A cable branch box insulation state monitoring and positioning method based on high-frequency pulse injection

Abstract

This invention discloses a method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection. The steps include: generating an injection excitation sequence using a high-frequency excitation source with pre-distortion compensation function; acquiring the response waveforms of each branch through a distributed virtual time synchronization array; inputting the waveform data into a physical information neural network and a sparse representation model, decoupling and extracting multi-dimensional feature fingerprint vectors and high-order singular value operators; associating and mapping the data with a multi-physics fingerprint database through a transfer learning model, outputting an evaluation data packet; comprehensively evaluating the data packet and the response waveforms, constructing a complex frequency domain topology matrix inversion model, extracting the node reflection coefficient tensor and calibrating the wave velocity in real time, outputting the location result; projecting the location result onto a digital twin visualization platform, matching it with an operation and maintenance knowledge graph, and outputting maintenance and handling instructions. This invention enables online monitoring and precise location of the insulation status of multiple branches in a cable branch box, significantly improving the operation and maintenance efficiency and power supply reliability of the distribution network.

Description

Technical Field

[0001] This invention belongs to the field of power cable condition monitoring and fault diagnosis technology, specifically, it relates to a method for monitoring and locating the insulation condition of cable branch boxes based on high-frequency pulse injection. Background Technology

[0002] Cable distribution boxes are critical nodes in urban power distribution networks, connecting main cables to branch lines for end users. They are widely deployed in underground utility tunnels and outdoor cabinet environments. With rapid urbanization, the service life of numerous high-voltage distribution cables continues to increase, leading to increasingly prominent insulation aging problems caused by factors such as temperature cycling, moisture intrusion, and mechanical stress. Distribution box connection nodes are characterized by both electrical stress concentration and weak sealing protection; abnormal insulation conditions in these boxes are often precursors to overall cable insulation failure. Therefore, real-time insulation monitoring and precise fault location for distribution boxes have become crucial requirements for power distribution network operation and maintenance projects.

[0003] Existing cable insulation monitoring technologies mainly include time-domain reflectometry (TDR), partial discharge detection, and frequency-domain dielectric spectroscopy. TDR injects a step or pulse signal into the cable and determines the location of impedance discontinuities based on the reflected waveform. Partial discharge detection captures high-frequency discharge pulses generated by insulation defects and identifies the degree of insulation degradation through pulse characteristic analysis. Frequency-domain dielectric spectroscopy applies a small-amplitude alternating voltage over a wide frequency band and measures the frequency response of the dielectric loss factor to assess the overall insulation condition. All of these methods have been applied in engineering projects to some extent, forming relatively complete single technical routes.

[0004] However, the above methods have significant limitations in multi-branch scenarios within cable distribution boxes. Traditional TDR's single-point resolution is limited by pulse width, making it difficult to effectively distinguish between multiple insulation defects located close together. Furthermore, it lacks methods to handle the multipath reflection superposition effect in branch topologies, leading to ambiguous or even misjudged location results. Partial discharge detection relies on the occurrence of defect discharge events and lacks sensitivity to the slowly evolving insulation aging process, resulting in a high false alarm rate in urban power distribution networks with strong electromagnetic interference. Frequency domain dielectric spectroscopy typically requires power outages, making continuous online monitoring impossible. For multi-branch topologies, the measurement results are a network equivalent superposition of the insulation responses of each branch, failing to distinguish the insulation contributions of each branch and severely limiting spatial location capabilities.

[0005] The fundamental reason for the above limitations is that existing methods generally adopt a single-end excitation-single-end reception or centralized acquisition architecture, which lacks the ability to globally perceive and collaboratively process the topology of multiple branches in the branch box; the excitation signal is affected by frequency-related attenuation in the cable, which limits the effective range; at the same time, existing feature extraction methods do not organically integrate the physical laws of electromagnetic wave propagation with the data-driven model, resulting in feature extraction accuracy and defect location resolution that are difficult to meet the actual operation and maintenance needs. Summary of the Invention

[0006] Currently, the field of cable branch box insulation monitoring faces problems such as difficulty in locating multiple branch topologies, limited effective operating distance due to excitation frequency attenuation, and insufficient accuracy in defect feature extraction. The following invention is proposed to address these issues: A method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection, comprising the following steps: Step S01: Using a high-frequency excitation source with pre-distortion compensation function, generate an encoded high-frequency pulse excitation sequence and inject it into the cable branch box; Step S02: Using a distributed virtual time synchronization array, time-aligned acquisition is performed on the response generated by the injected excitation sequence, and distributed synchronous impulse response waveform data is output. Step S03: Input the distributed synchronous impulse response waveform data into the physical information neural network and the sparse representation model respectively, decouple and output the multidimensional feature fingerprint vector and the high-order singular value operator; Step S04: Based on the multidimensional feature fingerprint vector and the high-order singular value operator, the transfer learning model is used to perform association mapping with the multi-physics fingerprint database to output a health evaluation data package containing insulation health level and evolution trend. Step S05: Combine the health assessment data packet and the distributed synchronous impulse response waveform data to construct a complex frequency domain topology matrix inversion model, extract the node reflection coefficient tensor and calibrate the wave velocity in real time, and output the fault branch code and accurate physical coordinates; Step S06: Project the fault branch code and precise physical coordinates onto the digital twin visualization platform, match them with the operation and maintenance knowledge graph, and output maintenance and handling instructions with topological semantics.

[0007] Furthermore, the high-frequency pulse excitation sequence in step S01 is generated using the maximum length sequence spread spectrum coding method, and the injection frequency band covers 100kHz to 30MHz; the high-frequency excitation source has a built-in pre-emphasis filter, which uses the inverse characteristic of cable frequency-related attenuation to pre-distort the excitation sequence to ensure that the spectral flatness error at the injection port does not exceed ±3dB.

[0008] Furthermore, in step S02, the distributed virtual time synchronization array uses the IEEE 1588 precise time protocol to achieve sub-microsecond node clock synchronization. Each node triggers the acquisition of pulse response waveforms at a sampling rate of not less than 100MHz, and then aggregates and outputs them with node identifiers and precise timestamps.

[0009] Furthermore, in step S03, the physical information neural network embeds the cable distribution parameter transmission line equation as a hard constraint into the network loss function, requiring the output to satisfy the voltage and current propagation equation; the sparse representation model uses a dictionary learning algorithm to construct a dictionary of three types of fault atoms, using the sparse coefficient vector as the basic quantity of the higher-order singular value operator.

[0010] Furthermore, the multi-physics fingerprint database in step S04 is constructed through accelerated aging experiments of cables covering the three-dimensional parameter space of temperature, humidity and mechanical stress, and the dielectric spectrum feature vector is used as the fingerprint of each aging state; the transfer learning model is fine-tuned with few samples for the target cable type, and outputs the health level score and trend prediction confidence interval.

[0011] Furthermore, in step S05, the complex frequency domain topology matrix inversion model uses an ABCD transmission matrix to model each branch, solves the overdetermined equations using the regularized least squares method, and inverts the lumped parameters of each branch; the real-time wave velocity calibration is dynamically corrected based on the transit time measurement results of the reference branch with known length.

[0012] Furthermore, the operation and maintenance knowledge graph nodes in step S06 cover fault types, historical locations, maintenance records, and recommended handling measures, and the edges include two types of associations: time correlation and spatial proximity. After semantic reasoning through the graph neural network, maintenance and handling instructions containing equipment numbers and operation procedures are output in priority order.

[0013] Compared with existing technologies, this invention has the following advantages: First, the pre-distortion compensation technology effectively overcomes the frequency-dependent attenuation of high-frequency pulses in cables, ensuring the effective coverage and operating distance of the excitation signal in a wide frequency band; Second, the distributed virtual time synchronization array realizes multi-node sub-microsecond collaborative acquisition, completely solving the spatial blind zone problem under multi-branch topology; Third, the physical information neural network introduces the physical constraints of transmission lines into the feature extraction process, significantly improving the accuracy and interpretability of feature decoupling; Fourth, the combination of complex frequency domain topology matrix inversion and real-time wave velocity calibration realizes sub-meter-level accurate physical positioning of fault points in multi-branch networks; Fifth, the deep integration of digital twins and knowledge graphs enables maintenance and handling instructions to have topological semantics, effectively supporting the intelligentization and standardization of operation and maintenance decisions. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a flowchart illustrating step S03 of the present invention. Detailed Implementation

[0015] The following detailed description, in conjunction with the accompanying drawings, illustrates specific embodiments of the present invention to enable those skilled in the art to more clearly understand the invention. A method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection is proposed, the steps of which are as follows: Step S01: Generation of the encoded high-frequency pulse excitation sequence and injection of pre-distortion compensation. High-frequency pulse injection is the core operation at the excitation source end of this method. The high-frequency excitation source consists of a direct digital synthesis (DDS) module and a broadband power amplifier, capable of generating controlled broadband excitation signals within the 100kHz to 30MHz frequency band. The encoded high-frequency pulse excitation sequence is generated based on the maximum length sequence (M-sequence) spread spectrum coding method: assuming the order of the feedback polynomial is... The sequence length is The autocorrelation function of the sequence approximates the impulse function, exhibiting excellent noise suppression and multipath resolution capabilities. The cross-correlation characteristics of the M-sequence ensure that it can effectively deconvolve the responses of each path even under multi-branch reflection superposition environments, overcoming the multipath aliasing problem of traditional single-pulse TDR in branch topologies.

[0016] Because cable insulation exhibits frequency-dependent attenuation characteristics for high-frequency components, directly injecting an uncompensated sequence will severely attenuate the high-frequency components of the far-end response, reducing the resolution of insulation feature extraction. Pre-distortion compensation is achieved through a built-in pre-emphasis filter: after the excitation sequence is generated, the sequence is spectrally shaped based on the inverse characteristics of the cable transfer function, ensuring that the output spectral flatness error at the injection port is controlled within ±3dB. Let the frequency domain estimation of the cable transfer function be... The frequency response of the pre-emphasis filter is designed as follows:

[0017] in For regularization parameters, to prevent... Excessive magnification occurs in extremely small areas. The frequency response is estimated from the injection-receiver frequency response data in the initial offline test and is automatically updated with the latest offline data before each measurement begins.

[0018] The excitation sequence is injected into the main incoming terminal of the cable branch box in a differential mode via a coupling network. The injection power does not exceed 5% of the cable's rated insulation withstand level to ensure a safety margin for non-invasive detection. Simultaneously, the excitation trigger signal is broadcast to each distributed acquisition node in step S02, serving as a unified time reference for multi-node synchronous acquisition. The coded high-frequency pulse excitation sequence generated and injected into the cable branch box in this step is denoted as follows: Frequency domain representation is All data are completely saved for use in subsequent steps S02 (response deconvolution) and S05 (complex frequency domain inversion).

[0019] Step S02: Distributed time alignment acquisition based on virtual time synchronization array The distributed virtual time synchronization array consists of high-speed acquisition units deployed at the ends of each branch and key intermediate nodes in the branch box. These units are interconnected via Ethernet to form a logical clock synchronization array. Clock synchronization adopts the IEEE 1588 Precision Time Protocol (PTP): the master clock node is located at the injection end, and each slave node measures the link propagation delay and corrects its local clock deviation through a two-step PTP handshake assisted by hardware timestamps. After iterative convergence, the residual time deviation of each node relative to the master clock can be stabilized within 100ns; considering a wave velocity of approximately... The positioning error corresponding to this synchronization accuracy does not exceed 20mm, which meets the spatial resolution requirements at the branch box scale.

[0020] The coded high-frequency pulse excitation sequence generated and injected into the cable branch box in step S01 is the injection excitation sequence described in this step. The signals acquired by each acquisition node are the response waveforms generated by the injected excitation sequence in the insulation medium of each branch. Each acquisition node starts waveform acquisition immediately after receiving the trigger signal broadcast by the injection end, with a sampling rate of not less than 100MHz, and records the response window for a duration not less than twice the round-trip delay of the maximum branch electrical length. The acquisition unit has a built-in programmable gain amplifier (PGA) that adaptively adjusts the gain according to the peak value of the previously acquired signal to cope with the dynamic range differences in signal amplitude under different branch lengths and insulation conditions.

[0021] The waveform data acquired by each node, after local preprocessing (removing DC bias and applying a low-pass anti-aliasing filter with a cutoff frequency of 60MHz), is then compared with the injected excitation sequence. Perform cross-correlation and deconvolution to obtain the impulse response estimate at that node location. :

[0022] in For the first The acquired waveform spectrum of the node. To be the conjugate of the excitation signal spectrum, To prevent division by zero regularization, the deconvolution result is appended with node IDs, physical location coordinates, and a precise PTP timestamp. This data is then aggregated via Ethernet to the central processing unit, forming distributed synchronization impulse response waveform data containing the responses of all distributed nodes. ,in This represents the total number of nodes. This waveform data serves as the input for feature extraction in step S03, and is also fully backed up for use in the complex frequency domain inversion in step S05.

[0023] Step S03: Decoupling of physical constraint neural network and sparse representation parallel features Feature decoupling employs a parallel processing architecture combining a Physical Information Neural Network (PINN) and a sparse representation model. Step S02 outputs distributed synchronous impulse response waveform data. Input two parallel modules respectively.

[0024] The Physical Information Neural Network (PINN) uses a multilayer perceptron as its backbone and receives impulse responses from each node. The time-frequency matrix obtained through short-time Fourier transform is used as input, and the output is a continuous function estimate representing the spatial distribution of dielectric parameters along the cable insulation. The network loss function consists of data fitting terms. With physical constraints It consists of two parts:

[0025] The data fitting term measures the mean square error between the network's predicted waveform and the measured waveform; the physical constraint term incorporates the cable's distributed parameters and transmission line equations.

[0026] The residual penalty is embedded in the loss function, forcing the network output voltage... With current The distribution satisfies the propagation relationship derived from Maxwell's equations, where , , , These represent resistance, inductance, conductance, and capacitance per unit length, respectively. These are the physical constraint weight hyperparameters. After training and convergence, the feature representations output by the hidden layers of the network constitute a multidimensional feature fingerprint vector reflecting the spatial distribution of cable insulation health. Each dimension corresponds to a dielectric characteristic component at a different location and frequency band.

[0027] The sparse representation model is based on the dictionary learning (K-SVD) algorithm. In the offline stage, an atomic dictionary is constructed from the simulation and measured response data of three typical faults (partial discharge, insulation moisture, and space charge accumulation). ,in The length of the sample vector. The number of atoms. The input impulse response vector. Solving for sparse coefficients using Orthogonal Matching Pursuit (OMP):

[0028] in For sparsity constraints. For all The third-order tensor formed by the responses of each node ( Calculate Higher-Order Singular Value Decomposition (HOSVD) for each time frame, obtaining the singular value set of the kernel tensor and the expansion matrix of each mode, which are then encapsulated into a Higher-Order Singular Value Operator. This captures the spatial correlation and temporal evolution information among multi-node responses. Multidimensional feature fingerprint vector. With higher-order singular value operators They are then transferred to step S04.

[0029] Step S04: Evaluation of the correlation between the transfer learning model and the multiphysics fingerprint database on insulation health Health assessment is achieved based on the association mapping mechanism between the transfer learning model and the multi-physics fingerprint database.

[0030] A multiphysics fingerprint database was constructed offline through accelerated cable aging experiments. Orthogonal experiments were arranged according to Box-Behnken design within a three-dimensional parameter space of temperature (20℃ to 90℃), humidity (30%RH to 95%RH), and mechanical stress (strain corresponding to 0 to twice the rated bending radius). Broadband dielectric spectrum measurements were performed on cable samples aged under each parameter combination, covering a frequency range from 0.1Hz to 10MHz. The dielectric loss factor was used as the basis for the measurement. The frequency response curve in The sampled values ​​at the frequency points constitute 3D fingerprint vector To form a multi-physics fingerprint database ,in This is the corresponding aging parameter label vector. The total number of samples.

[0031] The transfer learning model uses a Convolutional Long Short-Term Memory (Conv-LSTM) network as its backbone, first applying it to the full fingerprint database. The model is pre-trained to learn a universal representation of insulation aging characteristics. Then, its parameters are fine-tuned using a small number of labeled samples (usually no more than 50 sets of online detection data) for the target cable type to adapt the model to the insulation material characteristics of the target cable. The fine-tuned model receives the multi-dimensional feature fingerprint vector output from step S03. With higher-order singular value operators Calculate the Wasserstein distance between the fingerprint and each aging state feature in the fingerprint database, and output the health level score. And the corresponding confidence interval.

[0032] The insulation health level mapping rules are as follows: Level 1 (Excellent) corresponds to all physical field parameters deviating less than the benchmark value by 10%; Level 2 (Good) corresponds to local parameter deviations between 10% and 25%; Level 3 (Caution) corresponds to a significant increase in the dielectric loss factor in a certain frequency band; Level 4 (Warning) corresponds to simultaneous degradation of multiple parameters with a trend slope exceeding the threshold; Level 5 (Severe Aging) corresponds to any physical field parameter exceeding the specified limit. The evaluation model is also based on... Historical time series data are used to calculate the slope of change, tracking the evolution trend of insulation degradation. A linear extrapolation combined with an exponential correction model is then used to provide an estimated range for the remaining useful life (RUL). Insulation health class. The confidence interval, evolution trend slope, and RUL estimate are encapsulated together into a health assessment data package and passed to step S05.

[0033] Step S05: Accurate Fault Location Driven by Complex Frequency Domain Topology Matrix Inversion and Node Reflection Coefficients The complex frequency domain topology matrix inversion model models the network topology of the cable branch box as an interconnected multi-port system. For each branch, an ABCD transmission matrix is ​​used to describe its frequency domain characteristics:

[0034] in , , , Let be the propagation constant. Characteristic impedance, This is the length of the branch road.

[0035] The impulse response of each node in the distributed synchronization impulse response waveform data output in step S02 Frequency domain data obtained by Fourier transform For observation purposes, in At discrete frequency points, Kirchhoff's current law equations are written for all nodes of the branch box, and a system is constructed based on the propagation constants of each branch. Characteristic impedance and branch length Overdetermined system of equations with unknowns (dimension is) , For the number of nodes, (Total number of unknowns). Solved using Tikhonov regularized least squares method:

[0036] in The regularization parameter is automatically selected using generalized cross-validation (GCV). The inversion results provide the lumped parameters for each branch, thereby allowing the calculation of the equivalent impedance at each node. With reflection coefficient:

[0037] in The system reference impedance (typically taken as 50Ω). Nodal reflection coefficient tensor. Elements with abnormal mid-amplitude values ​​correspond to impedance mismatch nodes, i.e., insulation degradation or fault locations.

[0038] The method for real-time calibration of wave velocity is as follows: using the main incoming line segment as the known physical length. The reference branch, the reflection transit time difference generated at both ends of the injected pulse. Accurate measurement of cross-correlation peak value, real-time wave velocity In this way, the time delay difference of each abnormal node is corrected and converted into physical distance. Combined with the coordinate system of the branch box drawing, the fault branch code (composed of the branch box device ID and the branch sequence number) and the precise physical coordinates in meters are finally output and transmitted to step S06 along with the health evaluation data packet of step S04.

[0039] Step S06: Generation of topological semantic maintenance instructions based on the fusion of digital twin and knowledge graph The digital twin visualization platform is built upon an integrated framework of Building Information Modeling (BIM) and Geographic Information System (GIS), maintaining a real-time mirror of the internal topological 3D model of the branch box and the external utility tunnel spatial model. The fault branch code and precise physical coordinates output in step S05 are transformed (local equipment coordinate system → global geographic coordinate system) and then annotated in real-time on the corresponding 3D model nodes; the health level output in step S04... The thermal gradient chromatogram is superimposed along the cable route, and historical trend curves are displayed in the node pop-up window, forming an intuitive global insulation health status map.

[0040] The operation and maintenance knowledge graph is built on a graph database (Neo4j). Node types cover four categories: cable lines, branch box equipment, historical fault events, and maintenance work orders. Edge types include two semantic associations: "fault of type Y occurred at location X" (time-related edges) and "equipment A and equipment B were put into operation at the same time and have spatial proximity" (spatial proximity edges). Node attributes include equipment specifications, years of operation, and maintenance records. A graph neural network (GNN) uses fault branch codes and health assessment data packets as query inputs to perform multi-hop semantic reasoning on the knowledge graph, retrieves handling experience of similar historical faults, and integrates the graph reasoning results with the model prediction results according to confidence weights to generate a candidate set of handling solutions.

[0041] The final output maintenance and handling instructions are encapsulated in structured JSON, with fields including: fault level code, fault location description (branch number and geographical coordinates), a list of suggested operation procedures (including safety isolation procedures, testing and verification procedures, and power restoration procedures), historical reference work order numbers, and a comprehensive priority score. Maintenance and handling instructions with topological semantics are pushed to the mobile terminals of maintenance personnel via the platform interface, and are overlaid on the corresponding device nodes in the digital twin 3D view using augmented reality (AR) annotations, achieving closed-loop maintenance and handling.

Claims

1. A method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection, characterized in that, Includes the following steps: Step S01: Using a high-frequency excitation source with pre-distortion compensation function, generate an encoded high-frequency pulse excitation sequence and inject it into the cable branch box; Step S02: Using a distributed virtual time synchronization array, time-aligned acquisition is performed on the response generated by the injected excitation sequence, and distributed synchronous impulse response waveform data is output. Step S03: Input the distributed synchronous impulse response waveform data into the physical information neural network and the sparse representation model respectively, decouple and output the multidimensional feature fingerprint vector and the high-order singular value operator; Step S04: Based on the multidimensional feature fingerprint vector and the high-order singular value operator, the transfer learning model is used to perform association mapping with the multi-physics fingerprint database to output a health evaluation data package containing insulation health level and evolution trend. Step S05: Combine the health assessment data packet and the distributed synchronous impulse response waveform data to construct a complex frequency domain topology matrix inversion model, extract the node reflection coefficient tensor and calibrate the wave velocity in real time, and output the fault branch code and accurate physical coordinates; Step S06: Project the fault branch code and precise physical coordinates onto the digital twin visualization platform, match them with the operation and maintenance knowledge graph, and output maintenance and handling instructions with topological semantics.

2. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S01, the encoded high-frequency pulse excitation sequence is generated using the maximum length sequence spread spectrum coding method, and the injection frequency band covers 100kHz to 30MHz; the high-frequency excitation source has a built-in pre-emphasis filter, which uses the inverse characteristic of cable frequency-dependent attenuation to pre-distort the excitation sequence, ensuring that the spectral flatness error of the sequence at the injection port does not exceed ±3dB.

3. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S02, the distributed virtual time synchronization array uses the IEEE 1588 precise time protocol to establish sub-microsecond clock synchronization between each acquisition node; after receiving the excitation trigger signal, each node synchronously acquires the pulse response waveform at a sampling rate of not less than 100MHz, and the waveform data is aggregated and output after being accompanied by node identifiers and precise timestamps.

4. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S03, the physical information neural network embeds the cable distribution parameter transmission line equation as a hard constraint into the network loss function, requiring the voltage and current distribution output by the network to satisfy the propagation equation derived from Maxwell's equations; the sparse representation model uses a dictionary learning algorithm to construct an atomic dictionary for three types of faults: partial discharge, insulation damping, and space charge accumulation, and uses the sparse coefficient vector as the basic quantity of the higher-order singular value operator.

5. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S04, the multiphysics fingerprint database is constructed through cable accelerated aging experiments covering the three-dimensional parameter space of temperature, humidity and mechanical stress, and the dielectric spectrum feature vector is used as the fingerprint of each aging state. The transfer learning model is based on a backbone network pre-trained on a fingerprint database, fine-tuned with a few samples for the target cable type, and outputs a health level score and a confidence interval for trend prediction.

6. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S05, the complex frequency domain topology matrix inversion model models each branch of the branch box as an ABCD transmission matrix, uses the reflection coefficient tensor elements measured at each node as observations, solves the overdetermined equations using the regularized least squares method, and inverts the lumped parameters of each branch; the real-time wave velocity calibration is dynamically corrected based on the transit time measurement results of the reference branch with known length.

7. The method for monitoring and locating the insulation status of cable branch boxes based on high-frequency pulse injection according to claim 1, characterized in that: In step S06, the nodes of the operation and maintenance knowledge graph cover fault types, historical occurrence locations, existing maintenance records, and recommended handling measures, and the edges include two types of associations: time correlation and spatial proximity. After semantic reasoning on the knowledge graph through a graph neural network, maintenance handling instructions with equipment numbers and operation procedures are output according to the handling priority.

Images