A line fault rapid determination method and system based on feature point extraction

Abstract

The application provides a line fault rapid judgment method and system based on feature point extraction, relates to the technical field of power fault diagnosis, and receives a power line pulse signal detected by a Rohn coil and samples through a magneto-electric isolation unit; time synchronization marking is performed through a time service module; sampling data with time marking is stored in a double-queue buffer according to a shunt threshold and is processed in time to obtain feature point data; high-speed and low-speed double-channel feature extraction is performed on the feature point data and weighted fusion is performed to generate fault feature data; and the fault feature data is sent to a remote server for fault positioning analysis.

Description

Technical Field

[0001] This invention relates to power fault diagnosis technology, and more particularly to a method and system for rapid determination of line faults based on feature point extraction. Background Technology

[0002] Rapid detection and location of power line faults are key technologies for ensuring the safe and stable operation of the power grid. Traditional line fault detection methods mainly rely on the action information of protection devices or manual line inspection, which suffers from problems such as response lag and low accuracy. With the expansion and increasing complexity of the power grid, single monitoring points and simple algorithms are insufficient to meet the needs of rapid and accurate location. Especially in complex environments and situations with multiple concurrent faults, conventional methods struggle to effectively distinguish between primary and secondary fault sources and their propagation paths, leading to inaccurate fault judgment and imprecise location. Furthermore, existing technologies typically employ fixed sampling frequencies and uniform processing strategies, lacking a mechanism for prioritizing key characteristic signals, resulting in low processing efficiency when dealing with large amounts of data.

[0003] Current fault feature extraction techniques often employ a single-channel and fixed-weight fusion approach, failing to adequately consider the time-varying characteristics and complex fault features of fault signals. Especially under conditions of dynamic changes in grid load, external environmental interference, and multiple coupled faults, existing feature extraction algorithms are susceptible to interference, leading to misjudgments or missed detections. Furthermore, existing fault location methods primarily rely on simple time difference or amplitude attenuation principles, lacking comprehensive consideration of power line topology and wave propagation speed variations, making it difficult to accurately locate fault points in complex lines. Facing the higher demands of modern smart grids for rapid fault response and precise location, there is an urgent need to develop novel rapid fault determination methods for power lines based on advanced signal processing and intelligent algorithms. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for rapid line fault determination based on feature point extraction, which can solve the problems in existing technologies.

[0005] A first aspect of this invention provides a method for rapid determination of line faults based on feature point extraction, comprising:

[0006] The power line pulse signal detected by the Rohm coil is received by the magneto-electric isolation unit, and the power line pulse signal is sampled by the analog-to-digital converter to obtain sampled data; the sampled data is time-synchronized by the time synchronization module.

[0007] The time-stamped sampled data is stored in a dual-queue buffer according to the signal splitting threshold and then subjected to time-division feature processing to obtain feature point data containing the signal arrival time.

[0008] The feature point data is subjected to high-speed and low-speed dual-channel feature extraction and weighted fusion to generate fault feature point data;

[0009] The fault feature point data is sent to a remote server for fault location analysis.

[0010] In one optional implementation, the step of storing the time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and performing time-division feature processing includes:

[0011] Sampled data with amplitude greater than the signal shunting threshold are stored as feature signals in the first data buffer queue, and sampled data with amplitude less than the signal shunting threshold are stored as non-feature signals in the second data buffer queue.

[0012] The first data buffer queue and the second data buffer queue are processed using a time-sharing multiplexing method, wherein the processing time allocated to the first data buffer queue accounts for a higher proportion than that allocated to the second data buffer queue.

[0013] Calculate the periodic change trend characteristics and high-frequency abrupt change characteristics of the feature signal sequence for N consecutive sampling periods in the first data buffer queue; based on the periodic change trend characteristics and high-frequency abrupt change characteristics, determine the sampling window size, sliding rate and extraction interval of the feature point data;

[0014] When the periodic change trend characteristic exceeds the periodic characteristic judgment threshold, the processor resource ratio of the first data buffer queue is dynamically increased; when the high-frequency mutation characteristic exceeds the mutation characteristic judgment threshold, the sampling frequency of the characteristic signal is adaptively adjusted according to the transient response characteristic.

[0015] When the number of feature signals in the first data buffer queue exceeds the upper limit of the queue capacity, the retention priority of the feature signals is determined based on a comprehensive score of periodic change trend features and high-frequency mutation features.

[0016] In one optional implementation, the step of performing high-speed and low-speed dual-channel feature extraction and weighted fusion on the feature point data includes:

[0017] Instantaneous features of the signal are extracted from the high-speed channel, and statistical features of the signal are extracted from the low-speed channel to obtain initial feature point data; reliability indices of the features of the high-speed and low-speed channels are calculated, and weight coefficients of each channel are set based on the reliability indices; the weight coefficients are updated in real time according to the time-varying characteristics of the signal and the accuracy evaluation of the feature fusion results, and the initial feature point data is weighted and combined to obtain fused feature point data.

[0018] The fused feature point data is blindly separated using the independent component analysis algorithm to decompose the mixed fault signal into multiple independent fault source signals; features are extracted from each of the independent fault source signals to generate independent fault feature point data, and primary and secondary fault sources are identified based on signal energy distribution and time series correlation.

[0019] Construct a fault correlation matrix, analyze the propagation characteristics of the primary fault source and its triggering mechanism for secondary fault sources, and determine the propagation path and scope of influence between faults;

[0020] Based on the propagation path and the scope of influence, a multi-objective optimization algorithm is used to weight and fuse the independent fault feature point data, while simultaneously optimizing the detection accuracy, false alarm rate, and false negative rate to generate the final fault feature point data.

[0021] In one optional implementation, the step of using an independent component analysis algorithm to perform blind source separation on the fused feature point data, decomposing the mixed fault signal into multiple independent fault source signals, includes:

[0022] The fused feature point data is decomposed in the time-frequency domain to obtain the frequency components and phase features of the signal; a signal independence evaluation index is constructed, including mutual information minimization and non-Gaussian degree maximization optimization objectives; an iterative algorithm with an adaptive learning rate is used to separate the mixed fault signals, and the iteration stops when the change in the independence evaluation index between two adjacent iterations is less than a preset change threshold;

[0023] The cross-correlation coefficients between each separated signal component and the original mixed signal are calculated, and duplicate signal components are removed. Cluster analysis is performed on the remaining signal components, and signal components with similar time-frequency characteristics are merged to obtain independent fault source signals.

[0024] In one optional implementation, the steps of constructing a fault correlation matrix and analyzing the propagation characteristics of the main fault source include:

[0025] The arrival time difference and signal attenuation characteristics of the independent fault source signals are calculated to determine the spatial distance relationship between the fault sources; based on the spatial distance relationship and the signal strength change rate, the electrical connection relationship of the fault sources is constructed; based on the electrical connection relationship, the signal amplitude and phase characteristics are analyzed to calculate the coupling degree between the fault sources; the changing trend of the fault source signals is analyzed by using a sliding time window to identify the excitation-response relationship between the fault sources; based on the coupling degree and excitation-response relationship, the propagation path and influence range between faults are determined.

[0026] 6. The method according to claim 3, characterized in that the step of weighted fusion of the independent fault feature point data using a multi-objective optimization algorithm includes:

[0027] The detection accuracy, false alarm rate, and false negative rate are transformed into optimization objective functions; based on the diagnostic results of the independent fault feature point data, the weight coefficients of each feature point data are adjusted to determine the optimization range of the weight coefficients; within the optimization range, the optimal solution set is calculated, and the weight coefficient combination with the highest comprehensive score is selected; the selected weight coefficient combination is verified for fault diagnosis, and the final weight coefficient is determined when the diagnostic accuracy deviation of N consecutive verification rounds is less than the preset deviation threshold.

[0028] In one optional implementation, the step of sending the fault feature point data to a remote server for fault location analysis includes:

[0029] The fault feature point data are sorted and classified according to time sequence and spatial location to establish a fault feature retrieval structure;

[0030] The time difference and amplitude attenuation of fault signals recorded at different monitoring points are calculated based on the retrieval structure. Based on the time difference and amplitude attenuation of the fault signals, combined with the topology information of the power line, the propagation velocity of the fault wave in each line branch is calculated. The estimated value of the spatial coordinates of the fault source is constructed using the relationship between the propagation velocity of the fault wave and the distance between monitoring points. The estimated value of the spatial coordinates of the fault source is corrected through multiple rounds of iterative calculation until the position deviation between two adjacent rounds of calculation is less than a preset position deviation threshold, thus obtaining the precise location of the fault point.

[0031] Secondly, a rapid fault diagnosis system for lines based on feature point extraction is provided, including:

[0032] The first unit is used to receive the power line pulse signal detected by the Rohm coil through the magneto-electric isolation unit, and to sample the power line pulse signal through the analog-to-digital converter to obtain sampled data; and to perform time synchronization marking on the sampled data through the time synchronization module.

[0033] The second unit is used to store the time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and perform time-division feature processing to obtain feature point data containing the signal arrival time.

[0034] The third unit is used to perform high-speed and low-speed dual-channel feature extraction and weighted fusion on the feature point data to generate fault feature point data.

[0035] The fourth unit is used to send the fault feature point data to a remote server for fault location analysis.

[0036] Thirdly, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0037] This invention establishes a dual-queue caching and time-sharing feature processing mechanism, enabling priority processing of key feature signals. It can dynamically adjust processor resource allocation and sampling frequency based on signal characteristics, significantly improving the system's response speed to sudden faults. By using high-speed and low-speed dual-channel feature extraction and adaptive weight fusion, the method effectively balances the advantages of instantaneous and statistical features, improving the accuracy and anti-interference capability of fault feature extraction. In particular, the introduction of independent component analysis (ICM) algorithm for blind source separation solves the problem of complex fault analysis, effectively distinguishing primary and secondary fault sources and clarifying their propagation paths, significantly reducing the misjudgment rate in complex fault scenarios.

[0038] The fault correlation matrix analysis method constructed in this invention can deeply reveal the propagation mechanism and impact range of faults, providing a theoretical basis for fault early warning and preventive maintenance. By simultaneously considering multiple performance indicators such as detection accuracy, false alarm rate, and false negative rate through a multi-objective optimization algorithm, optimal fusion of fault feature data is achieved, significantly improving the overall system performance. In the fault location stage, this invention combines time difference, amplitude attenuation, and line topology information, employing an iterative calculation method to continuously correct the fault location estimate, greatly improving the accuracy of fault location.

[0039] Overall, this invention constructs a complete technical chain from signal acquisition, processing, feature extraction to fault location, providing an innovative solution for rapid identification and accurate location of power line faults, and has significant value for improving the reliability of power grid operation and reducing maintenance costs. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the method for rapid line fault determination based on feature point extraction according to an embodiment of the present invention.

[0041] Figure 2 This is a flowchart of dual-channel processing and fault propagation analysis for feature point data. Detailed Implementation

[0042] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0043] Figure 1 This is a flowchart illustrating the rapid line fault determination method based on feature point extraction of the present invention. Figure 1 As shown, the method includes:

[0044] The power line pulse signal detected by the Rohm coil is received by the magneto-electric isolation unit, and the power line pulse signal is sampled by the analog-to-digital converter to obtain sampled data; the sampled data is time-synchronized by the time synchronization module.

[0045] The time-stamped sampled data is stored in a dual-queue buffer according to the signal splitting threshold and then subjected to time-division feature processing to obtain feature point data containing the signal arrival time.

[0046] The feature point data is subjected to high-speed and low-speed dual-channel feature extraction and weighted fusion to generate fault feature point data;

[0047] The fault feature point data is sent to a remote server for fault location analysis.

[0048] For example, a magneto-electric isolation unit receives the power line pulse signal detected by the ROHM coil, and a digital-to-analog converter (DAC) samples the power line pulse signal to obtain sampled data. The magneto-electric isolation unit uses optocoupler technology to achieve electrical isolation, with a withstand voltage rating of 15kV and a common-mode rejection ratio greater than 80dB, effectively preventing damage to the detection equipment from high-voltage side faults. The ROHM coil is installed around the power line, with an inner diameter of 120mm, 1000 turns, and an inductance coefficient of 5mV / A, enabling it to accurately capture current pulse signals in the line. The DAC uses 16-bit resolution, with a sampling rate configurable from 10kHz to 100kHz and an input range of ±10V. In practical applications, it is typically set to a sampling rate of 50kHz. In a short-circuit fault case of a 10kV distribution line, the pulse signal amplitude detected by the ROHM coil reached 8.6V. The DAC samples this signal at a sampling rate of 50kHz, obtaining 50 data points per millisecond, forming a continuous digital signal sequence.

[0049] The timing module employs a combination of a GPS receiver and a high-precision crystal oscillator, achieving a time accuracy better than 1 microsecond. The GPS receiver outputs one PPS (pulses per second) signal per second as a time reference, while the high-precision crystal oscillator (temperature-compensated crystal oscillator with a stability of 0.1 ppm) handles the timing between two PPS signals. GPS time is correlated with the sampled data, and each data point is timestamped in the format "year-month-day hour:minute:second.millisecond", accurate to 0.001 milliseconds. For example, in the short-circuit fault case mentioned above, the sampling point at the time of the fault occurrence is marked as "2023-06-15 14:28:32.458". Time synchronization is particularly important for multi-point collaborative monitoring. When multiple monitoring points in the distribution network are operating simultaneously, the time difference of fault wave propagation can be calculated by comparing the timestamps at different locations, enabling accurate fault location.

[0050] Time-stamped sampled data is stored in a dual-queue buffer according to the signal shunting threshold and then subjected to time-division feature processing to obtain feature point data containing the signal arrival time. The dual-queue buffer mechanism includes a high-speed buffer queue and a low-speed buffer queue. The high-speed buffer queue has a capacity of 10,000 sample points (equivalent to 200ms of data) and is used to capture transient features; the low-speed buffer queue has a capacity of 500,000 sample points (equivalent to 10s of data) and is used to analyze steady-state features. The shunting threshold is set to three times the standard deviation. When the signal change rate exceeds this threshold, the data enters the high-speed buffer queue for real-time processing; otherwise, it enters the low-speed buffer queue for batch processing. In the short-circuit fault case mentioned above, the signal change rate at the moment of the fault reaches 2.8V / ms, which is much higher than the shunting threshold of 0.5V / ms. Therefore, this data is imported into the high-speed buffer queue. Feature points are extracted from the data in the high-speed buffer queue using a combination of wavelet transform and extremum detection, focusing mainly on signal amplitude abrupt changes, oscillation frequency, and waveform characteristics. Key detected features include: the initial rising edge (time 14:28:32.458, amplitude 0.2V), the first peak (time 14:28:32.462, amplitude 8.6V), and the first valley (time 14:28:32.467, amplitude -5.3V). For data in the low-speed buffer queue, a sliding window method was used to extract statistical features. The window length was 2500 points (50ms), and the step size was 500 points (10ms). The mean, variance, and spectral energy distribution within each window were calculated. The result of time-division feature processing is a set of feature point data containing time, amplitude, and feature attributes. These feature points accurately reflect the time and frequency domain characteristics of the fault signal.

[0051] The high-speed channel focuses on the instantaneous characteristics of the signal, such as rise time, peak amplitude, oscillation frequency, and attenuation coefficient. In the short-circuit fault case mentioned above, the features extracted by the high-speed channel include: a rise time of 4ms (from 0.2V to 8.6V), a peak amplitude of 8.6V, an oscillation frequency of 200Hz (calculated by measuring the time interval between adjacent peaks), and an attenuation coefficient of 0.15 (the signal amplitude attenuates by approximately 15% per cycle). The low-speed channel focuses on the statistical and frequency domain characteristics of the signal, such as mean, standard deviation, spectral components, and harmonic content. In the same fault case, the features extracted by the low-speed channel include: a signal mean of 0.05V before the fault and a signal mean of 0.12V after the fault; a standard deviation of 0.08V before the fault and a standard deviation of 1.85V after the fault; a dominant frequency component of 200Hz with an energy share of 62%; and a second harmonic (400Hz) energy share of 28%. The weighting coefficients are dynamically adjusted based on the reliability indicators of each channel feature. The reliability indicators are comprehensively evaluated through signal-to-noise ratio, feature stability, and historical matching degree. For example, in the initial stage of a fault, the signal-to-noise ratio (SNR) of the high-speed channel features is 35 dB, and the SNR of the low-speed channel features is 18 dB. The high-speed channel is assigned a weight of 0.7, and the low-speed channel a weight of 0.3. In the steady-state phase of the fault, the SNR of the high-speed channel features drops to 15 dB, while the SNR of the low-speed channel features rises to 28 dB, with corresponding weight adjustments of 0.4 and 0.6. The weighted fusion uses a normalized weighted average method, generating fault feature point data that encompasses comprehensive signal features, preserving both transient and steady-state information.

[0052] Data transmission employs Secure Sockets Layer (SSL) protocol to ensure secure transmission, achieving a data compression rate of 60% to reduce bandwidth consumption. Upon receiving the data, the remote server first performs data integrity verification and time synchronization correction to ensure accurate alignment of data from multiple monitoring points. The server uses an improved dual-end traveling wave localization algorithm to analyze the fault location. This algorithm calculates the fault point coordinates based on the time difference and waveform characteristics of the fault wave arriving at different monitoring points, combined with power line topology information. In the aforementioned short-circuit fault case, three monitoring points were deployed, located at the start, middle, and end points of the line, respectively. The arrival times of the fault wave at the three points were 14:28:32.458, 14:28:32.464, and 14:28:32.472, respectively. By calculating the time difference and combining it with the line length parameters (total length 6 km, wave speed 150 km / s), the fault point was located 0.9 km downstream of the line's starting point, with a positioning accuracy of ±50 meters. The server also identifies the fault type. By comparing the fault characteristics with a pre-established fault characteristic database, it determines that the case is a phase-to-phase short circuit fault, caused by insulation aging or external damage. The server generates a fault report, including the fault type, location, time of occurrence, cause, and handling suggestions, and pushes it to maintenance personnel via mobile communication network, greatly shortening fault handling time.

[0053] The reliability of fault diagnosis results is assessed using a confidence score mechanism. The confidence score is calculated based on three factors: feature matching degree, historical similar cases, and multi-point cross-validation. For example, in the short-circuit fault case mentioned above, the feature matching degree is 92% (the degree of matching with the short-circuit fault template), 5 historical similar cases were found, and the consistency of multi-point cross-validation is 95%. The overall calculated confidence score is 0.93, indicating that the diagnosis result is highly reliable. Cases with lower confidence are marked as "pending confirmation," prompting maintenance personnel to conduct further inspection and verification. The system also possesses self-learning capabilities, continuously optimizing the fault feature database and diagnosis algorithm based on feedback from maintenance personnel and actual repair results, thereby improving the accuracy and reliability of fault diagnosis.

[0054] The proposed method for rapid fault determination based on feature point extraction achieves rapid and accurate fault determination of power lines through magnetic and electrical isolation acquisition, high-precision time synchronization, dual-queue time-sharing processing, and dual-channel feature fusion technology. This significantly improves the efficiency and accuracy of power fault diagnosis, reduces power outage time and economic losses, and provides strong technical support for the safe and stable operation of power.

[0055] In one optional implementation, the step of storing the time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and performing time-division feature processing includes:

[0056] Sampled data with amplitude greater than the signal shunting threshold are stored as feature signals in the first data buffer queue, and sampled data with amplitude less than the signal shunting threshold are stored as non-feature signals in the second data buffer queue.

[0057] The first data buffer queue and the second data buffer queue are processed using a time-sharing multiplexing method, wherein the processing time allocated to the first data buffer queue accounts for a higher proportion than that allocated to the second data buffer queue.

[0058] Calculate the periodic change trend characteristics and high-frequency abrupt change characteristics of the feature signal sequence for N consecutive sampling periods in the first data buffer queue; based on the periodic change trend characteristics and high-frequency abrupt change characteristics, determine the sampling window size, sliding rate and extraction interval of the feature point data;

[0059] When the periodic change trend characteristic exceeds the periodic characteristic judgment threshold, the processor resource ratio of the first data buffer queue is dynamically increased; when the high-frequency mutation characteristic exceeds the mutation characteristic judgment threshold, the sampling frequency of the characteristic signal is adaptively adjusted according to the transient response characteristic.

[0060] When the number of feature signals in the first data buffer queue exceeds the upper limit of the queue capacity, the retention priority of the feature signals is determined based on a comprehensive score of periodic change trend features and high-frequency mutation features.

[0061] For example, a signal shunting threshold is set, which can be determined based on statistical analysis of historical fault data. In practical applications, it can be set to 1.5 times the peak value of the pulse signal in a normal power line. For instance, for a typical 10kV distribution line, the peak value of the pulse signal detected during normal operation is approximately 0.8V, so the signal shunting threshold can be set to 1.2V. After receiving time-stamped sampled data, the amplitude of each sample point is compared with this threshold. Sampled data with an amplitude greater than 1.2V is identified as a characteristic signal and stored in the first data buffer queue; sampled data with an amplitude less than or equal to 1.2V is identified as a non-characteristic signal and stored in the second data buffer queue. Both the first and second data buffer queues are implemented using a circular buffer structure, with capacities set to 8MB and 16MB respectively, supporting high-speed data read / write and first-in-first-out operations.

[0062] Two buffer queues are processed using a time-division multiplexing approach, specifically implemented through processor time slice allocation. By default, 75% of the processing time is allocated to the first data buffer queue, and 25% to the second. For example, in an embedded system with a 1.2GHz processor, within a 100ms processing cycle, 75ms are allocated to the first data buffer queue, and 25ms to the second. This allocation ensures priority processing of characteristic signals and improves the response speed to fault signals.

[0063] For the feature signals stored in the first data buffer queue, the signal sequence of N consecutive sampling periods is analyzed to calculate the periodic variation trend characteristics and high-frequency abrupt change characteristics. In practical implementation, the value of N can be set to 10, that is, the feature signals within 10 consecutive sampling periods are analyzed. The periodic variation trend characteristics are quantified by calculating the similarity of the signal waveforms of adjacent sampling periods. In specific implementation, the cross-correlation coefficient of adjacent period signals can be used as the similarity index. When the cross-correlation coefficient is greater than 0.85, it indicates that the signal has strong periodicity; when the cross-correlation coefficient is between 0.6 and 0.85, it indicates that the signal has moderate periodicity; when the cross-correlation coefficient is less than 0.6, it indicates that the signal has weak periodicity. The high-frequency abrupt change characteristics are quantified by calculating the rate of change of the signal in a short period of time. In specific implementation, the maximum value of the amplitude difference between adjacent sampling points can be used as the rate of change index. For example, if the signal amplitude changes from 1.2V to 3.5V in 1ms, the rate of change is 2.3V / ms, indicating that the signal has significant high-frequency abrupt change characteristics.

[0064] When the signal exhibits strong periodicity (cross-correlation coefficient greater than 0.85) and insignificant high-frequency abrupt changes (rate of change less than 1.5V / ms), a larger sampling window (e.g., 200ms), a slower sliding rate (e.g., 50% overlap), and a larger extraction interval (e.g., extracting feature points every 100ms) can be used. When the signal exhibits weak periodicity (cross-correlation coefficient less than 0.6) and significant high-frequency abrupt changes (rate of change greater than 2.5V / ms), a smaller sampling window (e.g., 50ms), a faster sliding rate (e.g., 75% overlap), and a smaller extraction interval (e.g., extracting feature points every 20ms) should be used. This dynamic adjustment mechanism can optimize the feature extraction process based on signal characteristics, improving the accuracy of fault feature capture.

[0065] When a periodic trend is detected exceeding a preset periodicity threshold, the processor resource allocation to the first data buffer queue is dynamically increased. For example, if the cross-correlation coefficient of a signal is greater than 0.9 for five consecutive periods (above the preset periodicity threshold of 0.85), the processing time allocation to the first data buffer queue is increased from 75% to 85%, while the processing time allocation to the second data buffer queue is reduced to 15%. This dynamic resource allocation mechanism allows for the concentration of more computing resources for analysis and processing when strong periodic signals (typically associated with faults) are detected. Similarly, when a high-frequency abrupt change exceeds a transient change threshold, the sampling frequency of the characteristic signal is adaptively adjusted based on the transient response characteristics. For example, if the signal rate of change consistently exceeds 3.0V / ms (above the preset transient change threshold of 2.5V / ms), the sampling frequency is increased from the default 10kHz to 20kHz to capture more transient change details. This adaptive sampling mechanism provides finer temporal resolution when high-speed changing signals occur.

[0066] When the number of characteristic signals in the first data buffer queue exceeds the queue capacity limit (8MB), the retention priority of the characteristic signals is determined based on a comprehensive score of periodic variation trend characteristics and high-frequency abrupt change characteristics. The comprehensive score can be calculated using a weighted summation method, for example: Comprehensive score = 0.6 × periodic characteristic value + 0.4 × high-frequency abrupt change characteristic value. For a typical short-circuit fault signal, its periodic characteristic value is approximately 0.75, and its high-frequency abrupt change characteristic value is approximately 0.85, resulting in a comprehensive score of 0.79; while for a typical load abrupt change signal, its periodic characteristic value is approximately 0.65, and its high-frequency abrupt change characteristic value is approximately 0.55, resulting in a comprehensive score of 0.61. The characteristic signals are sorted according to the comprehensive score, retaining signals with higher scores and discarding signals with lower scores. To prevent information loss due to discarding too many continuous signals, at least one highest-scoring signal point within every 100ms is retained to ensure basic tracking capability of the fault development process.

[0067] The dual-queue buffer time-sharing feature processing method implemented in this invention achieves intelligent processing of key feature signals through signal splitting, priority processing, and dynamic resource allocation mechanisms. It can adaptively adjust sampling parameters and processing strategies, significantly improving the identification ability and processing efficiency of complex fault signals, and providing high-quality feature point data for subsequent fault feature extraction and location analysis.

[0068] In one optional implementation, the step of performing high-speed and low-speed dual-channel feature extraction and weighted fusion on the feature point data includes:

[0069] Instantaneous features of the signal are extracted from the high-speed channel, and statistical features of the signal are extracted from the low-speed channel to obtain initial feature point data; reliability indices of the features of the high-speed and low-speed channels are calculated, and weight coefficients of each channel are set based on the reliability indices; the weight coefficients are updated in real time according to the time-varying characteristics of the signal and the accuracy evaluation of the feature fusion results, and the initial feature point data is weighted and combined to obtain fused feature point data.

[0070] The fused feature point data is blindly separated using the independent component analysis algorithm to decompose the mixed fault signal into multiple independent fault source signals; features are extracted from each of the independent fault source signals to generate independent fault feature point data, and primary and secondary fault sources are identified based on signal energy distribution and time series correlation.

[0071] Construct a fault correlation matrix, analyze the propagation characteristics of the primary fault source and its triggering mechanism for secondary fault sources, and determine the propagation path and scope of influence between faults;

[0072] Based on the propagation path and the scope of influence, a multi-objective optimization algorithm is used to weight and fuse the independent fault feature point data, while simultaneously optimizing the detection accuracy, false alarm rate, and false negative rate to generate the final fault feature point data.

[0073] Combination Figure 2The flowchart illustrates the dual-channel processing of feature point data and fault propagation analysis. This embodiment employs a parallel processing mechanism with high-speed and low-speed channels. The high-speed channel extracts the instantaneous characteristics of the signal, primarily including signal amplitude, rise time, and oscillation frequency. The signal amplitude is obtained by directly measuring the voltage value of the feature point; for example, the amplitude of a feature point in a short-circuit fault is 3.8V. The rise time is calculated by determining the time required for the signal to rise from 10% to 90% of its amplitude; for example, the rise time for the same short-circuit fault signal is 0.75ms. The oscillation frequency is calculated by taking the reciprocal of the time interval between two consecutive peaks in the signal waveform; for example, the oscillation frequency of this short-circuit fault signal is 450Hz. The low-speed channel extracts the statistical characteristics of the signal, primarily including the mean, variance, and spectral distribution. The mean is obtained by calculating the average of all sampling points within a time window, such as a signal mean of 2.1V within a 200ms window. The variance is obtained by calculating the average of the sum of squared deviations of the sampling points from the mean within the same window, such as a signal variance of 0.85 within that window. The spectral distribution is obtained by performing a frequency domain transformation on the signal and statistically analyzing the energy proportions of different frequency bands, such as 35% energy in the 0-200Hz band, 45% in the 200-500Hz band, and 20% in the 500-1000Hz band. Initial feature point data refers to the set of feature data obtained after extracting features from the high-speed and low-speed channels respectively, but before weighted fusion processing. Specifically, it includes the original feature vector or feature matrix composed of instantaneous features extracted from the high-speed channel (such as signal amplitude, rise time, and oscillation frequency) and statistical features extracted from the low-speed channel (such as mean, variance, and spectral distribution). These data are the basic inputs for subsequent reliability assessment, weight calculation, and feature fusion.

[0074] The reliability index of high-speed channel characteristics is comprehensively evaluated through signal-to-noise ratio (SNR) and characteristic stability. SNR is calculated as the ratio of the characteristic signal amplitude to the background noise amplitude, and characteristic stability is calculated using the coefficient of variation (COP) of multiple consecutive characteristic values. For example, for a high-speed channel characteristic experiencing a fault, the measured SNR is 18 dB, the COP is 0.12, and the overall reliability index is 0.85. The reliability index of low-speed channel characteristics is mainly evaluated through statistical significance and sample sufficiency. Statistical significance is calculated by the degree to which the characteristic value deviates from the normal range, and sample sufficiency is calculated by the ratio of the number of sampling points within the time window to the minimum required number of points. For example, for the corresponding low-speed channel characteristic, the statistical significance is 0.78, the sample sufficiency is 0.95, and the overall reliability index is 0.82. Based on these reliability indices, the weight coefficients of each channel are initialized, with the initial weight for the high-speed channel set to 0.55 and the initial weight for the low-speed channel set to 0.45.

[0075] The system monitors the time-varying characteristics of the monitoring signals and the accuracy of the feature fusion results, updating the weighting coefficients in real time. Time-varying characteristics are quantified by calculating the rate of change of feature values ​​within a continuous time window. For example, if the high-speed channel feature change rate is detected at 85% and the low-speed channel feature change rate at 25% in the early stages of a fault, it indicates that the high-speed channel features are more sensitive to fault response; therefore, the weight of the high-speed channel is adjusted to 0.65 and the weight of the low-speed channel to 0.35. In the steady-state phase of the fault, the rate of change of the high-speed channel features decreases to 15%, while the rate of change of the low-speed channel features remains at 20%; therefore, the weight of the high-speed channel is adjusted to 0.45 and the weight of the low-speed channel to 0.55. The accuracy of the fusion results is evaluated by calculating the matching degree with historical typical fault modes. For example, if a certain fusion result has a matching degree of 0.88 with a short-circuit fault mode and a matching degree of 0.23 with a line-breaking fault mode, it indicates a good fusion effect. The initial feature point data is then weighted and combined according to the updated weighting coefficients to obtain the fused feature point data.

[0076] The fused feature point data undergoes centering and whitening preprocessing to eliminate linear correlations between data points. Centering is achieved by subtracting the feature mean; for example, subtracting the mean 3.47V from the original feature values ​​[3.8V, 2.5V, 4.1V] yields the centered feature values ​​[0.33V, -0.97V, 0.63V]. Whitening is achieved by calculating the covariance matrix of the feature data, performing eigenvalue decomposition on the matrix, and then constructing a whitening transformation matrix using the eigenvalues ​​and eigenvectors. The centered data is then multiplied by this matrix to achieve whitening. For example, calculating the eigenvalues ​​[1.85, 0.92, 0.43] and corresponding eigenvectors from the covariance matrix of the centered data, and constructing the whitening matrix yields the whitened data [0.24, -0.71, 0.46].

[0077] Blind source separation is achieved using the FastICA algorithm. This algorithm finds the separation matrix that maximizes the non-Gaussianity of the mixed signals through an iterative approach. Specifically, negative entropy is chosen as the measure of non-Gaussianity, and the separation is achieved through a nonlinear function G(u) = u 3 Alternatively, tanh(u) can be used to approximate the negative entropy. The algorithm starts by randomly initializing the weight vector w and iteratively updates w through fixed points: first, w is calculated... T The expectation of the nonlinear transformation of x (representing the projection of the observed data onto the weight vector direction) is E{xg(w T·x)}, where x is the whitened observed data vector, g is the derivative of G, and then w is updated to the expectation minus the correction term; after the update, w is normalized; the above process is repeated until w converges. In practical applications, the iterative process of multiple weight vectors will run in parallel to extract multiple independent components. For the above case of short circuit and open circuit combined fault, two weight vectors are initialized. After 17 iterations, the convergence value of the first weight vector is [0.82, 0.15, 0.55], and the convergence value of the second weight vector is [-0.23, 0.91, 0.34], which correspond to the feature components of short circuit fault and open circuit fault, respectively.

[0078] The separated independent components are converted into fault source signals in the original signal space through back projection. This step is achieved by calculating the inverse (or pseudo-inverse) matrix of the mixing matrix. For example, in the above case, the estimated mixing matrix is ​​[[0.85,-0.21],[0.12,0.92],[0.48,0.36]]. After calculating its pseudo-inverse matrix, it is multiplied by the independent components to obtain the estimated fault source signals. To evaluate the separation quality, the source signal independence index, including mutual information and cross-correlation coefficient, is calculated. In this case, the cross-correlation coefficient of the two separated fault source signals is 0.08, and the mutual information is 0.12, indicating that the statistical dependence between the signals is very low and the separation effect is good. The reconstruction error is also used as a measure of separation quality. The mean square error between the original mixed signal and the signal reconstructed by the estimated mixing matrix and the source signals is calculated. In this case, the reconstruction error is 0.064, further verifying the effectiveness of the separation.

[0079] Taking the aforementioned combined short-circuit and open-circuit fault as an example, the features extracted from the separated short-circuit fault source signal are: amplitude 4.2V, rise time 0.65ms, oscillation frequency 480Hz, mean 2.3V, and variance 0.92; the features extracted from the separated open-circuit fault source signal are: amplitude 2.8V, rise time 1.25ms, oscillation frequency 120Hz, mean 1.7V, and variance 0.63. The primary and secondary fault sources are identified based on signal energy distribution and time series correlation. Energy distribution is obtained by calculating the integral value of the signal power spectrum, and time series correlation is obtained by calculating the time delay between the fault source signal and the response signal. In this case, the energy of the short-circuit fault source signal is 2.3 times that of the open-circuit fault source signal, and the short-circuit fault source signal appears 150ms earlier than the open-circuit fault source signal. Therefore, the short-circuit fault is identified as the primary fault source, and the open-circuit fault as the secondary fault source.

[0080] Design an n×n fault correlation matrix (where n is the number of fault sources), where the matrix element a_ij represents the correlation strength from fault i to fault j. The correlation strength is calculated by combining three factors: time-series correlation, energy transfer ratio, and spatial proximity. The time-series correlation is achieved through Granger causality tests. A time-series autoregressive model of fault i and fault j is constructed, and the F-statistic is calculated for the predictive ability of fault i on fault j, with and without fault i, and converted into a correlation probability value. For example, the Granger causality test F-value for short-circuit faults on open-circuit faults is 28.7, with a corresponding p-value of 0.0001, indicating a strong correlation. The energy transfer ratio is calculated as the ratio of the energy of fault source j to the energy of fault source i, adjusted for time delay. Spatial proximity is calculated based on the distance between the two fault points in the power grid topology; the closer the distance, the stronger the correlation. In the above case, the element a_12 (short circuit to open circuit) of the fault correlation matrix is ​​0.85, indicating a strong correlation; a_21 (open circuit to short circuit) is 0.12, indicating a weak correlation, which is consistent with the physical process of a short circuit fault leading to an open circuit fault.

[0081] For the correlation matrix, graph theory methods are applied, treating faults as nodes and correlation strength as directed edge weights to construct a fault propagation graph. A depth-first search algorithm is used to find all propagation paths originating from the primary fault source, and the path strength is calculated based on the cumulative product of the edge weights along the path. For example, in a case involving short circuits, open circuits, and ground faults, the propagation path "short circuit → open circuit → ground fault" is identified, with a path strength of 0.85 × 0.67 = 0.57. The fault impact range is also calculated by setting a correlation strength threshold (e.g., 0.3) to determine the set of faults affected by the primary fault source. In this case, the impact range of a short circuit fault includes both open circuit and ground faults, while the impact range of an open circuit fault only includes ground faults, further confirming that the short circuit fault is the root cause.

[0082] Based on the defined propagation path and scope of influence, the detection accuracy (objective f1), false alarm rate (objective f2), and false negative rate (objective f3) are set as optimization objectives. The objective function values ​​are verified and calculated on historical fault datasets. NSGA-II is used to implement multi-objective optimization. This algorithm finds the Pareto optimal solution set through non-dominated sorting and congestion distance calculation mechanisms. In specific implementation, weight variables are set for each fault source feature. For example, for the aforementioned short-circuit and open-circuit combined faults, the weight variables are w1 and w2, with the constraints w1+w2=1, 0≤w1, w2≤1. The initial population size is 50, and each individual contains weight variable values, such as [0.6, 0.4]. Offspring individuals are generated through crossover operations (such as simulated binary crossover) and mutation operations (such as polynomial mutation). After merging the parent and offspring generations, non-dominated sorting and congestion calculation are performed to select the next generation of individuals. In actual operation, the crossover probability is set to 0.9, the mutation probability to 0.1, and the maximum number of iterations to 100.

[0083] After iteration, a Pareto optimal solution set was obtained, containing 15 non-dominated solutions, each corresponding to a set of weight values ​​and its objective function value. For example, solution 1: [w1=0.75, w2=0.25], corresponding to objective values ​​[f1=0.92, f2=0.08, f3=0.06]; solution 2: [w1=0.65, w2=0.35], corresponding to objective values ​​[f1=0.88, f2=0.05, f3=0.09]. The TOPSIS method (Top-Closest Solution Ranking) was used to select the optimal solution from the non-dominated solution set. This method calculates the distance of each solution to the ideal solution and the negative ideal solution, and selects the solution with the highest relative proximity. In this case, the solution [w1=0.7, w2=0.3] has a relative proximity of 0.83, making it the optimal solution, with corresponding objective values ​​[f1=0.90, f2=0.06, f3=0.07]. The weighted combination is used to perform weighted fusion of the fault feature point data to generate the final fault feature point data. During the fusion process, the weights of each feature dimension are not completely identical and are fine-tuned according to the importance of the features. For example, the amplitude feature of short-circuit faults is assigned a weight of 0.75, and the amplitude feature of open-circuit faults is assigned a weight of 0.25; however, the frequency feature is assigned weights of 0.65 and 0.35 respectively to reflect the differences in the importance of different features in different types of faults.

[0084] The high-speed and low-speed dual-channel feature extraction and weighted fusion method of the present invention, combined with independent component analysis and multi-objective optimization technology, can effectively separate composite fault signals, identify primary and secondary fault sources and their propagation relationships, significantly improve the detection accuracy in complex fault scenarios, and reduce false alarm rate and false negative rate, providing strong technical support for the rapid and accurate determination of power line faults.

[0085] In one optional implementation, the step of using an independent component analysis algorithm to perform blind source separation on the fused feature point data, decomposing the mixed fault signal into multiple independent fault source signals, includes:

[0086] The fused feature point data is decomposed in the time-frequency domain to obtain the frequency components and phase features of the signal; a signal independence evaluation index is constructed, including mutual information minimization and non-Gaussian degree maximization optimization objectives; an iterative algorithm with an adaptive learning rate is used to separate the mixed fault signals, and the iteration stops when the change in the independence evaluation index between two adjacent iterations is less than a preset change threshold;

[0087] The cross-correlation coefficients between each separated signal component and the original mixed signal are calculated, and duplicate signal components are removed. Cluster analysis is performed on the remaining signal components, and signal components with similar time-frequency characteristics are merged to obtain independent fault source signals.

[0088] For example, the fused feature point data is decomposed in the time-frequency domain, and the signal is processed using a short-time Fourier transform. A Hamming window is selected as the window function, with a window length of 64 sampling points and a window overlap rate of 50%. For a detected composite fault signal, the time-domain waveform exhibits complex oscillation and attenuation characteristics, with an amplitude range of -4.5V to +5.2V. After the short-time Fourier transform, a time-frequency distribution map is obtained, revealing an energy concentration region in the 100Hz-150Hz band with an amplitude of approximately 2.8V; and another energy concentration region in the 350Hz-420Hz band with an amplitude of approximately 3.5V. These two bands correspond to different types of fault characteristics. Further phase information is extracted, and the phase angle at each frequency point is calculated to obtain the phase spectrum. In the 100Hz-150Hz band, the phase angle exhibits a slow change characteristic, with a change rate of approximately 0.08 radians / Hz; in the 350Hz-420Hz band, the phase angle changes more rapidly, with a change rate of approximately 0.25 radians / Hz. These time-frequency domain features provide important information for subsequent blind source separation.

[0089] Mutual information, a metric for signal independence, measures the statistical dependence between signal components; a smaller value indicates greater independence. Since direct calculation of mutual information is complex, the difference between the sum of joint entropy and marginal entropy is used as an approximation. In practical implementation, the signal is divided into 20 equally probable intervals, and the joint and marginal occurrence frequencies of each interval are statistically analyzed to calculate approximate mutual information values. For the aforementioned composite fault signal, the mutual information value for signals in different time-frequency regions under the initial state is 1.42 bits, indicating a strong statistical dependence between signal components. Non-Gaussianity is quantified using kurtosis or negative entropy; negative entropy is chosen as the metric, and it is approximated using a non-quadratic function. The log-cosine function is used as the non-quadratic function to calculate the approximate negative entropy value for each signal component. For the aforementioned composite fault signal, the negative entropy value for the 100Hz-150Hz band is 1.85, and the negative entropy value for the 350Hz-420Hz band is 2.23, both exhibiting significant non-Gaussianity, making them suitable for separation using independent component analysis algorithms.

[0090] An iterative algorithm with an adaptive learning rate is employed to separate mixed fault signals. This algorithm is based on the FastICA framework but introduces an adaptive learning rate mechanism to improve convergence speed and stability. First, the signal data is preprocessed, including centering and whitening. Centering is achieved by subtracting the signal mean, while whitening is achieved through eigenvalue decomposition and normalization, ensuring that the processed signal components have unit variance and are mutually orthogonal. For the aforementioned composite fault signal, the signal range after centering becomes -3.8V to +5.9V with a mean of 0; after whitening, the variance of all components is 1, and the correlation coefficient between the principal components is close to 0. The separation matrix W is randomly initialized with an initial learning rate of 0.01. In each iteration, the mutual information value and non-Gaussianity value corresponding to the current separation matrix are calculated, and an objective function is constructed based on the weighted sum of these two objectives. The separation matrix is ​​updated using gradient descent, with the gradient direction determined by the partial derivative of the objective function with respect to the separation matrix. Unlike traditional fixed learning rates, the learning rate is dynamically adjusted based on the changes in the objective function over several iterations: when the objective function continues to improve, the learning rate is increased (maximum not exceeding 0.05); when the objective function fluctuates or deteriorates, the learning rate is decreased (minimum not lower than 0.001).

[0091] The iteration termination condition is set as follows: the change in the independence evaluation index between two adjacent iterations is less than a preset change threshold or the maximum number of iterations is reached. The change is obtained by calculating the weighted average difference between the mutual information value and the non-Gaussian degree value of two iterations. The preset change threshold is set to 0.001, and the maximum number of iterations is set to 200. In the above complex fault case, the algorithm converged after 37 iterations, and the final mutual information value dropped to 0.08 bits, indicating that the separated signal components have high statistical independence. A total of 5 signal components were obtained, of which two main components correspond to the 100Hz-150Hz and 350Hz-420Hz frequency bands, respectively, with amplitudes of 2.6V and 3.3V, which are basically consistent with the time-frequency analysis results. The remaining three components have smaller amplitudes, representing noise or minor fault characteristics.

[0092] The cross-correlation coefficients of the separated signal components with the original mixed signal are obtained by calculating the maximum value of the standardized cross-correlation function of the two signals, reflecting the degree of linear correlation between the signals. Cross-correlation coefficients are calculated for each pair of signal components, constructing a cross-correlation matrix. For example, for the five separated components above, the diagonal elements of their cross-correlation matrix are all 1s (autocorrelation), and the off-diagonal elements range from 0.04 to 0.78. A cross-correlation threshold of 0.75 is set; when the cross-correlation coefficients of two components exceed this threshold, they are considered duplicate components. In this case, the cross-correlation coefficient between the third and fifth components is 0.78, exceeding the threshold. The fifth component, with the larger amplitude, is retained, while the third component is discarded. After the discarding operation, the remaining four signal components proceed to the next processing stage.

[0093] Feature vectors are extracted for each signal component, including time-frequency characteristic parameters such as the center frequency, bandwidth, peak frequency, and phase change rate. For example, the feature vector for component 1 is [center frequency = 125Hz, bandwidth = 45Hz, peak frequency = 132Hz, phase change rate = 0.075 radians / Hz]; the feature vector for component 2 is [center frequency = 385Hz, bandwidth = 65Hz, peak frequency = 402Hz, phase change rate = 0.23 radians / Hz]; the feature vector for component 4 is [center frequency = 390Hz, bandwidth = 70Hz, peak frequency = 405Hz, phase change rate = 0.28 radians / Hz]; and the feature vector for component 5 is [center frequency = 118Hz, bandwidth = 40Hz, peak frequency = 125Hz, phase change rate = 0.07 radians / Hz]. K-means clustering is used to cluster these feature vectors, with a cluster size of 2 (based on prior knowledge, it is estimated that there are two main fault types). Before clustering, the feature vectors are normalized to ensure that the weights of each feature parameter are balanced. The clustering results show that the first and fifth components are grouped into one class, and the second and fourth components are grouped into another class.

[0094] Signal components within the same cluster are combined to generate the final independent fault source signals. The combining process uses a weighted average method, with the weights proportional to the signal-to-noise ratio (SNR) of each component. The SNR of each component is calculated as follows: component 1: 15.3 dB; component 2: 17.8 dB; component 4: 16.2 dB; component 5: 14.7 dB. Based on these SNR values, the combining weights for component 1 and component 5 within cluster 1 are calculated to be 0.51 and 0.49, respectively; and the combining weights for component 2 and component 4 within cluster 2 are calculated to be 0.52 and 0.48, respectively. After combining, two independent fault source signals are obtained: the first fault source signal has a center frequency of 122 Hz, corresponding to a ground fault; the second fault source signal has a center frequency of 388 Hz, corresponding to a short-circuit fault. The effectiveness of the separation results was further verified by calculating the mutual information value (0.06 bits) of the merged signal and the reconstruction error (the root mean square error between the original mixed signal and the signal reconstructed by the estimated mixing matrix and the separated source signal, with a value of 0.15V), confirming that the separation quality was good.

[0095] The physical meaning of the isolated independent fault source signals was also interpreted. By comparing with a fault model library, the first fault source signal (122Hz) was identified as conforming to the typical characteristics of a single-phase ground fault, including low-frequency oscillation and slow decay; the second fault source signal (388Hz) conformed to the typical characteristics of a phase-to-phase short-circuit fault, including high-frequency oscillation and rapid decay. This physical interpretation further verified the rationality and accuracy of the blind source separation results.

[0096] The proposed method for blind source separation using independent component analysis effectively separates independent fault sources in mixed fault signals through time-frequency domain decomposition and adaptive learning rate iteration. Then, through cross-correlation analysis and cluster merging, redundant components are eliminated and fault source signals with clear physical meaning are generated, which significantly improves the accuracy and interpretability of composite fault identification and provides a reliable foundation for fault type judgment and fault propagation analysis.

[0097] In one optional implementation, the steps of constructing a fault correlation matrix and analyzing the propagation characteristics of the main fault source include:

[0098] The arrival time difference and signal attenuation characteristics of the independent fault source signals are calculated to determine the spatial distance relationship between the fault sources; based on the spatial distance relationship and the signal strength change rate, the electrical connection relationship of the fault sources is constructed; based on the electrical connection relationship, the signal amplitude and phase characteristics are analyzed to calculate the coupling degree between the fault sources; the changing trend of the fault source signals is analyzed by using a sliding time window to identify the excitation-response relationship between the fault sources; based on the coupling degree and excitation-response relationship, the propagation path and influence range between faults are determined.

[0099] For example, the time difference of arrival (TDOA) is calculated by comparing the start times of different fault source signals. A threshold cross-validation method is used to determine the signal start point, specifically the moment when the signal amplitude first exceeds the mean background noise plus three standard deviations. In a case of a complex fault, three independent fault source signals were detected: a phase A ground fault, a phase B short-circuit fault, and a phase C open-circuit fault. The start time of the phase A ground fault signal is t1 = 10.25 ms, the start time of the phase B short-circuit fault signal is t2 = 15.78 ms, and the start time of the phase C open-circuit fault signal is t3 = 42.36 ms. The calculated TDOAs are t2 - t1 = 5.53 ms, t3 - t2 = 26.58 ms, and t3 - t1 = 32.11 ms, respectively. The signal attenuation characteristics are obtained by measuring the attenuation rate of the signal amplitude over time. An exponential attenuation curve is fitted to each fault source signal, and the attenuation coefficient is extracted. The attenuation coefficients for the three fault source signals are α1 = 0.045 ms. -1 α2 = 0.078 ms -1 α3 = 0.032 ms -1 This indicates that the short-circuit fault signal in phase B decays the fastest, while the open-circuit fault signal in phase C decays the slowest.

[0100] Based on the time difference of arrival and attenuation characteristics, an electromagnetic wave propagation model is used to estimate the spatial distance between fault sources. In power systems, the speed at which fault signals propagate along conductors is approximately 60%-80% of the speed of electromagnetic waves in free space; a propagation speed of 2 × 10⁻⁶ is taken as the mean. 8m / s. Based on the propagation speed and the time difference of arrival, the distance between the fault sources is calculated: d 12 =v×(t2-t1)=2×10 8 ×5.53×10 -3 =1106m, meaning the distance from the A-phase ground fault point to the B-phase short-circuit fault point is approximately 1106m; similarly, d 23 =5316m, d 13 =6422m. The signal attenuation effect on the transmission line was also considered, and corrections were made based on the relationship between the attenuation coefficient and distance. The corrected distance estimate is: d 12 =1050m, d 23 =5280m, d 13 =6330m. These spatial distance relationships provide the basic data for subsequent construction of electrical connection relationships.

[0101] The rate of change of signal strength is obtained by calculating the amplitude ratio of the fault source signal at different measurement points. Multiple measurement points were set up in the power network. In the above case, signal data collected at four key nodes showed the following amplitudes: Phase A ground fault signal amplitudes at the four measurement points were 3.8V, 2.5V, 1.6V, and 0.9V, respectively; Phase B short-circuit fault signal amplitudes were 4.2V, 3.7V, 1.8V, and 1.1V, respectively; and Phase C open-circuit fault signal amplitudes were 2.1V, 1.8V, 1.5V, and 1.3V, respectively. Based on these amplitude data, the signal strength attenuation curve with distance was calculated, and a signal propagation attenuation equation was obtained through fitting. Based on the fitting results, a directed graph was constructed to represent the electrical connection relationships of the fault source. Nodes in the graph represent fault points, and edges represent electrical connections. The weight of the edges is inversely proportional to the distance and directly proportional to the signal transmission efficiency. In this directed graph, the connection weight from the phase A ground fault node to the phase B short-circuit fault node is 0.85, the connection weight from the phase B short-circuit fault node to the phase C open-circuit fault node is 0.72, while the connection weight from the phase A ground fault node to the phase C open-circuit fault node is only 0.25. This indicates that the electrical connections between the phase A ground fault and the phase B short-circuit fault, and between the phase B short-circuit fault and the phase C open-circuit fault, are relatively strong, while the direct electrical connection between the phase A ground fault and the phase C open-circuit fault is relatively weak.

[0102] Based on the electrical connection relationships, the amplitude envelope and phase change curves of each fault source signal are extracted. In the above case, the maximum amplitude of the phase A ground fault signal is 3.8V, and the phase of the main frequency component is 45°; the maximum amplitude of the phase B short-circuit fault signal is 4.2V, and the phase of the main frequency component is 120°; the maximum amplitude of the phase C open-circuit fault signal is 2.1V, and the phase of the main frequency component is 210°. The amplitude ratio and phase difference between the fault source signals are calculated, and the amplitude ratio matrix and phase difference matrix are constructed. For example, the amplitude ratio between the phase A ground fault and the phase B short-circuit fault is 3.8 / 4.2=0.9, and the phase difference is 120°-45°=75°; the amplitude ratio between the phase B short-circuit fault and the phase C open-circuit fault is 4.2 / 2.1=2.0, and the phase difference is 210°-120°=90°. The coupling coefficient is calculated by comprehensively considering the amplitude ratio, phase difference, and the aforementioned electrical connection weights. The coupling coefficient calculation considered the logarithmic transformation of the amplitude ratio and the cosine transformation of the phase difference, ensuring that its value was between 0 and 1. The calculated coupling coefficient matrix shows that the coupling coefficient between phase A ground fault and phase B short circuit fault is 0.72, the coupling coefficient between phase B short circuit fault and phase C open circuit fault is 0.68, and the coupling coefficient between phase A ground fault and phase C open circuit fault is 0.23. This indicates that there is a strong coupling between phase A ground fault and phase B short circuit fault, meaning that one fault induced the other.

[0103] The sliding time window was set to 20ms in length and 5ms in step. Characteristic parameters, including mean, standard deviation, peak value, and energy, were calculated for the fault source signal within each time window. These characteristic parameters were then plotted as trend curves over time to analyze the temporal relationship between the changes in characteristic parameters of different fault source signals. In the above case, it was observed that the energy of the phase A ground fault signal began to increase at t=10.25ms and reached a peak value of 3.8V at t=12.50ms. 2 Subsequently, at t=15.78ms, the energy of the phase B short-circuit fault signal began to increase, reaching a peak of 4.2V at t=17.25ms. 2 Finally, at t=42.36ms, the energy of the C-phase open-circuit fault signal began to increase, reaching a peak of 2.1V at t=45.10ms. 2 The Granger causality test was applied to analyze these time-series data to calculate the causal relationship strength between fault source signals. The test results showed that the causal relationship strength from the A-phase ground fault to the B-phase short-circuit fault was 0.83, the causal relationship strength from the B-phase short-circuit fault to the C-phase open-circuit fault was 0.76, while the causal relationship strength from the A-phase ground fault to the C-phase open-circuit fault was only 0.27. This indicates that the A-phase ground fault strongly triggered the B-phase short-circuit fault, which in turn further triggered the C-phase open-circuit fault, forming a fault chain.

[0104] A comprehensive weight matrix is ​​constructed by weighting the coupling coefficient and causal relationship strength, with a weight ratio of 0.4:0.6. For the above case, the calculated comprehensive weight matrix is ​​as follows: the comprehensive weight from phase A ground fault to phase B short circuit fault is 0.78, the comprehensive weight from phase B short circuit fault to phase C open circuit fault is 0.73, and the comprehensive weight from phase A ground fault to phase C open circuit fault is 0.25. The shortest path algorithm is applied to find the optimal propagation path on the directed graph constructed by the comprehensive weights. The analysis results show that the optimal propagation path is "phase A ground fault → phase B short circuit fault → phase C open circuit fault", with a total path weight of 0.78 × 0.73 = 0.57, significantly higher than the weight of 0.25 for the direct path "phase A ground fault → phase C open circuit fault". This indicates that the optimal propagation order of the fault is: phase A ground fault occurs first, then phase B short circuit fault occurs, and finally phase C open circuit fault occurs. The impact range of each fault was also calculated. By setting a comprehensive weight threshold of 0.3, the set of all fault points reachable from each fault was determined. The calculation results show that the impact range of the phase A ground fault includes the phase B short-circuit fault and the phase C open-circuit fault; the impact range of the phase B short-circuit fault includes the phase C open-circuit fault; and the phase C open-circuit fault does not affect other faults. This further confirms that the phase A ground fault is the root cause fault and has the widest impact range.

[0105] Based on the fault propagation path and impact range, a fault correlation matrix and a fault propagation diagram are generated to visually display the correlation relationships and propagation characteristics between faults. In the fault correlation matrix, rows represent source faults, columns represent target faults, and matrix element values ​​represent the overall correlation strength. In the fault correlation matrix of the above example, the element values ​​are: a 11 =1.0, a 12 =0.78, a 13 =0.25, a 21 =0.12, a 22 =1.0, a 23 =0.73, a 31 =0.05, a 32 =0.08, a 33 =1.0. The fault propagation diagram visually illustrates the path and direction of fault propagation, with the arrow width proportional to the overall weight. Centrality indices for each fault were also calculated, including out-degree centrality and in-degree centrality, to analyze the importance of fault nodes in the propagation network. The results show that the A-phase ground fault has the highest out-degree centrality, indicating it is the primary fault source; the C-phase open-circuit fault has the highest in-degree centrality, indicating it is mainly a result of other faults.

[0106] The fault correlation matrix construction and main fault source propagation characteristic analysis method proposed in this invention can accurately identify the fault propagation path and impact range in complex faults through multi-dimensional signal feature analysis and graph theory model, reveal the causal relationship and propagation mechanism between faults, and provide an effective tool for power fault root cause analysis and preventive maintenance, which greatly improves the accuracy and reliability of complex fault handling.

[0107] In one optional implementation, the step of weighted fusion of the independent fault feature point data using a multi-objective optimization algorithm includes:

[0108] The detection accuracy, false alarm rate, and false negative rate are transformed into optimization objective functions; based on the diagnostic results of the independent fault feature point data, the weight coefficients of each feature point data are adjusted to determine the optimization range of the weight coefficients; within the optimization range, the optimal solution set is calculated, and the weight coefficient combination with the highest comprehensive score is selected; the selected weight coefficient combination is verified for fault diagnosis, and the final weight coefficient is determined when the diagnostic accuracy deviation of N consecutive verification rounds is less than the preset deviation threshold.

[0109] For example, the implementation process of weighted fusion of independent fault feature point data using a multi-objective optimization algorithm involves several key technical aspects. The detection accuracy, false alarm rate, and false negative rate are transformed into optimization objective functions to achieve optimal fusion of fault features. Detection accuracy is defined as the ratio of the number of correctly identified fault samples to the total number of samples. This is achieved by inputting independent fault feature point data into a pre-trained fault classifier to obtain the identification results, which are then compared with the labeled true fault types to calculate the accuracy. For instance, in a power distribution network fault diagnosis, 200 samples containing four fault types—short circuit, open circuit, grounding, and overload—were collected. The initial fusion scheme had a detection accuracy of 0.82, meaning 164 samples were correctly identified. The false alarm rate is defined as the ratio of the number of samples incorrectly diagnosed as a certain type of fault to the total number of samples that do not actually belong to that type of fault. The false alarm rate is calculated separately for each fault type. In the above case, the false alarm rate for short-circuit faults was 0.06, for open-circuit faults 0.08, for ground faults 0.04, and for overload faults 0.07, with an average false alarm rate of 0.0625 for the four fault types. The false alarm rate is defined as the ratio of the number of undetected fault samples to the total number of actual fault samples of that type, and is calculated separately for each fault type. In the above case, the false alarm rate for short-circuit faults was 0.05, for open-circuit faults 0.07, for ground faults 0.03, and for overload faults 0.06, with an average false alarm rate of 0.0525.

[0110] Three objective functions are constructed: f1 is the negative of the detection accuracy (transforming the problem into a minimization problem), f2 is the average false positive rate, and f3 is the average false negative rate. The optimization objective is to simultaneously minimize these three objective functions, i.e., improve detection accuracy and reduce false positive and false negative rates. Since these three objectives conflict—for example, reducing the false negative rate may lead to an increase in the false positive rate—a multi-objective optimization algorithm is needed to find the Pareto optimal solution set. NSGA-II (Non-dominated sorting genetic algorithm II) based on a genetic algorithm is chosen to implement multi-objective optimization. During the optimization process, individuals in the population are sorted and selected using non-dominated sorting and crowding calculations to ensure the solution set is uniformly distributed in the objective space and gradually approaches the Pareto front. The population size is set to 50, the maximum number of iterations to 100, the crossover probability to 0.9, and the mutation probability to 0.1. In each iteration, parent individuals are selected through a binary tournament selection operation, and offspring are generated using simulated binary crossover and polynomial mutation. Then, the parent and offspring populations are merged, and non-dominated sorting and crowding calculations are performed to select high-quality individuals to form the new generation population.

[0111] An adaptive weight adjustment mechanism was designed to dynamically determine the weight coefficient range based on the contribution of each feature point's data to fault identification. First, an initial weight was set for each independent fault feature point. For example, in the above case, the initial weights for the four fault types (short circuit, open circuit, grounding, and overload) were 0.25, 0.25, 0.25, and 0.25, respectively. Sensitivity analysis was used to evaluate the impact of each fault feature on the identification results. This involved adjusting the weight of a single fault feature while keeping other weights constant, and observing changes in accuracy, false alarm rate, and false negative rate. In the above cases, when the weight of the short-circuit fault feature increases by 0.1, the detection accuracy improves by 0.03, the false alarm rate increases by 0.01, and the missed alarm rate decreases by 0.02; when the weight of the open-circuit fault feature increases by 0.1, the detection accuracy improves by 0.02, the false alarm rate decreases by 0.01, and the missed alarm rate increases by 0.01; when the weight of the ground fault feature increases by 0.1, the detection accuracy improves by 0.04, the false alarm rate decreases by 0.02, and the missed alarm rate decreases by 0.01; when the weight of the overload fault feature increases by 0.1, the detection accuracy decreases by 0.01, the false alarm rate increases by 0.02, and the missed alarm rate remains unchanged.

[0112] Based on the sensitivity analysis results, the comprehensive contribution of each feature is calculated as the basis for determining the weight optimization interval. The comprehensive contribution is calculated by subtracting the changes in false alarm rate and missed alarm rate from the changes in accuracy. The weight factors are set according to actual application requirements. In the above case, the weight factors for accuracy, false alarm rate, and missed alarm rate are 0.5, 0.25, and 0.25, respectively. The calculated comprehensive contributions of the four fault features are 0.0175, 0.0175, 0.0325, and -0.0175, respectively. Based on the comprehensive contribution, the optimization intervals for the weight coefficients are determined as follows: the weight interval for short-circuit fault features is [0.2, 0.35], the weight interval for open-circuit fault features is [0.2, 0.35], the weight interval for ground fault features is [0.25, 0.4], and the weight interval for overload fault features is [0.15, 0.25]. A constraint condition that the sum of weights is 1 is also set to ensure that the sum of all feature weights is equal to 1. In this way, during the multi-objective optimization process, any change in any weight will lead to a reverse change in at least one other weight, forming a mutually restrictive relationship.

[0113] Within a defined optimization interval, the optimal solution set is calculated, and the weight coefficient combination with the highest overall score is selected. The NSGA-II algorithm converges after 78 iterations, generating a Pareto optimal solution set containing 25 non-dominated solutions. Each of these 25 solutions represents a set of weight coefficient combinations, and none of them are dominant in the three optimization objectives (accuracy, false positive rate, and false negative rate). The objective function value for each non-dominated solution is calculated. For example, Solution 1: weight combination [0.28, 0.26, 0.35, 0.11], objective function value [-0.91, 0.04, 0.03], corresponding to an accuracy of 0.91, a false positive rate of 0.04, and a false negative rate of 0.03; Solution 2: weight combination [0.32, 0.22, 0.38, 0.08], objective function value [-0.93, 0.05, 0.02], corresponding to an accuracy of 0.93, a false positive rate of 0.05, and a false negative rate of 0.02. To select a final solution from the Pareto optimal set, the TOPSIS (Top-Approximation-Ideal-Solution Ranking) method was used to calculate the overall score of each solution. This method first defines an ideal solution (a virtual solution with the highest accuracy and lowest false positive and false negative rates) and a negative ideal solution (a virtual solution with the lowest accuracy and highest false positive and false negative rates). Then, it calculates the distance of each actual solution to the ideal and negative ideal solutions, and finally calculates the relative proximity as the overall score. In the above case, solution 1 has an overall score of 0.78, and solution 2 has an overall score of 0.82. After calculating the overall scores of all 25 solutions, the solution with the highest score was selected as the final weighted coefficient combination, i.e., [0.30, 0.25, 0.36, 0.09], corresponding to an accuracy of 0.92, a false positive rate of 0.045, and a false negative rate of 0.025.

[0114] The validation process employs cross-validation, randomly dividing the sample data into K parts. Each time, one part is used as the test set, and the remaining K-1 parts are used as the training set. This process is repeated K times to obtain K accuracy results. In the above case, 5-fold cross-validation is used, i.e., K=5, with each fold containing 40 samples. The selected weight coefficient combination [0.30, 0.25, 0.36, 0.09] is used to weight and fuse the independent fault feature point data, which is then input into the fault classifier for diagnosis. The accuracies obtained after 5 rounds of validation are 0.925, 0.900, 0.925, 0.950, and 0.900, respectively, with an average accuracy of 0.920 and a standard deviation of 0.020. With the number of consecutive validation rounds N=5 and the accuracy deviation threshold set to 0.03, the difference between the maximum and minimum accuracy is calculated to be 0.050, which is greater than the preset threshold, indicating that the stability of this weight combination is insufficient.

[0115] Continuing to search for more stable weight combinations within the Pareto optimal solution set, the combination with the second-highest overall score [0.32, 0.24, 0.34, 0.10] was selected for cross-validation. The 5-round accuracies were 0.915, 0.900, 0.925, 0.910, and 0.905, with an average accuracy of 0.911 and a standard deviation of 0.010. The difference between the maximum and minimum accuracy was 0.025, less than the preset threshold of 0.03, indicating that this weight combination has good stability. Further expanding the dataset to include 300 samples and adding composite fault types, another 5-fold cross-validation was performed, yielding accuracies of 0.917, 0.900, 0.908, 0.925, and 0.908, with an average accuracy of 0.912 and a standard deviation of 0.009. The difference between the maximum and minimum accuracy was still less than the preset threshold, confirming the generalization ability and stability of this weight combination. Therefore, the final weighting coefficient combination was determined to be [0.32, 0.24, 0.34, 0.10], which is the optimal scheme for weighted fusion of independent fault feature point data.

[0116] The independent fault feature point data is weighted and fused using the final determined weight coefficient combination to generate the final fault feature point data. Different weight coefficients are applied to each dimension of different types of fault features to form a complete weighted fusion scheme. In the feature space, this weighted fusion operation is equivalent to adjusting the importance of each feature dimension, making the classification boundary clearer and improving the accuracy and robustness of fault identification. The final generated fault feature point data provides a reliable foundation for subsequent fault diagnosis and prediction.

[0117] The multi-objective optimization weighted fusion method proposed in this invention simultaneously optimizes three key indicators: detection accuracy, false alarm rate, and false negative rate. Combined with adaptive weight interval determination and multi-round cross-validation, it improves the stability and generalization ability of the algorithm while ensuring the accuracy of fault diagnosis. It provides an efficient and reliable technical solution for the identification of multiple types of faults in complex power systems, and significantly improves the overall performance and practical value of fault diagnosis.

[0118] In one optional implementation, the step of sending the fault feature point data to a remote server for fault location analysis includes:

[0119] The fault feature point data are sorted and classified according to time sequence and spatial location to establish a fault feature retrieval structure;

[0120] The time difference and amplitude attenuation of fault signals recorded at different monitoring points are calculated based on the retrieval structure. Based on the time difference and amplitude attenuation of the fault signals, combined with the topology information of the power line, the propagation velocity of the fault wave in each line branch is calculated. The estimated value of the spatial coordinates of the fault source is constructed using the relationship between the propagation velocity of the fault wave and the distance between monitoring points. The estimated value of the spatial coordinates of the fault source is corrected through multiple rounds of iterative calculation until the position deviation between two adjacent rounds of calculation is less than a preset position deviation threshold, thus obtaining the precise location of the fault point.

[0121] For example, the process of sending fault feature point data to a remote server for fault location analysis involves several key steps. A secure transmission protocol is used to upload the fault feature point data to the remote server. Data transmission uses 256-bit encryption and is compressed to reduce transmission time and bandwidth consumption. After receiving the data, the remote server verifies its integrity and, upon confirmation, proceeds with the fault location analysis process.

[0122] In the steps of establishing the fault feature retrieval structure, time sorting is based on the timestamps of fault feature points, accurate to the millisecond level. For example, in the above case, the fault occurrence times recorded by the eight monitoring points are 13:45:27.328, 13:45:27.332, 13:45:27.334, 13:45:27.335, 13:45:27.337, 13:45:27.338, 13:45:27.340, and 13:45:27.342, respectively. Spatial sorting is based on the geographical coordinates of the monitoring points and their positions in the power network topology. A power network topology database is maintained, containing the coordinate information and connection relationships of all monitoring points. For example, monitoring point A is located at the main transformer outgoing line of the substation, with coordinates (0,0,0); monitoring point B is located on the first branch line, 1250 meters away from monitoring point A, with coordinates (1250,0,0); monitoring point C is located on the second branch line, 980 meters away from monitoring point A, with coordinates (850,480,0). A two-dimensional index structure is constructed based on time sequence and spatial location. The first dimension is the time index, and the second dimension is the spatial index. A tree structure is used to achieve fast retrieval. In this retrieval structure, the fault feature points are first divided into multiple time windows based on the timestamp, with a window size of 10 milliseconds. Then, a spatial sub-index is built within each time window based on the spatial location of the monitoring point. This two-dimensional index structure supports two types of efficient queries: querying data for all monitoring points given a time range, and querying the time series data for a given monitoring point.

[0123] The fault time difference recorded at different monitoring points is calculated based on the arrival time of the fault wavefront at each monitoring point. A threshold detection method is used to determine the arrival time of the fault wavefront, i.e., the time when the signal amplitude first exceeds a preset threshold. The preset threshold is set as the signal mean plus three standard deviations under normal operating conditions. In the above case, the calculated arrival times of the fault wavefront at the eight monitoring points are 13:45:27.328, 13:45:27.332, 13:45:27.334, 13:45:27.335, 13:45:27.337, 13:45:27.338, 13:45:27.340, and 13:45:27.342. Taking monitoring point A as the reference point, the calculated time differences are 0ms, 4ms, 6ms, 7ms, 9ms, 10ms, 12ms, and 14ms, respectively. Amplitude attenuation calculation is based on the maximum amplitude of the fault signal at each monitoring point. The maximum amplitude of the fault signal recorded at each monitoring point is extracted, and the attenuation ratio relative to the reference point is calculated. In the above case, the maximum amplitudes of the fault signal recorded at the eight monitoring points are 4.8V, 3.6V, 3.2V, 3.0V, 2.7V, 2.5V, 2.2V, and 1.9V, respectively, and the calculated attenuation ratios are 1.00, 0.75, 0.67, 0.63, 0.56, 0.52, 0.46, and 0.40, respectively.

[0124] Based on the fault signal time difference and amplitude attenuation, combined with the power line topology information, the propagation velocity of the fault wave in each line branch is calculated. According to the power line topology, the monitoring points are divided into multiple line branch groups. For example, in the above case, monitoring points A, B, E, and H belong to the main line, monitoring points C and D belong to the first branch, and monitoring points F and G belong to the second branch. For each line branch, two adjacent monitoring points are selected, and the propagation velocity is calculated based on the line distance between them and the fault wave arrival time difference. For example, the line distance between monitoring points A and B is 1250 meters, and the time difference is 4 milliseconds, so the propagation velocity on the main line is calculated to be 312.5 km / s; the line distance between monitoring points A and C is 980 meters, and the time difference is 6 milliseconds, so the propagation velocity on the first branch is calculated to be 163.3 km / s; the line distance between monitoring points E and F is 550 meters, and the time difference is 1 millisecond, so the propagation velocity on the second branch is calculated to be 550 km / s. Considering the differences in propagation speed caused by the material, installation method, and load conditions of different line branches, the propagation speed of each branch is calculated separately to improve positioning accuracy.

[0125] By utilizing the relationship between the propagation velocity of the fault wave and the distance between monitoring points, an estimate of the spatial coordinates of the fault source is constructed. The hyperbolic positioning method is employed, calculating the fault location based on the time difference between the arrival of the fault wave at different monitoring points. In three-dimensional space, each pair of monitoring points forms a hyperboloid, with the fault point located at the intersection of multiple hyperboloids. Due to errors in actual measurements, these hyperboloids will not intersect precisely at a single point; the optimal estimated location is determined using the least squares method. Specifically, an objective function is established, representing the sum of squared errors between the actual measured time difference and the theoretical time difference calculated based on the assumed fault location. This objective function is then minimized using gradient descent. In the above case, monitoring point A is selected as the reference point, and a system of equations is established using the time difference data from the other seven monitoring points. The initial guess of the fault location is (500, 300, 0). After gradient descent calculation, the estimated spatial coordinates of the fault source obtained in the first iteration are (620, 410, 0).

[0126] The estimated spatial coordinates of the fault source are corrected through multiple rounds of iterative calculations until the positional deviation between two adjacent rounds is less than a preset positional deviation threshold, thus obtaining the precise location of the fault point. During the iterative calculation process, after updating the fault point position estimate in each round, the theoretical propagation time and theoretical time difference from each monitoring point to the fault point are recalculated and compared with the actual measured time difference to calculate the residual. The residual is used to adjust the search direction and step size of the next round of iteration. The step size gradually decreases as the number of iterations increases to ensure algorithm convergence. The positional deviation threshold is set to 5 meters, and the maximum number of iterations is 50. In the above case, the estimated spatial coordinates of the fault source obtained in the second round of iteration are (635, 422, 0), with a positional deviation of 17.2 meters from the result of the first round, which is greater than the threshold, so the iteration continues; the result of the third round is (642, 430, 0), with a positional deviation of 10.0 meters; the result of the fourth round is (645, 433, 0), with a positional deviation of 4.2 meters, which is less than the threshold, so the iteration stops. The final fault location is output as (645,433,0), which corresponds to a connection point on the actual power line, 776 meters away from the substation in a straight line.

[0127] The reliability of the location results was assessed, and the location error range and confidence level were calculated. The reliability assessment was based on residual analysis and Monte Carlo simulation. Assuming random errors in the time difference measurement, the probability distribution of the fault location was obtained through multiple simulations, thus estimating the location error range. In the above case, 1000 Monte Carlo simulations were performed, yielding a 95% confidence interval for the fault location as a circular area with a radius of 15 meters, centered at (645, 433, 0). The location results were also corrected based on the power line topology, projecting the fault point onto the nearest actual line. The corrected fault location was (648, 435, 0), corresponding to a joint on the line. This matched the fault location found during subsequent on-site inspection (an aging cable joint), verifying the accuracy of the location results.

[0128] The fault location results are visualized on the power network topology map, generating a fault location report. The report includes the fault occurrence time, fault type, precise fault location, location error range, equipment information near the fault location, and recommended maintenance plan. For example, in the above case, the fault location report shows: the fault occurrence time is 13:45:27.328, the fault type is single-phase ground fault, the fault location is (648,435,0), located at cable joint J12 of the first branch line, the location error range is 15 meters, this joint connects 10kV aluminum core cable and copper core cable, installed 5 years ago, and the recommended maintenance plan is to replace the joint and inspect similar joints in the nearby area. The system also automatically sends the fault location report, including the precise coordinates of the fault point and navigation route, to maintenance personnel for rapid on-site repair.

[0129] The fault feature point data remote analysis and location method proposed in this invention can accurately determine the location of fault points in power lines by constructing a spatiotemporal retrieval structure, calculating the fault wave propagation characteristics and iteratively optimizing the location algorithm. This significantly shortens the fault search time, improves the reliability of power operation and maintenance efficiency, and provides strong technical support for the rapid handling and preventive maintenance of power grid faults.

[0130] Secondly, a rapid fault diagnosis system for lines based on feature point extraction is provided, including:

[0131] The first unit is used to receive the power line pulse signal detected by the Rohm coil through the magneto-electric isolation unit, and to sample the power line pulse signal through the analog-to-digital converter to obtain sampled data; and to perform time synchronization marking on the sampled data through the time synchronization module.

[0132] The second unit is used to store the time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and perform time-division feature processing to obtain feature point data containing the signal arrival time.

[0133] The third unit is used to perform high-speed and low-speed dual-channel feature extraction and weighted fusion on the feature point data to generate fault feature point data.

[0134] The fourth unit is used to send the fault feature point data to a remote server for fault location analysis.

[0135] Thirdly, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

Claims

1. A method for rapid fault determination in line circuits based on feature point extraction, characterized in that, include: The power line pulse signal detected by the Rohm coil is received by the magneto-electric isolation unit, and the power line pulse signal is sampled by the analog-to-digital converter to obtain sampled data. The sampling data is time-synchronized and marked using a time synchronization module; The time-stamped sampled data is stored in a dual-queue buffer according to the signal splitting threshold and then subjected to time-division feature processing to obtain feature point data containing the signal arrival time. The feature point data is subjected to high-speed and low-speed dual-channel feature extraction and weighted fusion to generate fault feature point data. This includes: extracting instantaneous features of the signal in the high-speed channel and statistical features of the signal in the low-speed channel to obtain initial feature point data; calculating reliability indices for the high-speed and low-speed channel features, and setting weight coefficients for each channel based on these reliability indices; updating the weight coefficients in real time according to the time-varying characteristics of the signal and the accuracy evaluation of the feature fusion result, and weighting and combining the initial feature point data to obtain fused feature point data; and performing blind fusion of the fused feature point data using an independent component analysis algorithm. Source separation decomposes the mixed fault signal into multiple independent fault source signals; feature extraction is performed on each independent fault source signal to generate independent fault feature point data, and primary and secondary fault sources are identified based on signal energy distribution and time series correlation; a fault correlation matrix is ​​constructed to analyze the propagation characteristics of the primary fault source and its triggering mechanism for secondary fault sources, and to determine the propagation path and influence range between faults; based on the propagation path and influence range, a multi-objective optimization algorithm is used to perform weighted fusion of the independent fault feature point data, while simultaneously optimizing the detection accuracy, false alarm rate, and false negative rate targets to generate the final fault feature point data; The fault feature point data is sent to a remote server for fault location analysis.

2. The method according to claim 1, characterized in that, The steps of storing time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and performing time-division feature processing include: Sampled data with amplitude greater than the signal shunting threshold are stored as feature signals in the first data buffer queue, and sampled data with amplitude less than the signal shunting threshold are stored as non-feature signals in the second data buffer queue. The first data buffer queue and the second data buffer queue are processed using a time-sharing multiplexing method, wherein the processing time allocated to the first data buffer queue accounts for a higher proportion than that allocated to the second data buffer queue. Calculate the periodic change trend characteristics and high-frequency abrupt change characteristics of the feature signal sequence for N consecutive sampling periods in the first data buffer queue; based on the periodic change trend characteristics and high-frequency abrupt change characteristics, determine the sampling window size, sliding rate and extraction interval of the feature point data; When the periodic change trend characteristic exceeds the periodic characteristic judgment threshold, the processor resource ratio of the first data buffer queue is dynamically increased; when the high-frequency mutation characteristic exceeds the mutation characteristic judgment threshold, the sampling frequency of the characteristic signal is adaptively adjusted according to the transient response characteristic. When the number of feature signals in the first data buffer queue exceeds the upper limit of the queue capacity, the retention priority of the feature signals is determined based on a comprehensive score of periodic change trend features and high-frequency mutation features.

3. The method according to claim 1, characterized in that, The steps of using the Independent Component Analysis (ICA) algorithm to perform blind source separation on the fused feature point data, and decomposing the hybrid fault signal into multiple independent fault source signals, include: The fused feature point data is decomposed in the time-frequency domain to obtain the frequency components and phase features of the signal; a signal independence evaluation index is constructed, including mutual information minimization and non-Gaussian degree maximization optimization objectives; an iterative algorithm with an adaptive learning rate is used to separate the mixed fault signals, and the iteration stops when the change in the independence evaluation index between two adjacent iterations is less than a preset change threshold; The cross-correlation coefficients between each separated signal component and the original mixed signal are calculated, and duplicate signal components are removed. Cluster analysis is performed on the remaining signal components, and signal components with similar time-frequency characteristics are merged to obtain independent fault source signals.

4. The method according to claim 1, characterized in that, The steps for constructing a fault correlation matrix and analyzing the propagation characteristics of the main fault source include: The arrival time difference and signal attenuation characteristics of the independent fault source signals are calculated to determine the spatial distance relationship between the fault sources; based on the spatial distance relationship and the signal strength change rate, the electrical connection relationship of the fault sources is constructed; based on the electrical connection relationship, the signal amplitude and phase characteristics are analyzed to calculate the coupling degree between the fault sources; the changing trend of the fault source signals is analyzed by using a sliding time window to identify the excitation-response relationship between the fault sources; based on the coupling degree and excitation-response relationship, the propagation path and influence range between faults are determined.

5. The method according to claim 1, characterized in that, The steps for weighted fusion of the independent fault feature point data using a multi-objective optimization algorithm include: The detection accuracy, false alarm rate, and false negative rate are transformed into optimization objective functions; based on the diagnostic results of the independent fault feature point data, the weight coefficients of each feature point data are adjusted to determine the optimization range of the weight coefficients; within the optimization range, the optimal solution set is calculated, and the weight coefficient combination with the highest comprehensive score is selected; the selected weight coefficient combination is verified for fault diagnosis, and the final weight coefficient is determined when the diagnostic accuracy deviation of N consecutive verification rounds is less than the preset deviation threshold.

6. The method according to claim 1, characterized in that, The steps of sending the fault feature point data to a remote server for fault location analysis include: The fault feature point data are sorted and classified according to time sequence and spatial location to establish a fault feature retrieval structure; The time difference and amplitude attenuation of fault signals recorded at different monitoring points are calculated based on the retrieval structure. Based on the time difference and amplitude attenuation of the fault signals, combined with the topology information of the power line, the propagation velocity of the fault wave in each line branch is calculated. The estimated value of the spatial coordinates of the fault source is constructed using the relationship between the propagation velocity of the fault wave and the distance between monitoring points. The estimated value of the spatial coordinates of the fault source is corrected through multiple rounds of iterative calculation until the position deviation between two adjacent rounds of calculation is less than a preset position deviation threshold, thus obtaining the precise location of the fault point.

7. A rapid line fault determination system based on feature point extraction, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to receive the power line pulse signal detected by the Rohm coil through the magneto-electric isolation unit, and to sample the power line pulse signal through the analog-to-digital converter to obtain sampled data; The sampling data is time-synchronized and marked using a time synchronization module; The second unit is used to store the time-stamped sampled data into a dual-queue buffer according to the signal splitting threshold and perform time-division feature processing to obtain feature point data containing the signal arrival time. The third unit is used to perform high-speed and low-speed dual-channel feature extraction and weighted fusion on the feature point data to generate fault feature point data. The fourth unit is used to send the fault feature point data to a remote server for fault location analysis.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images