A method and system for ultra-long distance cable fault monitoring

Abstract

The application discloses a kind of super-long distance cable fault monitoring method and system, belong to electric power system safety monitoring field, the method includes: obtaining the partial discharge high-frequency transient signal of cable, hyperspectral remote sensing image and atmospheric parameter data, the phase space reconstruction of partial discharge high-frequency transient signal is carried out, constructs chaos feature;High spectral image is corrected based on atmospheric parameter, generate ground reflectivity image, and the spectral unmixing of ground reflectivity image is carried out, extract spectral fingerprint.Local discharge high-frequency transient signal and hyperspectral remote sensing image are time-space alignment, generate time-space mapping index, and accordingly chaos feature and spectral fingerprint are paired, form time-space feature pair.The fault probability of each position point of cable in future time range is predicted by cross-modal fusion model, and latent fault is determined according to threshold value.Through implementation of the application, it can solve the problem of inaccurate monitoring caused by weak signal and lagging index in the prior art.

Description

Technical Field

[0001] This invention relates to the field of power system safety monitoring technology, specifically to a method and system for monitoring faults in ultra-long-distance cables. Background Technology

[0002] As critical arteries for modern energy and information transmission, fault monitoring of ultra-long-distance cables is a core task for ensuring system safety. Current monitoring of ultra-long-distance cables primarily relies on analyzing terminal electrical parameters to determine faults. However, this method struggles to capture early signs of faults. The fundamental reason is that the initial signals generated by early faults are extremely weak, and during long-distance transmission, attenuation and dispersion lead to energy dissipation and waveform distortion, ultimately resulting in them being completely submerged in system background noise. Traditional signal analysis tools are largely ineffective against such nonlinear transient characteristics with extremely low signal-to-noise ratios, leading to severely inadequate early warning capabilities.

[0003] To enhance monitoring capabilities, some technologies have begun to integrate internal electrical signals with temperature data measured by optical fibers laid parallel to the cable. However, this method requires the installation of additional physical sensing circuitry alongside the cable, making the retrofit project massive and costly for the vast number of existing cables. More importantly, temperature is a delayed indicator of faults—it can only be detected after insulation degradation has progressed to a certain extent and produced significant thermal effects, often by which time the optimal window for preventative maintenance has been missed. Therefore, existing fusion solutions still face significant technical bottlenecks in achieving truly accurate pre-fault prediction. Summary of the Invention

[0004] This invention provides a method and system for monitoring faults in ultra-long-distance cables, which can solve the problem of inaccurate monitoring caused by weak signals and lagging indicators in the prior art.

[0005] An embodiment of the present invention provides a method for monitoring faults in ultra-long-distance cables, comprising:

[0006] Acquire high-frequency transient signals of partial discharge of the cable under monitoring, hyperspectral remote sensing images covering the cable route, and atmospheric parameter data;

[0007] Phase space reconstruction is performed on the high-frequency transient signal of partial discharge to generate a phase space attractor, and a chaotic feature vector is constructed based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor.

[0008] Atmospheric correction is performed on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images; spectral demixing is performed on the surface reflectance images based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples;

[0009] The high-frequency transient signal of partial discharge, hyperspectral remote sensing image and preset GIS route map of the cable to be monitored are spatiotemporally aligned to generate a spatiotemporal mapping index. Based on the spatiotemporal mapping index, the chaotic feature vector is paired with the spectral fingerprint feature sequence to generate a spatiotemporal feature pair sequence.

[0010] The spatiotemporal feature pairs are input into a preset cross-modal fusion model to generate a probability sequence of faults at each location point of the cable to be monitored within a future time range. The probability sequences of all location points are combined to construct a spatiotemporal fault probability cloud map.

[0011] Determine whether there is any spatial location point in the spatiotemporal fault probability cloud map whose fault probability continuously exceeds a preset threshold in the future time period; if so, determine that the cable under monitoring has a latent fault and trigger a fault warning; if not, determine that the cable under monitoring does not have a latent fault.

[0012] Furthermore, the step of reconstructing the phase space of the partial discharge high-frequency transient signal to generate a phase space attractor includes:

[0013] The high-frequency transient signal of partial discharge is sampled at fixed time intervals to generate a sampling sequence;

[0014] The sampled sequence is detrended to obtain a detrended signal sequence;

[0015] The amplitude normalization process is performed on the detrended signal sequence to obtain the preprocessed signal sequence;

[0016] Based on the preprocessed signal sequence, the autocorrelation function curve is calculated to obtain the autocorrelation function data.

[0017] The first minimum point of the autocorrelation function data is determined, and the time corresponding to the first minimum point is determined as the reconstruction delay time;

[0018] Based on the preprocessed signal sequence, the optimal embedding dimension is determined using the pseudo-neighbor method.

[0019] Based on the reconstruction delay time and the optimal embedding dimension, the preprocessed signal sequence is subjected to delayed sampling to generate a multidimensional state vector sequence.

[0020] The multidimensional state vector sequence is arranged in chronological order to form a phase space trajectory, and the phase space trajectory is used as a phase space attractor.

[0021] Furthermore, the step of performing atmospheric correction on the hyperspectral remote sensing image based on atmospheric parameter data to generate a surface reflectance image includes:

[0022] Radiometric calibration is performed on the raw pixel values ​​of the hyperspectral remote sensing image to generate radiance data;

[0023] Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the air scattering effect on the radiance data is calculated to generate scattering correction data.

[0024] Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the atmospheric absorption effect on the radiance data is calculated, and absorption correction data is generated.

[0025] Based on the scattering correction data and absorption correction data, the radiance data is corrected to generate an atmospherically corrected surface reflectance image.

[0026] Furthermore, the step of performing spectral demixing on the surface reflectance image and extracting spectral fingerprint feature sequences based on a preset characteristic spectral model includes:

[0027] The light intensity values ​​of each spectral band are extracted from the preset characteristic spectral model to construct the target spectral vector;

[0028] Traverse each pixel along the path of the cable to be monitored in the surface reflectance image and extract the reflectance value of each pixel in each spectral band to generate the corresponding pixel spectral vector.

[0029] For each pixel spectral vector, a matched filtering algorithm is used to calculate the projection value of the current pixel spectral vector in the direction of the target spectral vector, and generate the corresponding projection result for each pixel.

[0030] The projection results are normalized to generate the corresponding abundance values.

[0031] The abundance values ​​are arranged in order of cable path to form a spectral fingerprint feature sequence.

[0032] Furthermore, the step of spatiotemporally aligning the partial discharge high-frequency transient signal, hyperspectral remote sensing image, and the preset GIS route map of the cable to be monitored to generate a spatiotemporal mapping index includes:

[0033] Event detection is performed on the high-frequency transient signal of partial discharge to determine the occurrence time of each partial discharge event and generate an event time list;

[0034] For each partial discharge event, based on the high-frequency transient signal of the partial discharge, the location of the event on the cable to be monitored is determined, and a list of event spatial locations is generated.

[0035] Based on the event time list and the event spatial location list, the time of each partial discharge event is paired with its corresponding spatial location to generate an event time-space pair list;

[0036] The spatial locations in the event spatiotemporal pair list are mapped to the corresponding hyperspectral image pixel locations using a preset GIS path map of the cable to be monitored, thereby generating an event pixel correspondence table.

[0037] By combining the event spatiotemporal pair list with the event pixel correspondence table, a spatiotemporal mapping index is generated.

[0038] Furthermore, the step of pairing the chaotic feature vector with the spectral fingerprint feature sequence according to the spatiotemporal mapping index to generate a spatiotemporal feature pair sequence includes:

[0039] Iterate through each record in the spatiotemporal mapping index and retrieve the occurrence time and corresponding pixel position of each event from the record;

[0040] For each event, based on the occurrence time of the current event, extract the chaotic feature vector corresponding to the current event from the chaotic feature vector time series; based on the pixel position of the current event, extract the spectral fingerprint feature value corresponding to the current event from the spectral fingerprint feature sequence.

[0041] The chaotic feature vector is paired with the spectral fingerprint feature value to form a spatiotemporal feature pair for each event;

[0042] Arrange all spatiotemporal feature pairs according to the spatial order of the cable path to be monitored, and generate a spatiotemporal feature pair sequence.

[0043] Furthermore, the cross-modal fusion model includes a chaotic feature encoding module, a spectral feature encoding module, an attention fusion module, and a fault probability prediction module;

[0044] The step of inputting the spatiotemporal feature pair sequence into a preset cross-modal fusion model to generate a probability sequence of fault generation at each location point of the cable to be monitored within a future time range includes:

[0045] The spatiotemporal feature pair sequence is input into a preset cross-modal fusion model, so that the cross-modal fusion model performs deep feature encoding on the chaotic feature vectors in the spatiotemporal feature pair sequence through the chaotic feature encoding module, generating a chaotic modal deep feature sequence.

[0046] The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence in the spatiotemporal feature pair sequence to generate a spectral modality deep feature sequence.

[0047] The attention fusion module performs weighted fusion of the chaotic mode deep feature sequence and the spectral mode deep feature sequence to generate a fused spatiotemporal feature vector sequence.

[0048] The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence to generate a probability sequence of faults occurring at each location point of the cable to be monitored within a future time range.

[0049] Furthermore, the training of the cross-modal fusion model includes:

[0050] Obtain a historical dataset of cable faults; wherein, the historical dataset of cable faults includes several training data, each training data including a sequence of historical spatiotemporal feature pairs, and a label of the actual fault state in the future time period corresponding to the historical spatiotemporal feature pair sequence in time;

[0051] According to the preset batch size, the historical cable fault dataset is randomly divided into several batches of training samples.

[0052] The training samples from each batch are sequentially input into the cross-modal fusion model, and end-to-end iterative training is performed on all learnable parameters in the cross-modal fusion model until a preset number of training rounds is reached. Among them, when the cross-modal fusion model receives a batch of training samples, it uses the chaotic feature encoding module to perform deep feature encoding on the chaotic feature vector of the historical spatiotemporal feature pair sequence in the current batch of training samples, generating a chaotic modal deep feature sequence during the training process.

[0053] The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence of the historical spatiotemporal feature pairs in the current batch of training samples to generate the spectral modality deep feature sequence during the training process.

[0054] The attention fusion module performs weighted fusion based on the chaotic mode deep feature sequence and the spectral mode deep feature sequence during the training process to generate a fused spatiotemporal feature vector sequence during the training process.

[0055] The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence during training to generate a predicted probability sequence of faults in the future time range corresponding to the current training samples. A preset loss function is used to calculate and generate a loss function value based on the predicted probability sequence and the corresponding real fault state label. A preset optimizer is then used to perform gradient backpropagation and update the learnable network parameters in the chaotic feature encoding module, the spectral feature encoding module, the attention fusion module, and the fault probability prediction module based on the loss function value.

[0056] Furthermore, the characteristic spectral model is constructed in the following manner:

[0057] Obtain laboratory spectral data; wherein the laboratory spectral data is obtained by applying excitation to a cable sample of the same type as the cable to be monitored in a controlled experimental device to induce local micro-discharge, and using a spectrometer to collect the optical radiation signal of the discharge phenomenon;

[0058] The laboratory spectral data are subjected to noise filtering and baseline correction to generate standardized spectral sample data;

[0059] In the standardized spectral sample data, characteristic spectral lines representing local micro-discharge phenomena are identified and extracted using a peak detection algorithm;

[0060] The wavelengths of the characteristic spectral lines and their corresponding relative light intensity values ​​are digitized and stored to construct a characteristic spectral model.

[0061] Based on the above method embodiments, the present invention provides corresponding system embodiments.

[0062] One embodiment of the present invention provides a fault monitoring system for ultra-long-distance cables, comprising: a data acquisition module, a chaotic feature extraction module, a spectral feature extraction module, a feature pairing module, a cross-modal fusion module, and a fault prediction module;

[0063] The data acquisition module is used to acquire high-frequency transient signals of partial discharge of the cable to be monitored, hyperspectral remote sensing images covering the cable route, and atmospheric parameter data.

[0064] The chaotic feature extraction module is used to reconstruct the phase space of the high-frequency transient signal of partial discharge, generate a phase space attractor, and construct a chaotic feature vector based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor.

[0065] The spectral feature extraction module is used to perform atmospheric correction on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images; and to perform spectral unmixing on the surface reflectance images based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples.

[0066] The feature pairing module is used to perform spatiotemporal alignment of the partial discharge high-frequency transient signal, hyperspectral remote sensing image and preset GIS route map of the cable to be monitored, generate a spatiotemporal mapping index, and pair the chaotic feature vector with the spectral fingerprint feature sequence according to the spatiotemporal mapping index to generate a spatiotemporal feature pair sequence.

[0067] The cross-modal fusion module is used to input the spatiotemporal feature pair sequence into the preset cross-modal fusion model to generate a probability sequence of faults generated at each location point of the cable to be monitored within a future time range, and to combine the probability sequences of all location points to construct a spatiotemporal fault probability cloud map.

[0068] The fault prediction module is used to determine whether there is a fault probability at any spatial location point in the spatiotemporal fault probability cloud map that is continuously higher than a preset threshold in the future time period; if so, it is determined that there is a latent fault in the cable to be monitored and a fault warning is triggered; if not, it is determined that there is no latent fault in the cable to be monitored.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] This invention provides a method and system for monitoring faults in ultra-long-distance cables. The method acquires high-frequency transient signals of partial discharge, hyperspectral remote sensing images, and atmospheric parameter data of the cable to be monitored; reconstructs the phase space of the partial discharge signal to construct a chaotic feature vector; performs atmospheric correction on the hyperspectral image and extracts a spectral fingerprint feature sequence based on a feature spectral model constructed from characteristic spectral lines obtained from laboratory micro-discharges; aligns the chaotic feature vector and the spectral fingerprint feature sequence spatiotemporally using the cable GIS path to generate a spatiotemporal feature pair sequence; inputs the spatiotemporal feature pair sequence into a cross-modal fusion model to predict the future fault probability at each location point and constructs a spatiotemporal fault probability cloud map; and determines whether the fault probability at any spatial point continuously exceeds a threshold based on the cloud map, thereby identifying latent faults and triggering an early warning.

[0071] Early cable faults typically generate weak and difficult-to-detect signals. This is fundamentally because micro-discharge is the initial physical process, altering the local electric field and signal fluctuations without causing significant changes in cable temperature or structure. Therefore, traditional monitoring based on macroscopic indicators struggles to detect potential faults early. To address this, this invention approaches the problem from a microscopic level, analyzing internal cable signals based on chaos theory to quantitatively characterize the dynamics of signal evolution from stationary to chaotic states, thus providing an early warning of internal signal dynamics. Simultaneously, using hyperspectral remote sensing imagery combined with a laboratory-obtained micro-discharge spectral model, local optical signals generated by micro-discharge are extracted through spectral demixing and matched filtering algorithms. These optical signals reflect the micro-discharge process itself, rather than cable structural damage, and can expose potential faults earlier than temperature indicators that rely on heat accumulation. Finally, the internal chaotic features and external spectral fingerprints are comprehensively analyzed through spatiotemporal alignment and cross-modal fusion to achieve early fault warning with dual microscopic precursors, improving monitoring sensitivity and accuracy. Attached Figure Description

[0072] Figure 1This is a flowchart illustrating a method for monitoring faults in ultra-long-distance cables according to an embodiment of the present invention.

[0073] Figure 2 This is a schematic diagram of the structure of an ultra-long-distance cable fault monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] It should be noted that this invention aims to solve a core problem in existing ultra-long-distance cable fault monitoring technologies: the early warning signals of faults are extremely weak, and traditional monitoring methods rely heavily on macroscopic and lagging physical indicators, resulting in delayed warnings and insufficient reliability. To fundamentally overcome this bottleneck, the core technical concept of this invention is to shift the monitoring focus from macroscopic fault phenomena to microscopic physical precursors, and to establish a dual early warning system based on the mutual corroboration of internal electrodynamic evolution and external physical environment products.

[0076] In this system, based on the analysis of internal electrical signals, this invention recognizes that the initial form of fault is microscopic discharge, a process that disrupts the original stable and linear patterns of the signal. Therefore, this invention introduces chaos theory as a nonlinear analysis tool, using methods such as phase space reconstruction to quantitatively characterize the dynamic process of the signal evolving from an ordered to a chaotic state, thereby capturing the earliest signs of system instability that traditional methods cannot identify.

[0077] Complementing this, to obtain independent external physical evidence, this invention innovatively applies hyperspectral remote sensing technology to the detection of microscopic signs. Instead of observing macroscopic damage to the cable itself, it aims to identify and extract the spectral fingerprint of specific chemical substances produced by the instantaneous ionization of surrounding air by microscopic discharges. This fingerprint is a direct, instantaneous physical product of the fault, offering unparalleled early detection and sensitivity compared to lagging indicators such as temperature, which require heat accumulation.

[0078] The ultimate technical effect of this invention is achieved through the deep fusion of the two independent microscopic precursor features mentioned above. By precisely pairing the internal chaotic features with the external spectral fingerprint through spatiotemporal alignment, and using a cross-modal fusion model for comprehensive analysis, cross-verification of dual evidence can be achieved. This fundamentally solves the problem of false alarms or missed alarms easily generated by single-source monitoring, significantly advancing the warning window while greatly improving the accuracy and reliability of fault monitoring.

[0079] like Figure 1 As shown, to address the problem of inaccurate monitoring caused by weak signals and lagging indicators in existing technologies, an embodiment of the present invention provides a method for monitoring faults in ultra-long-distance cables, comprising at least the following steps:

[0080] Step S1: Acquire the high-frequency transient signal of partial discharge of the cable to be monitored, the hyperspectral remote sensing image covering the cable route, and atmospheric parameter data;

[0081] Specifically, in one embodiment of the present invention, the initial step of the ultra-long-distance cable fault monitoring method is the synchronous acquisition of multi-source heterogeneous data, which provides a data foundation for subsequent fusion analysis.

[0082] Specifically, the monitoring system continuously collects high-frequency transient signals of partial discharge within the cable by deploying high-frequency current transformers or capacitively coupled sensors at the terminals or key nodes along the cable line. This signal is the direct current or voltage response generated in the electrical circuit when a micro-discharge occurs within the cable's insulation medium. Its transient waveform and timing characteristics contain nonlinear dynamic information characterizing the evolution of the cable insulation from stability to degradation, serving as the fundamental basis for subsequent chaotic feature analysis.

[0083] Simultaneously, the system invokes or controls an airborne or spaceborne remote sensing platform equipped with a hyperspectral imager to acquire hyperspectral remote sensing images that completely cover the geographical path of the cable to be monitored within a preset monitoring period. These hyperspectral remote sensing images, presented as image cubes, record the reflectance spectra of solar radiation along the cable's surface across hundreds of continuous, narrow bands. This information provides a core optical data source for accurately identifying and extracting specific physicochemical characteristics—i.e., spectral fingerprints—generated by microscopic discharge activity in the cable's external environment.

[0084] To ensure the accuracy of spectral fingerprint feature extraction, atmospheric parameter data covering the cable route also needs to be acquired simultaneously. This atmospheric parameter data includes key information such as ambient temperature, humidity, visibility, and atmospheric aerosol optical thickness corresponding to the time of hyperspectral remote sensing image imaging. This data can be obtained from the network of ground meteorological stations along the route or from public meteorological service data interfaces. This atmospheric parameter data is the core input for subsequent atmospheric radiative transfer model calculations and atmospheric corrections. Its role is to accurately isolate the effects of atmospheric scattering and absorption on the surface reflectance spectrum, which is a crucial prerequisite for restoring the true surface reflectance and ensuring the effectiveness of subsequent spectral unmixing.

[0085] Through the three data sources obtained in the above steps, this invention constructs a spatiotemporally synchronized multidimensional heterogeneous dataset consisting of internal electrical signals, external optical images, and environmental correction parameters, providing a comprehensive and reliable data foundation for subsequent deep fusion early warning through dual microscopic precursors.

[0086] Step S2: Reconstruct the phase space of the high-frequency transient signal of partial discharge, generate a phase space attractor, and construct a chaotic feature vector based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor.

[0087] In a preferred embodiment, the step of reconstructing the phase space of the partial discharge high-frequency transient signal to generate a phase space attractor includes:

[0088] The high-frequency transient signal of partial discharge is sampled at fixed time intervals to generate a sampling sequence;

[0089] The sampled sequence is detrended to obtain a detrended signal sequence;

[0090] The amplitude normalization process is performed on the detrended signal sequence to obtain the preprocessed signal sequence;

[0091] Based on the preprocessed signal sequence, the autocorrelation function curve is calculated to obtain the autocorrelation function data.

[0092] The first minimum point of the autocorrelation function data is determined, and the time corresponding to the first minimum point is determined as the reconstruction delay time;

[0093] Based on the preprocessed signal sequence, the optimal embedding dimension is determined using the pseudo-neighbor method.

[0094] Based on the reconstruction delay time and the optimal embedding dimension, the preprocessed signal sequence is subjected to delayed sampling to generate a multidimensional state vector sequence.

[0095] The multidimensional state vector sequence is arranged in chronological order to form a phase space trajectory, and the phase space trajectory is used as a phase space attractor.

[0096] Specifically, after acquiring the high-frequency transient signal of partial discharge characterizing the cable state, a core step of this invention lies in applying chaos theory to conduct in-depth dynamic analysis of the signal. Although the original one-dimensional time series signal contains fault information, its inherent and complex dynamic laws are difficult to reveal directly. Therefore, this embodiment uses phase space reconstruction technology to expand the one-dimensional signal into a high-dimensional state space, thereby recovering and visualizing the intrinsic structure of the original dynamic system, i.e., the phase space attractor.

[0097] In a preferred embodiment, the specific process of the phase space reconstruction described above first involves sampling the acquired continuous partial discharge high-frequency transient signal at a fixed sampling frequency, converting it into a discrete time-series signal, i.e., a sampling sequence. To eliminate the potential influence of low-frequency interference such as DC bias or power frequency trend on nonlinear analysis, the sampling sequence needs to be detrended to obtain a zero-mean detrended signal sequence. Subsequently, to facilitate subsequent mathematical calculations and eliminate the influence of signal amplitude variations, the detrended signal sequence is normalized to obtain a preprocessed signal sequence with amplitudes within a uniform range.

[0098] After signal preprocessing is completed, two key parameters for phase space reconstruction need to be determined: reconstruction delay time. and optimal embedding dimension In this embodiment, the delay time The autocorrelation function method is used to determine the reconstruction delay time. First, based on the preprocessed signal sequence, its autocorrelation function curve is calculated. The point where the function value first drops to a minimum on this curve best ensures the lowest linear correlation among the reconstructed components. Therefore, the time corresponding to this minimum point is determined as the reconstruction delay time. Optimal embedding dimension The method of determining the false neighbor ratio is based on the principle that when the embedding dimension is too low, the projection of the attractor of the original dynamic system into the low-dimensional space will produce false neighbor points. By calculating and observing the change in the false neighbor ratio as the dimension increases, the dimension at which the false neighbor ratio first drops to a preset low threshold (usually close to zero) is selected as the minimum integer dimension that can fully unfold the true structure of the attractor, i.e., the optimal embedding dimension. .

[0099] The reconstruction delay time determined by the above method and optimal embedding dimension This allows for the reconstruction of the preprocessed signal sequence. The reconstruction process is based on Takens' embedding theorem, using delayed sampling to construct a multidimensional state vector sequence, mathematically expressed as follows:

[0100]

[0101] In the formula, For the first phase space A state vector; For the first preprocessed signal sequence One sampling point; The optimal embedding dimension; This is the reconstruction delay time.

[0102] All calculated The multidimensional state vector sequence is arranged in chronological order, and the resulting trajectory is the phase space trajectory, which fully presents the attractor shape of the original dynamic system.

[0103] After successfully generating the phase space attractor, this embodiment further calculates its quantitative chaotic characteristic index.

[0104] To quantitatively assess the predictability of the system, this embodiment calculates the Maximum Lyapunov Exponent of the phase space attractor. This exponent quantifies the average exponential rate at which two infinitely close points on the attractor trajectory separate or merge over time. A positive Maximum Lyapunov Exponent is a clear indicator of chaotic behavior. This embodiment employs the Small Data Method, whose core idea is to track the average logarithmic distance growth rate of neighboring point pairs on the attractor. Its basic principle can be expressed by the following equation:

[0105]

[0106] In the formula, This is an estimate of the maximum Lyapunov exponent; For time evolution step size; This represents the total number of evolutionary steps. For the first For neighboring points in evolution time The distance after; For the first Initial distances to neighboring points. If the calculated... This indicates that the system trajectory has local instability and is in a chaotic state. The larger the value, the worse the predictability of the system and the higher the degree of signal disorder.

[0107] To quantitatively assess the complexity of the system, this embodiment calculates the fractal dimension of the phase space attractor, specifically using the correlation dimension as an effective estimate. The correlation dimension is calculated using the Grassberger-Procaccia Algorithm (GP), the core of which lies in the fact that the distance between any two points on the statistical attractor is less than a given scale. The probability, i.e., the correlation integral For an attractor with fractal properties, its correlation integral is related to the distance scale. There is a power-law relationship:

[0108]

[0109] In the formula, The correlation integral indicates that the distance between two points in phase space is less than 1 / 2. The probability of; The distance scale; This represents the correlation dimension.

[0110] In practice, calculations are performed in logarithmic coordinates. and The correlation dimension can be obtained by measuring the slope of the linear portion (i.e., the scaling region) of the relationship curve. Correlation dimension The value of reflects the complexity of the attractor geometry and the number of independent variables contained in the signal. A non-integer, finite correlation dimension is also strong evidence of chaotic properties in the system.

[0111] Finally, the scalar eigenvalue of the calculated maximum Lyapunov exponent is combined with the scalar eigenvalue of the correlation dimension to construct a two-dimensional chaotic eigenvector.

[0112] Through the above series of processes, the present invention successfully transforms the original, high-dimensional, and complex transient signal into a low-dimensional chaotic feature vector that can accurately quantify the stability of the system, providing a first-level, quantitative early warning feature that is highly sensitive to the internal state of the cable for subsequent cross-modal fusion.

[0113] Step S3: Perform atmospheric correction on the hyperspectral remote sensing image based on atmospheric parameter data to generate a surface reflectance image; perform spectral unmixing on the surface reflectance image based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples.

[0114] In a preferred embodiment, the step of performing atmospheric correction on the hyperspectral remote sensing image based on atmospheric parameter data to generate a surface reflectance image includes:

[0115] Radiometric calibration is performed on the raw pixel values ​​of the hyperspectral remote sensing image to generate radiance data;

[0116] Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the air scattering effect on the radiance data is calculated to generate scattering correction data.

[0117] Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the atmospheric absorption effect on the radiance data is calculated, and absorption correction data is generated.

[0118] Based on the scattering correction data and absorption correction data, the radiance data is corrected to generate an atmospherically corrected surface reflectance image.

[0119] In a preferred embodiment, the step of spectrally unmixing the surface reflectance image and extracting the spectral fingerprint feature sequence according to a preset characteristic spectral model includes:

[0120] The light intensity values ​​of each spectral band are extracted from the preset characteristic spectral model to construct the target spectral vector;

[0121] Traverse each pixel along the path of the cable to be monitored in the surface reflectance image and extract the reflectance value of each pixel in each spectral band to generate the corresponding pixel spectral vector.

[0122] For each pixel spectral vector, a matched filtering algorithm is used to calculate the projection value of the current pixel spectral vector in the direction of the target spectral vector, and generate the corresponding projection result for each pixel.

[0123] The projection results are normalized to generate the corresponding abundance values.

[0124] The abundance values ​​are arranged in order of cable path to form a spectral fingerprint feature sequence.

[0125] In a preferred embodiment, the characteristic spectral model is constructed in the following manner:

[0126] Obtain laboratory spectral data; wherein the laboratory spectral data is obtained by applying excitation to a cable sample of the same type as the cable to be monitored in a controlled experimental device to induce local micro-discharge, and using a spectrometer to collect the optical radiation signal of the discharge phenomenon;

[0127] The laboratory spectral data are subjected to noise filtering and baseline correction to generate standardized spectral sample data;

[0128] In the standardized spectral sample data, characteristic spectral lines representing local micro-discharge phenomena are identified and extracted using a peak detection algorithm;

[0129] The wavelengths of the characteristic spectral lines and their corresponding relative light intensity values ​​are digitized and stored to construct a characteristic spectral model.

[0130] Specifically, after obtaining hyperspectral remote sensing images and synchronized atmospheric parameter data, another key technical approach of this invention lies in extracting external physical evidence that can characterize micro-discharge phenomena, namely, spectral fingerprint feature sequences. The core of this process is to reconstruct and identify the weak spectral features, which are of true interest to us and caused by micro-discharges, through a series of refined processing steps, of the mixed light signals received by the sensor in space, which contain complex information about the atmosphere and the Earth's surface.

[0131] The processing first requires atmospheric correction to eliminate atmospheric interference with optical signals and generate a true surface reflectance image. In a preferred embodiment, this atmospheric correction process includes: firstly, radiometric calibration of the raw pixel values ​​(DN values) of the hyperspectral remote sensing image, converting them into physically meaningful radiance data at the sensor's entrance pupil. However, this radiance data contains information about the entire process of solar radiation passing through the atmosphere to reach the surface, being reflected by the surface, and then passing through the atmosphere again to be received by the sensor, and is severely affected by scattering and absorption by air molecules. Therefore, it is necessary to use a pre-defined atmospheric radiative transfer model (such as MODTRAN, 6S, etc.) combined with synchronously acquired atmospheric parameter data (including temperature, humidity, aerosol optical thickness, etc.) to correct the radiance data. The physical model of this correction process can be simplified as follows:

[0132]

[0133] In the formula, Let be the surface reflectance to be determined; This is the total radiance value received by the sensor after radiometric calibration; This refers to atmospheric path radiance, which is the portion that is directly emitted by the atmosphere itself and enters the sensor. The total atmospheric transmittance from the Earth's surface to the sensor; The total solar irradiance reaching the Earth's surface.

[0134] In this embodiment, It can be obtained directly from radiometrically calibrated images. Atmospheric path radiance... Total atmospheric transmittance and total solar irradiance All of these can be accurately calculated using the synchronously acquired atmospheric parameter data as input through the atmospheric radiative transfer model. Through the above calculations, atmospheric influences can be removed, yielding a true image of the Earth's surface reflectance.

[0135] After obtaining the surface reflectance image, the next step is to extract the spectral fingerprint feature sequence from it based on a preset characteristic spectral model. In a preferred embodiment, this extraction process employs a matched filtering algorithm. This algorithm is an ideal detector for detecting known target signals from a mixed signal background. First, the target spectral vector representing the micro-discharge phenomenon is extracted from the preset characteristic spectral model. Then, each pixel along the cable path in the surface reflectance image is traversed, and its reflectance values ​​in all spectral bands are extracted to form the spectral vector of the pixel to be detected. The core of the matched filtering algorithm is to maximize the response of the target signal while suppressing the response of background noise. Its output can be calculated by the following formula:

[0136]

[0137] In the formula, The output of the matched filter; This represents the spectral vector of the pixel to be detected. The target spectral vector extracted from the characteristic spectral model; The mean spectral vector of the background pixels; It is the inverse of the background pixel spectral covariance matrix.

[0138] In this embodiment, the pixel spectral vector Target spectral vector obtained from surface reflectance imagery Background statistics are obtained from a pre-defined characteristic spectral model. and This can be obtained through statistical calculations on the background area of ​​the image. The calculated projection result... After normalization, a spectral fingerprint feature sequence is formed.

[0139] The pre-defined characteristic spectral model is constructed by a method comprising the following steps executed by one or more processors: First, a set of laboratory spectral data is acquired. This laboratory spectral data is obtained by applying excitation to a cable sample of the same type as the cable to be monitored in a controlled experimental setup to induce local micro-discharge, and then using a spectrometer to collect the optical radiation signal of the discharge phenomenon. Subsequently, digital signal processing is performed on the laboratory spectral data, including noise filtering and baseline correction, to generate standardized spectral sample data. Next, in the standardized spectral sample data, one or more characteristic spectral lines that can uniquely characterize the local micro-discharge phenomenon are identified and extracted using peak detection algorithms and other means. Finally, the wavelengths of these characteristic spectral lines and their corresponding relative light intensity values ​​are digitized and stored to construct a pre-defined characteristic spectral model containing these digitized characteristic spectral lines.

[0140] In summary, through a series of refined data processing techniques such as atmospheric correction, characteristic spectral model construction, and matched filtering, this invention successfully transforms the original hyperspectral remote sensing image, which is severely affected by atmospheric interference, into a spectral fingerprint feature sequence that precisely corresponds to the spatial location of the cable line and can sensitively reflect microscopic discharge precursors. This provides a high-quality and highly sensitive second-level early warning feature for subsequent cross-modal fusion.

[0141] Step S4: Spatiotemporally align the partial discharge high-frequency transient signal, hyperspectral remote sensing image, and preset GIS path map of the cable to be monitored to generate a spatiotemporal mapping index. Based on the spatiotemporal mapping index, pair the chaotic feature vector with the spectral fingerprint feature sequence to generate a spatiotemporal feature pair sequence.

[0142] In a preferred embodiment, the step of spatiotemporally aligning the partial discharge high-frequency transient signal, hyperspectral remote sensing image, and a preset GIS route map of the cable to be monitored to generate a spatiotemporal mapping index includes:

[0143] Event detection is performed on the high-frequency transient signal of partial discharge to determine the occurrence time of each partial discharge event and generate an event time list;

[0144] For each partial discharge event, based on the high-frequency transient signal of the partial discharge, the location of the event on the cable to be monitored is determined, and a list of event spatial locations is generated.

[0145] Based on the event time list and the event spatial location list, the time of each partial discharge event is paired with its corresponding spatial location to generate an event time-space pair list;

[0146] The spatial locations in the event spatiotemporal pair list are mapped to the corresponding hyperspectral image pixel locations using a preset GIS path map of the cable to be monitored, thereby generating an event pixel correspondence table.

[0147] By combining the event spatiotemporal pair list with the event pixel correspondence table, a spatiotemporal mapping index is generated.

[0148] In a preferred embodiment, the step of pairing the chaotic feature vector with the spectral fingerprint feature sequence according to the spatiotemporal mapping index to generate a spatiotemporal feature pair sequence includes:

[0149] Iterate through each record in the spatiotemporal mapping index and retrieve the occurrence time and corresponding pixel position of each event from the record;

[0150] For each event, based on the occurrence time of the current event, extract the chaotic feature vector corresponding to the current event from the chaotic feature vector time series; based on the pixel position of the current event, extract the spectral fingerprint feature value corresponding to the current event from the spectral fingerprint feature sequence.

[0151] The chaotic feature vector is paired with the spectral fingerprint feature value to form a spatiotemporal feature pair for each event;

[0152] Arrange all spatiotemporal feature pairs according to the spatial order of the cable path to be monitored, and generate a spatiotemporal feature pair sequence.

[0153] Specifically, after extracting the chaotic feature vector characterizing the internal state of the cable and the spectral fingerprint feature sequence characterizing the physical precursors of the external environment, a core step of this invention involves precise spatiotemporal alignment and pairing of these two heterogeneous, time-divisional, and domain-divisional data streams. The fundamental purpose is to establish a unique and definitive correspondence between each internal electrical disturbance event (described by the chaotic feature vector) and its potential physical manifestations (described by the spectral fingerprint feature values) at the same spatiotemporal location outside the cable. This process is a crucial prerequisite for achieving subsequent effective cross-modal fusion and cross-validation of dual evidence.

[0154] In a preferred embodiment, this spatiotemporal alignment process is achieved by constructing a spatiotemporal mapping index. First, event detection is required on the continuously acquired high-frequency transient signals of partial discharge. This detection process can use digital signal processing methods such as short-time energy analysis, wavelet transform, or setting dynamic thresholds to identify and separate each independent transient pulse event caused by micro-discharge from the background noise, thereby determining the precise occurrence time of each partial discharge event and forming an event time list.

[0155] Next, spatial localization of each detected partial discharge event is required. When sensors are evenly distributed at both ends of the cable, the Time Difference of Arrival (TDOA) method can be used to determine the location of the event on the cable line. This method is based on the physical principle that electromagnetic waves propagate at an approximately constant speed in the cable medium. Its location calculation can be expressed by the following formula:

[0156]

[0157] In the formula, For fault events, distance cable The physical distance between the ends; The total length of the cable segment for which both ends are measured; This refers to the propagation speed of a transient signal in the cable. The same discharge event signal arrived at the cable End and The time difference between the two ends.

[0158] In this embodiment, the cable segment length and signal propagation speed All parameters are known or pre-calibrated, and the time difference is... This information can be directly measured from signals acquired synchronously at both ends. Through this calculation, the physical mileage location on the line can be determined for each discharge event, forming a list of event spatial locations. Subsequently, the event time list is paired with the event spatial location list to form a spatiotemporal event pair list. Finally, each physical mileage location in this list is converted into geospatial coordinates using a pre-set, geo-registered GIS route map of the cable to be monitored, and further mapped to the corresponding hyperspectral image pixel location, ultimately generating a spatiotemporal mapping index containing information such as event time, physical location, and pixel location.

[0159] After generating the spatiotemporal mapping index, the separated chaotic feature vectors can be paired with the spectral fingerprint feature sequence based on this index. In a preferred embodiment, the pairing process includes: traversing each record in the spatiotemporal mapping index to obtain the occurrence timestamp and corresponding pixel position of each local discharge-electrode event. For each record in the index, the system finds and extracts the chaotic feature vector that completely corresponds to the occurrence timestamp from the previously calculated time-axis sequence of chaotic feature vectors. Simultaneously, based on the pixel position of the record, the system finds and extracts the spectral fingerprint feature value that completely corresponds to the pixel position from the previously calculated spatial path-axis sequence of spectral fingerprint features. Combining the extracted chaotic feature vectors and spectral fingerprint feature values ​​forms a "spatiotemporal feature pair" that can completely describe the internal and external precursors of the discharge event. This process is repeated for all events recorded in the spatiotemporal mapping index, and all generated spatiotemporal feature pairs are arranged according to the spatial order of the cable path, ultimately forming a complete and structured sequence of spatiotemporal feature pairs.

[0160] Through the above-described series of steps of spatiotemporal alignment and feature pairing, this invention successfully transforms two independent original feature data streams organized in different dimensions into a unified, highly structured spatiotemporal feature pair sequence in which each data point contains dual precursor information. This lays a regular and reliable data foundation for subsequent input into a cross-modal fusion model for high-precision prediction and analysis.

[0161] Step S5: Input the spatiotemporal feature pair sequence into the preset cross-modal fusion model to generate a probability sequence of faults at each location point of the cable to be monitored within the future time range, and combine the probability sequences of all location points to construct a spatiotemporal fault probability cloud map.

[0162] In a preferred embodiment, the cross-modal fusion model includes a chaotic feature encoding module, a spectral feature encoding module, an attention fusion module, and a fault probability prediction module;

[0163] The step of inputting the spatiotemporal feature pair sequence into a preset cross-modal fusion model to generate a probability sequence of fault generation at each location point of the cable to be monitored within a future time range includes:

[0164] The spatiotemporal feature pair sequence is input into a preset cross-modal fusion model, so that the cross-modal fusion model performs deep feature encoding on the chaotic feature vectors in the spatiotemporal feature pair sequence through the chaotic feature encoding module, generating a chaotic modal deep feature sequence.

[0165] The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence in the spatiotemporal feature pair sequence to generate a spectral modality deep feature sequence.

[0166] The attention fusion module performs weighted fusion of the chaotic mode deep feature sequence and the spectral mode deep feature sequence to generate a fused spatiotemporal feature vector sequence.

[0167] The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence to generate a probability sequence of faults occurring at each location point of the cable to be monitored within a future time range.

[0168] In a preferred embodiment, training the cross-modal fusion model includes:

[0169] Obtain a historical dataset of cable faults; wherein, the historical dataset of cable faults includes several training data, each training data including a sequence of historical spatiotemporal feature pairs, and a label of the actual fault state in the future time period corresponding to the historical spatiotemporal feature pair sequence in time;

[0170] According to the preset batch size, the historical cable fault dataset is randomly divided into several batches of training samples.

[0171] The training samples from each batch are sequentially input into the cross-modal fusion model, and end-to-end iterative training is performed on all learnable parameters in the cross-modal fusion model until a preset number of training rounds is reached. Among them, when the cross-modal fusion model receives a batch of training samples, it uses the chaotic feature encoding module to perform deep feature encoding on the chaotic feature vector of the historical spatiotemporal feature pair sequence in the current batch of training samples, generating a chaotic modal deep feature sequence during the training process.

[0172] The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence of the historical spatiotemporal feature pairs in the current batch of training samples to generate the spectral modality deep feature sequence during the training process.

[0173] The attention fusion module performs weighted fusion based on the chaotic mode deep feature sequence and the spectral mode deep feature sequence during the training process to generate a fused spatiotemporal feature vector sequence during the training process.

[0174] The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence during training to generate a predicted probability sequence of faults in the future time range corresponding to the current training samples. A preset loss function is used to calculate and generate a loss function value based on the predicted probability sequence and the corresponding real fault state label. A preset optimizer is then used to perform gradient backpropagation and update the learnable network parameters in the chaotic feature encoding module, the spectral feature encoding module, the attention fusion module, and the fault probability prediction module based on the loss function value.

[0175] Specifically, after generating a structured sequence of spatiotemporal feature pairs through spatiotemporal alignment and pairing, the core task of this invention is to perform in-depth analysis and time-series prediction of this sequence using a pre-defined cross-modal fusion model. Its fundamental purpose is to uncover the complex, nonlinear intrinsic correlation between internal chaotic features and external spectral fingerprints—two microscopic precursors—and, based on this correlation, to make quantitative and probabilistic accurate predictions of cable failure risks over future timeframes.

[0176] In a preferred embodiment, the cross-modal fusion model is a deep neural network model with a carefully designed internal structure to adapt to the dual-source heterogeneous data characteristics of the present invention. The model mainly includes a chaotic feature encoding module, a spectral feature encoding module, an attention fusion module, and a fault probability prediction module. Upon receiving a sequence of spatiotemporal feature pairs, the model's inference (i.e., prediction) process is as follows: First, the chaotic feature encoding module and the spectral feature encoding module, acting as two parallel "feature extractors," process the input chaotic feature vector sequence and spectral fingerprint feature sequence, respectively. These two sub-modules can employ structures such as one-dimensional convolutional neural networks (1D-CNN) or recurrent neural networks (RNN), aiming to learn and extract higher-dimensional deep features from their respective input sequences that can characterize the spatial correlation of fault symptoms along the line, generating chaotic modal deep feature sequences and spectral modal deep feature sequences, respectively.

[0177] Subsequently, the attention fusion module dynamically weights and fuses the two deep feature sequences. The core function of this module is to enable the model to adaptively learn the relative importance of chaotic and spectral features at each location point on the cable. For example, under certain operating conditions, electrical signal disturbances may be more critical; while under other conditions, spectral anomalies in the external environment may provide more decisive evidence. This dynamic weighting process can be implemented using a gating mechanism:

[0178]

[0179]

[0180] In the formula, Index of location points along the cable path; For the chaotic mode at position point The depth feature vector; For the spectral mode at position point The depth feature vector; This indicates that two feature vectors are concatenated; This represents the learnable weight matrix in the gated network; For learnable bias terms in gated networks; The sigmoid activation function is used to generate a weight gate in the range of 0 to 1. ; This indicates element-wise multiplication; For the final location point The generated spatiotemporal feature vector after attention-weighted fusion.

[0181] All location points Combined, these elements form a fused spatiotemporal feature vector sequence. Finally, the fault probability prediction module (which can consist of one or more fully connected layers) receives this fused feature sequence, performs nonlinear transformation and temporal analysis on it, and ultimately outputs a probability sequence that corresponds one-to-one with the input sequence positions, representing the probability of fault occurrence at each location point within a future time range. Visualizing this probability sequence along the cable path creates an intuitive and dynamic spatiotemporal fault probability cloud map.

[0182] In a preferred embodiment, the aforementioned cross-modal fusion model is trained using supervised learning. The training process includes: first, acquiring a historical cable fault dataset, which consists of a large number of training samples. Each training sample includes a "historical spatiotemporal feature pair sequence" extracted from historical monitoring data for a known time period, and a corresponding "real fault state label" (e.g., 0 represents no fault, 1 represents a fault) for a future time window after that time period. The dataset is then randomly divided into several batches of training samples according to a preset batch size.

[0183] During training, training samples from each batch are sequentially input into the model for end-to-end iterative training. Upon receiving each batch of training samples, the model first performs a forward propagation step identical to the inference process described above, generating a "predicted probability sequence" of future fault states. Subsequently, a pre-defined loss function is used to quantify the difference between this "predicted probability sequence" and the "true fault state label." For this type of binary classification probability prediction problem, the binary cross-entropy loss function can be used.

[0184]

[0185] In the formula, The calculated loss function value; This represents the total number of locations in the current batch. Location point The actual fault status label (0 or 1); For the model to the location points The output is the predicted probability.

[0186] In this embodiment, Obtained from the cable fault history dataset The model outputs the value in real time during forward propagation. Finally, using a pre-defined optimizer (such as the Adam optimizer), the calculated loss function value is used to... The gradient backpropagation algorithm is used to synchronously and jointly update all learnable network parameters within all four modules of the model (chaotic feature encoding module, spectral feature encoding module, attention fusion module, and fault probability prediction module). This iterative process is repeated until the model's predictive performance converges on the validation set or meets the preset number of training rounds.

[0187] Through the construction, reasoning, and training of the aforementioned deep learning model, this invention can automatically learn and establish an accurate fault prediction model from complex, multimodal microscopic precursor data, ultimately achieving highly sensitive and reliable early warning of latent faults.

[0188] Step S6: Determine whether there is a fault probability at any spatial location point in the spatiotemporal fault probability cloud map that is continuously higher than a preset threshold in the future time period; if so, determine that the cable to be monitored has a latent fault and trigger a fault warning; if not, determine that the cable to be monitored does not have a latent fault.

[0189] Specifically, after generating a spatiotemporal fault probability cloud map characterizing the future fault risk at various locations along the cable using a cross-modal fusion model, the final step of this invention is to establish an objective and reliable automated fault diagnosis and early warning decision-making mechanism based on this probability cloud map. Its core objective is to transform the continuous probability values ​​output by the model into clear diagnostic conclusions of "latent fault" or "non-latent fault" that can be directly used by maintenance personnel, and to trigger corresponding operations.

[0190] Judging solely based on instantaneous fault probability values ​​is susceptible to random noise or non-faulty transient disturbances, potentially leading to false alarms. To address this issue, a preferred embodiment of the present invention employs a joint judgment criterion based on both "amplitude" and "duration" conditions. This criterion aims to filter out random, short-lived probability fluctuations, issuing warnings only for anomalies exhibiting persistent, high-risk characteristics, thereby ensuring high reliability of the warnings. This criterion requires presetting two key thresholds: a fault probability threshold... and duration threshold These two thresholds can be set by those skilled in the art based on the cable's historical operation and maintenance data, safe operating procedures, and a trade-off between early warning sensitivity and false alarm rate.

[0191] The logic of this joint judgment criterion can be formally described by the following conditional expression. For any spatial location point in the spatiotemporal fault probability cloud map... In the future Warning status It is triggered if and only if the following condition is met:

[0192]

[0193] In the formula, The spatial location point output by the cross-modal fusion model At some point in the past The predicted failure probability value; This is the current monitoring time; In order to monitor the current time Any historical moment within the sliding time window that marks the end point; The preset duration threshold is used to define the length of the sliding time window; This is a preset fault probability threshold.

[0194] In this embodiment, The value is continuously output by the cross-modal fusion model. When the system detects that, for any spatial location point... ,in the past The predicted failure probability at all times within the time period. All are higher than the preset probability threshold. When the condition is true, the result of the above conditional expression is true.

[0195] Once the warning status is established at any spatial location... If the status is determined to be false, the system determines that the monitored cable has a latent fault at that location. The system then automatically triggers a fault warning mechanism. This mechanism may include: generating a warning report containing the fault warning level, the precise coordinates (or physical mileage) of the suspected fault location, the current fault probability value, and key information such as the chaotic characteristics and spectral fingerprint upon which the warning was triggered; and sending the warning report to maintenance personnel via the monitoring system interface, SMS, or a specified network protocol. If the warning status at all spatial locations remains false throughout the monitoring period, the system determines that the monitored cable has no latent fault and continues to maintain a silent monitoring state.

[0196] By using the joint judgment criteria based on the dual conditions of "amplitude and duration", this invention constructs an intelligent "early warning filter". It ensures that only those highly deterministic fault precursors with a continuous deterioration trend are identified as latent faults. In this way, while ensuring extremely high monitoring sensitivity, it effectively suppresses the false alarm rate and provides maintenance personnel with highly reliable and operable early warning decision support.

[0197] Based on the above method embodiments, the present invention provides corresponding system embodiments.

[0198] like Figure 2 As shown, an embodiment of the present invention provides an ultra-long-distance cable fault monitoring system, including: a data acquisition module, a chaotic feature extraction module, a spectral feature extraction module, a feature pairing module, a cross-modal fusion module, and a fault prediction module;

[0199] The data acquisition module is used to acquire high-frequency transient signals of partial discharge of the cable to be monitored, hyperspectral remote sensing images covering the cable route, and atmospheric parameter data.

[0200] The chaotic feature extraction module is used to reconstruct the phase space of the high-frequency transient signal of partial discharge, generate a phase space attractor, and construct a chaotic feature vector based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor.

[0201] The spectral feature extraction module is used to perform atmospheric correction on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images; and to perform spectral unmixing on the surface reflectance images based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples.

[0202] The feature pairing module is used to perform spatiotemporal alignment of the partial discharge high-frequency transient signal, hyperspectral remote sensing image and preset GIS route map of the cable to be monitored, generate a spatiotemporal mapping index, and pair the chaotic feature vector with the spectral fingerprint feature sequence according to the spatiotemporal mapping index to generate a spatiotemporal feature pair sequence.

[0203] The cross-modal fusion module is used to input the spatiotemporal feature pair sequence into the preset cross-modal fusion model to generate a probability sequence of faults generated at each location point of the cable to be monitored within a future time range, and to combine the probability sequences of all location points to construct a spatiotemporal fault probability cloud map.

[0204] The fault prediction module is used to determine whether there is a fault probability at any spatial location point in the spatiotemporal fault probability cloud map that is continuously higher than a preset threshold in the future time period; if so, it is determined that there is a latent fault in the cable to be monitored and a fault warning is triggered; if not, it is determined that there is no latent fault in the cable to be monitored.

[0205] It should be noted that the embodiments of the system described above correspond to the embodiments of the present invention described above, and can realize the ultra-long-distance cable fault monitoring method of any one of the present invention described above. Furthermore, the embodiments of the system described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the system embodiment drawings provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0206] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0207] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for monitoring faults in ultra-long-distance cables, characterized in that, include: Acquire high-frequency transient signals of partial discharge of the cable under monitoring, hyperspectral remote sensing images covering the cable route, and atmospheric parameter data; Phase space reconstruction is performed on the high-frequency transient signal of partial discharge to generate a phase space attractor, and a chaotic feature vector is constructed based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor. Atmospheric correction is performed on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images; spectral demixing is performed on the surface reflectance images based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples; The high-frequency transient signal of partial discharge, hyperspectral remote sensing image and preset GIS route map of the cable to be monitored are spatiotemporally aligned to generate a spatiotemporal mapping index. Based on the spatiotemporal mapping index, the chaotic feature vector is paired with the spectral fingerprint feature sequence to generate a spatiotemporal feature pair sequence. The spatiotemporal feature pairs are input into a preset cross-modal fusion model to generate a probability sequence of faults at each location point of the cable to be monitored within a future time range. The probability sequences of all location points are combined to construct a spatiotemporal fault probability cloud map. Determine whether there is any spatial location point in the spatiotemporal fault probability cloud map whose fault probability continuously exceeds a preset threshold in the future time period; if so, determine that the cable under monitoring has a latent fault and trigger a fault warning; if not, determine that the cable under monitoring does not have a latent fault. The step of performing spectral demixing on the surface reflectance image and extracting spectral fingerprint feature sequences according to a preset characteristic spectral model includes: The light intensity values ​​of each spectral band are extracted from the preset characteristic spectral model to construct the target spectral vector; Traverse each pixel along the path of the cable to be monitored in the surface reflectance image and extract the reflectance value of each pixel in each spectral band to generate the corresponding pixel spectral vector. For each pixel spectral vector, a matched filtering algorithm is used to calculate the projection value of the current pixel spectral vector in the direction of the target spectral vector, and generate the corresponding projection result for each pixel. The projection results are normalized to generate the corresponding abundance values. The abundance values ​​are arranged in order of the cable path to form a spectral fingerprint feature sequence; The step of spatiotemporally aligning the partial discharge high-frequency transient signal, hyperspectral remote sensing image, and a preset GIS route map of the cable to be monitored to generate a spatiotemporal mapping index includes: Event detection is performed on the high-frequency transient signal of partial discharge to determine the occurrence time of each partial discharge event and generate an event time list; For each partial discharge event, based on the high-frequency transient signal of the partial discharge, the location of the event on the cable to be monitored is determined, and a list of event spatial locations is generated. Based on the event time list and the event spatial location list, the time of each partial discharge event is paired with its corresponding spatial location to generate an event time-space pair list; The spatial locations in the event spatiotemporal pair list are mapped to the corresponding hyperspectral image pixel locations using a preset GIS path map of the cable to be monitored, thereby generating an event pixel correspondence table. By combining the event spatiotemporal pair list with the event pixel correspondence table, a spatiotemporal mapping index is generated.

2. The method for monitoring faults in ultra-long-distance cables as described in claim 1, characterized in that, The step of reconstructing the phase space of the high-frequency transient signal of partial discharge to generate a phase space attractor includes: The high-frequency transient signal of partial discharge is sampled at fixed time intervals to generate a sampling sequence; The sampled sequence is detrended to obtain a detrended signal sequence; The amplitude normalization process is performed on the detrended signal sequence to obtain the preprocessed signal sequence; Based on the preprocessed signal sequence, the autocorrelation function curve is calculated to obtain the autocorrelation function data. The first minimum point of the autocorrelation function data is determined, and the time corresponding to the first minimum point is determined as the reconstruction delay time; Based on the preprocessed signal sequence, the optimal embedding dimension is determined using the pseudo-neighbor method. Based on the reconstruction delay time and the optimal embedding dimension, the preprocessed signal sequence is delayed and sampled to generate a multidimensional state vector sequence. The multidimensional state vector sequence is arranged in chronological order to form a phase space trajectory, and the phase space trajectory is used as a phase space attractor.

3. The method for monitoring faults in ultra-long-distance cables as described in claim 1, characterized in that, The step of performing atmospheric correction on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images includes: Radiometric calibration is performed on the raw pixel values ​​of the hyperspectral remote sensing image to generate radiance data; Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the air scattering effect on the radiance data is calculated to generate scattering correction data. Based on the atmospheric parameter data, and according to the preset atmospheric radiative transfer model, the atmospheric absorption effect on the radiance data is calculated, and absorption correction data is generated. Based on the scattering correction data and absorption correction data, the radiance data is corrected to generate an atmospherically corrected surface reflectance image.

4. The method for monitoring faults in ultra-long-distance cables as described in claim 1, characterized in that, The step of pairing chaotic feature vectors with spectral fingerprint feature sequences based on spatiotemporal mapping indexes to generate spatiotemporal feature pair sequences includes: Iterate through each record in the spatiotemporal mapping index and retrieve the occurrence time and corresponding pixel position of each event from the record; For each event, based on the occurrence time of the current event, extract the chaotic feature vector corresponding to the current event from the chaotic feature vector time series; based on the pixel position of the current event, extract the spectral fingerprint feature value corresponding to the current event from the spectral fingerprint feature sequence. The chaotic feature vector is paired with the spectral fingerprint feature value to form a spatiotemporal feature pair for each event; Arrange all spatiotemporal feature pairs according to the spatial order of the cable path to be monitored, and generate a spatiotemporal feature pair sequence.

5. The method for monitoring faults in ultra-long-distance cables as described in claim 1, characterized in that, The cross-modal fusion model includes a chaotic feature encoding module, a spectral feature encoding module, an attention fusion module, and a fault probability prediction module; The step of inputting the spatiotemporal feature pair sequence into a preset cross-modal fusion model to generate a probability sequence of fault generation at each location point of the cable to be monitored within a future time range includes: The spatiotemporal feature pair sequence is input into a preset cross-modal fusion model, so that the cross-modal fusion model performs deep feature encoding on the chaotic feature vectors in the spatiotemporal feature pair sequence through the chaotic feature encoding module, generating a chaotic modal deep feature sequence. The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence in the spatiotemporal feature pair sequence to generate a spectral modality deep feature sequence. The attention fusion module performs weighted fusion of the chaotic mode deep feature sequence and the spectral mode deep feature sequence to generate a fused spatiotemporal feature vector sequence. The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence to generate a probability sequence of faults occurring at each location point of the cable to be monitored within a future time range.

6. The method for monitoring faults in ultra-long-distance cables as described in claim 5, characterized in that, The training of the cross-modal fusion model includes: Obtain a historical dataset of cable faults; wherein, the historical dataset of cable faults includes several training data, each training data including a sequence of historical spatiotemporal feature pairs, and a label of the actual fault state in the future time period corresponding to the historical spatiotemporal feature pair sequence in time; According to the preset batch size, the historical cable fault dataset is randomly divided into several batches of training samples. The training samples from each batch are sequentially input into the cross-modal fusion model, and end-to-end iterative training is performed on all learnable parameters in the cross-modal fusion model until a preset number of training rounds is reached. Among them, when the cross-modal fusion model receives a batch of training samples, it uses the chaotic feature encoding module to perform deep feature encoding on the chaotic feature vector of the historical spatiotemporal feature pair sequence in the current batch of training samples, generating a chaotic modal deep feature sequence during the training process. The spectral feature encoding module performs deep feature encoding on the spectral fingerprint feature sequence of the historical spatiotemporal feature pairs in the current batch of training samples to generate the spectral modality deep feature sequence during the training process. The attention fusion module performs weighted fusion based on the chaotic mode deep feature sequence and the spectral mode deep feature sequence during the training process to generate a fused spatiotemporal feature vector sequence during the training process. The fault probability prediction module performs time-series prediction based on the fused spatiotemporal feature vector sequence during training to generate a predicted probability sequence of faults in the future time range corresponding to the current training samples. A preset loss function is used to calculate and generate a loss function value based on the predicted probability sequence and the corresponding real fault state label. A preset optimizer is then used to perform gradient backpropagation and update the learnable network parameters in the chaotic feature encoding module, the spectral feature encoding module, the attention fusion module, and the fault probability prediction module based on the loss function value.

7. The method for monitoring faults in ultra-long-distance cables as described in claim 1, characterized in that, The characteristic spectral model is constructed using the following method: Obtain laboratory spectral data; wherein the laboratory spectral data is obtained by applying excitation to a cable sample of the same type as the cable to be monitored in a controlled experimental device to induce local micro-discharge, and using a spectrometer to collect the optical radiation signal of the discharge phenomenon; The laboratory spectral data are subjected to noise filtering and baseline correction to generate standardized spectral sample data; In the standardized spectral sample data, characteristic spectral lines representing local micro-discharge phenomena are identified and extracted using a peak detection algorithm; The wavelengths of the characteristic spectral lines and their corresponding relative light intensity values ​​are digitized and stored to construct a characteristic spectral model.

8. A fault monitoring system for ultra-long-distance cables, characterized in that, include: The module includes a data acquisition module, a chaotic feature extraction module, a spectral feature extraction module, a feature pairing module, a cross-modal fusion module, and a fault prediction module. The data acquisition module is used to acquire high-frequency transient signals of partial discharge of the cable to be monitored, hyperspectral remote sensing images covering the cable route, and atmospheric parameter data. The chaotic feature extraction module is used to reconstruct the phase space of the high-frequency transient signal of partial discharge, generate a phase space attractor, and construct a chaotic feature vector based on the maximum Lyapunov exponent and fractal dimension of the phase space attractor. The spectral feature extraction module is used to perform atmospheric correction on hyperspectral remote sensing images based on atmospheric parameter data to generate surface reflectance images; and to perform spectral unmixing on the surface reflectance images based on a preset characteristic spectral model to extract spectral fingerprint feature sequences; wherein, the characteristic spectral model is a digital data model obtained by spectral measurement of micro-discharge of controlled cable samples. The feature pairing module is used to perform spatiotemporal alignment of the partial discharge high-frequency transient signal, hyperspectral remote sensing image and preset GIS route map of the cable to be monitored, generate a spatiotemporal mapping index, and pair the chaotic feature vector with the spectral fingerprint feature sequence according to the spatiotemporal mapping index to generate a spatiotemporal feature pair sequence. The cross-modal fusion module is used to input the spatiotemporal feature pair sequence into the preset cross-modal fusion model to generate a probability sequence of faults generated at each location point of the cable to be monitored within a future time range, and to combine the probability sequences of all location points to construct a spatiotemporal fault probability cloud map. The fault prediction module is used to determine whether there is a fault probability at any spatial location point in the spatiotemporal fault probability cloud map that is continuously higher than a preset threshold in the future time period; if so, it is determined that there is a latent fault in the cable to be monitored and a fault warning is triggered; if not, it is determined that there is no latent fault in the cable to be monitored. The step of performing spectral demixing on the surface reflectance image and extracting spectral fingerprint feature sequences according to a preset characteristic spectral model includes: The light intensity values ​​of each spectral band are extracted from the preset characteristic spectral model to construct the target spectral vector; Traverse each pixel along the path of the cable to be monitored in the surface reflectance image and extract the reflectance value of each pixel in each spectral band to generate the corresponding pixel spectral vector. For each pixel spectral vector, a matched filtering algorithm is used to calculate the projection value of the current pixel spectral vector in the direction of the target spectral vector, and generate the corresponding projection result for each pixel. The projection results are normalized to generate the corresponding abundance values. The abundance values ​​are arranged in order of the cable path to form a spectral fingerprint feature sequence; The step of spatiotemporally aligning the partial discharge high-frequency transient signal, hyperspectral remote sensing image, and a preset GIS route map of the cable to be monitored to generate a spatiotemporal mapping index includes: Event detection is performed on the high-frequency transient signal of partial discharge to determine the occurrence time of each partial discharge event and generate an event time list; For each partial discharge event, based on the high-frequency transient signal of the partial discharge, the location of the event on the cable to be monitored is determined, and a list of event spatial locations is generated. Based on the event time list and the event spatial location list, the time of each partial discharge event is paired with its corresponding spatial location to generate an event time-space pair list; The spatial locations in the event spatiotemporal pair list are mapped to the corresponding hyperspectral image pixel locations using a preset GIS path map of the cable to be monitored, thereby generating an event pixel correspondence table. By combining the event spatiotemporal pair list with the event pixel correspondence table, a spatiotemporal mapping index is generated.

Images