UHV cable shielding material performance detection method based on intelligent sensing technology

Abstract

The application discloses a method for detecting the performance of a UHV cable shielding material based on intelligent sensing technology, and belongs to the technical field of UHV cable detection. The method comprises the following steps: embedding a distributed intelligent sensing array into the surface of a shielding layer, and synchronously collecting multi-dimensional original response signals under the condition that the cable is in an energized operation state; performing joint sparse representation on the signals through an edge side preprocessing unit, separating transient impact components and steady-state background components, and triggering an abnormal event marker; extracting relaxation time distribution characteristics and conductivity temperature dependence characteristics from the steady-state background, and constructing an initial feature tensor; using a power frequency magnetic field coupling signal as a reference source to perform spatial phase correction on the initial feature tensor; inputting the corrected feature tensor into a twin asymmetric convolution network, calculating the topological invariant distance between the corrected feature tensor and multiple typical state templates in a health baseline state library, and obtaining a shielding material aging degree quantitative value and an interface defect type probability distribution.

Description

Technical Field

[0001] This invention relates to the field of ultra-high voltage cable testing technology, specifically a method for testing the performance of ultra-high voltage cable shielding materials based on intelligent sensing technology. Background Technology

[0002] Ultra-high voltage (UHV) cable shielding materials operate under complex conditions involving strong electromagnetic fields, alternating temperatures, and mechanical stress. Accurate online assessment of their insulation performance and interface condition is a significant challenge in cable condition-based maintenance. Existing technologies often employ offline sampling or conventional dielectric response testing after power outages. These methods cannot reproduce the true electrothermal coupling stress of cables under energized conditions, missing numerous subtle characteristics that can only be elicited under energized conditions. Some online monitoring solutions collect partial discharge or dielectric loss signals by deploying single-type sensors. Their signal processing links are typically based on fixed frequency domain filtering or threshold comparisons, lacking the ability to analyze the combined structure of multi-source heterogeneous signals. When faced with composite responses resulting from the superposition of different physical processes within the shielding material, this approach easily drowns out weak transient anomalies with strong background noise or power frequency interference, failing to effectively separate the impact components characterizing early degradation or interface defects. In live-line testing environments, the power frequency magnetic field coupling of the UHV cable body introduces complex spatial phase rotation and amplitude modulation into the sensor circuit. This interference is not globally uniform but varies non-uniformly with the cable's axial position and the distribution of stray parameters to ground in the shielding layer. Existing interference suppression methods mostly rely on fixed-coefficient hardware notch filtering or conventional digital filtering, which cannot track and correct this spatially distributed strong electromagnetic coupling disturbance. This causes the acquired dielectric characteristic tensor to be distorted due to spatial phase distortion, leading to deviations in subsequent aging assessment results. Furthermore, existing feature comparison methods mostly use geometric metrics such as Euclidean distance or cosine similarity. These metrics are insensitive to the global topological structure of the feature-embedded manifold and cannot distinguish between homogeneity changes caused by interface defects and scale shrinkage caused by overall thermal aging, thus limiting the accuracy of aging degree evaluation and defect type identification. Summary of the Invention

[0003] The purpose of this invention is to propose a method that can separate weak transient impact components from blind sources in multi-dimensional mixed signals and extract steady-state relaxation characteristics when ultra-high voltage cables are in operation, while effectively compensating for spatially distributed power frequency magnetic field coupling interference, and accurately assessing the aging degree of shielding material and interface defect type based on the topological invariants of characteristic manifolds.

[0004] To achieve the above objectives, this invention provides the following technical solution: This invention discloses a method for testing the performance of ultra-high voltage (UHV) cable shielding materials based on intelligent sensing technology. By embedding a distributed intelligent sensing array on the surface of the UHV cable shielding layer, multi-dimensional raw response signals of the shielding material are simultaneously acquired while the cable is energized, enabling online evaluation of the shielding material's condition. This method involves arranging thin-film impedance sensors and fiber optic temperature sensors along the cable axis at non-equidistant intervals on the shielding layer surface. Each location integrates an impedance measurement electrode and a temperature sensing unit, simultaneously acquiring broadband dielectric response signals, partial discharge pulse signals, and shielding layer temperature distribution signals. The broadband dielectric response signal is obtained by applying a swept-frequency sinusoidal excitation to the impedance measurement electrode using a high-frequency broadband impedance analyzer to acquire the real and imaginary parts of the complex impedance. The partial discharge pulse signal's amplitude, rise time, and equivalent duration are recorded by a coupling capacitance sensor. The temperature distribution signal is obtained by demodulating the wavelength offset using a fiber optic temperature sensor and calibrating the conversion relationship to obtain the instantaneous temperature value. This multi-physical quantity, multi-location synchronous acquisition method provides rich raw information for subsequent feature extraction.

[0005] The acquired multidimensional raw response signals undergo joint sparse representation processing in the edge-side preprocessing unit. First, the signals are aligned along the time axis to construct a multidimensional hybrid signal matrix. Then, a dictionary learning algorithm based on stochastic resonance is used to learn an overcomplete joint dictionary from this matrix. This dictionary learning process introduces a bistable potential function model to enhance weak transient impact components and dynamically updates the dictionary online to adapt to the time-varying characteristics of the signals. The overcomplete joint dictionary contains transient atoms representing transient impacts and steady-state atoms representing steady-state backgrounds. By performing orthogonal matching pursuit decomposition on the dictionary, the transient impact component and the steady-state background component can be separated. Based on the temporal distribution characteristics of transient impact components, the peak arrival time and energy integral of each impact pulse are calculated and mapped to a transient event time series. Then, a sliding window is used to statistically analyze the frequency of impact pulse occurrence and the cumulative energy. When the frequency exceeds the first threshold or the cumulative energy exceeds the second threshold, an abnormal event marker vector containing the start time, duration, and average energy is generated and pushed to the upper-level state assessment module in real time to trigger weighted amplification processing of the corresponding time window data, ensuring that abnormal information is not ignored.

[0006] The relaxation time distribution and conductivity temperature dependence characteristics of the shielding material were extracted from the steady-state background components. A Hough transform was performed on the broadband dielectric response subset in the steady-state background components, converting the curve of the real part of the complex impedance as a function of frequency into multiple characteristic peaks on the relaxation time spectrum. Peak position parameters and peak width parameters were extracted as relaxation time distribution characteristics. Simultaneously, a correlation analysis was performed between the shielding layer temperature distribution signal and the real part of the broadband dielectric response at low frequencies. An exponential relationship between conductivity and temperature was fitted to obtain the activation energy parameter and pre-exponential factor parameter, constituting the conductivity temperature dependence characteristics. The peak position parameters, peak width parameters, activation energy parameters, and pre-exponential factor parameters were arranged according to the sensor placement locations to construct a three-dimensional initial feature tensor, used to characterize the spatial distribution characteristics of the shielding material in its current state.

[0007] To eliminate interference from the power frequency magnetic field during UHV cable operation on feature extraction, an adaptive interference correction algorithm based on a compressed sensing framework is adopted. A miniature induction coil is integrated into the sensor array to acquire the power frequency magnetic field coupling signal in real time and demodulate the instantaneous phase and amplitude fluctuation envelope of the power frequency fundamental wave. Based on this, a spatial phase interference matrix with the same frequency and phase as the interference source is constructed. Each element in the initial feature tensor is considered a complex number, with its real part corresponding to the resistive component in the broadband dielectric response and its imaginary part corresponding to the reactive component. The corresponding elements of the spatial phase interference matrix are also considered complex numbers, with their magnitude and argument determined by the amplitude fluctuation envelope and instantaneous phase, respectively. A point-by-point complex division operation is performed to subtract the phase rotation and amplitude modulation effects introduced by the power frequency magnetic field coupling from the initial feature tensor, obtaining the corrected feature tensor. This correction process effectively suppresses power frequency environmental interference, improves the purity of the feature data, and enhances the accuracy of subsequent state assessment.

[0008] The corrected feature tensor is input into a pre-trained Siamese asymmetric convolutional network for state evaluation. This network includes a fixed-weight feature extraction branch and a learnable-weight metric learning branch. The fixed-weight branch processes typical state templates from the healthy baseline state library, while the learnable-weight branch processes the current corrected feature tensor, yielding corresponding feature embedding vectors. The persistent homology distance between the current state embedding vector and each baseline state embedding vector is calculated as the topological invariant distance. During calculation, a Vitoris-Reis complex is constructed in the point cloud space of the embedding vectors with different scale parameters, recording the generation and extinction times of homology groups in each dimension to form a persistent graph. The Wasserstein distance between the current state persistent graph and the baseline state persistent graph is then calculated. From the resulting set of topological invariant distances, several of the smallest values ​​are selected. These distances correspond to the aging age and interface defect type labels of the typical state templates. The reciprocal of the distance values ​​is then normalized to obtain the weight coefficients. The aging period is weighted and summed with weighting coefficients to obtain a quantitative value for the aging degree of the shielding material. Interface defect types are grouped by type, and the sum of weighting coefficients within each group is used as the probability of each type, resulting in a probability distribution of interface defect types. This evaluation method measures the differences in characteristic spatial structure through topological invariants, is robust to local deformation and noise, and can accurately reflect the aging state of the shielding material and the classification of interface defects.

[0009] Preferably, after completing the abnormal event marking, the edge-side preprocessing unit stores the abnormal event marking vector in the cache and pushes it in real time, so that the upper-layer module can apply weighted amplification to the data of the corresponding time period in the initial feature tensor according to the time window of the event occurrence, ensuring that sudden anomalies such as partial discharge are fully expressed in the feature tensor. As a technical solution of the present invention, the dictionary learning algorithm based on stochastic resonance uses a bistable potential function model to enhance the weak impact component during the iteration process, and dynamically updates the dictionary atoms through an online learning mechanism, so as to maintain separation accuracy under varying operating conditions. The above-mentioned interference adaptive correction algorithm completes spatial phase correction in the complex domain, has high computational efficiency, and directly corresponds to the complex physical meaning of the initial feature tensor, avoiding the loss of amplitude and phase information.

[0010] The method of this invention enables shielding material performance testing to be completed without interrupting cable operation, avoiding the limitations of traditional offline testing that requires power outages and sampling. Furthermore, through multi-dimensional sensing collaboration and interference adaptive correction, the accuracy of aging degree and defect type identification is significantly improved, providing a reliable basis for condition-based maintenance and lifespan management of UHV cables. Simultaneously, the combination of edge-side preprocessing and twin network evaluation automates the process from signal acquisition to condition determination, enabling online real-time evaluation.

[0011] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A distributed intelligent sensor array synchronously acquires multi-dimensional raw response signals, including broadband dielectric response, partial discharge pulses, and shielding temperature distribution, while the cable is energized. An edge-side preprocessing unit performs joint sparse representation and orthogonal matched pursuit decomposition on these signals based on stochastic resonance enhancement. By introducing a bistable potential function model, weak transient impulse components submerged in noise are intentionally resonantly enhanced during dictionary learning. Then, transient and steady-state atoms from an overcomplete joint dictionary accurately reconstruct the multi-dimensional mixed signal, separating the pure transient impulse component from the steady-state background component. This separation mechanism does not require pre-assuming signal bandwidth boundaries or establishing fixed thresholds; it directly utilizes the signal's own time-frequency structure differences to achieve blind source separation. This effectively distinguishes weak partial discharge impulse pulses from high-frequency interference pulses, providing a high signal-to-noise ratio signal foundation for subsequent accurate imaging of the relaxation time spectrum and anomaly event labeling. By utilizing miniature induction coils integrated into a sensor array to capture the power frequency magnetic field coupling signal during the operation of an ultra-high voltage cable in real time, and demodulating the instantaneous phase and amplitude fluctuation envelope of the power frequency fundamental wave, a spatial phase interference matrix that is in phase and frequency-aligned with the interference source is constructed. Each complex element in the initial characteristic tensor is then divided point-by-point with the corresponding complex element in the interference matrix, fundamentally correcting the non-uniform phase rotation and amplitude modulation effects introduced by power frequency electromagnetic coupling, outputting a corrected characteristic tensor with consistent spatial phase. This correction method treats power frequency magnetic field coupling as a measurable, modelable, and point-by-point compensable physical field, no longer relying on statistical assumptions or global averaging. It is particularly sensitive to differences in dielectric properties at different axial positions of the cable, effectively restoring the true dielectric response characteristics of the shielding material at various spatial locations.

[0012] The corrected feature tensors are input into a Siamese asymmetric convolutional network. The current state and the healthy baseline template are mapped to the feature embedding space via a metric learning branch with learnable weights and a feature extraction branch with fixed weights, respectively. Based on this, the persistent homology distance between the Vitoris-Lipps complexes of the point clouds generated by the two embedding vectors is calculated. Persistent homology, as a topological invariant metric, can capture global topological differences in the connectivity and hole structure of feature manifolds at different scales. The Wasserstein distance is used as the quantized topological invariant distance to accurately distinguish between topological changes caused by local interface defects and geometric shrinkage caused by overall material aging. By traversing typical state templates labeled with different aging years and interface defect types in the healthy baseline state library, the aging degree quantification value and defect type probability distribution are obtained by weighting with the inverse of the topological invariant distance, realizing the parallel identification of aging stages and defect patterns. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a flowchart of the performance testing method for ultra-high voltage cable shielding materials; Figure 2 This is a flowchart for evaluating the aging and defects of shielding materials based on twin asymmetric convolutional networks and persistent homology distance; Figure 3 It is a comparison of the spatial distribution of the activation energy parameters of the shielding material along the sensor layout locations before and after calibration. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0016] See Figure 1 This invention provides a method for performance testing of ultra-high voltage (UHV) cable shielding materials based on intelligent sensing technology. The method embeds a distributed intelligent sensing array onto the surface of the UHV cable shielding layer, synchronously acquiring multi-dimensional raw response signals of the shielding material while the cable is energized. An edge-side preprocessing unit performs joint sparse representation on the multi-dimensional raw response signals, separating transient impact components and steady-state background components from the signals. Anomaly event marking is triggered based on the temporal distribution characteristics of the transient impact components. Furthermore, the relaxation time distribution characteristics and conductivity temperature dependence characteristics of the shielding material are extracted from the steady-state background components to construct an initial feature tensor characterizing the current state of the shielding material. An interference adaptive correction algorithm based on a compressed sensing framework is used, employing the power frequency magnetic field coupling signal during UHV cable operation as a reference source to perform spatial phase correction on the initial feature tensor, obtaining a corrected feature tensor. Finally, the corrected feature tensor is input into a pre-trained twin asymmetric convolutional network, calculating the topological invariant distance between the feature tensor and multiple typical state templates in the health baseline state library, obtaining the quantified value of the shielding material's aging degree and the probability distribution of interface defect types.

[0017] Example 1: In practical implementation, a distributed intelligent sensor array is embedded in the surface of the UHV cable shielding layer to synchronously collect multi-dimensional raw response signals of the shielding material while the cable is energized. These multi-dimensional raw response signals include broadband dielectric response signals, partial discharge pulse signals, and shielding layer temperature distribution signals.

[0018] The distributed intelligent sensor array is deployed as follows: multiple thin-film impedance sensors and multiple fiber Bragg grating temperature sensors are embedded along the cable axis at non-equidistant points on the shielding layer surface. Each point integrates both an impedance measurement electrode and a temperature sensing unit. The non-equidistant spacing is determined based on the differences in electric and thermal field distributions in the stress cone region, joint transition region, and straight section of the cable. The spacing is reduced in the stress cone region and joint transition region, and increased in the straight section region. The thin-film impedance sensor employs a structure with interdigitated gold electrodes deposited on a flexible polyimide substrate. The linewidth and spacing of the interdigitated gold electrodes are both set to 50 micrometers, allowing the thin-film impedance sensor to conformally fit the shielding layer surface. The fiber Bragg grating temperature sensor uses a written fiber Bragg grating with a center wavelength of 1550 nanometers and is encapsulated in a thermally conductive flexible adhesive layer, ensuring tight thermal contact between the fiber Bragg grating temperature sensor and the shielding layer surface and insensitivity to stress.

[0019] The process of acquiring the broadband dielectric response signal is as follows: A swept-frequency sinusoidal excitation signal is applied to the impedance measurement electrodes using a high-frequency broadband impedance analyzer. The frequency range of the swept-frequency sinusoidal excitation signal is from 0.01 Hz to 1 MHz, and the excitation voltage amplitude is set to 100 mV to avoid nonlinear polarization interference to the shielding material. The high-frequency broadband impedance analyzer simultaneously acquires the real and imaginary parts of the complex impedance of the shielding material at different frequency points. The real and imaginary parts of the complex impedance at all frequency points are arranged in ascending order of frequency and combined to form a broadband dielectric response signal. The high-frequency broadband impedance analyzer is connected to the impedance measurement electrodes at each location via miniature coaxial shielded wires of equal length to eliminate the influence of lead inductance differences on impedance measurement.

[0020] The process of acquiring partial discharge pulse signals is as follows: Partial discharge pulse waveforms on the surface of the shielding layer are acquired using a coupling capacitor sensor within the same time window. The coupling capacitor sensor consists of a ring-shaped metal foil attached to the outside of the shielding layer, forming a 50-micron-thick polytetrafluoroethylene (PTFE) insulating dielectric layer between it and the shielding layer, constituting a high-frequency coupling capacitor. The 3 dB bandwidth of the coupling capacitor sensor is no less than 100 MHz. The output signal of the coupling capacitor sensor is amplified by 20 dB by a low-noise preamplifier and then sent to a high-speed data acquisition card with a 2 GHz sampling rate. The amplitude, rise time, and equivalent duration of each partial discharge pulse are recorded. The recorded amplitude, rise time, and equivalent duration of each partial discharge pulse are used as the partial discharge pulse signal. The rise time is defined as the time interval from the partial discharge pulse rising from 10% peak value to 90% peak value, and the equivalent duration is defined as the time width obtained by dividing the integral value of the partial discharge pulse energy by the square of the peak value.

[0021] The process of acquiring the shielding layer temperature distribution signal is as follows: Wavelength offsets at various locations on the shielding layer surface are demodulated using fiber Bragg grating temperature sensors. A scanning fiber Bragg grating demodulator is used to acquire the center wavelength of the reflection spectrum of each fiber Bragg grating temperature sensor at a scanning frequency of 1 Hz. The difference between the center wavelength of each fiber Bragg grating temperature sensor in operation and the reference wavelength pre-stored in the demodulator is calculated to obtain the wavelength offset at each location. Based on the pre-calibrated wavelength-temperature conversion relationship, the instantaneous temperature value at each location is calculated. The instantaneous temperature values ​​at each location are arranged in order of sensor location and combined to form the shielding layer temperature distribution signal. The pre-calibrated wavelength-temperature conversion relationship is obtained through a constant-temperature oil bath calibration experiment. The center wavelength of the fiber Bragg grating temperature sensor is recorded every 5 degrees Celsius within the range of 25 degrees Celsius to 120 degrees Celsius. The following formula is obtained by linear fitting using the least squares method:

[0022] in, This indicates the wavelength offset of the fiber Bragg grating temperature sensor, measured in picometers. This indicates the temperature change at the location of the fiber Bragg grating temperature sensor, in degrees Celsius. The value is obtained by subtracting the reference temperature of 25 degrees Celsius from the current temperature. The coefficient for the first term represents the temperature sensitivity of the fiber Bragg grating temperature sensor, and its value, obtained by fitting calibration data, is 10.2 picometers per degree Celsius. The coefficient for the quadratic term is used to compensate for nonlinear effects at high temperatures, and its value, obtained through fitting calibration data, is 0.003 picometers per square degree Celsius. For each fiber Bragg grating temperature sensor, the measured wavelength offset... Substitute into the above formula and solve for... The quadratic equation in one variable is taken, and the positive root that conforms to the physical meaning is added to the reference temperature of 25 degrees Celsius to obtain the instantaneous temperature value at the location of the fiber optic temperature sensor.

[0023] Example 2: In practical implementation, the multi-dimensional raw response signal is jointly sparsely represented by an edge-side preprocessing unit to separate the transient impact component and the steady-state background component from the signal. The edge-side preprocessing unit is deployed on a computing node located at the cable monitoring site. The computing node is connected to the data acquisition module of the distributed intelligent sensor array via an industrial Ethernet to receive broadband dielectric response signals, partial discharge pulse signals, and shielding temperature distribution signals.

[0024] The edge-side preprocessing unit aligns the synchronously acquired broadband dielectric response signal, partial discharge pulse signal, and shielding layer temperature distribution signal on the time axis to construct a multidimensional hybrid signal matrix. The time axis alignment method is as follows: using the second pulse synchronization signal shared by the high-frequency broadband impedance analyzer, high-speed data acquisition card, and fiber optic demodulator as a reference, a timestamp is added to each frame of acquired data. Based on the timestamp, the edge-side preprocessing unit groups the sampling points of the broadband dielectric response signal, the pulse waveform sampling points of the partial discharge pulse signal, and the temperature values ​​of the shielding layer temperature distribution signal within the same millisecond window into the same column of the multidimensional hybrid signal matrix. The rows of the multidimensional mixed signal matrix correspond to different types of signal channels. The first to the Kth rows store the real part of the complex impedance at K frequency points in the broadband dielectric response signal, the (K+1)th to the 2Kth rows store the imaginary part of the complex impedance at K frequency points, the (K+1)th row stores the amplitude sequence of the partial discharge pulse signal, the (K+2)th row stores the rise time sequence of the partial discharge pulse signal, the (K+3)th row stores the equivalent time width sequence of the partial discharge pulse signal, and the (K+4)th to (K+3+M)th rows store the instantaneous temperature values ​​at M sampling points. The columns of the matrix correspond to the time sampling points.

[0025] The edge-side preprocessing unit employs a stochastic resonance-based dictionary learning algorithm to learn an overcomplete joint dictionary from the multidimensional mixed signal matrix. This algorithm enhances the weak transient impulse components in the multidimensional mixed signal matrix by introducing a bistable potential function model. The bistable potential function model is a symmetric potential well system consisting of two stable states and an intermediate unstable potential barrier. Each signal in the multidimensional mixed signal matrix undergoes stochastic resonance preprocessing via the bistable potential function model before being input into the dictionary learning module. This stochastic resonance preprocessing is achieved by numerically solving the Langevin equation. The potential function term in the Langevin equation uses a fourth-order bistable potential function, and the noise intensity is adaptively adjusted according to the signal-to-noise ratio (SNR) of the input signal, thus improving the SNR of the weak transient impulse components at the output of the bistable system. The multidimensional mixed signal matrix after stochastic resonance preprocessing is then input into the dictionary learning module. The dictionary learning module dynamically updates the overcomplete joint dictionary using an online learning approach. The specific process is as follows: The multidimensional mixed signal matrix, after stochastic resonance preprocessing, is divided into consecutive data blocks in chronological order. Each data block contains data from N sampling times, where N is 500. For the first data block, an overcomplete joint dictionary is initialized using the K-SVD algorithm. For each subsequent data block, the overcomplete joint dictionary learned from the previous data block is used as the initial dictionary. A stochastic gradient descent iteration is performed on the initial dictionary using the current data block, updating each atom in the dictionary to adapt to the slow evolution of the shielding material state over time. The overcomplete joint dictionary contains transient atoms representing transient impacts and steady-state atoms representing steady-state backgrounds. Transient atoms are those with sharp time-domain waveforms, short durations, and wide spectra, while steady-state atoms are those with smooth time-domain waveforms, long durations, and narrow spectra. The total number of atoms in the overcomplete joint dictionary is set to four times the number of signal channels, with a transient-to-steady-state atom ratio of 1:3.

[0026] After obtaining the overcomplete joint dictionary, the edge-side preprocessing unit performs orthogonal matching pursuit decomposition on the multidimensional mixed signal matrix using the overcomplete joint dictionary. The sparsity parameter of the orthogonal matching pursuit decomposition is set to 10, and the iteration process stops when the residual energy is less than 1% of the original signal energy. The decomposition process selects the atoms with the highest correlation to the current residual from the overcomplete joint dictionary one by one, and updates the coefficients of the selected atoms using the least squares method. After satisfying the sparsity or residual conditions, the transient impact component reconstructed by the linear combination of transient atoms and the steady-state background component reconstructed by the linear combination of steady-state atoms are obtained.

[0027] Based on the temporal distribution characteristics of the transient impact components, abnormal event markers are triggered. The edge-side preprocessing unit calculates the peak arrival time and energy integral of each impact pulse in the transient impact components. The peak arrival time is defined as the time coordinate corresponding to the maximum absolute value of the impact pulse, and the energy integral is defined as the integral of the square of the impact pulse waveform along the time axis. The edge-side preprocessing unit maps the calculated peak arrival time and energy integral onto the time axis to form a transient event time series. Each element of the transient event time series is a two-dimensional vector containing a peak arrival time and an energy integral value.

[0028] The edge-side preprocessing unit uses a sliding window to scan the transient event time series. The window width is set to 10 milliseconds, and the sliding step size is set to 1 millisecond. Within each sliding window, the occurrence frequency and cumulative energy of the impact pulses are statistically analyzed. The occurrence frequency is defined as the number of impact pulses within the window divided by the window width, and the cumulative energy is defined as the sum of the energy integrals of all impact pulses within the window.

[0029] When the frequency of an event exceeds a preset first threshold or the cumulative energy exceeds a preset second threshold, the edge-side preprocessing unit generates an abnormal event marker vector containing the start time, duration, and average energy. The preset first threshold is set to 100 Hz based on the statistical law of partial discharge monitored over a long period under normal operating conditions of the cable. The preset second threshold is set to 5 picocoulombs squared based on the initial energy level of partial discharge in the shielding material. The start time is taken as the timestamp of the left endpoint of the sliding window when the threshold condition is triggered, the duration is taken as the window width of the sliding window (10 milliseconds), and the average energy is taken as the arithmetic mean of the energy integrals of all impact pulses within the window. The abnormal event marker vector is stored in the edge-side buffer, which is a circular buffer with a capacity set to store the most recent 1000 abnormal event marker vectors.

[0030] The anomaly event marker vector is pushed to the upper-layer state assessment module in real time via asynchronous messages based on the MQTT protocol. The message body contains the binary encoding of the anomaly event marker vector. After receiving the anomaly event marker vector, the upper-layer state assessment module locates the data corresponding to the time window in the initial feature tensor extracted from the steady-state background components based on the start time and duration in the anomaly event marker vector. It then multiplies the values ​​of the initial feature tensor elements within that time window by a weighting coefficient of 1.8 to complete the weighted amplification process.

[0031] Example 3: In practice, the relaxation time distribution characteristics and conductivity temperature dependence characteristics of the shielding material are extracted from the steady-state background components to construct the initial feature tensor of the current state of the shielding material.

[0032] The process of extracting relaxation time distribution features is as follows: A broadband dielectric response subset is extracted from the steady-state background components. This subset contains the real part data of the complex impedance at all sensor locations within the swept frequency range after separation by joint sparse representation. For each sensor location's broadband dielectric response subset, a Hough transform is performed. The Hough transform converts the curve of the real part of the complex impedance as a function of frequency into a relaxation time spectrum. Specifically, a relaxation time axis is established, consisting of a series of discrete relaxation time values ​​ranging from 10⁻⁶ seconds to 10⁶ seconds, with 1000 points taken at equal intervals on a logarithmic scale. For each frequency point in the broadband dielectric response subset... Calculate the relaxation time corresponding to this frequency point. The real part of the complex impedance at that frequency point is accumulated onto the relaxation time axis and interpolated. At the nearest discrete points, a Gaussian kernel function is used for smoothing during accumulation. The standard deviation of the Gaussian kernel function is set to be equal to the standard deviation of the discrete points. The distance between adjacent discrete points is 0.5 times the distance between them. After traversing all frequency points, a spectral line characterizing the relaxation time distribution of the dielectric response is obtained on the relaxation time axis.

[0033] Characteristic peaks are identified on the relaxation time spectrum using a continuous wavelet transform peak detection method. The Mexican hat wavelet is used to perform a multi-scale continuous wavelet transform on the relaxation time spectrum, with the wavelet scale range set from 1 to 64. The positions of the maximum values ​​in each column of the wavelet coefficient matrix are extracted, and the relaxation time positions that are detected as local maxima at different scales are determined as characteristic peak positions. For each identified characteristic peak, peak position parameters and peak width parameters are extracted. The peak position parameter is defined as the relaxation time value corresponding to the peak value of the characteristic peak, and the peak width parameter is defined as the full width at half maximum (FWHM) of the characteristic peak, i.e., the absolute value of the difference between the two relaxation time values ​​corresponding to half the peak height. The peak position parameters and peak width parameters of all characteristic peaks extracted at each sensor placement location are used together as the relaxation time distribution characteristics of that placement location.

[0034] The process of extracting the temperature dependence feature of conductivity is as follows: Extract the real part of conductivity in the low-frequency band from the shielding layer temperature distribution signal and the broadband dielectric response subset from the steady-state background components. The low-frequency band is defined as the interval from 0.01 Hz to 1 Hz in the swept frequency range. For each sensor placement location, extract the imaginary part of the complex impedance at all frequency points within the low-frequency band from the broadband dielectric response subset, and calculate the real part of conductivity at each frequency point according to the conversion relationship between the imaginary part of the complex impedance and the real part of conductivity. The conversion relationship satisfies... ,in Angular frequency The real part of the conductivity below, The vacuum permittivity, The imaginary part of the complex relative permittivity is obtained by converting the imaginary part of the complex impedance through geometric capacitance relationships. For each sensor placement location, the arithmetic mean of the real parts of the conductivity at all frequency points within the low-frequency band is taken as the representative conductivity value of that placement location at the current temperature. The instantaneous temperature value of the same placement location at the same moment is extracted from the shielding layer temperature distribution signal, and the instantaneous temperature value is converted into an absolute temperature.

[0035] For each sensor location, temperature-conductivity data pairs were collected at multiple different operating times. These operating times should cover various cable conditions from low load to full load to ensure the widest possible temperature variation range. The least squares method was used to fit the temperature-conductivity data pairs for each location into the exponential relationship shown in the following equation:

[0036] in, This indicates that when the absolute temperature is The actual part of the conductivity of the shielding material is given, in Siemens per meter. This is a pre-exponential factor parameter, measured in Siemens units per meter. Its value is determined by analyzing multiple sets of data at the same location. The data points were obtained by least-squares fitting. The activation energy parameter, in electron volts, is also obtained through least-squares fitting. Boltzmann's constant has a value of . Electron volts per Kelvin; The value is the thermodynamic temperature, expressed in Kelvin, obtained by adding 273.15 to the instantaneous temperature in Celsius at the same location. The fitting process... and When performing linear regression, the slope of the regression line is... The intercept is Thus, the activation energy parameters at each sensor placement location are obtained. and pre-exponential factor parameters The activation energy parameter and the pre-exponential factor parameter together constitute the temperature-dependent conductivity characteristics at this sampling point location.

[0037] The process of constructing the initial feature tensor is as follows: For each sensor location in the distributed intelligent sensor array, the peak position parameter and peak width parameter in the relaxation time distribution feature, and the activation energy parameter and pre-exponential factor parameter in the conductivity temperature dependence feature, are arranged into a feature vector in a fixed order. The peak position parameter is taken from the peak position parameter of the feature peak with the largest amplitude in the relaxation time spectrum at that location. If no feature peak is detected in the relaxation time spectrum, the peak position parameter is set to zero. The peak width parameter is taken from the full width at half maximum (FWHM) of the same feature peak. If no feature peak is detected, it is set to zero. The activation energy parameter and pre-exponential factor parameter are taken from the values ​​obtained by correlation analysis fitting. The arrangement order of the feature vector is uniformly as follows: the first element is the peak position parameter, the second element is the peak width parameter, the third element is the activation energy parameter, and the fourth element is the pre-exponential factor parameter.

[0038] The feature vectors corresponding to all sensor locations are arranged in spatial order along the cable axis of the distributed intelligent sensor array, pointing from one end of the cable to the other, forming a two-dimensional matrix. The number of rows in this matrix is ​​the total number of sensor locations, and the number of columns is 4. A third dimension is added to this two-dimensional matrix to represent the state snapshot time. In a single feature extraction, the third dimension has a size of 1, thus forming a three-dimensional initial feature tensor. The three dimensions of this initial feature tensor are: spatial location dimension, feature parameter dimension, and time snapshot dimension.

[0039] Example 4: In the specific implementation, an interference adaptive correction algorithm based on the compressed sensing framework is adopted. The power frequency magnetic field coupling signal of the UHV cable during operation is used as a reference source to perform spatial phase correction on the initial feature tensor, and the corrected feature tensor is obtained.

[0040] The acquisition of the power frequency magnetic field coupling signal is achieved through a miniature induction coil integrated into a distributed intelligent sensor array. The miniature induction coil is wound with a high-permeability ferrite core, has 500 turns, a cross-sectional area of ​​10 square centimeters, and its axis is perpendicular to the cable axis, maximizing its sensitivity to the tangential magnetic field component generated by the power frequency current in the cable. The output of the miniature induction coil is connected to a signal conditioning circuit, which includes a bandpass filter with a center frequency of 50 Hz and an instrumentation amplifier with a gain of 40 dB. The passband of the bandpass filter is 2 Hz to suppress harmonic interference and high-frequency noise. The output signal of the signal conditioning circuit is input to the high-speed analog-to-digital conversion channel of the edge-side preprocessing unit and digitized at a sampling rate of 1000 Hz to obtain the discrete-time sequence of the power frequency magnetic field coupling signal.

[0041] The instantaneous phase and amplitude fluctuation envelope of the fundamental power frequency wave are demodulated from the discrete-time sequence of a power frequency magnetic field coupling signal. The demodulation process employs an analytic signal construction method based on the Hilbert transform. A Hilbert transform is performed on the discrete-time sequence of the power frequency magnetic field coupling signal to obtain an orthogonal sequence. The original discrete-time sequence is used as the real part, and the orthogonal sequence as the imaginary part to construct an analytic signal. The instantaneous phase of the analytic signal is obtained by calculating the arctangent of the ratio of the imaginary part to the real part. Phase unwrapping is performed on the instantaneous phase to eliminate phase jumps, resulting in a continuous instantaneous phase sequence of the fundamental power frequency wave. The envelope of the analytic signal is obtained by calculating the square root of the sum of the squares of the real and imaginary parts. A moving average of the envelope sequence with a 1-second window is taken to obtain the amplitude fluctuation envelope of the fundamental power frequency wave.

[0042] Based on the instantaneous phase sequence and amplitude fluctuation envelope of the fundamental power frequency wave, a spatial phase interference matrix with the same frequency and phase as the interference source is constructed. The spatial phase interference matrix is ​​constructed as follows: the dimension of the spatial phase interference matrix is ​​consistent with the spatial dimension of the initial feature tensor, i.e., the number of rows in the spatial phase interference matrix equals the total number of deployment points of the distributed intelligent sensor array, and the number of columns equals the number of columns of the feature parameter dimension of the initial feature tensor (4). For each element in the spatial phase interference matrix, its value is set to a complex number. The modulus of the complex number is taken from the amplitude fluctuation value corresponding to the center time of the current time snapshot window in the amplitude fluctuation envelope, and the argument of the complex number is taken from the instantaneous phase value at the same moment in the instantaneous phase sequence of the fundamental power frequency wave. When elements in different rows of the spatial phase interference matrix correspond to the same column index, the modulus of the corresponding elements in each row uses the amplitude fluctuation value at the same moment. However, based on the spatial coordinates of each deployment point along the cable axis, a phase offset linearly related to the spatial coordinates is superimposed on the instantaneous phase. The phase offset satisfies the following formula:

[0043] in, Indicates the first Phase offset at each location, in radians; For the first The axial distance of each placement point from the reference point at the beginning of the cable is expressed in meters. The equivalent wavelength of the power frequency electromagnetic wave propagating along the cable axis is taken as 3000 kilometers. This value is based on the propagation speed calculated from the inductance and capacitance parameters per unit length of the cable. The value is obtained by dividing meters per second by the power frequency of 50 Hz. The first [unit / number] in the spatial phase interference matrix. Line 1 The final argument of the column element is taken as the instantaneous phase and phase offset. sum.

[0044] Pointwise complex division is performed between the initial feature tensor and the spatial phase interference matrix. The specific process of pointwise complex division is as follows: each element in the initial feature tensor is considered a complex number, where the real part represents the resistive component of the broadband dielectric response signal, and the imaginary part represents the reactive component. When the initial feature tensor contains four feature parameter dimensions, the slice containing the peak position parameter and the pre-exponential factor parameter is considered the holder of the real part; the real part of the complex number is taken as the peak position parameter value or the pre-exponential factor parameter value, and the imaginary part is zero. The slice containing the peak width parameter and the activation energy parameter is considered the holder of the imaginary part; the imaginary part of the complex number is taken as the peak width parameter value or the activation energy parameter value, and the real part is zero. The corresponding element in the spatial phase interference matrix is ​​also considered a complex number; the modulus of the complex number is the corresponding value in the amplitude fluctuation envelope, and the argument of the complex number is the sum of the instantaneous phase and the corresponding row phase offset. For each spatial location point and each feature parameter dimension, the complex elements in the initial feature tensor are divided by the complex elements in the spatial phase interference matrix. The quotient obtained by this complex division is used as the corrected complex eigenvalue for that location point and that feature parameter dimension. The complex division is calculated as follows: the complex elements in the initial feature tensor... Complex elements in the spatial phase interference matrix Divide them, the quotient is ,in Represents the imaginary unit. and Let represent the real and imaginary parts of the complex elements in the initial feature tensor, respectively. and These represent the real and imaginary parts of the corresponding complex elements in the spatial phase interference matrix, respectively. All spatial location points and all feature parameter dimension-corrected complex eigenvalues ​​are combined into a corrected feature tensor according to the same spatial arrangement and dimensional order as the initial feature tensor. The corrected feature tensor eliminates the phase rotation and amplitude modulation effects introduced by power frequency magnetic field coupling.

[0045] See Figure 3 In the figure, the horizontal axis represents the sensor placement location in meters, ranging from 0 to 500 meters; the vertical axis represents the activation energy parameter in electron volts (eV). The figure contains two curves: the dashed line represents the activation energy parameter distribution before correction, and the solid line represents the activation energy parameter distribution after processing based on the compressed sensing framework interference adaptive correction algorithm described in Example 4.

[0046] The curves show a clear spatial variation in the activation energy parameter along the cable axis. Within a distance of 0 meters to approximately 250 meters, the activation energy parameter gradually decreases, dropping from approximately 0.80 eV to a minimum of approximately 0.58 eV. Within a distance of 250 meters to 500 meters, the activation energy parameter gradually recovers, approaching the level of the initial position, reaching approximately 0.80 eV. This trend reflects spatial differences in the temperature dependence of the conductivity of the shielding material corresponding to the sensor placement location; the lowest value range may correspond to areas where the shielding material is severely aged or defective.

[0047] The two curves before and after correction highly overlap, indicating that the phase rotation and amplitude modulation interference introduced by power frequency magnetic field coupling were successfully eliminated through the construction of the spatial phase interference matrix and the point-by-point complex division correction in Example 4, without introducing significant signal distortion. The activation energy parameter curve after correction is smoother, and the noise component is reduced, demonstrating the effectiveness of the interference adaptive correction algorithm in restoring the true shielding material state characteristics.

[0048] Overall, this figure effectively verifies the ability of the interference adaptive correction algorithm based on the compressed sensing framework described in Example 4 to correct the conductivity temperature-dependent features in the initial feature tensor, ensuring the accuracy and reliability of the feature input in the subsequent state evaluation stage.

[0049] Example 5: In specific implementation, please refer to Figure 2 The corrected feature tensor is input into a pre-trained twin asymmetric convolutional network, and the topological invariant distance between the corrected feature tensor and multiple typical state templates in the health baseline state library is calculated to obtain the quantitative value of the aging degree of the shielding material and the probability distribution of interface defect types.

[0050] The Siamese asymmetric convolutional network comprises two input branches: a fixed-weight feature extraction branch and a learnable-weight metric learning branch. The fixed-weight feature extraction branch processes each typical state template in the healthy baseline state library, while the learnable-weight metric learning branch processes the corrected feature tensors. The fixed-weight feature extraction branch consists of three convolutional layers and two fully connected layers. The kernel sizes of the three convolutional layers are 3×3, 3×3, and 1×1, respectively, followed by a batch normalization layer and a modified linear unit activation function. The two fully connected layers have 256 and 128 neurons, respectively. The weight parameters of the fixed-weight feature extraction branch are pre-trained on a separate source task and remain unchanged during subsequent inference. The source task employs a self-supervised contrastive learning approach, training the fixed-weight feature extraction branch using a simulation dataset of the dielectric response of cable shielding materials with different operating years and defect types. The training optimization objective is to maximize the similarity of feature vectors of the same category while minimizing the similarity of feature vectors of different categories. After the fixed-weight feature extraction branch is trained, feature embedding vectors are extracted for each typical state template and stored in the template embedding table of the healthy baseline state library to avoid redundant calculations. The metric learning branch with learnable weights has the exact same network structure as the fixed-weight feature extraction branch, but the weight parameters of the convolutional and fully connected layers are learned during the overall training phase of the Siamese network. The training process of the Siamese asymmetric convolutional network adopts a metric learning strategy, using feature tensors of shielding material samples labeled with known aging years and interface defect types to form the training set. The loss function during training consists of two parts: one part is the contrast loss between paired samples, using cosine distance and a hinge loss with a marginal parameter of 0.5; the other part is the classification loss, trained using multi-class cross-entropy loss for defect types. During training, the weights of the fixed-weight feature extraction branch are frozen, and only the weights of the metric learning branch with learnable weights and the shared batch normalization layer statistics are updated. The optimizer used is an adaptive moment estimation optimizer with an initial learning rate of 0.001. The training rounds are set to 200 rounds. After each round of training, the intra-class compactness of the feature embedding vectors is evaluated on the validation set. Training is terminated early when the intra-class compactness of the validation set no longer improves within 10 consecutive rounds.

[0051] During inference, the Siamese asymmetric convolutional network inputs the corrected feature tensor into a metric learning branch with learnable weights. The output of this branch undergoes global average pooling and L2 normalization to obtain the feature embedding vector of the current state, with a vector dimension of 128. Simultaneously, a typical state template is retrieved from the health baseline state database and input into a fixed-weight feature extraction branch. This template also undergoes global average pooling and L2 normalization to obtain the feature embedding vector of the baseline state. The health baseline state database stores multiple typical state templates. Each template corresponds to a known aging period label and an interface defect type label. The aging period label is determined by the actual number of years the cable has been in operation, and the interface defect type label includes five types: no defects, micropore defects, delamination defects, impurity defects, and crack defects.

[0052] For the feature embedding vector of the current state and the feature embedding vector of a baseline state, the persistent homology distance between them is calculated and used as the topological invariant distance. The calculation process of the persistent homology distance is as follows: Point clouds are constructed in the point cloud space of the feature embedding vector of the current state and the point cloud space of the feature embedding vector of the baseline state, respectively. The point cloud is constructed by splitting the feature embedding vector into multiple local vectors according to the sensor placement positions. Each local vector corresponds to a spatial location point, and the first three elements of the local vector are used as the coordinates of that location point in three-dimensional space to form a point cloud for each. For the point cloud of the current state and the point cloud of the baseline state, a Vitoris-Lipps complex is constructed with different scale parameters. The scanning range of the scale parameter starts from zero and increases in increments of 0.01 until the distance between the farthest point pairs in the point cloud. Under each scale parameter, the generators of the homology groups in each dimension of the point cloud are calculated, and the generation time and the disappearance time of each generator are recorded. The generation time is the scale parameter value when the generator first appears, and the disappearance time is the scale parameter value when the generator merges into a higher-order structure or disappears. Plot the generation and destruction times of all generators as points in a two-dimensional plane, where the x-axis represents the generation time and the y-axis represents the destruction time, to obtain a persistent graph. The point cloud of the current state yields the current state persistent graph, and the point cloud of the baseline state yields the baseline state persistent graph.

[0053] The Wasserstein distance between the current state persistent graph and the baseline state persistent graph is calculated using a one-dimensional Wasserstein distance. The persistent graph is considered a set of points, and the Wasserstein distance measures the optimal transport cost between the two sets of points, expressed as the square of the L2 norm. The Wasserstein distance is calculated by solving an optimal matching problem: points in the current state persistent graph are matched with points in the baseline state persistent graph or their diagonal projections. After matching, the Wasserstein distance is obtained by taking the square root of the sum of the squares of the Euclidean distances between the matched points. This Wasserstein distance is the persistent homology distance between the feature embedding vectors of the current state and the feature embedding vectors of the baseline state. The above persistent homology distance calculation process is then performed sequentially on all typical state templates in the healthy baseline state library, obtaining a persistent homology distance corresponding to each typical state template. All persistent homology distances form a set of topological invariant distances.

[0054] After obtaining a set of topological invariant distances, the smallest distance values ​​are identified from this set. The number K of these distance values ​​is dynamically determined based on the total number M of typical state templates in the health baseline state library. K is a minimum integer not less than M / 10, and its minimum value is 3. The aging period and interface defect type labels of the typical state templates corresponding to these K distance values ​​are recorded. The reciprocals of these K distance values ​​are taken to obtain K reciprocal distance values. These K reciprocal distance values ​​are then normalized so that the sum of the K normalized values ​​is 1. Each normalized value is the weight coefficient of the corresponding typical state template.

[0055] The process of calculating the quantification value of the aging degree of the shielding material is as follows: The aging years of K typical state templates are weighted and summed with their corresponding weighting coefficients. The weighted sum is the quantification value of the aging degree of the shielding material, in years. The process of calculating the probability distribution of interface defect types is as follows: The K typical state templates are grouped according to all five types of interface defect type labels. The weighting coefficients of the typical state templates belonging to the same interface defect type label group are summed. This sum is the probability of that interface defect type. The sum of the probability values ​​of the five interface defect types is 1. These five probability values ​​together constitute the interface defect type probability distribution.

[0056] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for testing the performance of ultra-high voltage cable shielding materials based on intelligent sensing technology, characterized in that, Includes the following steps: A distributed intelligent sensor array is embedded in the surface of the shielding layer of an ultra-high voltage cable to synchronously collect multi-dimensional original response signals of the shielding material while the cable is energized. The multi-dimensional original response signal is jointly sparsely represented by the edge-side preprocessing unit, the transient impact component and the steady-state background component in the signal are separated, and the abnormal event is triggered based on the time-domain distribution characteristics of the transient impact component. The relaxation time distribution characteristics and conductivity temperature dependence characteristics of the shielding material are extracted from the steady-state background components to construct the initial feature tensor of the current state of the shielding material. An interference adaptive correction algorithm based on a compressed sensing framework is adopted, using the power frequency magnetic field coupling signal of the UHV cable during operation as a reference source to perform spatial phase correction on the initial feature tensor, thereby obtaining the corrected feature tensor. The corrected feature tensor is input into a pre-trained twin asymmetric convolutional network, and the topological invariant distance between it and multiple typical state templates in the health baseline state library is calculated to obtain the quantitative value of the aging degree of the shielding material and the probability distribution of interface defect types.

2. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 1, characterized in that, The multi-dimensional raw response signal includes a broadband dielectric response signal, a partial discharge pulse signal, and a shielding layer temperature distribution signal.

3. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 2, characterized in that, A distributed intelligent sensor array is embedded in the surface of the UHV cable shielding layer to synchronously collect multi-dimensional raw response signals of the shielding material while the cable is energized, including: Multiple thin-film impedance sensors and fiber optic temperature sensors are embedded along the cable axis in a non-equal interval pattern on the surface of the shielding layer. Each point integrates an impedance measurement electrode and a temperature sensing unit. A sweeping sinusoidal excitation signal is applied to the impedance measurement electrode by a high-frequency broadband impedance analyzer, and the real and imaginary parts of the complex impedance of the shielding material at different frequency points are collected simultaneously and combined to form the broadband dielectric response signal. The partial discharge pulse waveform on the surface of the shielding layer is acquired by a coupling capacitor sensor within the same time window, and the amplitude, rise time and equivalent duration of each pulse are recorded as the partial discharge pulse signal. The wavelength offset at each point on the surface of the shielding layer is demodulated by the fiber optic grating temperature sensor. Based on the pre-calibrated wavelength-temperature conversion relationship, the instantaneous temperature value at each point is calculated and combined to form the temperature distribution signal of the shielding layer.

4. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 3, characterized in that, The multi-dimensional original response signal is jointly sparsely represented by an edge-side preprocessing unit to separate the transient impact component and the steady-state background component in the signal, including: The synchronously acquired broadband dielectric response signal, partial discharge pulse signal and shielding layer temperature distribution signal are aligned on the time axis to construct a multidimensional hybrid signal matrix. A dictionary learning algorithm based on stochastic resonance is used to learn an overcomplete joint dictionary from the multidimensional mixed signal matrix. The overcomplete joint dictionary contains transient atoms for representing transient impacts and steady-state atoms for representing steady-state backgrounds. The multidimensional mixed signal matrix is ​​decomposed by orthogonal matching pursuit on the overcomplete joint dictionary to obtain the transient impact component reconstructed by the linear combination of the transient atoms and the steady-state background component reconstructed by the linear combination of the steady-state atoms.

5. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 4, characterized in that, The dictionary learning algorithm based on stochastic resonance enhances the weak transient impulse components in the multidimensional mixed signal matrix by introducing a bistable potential function model, and dynamically updates the overcomplete joint dictionary using an online learning method.

6. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 4, characterized in that, Based on the time-domain distribution characteristics of the transient impact components, anomaly event markers are triggered, including: Calculate the peak arrival time and energy integral of each impact pulse in the transient impact component, and map the peak arrival time and energy integral onto the time axis to form a transient event time series; The transient event time series is scanned using a sliding window, and the occurrence frequency and cumulative energy of the impact pulses within the window are statistically analyzed. When the occurrence frequency exceeds a preset first threshold or the energy accumulation exceeds a preset second threshold, an abnormal event marker vector containing the start time, duration, and average energy is generated and stored in the edge-side buffer.

7. The method for testing the performance of ultra-high voltage cable shielding material based on intelligent sensing technology according to claim 6, characterized in that, The abnormal event marker vector is pushed to the upper-level state evaluation module in real time to trigger weighted amplification processing of the corresponding time window data in the initial feature tensor.

8. The method for testing the performance of UHV cable shielding material based on intelligent sensing technology according to claim 6, characterized in that, The relaxation time distribution characteristics and conductivity temperature dependence characteristics of the shielding material are extracted from the steady-state background components to construct the initial feature tensor of the current state of the shielding material, including: The broadband dielectric response subset in the steady-state background component is processed by Hough transform, and the curve of the real part of the complex impedance changing with frequency is converted into multiple characteristic peaks on the relaxation time spectrum. The peak position parameter and peak width parameter of each characteristic peak are extracted as the relaxation time distribution feature. A correlation analysis was performed on the shielding layer temperature distribution signal in the steady-state background component and the real part of the conductivity of the broadband dielectric response subset in the low-frequency band. The activation energy parameter and pre-exponential factor parameter in the exponential relationship of conductivity with temperature were fitted and used as the temperature-dependent characteristics of conductivity. The peak position parameter and peak width parameter in the relaxation time distribution feature, as well as the activation energy parameter and pre-exponential factor parameter in the conductivity temperature dependence feature, are arranged according to the sensor placement locations to construct a three-dimensional initial feature tensor.

9. The method for testing the performance of ultra-high voltage cable shielding material based on intelligent sensing technology according to claim 8, characterized in that, An interference adaptive correction algorithm based on a compressed sensing framework is adopted. Using the power frequency magnetic field coupling signal of the UHV cable during operation as a reference source, spatial phase correction is performed on the initial feature tensor to obtain the corrected feature tensor, including: A miniature induction coil is integrated into the distributed intelligent sensor array to collect the power frequency magnetic field coupling signal generated by the ultra-high voltage cable in real time, and demodulate the instantaneous phase and amplitude fluctuation envelope of the power frequency fundamental wave from the power frequency magnetic field coupling signal. Based on the instantaneous phase and the amplitude fluctuation envelope, a spatial phase interference matrix that is in phase and frequency with the interference source is constructed, wherein the dimension of the spatial phase interference matrix is ​​consistent with the spatial dimension of the initial feature tensor. The initial feature tensor is divided point-by-point by the spatial phase interference matrix to subtract the phase rotation and amplitude modulation effects introduced by the power frequency magnetic field coupling from the initial feature tensor, and the corrected feature tensor is output.

10. The method for testing the performance of ultra-high voltage cable shielding material based on intelligent sensing technology according to claim 1, characterized in that, The corrected feature tensor is input into a pre-trained Siamese asymmetric convolutional network, and the topological invariant distance between it and multiple typical state templates in the healthy baseline state library is calculated to obtain the quantified value of the aging degree of the shielding material and the probability distribution of interface defect types, including: The twin asymmetric convolutional network includes a feature extraction branch with fixed weights and a metric learning branch with learnable weights. The feature extraction branch with fixed weights is used to process each typical state template in the healthy baseline state library, and the metric learning branch with learnable weights is used to process the corrected feature tensor. The corrected feature tensor is input into the metric learning branch with learnable weights to obtain the feature embedding vector of the current state. At the same time, a typical state template from the health baseline state library is input into the feature extraction branch with fixed weights to obtain the feature embedding vector of the baseline state. Calculate the persistent homology distance between the feature embedding vector of the current state and the feature embedding vector of the baseline state, use the persistent homology distance as the topological invariant distance, and traverse all typical state templates in the healthy baseline state library to obtain a set of topological invariant distances.

Images