A power distribution network insulation fault detection and positioning method and system based on multi-source transient data

Abstract

This invention relates to the field of power system relay protection and fault detection technology, and discloses a method and system for detecting and locating insulation faults in distribution networks based on multi-source transient data. The method includes the following steps: S1, synchronously acquiring transient traveling wave signals generated by the fault at monitoring points in the distribution network, and capturing the initial traveling wave front data arriving at each monitoring point; S2, performing time-frequency analysis on the captured initial traveling wave front data. This method and system for detecting and locating insulation faults in distribution networks based on multi-source transient data directly identifies faulty branches through the unique mapping relationship between the high-frequency attenuation spectrum of the traveling wave front and the distribution network path topology. This achieves unique location results without false point interference. Furthermore, through dynamic topology mapping and a branch fingerprint database, it can automatically adapt to changes in network operation modes. It exhibits strong robustness and practical value for weak fault signals such as high-resistance grounding and complex network structures, significantly improving the reliability of distribution network fault location and operation and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection and fault detection technology, and in particular to a method and system for detecting and locating insulation faults in distribution networks based on multi-source transient data. Background Technology

[0002] Traveling wave (TW) location is an important technology for fault location in distribution networks. It calculates the fault distance by measuring the time difference of the arrival of the high-frequency transient traveling wave signal generated by the fault at different monitoring points. However, in the complex tree-like distribution network structure, the traveling wave undergoes complex reflection and refraction phenomena at impedance discontinuities such as branch points and ends, which may cause the monitoring device to capture reflected waves from non-faulty paths. When the arrival time of the reflected wave is confused with the arrival time of the direct wave from the actual fault point, the time difference-based calculation will produce erroneous distance information, thus creating "false fault points" at non-faulty locations. This problem has been a long-standing bottleneck restricting the practical application of TW location technology.

[0003] Existing technologies primarily focus on improving wavefront detection accuracy, optimizing time synchronization, or fusion of multi-terminal data. However, none of these approaches depart from the traditional paradigm of measuring time, calculating distance, and mapping position, and are essentially still limited by interference from reflected waves. Some methods employing artificial intelligence attempt to bypass precise modeling, but their "black box" nature leads to poor interpretability and difficulty in adapting to topological changes. Therefore, there is an urgent need for a novel positioning principle and method that can fundamentally resist interference from traveling wave reflections. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing technology relies on the arrival time of traveling waves and is susceptible to interference from reflected waves, which can lead to false fault points. To address this, we propose a method and system for detecting and locating insulation faults in distribution networks based on multi-source transient data.

[0005] To achieve the above objectives, this application adopts the following technical solution: a method for detecting and locating insulation faults in a distribution network based on multi-source transient data, comprising the following steps:

[0006] S1. Synchronously collect transient traveling wave signals generated by faults at monitoring points in the distribution network, and capture the initial traveling wave front data of the fault arriving at each monitoring point;

[0007] S2. Perform time-frequency analysis on the captured initial traveling wave wavefront data, extract the energy values ​​of the wavefront signal in at least two preset characteristic high-frequency bands, and the energy value in a reference low-frequency band; calculate the attenuation coefficient of the energy in each characteristic high-frequency band relative to the energy in the reference low-frequency band, and construct a high-frequency attenuation spectrum feature vector characterizing the characteristics of the traveling wave propagation path at the monitoring point.

[0008] S3. Perform mode decomposition on the initial traveling wavefront data to obtain linear mode components and zero mode components. Perform step S2 independently on the wavefronts of the linear mode components and zero mode components to obtain the linear mode high-frequency attenuation spectrum feature vector and the zero mode high-frequency attenuation spectrum feature vector, forming a multi-mode verification feature pair.

[0009] S4. Generate a dynamic topology based on the real-time switching status of the distribution network. Based on the dynamic topology, generate a standard branch attenuation fingerprint for each potential fault branch in advance. The standard branch attenuation fingerprint defines the set of theoretical high-frequency attenuation spectrum feature vectors that propagate to each monitoring point when a fault occurs in the corresponding branch.

[0010] S5. Collect the measured multimodal verification feature pairs of all monitoring points to form a multi-source feature set. Calculate the matching degree between the multi-source feature set and the theoretical feature set of each candidate branch in the fingerprint database. Based on the matching degree calculation results, determine the faulty branch.

[0011] Preferably, in step S2, the construction process of the high-frequency attenuation spectrum feature vector is as follows: Let the energy of the reference low-frequency band be... , No. The energy of each characteristic high-frequency band is The corresponding attenuation coefficient The calculation formula is: The high-frequency attenuation spectrum feature vector of this monitoring point is .

[0012] Preferably, in step S5, the matching degree calculation uses a weighted multimodal distance metric to calculate the multi-source feature set. With candidate branches Theoretical fingerprint Overall distance between The formula is: ,in and These are the Mahalanobis distance and Euclidean distance between the eigenvectors of the linear mode and the zero mode and their theoretical values, respectively. These are the weighting coefficients. Match score The calculation formula is .

[0013] Preferably, a joint criterion based on the energy mutation rate of wavelet transform detail coefficients and the instantaneous amplitude of high-frequency components is used to capture the initial traveling wave front, and a time window is set to isolate subsequent reflected waves.

[0014] Preferably, in step S2, the time-frequency analysis employs S-transform or wavelet transform; the characteristic high-frequency band is located in the frequency range of 150kHz to 1MHz.

[0015] Preferably, the standard branch attenuation fingerprint is generated through electromagnetic transient simulation or a machine learning model trained based on historical data and simulation data.

[0016] A distribution network insulation fault detection and location system based on multi-source transient data includes:

[0017] Multiple distributed monitoring terminals are deployed at key nodes of the power distribution network to synchronously acquire transient traveling wave signals and complete initial wavefront acquisition and local feature extraction.

[0018] The master station analysis system is connected to all the monitoring terminals through a communication network. It is used to realize dynamic topology management, construction and updating of branch attenuation fingerprint database, multi-source feature matching calculation, and fault branch identification. The monitoring terminal includes at least a broadband high-speed waveform recording module, an initial wavefront triggering module, and a local feature extraction module. The master station analysis system includes at least a dynamic topology management module, a branch fingerprint database management module, and a fault identification and analysis module.

[0019] Preferably, the local feature extraction module is used to perform high-frequency attenuation spectrum feature vector extraction and multimodal coupling feature generation, and the fault identification and analysis module is used to perform multi-source feature matching calculation and output fault branch location results and confidence levels.

[0020] Preferably, the system further includes a high-speed communication network for reliable transmission of time synchronization signals and data information between the monitoring terminal and the main station analysis system.

[0021] Preferably, the high-speed communication network adopts a hybrid communication architecture of fiber optic private network and 5G wireless network to ensure time synchronization between monitoring terminals and data transmission between monitoring terminals and the main station analysis system.

[0022] The technical effects and advantages of this invention are as follows:

[0023] In this invention, the faulty branch is directly identified by the unique mapping relationship between the high-frequency attenuation spectrum of the traveling wave front and the distribution network path topology, achieving unique location results without false point interference. Furthermore, through dynamic topology mapping and branch fingerprint database, it can automatically adapt to changes in network operation mode. It exhibits strong robustness and practical value for weak fault signals such as high-resistance grounding and complex network structures, significantly improving the reliability of distribution network fault location and operation and maintenance efficiency. Attached Figure Description

[0024] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0025] Figure 1This is a flowchart illustrating the execution process of the method of the present invention. Detailed Implementation

[0026] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0027] Reference Figure 1 As shown, the present invention provides a technical solution: a method for detecting and locating insulation faults in a distribution network based on multi-source transient data, comprising the following steps:

[0028] S1. Simultaneously collect transient traveling wave signals generated by faults at monitoring points in the distribution network, and capture the initial traveling wave front data of the fault arriving at each monitoring point. Deploy broadband monitoring terminals at key nodes in the distribution network, such as outgoing lines and branch points, to simultaneously collect transient signals of three-phase voltage and current generated by faults. Adaptive criteria based on high-frequency energy mutation are used to reliably capture the first arriving initial traveling wave front and extract short time window data to isolate subsequent reflected wave interference.

[0029] Each monitoring terminal's waveform recording module continuously and synchronously acquires three-phase voltage and current signals at a high sampling rate (e.g., 10MHz) to ensure high data accuracy and real-time performance. The acquired data is cyclically stored in a high-speed cache memory to provide sufficient data support for subsequent processing. The built-in high-performance processor performs multi-scale wavelet decomposition on the current signal in real time to extract detailed information from different frequency bands and continuously calculates the detailed signal energy value of specific high-frequency bands (e.g., from hundreds of kHz to several MHz). The system dynamically monitors the changing trend of this energy value. Once a sharp jump of several times or even tens of times is detected in a very short time, an energy mutation flag is immediately generated and recorded as a preliminary trigger condition.

[0030] Meanwhile, the system performs high-speed digital filtering on the original voltage and current signals, extracts ultra-high frequency components above 1MHz in real time, and dynamically tracks their instantaneous amplitude changes. By continuously updating the background noise level, the system can accurately determine whether there are significant anomalies in the current high-frequency components. Only when an energy mutation flag is generated and the instantaneous amplitude of the ultra-high frequency component significantly exceeds the background noise threshold is the system determined to be a valid fault initial traveling wave front arrival event.

[0031] Once both conditions are met, the waveform recording module immediately locks the current buffer area and uses the wavefront start point as the time starting point to extract a short time window of data of a preset length. The setting of the window length follows two key principles: first, it must be long enough to completely cover all the characteristic information of the initial wavefront; second, it must be strictly shorter than the earliest possible arrival time of the first reflected wave of the fault traveling wave, so as to physically avoid interference from subsequent reflected waves or other subsequent disturbances on the initial wavefront data and ensure the accuracy and reliability of the extracted waveform data.

[0032] S2. For the captured initial traveling wave front signal data, first perform high-resolution time-frequency joint analysis to extract the energy distribution value of the wave front signal in multiple preset characteristic high frequency bands, and at the same time obtain the energy reference value in a selected reference low frequency band.

[0033] Subsequently, by calculating the attenuation ratio of each characteristic high-frequency band energy relative to the reference low-frequency band energy, a high-frequency attenuation spectrum feature vector that can effectively characterize the traveling wave propagation path characteristics at the monitoring point is constructed. In the initial wavefront identification stage, a dual criterion based on the energy mutation rate of wavelet transform detail coefficients and the instantaneous amplitude change of high-frequency components is adopted to achieve accurate capture of the wavefront. A reasonable time window width is set to isolate interference from subsequent reflected waves and ensure the purity of the analyzed wavefront data. Time-frequency analysis methods can employ S-transform or wavelet transform, which possess good time-frequency localization characteristics. The defined characteristic high-frequency band is typically located within a wide frequency range of 150kHz to 1MHz. For the extracted initial wavefront data segment, S-transform is used for refinement, generating a time-spectrum diagram that clearly shows the evolution of signal frequency components over time. In this diagram, the time period in which the wavefront signal energy is most concentrated and stable is first identified. Within this time period, the spectral amplitude is integrated segment by segment for a pre-defined reference low-frequency band, such as 50–100kHz, and multiple characteristic high-frequency bands distributed within the 150kHz to 1MHz frequency band, to obtain the energy value corresponding to each frequency band. Finally, the energy values ​​obtained by integrating each characteristic high-frequency band are compared and analyzed with the baseline energy value of the reference low-frequency band.

[0034] Specifically, during signal propagation, the energy value of the reference low-frequency band is extracted as a benchmark and divided by the energy value of each high-frequency band to calculate a series of corresponding attenuation coefficients. The larger these attenuation coefficients are, the more severe the attenuation effect experienced by the corresponding high-frequency components in the signal propagation path, reflecting a stronger absorption, scattering, or reflection of the high-frequency signal by the propagation medium or environment. Subsequently, all the calculated attenuation coefficients are arranged in ascending order according to their corresponding frequency bands, forming an ordered numerical sequence. This sequence constitutes a high-frequency attenuation spectrum feature vector that represents the characteristics of the signal propagation path received at the monitoring point. This feature vector can be used for further analysis of signal propagation characteristics or path identification.

[0035] The process of constructing the high-frequency attenuation spectrum eigenvector is as follows: Let the energy of the reference low-frequency band be... , No. The energy of each characteristic high-frequency band is The corresponding attenuation coefficient The calculation formula is: The high-frequency attenuation spectrum feature vector of this monitoring point is ;

[0036] S3. Modal decomposition is performed on the initial traveling wavefront data. The original signal is decomposed into linear mode components and zero-mode components using mathematical transformation methods. Then, for the wavefront information of the decomposed linear mode components and zero-mode components, the analysis process defined in step S2 is executed independently. Through this process, the high-frequency attenuation spectrum feature vectors corresponding to the linear mode components and the high-frequency attenuation spectrum feature vectors corresponding to the zero-mode components are extracted, and the two are combined to form a multi-mode verification feature pair to enhance the diversity of feature expression and discrimination ability.

[0037] In addition, the Kelenberger transform matrix is ​​used to perform a linear transformation on the acquired three-phase initial traveling wavefront data to achieve mathematical decoupling from the phase domain to the mode domain, thereby separating the independent linear mode component and zero mode component. The linear mode component waveform and zero mode component waveform obtained after decoupling are processed as two independent signal sources. For these two signals, the entire processing flow included in the aforementioned S2 step is completely repeated to obtain their respective high-frequency attenuation spectrum characteristics, providing a more sufficient data foundation for subsequent analysis.

[0038] That is, time-frequency analysis, frequency band energy integration, and attenuation coefficient calculation are performed separately.

[0039] After the parallel processing described above, the system ultimately obtains two sets of highly representative high-frequency attenuation spectrum feature vectors: one set originates from the in-depth analysis and extraction of the linear mode components, while the other set comes from the precise analysis and conversion of the zero-mode components. Although the physical paths traversed by the linear and zero modes during propagation in the power grid are essentially the same, their fundamental differences in electromagnetic propagation mechanisms lead to significant inherent differences in their attenuation characteristics and dispersion behavior. These differences are not only reflected in the speed and frequency dependence of energy attenuation but also in phase propagation and waveform distortion. Therefore, the two sets of vectors obtained exhibit a systematic, mutually referential, and verifiable intrinsic correlation in terms of numerical distribution, trends, and structural composition. Outputting these two sets of vectors as a complementary feature pair not only enhances the expressive dimension of the features but also constitutes a multimodal verification feature pair with a self-verification mechanism, significantly improving the robustness and reliability of state identification.

[0040] S4. A dynamic topology is generated based on the real-time switching status of the distribution network. Based on this dynamic topology, the system pre-generates a standard branch attenuation fingerprint for each potential faulty branch. The standard branch attenuation fingerprint precisely defines the set of theoretical high-frequency attenuation spectrum feature vectors propagating to each monitoring point when a fault occurs in the corresponding branch. Its generation methods mainly include two types: one is to calculate it through high-precision electromagnetic transient simulation, and the other is to predict and generate it using a machine learning model trained with historical fault data and a large amount of simulation data.

[0041] The master station system continuously monitors the switch position signals transmitted by the distribution automation system with high precision in real time. Once the system detects any change in the network topology, it immediately activates a response mechanism. Based on the latest switch position information, it uses advanced graph theory analysis methods to recalculate the network topology and quickly generate a dynamic network topology diagram that accurately reflects the current actual connection state of the power grid. On this basis, for each line branch in the new topology that may experience a fault, the system constructs a corresponding refined numerical model in professional electromagnetic transient simulation software. Multiple key locations are pre-set on this branch, and various fault types, including short circuits, open circuits, and grounding faults, are configured to highly simulate the traveling wave generation mechanism and propagation characteristics under real fault scenarios. During the fault simulation process, the system collects theoretical traveling wave data at virtual observation locations that correspond one-to-one with the actual physical monitoring points. Based on this data, it calculates the set of theoretical multimodal traveling wave feature vectors that can be obtained from various monitoring points distributed throughout the power grid when a fault occurs on that branch.

[0042] As an alternative or supplementary implementation, the system can pre-train a machine learning model with good generalization ability using massive amounts of historical real fault data or high-fidelity simulation datasets. This model can learn the electrical characteristic patterns under different power grid topologies, thereby establishing a mapping relationship from topology information to standard feature vectors. When the system detects a change in the power grid topology, it only needs to input the new topology information and the electrical parameters of the relevant branches into the pre-trained model. The model can then quickly infer based on existing knowledge and directly output the standard feature vector set of the corresponding branch. This method not only significantly improves the efficiency of feature calculation and greatly shortens the processing time, but also effectively reduces the dependence on high-cost simulation resources, providing a feasible technical path for real-time or near-real-time topology updates.

[0043] The system will bind the set of theoretical feature vectors obtained through any of the above methods, whether based on traditional calculation or machine learning inference, with the unique identification number of the branch and the topology snapshot identifier that triggered the calculation, and finally form a complete standard branch attenuation fingerprint record. This record will be persistently stored in the core database of the main station. Through this systematic management mechanism, the system can always ensure that the fingerprint database is strictly consistent with the actual topology of the current power grid, thereby providing a stable and reliable data foundation for subsequent fault diagnosis, intelligent analysis and precise location.

[0044] S5. Collect the measured multimodal verification feature pairs from all monitoring points to form a multi-source feature set. Calculate the matching degree between the multi-source feature set and the theoretical feature sets of each candidate branch in the fingerprint database. Based on the matching degree calculation results, determine the faulty branch. The matching degree calculation uses a weighted multimodal distance metric to calculate the multi-source feature set. With candidate branches Theoretical fingerprint Overall distance between The formula is: ,in and These are the Mahalanobis distance and Euclidean distance between the eigenvectors of the linear mode and the zero mode and their theoretical values, respectively. These are the weighting coefficients. Match score The calculation formula is .

[0045] After a fault occurs, the main station system receives measured multimodal verification feature pairs uploaded from all monitoring terminals, forming a global multi-source feature dataset. The localization engine retrieves fingerprints of all branches that match the current topology from the fingerprint database. Then, iteratively, the measured multi-source feature dataset is compared one by one with the theoretical feature vector set of each candidate branch. The comparison process uses a weighted distance metric algorithm, which calculates the overall distance between the measured data and the theoretical data in the online modal feature space and the zero modal feature space, respectively.

[0046] Subsequently, the distances between the two spaces are weighted and fused according to preset weights to obtain a comprehensive distance representing the difference between the candidate branch and the measured data. Finally, this comprehensive distance is converted into an intuitive matching score; the higher the score, the better the match. After comparing all branches, the system selects the branch with the highest matching score as the final fault location result. At the same time, the system evaluates the difference between the highest and second-highest scores, as well as the absolute score, to generate a location reliability index, which is pushed to the operators along with the fault branch information.

[0047] Example 1: A simple 10kV radial feeder starts from the substation busbar, passes through tower nodes, branch points B1 and B2, and finally connects to the end load. Three broadband monitoring terminals are deployed on this line: IMU1 at the busbar, IMU2 at branch point B1, and IMU3 at branch point B2. If a phase A metallic ground fault occurs on the last span L3 downstream of branch point B2 at the end of the line, the traveling wave generated by the fault first reaches monitoring terminal IMU3, i.e., branch point B2, and then continues to propagate towards the end of the line, undergoing total reflection at the insulated end. This reflected wave returns to branch point B2 and continues to propagate upstream, possibly being captured again by monitoring terminals IMU2 and IMU1. Traditional time-of-arrival methods are prone to misinterpreting the wavefront reflected from the end as a new fault wavefront from an upstream location, thus calculating a false fault point on the non-faulty L1 or L2 segment, leading to location failure.

[0048] After the fault occurred, monitoring terminals IMU1, IMU2, and IMU3 were synchronously triggered, each capturing the first arriving initial traveling wavefront. Each terminal independently performed time-frequency analysis and mode decomposition, extracting the high-frequency attenuation spectrum feature vector. For monitoring terminal IMU3, it captured the wavefront propagating directly from the fault point, which has a short path and relatively small high-frequency attenuation. The calculated attenuation coefficient vector... The values ​​are relatively small. For monitoring terminals IMU1 and IMU2, the wavefronts they capture continue to travel a round-trip reflection path from the end of branch point B2 to branch point B2 after reaching branch point B2 (i.e., monitoring terminal IMU3). Therefore, the actual propagation path is longer and involves more nodes. Although the arrival time may be earlier, its high-frequency attenuation spectrum eigenvector... and The attenuation coefficient in the path will be significantly greater than the theoretical value under the direct path.

[0049] The main station system generates a topology based on the current operating mode and retrieves the standard fingerprints of each branch under that topology from the fingerprint database, such as L1, L2, and L3. The system then uses the measured multi-source feature set. Match the fingerprints of each branch.

[0050] The matching calculation results show that only when the fault is assumed to occur in the L3 branch, the theoretical characteristics of each monitoring point predicted by the fingerprint, the feature value of monitoring terminal IMU3 is small, while the feature values ​​of monitoring terminal IMU1 and IMU2 are large due to the long path, which highly matches the measured feature pattern. The matching degree score is much higher than that of other branches, such as L1 or L2. Based on this, the fault location result is uniquely determined to be located at the end of the L3 branch.

[0051] By analyzing the physical attenuation characteristics of the wavefront rather than the arrival time, the terminal reflected wave is identified as a propagation result with a longer path, thus clearly distinguishing it from the characteristics of the real fault path and fundamentally eliminating the false point problem caused by simple reflection.

[0052] Example 2: A 10kV tree-structured distribution network, comprising multi-level T-connected branches and parallel branches. At key nodes of the network, five monitoring terminals are deployed, namely IMU1 to IMU5, including the outgoing line, primary branch point B1, secondary branch point B2, tertiary branch point B3, and terminal node B4.

[0053] A single-phase ground fault occurs on a relatively concealed branch Lx, originating from the third-level branch point B3. Traditional traveling wave location methods encounter the following problems: The network topology is complex, and the traveling wave undergoes multiple reflections at various branch points, resulting in a mixed wave sequence received by each monitoring point, containing numerous reflected and refracted waves from different paths. Traditional methods struggle to accurately identify the first wavefront corresponding to the true fault point from this complex wave sequence, or may fail to determine the fault location due to incorrect wavefront identification, or generate multiple difficult-to-distinguish false fault points, leading to low reliability of the location results.

[0054] In the application process of this invention, each monitoring terminal uses a joint criterion based on high-frequency energy mutation to reliably lock the first arriving initial wavefront with the strongest energy in a mixed waveform. This wavefront is usually the shortest path wave from the fault point to the IMU, and short time window data is extracted, effectively avoiding interference from subsequent complex reflected waves.

[0055] IMU1 to IMU5 extract the high-frequency attenuation spectrum feature vectors of their respective wavefronts. These feature vectors together constitute a multi-view feature set describing the topological connection relationship between the fault point and the monitoring points of the entire network. Although IMU4 and IMU5 both pass through the third-level branch point B3, their subsequent paths to the fault point are different, and their high-frequency attenuation spectrum feature vectors will have subtle but distinguishable differences.

[0056] The main station system possesses a complete branch attenuation fingerprint database for this complex network. The matching algorithm does not need to analyze the complex traveling wave reflection path, but instead treats the measured features of the five IMUs as a whole pattern and performs a global comparison with each candidate branch in the fingerprint database, including the overall theoretical pattern of Lx.

[0057] Calculations revealed that only the multi-point theoretical feature pattern predicted by the standard fingerprint of branch Lx achieved optimal matching with the overall measured pattern of the five IMUs. The fingerprints of any other branch could not simultaneously and coherently interpret the measured features of all five monitoring points. Based on this, the system outputs a uniquely identified faulty branch Lx.

[0058] By constructing a global feature pattern using data from multiple monitoring points and performing overall matching with a pre-built fingerprint database, the method achieves unique identification of faulty branches in complex topologies, avoiding misjudgments caused by complex wave sequences and demonstrating its strong adaptability to complex distribution networks.

[0059] Example 3: 10kV distribution network, with the same topology as Example 1 and Example 2. Due to reasons such as trees touching the line, an insulation fault occurs that is grounded through a transition resistor. The grounding resistance is high, such as several thousand ohms. When a high-resistance grounding fault occurs, the initial traveling wave amplitude generated by the high-resistance grounding fault is weak, and the wavefront steepness is reduced. Traditional methods rely on the accurate detection of the wavefront arrival time. When the signal is weak, wavefront identification is difficult, and it is easy to miss the detection or produce a large error, resulting in positioning failure or a serious decrease in accuracy.

[0060] Based on this, the robustness of feature extraction to weak signals is demonstrated. Even though the amplitude of the fault traveling wave is small, the relative relationships of different frequency components in its waveform, i.e., the spectral morphology, still retain the characteristic of being related to the propagation path. The high-frequency attenuation spectrum feature vector extracted in this invention... It is a ratio relationship, mainly dependent on the spectral shape rather than the absolute amplitude. Therefore, even if the signal is weak, as long as it can be reliably triggered and captured, the attenuation ratio of its high-frequency components relative to the low-frequency components remains stable and can be effectively extracted. The standard fingerprints in the fingerprint database are also theoretically generated based on this physical ratio relationship. During the matching calculation, the algorithm compares the relative positions of the measured and theoretical feature vectors in space. It is not sensitive to the overall scaling of the vector corresponding to changes in signal amplitude, but is sensitive to the spectral shape corresponding to the ratio of each component of the vector. For this high-impedance fault, each monitoring terminal can still extract effective high-frequency attenuation spectrum feature vectors. After the main station performs the matching calculation, it can clearly find a candidate branch with a matching score significantly higher than other branches. Although the absolute matching score may be slightly lower than that of the low-impedance fault due to signal noise, the difference between its highest and second-highest relative scores is still significant, which is sufficient to support reliable location judgment. By utilizing the relative spectral characteristics of the traveling wave rather than the absolute amplitude or steep time point, the detection and location capabilities of weak fault signals such as high-impedance grounding are significantly improved, expanding the applicability of the traveling wave location method and improving its practicality.

[0061] A distribution network insulation fault detection and location system based on multi-source transient data includes:

[0062] Multiple distributed monitoring terminals are deployed at key nodes of the power distribution network to synchronously acquire transient traveling wave signals and complete initial wavefront acquisition and local feature extraction.

[0063] The GPS synchronization clock module within the terminal receives satellite timing signals, generating high-precision second pulses and absolute time information. The high-speed waveform recording module uses this synchronization signal as a reference to drive a multi-channel ADC to sample and digitize the three-phase voltage and current in strict synchronization. The initial wavefront trigger module analyzes the sampled data stream in real time, using preset joint criteria for monitoring. Once triggered, the loop buffer is immediately frozen, and the local feature extraction module is activated. The feature extraction module performs modal decomposition, time-frequency analysis, and high-frequency attenuation spectrum feature vector calculation according to a predetermined process, ultimately packaging and generating a multi-modal verification feature pair data packet. After processing, the terminal uploads the feature data packet along with a precise fault absolute time stamp to the main station analysis system via the communication network.

[0064] The main station analysis system connects to all monitoring terminals via a communication network to achieve dynamic topology management, construction and updating of the branch attenuation fingerprint database, multi-source feature matching calculation, and fault branch identification. The monitoring terminal includes at least a broadband high-speed waveform recording module, an initial wavefront triggering module, and a local feature extraction module. The main station analysis system includes at least a dynamic topology management module, a branch fingerprint database management module, and a fault identification and analysis module.

[0065] The dynamic topology management module, acting as a digital image maintenance unit for the power grid, parses SCADA communication messages in real time. Any switch change triggers an online update of its internal topology connection matrix and generates a new topology version number. The branch fingerprint database management module monitors topology version changes. When a version is updated, it first queries the existing fingerprint database. If the fingerprint for a branch under the new topology has not yet been generated, it automatically initiates a fingerprint construction task, either by calling a simulation software interface for simulation calculation or by calling a trained AI model for inference, and then stores the new fingerprint in association with the topology version number. The fault identification and analysis module is the location decision center. After receiving data reported by various terminals, it first performs time-stamp alignment and validity verification. Subsequently, it calls a matching algorithm to perform a full comparison, scoring, and sorting of the measured multi-source data with the fingerprint database under the current topology version. Finally, it not only outputs the most likely faulty branch but also calculates and adds a confidence assessment based on the score distribution, forming a complete location report.

[0066] The local feature extraction module is primarily used to extract high-frequency attenuation spectrum feature vectors from the power system and further generate multi-modal coupled features. This process effectively captures subtle abnormal signals during equipment operation. The fault identification and analysis module is responsible for matching and calculating multi-source features. Through comprehensive comparison and analysis, it accurately outputs the location results of the faulty branch and its corresponding confidence level, thereby improving the accuracy and reliability of fault diagnosis. Furthermore, the entire system integrates a high-speed communication network. This network adopts a hybrid communication architecture combining a dedicated fiber optic network and a 5G wireless network. This ensures a high degree of consistency in time synchronization signals between the monitoring terminal and the main station analysis system, while also guaranteeing the real-time performance and stability of various data transmissions, thus providing a solid communication foundation for the overall system performance.

[0067] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for detecting and locating insulation faults in a distribution network based on multi-source transient data, characterized in that, Includes the following steps: S1. Synchronously collect transient traveling wave signals generated by faults at monitoring points in the distribution network, and capture the initial traveling wave front data of the fault arriving at each monitoring point; S2. Perform time-frequency analysis on the captured initial traveling wave wavefront data, extract the energy values ​​of the wavefront signal in at least two preset characteristic high-frequency bands, and the energy value in a reference low-frequency band; calculate the attenuation coefficient of the energy in each characteristic high-frequency band relative to the energy in the reference low-frequency band, and construct a high-frequency attenuation spectrum feature vector characterizing the characteristics of the traveling wave propagation path at the monitoring point. S3. Perform mode decomposition on the initial traveling wavefront data to obtain linear mode components and zero mode components. Perform step S2 independently on the wavefronts of the linear mode components and zero mode components to obtain the linear mode high-frequency attenuation spectrum feature vector and the zero mode high-frequency attenuation spectrum feature vector, forming a multi-mode verification feature pair. S4. Generate a dynamic topology based on the real-time switching status of the distribution network. Based on the dynamic topology, generate a standard branch attenuation fingerprint in advance for each potential fault branch. The standard branch attenuation fingerprint defines the set of theoretical high-frequency attenuation spectrum feature vectors that propagate to each monitoring point when a fault occurs in the corresponding branch. S5. Collect the measured multimodal verification feature pairs of all monitoring points to form a multi-source feature set. Calculate the matching degree between the multi-source feature set and the theoretical feature set of each candidate branch in the fingerprint database. Based on the matching degree calculation results, determine the faulty branch.

2. The method for detecting and locating insulation faults in a power distribution network based on multi-source transient data according to claim 1, characterized in that: In step S2, the construction process of the high-frequency attenuation spectrum feature vector is as follows: Let the energy of the reference low-frequency band be... , No. The energy of each characteristic high-frequency band is The corresponding attenuation coefficient The calculation formula is: The high-frequency attenuation spectrum feature vector of this monitoring point is .

3. The method for detecting and locating insulation faults in a power distribution network based on multi-source transient data according to claim 1, characterized in that: In step S5, the matching degree calculation uses a weighted multimodal distance metric to calculate the multi-source feature set. With candidate branches Theoretical fingerprint Overall distance between The formula is: ,in and These are the Mahalanobis distance and Euclidean distance between the eigenvectors of the linear mode and the zero mode and their theoretical values, respectively. These are the weighting coefficients. Match score The calculation formula is .

4. The method for insulating fault detection and location of power distribution network based on multi-source transient data according to claim 1, characterized in that: The initial traveling wave front is captured by a joint criterion based on the energy mutation rate of wavelet transform detail coefficients and the instantaneous amplitude of high-frequency components, and a time window is set to isolate subsequent reflected waves.

5. The method for detecting and locating insulation faults in a distribution network based on multi-source transient data according to claim 1, characterized in that: In step S2, the time-frequency analysis employs S-transform or wavelet transform; the characteristic high-frequency band is located in the frequency range of 150kHz to 1MHz.

6. The method for detecting and locating insulation faults in a distribution network based on multi-source transient data according to claim 1, characterized in that: The standard branch attenuation fingerprint is generated through electromagnetic transient simulation or a machine learning model trained based on historical data and simulation data.

7. A multi-source transient data based distribution network insulation fault detection and location system for implementing the method of any one of claims 1-6, characterized by: include: Multiple distributed monitoring terminals are deployed at key nodes of the power distribution network to synchronously acquire transient traveling wave signals and complete initial wavefront acquisition and local feature extraction. The master station analysis system is connected to all the monitoring terminals through a communication network. It is used to realize dynamic topology management, construction and updating of branch attenuation fingerprint database, multi-source feature matching calculation, and fault branch identification. The monitoring terminal includes at least a broadband high-speed waveform recording module, an initial wavefront triggering module, and a local feature extraction module. The master station analysis system includes at least a dynamic topology management module, a branch fingerprint database management module, and a fault identification and analysis module.

8. The multi-source transient data based distribution network insulation fault detection and location system of claim 7, wherein: The local feature extraction module is used to perform high-frequency attenuation spectrum feature vector extraction and multimodal coupling feature generation. The fault identification and analysis module is used to perform multi-source feature matching calculation and output the fault branch location result and confidence level.

9. The multi-source transient data based distribution network insulation fault detection and location system of claim 7, wherein: The system also includes a high-speed communication network for reliable transmission of time synchronization signals and data information between the monitoring terminal and the main station analysis system.

10. The multi-source transient data based distribution network insulation fault detection and location system of claim 9, wherein: The high-speed communication network adopts a hybrid communication architecture of fiber optic private network and 5G wireless network to ensure time synchronization between monitoring terminals and data transmission between monitoring terminals and the main station analysis system.

Images