Range

[0136] To achieve collaborative modeling and fault section identification using multi-point measurement data in distribution networks, this paper proposes a fault feature matrix construction method based on high-frequency injected response signals. This method uses synchronously acquired electrical signals from multiple points as a foundation, and through feature extraction, normalization, and distance modeling, forms a topological matrix expressing the response differences between measurement points, providing effective support for subsequent fault identification and location. Specifically, it includes:

[0137] S2.2.1 Simultaneous sampling and feature extraction at multiple measurement points:

[0138] Suppose there are M measurement points in the entire system, and the time-domain waveform acquired by each measurement point is of length N. For each record, feature extraction is performed to extract a K-dimensional feature vector. Then, the feature vector of the m-th measurement point under the nth sampling is expressed as:

[0139] , (3)

[0140] Where m=1,…,M; n=1,…,N; This represents the k-th feature extracted from the m-th measurement point in the n-th sampling; fault features are filtered by the XGBoost algorithm.

[0141] S2.2.2 Feature Normalization and Matrix Construction:

[0142] To eliminate the influence of dimensions, each feature is normalized:

[0143] (4)

[0144] Define the normalized feature matrix for the nth failure event as:

[0145] (5).

[0146] in, This indicates that the matrix has M rows and K columns, and each element is a real number;

[0147] S2.2.3 Feature distance calculation and topological matrix generation:

[0148] To quantify the response differences between different measurement points, Euclidean distance is introduced to construct the difference matrix D between the measurement points. (n) Then we have:

[0149] (6)

[0150] The sum of the variances of the optimal fault features between different measurement points is the sum of the variances of fault features where several optimal fault features are the same for two measurement points. The number of fault features at each measurement point has already been calculated in the matrix of formula (5). and It is in this matrix The set represents several optimal fault characteristics of a measurement point, which is the difference measure between measurement point i and measurement point j. It is an element in the difference matrix.

[0151] The final generated symmetric matrix This refers to the topological structure of the measurement point response under this fault, which has the following form:

[0152] (7)

[0153] Where, d ij This indicates the degree of difference between the characteristic responses of measurement point i and measurement point j.

[0154] S2.2.4 Fault Section Identification and Pattern Analysis:

[0155] After completing the synchronous feature extraction at multiple measurement points, the system constructs a high-dimensional symmetric feature Euclidean distance matrix to measure the degree of difference in fault response among the measurement points; each element in this matrix... It reflects the similarity between the extracted features of the i-th and j-th measurement points. The larger the value, the more significant the difference. Under normal circumstances, if all measurement points are in the same segment and the feature changes are stable, the elements in the Euclidean distance matrix are evenly distributed and there are no obvious anomalies. However, once a certain part of the measurement points are in the fault-affected area, the feature difference between them and other measurement points will increase significantly, and a prominent "strip jump" area will appear in the matrix.

[0156] Although the symmetric Euclidean distance matrix can be used to roughly observe which regions have characteristic changes, the matrix is ​​essentially a two-dimensional symmetric structure. It is difficult to accurately identify the location and magnitude of the transition boundary manually or by thresholding, especially under weak fault, background noise or complex grounding conditions, the transition edge is often blurred.

[0157] To accurately quantify the location of structural transitions and further infer the upper and lower boundaries of the faulty segment, this invention introduces a sliding window change detection algorithm. Its core function is to detect the changing trend of each row (or column) of the matrix, thereby accurately identifying the location of the change. This method can significantly improve the robustness and accuracy of faulty segment boundary identification. The specific steps of the change detection algorithm include:

[0158] ① Construction of the difference change:

[0159] For each row or column of the matrix Calculate the Euclidean difference with the next row:

[0160] (8)

[0161] in, This represents the distance vector between the i-th measuring point and other measuring points. G represents the distance vector between the (i+1)th measuring point and other measuring points; i It represents the degree of jump between the i-th and i+1-th measurement points in the "global feature difference distribution", which is an indicator of whether these two points belong to different fault sections;

[0162] ②Mutation threshold setting:

[0163] Because different fault conditions, loads, and line structures can lead to diverse differential amplitude distributions, a fixed threshold cannot be used to determine sudden changes. Therefore, an adaptive threshold strategy is introduced. :

[0164] (9)

[0165] in, Represents all g i The mean reflects the level of normal fluctuation; For all g i The standard deviation of the difference represents the overall fluctuation of the difference. These parameters are used to adjust sensitivity.

[0166] ③ Fault boundary determination:

[0167] when When a point is identified as a sudden change, it means that there may be a significant change in structural characteristics between measurement point i and i+1, which is one of the boundaries of the fault section; by statistically analyzing all points that satisfy the condition... The location set is used to identify the start and end boundary intervals of the faulty section. , An index for the first occurrence of a significant mutation location. This serves as an index for adjacent locations after the last significant mutation, thus dividing the original line into "potentially affected areas" and "unaffected areas," significantly reducing the search space for subsequent traveling wave precision localization algorithms.

[0168] The core value of this mutation detection algorithm lies in:

[0169] 1. Structural extraction of transition boundaries: Transforming the "fuzzy transitions" in the original feature matrix into clear numerical abrupt change indicators;

[0170] 2. Strong robustness: The adaptive threshold strategy avoids the subjectivity of manual setting and improves the accuracy under different noise conditions;

[0171] 3. High efficiency: It can quickly identify boundaries in the feature matrix without large-scale training or model involvement, making it suitable for on-site deployment;

[0172] 4. Provides high-quality initial judgment sections for traveling wave ranging: reduces ranging errors and improves reliability.

[0173] Step S3. Cloud-based model evolution and self-learning (continuous optimization phase):

[0174] After fault location is completed, all data collected on-site and model recognition results will be uploaded to a cloud platform for unified management. The system continuously accumulates real-world operating samples and updates the training feature library, enabling online model updates and migration optimization. This continuously improves the model's adaptability and robustness to unknown operating conditions, forming an intelligent closed loop of "on-site data feedback → cloud-based model evolution → model deployment optimization," ensuring the model has reliable segment judgment capabilities under different regions and conditions. Specifically, this includes:

[0175] S3.1. After completing the initial identification of the fault section and the precise location by traveling wave ranging, all the field measurement data (including injected signal characteristics, fault response waveforms, traveling wave propagation time, etc.) and identification results (such as section determination information and precise fault distance) provided by this invention will be automatically uploaded to the remote cloud platform to form a unified cloud data warehouse.

[0176] S3.2. The cloud platform has a built-in data management module and a model training module. The data management module is responsible for structuring the uploaded data, including data cleaning, feature extraction and normalization, and building a unified fault feature database and corresponding label library (including fault type, specific location, and waveform features). As this data continues to accumulate, it can provide diverse and high-quality sample support for model training.

[0177] S3.3. The model training module utilizes uploaded data to continuously update the fault diagnosis and localization AI model in the cloud. Specifically, through periodic or real-time transfer learning and incremental learning processes, the model can proactively adapt to new fault characteristics and changes in operating conditions, overcoming the lack of generalization caused by traditional offline training methods, and significantly improving the model's recognition accuracy and robustness in unknown fault scenarios.

[0178] S3.4. The cloud platform also has an online evaluation and verification mechanism. After the new model completes training and self-optimization in the cloud, it will first undergo automated performance evaluation. Only when key performance indicators such as accuracy and recall meet the set thresholds will the model be sent to the field device in a differential update manner, so as to achieve rapid deployment and low-cost upgrade.

[0179] This invention establishes a complete intelligent closed-loop feedback mechanism: "Field feedback data → Cloud model evolution (transfer learning and adaptive optimization) → Online model evaluation → Optimized model redeployment in the field." This closed-loop mechanism ensures the continuous growth and self-improvement of the AI ​​model, enabling it to gradually evolve from a "theoretically derived model" into a "fault location expert" with rich field experience as actual operational data accumulates. This achieves long-term sustainable optimization of intelligent fault location technology, truly realizing the enormous application potential of artificial intelligence in the power sector.

[0180] Simulation and fault location experiment:

[0181] To verify the effectiveness of the fault location method proposed in this invention, a typical 10 kV three-phase four-wire overhead distribution network simulation model built above was used, with the following topology: Figure 9 As shown, the model is divided into four independent sections, each connected by π-type line elements, simulating the typical configuration and operating conditions of actual power distribution lines.

[0182] In this simulation experiment, the fault module (Three-Phase Fault) was placed in the middle of section two to simulate a typical line short-circuit fault. During the experiment, a 10kV power supply was used to apply a high-frequency transient signal to the line instead of a high-frequency pulse injection device. The fault response waveforms at each measurement point in the simulation were recorded. Feature extraction and data analysis were performed using four optimal feature parameters previously obtained based on the XGBoost algorithm. Finally, the feature Euclidean distance matrix (i.e., the fault matrix) shown below was constructed.

[0183]

[0184] The numerical distribution of the above fault feature matrix clearly shows that:

[0185] The characteristic Euclidean distance between measuring point ② (section 2) and other measuring points is significantly higher (such as the highlighted values ​​5.32, 4.87, 5.05, etc.), indicating that section 2 is most significantly affected by the fault.

[0186] Conversely, the characteristic distances between measuring points in other sections (measuring point ① and measuring point ③, measuring point ① and measuring point ④, etc.) are relatively small, indicating that these sections are less affected by the fault or are not affected at all.

[0187] The "abrupt stripe" feature corresponding to measurement point ② in the fault feature matrix is ​​particularly obvious, which can determine that the fault location in this experiment is located in section two.

[0188] Because the 10kV three-phase four-wire overhead distribution network model selected in this invention is typical and representative in terms of topology, with a clear and relatively simple structural division, and the location of the fault point is clear and easy to distinguish, the fault section can be intuitively observed through the method described above (i.e., the fault feature matrix constructed after XGBoost feature extraction), and the fault location can be accurately determined without further employing a mutation detection algorithm. This intuitive approach is suitable for scenarios with simple topologies, obvious fault characteristics, and minimal background interference, and can efficiently and quickly complete the fault location task.

[0189] However, in actual power distribution networks, due to more complex line structures, variable load conditions, strong noise interference, or unclear fault characteristics, the difference bands in the constructed fault feature matrix may not be obvious, making it difficult to directly determine the fault location based on intuition or simple observation. In this case, relying solely on visual inspection or simple threshold methods can easily lead to misjudgment or missed judgment.

[0190] To address the aforementioned complex situations, the mutation detection algorithm proposed in this invention demonstrates its significant value. By quantifying the changing features in the feature matrix into specific numerical difference indices (difference sequences) and adaptively determining the threshold using statistical methods, the robustness and accuracy of fault segment identification are significantly improved. Under complex background conditions, the mutation detection algorithm can help engineers more clearly identify the boundary locations of fault segments, supplementing the shortcomings of relying solely on intuitive observation, and effectively complementing the comprehensiveness, robustness, and universality of this method.

[0191] Therefore, this invention can efficiently and quickly locate faulty sections intuitively under simple and typical network structures. In complex network conditions or situations where features are not obvious, it can be combined with a mutation detection algorithm to further improve the accuracy and reliability of fault location, making the method more widely applicable and flexible in engineering. Based on simulation settings and the analysis results of the fault matrix, the fault section location method proposed in this invention, based on high-frequency injection, XGBoost feature extraction, and mutation detection algorithms, can quickly and accurately identify the actual faulty section, achieving accurate and reliable fault location, fully verifying the effectiveness and engineering practicality of this invention.

[0192] Once the fault location is complete, key data such as voltage and current waveforms, traveling wave reflection characteristics, and fault location information will be uploaded to a cloud platform to further enrich the training samples and optimize the feature database. This process not only helps improve the model's memory of typical fault scenarios but also enhances its generalization ability under complex operating conditions, thereby continuously improving the overall system's intelligence level and location accuracy.

[0193] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for intelligent fault section location of a power distribution line, the method comprising: The method is realized based on the following steps: ​ S1. Training stage: simulation modeling and feature pre-extraction: A variety of typical operation and fault conditions models of 10kV distribution system are constructed by using simulation tools, generating a large number of voltage and current waveform samples; through time domain, frequency domain and time-frequency joint analysis method, key fault feature parameters are extracted, and multi-dimensional feature vector is constructed, forming a training sample library, providing high-quality data support for AI model; S2. Core positioning stage: AI intelligent section preliminary judgment: After the fault occurs and the power is cut off, the system actively applies a high-frequency voltage pulse signal to the fault line, simulates the excitation of the fault response; the synchronous sampling device deployed at each measuring point collects and uploads the response waveform in real time, identifies the response difference between each measuring point through the pre-trained AI classification model, and realizes the rapid and intelligent identification of the fault section through the constructed fault feature matrix; S3. Continuous optimization stage: cloud model evolution and self-learning: After the fault location is completed, all the data collected on site and the model recognition results are uploaded to the cloud platform for unified management; real operation samples are continuously accumulated, the training feature library is updated, the model is updated and migrated and optimized online, and the adaptability and robustness of the model to unknown conditions are continuously improved.

2. The method of claim 1, wherein: Step S1 in the early stage of power failure in the distribution network, according to the field characteristics, a simulation distribution network model is built on the MATLAB / Simulink platform.

3. The method of claim 2, wherein: In step S1, during data generation, each fault condition is configured by randomly specifying initial phase angle, grounding resistance and fault location parameters. Under each fault scenario, the three-phase current and zero-sequence current waveforms of multiple measurement points are collected in turn. To ensure the representativeness and uniqueness of the data set, the data labeling strategy is as follows: for the nth fault data, the waveform sample label of the mth measurement point is set to 10xm+n, ensuring that each data can be uniquely mapped to the corresponding measurement point and fault condition.

4. The method of claim 3, wherein: In step S1, in order to fully characterize the waveform variation characteristics under different fault conditions, three types of typical features are extracted: time domain features, frequency domain features and time-frequency domain features. The time domain features include voltage instantaneous amplitude, mean value, peak value, kurtosis, skewness, margin factor, pulse factor and waveform factor. The frequency domain features include frequency spectrum energy distribution, main frequency component, frequency shift and power spectrum entropy. The time-frequency domain features include energy distribution obtained by S transform, instantaneous energy, center frequency, sample entropy and wavelet singular entropy.

5. The method of claim 4, wherein: In step S1, before modeling, the fault features need to be effectively selected, including: XGBoost algorithm is used to sort and select the features, and the most discriminative feature subset is selected as the core input for building fault feature matrix and intelligent identification model; The goal of the XGBoost algorithm is to minimize the overall loss function which consists of a prediction error term and a regularization term: （1） wherein, represents the true label of the i-th sample; represents the prediction value of XGBoost at the t-th iteration; represents the loss function; represents the regularization term of the k-th tree, which controls the complexity and prevents overfitting; is the number of newly added trees in this round, is the number of leaves, is the leaf weight, and is the regularization term coefficient, is the regularization term; Each round of iteration adds a new regression tree, fitting the last residual, through the cumulative gain of the feature split, the feature importance score is obtained; the total gain of a certain feature is defined as: （2） wherein, denotes the total gain, denotes the feature total number of nodes used for splitting, denotes the loss function reduction value brought by the s-th node splitting; the final gain value corresponding to each feature is obtained, and the greater the value is, the greater the contribution of the feature to the model discrimination ability is.

6. The method of claim 1, wherein: In step S2, AI intelligent section preliminary judgment includes: S2.

1. A controllable frequency and amplitude high-frequency voltage pulse signal is actively applied to the target fault line, and a series of electrical responses with fault characteristics are excited when the fault point is encountered, and the excitation signal is fully captured and collected; S2.

2. Based on the electrical quantity signals collected synchronously at multiple points, a topological matrix expressing the response difference between measuring points is formed through feature extraction, normalization processing and distance modeling to determine the fault section.

7. The method of claim 6, wherein: In step S2.1, in order to comprehensively capture the excitation signal and ensure high precision and high synchronism of data collection, signal collection devices for synchronous sampling are arranged at multiple key nodes. The signal collection device of each sampling node is internally provided with a high-precision GPS module to provide a unified clock source and ensure strict alignment of collected data on the time axis.

8. The method of claim 6, wherein: Step S2.2 specifically includes: S2.2.1 Multi-point synchronous sampling and feature extraction: Suppose there are M measuring points in the entire system, and the length of the time-domain waveform collected by each measuring point is N. Feature extraction is performed on each record to extract a K-dimensional feature vector. The feature vector of the mth measuring point at the nth sampling is represented as: ， （3） wherein m = 1, …, M; n = 1, …, N; Xm,n,k represents the kth feature extracted from the mth measuring point in the nth sampling; the fault feature is screened by an XGBoost algorithm; S2.2.2 Feature normalization and matrix construction: In order to eliminate the dimension effect, each feature is normalized: （4） The normalized feature matrix under the nth failure event is defined as: （5）； wherein denotes that the size of the matrix is M rows, K columns, and each element is a real number; S2.2.3 Feature distance calculation and topological matrix generation: To quantify the response difference between different measurement points, the Euclidean distance is introduced to construct the difference matrix D between measurement points (n) Then we have: （6） wherein, represents the sum of the variance between the optimal fault features of different measurement points; and is the set of in the matrix of formula (5), representing several optimal fault features of a measurement point, that is, the difference measurement value between measurement point i and measurement point j, which is an element in the difference matrix; The final generated symmetric matrix , which is the topological structure of the response of the measuring point under this fault, has the following form: （7） wherein d ij represents the difference between the characteristic responses of measurement point i and measurement point j; S2.2.4 Fault section identification and pattern analysis: After the multi-point synchronous feature extraction is completed, the system constructs a high-dimensional symmetric feature Euclidean distance matrix, which is used to measure the difference degree of fault response between each measuring point. Each element in the matrix reflects the similarity between the features extracted from the ith and jth measuring points. The larger the value, the more significant the difference. Under normal circumstances, if all measuring points are in the same section and the feature changes smoothly, the elements in the Euclidean distance matrix are evenly distributed and there is no obvious anomaly. Once a part of the measuring points is in the fault influence area, the feature difference between them and other measuring points will increase significantly, and there will be a prominent "strip jump" area in the matrix. In order to accurately quantify the jump position from a structural point of view and further infer the upper and lower boundaries of the fault section, a sliding window mutation detection algorithm is introduced to detect the change trend of each row or column of the matrix. The specific steps include: ① Difference change quantity construction: for each row or column of the matrix compute the euclidean difference with the next row: （8） wherein, represents the distance vector of the i-th measurement point to other measurement points, represents the distance vector of the i+1-th measurement point to other measurement points; g i represents the degree of jump of the i-th and i+1-th measurement points on the "global feature difference distribution", that is, an index of whether the two points belong to different fault sections; ② Mutation threshold setting: Because of the diversity of the distribution of the amplitude of the differential due to different fault conditions, different loads and line structures, a fixed threshold cannot be used to determine the mutation; for this purpose, an adaptive threshold strategy is introduced : （9） wherein, represents the mean of all g i reflecting the normal fluctuation level; is the standard deviation of all g i representing the overall fluctuation of the difference; is an adjustment parameter for controlling the sensitivity; ③ Fault boundary determination: When is judged as a mutation point, it means that there is a possible obvious structural feature change between the measuring points i and i+1, which is one of the fault section boundaries. By counting all the position sets satisfying , the start and end boundary intervals of the fault section are identified , is the index of the first position where a significant mutation occurs, is the index of the adjacent position after the last significant mutation; thus, the original line is divided into "possible fault-affected area" and "unaffected area", which greatly reduces the search space for the subsequent traveling wave fine positioning algorithm.

9. The method of claim 1, wherein: In step S3, cloud model evolution and self-learning specifically includes: S3.

1. After completing the preliminary identification of the fault section and the precise positioning of the traveling wave, all field measurement data and identification results including the injected signal features, fault response waveforms, and traveling wave propagation times are automatically uploaded to the remote cloud platform to form a unified cloud data warehouse; S3.

2. Structured processing of uploaded data, including data cleaning, feature extraction and standardization, and construction of a unified fault feature database and corresponding label library including fault type, specific location, and waveform features; S3.

3. Using the uploaded data, continuously update the cloud fault diagnosis and positioning AI model: through the migration learning and incremental learning process started periodically or in real time, the model actively adapts to new fault features and working condition changes; S3.

4. After the new model is trained and self-optimized in the cloud, it will first undergo automatic performance evaluation. Only when the accuracy and recall rate key performance indicators meet the set threshold, the model will be updated in a differential manner to the field device, realizing rapid deployment and low-cost upgrade.

Images