A method for all-around monitoring and fault location of high voltage cable

Abstract

The application provides a kind of all-around monitoring and fault location method for high-voltage cable, the method synchronously collects multiple physical quantities to obtain original data set, obtains multiple parameter synchronous data set after preprocessing;Then classify according to the physical properties of parameters to extract features, obtain standardized multi-domain feature set;Through the analytic hierarchy process entropy weight combination algorithm, a fault risk composite index is constructed and a standard feature fingerprint is generated, a fault feature matrix is obtained through dimension fusion and space matching, and an adaptive fault recognition classification model is input, fault type recognition, section initial positioning and risk classification are realized;Finally, combined with the fault feature fingerprint and the multi-parameter constraint, the fault coordinates are accurately obtained. This method realizes the all-around monitoring of high-voltage cable full link, improves the accuracy of fault identification and the accuracy of positioning, effectively solves the problem of incomplete monitoring and large error in identification and positioning of the prior art.

Description

Technical Field

[0001] This invention relates to the field of power system transmission and distribution technology, and in particular to a method for comprehensive monitoring and fault location of high-voltage cables. Background Technology

[0002] With the development of power systems towards higher voltage, larger capacity, and smarter operation, high-voltage cables have become a core component of urban power grids, new energy power plants, and industrial park transmission and distribution networks. Their laying environments are mostly enclosed spaces such as underground pipe corridors and cable trenches. During operation, they are susceptible to faults caused by factors such as environmental temperature and humidity, mechanical stress, insulation aging, and poor joint contact. Furthermore, fault location is difficult to pinpoint and maintenance cycles are long, posing a serious threat to the safe and stable operation of the power system. Currently, high-voltage cable monitoring technologies are mainly divided into offline preventative testing and online monitoring. Offline testing methods, such as DC withstand voltage testing and partial discharge offline detection, suffer from long testing intervals and the inability to monitor cable operating status in real time. Online monitoring technologies often employ single-parameter monitoring modes, such as monitoring only partial discharge signals or cable temperature. Some multi-parameter monitoring technologies also suffer from scattered sensor deployment and a lack of full-link coverage, and the spatiotemporal synchronization accuracy of multi-parameter acquisition is low, resulting in poor data correlation. In terms of fault identification and location, existing fault identification methods are mostly based on traditional machine learning models, which only extract single-domain or a few-dimensional features of the signal. The model training lacks the support of the physical mechanism of the cable, resulting in low accuracy of fault type identification and section location. The mainstream fault location methods are traveling wave method and impedance method. The impedance method is suitable for low-resistance faults but has a large location error. The traveling wave method has problems such as wavefront identification being easily affected by reflection and refraction, and the traveling wave velocity using a fixed value without combining real-time cable operating conditions for correction. As a result, the fault point location accuracy is difficult to meet the actual engineering needs. Summary of the Invention

[0003] The purpose of this invention is to provide a comprehensive monitoring and fault location method for high-voltage cables. By using an adaptive fault identification and classification model, fault feature fingerprints, and multi-parameter constraints, the method can improve the comprehensiveness of high-voltage cable fault monitoring, the accuracy of fault identification, and the precision of fault location.

[0004] To achieve the above objectives, the present invention provides the following solution: A method for comprehensive monitoring and fault location of high-voltage cables includes the following steps: According to the preset spatial layout rules, multiple types of sensors are set up around the high-voltage cable, and multiple physical quantities are collected synchronously through the multiple types of sensors to obtain the original dataset. Spatiotemporal registration, outlier removal, baseline correction, and signal enhancement are performed on the original dataset to obtain a multi-parameter synchronous dataset; The multi-parameter synchronization dataset is classified according to the physical properties of the parameters, and the classification results are subjected to multi-dimensional and multi-domain feature extraction to obtain a standardized multi-domain feature set. The standardized multi-domain feature set includes: time domain features, frequency domain features, time-frequency domain features, and nonlinear entropy features. Based on a standardized multi-domain feature set, a composite index of cable fault risk is constructed using the hierarchical analysis entropy weight combination algorithm. Standard feature fingerprints corresponding to various typical faults are generated, and the composite index of cable fault risk is dimensionally fused and spatially matched with the standard feature fingerprints to obtain a fault feature matrix. Based on the fault feature matrix, a pre-trained adaptive fault identification and classification model is used to identify cable fault types, locate fault sections initially, and classify fault risk levels, resulting in fault identification results, initial fault location results, and fault risk classification results. Based on the initial fault location results, the fault traveling wave head and traveling wave velocity of the fault area are identified and dynamically corrected by fault feature fingerprint and multi-parameter constraints to obtain the fault coordinate results.

[0005] Optionally, multiple types of sensors are installed around the high-voltage cable according to preset spatial layout rules, and multiple physical quantities are simultaneously collected through these sensors to obtain the original dataset, including: Temperature integrated sensors are installed along the high-voltage cable route. UHF partial discharge sensors, contact temperature sensors, joint grounding current sensors, and ultrasonic sensors are installed at the intermediate joints of the high-voltage cable. Terminal temperature sensors, gas density sensors, and terminal grounding current sensors are installed at both ends of the high-voltage cable. Sheath circulating current sensors, zero-sequence current sensors, and box temperature and humidity sensors are installed in cross-connection boxes and grounding boxes. Ambient temperature and humidity sensors, water immersion sensors, and harmful gas sensors are installed in cable trenches and pipe corridors. The parameters to be collected are divided into electrical parameters, temperature and strain parameters, acoustic parameters, and environmental parameters. Electrical parameters include: partial discharge signal, sheath circulating current, zero-sequence current, joint and terminal grounding current; temperature and strain parameters include: cable distributed temperature and strain, joint temperature and terminal temperature; acoustic parameters include: joint partial discharge ultrasonic signal and cable body vibration signal; environmental parameters include: ambient temperature and humidity, water immersion status, and concentration of harmful gases. Based on the preset differential sampling, the parameters to be collected are collected synchronously to obtain the original dataset.

[0006] Optionally, the original dataset undergoes spatiotemporal registration, outlier removal, baseline correction, and signal enhancement processing to obtain a multi-parameter synchronization dataset, including: Spatial location matching of the original dataset is performed based on a unified timestamp and sensor spatial coordinates to obtain a multi-channel signal set; Outlier identification and removal are performed on the multi-channel signal set based on the 3σ criterion and the median value within the sliding window to obtain an outlier-free signal set. The length of the sliding window is determined by the sampling frequency, and outliers are replaced by the median value within the sliding window. The DC drift and low-frequency baseline drift of the power frequency are removed from the set of signals without anomalies by a fifth-order polynomial fitting algorithm to obtain the baseline correction signal set. The baseline correction signal set is enhanced by a combined adaptive variational mode decomposition and adaptive wavelet thresholding denoising algorithm to obtain a multi-parameter synchronization dataset.

[0007] Optionally, the baseline correction signal set is subjected to feature enhancement processing using a joint denoising algorithm of adaptive variational mode decomposition and adaptive wavelet thresholding to obtain a multi-parameter synchronization dataset, including: The optimal number of modes for variational mode decomposition is determined based on the energy entropy extreme value criterion, and the baseline correction signal set is decomposed into multiple intrinsic mode components based on the optimal number of modes. Calculate the Pearson correlation coefficient between the intrinsic modal components and the original dataset, and filter and remove the intrinsic modal components according to the preset correlation threshold to obtain the effective modal components; The optimal threshold for each wavelet decomposition level in the effective modal components is calculated based on the unbiased risk estimation wavelet threshold function, and threshold denoising is performed on the effective modal components based on the optimal threshold to obtain the denoised components. The signal is reconstructed based on the denoised components, and the reconstructed signal is enhanced with prior information on fault features to obtain a multi-parameter synchronous dataset.

[0008] Optionally, the time-domain characteristics of the electrical parameters include: peak value, kurtosis, pulse count, rise time, fall time, RMS value, rectified average value, and pulse width; the frequency-domain characteristics include: center frequency, spectral energy distribution, total harmonic distortion rate, fundamental frequency ratio, characteristic frequency band energy ratio, and spectral kurtosis; the time-frequency domain characteristics include: wavelet packet decomposition energy entropy, Hilbert-Huang transform marginal spectral characteristics, instantaneous frequency, and instantaneous amplitude characteristics; the nonlinear entropy characteristics include: sample entropy, fuzzy entropy, permutation entropy, approximate entropy, and multi-scale entropy. The time-domain characteristics of temperature strain parameters include: mean temperature, rate of temperature rise, temperature gradient, extreme values ​​of axial temperature difference, strain change, and strain fluctuation variance; the nonlinear entropy characteristics include: permutation entropy of the temperature sequence and sample entropy of the strain sequence. The time-domain characteristics of acoustic parameters include: mean amplitude, peak value, vibration frequency, pulse duration, and signal energy; the frequency-domain characteristics include characteristic frequency components and spectral energy distribution.

[0009] Optionally, based on a standardized multi-domain feature set, a composite index of cable fault risk is constructed using the analytic hierarchy process (AHP) entropy weight combination algorithm. Standard feature fingerprints corresponding to various typical faults are generated, and the composite index of cable fault risk is dimensionally fused and spatially matched with the standard feature fingerprints to obtain a fault feature matrix, including: The standardized multi-domain feature set is screened for fault causal features by using a convergent cross-mapping algorithm, and redundant features and pseudo-correlated features with causal correlation strength below the preset causal threshold are removed to obtain the core feature set. Calculate the dynamic combination weights of each core feature in the core feature set, and construct a dynamic weight set; The composite index of cable fault risk is calculated based on the core feature set and dynamic weight set. By calculating the temporal change rate and future trend of fault risk at the point, a fault evolution trend term is introduced into the composite index of cable fault risk to obtain the spatiotemporal sequence of the composite index of fault risk. Based on the physical mechanism of cable faults, a typical fault basic feature fingerprint database is constructed. The core feature set is then calibrated and adapted in terms of amplitude and dimension by using basic cable parameters to obtain a standard feature fingerprint database. Based on spatial coordinates, the spatiotemporal sequence of the fault risk composite index is spatially matched with the standard feature fingerprint database, and multi-dimensional fusion is performed through a two-way constraint mechanism to obtain the fault feature matrix.

[0010] Optionally, based on the fault feature matrix, a pre-trained adaptive fault identification and classification model is used to identify cable fault types, initially locate fault sections, and classify fault risk levels, obtaining fault identification results, initial fault location results, and fault risk classification results, including: An adaptive fault identification and classification model is constructed, which is set up in sequence as follows: physical mechanism embedding layer, spatial topology attention branch, temporal evolution attention branch, dual-branch feature fusion layer, adaptive classification and localization layer, and dynamic risk classification layer; The adaptive fault identification and classification model is pre-trained in three stages. The fault feature matrix is ​​input into the trained adaptive fault identification and classification model. First, it passes through the physical mechanism embedding layer to complete the rigid fusion of physical priors and data features. Then, it is input into the spatial topology attention branch and the temporal evolution attention branch respectively to obtain the spatial fault attention weight map and the temporal fault evolution feature map. The fault depth features are obtained by weighted fusion through the dual-branch feature fusion layer. The fault type identification and initial fault segment localization are performed by using the dual-constraint reasoning mechanism of the adaptive classification and localization layer to obtain the fault identification result and the initial fault localization result. Based on the dynamic risk classification layer, the risk classification score is calculated and the risk level is divided according to the fault risk composite index, fault evolution rate and cable section importance coefficient, so as to obtain the fault risk classification result.

[0011] Optionally, the adaptive fault identification and classification model undergoes a three-stage progressive pre-training process, including: The adaptive fault identification and classification model is pre-trained under basic operating conditions using long-term normal operation data of cables and conventional environmental interference data. The adaptive fault identification and classification model was pre-trained using fault mechanism transfer data from physical simulation data of cable faults and laboratory simulated fault data. The adaptive fault identification and classification model is fine-tuned using small-sample incremental adjustments based on historical fault samples from the target cable site.

[0012] Optionally, based on the initial fault location results, fault traveling wave front identification and traveling wave velocity dynamic correction are performed on the fault area using fault feature fingerprints and multi-parameter constraints to obtain fault coordinate results, including: The target fault section is determined based on the initial fault location results, and the fault traveling wave signal of the target fault section is retrieved. Based on the fault feature fingerprint, the fault traveling wave signal is subjected to Hilbert-Huang transform to obtain the instantaneous amplitude and instantaneous phase. Then, the candidate set of traveling wave fronts is screened based on the jump point, and the interfering wave fronts are eliminated to obtain the double-ended traveling wave time difference between the time when the fault traveling wave signal arrives at the beginning of the segment and the time when it arrives at the end of the segment. An adaptive correction formula for traveling wave velocity with multi-parameter constraints is constructed based on a multi-parameter synchronous dataset to calculate the corrected traveling wave velocity. The distance from the fault point to the beginning of the section is calculated based on the time difference between the two-end traveling waves and the corrected traveling wave velocity. Based on the geospatial coordinates of the faulty section, the distance from the fault point to the beginning of the section is converted into the geographical coordinates of the fault point, thus obtaining the fault coordinate result.

[0013] A comprehensive monitoring and fault location system for high-voltage cables, comprising: The data acquisition module is used to set up multiple types of sensors around the high-voltage cable according to the preset spatial layout rules, and to collect multiple physical quantities synchronously through the multiple types of sensors to obtain the raw dataset. The preprocessing module is used to perform spatiotemporal registration, outlier removal, baseline correction, and signal enhancement on the original dataset to obtain a multi-parameter synchronous dataset. The classification and feature extraction module is used to classify multi-parameter synchronization datasets based on the physical properties of the parameters, and to extract multi-dimensional and multi-domain features from the classification results to obtain a standardized multi-domain feature set. The standardized multi-domain feature set includes: time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear entropy features. The index calculation module is used to construct a composite index of cable fault risk based on a standardized multi-domain feature set and through the hierarchical analysis entropy weight combination algorithm, and generate standard feature fingerprints corresponding to various typical faults. The composite index of cable fault risk is then dimensionally fused and spatially matched with the standard feature fingerprints to obtain a fault feature matrix. The fault identification module is used to identify cable fault types, initially locate fault sections, and classify fault risk levels based on the fault feature matrix and a pre-trained adaptive fault identification classification model, so as to obtain fault identification results, initial fault location results, and fault risk classification results. The fault location module is used to identify the traveling wave head and dynamically correct the traveling wave velocity of the fault area based on the initial fault location results, through fault feature fingerprints and multi-parameter constraints, to obtain the fault coordinate results.

[0014] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a comprehensive monitoring and fault location method for high-voltage cables, the method comprising: setting up multiple types of sensors around the high-voltage cable according to preset spatial layout rules, and synchronously acquiring multiple physical quantities through the multiple types of sensors to obtain an original dataset; performing spatiotemporal registration, outlier removal, baseline correction, and signal enhancement processing on the original dataset to obtain a multi-parameter synchronous dataset; classifying the multi-parameter synchronous dataset according to the physical properties of the parameters, and extracting multi-dimensional and multi-domain features from the classification results to obtain a standardized multi-domain feature set; based on the standardized multi-domain feature set, through layer... This method employs an entropy weight combination algorithm to construct a composite index for cable fault risk and generates standard feature fingerprints corresponding to various typical faults. The composite index and standard feature fingerprints are then dimensionally fused and spatially matched to obtain a fault feature matrix. Based on this matrix, a pre-trained adaptive fault identification and classification model is used to identify cable fault types, initially locate fault sections, and classify fault risk levels, yielding fault identification results, initial fault location results, and fault risk classification results. Based on the initial fault location results, fault traveling wave front identification and dynamic correction of traveling wave velocity are performed on the fault area using fault feature fingerprints and multi-parameter constraints, resulting in fault coordinates. This method constructs a multi-parameter synchronous monitoring system for the entire high-voltage cable link, performs data preprocessing, multi-dimensional and multi-domain feature extraction, constructs and fuses composite fault risk indicators, and combines an adaptive fault identification and classification model to achieve fault type identification, initial section location, and risk classification. It also utilizes fault feature fingerprints and multi-parameter constraints to achieve precise fault traveling wave location. This approach enables comprehensive and accurate perception of the entire high-voltage cable link's operating status and efficient fault identification and location, improving the comprehensiveness of high-voltage cable fault monitoring, the accuracy of fault identification, and the precision of fault location. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the comprehensive monitoring and fault location method for high-voltage cables according to the present invention. Figure 2 This is a schematic diagram of the structure of the all-round monitoring and fault location system for high-voltage cables according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides a method for comprehensive monitoring and fault location of high-voltage cables, comprising the following steps: Step 100: According to the preset spatial layout rules, set up multiple types of sensors around the high-voltage cable, and collect multiple physical quantities synchronously through the multiple types of sensors to obtain the original dataset; Step 200: Perform spatiotemporal registration, outlier removal, baseline correction, and signal enhancement on the original dataset to obtain a multi-parameter synchronization dataset; Step 300: Classify the multi-parameter synchronization dataset according to the physical properties of the parameters, and extract multi-dimensional and multi-domain features from the classification results to obtain a standardized multi-domain feature set; the standardized multi-domain feature set includes: time domain features, frequency domain features, time-frequency domain features, and nonlinear entropy features; Step 400: Based on the standardized multi-domain feature set, construct the cable fault risk composite index through the hierarchical analysis entropy weight combination algorithm, generate standard feature fingerprints corresponding to various typical faults, and perform dimensional fusion and spatial matching between the cable fault risk composite index and the standard feature fingerprints to obtain the fault feature matrix. Step 500: Based on the fault feature matrix, identify cable fault types, locate fault sections initially, and classify fault risk levels using a pre-trained adaptive fault identification and classification model to obtain fault identification results, initial fault location results, and fault risk classification results. Step 600: Based on the initial fault location results, the fault traveling wave head is identified and the traveling wave velocity is dynamically corrected in the fault area by using fault feature fingerprints and multi-parameter constraints to obtain the fault coordinate results.

[0020] In the specific implementation process, step 100 first completes the precise deployment of multiple types of sensors according to the preset spatial layout rules. Specifically, an integrated temperature sensor is deployed along the high-voltage cable route; ultra-high frequency partial discharge sensors, contact temperature sensors, joint grounding current sensors, and ultrasonic sensors are deployed at the intermediate joints of the high-voltage cable; terminal temperature sensors, gas density sensors, and terminal grounding current sensors are deployed at both ends of the high-voltage cable; sheath circulating current sensors, zero-sequence current sensors, and box temperature and humidity sensors are deployed in cross-connection boxes and grounding boxes; and environmental temperature and humidity sensors, water immersion sensors, and hazardous gas sensors are deployed in cable trenches and pipe racks. After completing full-link, full-scenario sensor coverage, the physical parameters to be collected are divided into four categories according to their physical properties: electrical parameters, temperature and strain parameters, acoustic parameters, and environmental parameters. Among these, electrical parameters include partial discharge signals, sheath circulating current sensors, zero-sequence current sensors, and box temperature and humidity sensors. The data includes layer circulating current, zero-sequence current, joint and terminal grounding current; temperature strain parameters including cable distributed temperature and strain, joint temperature and terminal temperature; acoustic parameters including joint partial discharge ultrasonic signals and cable body vibration signals; and environmental parameters including ambient temperature and humidity, water immersion status and harmful gas concentration. Based on the signal characteristics and monitoring requirements of each parameter, a pre-set differentiated sampling strategy is adopted, configuring matching sampling frequencies and sampling accuracies for different types of parameters. A unified time synchronization protocol is used to assign synchronization timestamps to all sensors, enabling simultaneous acquisition of multiple physical quantities of the four types of parameters. Simultaneously, the acquired data is associated and bound with the spatial coordinate information of the corresponding sensors. Finally, all the acquired unprocessed data is structured and integrated and stored according to parameter type, acquisition time, and sensor deployment location, ultimately obtaining a raw dataset for high-voltage cable monitoring containing spatiotemporal attributes and parameter type attributes.

[0021] In the specific implementation process, step 200 first performs spatial location matching on the original dataset based on the unified timestamp and sensor spatial coordinates to obtain a multi-channel signal set. Specifically, the built-in timestamps of the data collected by each sensor in the original dataset are extracted and matched with the preset precise spatial coordinate information of the sensors. A unified timestamp calibration operation is performed on all collected data to map the asynchronous timestamps caused by acquisition delays and clock deviations of different types of sensors to the same preset standard time axis, completely eliminating the problem of time dimension asynchrony. Then, based on the three-dimensional spatial coordinates of each sensor, including the axial coordinates of the high-voltage cable route, the radial coordinates of underground laying, and the elevation coordinates, as well as the full-link topology of the high-voltage cable, the single-channel physical quantity data collected by each sensor is precisely bound to the corresponding cable monitoring point and specific equipment component. Finally, according to the high-voltage cable monitoring section division and physical parameter type, all collected data after time calibration and spatial binding are structured and integrated to form a multi-channel signal set with standard time as the sequence, cable spatial monitoring point as the dimension, and various physical parameters as the channels, thereby ensuring that all data form a strict correspondence in the time and spatial dimensions.

[0022] Next, outlier identification and removal are performed on the multi-channel signal set based on the 3σ criterion and the moving window median, resulting in an outlier-free signal set. Specifically, outlier removal is achieved jointly based on the 3σ criterion and the moving window median, targeting each independent signal sequence {x} in the multi-channel signal set. i |i=1,2,...,n} (n is the number of sampling points for a single signal), first calculate the mean μ and standard deviation σ of the signal sequence, and then identify the samples that satisfy |x i -μ|>3σ is a suspected outlier, mainly caused by factors such as instantaneous sensor failure and strong random environmental interference. A sliding window matching the sampling frequency is configured for each signal channel. The length L of the sliding window is determined by the sampling frequency f, specifically calculated as L=k / f (k is a preset time constant, ranging from 0.1 to 1s). The sliding window slides along the standard time axis with a step size of 1. The median M is calculated for the signal data within each window. For suspected outliers identified by the 3σ criterion, they are not directly removed, but replaced with the median M within their respective sliding window to avoid breakpoints in the signal sequence. A second outlier check is then performed on the replaced signal sequence. After confirming that no residual suspected outliers remain, the signals from all channels in the multi-channel signal set after outlier processing are integrated to obtain a set of signals without outliers, thus ensuring the validity and continuity of the signal sequence.

[0023] Then, the DC drift and power frequency low-frequency baseline drift in the anomaly-free signal set are removed through a fifth-order polynomial fitting algorithm to obtain a baseline-corrected signal set. Specifically, for each continuous signal y(t) in the anomaly-free signal set, it is decomposed into an effective feature signal y s (t) and a baseline drift signal y d (t), and the decomposition relationship is y(t)=y s (t)+y d (t), where the baseline drift signal y d (t) is the interference signal to be removed by fitting; a fifth-order polynomial is used to fit the baseline drift signal, and the fitting formula is y d (t)=a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5 , where a0 is the constant term of the polynomial, a1 to a5 are the fitting coefficients of the first to fifth orders of the polynomial respectively, t is the time variable of the standard time axis, and the optimal fitting coefficients of the fifth-order polynomial are obtained by fitting and solving the sampling data of the anomaly-free signal through the least squares method; then the original anomaly-free signal y(t) is subtracted from the fitted baseline drift signal y d (t) to obtain the signal y c (t) after baseline correction, and the calculation relationship is y c (t)=y(t)-y d (t). The fifth-order polynomial fitting and drift signal subtraction operations are sequentially performed on the signals of all channels in the anomaly-free signal set to completely eliminate the DC drift and power frequency low-frequency baseline drift interference in the signals. After integrating all the processed signals, a baseline-corrected signal set is obtained.

[0024] Finally, the feature enhancement process is performed on the baseline-corrected signal set through an adaptive variational mode decomposition and adaptive wavelet threshold joint denoising algorithm to obtain a multi-parameter synchronous data set. Specifically, the optimal mode number K of the variational mode decomposition (VMD) is determined according to the energy entropy extreme value criterion, and each signal in the baseline-corrected signal set is decomposed into K mutually independent intrinsic mode components IMF k (k = 1, 2,..., K); then the Pearson correlation coefficient r k between each intrinsic mode component IMF k and the corresponding signal of the original data set is calculated. A preset correlation threshold r0 (the value range is 0.3 - 0.5) is set, and the irrelevant mode components with the Pearson correlation coefficient satisfying r k <r0 are removed, and the effective mode components IMF e with strong correlation with the original signal are retained; subsequently, based on the unbiased risk estimation wavelet threshold function, each effective mode component IMF ePerform multi-scale wavelet decomposition and calculate the optimal threshold λ for each wavelet decomposition level. j (j is the wavelet decomposition level), based on the optimal threshold λ j Threshold denoising is performed on the wavelet coefficients at each level. After removing the wavelet coefficients corresponding to noise, the denoised component IMF is reconstructed. d Finally, IMFs are applied to all denoised components of the same signal. d Signal reconstruction is performed to obtain a preliminary denoised signal. Then, combined with the prior information of high-voltage cable fault characteristics, including typical signal characteristic frequency bands and characteristic points of partial discharge and temperature anomalies, the fault characteristic frequency bands of the reconstructed signal are subjected to amplitude enhancement and feature point highlighting processing. The signals of all channels after feature enhancement are finally structured and integrated according to three dimensions: standard time, cable spatial coordinates, and physical parameter type, to obtain a multi-parameter synchronization dataset.

[0025] Furthermore, the core of adaptive variational mode decomposition (VMD) is to adaptively decompose the baseline correction signal into K narrowband, independent intrinsic mode components (IMFs), each IMF revolving around a center frequency and having a finite bandwidth. First, a variational problem is constructed, with the objective function being to minimize the sum of squared bandwidths of all IMF components, and the constraint being that the sum of the IMF components equals the original baseline correction signal y. c (t), for each IMF k (t) is subjected to Hilbert transform to obtain the analytic signal. ,in The Hilbert transform operator is used to correlate the analytic signal with its corresponding center frequency ω. k Complex exponential mixing yields the frequency-shifted analytic signal. The bandwidth of each IMF is measured by calculating the squared L2 norm of the frequency-shifted signal gradient. The variational problem is then solved iteratively using the Alternating Direction Multiplier Method (ADMM), with the iterative update formula including: Modal component update: ; Center frequency update: ; Lagrange multipliers update: ; Among them, u k For the k-th IMF component, ω k Let the center frequency of the k-th IMF be... For Lagrange multipliers, It is a secondary penalty factor (values ​​range from 1000 to 2000). U is the iteration step size (ranging from 0.1 to 0.5), n is the number of iterations, and U k (ω) is u k The Fourier transform of (t) is iterated until the convergence condition is met. (Take 1×10 ), and finally obtain K intrinsic mode functions IMF -6 (t). Subsequently, set the differential threshold of physical parameter types. For electrical parameters (partial discharge, current, etc.), take r0 = 0.4, for temperature strain parameters, take r0 = 0.35, and for acoustic parameters, take r0 = 0.38. Retain the intrinsic mode functions that satisfy r k ≥r0 as the effective mode functions IMF e (t), and eliminate the components with r k <r0 to ensure that subsequent processing focuses on the signal components strongly related to the operating state of high-voltage cables and reduces interference from irrelevant noise. e e

[0026] The optimal threshold is calculated using the unbiased risk estimation (SURE) wavelet threshold criterion. For the high-frequency coefficients of each wavelet decomposition level of each effective mode function IMF e (t), adaptively determine the optimal denoising threshold. First, perform multi-scale discrete wavelet transform (DWT) on each effective mode function IMF e (t), select the db4 wavelet basis, and set the decomposition level J according to the sampling frequency difference: when the sampling frequency f > 1 MHz, J = 7; when 100 kHz ≤ f ≤ 1 MHz, J = 6; when f < 100 kHz, J = 5. Decompose to obtain the high-frequency detail coefficients d j (j = 1, 2,..., J) of each level and the low-frequency approximation coefficient a J of the Jth level. For the high-frequency detail coefficient d j =[d j (1), d j (2),..., d j (N)] (N is the number of coefficients at this level), calculate the optimal threshold λ j using the SURE criterion. The SURE function is defined as , where is the counting function, which counts the number of coefficients whose absolute value does not exceed λ, is the sum of the coefficients that meet the conditions, and m is the coefficient index. Search through the grid in the interval [0, max(|d j |)] to find the λ value that minimizes SURE(λ, d j ), which is the optimal denoising threshold λ j for this level.

[0027] Signal reconstruction and feature enhancement are divided into three stages: threshold denoising, component reconstruction, and prior feature enhancement. First, perform threshold processing on the high-frequency detail coefficients d e (t) of each level of each effective mode function IMF using the soft threshold function. The formula is d j ​j '(m)=sign(d j (m))·max(|d j (m)|-λ j ,0), where sign(·) is the sign function, d j '(m) represents the m-th high-frequency detail coefficient of the j-th layer after denoising; the low-frequency approximation coefficient a of the J-th layer is retained. J The high-frequency detail coefficients d after noise reduction remain unchanged. j 'Approximation coefficient a with low frequency J Perform inverse discrete wavelet transform (IDWT) to reconstruct the denoised component IMF. d (t). IMF of all denoised components corresponding to the same baseline correction signal. d (t) superimposed to obtain the initial denoised signal y d (t)=ΣIMF d (t). Finally, feature enhancement is performed by combining prior information on high-voltage cable fault characteristics: For the characteristic frequency band of 300MHz~1.5GHz of partial discharge ultra-high frequency signal, the characteristic segment of joint temperature abnormality with a temperature rise rate >0.5℃ / min, and the characteristic point of sheath circulating current change, the target feature region is first extracted by bandpass filter or wavelet mode maxima method, the feature region signal is multiplied by gain coefficient G (G is 1.2~1.5, adjusted according to feature intensity), and then the enhanced feature region is combined with the preliminary denoised signal y. d (t) superimposed to obtain the signal with enhanced features; the signals processed by all channels are structured and integrated according to three dimensions: standard time, cable spatial coordinates, and physical parameter type to obtain a multi-parameter synchronization dataset.

[0028] It should be noted that through a progressive data preprocessing operation involving spatiotemporal registration, outlier removal, baseline correction, and feature enhancement, synchronous calibration of the original high-voltage cable acquisition data was achieved in both time and space dimensions, solving the problems of low spatiotemporal synchronization accuracy and poor data correlation in multi-parameter acquisition. Simultaneously, by combining the 3σ criterion with the moving window midpoint, outliers caused by sensor faults and strong random environmental interference were effectively removed while ensuring the continuity of the signal sequence. Furthermore, a fifth-order polynomial fitting method was used to accurately eliminate the masking of effective signals by DC drift and low-frequency power frequency baseline drift. Finally, adaptive variational mode decomposition effectively solved the problems associated with traditional wavelet... To address the common problem of mode aliasing, the baseline correction signal is adaptively decomposed into multiple independent narrowband components. Pearson correlation coefficients are then used to filter out noise-dominant components at their source. Furthermore, unbiased risk estimation wavelet thresholding achieves hierarchical and refined denoising for each component, avoiding the defects of past noise or insufficient denoising caused by fixed thresholds. Combined with feature enhancement operations based on prior physical information of high-voltage cable faults, effective features strongly correlated with cable insulation status, joint temperature anomalies, and current surges are accurately preserved and highlighted. The resulting multi-parameter synchronous dataset possesses both high signal-to-noise ratio and strong feature discriminability, reducing the misjudgment rate in subsequent analyses and improving the distinguishability of fault features.

[0029] Specifically, in step 300, for the discrete-time series signals of the electrical parameter classification subset, features in four dimensions—time domain, frequency domain, time-frequency domain, and nonlinear entropy—are extracted sequentially. All features are bound to the spatial coordinates and time window labels of the corresponding monitoring points. The extracted time-domain features include: peak value (characterizing the impulse amplitude characteristics of the signal), kurtosis (characterizing the non-Gaussian impulse characteristics of the partial discharge pulse), pulse count (the number of pulses with amplitudes exceeding twice the background noise threshold per unit time), rise time (the time interval during which the pulse amplitude rises from 10% to 90% of the peak value), fall time (the time interval during which the pulse amplitude falls from 90% to 10% of the peak value), RMS value, rectified average value, and pulse width (the duration during which the pulse amplitude exceeds the threshold). The extracted frequency domain features include: center frequency (characterizing the frequency domain concentration of signal energy), spectral energy distribution, total harmonic distortion rate (characterizing the degree of waveform distortion of the signal), fundamental frequency ratio (the ratio of fundamental frequency energy to total spectral energy), characteristic frequency band energy ratio (the ratio of energy of partial discharge UHF 300MHz~1.5GHz and sheath current power frequency harmonics 100~1000Hz to total energy), and spectral kurtosis (the kurtosis value of the signal at each frequency point). The extracted time-frequency domain features include: wavelet packet decomposition energy entropy (characterizing the distribution complexity of signal energy in the time-frequency domain), Hilbert-Huang transform marginal spectral features (used to capture the time-frequency evolution of non-stationary fault signals), instantaneous frequency, and instantaneous amplitude features. The extracted nonlinear entropy features include: sample entropy (used to measure the complexity and irregularity of signal sequences), fuzzy entropy, permutation entropy (used to capture the dynamic abrupt changes in signal characteristics), approximate entropy, and multi-scale entropy.

[0030] This study focuses on the temperature and strain time series of a subset of temperature-strain parameters. Extracted time-domain features include: mean temperature (characterizing the overall thermal state of the cable), temperature rise rate (used to capture abnormal upward trends in localized overheating), temperature gradient (characterizing the temperature change per unit length along the cable's direction), extreme axial temperature differences (the difference between the maximum and minimum temperature values ​​at all monitoring points along the cable's axial direction within the same time section), strain change (the maximum amplitude of the strain sequence within a time window, characterizing abrupt changes in mechanical stress within the cable body), and strain fluctuation variance (the squared standard deviation of the strain sequence, used to measure the degree of strain fluctuation). Extracted nonlinear entropy features include: permutation entropy of the temperature sequence (used to capture abnormal fluctuations and abrupt changes in the temperature sequence) and sample entropy of the strain sequence (used to measure the dynamic complexity of the strain sequence).

[0031] For discrete-time series signals of acoustic parameter classification subsets, the extracted time-domain features include: mean amplitude (characterizing the overall level of the acoustic signal), peak value (the maximum amplitude of the sequence, corresponding to the impact intensity of the partial discharge ultrasonic pulse), vibration frequency (the number of vibrations / pulses with amplitude exceeding a preset noise threshold per unit time), pulse duration (the length of time the ultrasonic pulse amplitude exceeds the threshold, characterizing the time-domain morphology of the pulse), and signal energy (characterizing the total energy level of the ultrasonic signal). The extracted frequency-domain features include characteristic frequency components and spectral energy distribution.

[0032] In the specific implementation process, step 400 first uses a convergent cross-mapping algorithm to screen fault causality features in the standardized multi-domain feature set, and removes redundant features and pseudo-correlated features whose causal correlation strength is lower than a preset causal threshold, thus obtaining the core feature set. Specifically, the standardized multi-domain feature set is denoted as F=[f1,f2,...,f M ] T Where M is the total feature dimension, and each feature vector f i =[f i (t1),f i (t2),...,f i (t T [] represents a standardized time series of length T, corresponding to continuous monitoring data at a specific monitoring point on the cable, and is matched with a cable fault state reference sequence Y=[Y(t1),Y(t2),...,Y(t...]]. T Y(t) represents the quantified fault state value of the cable at time t, with a value of 0 for normal operation and a larger value for higher fault severity. The optimal time delay is determined using the mutual information method, and the optimal embedding dimension is determined using the spurious nearest neighbor method. The phase space M is then reconstructed from the fault state sequence Y. Y And for each feature sequence f i The phase space is reconstructed synchronously to obtain the corresponding phase space M. fi For M Y For each phase point in the matrix, find its nearest neighbors with the Euclidean distance of the optimal embedding dimension + 1, record the time index and distance weight of the nearest neighbors, and map this set of nearest neighbors to M. fi In the process, the estimated value Y of the fault state sequence is reconstructed through weighted mapping. i Calculate the estimated sequence Y i The Pearson correlation coefficient with the original fault state sequence Y is the characteristic f. iThe causal correlation strength between the feature and the cable fault state is determined by the coefficient; the closer the coefficient is to 1, the stronger the causal driving force of the feature on the fault state. A preset causal correlation strength threshold is set, ranging from 0.6 to 0.75, with different values ​​depending on the cable voltage level: 0.7 for 110kV and above, and 0.65 for 35kV and below. All features that satisfy the coefficient ≥ the causal correlation strength threshold are retained, while redundant features and spurious correlation features with causal correlation strength below the threshold are removed, ultimately resulting in a K-dimensional core feature set.

[0033] Next, the dynamic combination weights of each core feature in the core feature set are calculated, and a dynamic weight set is constructed. Specifically, the first step is to divide the core feature set into four criterion layers according to physical attributes: electrical features, temperature and strain features, acoustic features, and environmental features. Experts in high-voltage cable operation and maintenance are invited to construct a 9-scale judgment matrix A based on the importance of each criterion layer to cable faults. The consistency ratio CR = CI / RI is calculated, where CI is the consistency index and RI is the average random consistency index. When CR < 0.1, the judgment matrix meets the consistency requirements; otherwise, the judgment matrix needs to be corrected. The eigenvector corresponding to the largest eigenvalue of the judgment matrix is ​​solved using the maximum eigenvalue method. After normalization, the weights of the criterion layers are obtained, and then the subjective weights of each core feature are obtained by further decomposition. The second step is to calculate the weight of the t-th sample of the i-th feature in the core feature set and the information of that feature. Entropy, the smaller the information entropy, the higher the dispersion of the feature and the greater the amount of information for fault differentiation. Then, the objective weight of the i-th feature is calculated using the following formula: The third step is to construct the dynamic weight calculation formula. , This is a dynamic adjustment coefficient, with a value range of 0.3 to 0.7, under normal operating conditions. A value of 0.6 to 0.7 is used, focusing on fault mechanism constraints based on expert experience; under abnormal operating conditions (risk values ​​exceeding the warning threshold). Take a value of 0.3 to 0.4.

[0034] Then, based on the core feature set and dynamic weight set, a composite index of cable fault risk is calculated. Furthermore, by calculating the temporal change rate and future trend of fault risk at specific locations, a fault evolution trend term is introduced into the composite index, resulting in a spatiotemporal sequence of the composite index. Specifically, the calculation formula for the composite index of fault risk is: ; ; ; in Let be the standardized value of the i-th core feature at point L at time t. For the dynamic combination weights of the corresponding features, The rate weighting coefficient is set to 0.4~0.6. The predicted offset weighting coefficient is set to 0.3~0.5. For the rate of change of risk over time, This represents the risk trend offset. The composite index values ​​from all monitoring points of the cable over the entire time series are integrated to obtain a spatiotemporal sequence of the fault risk composite index, with time t as the row and the cable axial spatial coordinate L as the column. , This represents the total number of cable monitoring points.

[0035] Subsequently, a typical fault feature fingerprint database was constructed based on the physical mechanism of cable faults. The core feature set was then calibrated and its dimensions adapted to the typical fault feature fingerprint database using basic cable parameters to obtain a standard feature fingerprint database. Specifically, the typical fault types of the entire high-voltage cable link were first identified, including six categories: intermediate joint insulation degradation faults, cable body overheating faults, sheath grounding system faults, cable terminal sealing / insulation faults, partial discharge insulation breakdown faults, and cable body mechanical strain faults. For each type of fault, its physical evolution mechanism and multi-parameter characteristic response laws were analyzed. For example, the physical mechanism of intermediate joint insulation degradation faults is insufficient joint interface pressure and partial discharge caused by insulation moisture. The accompanying characteristic responses include an increase in the UHF partial discharge pulse count, a faster rate of temperature rise in the joint contact area, an increase in the joint grounding current, and an increase in the amplitude of the ultrasonic characteristic frequency band. Then, for each type of typical fault, its core feature set, relative feature change thresholds, inter-feature correlation constraint rules, temporal evolution laws, and spatial distribution characteristics were determined based on the physical mechanism. A basic feature fingerprint vector was constructed, and the basic fingerprint vectors of all fault types were integrated to obtain the typical fault feature fingerprint database. Furthermore, for cables with different voltage levels and cross-sections, the amplitude reference values ​​and threshold ranges of each feature are adjusted. At the same time, it is ensured that the feature dimensions of the basic fingerprint are completely consistent with the dimensions of the core feature set, the reference values ​​of missing dimensions are supplemented, and redundant dimensions outside the core feature set are eliminated, so as to finally obtain a standard feature fingerprint library that is fully compatible with the target cable.

[0036] Finally, the spatiotemporal sequence of the fault risk composite index is spatially matched with the standard feature fingerprint database based on spatial coordinates, and multi-dimensional fusion is performed through a two-way constraint mechanism to obtain the fault feature matrix. Specifically, the first step uses the cable axial spatial coordinate as the sole matching benchmark, and spatially binds the time-series data of each monitoring point in the spatiotemporal sequence of the fault risk composite index with the fault fingerprint of the corresponding point in the standard feature fingerprint database. For example, the monitoring data of the joint point only matches the standard fingerprint of the joint-related fault, and the monitoring data of the cable body point only matches the standard fingerprint of the body fault, completely eliminating the invalid matching interference caused by spatial mismatch and forming spatially corresponding feature-fingerprint matching pairs. The second step uses cosine similarity to calculate the amplitude similarity between the real-time feature vector and the standard fingerprint benchmark vector for static feature amplitude; and uses the dynamic time warping (DTW) algorithm to calculate the temporal similarity between the real-time risk time sequence and the standard fault evolution time sequence for temporal evolution law; finally, the comprehensive matching similarity of each point and each type of fault is obtained. The third step involves constructing an initial fusion matrix using each monitoring point as the basic unit. The matrix's row dimension represents the time series, and its column dimensions represent the core feature dimension, fault risk composite index, and similarity matching of various fault standard fingerprints. The initial matrix is ​​then weighted and corrected using bidirectional constraints. The correction weights for physical mechanism constraints are derived from the subjective weights of the AHP method, while the correction weights for data feature constraints are derived from the objective weights of the entropy weight method. Feature regions that simultaneously satisfy both constraints are weighted more, while feature regions that violate physical mechanism constraints are weighted less. The fourth step involves stitching and integrating the fusion matrices of all monitoring points along the entire cable link in spatial coordinate order to obtain a three-dimensional fault feature matrix. This matrix fully integrates all effective information from the time dimension, spatial dimension, feature dimension, risk dimension, and fault fingerprint matching dimension.

[0037] It should be noted that, through the causal feature screening of the convergent cross-mapping algorithm, redundant and pseudo-correlated features that are not causally related to cable faults are eliminated from the root, reducing the curse of feature dimensionality and improving the fault representation efficiency of the feature set. The dynamic weight set constructed based on the analytic hierarchy process (AHP) entropy weight combination algorithm takes into account both the expert prior experience of the physical mechanism of high-voltage cable faults and the objective information contribution of real-time monitoring data. It can adaptively adjust according to the cable operating conditions, solving the defect that fixed weights cannot adapt to all operating conditions. The constructed fault risk composite index introduces a fault evolution trend term, realizing the transformation from static risk assessment to dynamic risk assessment. The upgraded state evolution early warning system can capture weak evolution signals of early cable faults; the standard feature fingerprint library built based on physical mechanisms provides rigid physical constraints for fault analysis, solving the problems of lack of interpretability and susceptibility to environmental interference and misjudgment in pure data-driven models; finally, the multi-dimensional fusion and spatial matching completed through the physical mechanism-data feature bidirectional constraint mechanism generates a fault feature matrix with high information density, strong physical interpretability and strict spatiotemporal correlation, which greatly improves the accuracy, anti-interference ability and engineering practicality of subsequent fault type identification, initial section location and risk classification from the feature fusion level.

[0038] In the specific implementation process, step 500 first constructs an adaptive fault identification and classification model. The model, from input to output, sequentially sets up a physical mechanism embedding layer, a spatial topology attention branch, a temporal evolution attention branch, a dual-branch feature fusion layer, an adaptive classification and localization layer, and a dynamic risk classification layer. The core of the physical mechanism embedding layer is to rigidly fuse fault physical priors with data features, first constructing a fault physical rule constraint matrix. J represents the total number of typical fault types. Matrix elements represent the physical correlation between the d-th feature and the j-th fault type, with a strong correlation of 1, a weak correlation of 0.3, and no correlation of 0. The input feature matrix is ​​then compared with the normalized physical correlation weight matrix. We perform a weighted Hadamard product to obtain the feature matrix embedded with the physical prior. Spatial topology attention branch constructs an adjacency matrix based on cable laying topology. The matrix elements represent the topological correlation between monitoring points. A multi-head graph attention network (GAT) is used to calculate the attention weight of each point with its neighboring points, and the final output is a spatial fault attention weight map. The temporal evolution attention branch employs a bidirectional long short-term memory (Bi-LSTM) network combined with a multi-head self-attention mechanism. First, the Bi-LSTM captures the long and short-term temporal dependencies of features. Then, the self-attention mechanism enhances the weights of fault mutation nodes, ultimately outputting a temporal fault evolution feature map. The dual-branch feature fusion layer employs a gated weighted fusion mechanism, first... and Linearly map to the same feature dimension, then calculate the fusion gate weights. ,in It is the sigmoid activation function. and The weights and biases of the gating network are used to ultimately obtain the fault depth features. The adaptive classification and localization layer is a multi-task output layer with a built-in dual-constraint reasoning mechanism of physical mechanism and data features. The classification branch outputs the probability distribution of various types of faults at each point, and the localization branch outputs the probability distribution of faults in each cable section. The dynamic risk classification layer completes the quantitative classification of fault risk based on fault depth features, risk composite indicators, and localization identification results.

[0039] Next, the adaptive fault identification and classification model was pre-trained in three stages. The first stage was pre-training under basic operating conditions. The training dataset consisted of long-term normal operation data (≥12 months) of the same type of cable at the voltage level of the target cable and conventional environmental interference data, including lightning interference, load fluctuations, changes in ambient temperature and humidity, and on-site electromagnetic interference, with an effective sample size of ≥1 million. The training objective was to enable the model to learn the characteristic distribution patterns of cable under normal operating conditions and to have the basic feature extraction ability to distinguish between normal operating conditions and environmental interference. During the training process, the classification and localization layer and the risk classification layer of the model were frozen, and only the physical mechanism embedding layer, the dual-branch attention layer, and the feature fusion layer were trained. The loss function was the mean squared error loss, the optimizer was AdamW, the initial learning rate was set to 1e-4, the batch size was 64, and the training epochs were 50. Early stopping was triggered when the loss on the validation set decreased by less than 1e-5 for 5 consecutive epochs, thus completing the pre-training of the basic feature extraction ability. The second stage is pre-training for fault mechanism transfer. The training dataset uses multi-physics simulation data of cable faults and laboratory simulated fault test data, including typical types such as insulation degradation, partial discharge, overheating, and sheath grounding faults. The effective sample size is ≥500,000, and the samples of each type of fault are balanced. The training objective is to enable the model to learn the characteristic representation rules and physical mechanisms of various typical faults, and to have the core capabilities of fault type identification and segment location. During the training process, the classification and localization layer is unfrozen, the basic weights of the physical mechanism embedding layer are frozen, the dual-branch attention layer and feature fusion layer are fine-tuned, and the classification and localization layer is trained with emphasis. The loss function adopts multi-task joint loss, the optimizer adopts AdamW, the learning rate is set to 5e-5, the batch size is 32, and the training rounds are 80. When the fault type identification accuracy on the validation set is ≥98% for 5 consecutive rounds and the segment location accuracy is ≥95% for 5 consecutive rounds, early stopping is triggered to complete the knowledge transfer of fault mechanism. The third stage is small-sample incremental fine-tuning. The training dataset uses historical fault samples from the target cable, abnormal operating condition samples from the same line under the same conditions, and manually labeled suspected fault samples from the field, with a small sample size (≥20 samples per type of fault). The training objective is to solve the domain adaptation problem of the model to the actual field scenario, matching the laying environment, operating conditions, and equipment characteristics of the target cable. During training, the bottom feature extraction layer is frozen, and only the top-level parameters of the classification and localization layer and the risk grading layer are fine-tuned. The loss function is labeled smooth cross-entropy loss, the optimizer is AdamW, the learning rate is set to 1e-5, the batch size is 16, the training epochs are 30, and an early stopping strategy is adopted to avoid overfitting. Training is stopped when the recognition accuracy of field samples on the validation set is ≥99% for 3 consecutive epochs. Finally, the full-process pre-training of the model and the adaptation to the target line are completed.

[0040] Then the fault feature matrix The input is fed into the trained adaptive fault identification and classification model, entering the physical mechanism embedding layer. The model automatically loads the pre-trained physical rule constraint matrix. The input feature matrix is ​​weighted by physical priors, and a physical constraint loss term is used to filter out physically meaningless noise features and pseudo-anomalies, thus achieving a rigid fusion of physical priors and data features, and outputting a feature matrix that embeds the physical mechanism. Then will The data is fed in parallel to the spatial topology attention branch and the temporal evolution attention branch. In the spatial topology attention branch, the model calculates the spatial attention weights of each monitoring point and its adjacent points based on the actual cable topology adjacency matrix using a multi-head graph attention mechanism. Feature weights are enhanced for high-probability fault points and adjacent points strongly correlated with fault points, while weights are attenuated for normally operating points. The output is a spatial fault attention weight map representing the spatial distribution characteristics of the fault. In the temporal evolution attention branch, the model captures the long-term and short-term temporal dependencies of each feature point through Bi-LSTM, and then uses a multi-head self-attention mechanism to enhance the weights of the time nodes of fault abrupt changes and the time segments with obvious degradation trends, outputting a temporal fault evolution feature map that characterizes the temporal evolution law of the fault. After that, and The input is fed into a dual-branch feature fusion layer. The model first aligns the dimensions of the two branches of features through a linear transformation. Then, through a gated weighted fusion mechanism, the gating weights are adaptively calculated based on the spatial and temporal information strength of the input features. The spatial and temporal features are then weighted and fused to generate a high-dimensional fault depth feature that simultaneously integrates fault physical priors, spatial distribution characteristics, and temporal evolution patterns. .

[0041] After that The input is fed into the adaptive classification and localization layer. The classification branch of this layer inputs the fault depth features into the pre-trained fully connected network and outputs the probability distribution of typical fault type J for each monitoring point in the entire cable link through the softmax activation function. Then, physical mechanism constraints are introduced to zero out the probability values ​​that do not conform to the fault-spatial location correspondence. For example, the probability of insulation degradation fault at intermediate joints is only retained at joint points, while the probability of this type of fault at cable body points is forcibly zeroed out. The probability of sheath grounding system faults is only retained at cross-connection boxes and grounding boxes, while the probability at other points is forcibly zeroed out. After correction, the fault type corresponding to the maximum probability at each monitoring point is taken as the fault type identification result for that point. At the same time, the fault identification probability threshold is set to 0.7. Only when the maximum probability value is ≥0.7 is the corresponding type of fault determined to exist at that point; otherwise, it is determined to be a normal operating condition. Finally, the identification results of all points in the entire link are integrated to obtain a complete fault type identification result, including fault type, fault location, and fault occurrence probability.

[0042] Subsequently, initial fault location inference is performed. Based on the actual cable laying structure, the entire cable link is divided into multiple continuous monitoring segments. Each segment contains a fixed number of monitoring points. The segment division follows the principle of "the same cable body, the same set of joints / terminals, and the same set of cross-connection boxes" to ensure that each segment is an independent physical unit. The location branch, based on the fault depth characteristics and the fault probability distribution of all points in the entire link, outputs the fault occurrence probability of each segment through a fully connected network. Furthermore, by introducing data feature constraints, and based on the mean, maximum, and spatial continuity of the fault probability of all monitoring points within the section, a second correction is made to the section's fault probability. The correction formula is as follows: ,in This represents the maximum probability of a fault at a point within the s-th segment. This represents the mean probability of point failure within the section. The spatial continuity of the fault locations within the section is determined. After correction, the section corresponding to the maximum probability is taken as the initial fault location result. At the same time, the starting and ending spatial coordinates of the fault section and the distribution of fault locations within the section are output to complete the initial location of the fault section.

[0043] Finally, the fault depth features, fault risk composite index, fault type identification results, and initial fault location results are simultaneously input into the dynamic risk classification layer to complete the quantitative calculation and classification of fault risk levels, thus obtaining the fault risk classification results. Specifically, the basic risk classification score is first calculated. ,in The maximum probability value for fault type identification is then used to introduce a cable section importance coefficient. This coefficient is determined based on the power supply importance, load level, and fault impact range of the cable section. The coefficient is 1.2 for Level 1 important sections (power supply to the urban core area, dedicated lines for important users), 1.0 for Level 2 important sections, and 0.8 for Level 3 important sections. High-importance sections are then weighted by risk to obtain a comprehensive risk classification score. Subsequently, in accordance with power industry standards and cable operation and maintenance specifications, the risk level was divided into four levels: Level I (emergency risk). ∈[0.8,1.0] indicates that a fault has occurred or insulation breakdown is about to occur, requiring immediate power outage for repair; Level II (major risk). ∈[0.6,0.8), indicating a significant trend of fault deterioration, requiring on-site verification and rectification within 24 hours; Level III (general risk). ∈[0.3,0.6), indicating early signs of degradation, requiring inclusion in the recent maintenance plan; Level IV (normal condition). ∈[0,0.3), indicating that the cable is in normal operation and there is no obvious risk of failure; finally, the comprehensive risk rating score, corresponding risk level, and risk handling suggestions for each fault point and fault section are output to complete the fault risk level classification.

[0044] It should be noted that the adaptive fault identification and classification model constructed in step 500 achieves a rigid fusion of fault physical priors and data features through a physical mechanism embedding layer, solving the problems of black box nature, lack of physical interpretability, and susceptibility to environmental interference and misjudgment in traditional machine learning models. The dual-branch attention architecture simultaneously captures the spatial topological correlation and temporal evolution law of cable faults, significantly improving the model's ability to perceive early weak fault features and its anti-interference ability under complex working conditions. The three-stage progressive pre-training strategy solves the problems of scarce field fault samples and insufficient model generalization ability of high-voltage cables, achieving accurate adaptation from general working conditions to target lines. The fault type identification and initial segment localization based on the dual-constraint reasoning mechanism takes into account the rigid constraints of physical mechanisms and the feature information of real-time data, significantly improving the accuracy of fault type identification and the precision of initial fault segment localization, and can quickly locate the fault target area. The dynamic risk classification layer combines fault risk composite indicators, evolution trends, and line importance to achieve refined and differentiated classification of fault risks, providing accurate decision-making basis for cable operation and maintenance.

[0045] In the specific implementation process, step 600 sorts all candidate fault sections output in step 500 according to the probability of fault occurrence from high to low, selects the section with the highest probability as the initial target fault section, and simultaneously retrieves all basic parameters of the target section, including the rated voltage, insulation type, cross-sectional parameters, design total length L of the cable body of the section, the monitoring point numbers of the first end S and the last end M of the section, three-dimensional spatial coordinates, and the sampling frequency and synchronization timestamp configuration of the traveling wave acquisition channel; then, combined with the fault type identification results, it retrieves the standard feature fingerprint of the corresponding fault type and extracts the multi-feature fingerprint of the fault type. The parameter characteristic response law is used to perform a secondary verification of the matching degree between the multi-parameter synchronous data of all monitoring points in the target section and the standard feature fingerprint. If the matching degree is lower than the preset threshold of 0.7, it is judged as an initial positioning misjudgment. The next high-probability section is selected for verification in sequence until the matching degree meets the requirements. After the verification is passed, the original voltage and current traveling wave signals of 5 power frequency cycles before and after the fault time are retrieved from the first and last ends of the target section. It is confirmed that the time synchronization accuracy of the double-end traveling wave signal is ≤100ns, which meets the time synchronization requirements of double-end traveling wave positioning. Finally, the target fault section that has passed the verification is locked.

[0046] Next, for the original traveling wave signals at the beginning and end, the adaptive variational mode decomposition and adaptive wavelet thresholding joint denoising algorithm from step 200 is reused to remove power frequency interference, on-site electromagnetic noise, and reflection clutter, resulting in a denoised clean traveling wave signal. Subsequently, combining the traveling wave characteristic frequency band, instantaneous amplitude jump threshold, and phase change characteristic constraint corresponding to the fault type in the fault feature fingerprint, Hilbert-Huang transforms are performed on the denoised traveling wave signals. First, the traveling wave signal is decomposed into multiple intrinsic mode components through empirical mode decomposition, and effective components matching the fault traveling wave characteristic frequency band are selected. Hilbert transforms are then performed on the effective components to obtain the continuous instantaneous amplitude sequence and instantaneous phase sequence of the signal. After that, the traveling wave is set... The wavefront jump determination rule uses three times the average steady-state amplitude of the traveling wave within one power frequency cycle before the fault occurs as the jump threshold. When the instantaneous amplitude increases beyond the jump threshold within three consecutive sampling points, and the instantaneous phase synchronously exhibits a π-step abrupt change, this time point is marked as a candidate point for the traveling wave front, generating candidate sets for the first and last wavefronts respectively. Combining the traveling wave propagation timing constraints and reflection and refraction patterns in the fault feature fingerprint, interference wavefronts are eliminated from the candidate sets. Then, combining the wavefront amplitude and phase characteristics in the fault feature fingerprint, the remaining candidate wavefronts are precisely matched to finally determine the accurate time when the initial traveling wave of the fault reaches the first and last ends, and the time difference between the two-end traveling waves is calculated.

[0047] Then, based on the multi-parameter synchronous dataset, an adaptive correction formula for the traveling wave velocity under multi-parameter constraints is constructed to calculate the corrected traveling wave velocity. The correction formula is as follows: ,in This refers to the theoretical traveling wave velocity of the cable under standard operating conditions at 20℃ and without insulation aging. This represents the real-time average temperature of the cable within the faulty section. The standard operating temperature is 20℃. These are the characteristic values ​​of cable insulation aging within the faulty section. This represents the dispersion characteristic value of the traveling wave signal. The distance from the fault point to the beginning of the section is calculated based on the time difference between the two ends of the traveling wave and the corrected traveling wave velocity.

[0048] Finally, the GIS geospatial database of the entire cable path in the target fault section is retrieved. This database contains continuous sampling point data of the actual cable laying path. Each sampling point corresponds to a unique cumulative line length and three-dimensional geographic coordinates. It also marks the locations of key equipment such as cable joints, bends, and cross-connection boxes, forming a mapping relationship of "cumulative length - geographic coordinates" that perfectly matches the actual cable laying path. Then, the calculated distance from the fault point to the beginning end is matched with the cumulative line length sequence in the GIS database to find the two closest adjacent sampling points. The precise three-dimensional geographic coordinates of the fault point are calculated using linear interpolation. At the same time, the three-dimensional geographic coordinates of the fault point, line location markings, fault type, fault risk level, and positioning error assessment data are integrated to form the final fault coordinate result.

[0049] It should be noted that step 600, based on the initial fault location results of step 500, significantly narrows the target range of traveling wave analysis, avoids clutter interference and redundant calculations in full-line traveling wave analysis, and achieves precise focusing of fault location. By combining the Hilbert-Huang transform wavefront identification method with fault feature fingerprints, false wavefronts caused by traveling wave reflections and on-site electromagnetic interference are effectively eliminated, accurately locking the arrival time of the initial traveling wave of the fault, solving the problems of easy interference and low accuracy of traditional traveling wave wavefront identification. The constructed multi-parameter constrained traveling wave velocity adaptive correction formula achieves dynamic and accurate correction of the traveling wave velocity, eliminating the systematic positioning error caused by the traditional fixed wave velocity method which does not consider the real-time operating conditions of the cable. Through double-end traveling wave positioning... The linear interpolation conversion between the formula and GIS geospatial data enables precise mapping of fault points from relative distance to three-dimensional geographic coordinates, improving the fault location accuracy of high-voltage cables from tens of meters to within meters. Simultaneously, the entire process, along with the preceding multi-parameter synchronous acquisition, data preprocessing, multi-domain feature extraction, and initial fault identification and location stages, forms a complete technical closed loop. This achieves fully automated processing of high-voltage cables from comprehensive monitoring across the entire link, fault type identification, initial section location, to precise fault point coordinate location. This significantly improves the accuracy, reliability, and engineering practicality of high-voltage cable fault location, effectively shortens the time for fault point investigation and repair, and reduces the economic losses and power grid safety risks caused by cable fault power outages.

[0050] like Figure 2 As shown, the present invention also provides a comprehensive monitoring and fault location system for high-voltage cables, comprising: The data acquisition module is used to set up multiple types of sensors around the high-voltage cable according to the preset spatial layout rules, and to collect multiple physical quantities synchronously through the multiple types of sensors to obtain the raw dataset. The preprocessing module is used to perform spatiotemporal registration, outlier removal, baseline correction, and signal enhancement on the original dataset to obtain a multi-parameter synchronous dataset. The classification and feature extraction module is used to classify multi-parameter synchronization datasets based on the physical properties of the parameters, and to extract multi-dimensional and multi-domain features from the classification results to obtain a standardized multi-domain feature set. The standardized multi-domain feature set includes: time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear entropy features. The index calculation module is used to construct a composite index of cable fault risk based on a standardized multi-domain feature set and through the hierarchical analysis entropy weight combination algorithm, and generate standard feature fingerprints corresponding to various typical faults. The composite index of cable fault risk is then dimensionally fused and spatially matched with the standard feature fingerprints to obtain a fault feature matrix. The fault identification module is used to identify cable fault types, initially locate fault sections, and classify fault risk levels based on the fault feature matrix and a pre-trained adaptive fault identification classification model, so as to obtain fault identification results, initial fault location results, and fault risk classification results. The fault location module is used to identify the traveling wave head and dynamically correct the traveling wave velocity of the fault area based on the initial fault location results, through fault feature fingerprints and multi-parameter constraints, to obtain the fault coordinate results.

[0051] The beneficial effects of this invention are as follows: 1) This invention completes the full-scene sensor deployment along the high-voltage cable body, intermediate joints, terminals, cross-connection boxes, grounding boxes and laying environment, and synchronously collects four major categories of core parameters: electrical, temperature strain, acoustic and environmental. Through differentiated sampling strategies and unified timestamp synchronization mechanisms, it achieves seamless perception of the cable's operating status throughout the entire operating cycle, the entire link space and multiple physical dimensions, and solves the core defects of existing technologies such as single monitoring dimensions, incomplete coverage and poor spatiotemporal synchronization from the source. 2) Spatiotemporal registration was used to unify the time axis and bind the spatial location of multi-source data. The 3σ criterion and the sliding window mean method were used to remove outliers while ensuring signal continuity. The DC and power frequency baseline drift was eliminated by fifth-order polynomial fitting. Finally, the adaptive variational mode decomposition and adaptive wavelet thresholding combined denoising algorithm were used to remove noise interference while accurately preserving and enhancing the weak feature signals of early faults. This solved the problems of high noise, baseline drift and masking of effective features in field-acquired data. It also enabled the construction of a high signal-to-noise ratio, high consistency multi-parameter synchronous dataset, effectively avoiding the interference of low-quality data on subsequent analysis.

[0052] 3) Based on the signal characteristics of different physical property parameters, features in four dimensions—time domain, frequency domain, time-frequency domain, and nonlinear entropy—are extracted in a differentiated manner, comprehensively covering the full-dimensional characterization of various faults such as insulation degradation, overheating, and abnormal mechanical stress in high-voltage cables. At the same time, the convergent cross-mapping algorithm is used to complete the screening of fault causal features, eliminating redundant features and pseudo-correlated features that are not causally related to the fault. This solves the problems of incomplete single-domain feature extraction and redundant features interfering with the analysis results in traditional techniques. The constructed standardized multi-domain feature set has both strict physical interpretability and strong fault discrimination ability, reducing the misjudgment rate of subsequent fault identification from the feature level. 4) Combining the expert prior experience of the analytic hierarchy process (AHP) with the objective data information of the entropy weight method, a dynamic weight set that adapts to the working conditions is constructed to calculate the composite index of fault risk. At the same time, a fault evolution trend term is introduced to capture the temporal change rate and future evolution trend of risk. This solves the problem that traditional static risk assessment cannot identify early and weak fault hazards. It can provide early warning of early faults such as cable insulation deterioration and joint overheating, providing accurate quantitative decision-making basis for condition-based maintenance and defect elimination of high-voltage cables, and significantly reducing the probability of sudden faults. 5) The designed model incorporates a physical mechanism embedding layer, achieving a rigid fusion of fault physical priors and data features, completely solving the defects of traditional machine learning models such as being black-boxed, lacking physical constraints, and being susceptible to environmental interference and misjudgment; through a spatial topology + temporal evolution dual-branch attention architecture, it simultaneously captures the spatial distribution correlation and temporal evolution law of faults, improving the ability to perceive early and weak faults; the innovative three-stage progressive pre-training strategy effectively solves the industry problem of scarce field fault samples and poor model field adaptability of high-voltage cables, achieving high accuracy, strong generalization, and high anti-interference fault type identification and initial segment localization; 6) Accurate identification of traveling wave fronts is achieved by combining fault feature fingerprints, effectively eliminating false wave fronts caused by traveling wave refraction and reflection, and on-site electromagnetic interference; an adaptive correction formula for traveling wave velocity is constructed by using a multi-parameter synchronous dataset, and dynamic correction of wave velocity is achieved by combining real-time cable temperature, strain, insulation aging status, and environmental conditions, eliminating systematic positioning errors caused by fixed wave velocity; finally, through GIS geospatial coordinate mapping, the accurate conversion of fault point from relative distance to three-dimensional geographic coordinates is achieved, which greatly shortens the fault point investigation and maintenance cycle and significantly reduces the economic losses and power grid operation risks caused by cable fault power outages.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0054] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for comprehensive monitoring and fault location of high-voltage cables, characterized in that, Includes the following steps: According to the preset spatial layout rules, multiple types of sensors are set up around the high-voltage cable, and multiple physical quantities are collected synchronously through the multiple types of sensors to obtain the original dataset. The original dataset is subjected to spatiotemporal registration, outlier removal, baseline correction, and signal enhancement processing to obtain a multi-parameter synchronous dataset; The multi-parameter synchronization dataset is classified according to the physical properties of the parameters, and multi-dimensional, multi-domain features are extracted from the classification results to obtain a standardized multi-domain feature set. The standardized multi-domain feature set includes: time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear entropy features. Based on the standardized multi-domain feature set, a composite index of cable fault risk is constructed by the hierarchical analysis entropy weight combination algorithm, and standard feature fingerprints corresponding to various typical faults are generated. The composite index of cable fault risk and the standard feature fingerprints are then dimensionally fused and spatially matched to obtain a fault feature matrix. Based on the fault feature matrix, a pre-trained adaptive fault identification and classification model is used to identify cable fault types, locate fault sections initially, and classify fault risk levels, resulting in fault identification results, initial fault location results, and fault risk classification results. Based on the initial fault location results, the fault region is identified by fault traveling wave head and dynamically corrected by traveling wave velocity through the fault feature fingerprint and multi-parameter constraints, and the fault coordinate results are obtained.

2. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 1, characterized in that, According to a preset spatial layout rule, multiple types of sensors are installed around the high-voltage cable, and multiple physical quantities are simultaneously collected through these sensors to obtain a raw dataset, including: Temperature integrated sensors are installed along the route of the high-voltage cable. UHF partial discharge sensors, contact temperature sensors, joint grounding current sensors, and ultrasonic sensors are installed at the intermediate joints of the high-voltage cable. Terminal temperature sensors, gas density sensors, and terminal grounding current sensors are installed at both ends of the high-voltage cable. Sheath circulating current sensors, zero-sequence current sensors, and box temperature and humidity sensors are installed in the cross-connection boxes and grounding boxes. Ambient temperature and humidity sensors, water immersion sensors, and harmful gas sensors are installed in the cable trenches and pipe corridors. The parameters to be collected are divided into electrical parameters, temperature strain parameters, acoustic parameters, and environmental parameters. The electrical parameters include: partial discharge signal, sheath circulating current, zero-sequence current, joint and terminal grounding current; the temperature strain parameters include: cable distributed temperature and strain, joint temperature, and terminal temperature; the acoustic parameters include: joint partial discharge ultrasonic signal and cable body vibration signal; the environmental parameters include: ambient temperature and humidity, water immersion status, and concentration of harmful gases. The parameters to be collected are synchronously collected according to the preset differential sampling to obtain the original dataset.

3. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 1, characterized in that, The original dataset is subjected to spatiotemporal registration, outlier removal, baseline correction, and signal enhancement processing to obtain a multi-parameter synchronization dataset, including: The original dataset is spatially matched based on a unified timestamp and sensor spatial coordinates to obtain a multi-channel signal set. The multi-channel signal set is used to identify and remove outliers based on the 3σ criterion and the median value within the sliding window to obtain an outlier-free signal set. The length of the sliding window is determined by the sampling frequency, and the outliers are replaced by the median value within the sliding window. The DC drift and low-frequency baseline drift of the power frequency are removed from the set of signals without anomalies by a fifth-order polynomial fitting algorithm to obtain a baseline correction signal set. The baseline correction signal set is subjected to feature enhancement processing using an adaptive variational mode decomposition and adaptive wavelet thresholding joint denoising algorithm to obtain the multi-parameter synchronization dataset.

4. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 3, characterized in that, The baseline correction signal set is subjected to feature enhancement processing using a joint denoising algorithm of adaptive variational mode decomposition and adaptive wavelet thresholding to obtain the multi-parameter synchronization dataset, which includes: The optimal number of modes for variational mode decomposition is determined according to the energy entropy extreme value criterion, and the baseline correction signal set is decomposed into multiple intrinsic mode components according to the optimal number of modes. Calculate the Pearson correlation coefficient between the intrinsic modal components and the original dataset, and filter and remove the intrinsic modal components according to a preset correlation threshold to obtain effective modal components; The optimal threshold of each wavelet decomposition level in the effective modal components is calculated based on the unbiased risk estimation wavelet threshold function, and the effective modal components are then subjected to threshold denoising based on the optimal threshold to obtain denoised components. The signal is reconstructed based on the denoised components, and the reconstructed signal is enhanced with prior information on fault features to obtain the multi-parameter synchronization dataset.

5. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 2, characterized in that, The time-domain characteristics of the electrical parameters include: peak value, kurtosis, pulse count, rise time, fall time, RMS value, rectified average value, and pulse width; the frequency-domain characteristics include: center frequency, spectral energy distribution, total harmonic distortion, fundamental frequency ratio, characteristic frequency band energy ratio, and spectral kurtosis; the time-frequency domain characteristics include: wavelet packet decomposition energy entropy, Hilbert-Huang transform marginal spectral characteristics, instantaneous frequency, and instantaneous amplitude characteristics; the nonlinear entropy characteristics include: sample entropy, fuzzy entropy, permutation entropy, approximate entropy, and multi-scale entropy; The time-domain characteristics of the temperature strain parameters include: mean temperature, rate of temperature rise, temperature gradient, extreme values ​​of axial temperature difference, strain change, and strain fluctuation variance; the nonlinear entropy characteristics include: permutation entropy of the temperature sequence and sample entropy of the strain sequence. The time-domain characteristics of the acoustic parameters include: mean amplitude, peak value, vibration frequency, pulse duration, and signal energy; the frequency-domain characteristics include characteristic frequency components and spectral energy distribution.

6. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 1, characterized in that, Based on the standardized multi-domain feature set, a composite index of cable fault risk is constructed using the hierarchical analysis entropy weight combination algorithm, and standard feature fingerprints corresponding to various typical faults are generated. The composite index of cable fault risk is then dimensionally fused and spatially matched with the standard feature fingerprints to obtain a fault feature matrix, including: The standardized multi-domain feature set is screened for fault causality features by using a convergent cross-mapping algorithm, and redundant features and pseudo-correlation features with causal correlation strength lower than a preset causal threshold are removed to obtain the core feature set. Calculate the dynamic combination weights of each core feature in the core feature set, and construct a dynamic weight set; The cable fault risk composite index is calculated based on the core feature set and the dynamic weight set. By calculating the temporal change rate and future trend of the fault risk at the point, a fault evolution trend term is introduced into the cable fault risk composite index to obtain the spatiotemporal sequence of the fault risk composite index. A typical fault basic feature fingerprint database is constructed based on the physical mechanism of cable faults. The core feature set is then calibrated and adapted in terms of amplitude and dimension by using basic cable parameters to obtain a standard feature fingerprint database. The spatiotemporal sequence of the fault risk composite index is spatially matched with the standard feature fingerprint database based on spatial coordinates, and multi-dimensional fusion is performed through a two-way constraint mechanism to obtain the fault feature matrix.

7. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 1, characterized in that, Based on the fault feature matrix, a pre-trained adaptive fault identification and classification model is used to identify cable fault types, initially locate fault sections, and classify fault risk levels, resulting in fault identification results, initial fault location results, and fault risk classification results, including: An adaptive fault identification and classification model is constructed, which is configured in sequence as follows: physical mechanism embedding layer, spatial topology attention branch, temporal evolution attention branch, dual-branch feature fusion layer, adaptive classification and localization layer, and dynamic risk classification layer; The adaptive fault identification and classification model is pre-trained in three stages. The fault feature matrix is ​​input into the trained adaptive fault identification and classification model. First, it passes through the physical mechanism embedding layer to complete the rigid fusion of physical priors and data features. Then, the spatial topology attention branch and the temporal evolution attention branch are input respectively to obtain the spatial fault attention weight map and the temporal fault evolution feature map. The model is then weighted and fused through the dual-branch feature fusion layer to obtain the fault depth features. The fault type identification and initial fault segment localization are performed on the fault depth features through the dual-constraint reasoning mechanism of the adaptive classification and localization layer to obtain the fault identification result and the initial fault localization result. Based on the dynamic risk classification layer, the risk classification score is calculated and the risk level is divided according to the fault risk composite index, fault evolution rate and cable section importance coefficient, so as to obtain the fault risk classification result.

8. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 7, characterized in that, The adaptive fault identification and classification model is pre-trained in three stages, including: The adaptive fault identification and classification model is pre-trained under basic operating conditions using long-term normal operation data of the cable and conventional environmental interference data. The adaptive fault identification and classification model is pre-trained using fault mechanism transfer data from physical simulation data of cable faults and laboratory simulated fault data. The adaptive fault identification and classification model is fine-tuned using small-sample incremental adjustments based on historical fault samples from the target cable site.

9. The method for comprehensive monitoring and fault location of high-voltage cables according to claim 1, characterized in that, Based on the initial fault location results, the fault region is identified by fault traveling wave front identification and dynamic correction of traveling wave velocity using the fault feature fingerprint and multi-parameter constraints, resulting in fault coordinates, including: The target fault section is determined based on the initial fault location results, and the fault traveling wave signal of the target fault section is retrieved. Based on the fault feature fingerprint, the fault traveling wave signal is subjected to Hilbert-Huang transform to obtain the instantaneous amplitude and instantaneous phase. Then, a candidate set of traveling wave fronts is selected based on the jump point, and interfering wave fronts are eliminated to obtain the double-ended traveling wave time difference between the time when the fault traveling wave signal arrives at the beginning of the segment and the time when it arrives at the end of the segment. Based on the multi-parameter synchronous dataset, an adaptive correction formula for traveling wave velocity with multi-parameter constraints is constructed to calculate the corrected traveling wave velocity. The distance from the fault point to the beginning of the section is calculated based on the time difference between the two-end traveling waves and the corrected traveling wave velocity. The distance from the fault point to the beginning of the fault section is converted into the geographical coordinates of the fault point based on the geospatial coordinates of the fault section, thus obtaining the fault coordinate result.

10. A comprehensive monitoring and fault location system for high-voltage cables, characterized in that, include: The data acquisition module is used to set up multiple types of sensors around the high-voltage cable according to the preset spatial layout rules, and to collect multiple physical quantities synchronously through the multiple types of sensors to obtain the raw dataset. The preprocessing module is used to perform spatiotemporal registration, outlier removal, baseline correction, and signal enhancement on the original dataset to obtain a multi-parameter synchronization dataset. The classification and feature extraction module is used to classify the multi-parameter synchronization dataset according to the physical properties of the parameters, and to extract multi-dimensional and multi-domain features from the classification results to obtain a standardized multi-domain feature set; the standardized multi-domain feature set includes: time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear entropy features; The index calculation module is used to construct a composite index of cable fault risk based on the standardized multi-domain feature set through the hierarchical analysis entropy weight combination algorithm, generate standard feature fingerprints corresponding to various typical faults, and perform dimensional fusion and spatial matching between the composite index of cable fault risk and the standard feature fingerprints to obtain a fault feature matrix. The fault identification module is used to identify cable fault types, initially locate fault sections, and classify fault risk levels based on the fault feature matrix using a pre-trained adaptive fault identification classification model, thereby obtaining fault identification results, initial fault location results, and fault risk classification results. The fault location module is used to identify the traveling wave head and dynamically correct the traveling wave velocity of the fault area based on the initial fault location result, through the fault feature fingerprint and multi-parameter constraints, to obtain the fault coordinate result.

Images