A cable fault locating method and system based on multi-source data space-time correlation

Abstract

The present application relates to the technical field of cable fault diagnosis and positioning, in particular to a cable fault positioning method and system based on multi-source data space-time correlation, comprising: obtaining partial discharge, traveling wave and temperature multi-source monitoring data from a high-voltage cable-overhead line hybrid power transmission network; based on space-time correlation analysis of partial discharge and temperature data, calibrating thermal-electric associated defect points; comparing cable and overhead line traveling waves, identifying and suppressing cross-domain propagation interference; positioning the unmarked traveling wave data to obtain a provisional coordinate, fusing sheath circulating current and other information to determine fault characteristics, and comparing the provisional coordinate with the defect points to generate a matching identifier; inputting the fault characteristics, defect matching identifier and temperature and other multi-dimensional information into a fusion model, calculating the fault mode confidence, and outputting the cross-validation positioning result. The present application realizes high-reliability positioning of hybrid line faults and active early warning of insulation defects through space-time correlation and fusion decision of multi-source monitoring quantities such as traveling wave, partial discharge, circulating current and temperature.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable fault diagnosis and positioning, in particular to a cable fault positioning method and system based on spatiotemporal correlation of multi-source data. BACKGROUND

[0002] In the existing technology of cable fault positioning, the traveling wave method is a mainstream technology. However, when this method is applied to the mixed line of cables and overhead lines commonly seen in urban power grids, the reliability of its positioning faces severe challenges. Due to the significant wave impedance mismatch at the connection point of cables and overhead lines, the fault transient traveling wave will undergo complex reflection and refraction at this point, resulting in serious distortion of the monitoring signal waveform, making it difficult to accurately identify the initial fault wave head, and thus easily causing positioning errors.

[0003] At the same time, industry demand is upgrading from passive fault finding to active fault prediction and health management. The key connection point of the cable terminal in the mixed network itself is a weak insulation and accident-prone part, and early state assessment of it has great value. However, the traditional traveling wave method has difficulty in coping with complex waveform distortion, and is powerless against such slowly changing and precursory insulation degradation defects. Therefore, there is an urgent need in the field for a new technology that can integrate multi-source information for cross-validation to overcome the positioning difficulties in mixed networks and achieve active health assessment.

[0004] To this end, the present application proposes a cable fault positioning method and system based on spatiotemporal correlation of multi-source data. SUMMARY

[0005] The purpose of the present application is to provide a cable fault positioning method and system based on spatiotemporal correlation of multi-source data, which realizes high-reliability positioning of mixed line faults and active early warning of insulation defects by spatiotemporal correlation and fusion decision of multi-source monitoring quantities such as traveling wave, partial discharge, circulating current, and temperature.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] A cable fault positioning method based on spatiotemporal correlation of multi-source data, comprising:

[0008] Obtaining cable monitoring data streams containing partial discharge data, load current data, temperature data, cable traveling wave data, and sheath circulating current data from a high-voltage cable-overhead line hybrid power transmission network, and overhead line monitoring data streams containing overhead line traveling wave data;

[0009] Based on partial discharge data and load current data, the PD source coordinates and the phase-resolved partial discharge (PRPD) pattern are obtained by TDOA algorithm phase fusion positioning; based on the PD source coordinates, the temperature data is analyzed to extract the temperature rise cumulative field; the thermal-electric associated defect points are determined by combining the PRPD pattern and the temperature rise cumulative field;

[0010] The timestamp comparison is performed on the cable traveling wave data and the overhead line traveling wave data to obtain the cross-domain propagation interference identification and generate the suppression alarm;

[0011] For the cable traveling wave data without identification, the provisional coordinates are obtained by the traveling wave positioning algorithm; based on the provisional coordinates and the sheath circulating current instantaneous amplitude obtained at the cable terminal monitoring point, the fault characteristics are determined by the feature classifier; the defect matching identification is generated by comparing the provisional coordinates and the thermal-electric associated defect points;

[0012] The fault characteristics, the defect matching identification, and the instantaneous temperature data at the provisional coordinates are input into the multi-dimensional feature fusion model to calculate the fault mode confidence and output the cross-validation positioning result.

[0013] Preferably, the process of obtaining the PD source coordinates and the phase-resolved partial discharge (PRPD) pattern by TDOA algorithm phase fusion positioning based on partial discharge data and load current data comprises:

[0014] The time difference of arrival of the high-frequency PD pulse signal in the partial discharge data is calculated by the TDOA algorithm to obtain the PD source coordinates; the occurrence phase of the high-frequency PD pulse signal is associated with the power frequency phase of the load current data to generate the PRPD pattern, which is used to represent the discharge amount and discharge frequency distribution of the PD signal within the power frequency cycle.

[0015] Preferably, the process of extracting the temperature rise cumulative field based on the PD source coordinates, analyzing the temperature data, and determining the thermal-electric associated defect points by combining the PRPD pattern and the temperature rise cumulative field comprises:

[0016] The temperature data of the PD source coordinate position is subjected to time series analysis through a space-time weighting algorithm, and after the temperature data is subjected to thermal effect compensation using the load current data to suppress thermal disturbance caused by load current changes, a local abnormal temperature rise trend that is consistent with the PD source coordinate position in space and persists in time is identified, and the temperature rise accumulation field is established; wherein the space-time weighting algorithm is spatially weighted according to the physical distance between the PD source coordinate and the temperature sensor, and the time series data points are time-weighted according to the occurrence time; the phase analysis partial discharge spectrum is input into the convolutional neural network model as a two-dimensional frequency spectrum graph through a convolutional neural network model, and the PD defect type is qualitatively determined; wherein the convolutional neural network model includes a convolutional layer for extracting spectrum features, a pooling layer for dimension reduction, and a fully connected layer for outputting classification results; when the PD defect type is qualitatively determined as internal void discharge and the temperature rise accumulation field exceeds a preset thermal effect threshold, the thermal-electric associated defect point is determined.

[0017] Preferably, the cable traveling wave data and overhead line traveling wave data are subjected to timestamp comparison, a cross-domain propagation interference identifier is obtained, and a suppression alarm is generated, including:

[0018] The first traveling wave waveform feature and the first arrival time of the cable traveling wave data are extracted, the second traveling wave waveform feature and the second arrival time of the overhead line traveling wave data are extracted, time-space feature comparison is performed, the first arrival time and the second arrival time are compared, and the propagation correlation between the first traveling wave waveform feature and the second traveling wave waveform feature is analyzed; when the second arrival time is earlier than the first arrival time, and the propagation correlation between the waveform features exceeds a preset correlation threshold, the cross-domain propagation interference identifier is generated.

[0019] Preferably, the un-identified cable traveling wave data is subjected to a traveling wave positioning algorithm to obtain a provisional coordinate, including:

[0020] The un-identified cable traveling wave data is preprocessed through a wavelet transform algorithm to enhance the rising edge feature of the wave head of the cable traveling wave data and suppress background noise; based on a double-end traveling wave positioning algorithm, the provisional coordinate is calculated by comparing the synchronous absolute time difference of the enhanced wave head rising edge feature arriving at both ends of the cable.

[0021] Preferably, the process of determining the fault feature through the feature classifier includes:

[0022] extracting a waveform morphology feature of the unidentified cable traveling wave data, the waveform morphology feature including a waveform distortion degree and an energy attenuation coefficient of the traveling wave; extracting a sheath loop current instantaneous amplitude corresponding to the unidentified cable traveling wave data in a fault occurrence time at a cable terminal monitoring point, and extracting a high-frequency spectral component to form a loop current feature; constructing a fused fault feature vector, the fault feature vector including the waveform morphology feature and the loop current feature; and inputting the fused fault feature vector into the feature classifier to classify the fault feature.

[0023] Preferably, the inputting the fault feature, the defect matching identifier, and the instantaneous temperature data at the provisional coordinate into the multi-dimensional feature fusion model, calculating a fault mode confidence, and outputting a cross-validation positioning result comprises:

[0024] calculating a physical distance between the provisional coordinate and the location information of the thermal-electric associated defect point to generate a quantitative defect matching identifier; the multi-dimensional feature fusion model is a graph neural network model, the graph neural network model is constructed based on a preset physical topology structure corresponding to a high-voltage cable-overhead line hybrid power transmission network, and a power grid topology graph model reflecting a current power flow distribution is generated by analyzing load current data in the cable monitoring data stream to identify a current load level of each power transmission path in the physical topology structure and giving a running state weight; a high-dimensional diagnostic feature vector including the fault feature, the defect matching identifier, and the instantaneous temperature data at the provisional coordinate is constructed, the high-dimensional diagnostic feature vector is input as a node feature into the power grid topology graph model reflecting the current running state, the node feature is associated to a corresponding node of the provisional coordinate in the topology graph model; the graph neural network model analyzes the association of the node feature in the topology graph model to calculate the fault mode confidence; when the fault mode confidence exceeds a preset high-confidence threshold, the provisional coordinate is output as the cross-validation positioning result.

[0025] Preferably, the calculating the fault mode confidence further comprises:

[0026] a physical constraint term based on a cable transmission line model is introduced, the physical constraint term is used to quantify a consistency degree between the space-time data including a traveling wave arrival time and a PD source coordinate in the high-dimensional diagnostic feature vector and cable physical parameters including a wave speed and a length; the physical constraint term is added to a loss function of the graph neural network model, and the fault mode confidence is corrected based on the physical constraint term by minimizing the loss function.

[0027] A cable fault location system based on multi-source data space-time association, comprising:

[0028] a data acquisition module, configured to acquire cable monitoring data streams containing partial discharge data, load current data, temperature data, cable traveling wave data and sheath circulating current data from the high-voltage cable-overhead line hybrid transmission network, and overhead line monitoring data streams containing overhead line traveling wave data;

[0029] a potential defect determination module, configured to acquire PD source coordinates and phase-resolved partial discharge maps through TDOA algorithm phase fusion positioning based on the partial discharge data and the load current data, analyze the temperature data based on the PD source coordinates, extract a temperature rise cumulative field, and determine a thermal-electric associated defect point by combining the partial discharge maps and the temperature rise cumulative field;

[0030] a network-level identification module, configured to compare time stamps of the cable traveling wave data and the overhead line traveling wave data, acquire cross-domain propagation interference identification, and generate a suppression alarm;

[0031] a fault feature determination module, configured to obtain tentative coordinates through a traveling wave positioning algorithm for cable traveling wave data that has not been identified, determine a fault feature through a feature classifier based on the tentative coordinates and instantaneous amplitudes of sheath circulating currents acquired at cable terminal monitoring points, and generate a defect matching identification by comparing the tentative coordinates and the thermal-electric associated defect point;

[0032] a fusion decision module, configured to input the fault feature, the defect matching identification, and instantaneous temperature data at the tentative coordinates into a multi-dimensional feature fusion model, calculate a fault mode confidence, and output a cross-verification positioning result.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] 1. The present application performs spatio-temporal correlation between PD source coordinates acquired through a TDOA algorithm and an abnormal temperature rise cumulative field at the coordinates position extracted based on load current compensation. This cooperative measurement and determination method of electric-thermal features can accurately identify and lock the insulation weak points (i.e. thermal-electric associated defect points) caused by the accumulation of local thermal effects due to continuous discharge. This improves fault diagnosis from passive response to transient events to active monitoring and evaluation of slowly changing insulation degradation processes, overcoming the inability of traditional traveling wave methods to identify such precursor defects.

[0035] 2. In this specific scenario of a hybrid line, the present application uses synchronous time technology to strictly compare the arrival time sequence and waveform propagation correlation of cable traveling waves and overhead line traveling waves. This method can clearly distinguish between external interference waves from the overhead line side and real fault waves inside the cable, and generate a suppression alarm when external interference is determined. This solves the fundamental problem of complex reflection and refraction signals caused by mismatched wave impedance at the hybrid connection point seriously interfering with the identification of traveling wave heads, ensuring that the signals used for positioning calculation are effective internal fault signals.

[0036] 3、The application comprehensively considers the provisional coordinates, the waveform distortion degree of the traveling wave itself, the high-frequency component of the fault time sheath circulating current, and the physical proximity of the coordinates and the pre-marked thermal-electric associated defect points. This determination method that combines transient waveform characteristics, electromagnetic coupling effect (circulating current), and physical entity defect state (fitting degree) constructs a multi-dimensional fault feature vector, greatly improving the accuracy of fault feature classification.

[0037] 4、The high-dimensional diagnostic feature vector containing fault features, defect fitting identifiers, and instantaneous temperatures at provisional coordinates is input into a graph neural network model constructed according to the physical topology of the power grid and real-time power flow. This model not only analyzes the internal relationship of each feature in the network structure, but also introduces physical parameters such as wave speed and cable length as constraint terms to correct the positioning results. This method combines data-driven pattern recognition with physical laws to ensure that the final fault location results have been thoroughly cross-validated, have high confidence and physical interpretability, and perfectly meet the precise positioning needs of complex working conditions such as hybrid lines. BRIEF DESCRIPTION OF DRAWINGS

[0038] Fig. 1 A flowchart of a cable fault location method based on multi-source data space-time association according to the present application;

[0039] Fig. 2 A fault mode confidence calculation and decision-making process according to an embodiment of the present application;

[0040] Fig. 3 A module diagram of a cable fault location system based on multi-source data space-time association according to the present application. DETAILED DESCRIPTION

[0041] To make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application. Other embodiments obtained by those skilled in the art based on the ideas in the present application without creative labor also fall within the scope of protection of the present application.

[0042] REFERENCE Figs. 1 to 3 The present application provides a cable fault location method and system based on multi-source data space-time association, and the specific technical solutions are as follows.

[0043] REFERENCE Fig. 1 A cable fault location method based on multi-source data space-time association, comprising:

[0044] obtain a cable monitoring data stream containing partial discharge data, load current data, temperature data, cable wave propagation data and sheath loop current data from a high-voltage cable-overhead line hybrid power transmission network, and an overhead line monitoring data stream containing overhead line wave propagation data;

[0045] Based on the partial discharge data and the load current data, the PD source coordinates and the phase analysis partial discharge spectrum are obtained by TDOA algorithm phase fusion positioning; based on the PD source coordinates, the temperature data is analyzed to extract the temperature rise cumulative field; the thermal-electric associated defect points are determined by combining the partial discharge spectrum and the temperature rise cumulative field;

[0046] The timestamp comparison is performed on the cable wave propagation data and the overhead line wave propagation data to obtain the cross-domain propagation interference identification and generate the suppression alarm;

[0047] For the cable wave propagation data that has not been identified, the provisional coordinates are obtained by the wave propagation positioning algorithm; based on the provisional coordinates and the sheath loop current instantaneous amplitude obtained at the cable terminal monitoring point, the fault feature is determined by the feature classifier; the defect fitting identification is generated by comparing the provisional coordinates and the thermal-electric associated defect points;

[0048] The fault feature, the defect fitting identification and the instantaneous temperature data at the provisional coordinates are input into the multi-dimensional feature fusion model to calculate the fault mode confidence and output the cross-validation positioning result.

[0049] Embodiment 1:

[0050] The embodiment of the application provides a cable fault positioning method based on multi-source data space-time association. The method of the embodiment is applied to a high-voltage level (for example, 110kV) urban power grid cable-overhead line hybrid power transmission line which contains both underground cable sections and overhead power transmission sections. A specific implementation process of the method starts from the step of obtaining monitoring data streams from the high-voltage cable-overhead line hybrid power transmission network.

[0051] The data acquisition process is realized by multiple sensors deployed along the high-voltage cable-overhead line hybrid transmission network, all of which collect data that is time-synchronized by a unified, Beidou satellite signal-based time service module, ensuring that each data point is associated with a high-precision absolute timestamp with nanosecond resolution. In the cable line section, data acquisition devices are deployed at the start and end of the cable and at intermediate joints to obtain cable monitoring data streams. The data streams specifically include partial discharge data collected by open high-frequency current transformers (HFCTs) installed on the ground wires at the cable terminals or joints, which monitor a frequency band of 0.5 MHz to 50 MHz; load current data collected by flexible Rogowski coils installed on the three-phase main cable cores, which reflect the 50 Hz power frequency load level; temperature data continuously recorded by contact PT100 temperature sensors fixed to the surfaces of the cable terminals and intermediate joints at a 2-second sampling interval; cable traveling wave data captured by special traveling wave sensors installed on the ground wires at both ends of the cable, which have a monitoring bandwidth of 1 kHz to 2 MHz; and sheath loop current data measured by current transformers installed in the cross-interconnected ground boxes in the cable sheath. At the same time, in the overhead line section, overhead line traveling wave data is collected by installing self-powered traveling wave sensors on key transmission towers, which constitute the overhead line monitoring data stream.

[0052] Further, the process of obtaining the PD source coordinates and the phase-resolved partial discharge map based on the partial discharge data and the load current data through TDOA algorithm phase fusion positioning includes: calculating the time difference of arrival of the high-frequency PD pulse signals in the partial discharge data through the TDOA algorithm to obtain the PD source coordinates; correlating the occurrence phase of the high-frequency PD pulse signals with the power frequency phase of the load current data to generate the phase-resolved partial discharge map, which is used to represent the discharge amount and discharge frequency distribution of the PD signals within a power frequency cycle.

[0053] Specifically, this process is performed after obtaining the partial discharge data containing high-frequency PD pulse signals. Assuming that high-frequency current sensors A and B are respectively arranged at the start and end of the cable line, when a partial discharge event occurs inside the cable, the high-frequency PD pulse signals generated by the event will propagate along the cable medium to both ends. Sensors A and B record the absolute arrival times tA and tB of the pulse signals, respectively, with nanosecond-level synchronization timestamps. The TDOA algorithm calculates the time difference of arrival of the same signal received by the two sensors based on this. The calculation of the PD source coordinates is based on the time difference of arrival, the known propagation speed of electromagnetic waves in the cable medium (for example, 1.7×10 8m / s) and the total length of the cable between the two sensors. In the calculation, the propagation speed is first multiplied by the time difference of arrival to obtain the length difference of the pulse signal propagation path; then, the length difference of the path is subtracted from the total length of the cable, and the result is divided by two, and the obtained value is the position distance of the discharge point from one of the sensors (for example, sensor A), which is established as the PD source coordinate.

[0054] At the same time of obtaining the PD source coordinate, the method performs phase correlation on each positioned PD pulse signal. This process uses the synchronously obtained 50Hz load current data as the phase reference of the power frequency voltage, and maps the absolute occurrence time stamp of each high-frequency PD pulse signal to its phase angle position in the corresponding power frequency cycle (0-360 degrees). By counting thousands of discharge events within a predetermined time (for example, 10 minutes), and correlating the discharge amplitude (unit: pC) of all PD signals with the corresponding phase angle, a two-dimensional phase-resolved partial discharge spectrum (PRPD spectrum) is finally generated. The horizontal axis of the spectrum is the power frequency phase angle, the vertical axis is the discharge amplitude, and the density of data points in the graph directly represents the discharge frequency.

[0055] This process combines the TDOA algorithm and the phase fusion technology, not only realizes the physical space positioning of the potential insulation defect source, but also generates a PRPD spectrum that can reflect the electrical characteristics of the defect, providing accurate location and feature information for subsequent qualitative analysis of defect types.

[0056] Further, the process of analyzing temperature data based on the PD source coordinate to extract a temperature rise accumulation field includes: performing time series analysis on the temperature data at the PD source coordinate position by a spatiotemporal weighting algorithm, identifying a local abnormal temperature rise trend that is consistent with the PD source coordinate position in space and exists continuously in time after compensating the temperature data for thermal effects using the load current data to suppress thermal disturbances caused by changes in load current, and establishing the temperature rise accumulation field; wherein the spatiotemporal weighting algorithm performs spatial weighting according to the physical distance between the PD source coordinate and the temperature sensor, and temporal weighting on time series data points according to occurrence time; inputting the phase-resolved partial discharge spectrum as a two-dimensional spectrum to a convolutional neural network model to qualitatively classify the PD defect type; wherein the convolutional neural network model includes a convolutional layer for extracting spectrum features, a pooling layer for dimension reduction, and a fully connected layer for outputting classification results; when the PD defect type is qualitatively classified as internal void discharge and the temperature rise accumulation field exceeds a preset thermal effect threshold, the thermal-electric concomitant defect point is determined.

[0057] Specifically, the process first performs a thermal effect analysis for the determined PD source coordinate. Since temperature sensors are installed at cable terminations and intermediate joints, this thermal effect analysis method is mainly applicable to scenarios where the PD source coordinate is adjacent to these pre-set sensor installation locations. In this scenario, the method retrieves the time-series temperature data collected by the temperature sensor that is physically closest to the PD source coordinate. To exclude the overall temperature fluctuations caused by line load changes, which are the main thermal disturbance, the method performs thermal effect compensation on the temperature data using the synchronously acquired load current data. This compensation process establishes a baseline relationship model between the load current and the normal temperature rise of the cable. For example, this baseline relationship model can be established by collecting the load current data and the corresponding cable temperature data of the cable in a healthy state (without significant PD activity) for at least one complete operating period (e.g., one week). A multiple linear regression model can be established using the least squares method, where the dependent variable is the cable temperature and the independent variables can include: the square of the load current (to represent the main Joule heat effect), the ambient temperature (if available), and historical temperature data (to represent thermal inertia). When analyzing the temperature data of a specific period, the actual measured temperature is compared with the temperature predicted by the baseline relationship model according to the load current at that time, and the difference is the net temperature rise value that has stripped away the load effect. To extract the temperature rise accumulation field, the present application uses a spatio-temporal correlation integration method: in the spatial dimension, if the PD source coordinate does not completely coincide with the sensor location, a distance inverse weighting method can be used to calculate the interpolated temperature sequence at the coordinate location based on the physical distance between the PD source coordinate and the adjacent multiple (e.g., two) temperature sensors; in the time dimension, the net temperature rise value time series (e.g., data of the past 72 consecutive hours) at the coordinate location is time-weighted integrated. This integration can use an exponentially weighted moving average, i.e., giving higher weights to more recent time points, so as to more sensitively reflect abnormal temperature rise trends. By performing this spatio-temporal correlation integration on the net temperature rise value of the time series, a quantified temperature rise accumulation field with the unit of "degrees Celsius·hours" is obtained.

[0058] To cover scenarios where the PD source coordinate is far away from the sensor installation location, the method also includes a supplementary calculation method based on a heat conduction model. When the physical distance between the PD source coordinate and the nearest temperature sensor exceeds a pre-set interpolation effective distance (e.g., 5 meters), the system will enable the heat conduction finite element model of the cable body. This model takes the measured data of all temperature sensors as boundary conditions, the Joule heat calculated from the load current data as the global heat source, and the continuous discharge at the PD source coordinate location (the discharge amount is quantified by the PRPD spectrum) as the local point heat source. By solving the steady-state or transient temperature distribution of this heat conduction model, the net temperature rise value at the PD source coordinate location is extracted, and further the temperature rise accumulation field is calculated.

[0059] In parallel, the process performs a qualitative analysis of the PD defect type. The method inputs the phase-resolved partial discharge pattern (PRPD pattern) generated in the previous step into a pre-trained convolutional neural network (CNN) model. The CNN model has the ability to accurately classify unknown patterns by learning the characteristics of PRPD patterns corresponding to a large number of known defect types (such as internal void discharge, surface discharge, corona discharge, etc.). After the model processes the input PRPD pattern, it outputs a qualitative PD defect type, for example, it is determined to be "internal void discharge".

[0060] To achieve the above-mentioned qualitative function, the convolutional neural network model can adopt a structure containing two convolutional pooling groups and two fully connected layers. Specifically, the phase-resolved partial discharge pattern (PRPD pattern) is first processed into a 128x128 pixel grayscale image as the model input; Specifically, the PRPD pattern is first gridded into a 128x128 matrix, and the value of each element in the matrix is the cumulative discharge frequency in the phase angle-amplitude interval, then the frequency matrix is normalized to the range of 0 to 255, generating a 128x128 pixel grayscale image; The first convolutional layer uses 32 5x5 convolutional kernels, the padding method is Same, the stride is (1, 1), and the activation function is ReLU, followed by 2x2 max pooling (stride (2, 2)); The second convolutional layer uses 64 3x3 convolutional kernels, the padding method is Same, the stride is (1, 1), and the activation function is ReLU, followed by 2x2 max pooling (stride (2, 2)); Then flatten the features and connect to a fully connected layer containing 256 neurons, and finally through a Softmax output layer, the output layer contains the same number of neurons as the preset defect type (for example, 3 neurons corresponding to "internal void discharge", "surface discharge", and "corona discharge"), and outputs the probability corresponding to the preset defect type "internal void discharge", "surface discharge", and "corona discharge". The model can be trained based on labeled PRPD pattern samples using a cross-entropy loss function. Finally, the method performs a determination step. This step sets two determination conditions: one is that the PD defect type output by the convolutional neural network model is qualitatively "internal void discharge"; The second is that the temperature rise cumulative field corresponding to the PD source coordinate position exceeds a preset thermal effect threshold (for example, 200 degrees Celsius hours). When and only when both conditions are met, the PD source coordinate is finally determined to be a thermal-electric associated defect point, indicating that there is a serious insulation defect at this location with significant discharge activity and abnormal thermal effect.

[0061] As a preferred implementation, the process of jointly determining the hot-electric associated defect point based on the partial discharge pattern and the temperature rise accumulation field further comprises: time series sampling of the phase-resolved partial discharge pattern to obtain a plurality of partial discharge patterns generated at different time points; generating a quantitative pattern time series stability index by calculating the shape similarity and discharge energy trend between the plurality of partial discharge patterns; and performing the joint determination when the pattern time series stability index exceeds a preset stability threshold.

[0062] Specifically, the process does not make a determination based on a single generated partial discharge pattern, but introduces a time dimension consideration. The method generates and stores a phase-resolved partial discharge pattern once every continuous monitoring period (for example, every 1 hour), thereby obtaining a sequence of time-ordered patterns. Then, the method uses the structural similarity algorithm in the field of image processing to calculate the shape similarity between adjacent patterns in the sequence one by one. At the same time, it also calculates the total discharge energy contained in each pattern and performs trend analysis on the time series of energy values. Finally, the average value of the shape similarity and the trend slope of the energy change are combined by weighting to form a comprehensive pattern time series stability index.

[0063] Specifically, the shape similarity can be obtained by calculating the structural similarity index (SSIM) between adjacent patterns and taking the average value; the energy change trend can be obtained by performing least squares linear fitting on the time series of energy values to obtain its slope. The comprehensive pattern time series stability index is generated by weighted combination: first, the average value of the shape similarity is taken as the first component; second, the absolute value of the energy trend slope is calculated, and 1 is subtracted from the absolute value (to represent the stability of the trend) as the second component; finally, the two components are multiplied by their respective preset weights (for example, the first component weight is 0.7 and the second component weight is 0.3), and the results of the multiplication are summed to obtain the final comprehensive index. The preset stability threshold can be set according to experimental data, for example, when the comprehensive index is greater than 0.8, it is determined to be time series stable. Only when the index shows that the pattern shape is highly stable and the discharge energy presents a stable or slow growth trend, indicating that the signal source is a persistent physical entity rather than a transient interference, the method will continue to perform the subsequent joint determination with the temperature rise accumulation field.

[0064] This method introduces dynamic stability evaluation of discharge characteristics, adds a key time dimension verification to defect determination, effectively filters out false positives caused by transient noise, and greatly improves the accuracy of early defect identification.

[0065] The process establishes a double-confirmation mechanism through cross-verification of electrical characteristics (PD pattern) and physical characteristics (temperature rise cumulative field). This mechanism can effectively distinguish between serious defects caused by insulation deterioration and those affected by transient interference or external heat sources with only single electrical signal characteristics, thereby improving the accuracy and reliability of early defect determination.

[0066] Further, the timestamp comparison of the cable and overhead line traveling wave data, the acquisition of cross-domain propagation interference identification and the generation of suppression alarm, comprises: extracting the first traveling wave waveform feature and the first arrival time of the cable traveling wave data; extracting the second traveling wave waveform feature and the second arrival time of the overhead line traveling wave data; performing space-time feature comparison, comparing the first arrival time and the second arrival time, and analyzing the propagation correlation between the first traveling wave waveform feature and the second traveling wave waveform feature; when the second arrival time is earlier than the first arrival time, and the propagation correlation between the waveform features exceeds the preset correlation threshold, the cross-domain propagation interference identification is generated.

[0067] Specifically, the process is triggered when transient traveling wave signals are detected in both cable monitoring data streams and overhead line monitoring data streams. The method extracts the first nanosecond-level arrival time of the initial wave head and the first traveling wave waveform feature representing the signal pattern from the cable traveling wave data collected by the traveling wave sensors at both ends of the cable. The waveform feature includes multiple parameters such as wave head steepness, main frequency component, and polarity. In parallel, the method extracts the second nanosecond-level arrival time of the initial wave head and the corresponding second traveling wave waveform feature from the overhead line traveling wave data collected by the traveling wave sensors on the overhead line towers.

[0068] Subsequently, the method performs a spatio-temporal feature comparison. In the time dimension, the first nanosecond-level arrival time is directly compared with the second nanosecond-level arrival time. In the space and physical feature dimension, the propagation correlation between the first traveling wave waveform feature and the second traveling wave waveform feature is analyzed. The correlation analysis is based on a physical propagation model that describes the wave shape distortion rules that must occur due to wave impedance mutations when the traveling wave signal propagates from the overhead line to the cable line. The method takes the second traveling wave waveform feature actually monitored as the model input, predicts the theoretical waveform feature that should be presented after it propagates to the cable end, and performs cross-correlation calculation on the theoretical feature and the first traveling wave waveform feature actually monitored to obtain a quantitative propagation correlation coefficient value. The physical propagation model is established based on the known characteristic wave impedance of the cable and the overhead line. The model uses a standard waveform transmission coefficient, i.e., twice the cable characteristic wave impedance divided by the sum of the cable and overhead line characteristic wave impedances, to predict the theoretical change in wave amplitude when propagating from the overhead line into the cable. Meanwhile, the model can also include a low-pass filter to simulate the attenuation effect of the cable-overhead line connection point and the cable itself on high-frequency components, thereby generating the theoretical waveform feature. Specifically, the physical propagation model takes the measured waveform of the overhead line traveling wave as input. First, the amplitude is multiplied by the theoretical transmission coefficient, which is calculated by dividing twice the cable characteristic wave impedance (e.g., 50 ohms) by the sum of the cable characteristic wave impedance and the overhead line characteristic wave impedance (e.g., 400 ohms). Subsequently, the amplitude-transformed signal is passed through a second-order Butterworth low-pass filter that simulates the high-frequency attenuation effect of the connection point, with the cutoff frequency set according to the transient response characteristics of a typical cable terminal (e.g., 5 megahertz). The waveform obtained after filtering is the "theoretical waveform feature".

[0069] The propagation correlation coefficient value is obtained by calculating the normalized cross-correlation coefficient peak value of the "theoretical waveform feature" and the first traveling wave waveform feature (i.e., the cable traveling wave) actually monitored within a 2-microsecond time window after the initial wave head arrives.

[0070] As a preferred implementation, when the absolute difference between the first arrival time and the second arrival time is less than a preset time window threshold, the process of performing the spatio-temporal feature comparison further includes: calculating the signal energy of the cable traveling wave data and the overhead line traveling wave data within a preset microsecond time window after the initial wave head arrives, respectively; calculating the initial energy injection ratio according to the signal energy and the known characteristic wave impedance of the cable and the overhead line; determining the attribution domain of the fault source based on the initial energy injection ratio, and generating the cross-domain propagation interference identification.

[0071] In particular, the process is specifically designed to handle the ambiguous situation where the fault occurs near the cable-overhead line junction. When the time difference between the arrival of the cable and overhead line traveling waves is too small to determine the order, the method initiates the energy criterion. It extracts the waveform data within a very short time window (e.g. 2 microseconds) after the initial wave front of both signals arrives, and calculates the energy of each signal. Due to the physical property difference between overhead line (high wave impedance) and cable (low wave impedance), the fault energy from overhead line will change significantly when injected into the cable, and vice versa. The method utilizes this physical principle, compares the two energy values actually calculated with the energy transmission / reflection coefficient calculated based on the theoretical wave impedance, and thus obtains an initial energy injection ratio. The ratio can clearly indicate whether the energy is from the high impedance side (overhead line) or the low impedance side (cable), thus making a reliable determination of the fault source domain.

[0072] The method effectively solves the problem of interference identification in the boundary ambiguous area by introducing an energy physical criterion independent of the propagation time, ensuring that the system can still make accurate judgments on the fault source even in the most complex locations.

[0073] Finally, the method performs a determination. When the two conditions are met, a cross-domain propagation interference flag is generated: first, in time, the second arrival time is earlier than the first arrival time, indicating that the signal is from the overhead line side; second, the calculated propagation correlation coefficient value between the waveform features exceeds a preset correlation threshold (e.g. 0.8), indicating that the traveling wave monitored in the cable is highly related in physical origin to the overhead line traveling wave. Once the flag is generated, the system generates an inhibition alarm, indicating that the subsequent fault location algorithm will not process the identified cable traveling wave data, regarding it as an external interference event.

[0074] The process can accurately identify external interference traveling waves caused by overhead line side (such as lightning, switch operation) and transmitted into the cable by strictly comparing the time sequence and physical feature causality of traveling wave signals from different physical domains (cable, overhead line), effectively avoiding confusion and misjudgment caused by such interference on cable body fault location.

[0075] Further, the cable traveling wave data that has not been identified is temporarily located by a traveling wave location algorithm, including: pre-processing the cable traveling wave data that has not been identified by a wavelet transform algorithm, enhancing the rising edge features of the cable traveling wave data and suppressing background noise; based on a double-end traveling wave location algorithm, calculating the temporary coordinates by comparing the synchronous absolute time difference of the enhanced rising edge features arriving at both ends of the cable.

[0076] Specifically, the process is only performed for the cable traveling wave data identified via the foregoing step, which is not identified as a cross-domain propagation interference. First, the method uses a wavelet transform algorithm to preprocess the original waveform of the set of cable traveling wave data. The processing selects a wavelet basis function (for example, Daubechies 4 wavelet) sensitive to transient mutation signals, and performs multi-scale decomposition on the original signal. In the high-frequency coefficients after decomposition, the modulus maxima of singular points caused by fault wave fronts are very prominent, and the energy of background noise is dispersed to each scale. By reconstructing the signal, a pure waveform can be obtained, in which the rising edge feature of the wave front is significantly enhanced and the background noise is effectively suppressed.

[0077] After obtaining the preprocessed waveform, the method calculates a provisional coordinate based on a double-end traveling wave positioning algorithm. The algorithm accurately extracts the enhanced wave front rising edge feature to reach the nanosecond-level absolute time stamps of the two ends (the starting end and the terminal end) of the cable. By calculating the difference between the two absolute time stamps, the time difference of the traveling wave reaching the two ends is obtained. The calculation of the provisional coordinate is based on the time difference, the known propagation speed of electromagnetic waves in the cable, and the total length between the monitoring points at the two ends of the cable. The calculation logic is: multiply the traveling wave propagation speed by the arrival time difference to obtain the path length difference of the traveling wave propagation; subtract the path length difference from the total length of the cable, and divide the result by two. The calculation result is the distance from the fault point to one of the monitoring points, which is determined as the provisional coordinate.

[0078] The process solves the technical problem that the wave front is submerged in the original traveling wave signal and difficult to accurately identify through wavelet transform preprocessing, ensuring the accuracy of the arrival time stamp extraction. Combined with the high-precision double-end positioning algorithm, a provisional coordinate with high reliability can be obtained, providing an accurate initial spatial positioning basis for subsequent multi-dimensional feature fusion decision-making.

[0079] Further, the process of determining the fault feature through the feature classifier includes: extracting the waveform shape feature of the unmarked cable traveling wave data, the waveform shape feature including the waveform distortion degree and energy attenuation coefficient of the traveling wave; extracting the instantaneous amplitude of the sheath circulating current corresponding to the unmarked cable traveling wave data in the fault occurrence time at the cable terminal monitoring point, and extracting the high-frequency spectral component to form a circulating current feature; constructing a fused fault feature vector, the fault feature vector including the waveform shape feature and the circulating current feature; inputting the fused fault feature vector into the feature classifier to classify the fault feature.

[0080] Specifically, the process first extracts waveform morphology features from the unlabelled cable traveling wave data. The waveform distortion degree is calculated by comparing the waveform difference when the traveling wave arrives at the near-end and far-end monitoring points. After aligning the two waveforms in time, the method compares them using a structural similarity algorithm, outputting a quantitative distortion degree index. The energy attenuation coefficient is calculated by calculating the total energy of the traveling wave at the near-end and far-end, respectively, and calculating a coefficient value representing the degree of energy attenuation during propagation based on these two energy values. These two values together constitute the waveform morphology features.

[0081] In parallel, the process extracts loop current features. The method locks the fault occurrence time according to the time stamp of the fault traveling wave, and retrieves the sheath loop current data of the cable terminal monitoring point within a very short time window before and after that time. From this data, the method extracts the instantaneous amplitude of the sheath loop current at the fault instant. Subsequently, the method performs spectral analysis on the loop current signal within this time window and calculates the energy integral in a pre-set high-frequency band to obtain a high-frequency spectral component. The pre-set high-frequency band is determined according to the typical spectral characteristics of the sheath loop current signal under cable fault transient, which should be higher than the power frequency and its low-order harmonics, while capturing the main energy of the transient signal, for example, it can be set to 100 kHz to 1 MHz. These two values together constitute the loop current features.

[0082] Subsequently, the method combines the aforementioned waveform morphology features, loop current features, and the obtained tentative coordinates to construct a multi-dimensional fused fault feature vector. Specifically, this vector is a five-dimensional vector, whose components are in turn: normalized waveform distortion degree (between 0 and 1), energy attenuation coefficient (unit: dB / km), sheath loop current instantaneous amplitude (unit: ampere), loop current high-frequency spectral energy integral in the 100 kHz to 1 MHz band (unit: A²·s), and the tentative coordinate relative position value normalized to 0 to 1. Before inputting into the SVM model, the remaining components in the vector except for the normalized components (distortion degree, coordinate) are processed using Z-score standardization.

[0083] The feature vector is input into a support vector machine model as a feature classifier. The model uses a radial basis function kernel and is pre-trained. The model determines key parameters through grid search optimization, for example, the penalty coefficient C is set to 10 and the gamma parameter is set to 0.1. The training data set used by the model contains at least 1000 groups of labeled fault samples obtained through simulation or field measurement, and the sample labels cover the three main fault characteristics of "single-phase grounding", "phase-to-phase short circuit" and "high-resistance grounding".

[0084] After receiving the input vector, the model outputs a classification result according to its internal decision boundary, which is the judgment of the current fault feature. Specifically, the output of the model is a probability vector containing three components corresponding to the probabilities of the three fault features, which is subsequently input into the multi-dimensional feature fusion model as the "fault feature".

[0085] This process provides multi-dimensional physical evidence for fault nature judgment by fusing the propagation characteristics (distortion and attenuation) of traveling waves inside the cable and the performance (sheath circulating disturbance) of faults in the external electromagnetic environment. Compared with methods that rely only on single waveform analysis, this fusion analysis method can more accurately distinguish different physical causes of fault types, improving the accuracy of fault feature judgment.

[0086] Further, the input of the fault feature, the defect matching identifier, and the instantaneous temperature data at the temporary coordinate into the multi-dimensional feature fusion model, the calculation of the fault mode confidence, and the output of the cross-validation positioning result include: calculating the physical distance between the temporary coordinate and the position information of the thermal-electric associated defect point to generate a quantitative defect matching identifier; the multi-dimensional feature fusion model is a graph neural network model, which is constructed based on a pre-set physical topology structure corresponding to the high-voltage cable-overhead line hybrid power transmission network, and generates a power grid topology graph model reflecting the current load distribution by analyzing the load current data in the cable monitoring data stream to identify the current load level of each power transmission path in the physical topology structure; a high-dimensional diagnostic feature vector containing the fault feature, the defect matching identifier, and the instantaneous temperature data at the temporary coordinate is constructed, and the high-dimensional diagnostic feature vector is input into the power grid topology graph model reflecting the current operating state as node features associated with the corresponding node of the temporary coordinate in the topology graph model; the graph neural network model analyzes the relevance of the node features in the topology graph model to calculate the fault mode confidence; when the fault mode confidence exceeds a pre-set high confidence threshold, the temporary coordinate is output as the cross-validation positioning result.

[0087] Please refer to Fig. 2which shows the specific procedure for calculating the fault mode confidence and making the final location decision. The procedure first generates a quantized defect coincidence indicator. The method calls the tentative coordinates output by the traveling wave location algorithm, and the coordinates of the thermal-electric concomitant defect points output by the early defect judgment procedure. By calculating the physical distance between the two coordinate points along the cable path, a distance value is obtained. The distance value is input into a conversion function to generate a quantized coincidence indicator. The indicator is designed to be inversely proportional to the physical distance, that is, the closer the physical distance, the higher the coincidence indicator, and vice versa. Specifically, the quantized defect coincidence indicator is calculated by a Gaussian function, which is calculated as follows: first, the physical distance (unit: meter) between the tentative coordinates and the thermal-electric concomitant defect points is obtained; then, the square of the physical distance is calculated, and then divided by twice the square of the preset distance scale parameter (for example, 5 meters); finally, the calculation result is taken as the negative value, and the power of the natural constant e is calculated. The coincidence indicator calculated is a scalar value between 0 and 1.

[0088] Subsequently, the method constructs and applies a graph neural network model as a multi-dimensional feature fusion model. The graph structure of the model is pre-constructed based on the physical topology of the high-voltage cable-overhead line hybrid power transmission network, in which the nodes represent key devices such as substations, cable terminals, intermediate joints, etc., and the edges represent the cable or overhead line sections connecting these nodes. The graph structure is not static, but by analyzing the real-time acquired load current data, the current load level of each transmission path in the network is calculated, and these load levels are assigned as weights to the corresponding edges in the graph, thereby generating a power grid topology graph model that dynamically reflects the current power flow distribution of the power grid.

[0089] In making the decision, the method constructs a high-dimensional diagnostic feature vector. The vector is a 5-dimensional vector constructed by concatenation, and its components are, in order: the fault feature probability vector (containing three components) output by the aforementioned feature classifier, the quantized defect coincidence indicator (scalar between 0 and 1) calculated in the previous step, and the instantaneous temperature data (unit: Celsius) at the tentative coordinates synchronized with the time of fault occurrence obtained from the temperature sensor. Before inputting the graph neural network model, the instantaneous temperature data component in the vector is standardized.

[0090] The high-dimensional vector is input into a dynamic power grid topology graph model as node features. When inputting, the system associates the high-dimensional diagnostic feature vector to the corresponding node or edge according to the physical position of the tentative coordinates in the power grid topology graph model: if the tentative coordinates are located on a node (such as an intermediate joint), the vector is input as the node feature of the node; if the tentative coordinates are located on an edge (i.e. a cable line segment) connecting two nodes (for example, node A and node B), the vector can be input into the model as an edge feature of the edge, or according to the physical distance of the tentative coordinates from node A and node B, the vector is distributed and updated to the node features of node A and node B respectively through an inverse distance weighting function (i.e. a distribution mode in which the closer the distance, the higher the weight). The graph neural network model (for example, a graph convolutional network) incorporates the edge features or updated node features into the calculation when aggregating neighborhood information through its message passing mechanism, to analyze the relevance of the fault features in the topology graph.

[0091] The graph neural network model then analyzes the topology graph with the input features. Specifically, the graph neural network model can adopt a graph convolutional network (GCN). Its neighborhood information aggregation mechanism follows the propagation rule of GCN, that is, the feature vector of each node (for example, a cable joint) is updated in the next layer by weighted averaging (the weight is determined by the normalized Laplacian matrix of the power grid topology graph and the current flow distribution weight) of its own feature and the feature of its first-order neighbor node, and then passing through a nonlinear activation function (for example, ReLU). The graph neural network model can stack two graph convolutional layers to allow information to propagate two-hop distance on the topology graph. Through its inherent neighborhood information aggregation mechanism, the model can evaluate the rationality and relevance of the node features (i.e. the multi-dimensional information of the fault) in the entire power grid topology structure and the current operating state. The model finally outputs a comprehensive fault mode confidence. The confidence is used to compare with a preset high confidence threshold. The fault mode confidence is a probability value between 0 and 1 calculated by the output layer (for example, Softmax layer or Sigmoid layer) of the graph neural network model. The preset high confidence threshold can be determined by ROC curve analysis to determine the best threshold point according to historical fault data and model verification results, for example, it can be set to 0.9 or 0.95 to ensure that the output result has high accuracy. When the calculated confidence exceeds the threshold, it indicates that the traveling wave positioning result is strongly supported and cross-validated by multi-source information, at which point the method outputs the tentative coordinates as the final cross-validation positioning result.

[0092] As a preferred implementation, the process of outputting the cross-validated positioning result comprises: discretizing the cable line into a plurality of physical segments along its length; the graph neural network model outputs the probability of the fault occurring in each physical segment, forming a fault location probability distribution along the cable line; the location of the physical segment with the highest probability is taken as the main positioning result, and other physical segments with a probability exceeding a preset suboptimal threshold are taken as alternative positioning results, which are output together.

[0093] Specifically, the final output of the method is not a single coordinate point. In the model construction stage, the digital twin model of the entire cable line has been pre-discretized into a series of continuous physical segments with unique numbers (for example, one segment every 10 meters). The final output layer of the graph neural network model is designed as a Softmax classifier, and the number of output nodes corresponds to the number of physical segments. Therefore, after the model completes the comprehensive analysis of the high-dimensional diagnostic feature vector, the output is no longer a single confidence value, but a probability vector. Each element in the vector corresponds to a physical segment, and its value represents the posterior probability of the fault occurring in that segment. Finally, the system visualizes the probability vector as a fault location probability distribution along the cable line. The center point of the segment with the highest probability value in the graph is reported as the main fault location, and all segments with probability values that are not the highest but still exceed a certain suboptimal threshold are also highlighted as alternative fault locations that need to be focused on. The setting of the preset suboptimal threshold aims to balance recall rate and precision rate, ensuring that potential fault locations are not missed. The threshold can be set as a fixed value (for example, all segments exceeding 0.2 in the probability distribution) or a relative value (for example, all segments with a probability exceeding 30% of the highest probability value).

[0094] By outputting a probability distribution instead of a single coordinate, the method quantifies and presents the uncertainty of the positioning result, providing decision-making information including the main and alternative locations for the operation and maintenance personnel, greatly improving the strategic flexibility of fault repair.

[0095] The process establishes a final decision-making mechanism based on the global topology of the power grid and the real-time operating state, which deeply integrates the transient characteristics of the fault, the precursor information of the slowly varying defects, and the local physical environment state. This system-level comprehensive analysis goes beyond the local interpretation of a single signal and greatly improves the final confidence of the fault positioning result and its reliability in complex working conditions through cross-validation of multiple sources of information.

[0096] Further, the calculating the fault mode confidence further includes: introducing a physical constraint term based on a cable transmission line model, the physical constraint term being used to quantify a consistency degree between the space-time data containing the traveling wave arrival time and the PD source coordinates in the high-dimensional diagnostic feature vector and the cable physical parameters containing the wave speed and the length; and adding the physical constraint term to a loss function of the graph neural network model, and correcting the fault mode confidence based on the physical constraint term by minimizing the loss function.

[0097] Specifically, the process introduces a physical constraint term in the training and inference stages of the graph neural network model to enhance the physical rationality of its decision. The physical constraint term is constructed based on a standard cable transmission line theory model. When the model is applied, it extracts two types of space-time data in the high-dimensional diagnostic feature vector: one is the arrival time data of the fault traveling wave, and the other is the previously determined PD source coordinates. At the same time, it obtains the known physical parameters of the cable from the system parameter library, mainly including the theoretical propagation speed of the electromagnetic wave and the accurate length of the cable section.

[0098] The core function of the physical constraint term is to calculate the consistency degree between the above space-time measurement data and the cable physical parameters. For example, it verifies whether the fault location calculated according to the traveling wave arrival time difference is consistent with the physical position of the PD source coordinates within the allowed error range. It also verifies whether the time required for the traveling wave to theoretically propagate to the two end monitoring points from the PD source coordinate position is consistent with the actually monitored traveling wave arrival time if a PD source is considered as the source of the traveling wave fault. The results of all these verifications are quantified as one or more consistency indicators.

[0099] In the training stage of the graph neural network model, the quantified consistency indicator is added to the loss function of the model as an independent term. For example, the physical constraint term can be defined as: when the defect consistency identifier is greater than a preset threshold, the position difference between the traveling wave positioning coordinates and the PD source coordinates is calculated, and the square of the difference value (i.e. mean square error) is taken. The complete composite loss function is defined as the sum of two terms: the first term is the cross-entropy loss between the fault location probability distribution output by the graph neural network model and the true fault location label; and the second term is the square of the above physical constraint term (position difference) multiplied by a hyperparameter (e.g. 0.1) for balancing the classification accuracy and the physical consistency.

[0100] By minimizing this composite loss function containing the physical constraint term during the training process, the weight parameters inside the model will be adjusted so that the output fault mode confidence will naturally tend to those more consistent with the cable transmission physics. In the final inference decision stage, the fault mode confidence output by this physically constrained trained model is a more reliable result that has been corrected by physical laws.

[0101] This process deeply couples the data-driven machine learning model with the first-principle-based physical model. This method effectively avoids the pure data model that may produce solutions that violate common sense, ensuring that the final output of the fault mode confidence is not only statistically optimal in data but also highly self-consistent and interpretable in physical logic.

[0102] The cable fault location method based on multi-source data space-time correlation of the present application deeply fuses and cross- verifies the prior diagnostic information for slowly varying physical defects with real-time positioning information for transient electrical events. This method effectively solves the fundamental defect that traditional technology only relies on a single transient signal and weak decision basis and is easily disturbed. Through this fusion, the final output of the fault location conclusion is no longer an isolated mathematical calculation result, but a high-confidence diagnosis with a high degree of physical causal self-consistency that has been pre-verified by the long-term health status of the line, thereby achieving a substantial improvement in positioning reliability.

[0103] Embodiment 2:

[0104] The embodiment of the present application provides a cable fault location system based on multi-source data space-time correlation, which is applied to a high-voltage level urban power grid cable-overhead line hybrid transmission line to realize fault location and health status evaluation of the line. Referring to Fig. 3 The system includes the following modules:

[0105] The data acquisition module is the data basis of the entire system. The module is composed of multiple hardware sensors and data acquisition units deployed in the field of high-voltage cable-overhead line hybrid transmission networks. Specifically, at the key positions of the cable section (such as cable terminals and intermediate joints), high-frequency current sensors, flexible Rogowski coils, contact temperature sensors, special traveling wave sensors, and current transformers are deployed. These sensors are responsible for collecting partial discharge signals, load currents, surface temperatures, cable traveling waves, and sheath circulating currents, respectively, to form cable monitoring data streams. On the towers of overhead transmission sections, self-powered traveling wave sensors are deployed to collect overhead line monitoring data streams. All acquisition units are equipped with high-precision timing units to ensure that all data points in the data stream have uniform, nanosecond-resolution absolute time stamps, providing a basis for subsequent space-time correlation analysis.

[0106] The potential defect judgment module is responsible for early identification and positioning of the slowly varying insulation defects of the cable. The module receives partial discharge data and load current data from the data acquisition module. It first performs TDOA algorithm phase fusion positioning, calculates the physical coordinates of the PD source by analyzing the time difference of the partial discharge pulse signal arriving at the sensors at both ends of the cable, and combines the power frequency phase information of the load current to generate a phase analysis partial discharge spectrum representing the discharge characteristics of the discharge source; then, the module calls the temperature data corresponding to the PD source coordinate position, and uses the load current data for thermal effect compensation to strip the influence of normal operation temperature rise, so as to extract the local abnormal temperature rise accumulation field caused by continuous discharge activity; finally, the module jointly analyzes the partial discharge spectrum and the temperature rise accumulation field, when a defect shows specific discharge characteristics (for example, internal void discharge) and significant abnormal thermal effect at the same time, it is judged as a thermal-electricity associated defect point, and the position and characteristic information of the defect point are stored as prior knowledge for subsequent modules.

[0107] The network level discrimination module is the core function of distinguishing the faults originating from the cable body and the external interference from the overhead line side. When the system monitors the transient traveling wave event, the module receives the cable traveling wave data and the overhead line traveling wave data at the same time, it compares the nanosecond level time stamps of the two groups of data, and analyzes the physical propagation correlation between the waveform characteristics; based on the logic of cause and effect, if the arrival time of the overhead line traveling wave is significantly earlier than that of the cable traveling wave, and the waveform characteristics of the two meet the physical propagation law from the overhead line to the cable, the module determines that this event is external interference; at this time, it generates a cross-domain propagation interference identifier and triggers a suppression alarm to prevent the contaminated cable traveling wave data from entering the subsequent fault location process.

[0108] The fault feature judgment module is responsible for deep feature extraction and preliminary characterization of the confirmed internal fault traveling wave event. The module only processes the cable traveling wave data that is not identified by the network level discrimination module. It first applies the traveling wave positioning algorithm to calculate a preliminary fault location, i.e. the temporary coordinate; then, based on the temporary coordinate and the instantaneous amplitude of the sheath circulating current at the time of fault occurrence, it uses a pre-trained feature classifier to determine the basic physical characteristics of the fault (such as single-phase grounding or phase-to-phase short circuit); at the same time, the module also performs physical position comparison between the temporary coordinate and the thermal-electricity associated defect points output by the potential defect judgment module, generates a quantitative defect matching identifier to represent the spatial correlation degree between the transient fault and the known chronic defect.

[0109] The fusion decision module is the decision core of the whole system, and is responsible for the final comprehensive research and judgment of all information. The module receives the fault features output by the fault feature judgment module, the defect matching identifier, and the instantaneous temperature data at the tentative coordinates provided by the data acquisition module. It inputs these multi-dimensional information as a high-dimensional diagnostic feature vector into a multi-dimensional feature fusion model based on a graph neural network. The model can combine the real-time topology structure and the power flow distribution state of the power grid to calculate a comprehensive fault mode confidence. The confidence reflects the possibility of the tentative coordinates as the real fault point, and how many independent information sources support it. When the confidence exceeds a preset high confidence threshold, the module outputs the tentative coordinates as the final cross-validation positioning result, thereby completing a high-reliability fault location.

[0110] The cable fault location system based on multi-source data space-time association of the application constructs a complete closed loop from data acquisition, early defect discovery, external interference elimination, to multi-dimensional feature extraction and final fusion decision through the sequential cooperative work of the above modules, and realizes high-precision and high-confidence positioning of the mixed line cable fault.

[0111] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto. Any changes or replacements within the technical range disclosed by the application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the application. Therefore, the protection scope of the application should be limited by the protection scope defined in the claims.

Claims

1. A method for cable fault location based on spatio-temporal correlation of multi-source data, characterized in that, The method comprises the following steps: obtaining cable monitoring data stream containing partial discharge data, load current data, temperature data, cable traveling wave data and sheath loop current data from a high-voltage cable-overhead line hybrid power transmission network, and overhead line monitoring data stream containing overhead line traveling wave data; based on the partial discharge data and the load current data, obtaining the PD source coordinates and the phase-resolved partial discharge map through TDOA algorithm phase fusion positioning; calculating the time difference of arrival of high-frequency PD pulse signals in the partial discharge data through the TDOA algorithm to obtain the PD source coordinates; correlating the occurrence phase of the high-frequency PD pulse signals with the power frequency phase of the load current data to generate the phase-resolved partial discharge map, which is used to represent the discharge amount and discharge frequency distribution of PD signals within a power frequency cycle; based on the PD source coordinates, analyzing the temperature data to extract the temperature rise accumulation field; jointly determining the thermal-electric associated defect point based on the partial discharge map and the temperature rise accumulation field; through the space-time weighted algorithm, performing time series analysis on the temperature data at the PD source coordinate position, and after using the load current data to compensate the temperature data for thermal effect to suppress the thermal disturbance caused by the change of load current, identifying the local abnormal temperature rise trend that is consistent with the PD source coordinate position in space and continuously exists in time to establish the temperature rise accumulation field; wherein the space-time weighted algorithm performs spatial weighting according to the physical distance between the PD source coordinates and the temperature sensor, and performs time weighting on the time series data points according to the occurrence time; inputting the phase-resolved partial discharge map as a two-dimensional frequency spectrum into a convolutional neural network model through the convolutional neural network model to qualitatively determine the PD defect type; wherein the convolutional neural network model comprises a convolutional layer for extracting map features, a pooling layer for dimension reduction, and a fully connected layer for outputting classification results; when the PD defect type is qualitatively determined as internal void discharge and the temperature rise accumulation field exceeds a preset thermal effect threshold, the thermal-electric associated defect point is determined; performing timestamp comparison on the cable traveling wave data and the overhead line traveling wave data to obtain cross-domain propagation interference identification and generate suppression alarm; for the cable traveling wave data that has not been identified, obtaining a tentative coordinate through a traveling wave positioning algorithm; based on the tentative coordinate and the instantaneous amplitude of the sheath loop current obtained at the cable terminal monitoring point, determining the fault characteristics through a feature classifier; comparing the tentative coordinate with the thermal-electric associated defect point to generate defect matching identification; The fault feature, the defect matching identifier, and the instantaneous temperature data at the temporary coordinate are input into a multi-dimensional feature fusion model, the fault mode confidence is calculated, and the cross-validation positioning result is output; the physical distance between the temporary coordinate and the position information of the thermal-electric associated defect point is calculated, and a quantitative defect matching identifier is generated; the multi-dimensional feature fusion model is a graph neural network model, the graph neural network model is constructed based on a preset physical topology structure corresponding to a high-voltage cable-overhead line hybrid power transmission network, and the current load level of each power transmission path in the physical topology structure is identified by analyzing the load current data in the cable monitoring data stream, and a power grid topology graph model with an operating state weight reflecting the current power flow distribution is generated; a high-dimensional diagnostic feature vector containing the fault feature, the defect matching identifier, and the instantaneous temperature data at the temporary coordinate is constructed, the high-dimensional diagnostic feature vector is input as a node feature into the power grid topology graph model reflecting the current operating state, and the node feature is associated with the corresponding node of the temporary coordinate in the topology graph model; the graph neural network model analyzes the association of the node feature in the topology graph model, and calculates the fault mode confidence; when the fault mode confidence exceeds a preset high confidence threshold, the temporary coordinate is output as the cross-validation positioning result.

2. The cable fault locating method based on multi-source data space-time correlation according to claim 1, characterized in that, The timestamp comparison of the cable traveling wave data and the overhead line traveling wave data is performed, the cross-domain propagation interference identifier is obtained, and the suppression alarm is generated, including: The first traveling wave waveform feature and the first arrival time of the cable traveling wave data are extracted; the second traveling wave waveform feature and the second arrival time of the overhead line traveling wave data are extracted; the space-time feature comparison is performed, the first arrival time and the second arrival time are compared, and the propagation association between the first traveling wave waveform feature and the second traveling wave waveform feature is analyzed; when the second arrival time is earlier than the first arrival time, and the propagation association between the waveform features exceeds a preset association threshold, the cross-domain propagation interference identifier is generated.

3. The method of claim 1, wherein, The temporary coordinate of the unmarked cable traveling wave data is obtained by the traveling wave positioning algorithm, including: The unmarked cable traveling wave data is preprocessed by the wavelet transform algorithm, the rising edge feature of the cable traveling wave data is enhanced, and the background noise is suppressed; based on the double-end traveling wave positioning algorithm, the synchronous absolute time difference of the enhanced rising edge feature arriving at both ends of the cable is compared, and the temporary coordinate is calculated.

4. The method of claim 1, wherein, The process of determining the fault feature by the feature classifier includes: The waveform shape feature of the unmarked cable traveling wave data is extracted, the waveform shape feature includes the waveform distortion degree and the energy attenuation coefficient of the traveling wave; the instantaneous amplitude of the sheath circulating current corresponding to the unmarked cable traveling wave data in the fault occurrence time obtained at the cable terminal monitoring point is extracted, and the high-frequency spectral component is extracted to form a circulating current feature; a fused fault feature vector is constructed, the fused fault feature vector includes the waveform shape feature and the circulating current feature; the fused fault feature vector is input into the feature classifier, and the fault feature is classified.

5. The method of claim 1, wherein, The calculating the fault mode confidence further includes: A physical constraint term based on a cable transmission line model is introduced, which is used to quantify the consistency between the space-time data containing the traveling wave arrival time and the PD source coordinates in the high-dimensional diagnostic feature vector and the cable physical parameters containing the wave speed and length; the physical constraint term is added to the loss function of the graph neural network model, and the fault mode confidence is corrected based on the physical constraint term by minimizing the loss function.

6. A cable fault locating system based on spatio-temporal correlation of multi-source data, characterized in that, The method of claim 1 is executed, including: A data acquisition module acquires cable monitoring data streams containing partial discharge data, load current data, temperature data, cable traveling wave data and sheath loop current data from a high-voltage cable-overhead line hybrid power transmission network, and overhead line monitoring data streams containing overhead line traveling wave data; A potential defect determination module obtains PD source coordinates and a phase-resolved partial discharge map through TDOA algorithm phase fusion positioning based on the partial discharge data and the load current data; analyzes the temperature data based on the PD source coordinates to extract a temperature rise cumulative field; and determines a thermal-electric associated defect point in combination with the partial discharge map and the temperature rise cumulative field; A network-level identification module compares time stamps of the cable traveling wave data and the overhead line traveling wave data to obtain cross-domain propagation interference identification and generate suppression alarms; A fault feature determination module obtains tentative coordinates through a traveling wave positioning algorithm for cable traveling wave data that has not been identified; determines fault features through a feature classifier based on the tentative coordinates and the sheath loop current instantaneous amplitude obtained at the cable terminal monitoring point; and compares the tentative coordinates with the thermal-electric associated defect point to generate defect matching identification; A fusion decision module inputs the fault features, the defect matching identification and the instantaneous temperature data at the tentative coordinates into a multi-dimensional feature fusion model, calculates a fault mode confidence, and outputs cross-validation positioning results.

Images