Multi-source data driven cable operation state comprehensive evaluation method

Abstract

The invention relates to the technical field of cable operation state detection, and particularly discloses a multi-source data driven cable operation state comprehensive evaluation method, which comprises the following steps of S1, adopting a layered distributed sensing network architecture, and deploying three types of core sensors at key nodes of a cable, through space-time calibration of the multi-source heterogeneous sensor, data consistency is improved, fusion deviation is eliminated, the problem of data islands of a traditional system is solved, and a precise evaluation foundation is laid; noise suppression and dynamic correlation modeling are adopted, environmental interference is stripped, a vibration and displacement coupling relation is quantified, limitation of a single parameter is broken through, heterogeneous fault features are captured, and evaluation comprehensiveness and sensitivity are improved; a self-adaptive threshold mechanism is constructed based on environment weight and historical data, the bottleneck of a fixed threshold is broken through, an evaluation standard is corrected along with equipment aging and environment change, misjudgment is avoided, and diagnosis robustness in different scenes is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable operation state detection, and particularly relates to a multi-source data driven comprehensive evaluation method for cable operation state. BACKGROUND

[0002] At present, as the core carrier of urban power grid and power transmission and distribution system, the operation reliability of power cable directly affects power supply safety. Cable state evaluation needs to monitor multi-dimensional parameters such as mechanical vibration, deformation displacement and environmental conditions in real time, and potential fault risks are predicted through multi-source data analysis. The existing technology is usually based on time domain or frequency domain feature analysis of a single sensor (such as a vibration sensor or a temperature sensor), combined with a fixed threshold value to trigger an alarm mechanism. Some schemes use a multi-sensor sub-item monitoring strategy, respectively establish independent analysis models for vibration and displacement, and make a comprehensive judgment after superimposing the results.

[0003] However, due to the difference in sampling frequency or the asynchronization of time and space reference, the fusion accuracy of different sensors is insufficient, and it is difficult to accurately associate the multi-source response characteristics of the same physical event. False feature signals are easily generated under environmental noise interference, making it difficult to distinguish local defects from external disturbances. The fixed threshold value mechanism cannot adapt to the state drift caused by cable material aging and extreme environment, and frequent false alarms or false alarms occur. The evaluation model lacks adaptability to feature evolution caused by newly connected sensors or long-term operation, and the parameters need to be frequently adjusted manually. SUMMARY

[0004] The purpose of the present application is to provide a multi-source data driven comprehensive evaluation method and system for cable operation state, which adopts multi-source data time and space synchronization, dynamic coupling feature modeling and adaptive threshold generation technology, and can realize accurate dynamic evaluation of cable mechanical performance and intelligent early fault prediction.

[0005] A multi-source data driven comprehensive evaluation method for cable operation state, comprising the following steps:

[0006] S1: A hierarchical distributed sensor network architecture is adopted, and three types of core sensors are deployed at key nodes of the cable: a piezoelectric sensor is selected for a high-frequency micro-vibration sensor, a temperature compensation module is built-in, and axial and radial vibration acceleration data are synchronously collected; a wavelength demodulation technology is used for a fiber grating micro-displacement sensor, 360° circumferential displacement monitoring is realized through distributed optical fiber arrangement, and displacement vector data are generated; an array of environmental sensors integrates temperature, humidity, wind speed, air pressure and corrosion gas sensors, and an environmental parameter time and space distribution field is constructed, all sensors realize real-time data uploading through a 5G edge node, a time stamp encryption mechanism is used to ensure the traceability of original data,

[0007] S2: A spatiotemporal dual-dimensional adaptive calibration algorithm is proposed. For time calibration, the optical frequency clock of the fiber optic grating sensor is used as the reference. The sliding window cross-correlation correction method is used for vibration sensor data to calculate the time offset Δt between adjacent data. A dynamic compensation curve is generated by quadratic polynomial fitting to improve the accuracy of vibration data timestamp correction. For spatial calibration, a three-dimensional coordinate model of the cable is established. The spatial interpolation mapping method is used for sensor data at different locations. Based on the Delaunay triangulation algorithm, discrete monitoring point data is mapped to grid cells. Spatial heterogeneity is eliminated by smoothing with a Gaussian kernel function. The calibrated data can only proceed to the next stage after passing the spatiotemporal consistency test.

[0008] S3: A composite denoising strategy is designed for different types of data characteristics. For vibration data, an improved variational mode decomposition-wavelet thresholding joint algorithm is used. First, the optimal decomposition level K and penalty factor α of VMD are determined by particle swarm optimization algorithm to separate modal components. An improved soft and hard threshold trade-off function is used to process noisy components to improve the signal-to-noise ratio. For displacement data, a Kalman filter-empirical mode decomposition fusion algorithm is used. First, high-frequency random noise is removed by Kalman filtering, and then EEMD decomposition is performed on the residual signal to remove noisy intrinsic mode functions. For environmental data, a moving average-outlier detection joint algorithm is used. Local weighted regression algorithm is used to repair outlier data to improve data smoothness.

[0009] S4: Construct a multi-domain feature set. In the time domain, extract statistics such as peak factor, kurtosis, and impulse index of the vibration signal, and indicators such as mean, variance, and maximum deviation of the displacement data. In the frequency domain, obtain the main frequency distribution and spectral entropy of the vibration signal through fast Fourier transform, and extract the energy proportion of the frequency band using wavelet packet decomposition. In the time-frequency domain, generate the time-frequency matrix of the vibration signal using short-time Fourier transform, extract features such as singular value entropy and energy moment of the matrix, and perform feature dimensionality reduction using the Relief-F algorithm to retain the core features.

[0010] S5: A dynamic update mechanism for the spatiotemporal correlation matrix is ​​proposed. A cross-correlation matrix R(i,j) between vibration acceleration and displacement is constructed using a set time window, where i is the vibration feature dimension and j is the displacement feature dimension. A time decay factor θ is introduced to perform weighted iterative updates on historical matrix data, enhancing the weight of the latest data. The principal eigenvalues ​​are extracted through matrix singular value decomposition to construct a dynamically coupled eigenvector, reflecting the dynamic correlation strength between vibration and displacement.

[0011] S6: Design a dynamic coupling model for environmental factors, use partial least squares regression algorithm to calculate the correlation coefficient between dynamic coupling characteristics and each environmental parameter, and establish a coupling coefficient matrix; introduce environmental impact weight factors, determine the weight ratio of each environmental factor through analytic hierarchy process; calculate the comprehensive coupling coefficient to quantify the degree of influence of environmental factors on cable condition;

[0012] S7: Construct a dynamic threshold adjustment mechanism, generate an initial threshold range based on historical normal state data using a kernel density estimation algorithm; combine real-time environmental coupling coefficients and use a piecewise function to dynamically correct the thresholds; introduce an equipment aging coefficient to make annual corrections to ensure that the thresholds dynamically adapt to the equipment lifecycle.

[0013] S8: An improved deep learning fusion model is adopted, with the model architecture based on a bidirectional long short-term memory network. The input layer receives dynamically coupled features, and multiple hidden layers are set. An attention mechanism is added to strengthen the weights of key features. Training optimization uses a hybrid optimization algorithm to optimize model parameters and early stopping is used to prevent overfitting. Model validation uses cross-validation to ensure model accuracy.

[0014] S9: Design a five-level evaluation system. Level 1 is when the dynamic coupling feature is in the low threshold range and there is no abnormal fluctuation. Level 2 is when the feature is in the low threshold range and there is slight fluctuation. Level 3 is when the feature is in the medium threshold range and needs attention. Level 4 is when the feature is in the high threshold range and maintenance is recommended. Level 5 is when the feature exceeds the threshold range and an emergency warning is issued. When the model outputs the level, a feature contribution report is generated simultaneously to identify the main factors that lead to the state deterioration.

[0015] S10: Establish a closed-loop optimization mechanism, regularly collect the deviation between actual maintenance data and evaluation results, and calculate the correction coefficient; adopt an online incremental learning algorithm to accumulate new data to fine-tune the model, and update the training set through a sliding window; regularly calibrate the threshold generation mechanism, and optimize the coupling coefficient calculation model in combination with environmental change trends to maintain the evaluation accuracy.

[0016] Furthermore, S2 specifically includes employing a dual-reference time anchoring method for vibration and displacement data. The optical frequency clock of the fiber optic grating sensor serves as the primary reference, while the timing signal from the 5G base station is introduced as the secondary reference. A sliding window cross-correlation algorithm is used to calculate the time offset Δt between adjacent data points. A dynamic correction curve is generated using a temperature drift compensation model. The vibration data timestamps are calibrated in real time. The calibrated data undergoes time consistency verification to generate time-synchronized data with confidence markers. Based on a preset three-dimensional coordinate model of the cable, a layered mesh mapping algorithm is used. A uniform mesh is applied to straight cable sections, while a denser mesh is used for complex structures such as bends and joints. This is achieved through Delaunay... Nay triangulation is used to construct spatial topological relationships. When converting the discrete monitoring point coordinates of time-synchronized data into grid node coordinates, a spatial attenuation factor is introduced to weight and correct the data in the edge region, eliminating spatial heterogeneity errors. For environmental data and spatiotemporally calibrated data, a dynamic time bending-sliding window fusion algorithm is adopted. First, the optimal time matching path is calculated using the DTW algorithm. Then, the environmental data is interpolated or downsampled based on the sampling frequency of the vibration data. Trend consistency verification is introduced during the alignment process. When the deviation of the trend exceeds the threshold, a local realignment mechanism is triggered. Finally, multi-source synchronized data containing spatiotemporal coordinates, data values, and quality scores are generated, providing a reliable basis for subsequent processing.

[0017] Furthermore, S2 also includes noise suppression processing of multi-source synchronous data to generate denoised multi-source feature data. Specifically, this includes using an adaptive frequency band optimization partitioning method for vibration and displacement data, combining the spectral feature library of cable operation history data, and performing cluster analysis on the signal spectrum after fast Fourier transform using an improved spectral clustering algorithm to automatically identify key frequency bands and noise-dominant frequency bands related to cable faults, achieving accurate partitioning of 3-6 dynamic frequency bands. Differentiated denoising strategies are adopted for the characteristics of each frequency band. For key frequency bands, an improved variational mode decomposition algorithm is used to introduce a kurtosis criterion to optimize decomposition parameters and separate effective signals from noise. For noise-dominant frequency bands, wavelet packet threshold denoising is used, with the threshold dynamically adjusted according to the frequency band energy. The signal is then decomposed into frequency bands. A dynamic compensation model for environmental interference is constructed, and a three-dimensional interference coefficient matrix is ​​established based on temperature, humidity, and wind speed data. The temporal correlation between environmental parameters and noise energy is learned through a long short-term memory network, and the environmental interference compensation amount corresponding to each frequency band is output in real time. Piecewise linear compensation is used for temperature interference, a hysteresis correction factor is introduced for humidity interference, and a turbulence intensity correction model is used for wind speed interference. The frequency band decomposed signal is subjected to two-stage interference subtraction. First, the environmental interference compensation amount is subtracted, and then residual impulse noise is removed by verifying the change rate of adjacent sampling points is less than a threshold. Finally, denoised multi-source feature data is generated, along with noise suppression efficiency indicators for each frequency band: effective signal retention rate ≥90%, and noise removal rate ≥85%.

[0018] Furthermore, the construction of the dynamically coupled feature vector in S5 specifically includes the following steps:

[0019] 1) Feature extraction: From the denoised multi-source feature data, an improved time-frequency domain fusion extraction method is used to extract vibration energy spectrum features and displacement time-domain waveform features. For vibration energy spectrum features, a short-time Fourier transform is performed on the vibration signal to obtain the time-frequency matrix, and the energy proportion of each frequency unit is calculated (formula: E_k=∫(t1 to t2)|S(k,t)|2dt / ∫(t1 to t2)∑|S(k,t)|2dt, where S(k,t) is the time-frequency amplitude at frequency k). (The analysis period is from t1 to t2). The top 5 frequency units by energy percentage and their corresponding energy change rates are selected to form the vibration energy spectrum feature vector E = [E1, E1', E2, E2', ..., E5, E5']. For displacement time-domain waveform features, the peak value, mean, variance, kurtosis, and waveform slope change rate of the displacement signal are extracted to form the displacement time-domain waveform feature vector D = [D_peak, D_mean, D_var, D_kurt, D_slope].

[0020] 2) Correlation index calculation: The correlation index within a preset time period is calculated using a sliding window weighted correlation algorithm. The time window length is set to T, and the sliding step size is T / 5. Within each window, the correlation index is calculated through the following steps:

[0021] 3) Calculate the Pearson correlation coefficient matrix R between the vibration energy spectrum eigenvector E and the displacement time-domain waveform eigenvector D, where R(i,j)=Cov(E_i,D_j) / (σ_Ei·σ_Dj), Cov is the covariance, and σ is the standard deviation;

[0022] 4) Introduce feature importance weights ω (obtained by training historical fault data using the random forest algorithm, where ω_i represents the weight of the i-th vibration feature and ω_j represents the weight of the j-th displacement feature);

[0023] 5) Calculate the comprehensive correlation index C within the window = ∑(i = 1 to 10)∑(j = 1 to 5)ω_i·ω_j·|R(i,j)|, to obtain the correlation index sequence {C1,C2,...,Cn} within the preset time period;

[0024] 6) Dynamic coupling feature generation: Based on the correlation index sequence, a multi-scale coupling feature vector is constructed. Wavelet packet decomposition is performed on the correlation index sequence to obtain the decomposition coefficients of three scales. The energy values ​​E_scale1, E_scale2, and E_scale3 of each scale are calculated. Combined with the mean μ_C, standard deviation σ_C, and maximum fluctuation value ΔC_max of the correlation index, a dynamic coupling feature vector V=[μ_C,σ_C,ΔC_max,E_scale1,E_scale2,E_scale3] is generated.

[0025] 7) Pattern segmentation: The adaptive threshold clustering method is used to divide the dynamic coupling features into local defect patterns and structural loosening patterns according to the time window;

[0026] 8) Based on historical data, K-means clustering is used to obtain the central feature vectors V_defect (local defects) and V_loose (structural looseness) of the two types of patterns;

[0027] Calculate the Euclidean distance between the current window's dynamically coupled feature vector V and the center vectors of the two classes: d_defect = ||V - V_defect||, d_loose = ||V - V_loose||;

[0028] 9) Introduce a dynamic discrimination threshold θ = 0.5 × (d_defect_avg + d_loose_avg), where d_defect_avg and d_loose_avg are the historical average distances;

[0029] 10) When d_defect < θ, it is determined to be a local defect mode; when d_loose < θ, it is determined to be a structural loosening mode; when neither of them is satisfied, it is marked as a mixed mode and a secondary check is triggered to increase the window length and recalculate.

[0030] Furthermore, the step S7, which generates the initial threshold interval based on historical normal state data using a kernel density estimation algorithm, includes the following steps:

[0031] 1): Calculation of environmental coupling weight coefficients: The improved analytic hierarchy process (AHP)-entropy weight fusion algorithm is used to calculate the environmental coupling weight coefficients of temperature, humidity and wind speed data.

[0032] 2): Construct a judgment matrix. Based on the significance of the influence of temperature, humidity and wind speed on cable condition in historical data, establish a 3×3 judgment matrix A, where A(i,j) represents the ratio of the influence of the i-th environmental parameter on the j-th parameter (1-9 scale method).

[0033] 3): The weight calculation of the analytic hierarchy process (AHP) is carried out by solving the maximum eigenvalue λmax and the corresponding eigenvector of the judgment matrix through the eigenvalue method. After the consistency test (CR<0.1), the subjective weight ω_sub=[ω_T,ω_H,ω_W] (weight of temperature, humidity and wind speed) is obtained.

[0034] 4): Entropy weight method for weight calculation. The environmental parameter data is standardized (x_ij'=(x_ij-minx_j) / (maxx_j-minx_j)), and the information entropy H_j=-k∑x_ij'lnx_ij' (k=1 / lnn, n is the number of samples) is calculated to obtain the objective weight ω_obj=[ω_T',ω_H',ω_W'], where ω_j'=(1-H_j) / ∑(1-H_j);

[0035] 5): Weight fusion, introducing a dynamic balance coefficient α (dynamically adjusted according to environmental stability, 0.3≤α≤0.7), and calculating the comprehensive coupling weight coefficient ω=[αω_T+(1-α)ω_T',αω_H+(1-α)ω_H',αω_W+(1-α)ω_W'], to ensure that the weights simultaneously reflect subjective experience and objective data characteristics;

[0036] 6): The dynamic threshold range is generated by nonlinear superposition, using a hyperbolic tangent function nonlinear superposition model instead of simple addition;

[0037] 7): The basic threshold is set to the upper limit of the 95% confidence interval of historical normal data estimated by Th_base using kernel density estimation;

[0038] 8): Calculate the environmental impact correction factor f(ω,T,H,W)=tanh(ω_T·T_norm+ω_H·H_norm+ω_W·W_norm), where T_norm, H_norm, and W_norm are the standardized values ​​of temperature, humidity, and wind speed (range 0-1). The tanh function ensures that the correction factor changes smoothly between 0 and 1.

[0039] 9): Dynamic threshold range generation, upper limit Th_up=Th_base×(1+k1·f(ω,T,H,W)), lower limit Th_low=Th_base×(1-k2·f(ω,T,H,W)), where k1 and k2 are nonlinear adjustment coefficients (k1=0.2-0.5, k2=0.1-0.3, set according to equipment type). This model realizes the nonlinear mapping of environmental influences and avoids threshold mutations caused by linear superposition.

[0040] 10): The upper and lower limits of the historical fault data correction threshold are adopted using a fault density weighted correction algorithm. Constructing a fault-environment correlation matrix: Historical fault data are categorized by temperature, humidity, and wind speed intervals, and the fault frequency density ρ(i,j,k) (number of faults in the i-th temperature interval, j-th humidity interval, and k-th wind speed interval / total number of samples) is calculated within each interval. Correction coefficients are calculated: Correction coefficients Δ_up=∑ρ(i,j,k)×δ_up(i,j,k) and Δ_low=∑ρ(i,j,k)×δ_low(i,j,k) are calculated for the upper and lower limits of the dynamic threshold, respectively, where δ_up(i,j,k) is the upper limit correction amount corresponding to the fault in that interval (the average difference between the feature value at the time of the fault and the original upper limit), and δ_low(i,j,k) is the lower limit correction amount. Corrected thresholds: The final adaptive evaluation threshold has an upper limit Th_up'=Th_up+Δ_up and a lower limit Th_low'=Th_low-Δ_low, ensuring that the threshold is more stringent in high-fault-risk environmental intervals and more lenient in low-risk intervals.

[0041] Furthermore, in S10, a sliding window clustering algorithm is used to classify the dynamic coupling features in real time to generate a preliminary state level; the preliminary state level is then jointly determined with the adaptive evaluation threshold to generate a cable operating state level.

[0042] Furthermore, the method also includes: obtaining the difference between the cable operating status level and the actual fault record within a preset period; adjusting the weight distribution of the dynamic coupling feature in reverse according to the difference value to generate an updated dynamic evaluation model; and expanding the input dimension of the dynamic evaluation model and retraining the classification boundary when a new sensor is connected.

[0043] Furthermore, the determination of the structural loosening mode includes identifying the amplitude of the low-frequency main peak of the vibration energy spectrum and the displacement time-domain increment; when the amplitude of the low-frequency main peak exceeds a preset proportion of the displacement time-domain increment, a structural loosening warning signal is triggered.

[0044] Furthermore, the method also includes: generating a three-dimensional stress distribution simulation map of the cable based on the structural loosening early warning signal, combining the wind speed and direction parameters in the environmental data, calibrating the preset offset of the loosening physical location, comparing the preset offset with a preset safety margin threshold, and outputting maintenance priority suggestions.

[0045] Furthermore, a multi-dimensional synchronous acquisition module is used to acquire vibration data output by a preset high-frequency micro-vibration sensor, displacement data output by a preset fiber optic grating micro-displacement sensor, and temperature, humidity, and wind speed data from a preset environmental sensor at key cable nodes.

[0046] The spatiotemporal calibration module is used to perform spatiotemporal calibration on the vibration data, displacement data and environmental data to generate multi-source synchronous data;

[0047] The noise suppression module is used to perform noise suppression processing on the multi-source synchronous data to generate denoised multi-source feature data.

[0048] The dynamic coupling analysis module is used to construct a vibration-displacement correlation matrix based on the denoised multi-source feature data and generate dynamic coupling features.

[0049] An adaptive threshold module is used to generate an adaptive evaluation threshold based on the coupling coefficient between the dynamic coupling characteristics and the environmental data.

[0050] The status assessment module is used to input the dynamic coupling characteristics and adaptive assessment threshold into a preset dynamic assessment model and output the cable operating status level.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] 1. This invention significantly improves the spatiotemporal consistency of vibration, displacement, and environmental data through spatiotemporal calibration and joint calibration of multi-source heterogeneous sensors, effectively eliminating feature fusion deviations caused by differences in sampling frequencies or inconsistent measurement benchmarks. It solves the problems of data silos and collaborative failures in traditional multi-sensor systems. On this basis, it innovatively introduces a multi-source synchronous data mechanism with confidence labels. When matching environmental data and physical quantity data through a dynamic time bending-sliding window fusion algorithm, a trend consistency check and local realignment mechanism are set. When the trend deviation exceeds the threshold, a secondary calibration is automatically triggered, improving the accuracy of data quality scoring by more than 25%. For complex structures such as cable bending sections and joints, a denser grid is used, and the edge area data is weighted and corrected by a spatial attenuation factor. Compared with traditional uniform grid mapping, the spatial heterogeneity error is further reduced by 15-20%, laying a more accurate data foundation for the comprehensive evaluation of cable mechanical performance.

[0053] 2. This invention employs a hierarchical analysis strategy combining noise suppression and dynamic correlation feature modeling to extract the mechanical response signal of the cable body from complex environmental interference and quantify the time-varying coupling relationship between vibration and displacement. This overcomes the limitations of single-parameter analysis and can simultaneously capture subtle features of heterogeneous fault modes such as partial discharge and structural loosening. Furthermore, a dynamic compensation model for environmental interference is constructed, establishing a three-dimensional interference coefficient matrix based on temperature, humidity, and wind speed data. By learning the temporal correlation between environmental parameters and noise energy through a long short-term memory network, accurate deduction of environmental interference in each frequency band is achieved. Piecewise linear compensation is used for temperature interference, a hysteresis correction factor is introduced for humidity interference, and a turbulence intensity correction model is combined with wind speed interference, resulting in an effective signal retention rate of over 90% and a noise removal rate of over 85%. In addition, the construction of the vibration-displacement joint time-varying feature matrix incorporates kurtosis criteria and energy proportion analysis, enabling the differentiation between high-frequency pulses of partial discharge and low-frequency vibration characteristics of structural loosening. The identification response speed for heterogeneous fault modes is improved to the second level, detecting potential hazards 3-5 monitoring cycles earlier than traditional methods, further enhancing the comprehensiveness and sensitivity of condition assessment.

[0054] 3. This invention utilizes an adaptive threshold dynamic generation mechanism based on environmental weight factors and historical operating data, breaking through the bottleneck of traditional static evaluation with fixed thresholds. This allows the evaluation standard to autonomously adjust with cable aging and changes in the external environment, avoiding misjudgments and omissions under extreme conditions. Simultaneously, it achieves dual dynamic threshold adjustment: on the one hand, through a fault density weighted algorithm, it automatically tightens the threshold in high-fault-risk environments and moderately widens it in low-risk environments, reducing the misjudgment rate under extreme conditions by more than 40%; on the other hand, it introduces an annual correction mechanism for equipment aging coefficients, dynamically adjusting the threshold baseline in conjunction with the cable's service life. For cables over 10 years old, it automatically increases threshold sensitivity by 15-20%, avoiding lag in evaluation standards due to equipment aging. The system also supports dimensional expansion and classification boundary retraining when new sensors are added, requiring only 300-800 sets of new data to complete model adaptation, significantly enhancing the robustness of cable condition diagnosis in different regions and service stages. Attached Figure Description

[0055] Fig. 1 This is a flowchart illustrating the multi-source data-driven comprehensive evaluation method for cable operating status according to an embodiment of the present invention.

[0056] Fig. 2 This is a schematic diagram of the structure of a multi-source data-driven comprehensive evaluation system for cable operating status according to an embodiment of the present invention. Detailed Implementation

[0057] Example

[0058] In the embodiments of this invention, please refer to Figs. 1-2A multi-source data-driven comprehensive evaluation method for cable operating status includes the following steps:

[0059] S1: Employing a hierarchical distributed sensor network architecture, three types of core sensors are deployed at key cable nodes. The high-frequency micro-vibration sensor is a piezoelectric sensor with a sampling rate of 12-20kHz, featuring a built-in temperature compensation module and a sampling period of 30-100ms. It simultaneously acquires axial and radial vibration acceleration data, enabling it to sensitively capture minute vibrations caused by internal structural changes or external interference during cable operation. The fiber optic grating micro-displacement sensor utilizes wavelength demodulation technology, achieving a spatial resolution of 0.05-0.2mm. Through distributed fiber optic deployment, it enables 360° circumferential displacement monitoring, generating a set of displacement vector data every 50-200ms, comprehensively reflecting displacement changes at various parts of the cable. The environmental sensor array integrates temperature, humidity, wind speed, air pressure, and corrosive gas sensors, with temperature and humidity accuracy of ±0. With a temperature range of 0.3-0.8℃ / ±1-3%RH and a wind speed measurement range of 0-40m / s with an accuracy of ±0.2-0.5m / s, and a sampling frequency of 0.5-2Hz, an environmental parameter spatiotemporal distribution field is constructed to provide comprehensive data for subsequent analysis of the impact of the environment on cable conditions. All sensors achieve real-time data upload through 5G edge nodes, and a timestamp encryption mechanism is used to ensure the traceability of the original data, guaranteeing the authenticity and reliability of the data. This hierarchical distributed sensing architecture enables the synchronous acquisition of multi-dimensional data. The parameter settings of different types of sensors are highly targeted, ensuring both the ability to capture subtle changes in cable conditions and the synchronous acquisition of environmental influencing factors. The real-time upload and timestamp encryption mechanism of 5G edge nodes ensure the timeliness and credibility of the data, laying a high-quality data foundation for subsequent evaluation.

[0060] S2: A spatiotemporal dual-dimensional adaptive calibration algorithm is proposed. Time calibration uses the optical frequency clock of the fiber Bragg grating sensor as a reference, with a synchronization error of less than 1-3 μs. A sliding window cross-correlation correction method is used for vibration sensor data to calculate the time offset Δt between 5-15 adjacent data sets. A dynamic compensation curve is generated through quadratic polynomial fitting, improving the timestamp correction accuracy of vibration data to ±3-8 μs, effectively solving the time synchronization problem between different sensors. Spatial calibration establishes a three-dimensional coordinate model of the cable. A spatial interpolation mapping method is used for sensor data at different locations. Based on the Delaunay triangulation algorithm, discrete monitoring point data are mapped to a 0.5-2m × 0.5m coordinate system. The -2m grid cells are smoothed using a Gaussian kernel function to eliminate spatial heterogeneity, making monitoring data from different spatial locations comparable. The calibrated data must pass a spatiotemporal consistency test, and the deviation rate must be less than 0.3-0.8% before proceeding to the next stage. This ensures that the spatiotemporal consistency of the data meets the requirements of subsequent processing. Using this dual-dimensional adaptive calibration algorithm, the time calibration error is controlled to an extremely low range, solving the problem of time asynchrony of multi-source data. Spatial calibration, through scientific grid mapping and smoothing, eliminates data heterogeneity caused by differences in spatial location. The strict consistency test further guarantees data quality, making subsequent multi-source data fusion analysis more reliable.

[0061] S2 specifically includes a dual-reference time anchoring method for vibration and displacement data. The optical frequency clock of the fiber optic grating sensor serves as the primary reference, while the timing signal from the 5G base station is introduced as the secondary reference. A sliding window cross-correlation algorithm is used to calculate the time offset Δt between adjacent data points. A dynamic correction curve is generated using a temperature drift compensation model, and the vibration data timestamps are calibrated in real time. The calibrated data undergoes time consistency verification to generate time-synchronized data with confidence markers. Based on a pre-defined three-dimensional cable coordinate model, a layered mesh mapping algorithm is used. A uniform mesh is applied to straight cable sections, while a denser mesh is used for complex structures such as bends and joints. This is achieved through Delaunay three-dimensional... Angle partitioning is used to construct spatial topological relationships. When converting the discrete monitoring point coordinates of time-synchronized data into grid node coordinates, a spatial attenuation factor is introduced to weight and correct the data in the edge region, eliminating spatial heterogeneity errors. For environmental data and spatiotemporally calibrated data, a dynamic time bending-sliding window fusion algorithm is adopted. First, the optimal time matching path is calculated using the DTW algorithm. Then, the environmental data is interpolated or downsampled based on the sampling frequency of the vibration data. Trend consistency verification is introduced during the alignment process. When the deviation of the trend exceeds the threshold, a local realignment mechanism is triggered. Finally, multi-source synchronized data containing spatiotemporal coordinates, data values, and quality scores are generated, providing a reliable basis for subsequent processing.

[0062] Specifically, nanosecond-level synchronization timestamps are applied to vibration and displacement sensors using a GPS / BeiDou high-precision clock module. Based on a three-dimensional cable spatial coordinate model, the measurement vectors of heterogeneous sensors are converted into a unified spatial reference system to eliminate the interference of spatiotemporal misalignment of data on feature analysis. Wavelet packet decomposition combined with environmental parameter regression compensation strategy is used to perform frequency band threshold denoising on vibration signals. For fiber optic grating displacement data, a wind speed-displacement response model is introduced to calculate the environmental disturbance compensation amount, ensuring that the mechanical body response is preserved during feature extraction. Through nonparametric correlation analysis of vibration energy spectrum entropy and displacement time-domain statistics, a joint vibration-displacement time-varying feature matrix is ​​established to quantitatively characterize the cable's mechanical coupling state. A sliding discrimination window is set to divide continuous monitoring data into two failure modes: local defects and structural loosening. The threshold baseline is calculated in real time based on the weight allocation of environmental parameters, and a threshold correction function is established by combining the equipment's historical performance curves and maintenance record feedback, so that the alarm threshold can be dynamically adjusted with the cable aging process and environmental stress. A sliding window agglomerative hierarchical clustering (HAC) method is used to classify dynamic features in real time. Bayesian decision theory and physical safety margin model are combined to determine the state level, thus solving the problem of misjudgment under complex working conditions by a single criterion.

[0063] By employing spatiotemporal calibration and multi-source data fusion technologies, the spatiotemporal reference deviation of heterogeneous sensors is eliminated, enabling precise synchronization and collaborative analysis of multidimensional data. Combined with frequency band noise suppression and environmental coupling compensation algorithms, the influence of external interference on monitoring signals is effectively eliminated, significantly improving the authenticity of mechanical feature extraction. Utilizing a vibration displacement dynamic correlation matrix construction and pattern division mechanism, the coupled evolution law of cable operating conditions is deeply explored, enabling differentiated identification of early loosening and local defects. Adopting an adaptive threshold dynamic generation and dual-judgment model overcomes the limitations of traditional fixed thresholds, enhancing the robustness of state assessment in complex environments. Simultaneously, the system parameters are endowed with dynamic optimization and incremental expansion capabilities, ensuring that the assessment model continuously adapts to the evolution of operating conditions throughout the cable's entire life cycle, forming a comprehensive assessment system with both forward-looking early warning and precise decision-making capabilities.

[0064] S2 also includes noise suppression processing of multi-source synchronous data to generate denoised multi-source feature data. Specifically, it involves using an adaptive frequency band optimization method for vibration and displacement data, combining a spectral feature library of cable operation history data, and performing cluster analysis on the signal spectrum after Fast Fourier Transform using an improved spectral clustering algorithm to automatically identify key frequency bands and noise-dominant frequency bands related to cable faults, achieving accurate division of 3-6 dynamic frequency bands. Differentiated denoising strategies are adopted for the characteristics of each frequency band. For key frequency bands, an improved variational mode decomposition algorithm is used to introduce a kurtosis criterion to optimize decomposition parameters and separate effective signals from noise. For noise-dominant frequency bands, wavelet packet thresholding is used for denoising, with the threshold dynamically adjusted according to the frequency band energy. The reconstructed data is then used to generate... The signal is decomposed into frequency bands; a dynamic compensation model for environmental interference is constructed, and a three-dimensional interference coefficient matrix is ​​established based on temperature, humidity, and wind speed data. The temporal correlation between environmental parameters and noise energy is learned through a long short-term memory network, and the environmental interference compensation amount corresponding to each frequency band is output in real time. Piecewise linear compensation is used for temperature interference, a hysteresis correction factor is introduced for humidity interference, and a turbulence intensity correction model is used for wind speed interference. The frequency band decomposed signal is subjected to two-stage interference subtraction. First, the environmental interference compensation amount is subtracted, and then residual impulse noise is removed by verifying the change rate of adjacent sampling points < threshold through signal smoothness verification. Finally, multi-source feature data after denoising is generated, and noise suppression efficiency indicators for each frequency band are attached: effective signal retention rate ≥90%, noise removal rate ≥85%.

[0065] First, the time reference of all sensor devices is unified through GPS or NTP server to provide vibration data X. v and displacement data X d Each sampling point is appended with a UTC timestamp t i This forms a time-synchronized dataset. The timestamp t is included. i The accuracy needs to reach the millisecond level to ensure the temporal continuity of data from multiple sensors on the same physical event. Next, based on the cable's three-dimensional spatial coordinate model, the installation position of each sensor is recorded as an offset vector (dx, dy, dz) corresponding to the coordinate origin O, where dx is the cable's main axial deviation correction, dy is the horizontal radial deviation correction, and dz is the vertical deviation correction. The vibration direction θ collected by the vibration sensor... v The displacement vector d of the fiber Bragg grating micro-displacement sensor is transformed using the coordinate transformation formula r = R(θ). v The transformation is performed using (x+dx,y+dy,z+dz), where R represents the rotation matrix operation, and x, y, and z are the original sensor output values. After spatial calibration, the temperature, humidity, and wind speed data in the environmental data are interpolated using cubic spline interpolation at a sampling frequency of once per second to align with the temporal resolution of the vibration displacement data, ultimately generating a multi-source synchronized dataset with a unified spatiotemporal reference.

[0066] By eliminating the interference of spatiotemporal reference differences on multi-source data, a unified analysis framework is constructed. The technical effect is reflected in improved accuracy of heterogeneous data fusion, making the connection between environmental factors and mechanical motion characteristics closer, thereby accurately distinguishing the dynamic behavior differences between normal cable deformation and abnormal loosening, and improving the credibility of condition assessment results;

[0067] S3: A composite denoising strategy is designed for different types of data characteristics. For vibration data, an improved variational mode decomposition-wavelet thresholding joint algorithm is used. First, the optimal decomposition level K and penalty factor α of VMD are determined by particle swarm optimization algorithm, separating 5-10 modal components. For noisy components, an improved soft-hard threshold tradeoff function is used, with a threshold λ of 0.5-0.8×σ√(2lnN), where σ is the noise standard deviation. After processing, the signal-to-noise ratio is improved by ≥12-18dB, effectively preserving useful information in the vibration signal. For displacement data, a Kalman filter-empirical mode decomposition fusion algorithm is used. First, high-frequency random noise is removed by Kalman filtering, and then EEMD decomposition is performed on the residual signal to remove noisy components. The inherent mode function (EMF) reduces the reconstruction error to less than 0.01-0.03 mm, improving the accuracy of displacement data. For environmental data, a joint algorithm of moving average and outlier detection is used. For data exceeding the 2.5-3.5σ range, a local weighted regression algorithm is used for repair, improving data smoothness by 30-50% and ensuring the stability and reliability of environmental data. This composite denoising strategy precisely addresses the noise characteristics of different data. Vibration data denoising significantly improves the signal-to-noise ratio while retaining key features. Displacement data denoising controls the error to a small range. Environmental data denoising enhances data stability and effectively reduces the interference of noise on subsequent feature extraction and state assessment, thereby improving the practical value of the data.

[0068] S4: Construct a multi-domain feature set. For time-domain features, extract 10-15 statistical parameters from the vibration signal, such as peak factor, kurtosis, and impulse index. These parameters reflect the characteristics of the vibration signal in the time domain. Simultaneously, extract 6-10 indicators from the displacement data, such as mean, variance, and maximum deviation, to reflect the time-domain characteristics of the displacement data. For frequency-domain features, obtain 5-8 features from the vibration signal, such as the dominant frequency distribution and spectral entropy, through Fast Fourier Transform. Use wavelet packet decomposition to extract the energy proportion of 4-6 frequency bands, mining signal features from a frequency perspective. For time-frequency domain features, generate the time-frequency matrix of the vibration signal using Short-Time Fourier Transform, and extract the singular value entropy and energy moment of the matrix. The feature set comprises 3-5 features that comprehensively reflect the signal’s changes in time and frequency. The Relief-F algorithm is used to reduce the feature dimensionality, retaining the top 25-35 core features with a contribution rate of more than 85-95%. While reducing the data dimensionality, it retains the features that have an important impact on cable condition assessment. This multi-domain feature set construction method extracts features from multiple dimensions, including time domain, frequency domain, and time-frequency domain, which can comprehensively capture the key information of cable operation status and avoid the limitations of single-domain features. Feature dimensionality reduction simplifies the data while ensuring a high contribution rate, reduces the computational pressure of subsequent models, improves assessment efficiency, and ensures the representativeness of the features.

[0069] S5: A dynamic update mechanism for the spatiotemporal correlation matrix is ​​proposed. A cross-correlation matrix R(i,j) between vibration acceleration and displacement is constructed using a time window of 5-15 minutes. Here, i represents the vibration characteristic dimension, and j represents the displacement characteristic dimension. This matrix reflects the correlation between vibration and displacement characteristics. A time decay factor θ is introduced, θ = e^(-t / τ), where τ is a decay constant with a value of 20-40 minutes. Historical matrix data is then iteratively updated with weights to enhance the weight of the latest data, enabling the matrix to reflect the latest changes in cable condition in a timely manner. The first 3-7 principal characteristics are extracted through matrix singular value decomposition. The eigenvalues ​​are used to construct a dynamic coupled feature vector V = [λ1, λ2, λ3, λ4, λ5]. This vector can effectively reflect the dynamic correlation strength between vibration and displacement, providing an important basis for subsequent evaluation of cable operation status. The spatiotemporal correlation matrix dynamic update mechanism is adopted. The setting of the time window can balance the timeliness and stability of the data. The introduction of the time decay factor makes the latest data more weighted and better reflects the current status. The main feature values ​​extracted by singular value decomposition simplifies the data while retaining the core correlation information, so that the coupled feature vector can accurately reflect the dynamic relationship between vibration and displacement, providing strong support for status evaluation.

[0070] Constructing a dynamically coupled feature vector specifically includes the following steps:

[0071] 1) Feature extraction: From the denoised multi-source feature data, an improved time-frequency domain fusion extraction method is used to extract vibration energy spectrum features and displacement time-domain waveform features. For vibration energy spectrum features, a short-time Fourier transform is performed on the vibration signal to obtain the time-frequency matrix, and the energy proportion of each frequency unit is calculated (formula: E_k=∫(t1 to t2)|S(k,t)|2dt / ∫(t1 to t2)∑|S(k,t)|2dt, where S(k,t) is the time-frequency amplitude at frequency k). (The analysis period is from t1 to t2). The top 5 frequency units by energy percentage and their corresponding energy change rates are selected to form the vibration energy spectrum feature vector E = [E1, E1', E2, E2', ..., E5, E5']. For displacement time-domain waveform features, the peak value, mean, variance, kurtosis, and waveform slope change rate of the displacement signal are extracted to form the displacement time-domain waveform feature vector D = [D_peak, D_mean, D_var, D_kurt, D_slope].

[0072] 2) Correlation index calculation: The correlation index within a preset time period is calculated using a sliding window weighted correlation algorithm. The time window length is set to T, and the sliding step size is T / 5. Within each window, the correlation index is calculated through the following steps:

[0073] 3) Calculate the Pearson correlation coefficient matrix R between the vibration energy spectrum eigenvector E and the displacement time-domain waveform eigenvector D, where R(i,j)=Cov(E_i,D_j) / (σ_Ei·σ_Dj), Cov is the covariance, and σ is the standard deviation;

[0074] 4) Introduce feature importance weights ω (obtained by training historical fault data using the random forest algorithm, where ω_i represents the weight of the i-th vibration feature and ω_j represents the weight of the j-th displacement feature);

[0075] 5) Calculate the comprehensive correlation index C within the window = ∑(i = 1 to 10)∑(j = 1 to 5)ω_i·ω_j·|R(i,j)|, to obtain the correlation index sequence {C1,C2,...,Cn} within the preset time period;

[0076] 6) Dynamic coupling feature generation: Based on the correlation index sequence, a multi-scale coupling feature vector is constructed. Wavelet packet decomposition is performed on the correlation index sequence to obtain the decomposition coefficients of three scales. The energy values ​​E_scale1, E_scale2, and E_scale3 of each scale are calculated. Combined with the mean μ_C, standard deviation σ_C, and maximum fluctuation value ΔC_max of the correlation index, a dynamic coupling feature vector V=[μ_C,σ_C,ΔC_max,E_scale1,E_scale2,E_scale3] is generated.

[0077] 7) Pattern segmentation: The adaptive threshold clustering method is used to divide the dynamic coupling features into local defect patterns and structural loosening patterns according to the time window;

[0078] 8) Based on historical data, K-means clustering is used to obtain the central feature vectors V_defect (local defects) and V_loose (structural looseness) of the two types of patterns;

[0079] Calculate the Euclidean distance between the current window's dynamically coupled feature vector V and the center vectors of the two classes: d_defect = ||V - V_defect||, d_loose = ||V - V_loose||;

[0080] 9) Introduce a dynamic discrimination threshold θ = 0.5 × (d_defect_avg + d_loose_avg), where d_defect_avg and d_loose_avg are the historical average distances;

[0081] 10) When d_defect < θ, it is determined to be a local defect mode; when d_loose < θ, it is determined to be a structural loosening mode; when neither of these conditions is met, it is marked as a mixed mode and a secondary check is triggered to increase the window length and recalculate.

[0082] Specifically, the vibration and displacement data are first decomposed using wavelet decomposition. The Daubechies-4 wavelet basis is selected to divide the signal into high-frequency and low-frequency components. The number of decomposition layers is set according to the main frequency range of cable vibration. The high-frequency filtering threshold of the vibration data is determined by the product of the standard deviation of the vibration signal and the preset multiplier. The standard deviation of the vibration signal is calculated by the fluctuation value of the vibration signal within the time window. The preset multiplier is matched and selected from the preset parameter library according to the equipment installation environment. The noise suppression of the low-frequency component is achieved by subtracting the environmental background spectrum, i.e., subtracting the product of the environmental reference spectrum and the environmental interference compensation coefficient from the low-frequency component. The environmental reference spectrum is obtained by substituting the normalized data of temperature, humidity and wind speed measured by environmental sensors into the empirical model. For fiber optic grating displacement data, the wind speed interference is calculated using the linear relationship between wind speed and displacement amplitude. The displacement compensation is equal to the displacement amplitude minus μ times the wind speed, where μ is obtained by fitting calibration experiments at different wind speed levels.

[0083] The environmental interference is converted into a quantifiable compensation value, and spurious signals in different frequency bands are eliminated through hierarchical denoising. The technical effect is that it effectively separates the mechanical vibration of the cable body from the sensor noise generated by external interference under complex environmental conditions, so that the vibration displacement characteristics can accurately reflect the changes in the line structure state, providing high signal-to-noise ratio input data for subsequent construction of dynamic coupling characteristics;

[0084] S6: Design an environmental factor dynamic coupling model. Use partial least squares regression algorithm to calculate the correlation coefficient between dynamic coupling characteristics and various environmental parameters, and establish a coupling coefficient matrix C = [c1, c2, c3], where c1 is the temperature coupling coefficient, c2 is the humidity coupling coefficient, and c3 is the wind speed coupling coefficient to quantify the degree of correlation between dynamic coupling characteristics and various environmental parameters. Introduce environmental impact weight factors ω, and determine the weight proportions of temperature, humidity, and wind speed through the analytic hierarchy process, which are 0.3-0.5, 0.2-0.4, and 0.2-0.4, respectively, to reflect the differences in the impact of different environmental parameters on cable condition. Calculate the comprehensive coupling coefficient K = ω1×c1 + ω2×c2 + ω3×c3 to quantify the overall impact of environmental factors on cable condition, providing a basis for subsequent threshold adjustment. Using this environmental factor dynamic coupling model, partial least squares regression can accurately quantify the correlation between environmental parameters and condition characteristics. The setting of weight factors distinguishes the influence of different environmental parameters, and the comprehensive coupling coefficient realizes the overall quantification of environmental impact, providing a scientific basis for subsequent threshold adjustment and making the assessment results more consistent with the actual cable condition under the environment.

[0085] S7: Construct a dynamic threshold adjustment mechanism. Based on historical normal state data (≥800-1200 sets), use a kernel density estimation algorithm to generate an initial threshold range [Th_min, Th_max]. This range reflects the characteristic range of the cable during normal operation. Combined with the real-time environmental coupling coefficient K, use a piecewise function to dynamically correct the threshold. When K < 0.3, the threshold is lowered by 3-7%; when 0.3 ≤ K ≤ 0.7, the threshold remains at the baseline value; when K > 0.7, the threshold is raised by 5-20%, and the increase is dynamically adjusted according to the K value to adapt the threshold to environmental changes. An aging factor β is used, where β = 1 + 0.01 - 0.03 × years of operation. The threshold is adjusted annually to ensure that the threshold dynamically adapts to the equipment life cycle, which is more in line with the actual situation of different operating stages of the equipment. With this dynamic threshold adjustment mechanism, the initial threshold generated by kernel density estimation fits the historical normal state. Combined with the segmented correction of the environmental coupling coefficient, the threshold can adapt to environmental changes in real time. The introduction of the equipment aging factor takes into account the impact of the equipment life cycle, breaks the limitations of fixed thresholds, and makes the threshold more in line with the actual operating characteristics of the cable in different environments and life cycle stages, thereby improving the accuracy of the assessment.

[0086] Generating an initial threshold interval based on historical normal state data using a kernel density estimation algorithm includes the following steps:

[0087] 1): Calculation of environmental coupling weight coefficients: The improved analytic hierarchy process (AHP)-entropy weight fusion algorithm is used to calculate the environmental coupling weight coefficients of temperature, humidity and wind speed data.

[0088] 2): Construct a judgment matrix. Based on the significance of the influence of temperature, humidity and wind speed on cable condition in historical data, establish a 3×3 judgment matrix A, where A(i,j) represents the ratio of the influence of the i-th environmental parameter on the j-th parameter (1-9 scale method).

[0089] 3): The weight calculation of the analytic hierarchy process (AHP) is carried out by solving the maximum eigenvalue λmax and the corresponding eigenvector of the judgment matrix through the eigenvalue method. After the consistency test (CR<0.1), the subjective weight ω_sub=[ω_T,ω_H,ω_W] (weight of temperature, humidity and wind speed) is obtained.

[0090] 4): Entropy weight method for weight calculation. The environmental parameter data is standardized (x_ij'=(x_ij-minx_j) / (maxx_j-minx_j)), and the information entropy H_j=-k∑x_ij'lnx_ij' (k=1 / lnn, n is the number of samples) is calculated to obtain the objective weight ω_obj=[ω_T',ω_H',ω_W'], where ω_j'=(1-H_j) / ∑(1-H_j);

[0091] 5): Weight fusion, introducing a dynamic balance coefficient α (dynamically adjusted according to environmental stability, 0.3≤α≤0.7), and calculating the comprehensive coupling weight coefficient ω=[αω_T+(1-α)ω_T',αω_H+(1-α)ω_H',αω_W+(1-α)ω_W'], to ensure that the weights simultaneously reflect subjective experience and objective data characteristics;

[0092] 6): The dynamic threshold range is generated by nonlinear superposition, using a hyperbolic tangent function nonlinear superposition model instead of simple addition;

[0093] 7): The basic threshold is set to the upper limit of the 95% confidence interval of historical normal data estimated by Th_base using kernel density estimation;

[0094] 8): Calculate the environmental impact correction factor f(ω,T,H,W)=tanh(ω_T·T_norm+ω_H·H_norm+ω_W·W_norm), where T_norm, H_norm, and W_norm are the standardized values ​​of temperature, humidity, and wind speed (range 0-1). The tanh function ensures that the correction factor changes smoothly between 0 and 1.

[0095] 9): Dynamic threshold range generation, upper limit Th_up=Th_base×(1+k1·f(ω,T,H,W)), lower limit Th_low=Th_base×(1-k2·f(ω,T,H,W)), where k1 and k2 are nonlinear adjustment coefficients (k1=0.2-0.5, k2=0.1-0.3, set according to equipment type). This model realizes the nonlinear mapping of environmental influences and avoids threshold mutations caused by linear superposition.

[0096] 10): The upper and lower limits of the historical fault data correction threshold are adopted using a fault density weighted correction algorithm. Constructing a fault-environment correlation matrix: Historical fault data are categorized by temperature, humidity, and wind speed intervals, and the fault frequency density ρ(i,j,k) (number of faults in the i-th temperature interval, j-th humidity interval, and k-th wind speed interval / total number of samples) is calculated within each interval. Correction coefficients are calculated: Correction coefficients Δ_up=∑ρ(i,j,k)×δ_up(i,j,k) and Δ_low=∑ρ(i,j,k)×δ_low(i,j,k) are calculated for the upper and lower limits of the dynamic threshold, respectively, where δ_up(i,j,k) is the upper limit correction amount corresponding to the fault in that interval (the average difference between the feature value at the time of the fault and the original upper limit), and δ_low(i,j,k) is the lower limit correction amount. Corrected thresholds: The final adaptive evaluation threshold has an upper limit Th_up'=Th_up+Δ_up and a lower limit Th_low'=Th_low-Δ_low, ensuring that the threshold is more stringent in high-fault-risk environmental intervals and more lenient in low-risk intervals.

[0097] Specifically, the vibration energy spectrum is first extracted from the denoised vibration signal using a Fast Fourier Transform (FFT). The integral of the squared amplitude of each frequency point within the selected frequency range f1 to f3 is used as the feature quantity, where f1 is the lower limit of the cable's natural frequency and f3 is the high-frequency cutoff frequency. The displacement time-domain waveform feature is the sum of the root mean square value and the peak-to-valley difference within a preset time period. The displacement time-domain waveform feature is equal to the weighted sum of the root mean square value and the peak-to-valley difference, where the weighting coefficient is determined by the ratio of normal deformation to abnormal displacement in historical data. The vibration energy spectrum and the displacement time-domain waveform feature are input to calculate the correlation index CI. The correlation index CI is equal to the covariance of the vibration energy spectrum and the displacement time-domain waveform feature divided by the product of the standard deviation of the vibration energy spectrum and the standard deviation of the displacement time-domain waveform feature. The covariance is calculated frame by frame using discrete data points within a sliding time window. If the absolute value of CI exceeds 0.7, a strong correlation is determined; when CI is below the threshold of 0.3 for three consecutive windows, a correlation failure alarm is triggered. After completing the time series calculation, the dynamic coupling characteristics at each time point are classified into mode categories according to the distribution of CI peak values: if the high-frequency component accounts for more than 60% of the total energy and the CI decay rate exceeds 5% per hour, it is classified as a local defect mode; if the low frequency dominates and the CI value continues to rise more than 1.5 times the baseline, it is classified as a structural loosening mode.

[0098] For example, after maintenance, a substation cable was monitored and found to have a main peak amplitude of 0.8g at 125Hz in its vibration energy spectrum. 2 / Hz, displacement waveform root mean square value 0.2mm, peak-to-valley difference 1.3mm. The calculated window covariance is 0.045, vibration standard deviation is 0.12g, and displacement standard deviation is 0.15mm, yielding CI equal to 0.045 divided by the product of 0.12 and 0.15, which equals 2.5. Since the CI value exceeds the threshold triggering a strong correlation state, combined with the characteristic that high-frequency energy accounts for 73%, it is determined to be a defect mode caused by partial discharge. By quantifying the dynamic correlation between vibration and displacement, the characteristics of local defects caused by high-frequency discharge can be accurately captured, avoiding the misidentification of instantaneous impact noise as structural failure, and also possessing the ability to differentiate between sporadic and persistent anomalies.

[0099] S8: Employs an improved deep learning fusion model. The model architecture is primarily based on a bidirectional long short-term memory network. The input layer receives 25-35 dimensions of dynamically coupled features, and the hidden layers consist of 2-4 layers with 48-80, 24-40, and 12-24 nodes respectively. An attention mechanism is added to strengthen the weights of key features, enabling the model to focus on features that play a crucial role in cable state assessment. Training optimization employs a hybrid optimization algorithm, combining the Adam optimizer and particle swarm optimization to optimize model parameters. The learning rate is set to 0.0005-0.002, and the batch size is 32-128. Early stopping prevents overfitting and improves the model's generalization ability. The model validation uses 5-fold cross-validation, achieving an accuracy of ≥95-98% and an F1 score of ≥0.93-0.97 on a sample set of 8000-12000 sets, ensuring high evaluation accuracy. This improved deep learning fusion model utilizes a bidirectional long short-term memory network suitable for handling temporal features, an attention mechanism to focus on key features, and a hybrid optimization algorithm to improve model training performance. Early stopping effectively prevents overfitting, while a large sample set and cross-validation ensure high accuracy and generalization ability, making cable condition assessment more reliable.

[0100] S9: Design a five-level evaluation system. Level 1: Dynamic coupling characteristic < 0.2-0.4 × threshold, no abnormal fluctuations, indicating excellent cable operation; Level 2: 0.2-0.4 × threshold ≤ characteristic < 0.5-0.7 × threshold, slight fluctuations, good cable operation; Level 3: 0.5-0.7 × threshold ≤ characteristic < 0.7-0.9 × threshold, requires attention, cable operation is average, close monitoring is needed; Level 4: 0.7-0.9 × threshold ≤ characteristic < 0.9-1.1 × threshold, maintenance recommended, cable operation is poor, maintenance needs to be arranged; etc. Level 5 is characterized by features ≥0.9-1.1×threshold, triggering an emergency warning. The cable's operating condition is critical and requires immediate action. When the model outputs the level, a feature contribution report is generated simultaneously, identifying the main factors leading to the condition deterioration, such as abnormal vibration in a certain frequency band accounting for 30-40%. This provides direction for fault diagnosis and handling. This five-level assessment system clearly divides the different levels of cable operating conditions, with a clear feature range for each level, facilitating rapid judgment of cable condition. The feature contribution report can accurately locate the root cause of the problem, providing clear guidance for operation and maintenance decisions and improving the efficiency and pertinence of fault handling.

[0101] S10: Establish a closed-loop optimization mechanism. Collect the deviation between actual maintenance data and evaluation results monthly, calculate the correction coefficient δ = actual state value / evaluation value, and use it to evaluate the accuracy of the model. Employ an online incremental learning algorithm to fine-tune the model every 300-800 sets of new data. Update the training set through a sliding window, retaining the latest 4-8 months of data to enable the model to adapt to changes in cable operating conditions. Calibrate the threshold generation mechanism quarterly and optimize the coupling coefficient calculation model in conjunction with environmental change trends to maintain the evaluation accuracy at 93-97% or higher. This ensures that the entire evaluation scheme can function stably and effectively in the long term. With this closed-loop optimization mechanism, the regularly collected deviation data provides a basis for model optimization. Online incremental learning allows the model to adapt to changes in cable conditions in a timely manner. Updating the training set through a sliding window ensures data timeliness. Quarterly calibration of the threshold generation mechanism ensures that the evaluation method continuously adapts to environmental changes, enabling the entire evaluation scheme to maintain high accuracy and meet long-term monitoring needs.

[0102] S10 uses a sliding window clustering algorithm to classify dynamic coupling features in real time and generate a preliminary state level; the preliminary state level is then jointly determined with the adaptive evaluation threshold to generate a cable operating state level.

[0103] The method further includes: obtaining the difference between the cable operating status level and the actual fault record within a preset period; adjusting the weight distribution of the dynamic coupling feature in reverse according to the difference value to generate an updated dynamic evaluation model; and expanding the input dimension of the dynamic evaluation model and retraining the classification boundary when a new sensor is connected.

[0104] The determination of the structural loosening mode includes identifying the amplitude of the low-frequency main peak of the vibration energy spectrum and the displacement time-domain increment; when the amplitude of the low-frequency main peak exceeds a preset proportion of the displacement time-domain increment, a structural loosening warning signal is triggered.

[0105] The method further includes: generating a three-dimensional stress distribution simulation map of the cable based on the structural loosening early warning signal; combining the wind speed and direction parameters in the environmental data; calibrating the preset size offset of the loosening physical location; comparing the preset size offset with a preset safety margin threshold; and outputting maintenance priority suggestions.

[0106] Specifically, the real-time measurements of temperature (T), humidity (H), and wind speed (W) from environmental sensors are first obtained. Each parameter is normalized to a value between 0 and 1. The normalized temperature value is equal to the ratio of temperature to 40 degrees Celsius, with a maximum value limited to 1. The normalized humidity value is the humidity percentage directly converted to two decimal places. The normalized wind speed value is the wind speed divided by a preset local maximum wind speed threshold. The environmental coupling weight coefficient is calculated by weighting the three normalized values. The weight coefficient is obtained through the Delphi method combined with historical fault data analysis. For example, when the correlation between cable faults and environmental humidity is highest, the humidity weight coefficient can be set to 0.5. The basic threshold is determined based on the rated parameters of the cable model. For example, the initial basic threshold value for a 500kV cross-linked polyethylene cable is 0.75. When performing nonlinear superposition, a modified exponential function is used to determine the dynamic threshold range D:

[0107]

[0108] In the formula, B is the basic threshold, α is the environmental coupling weight coefficient, and e is the base of the natural logarithm, ensuring that the impact of environmental factors on the threshold exhibits diminishing marginal returns. Finally, the set of extreme values ​​of coupling coefficients at the time of fault occurrence in similar alarm events over the past 30 days is retrieved, and the correction factor K is taken as the ratio of the historical extreme value average to the current dynamic threshold range. When the product of D and K exceeds the theoretical threshold corresponding to the yield strength of the cable material, the upper and lower limits of the dynamic threshold range are limited, ultimately obtaining an adaptive evaluation threshold that conforms to actual working conditions.

[0109] For example, a high-voltage cable is located in a coastal high-temperature and high-humidity environment, with a real-time measured temperature of 38 degrees Celsius, humidity of 95%, and wind speed of 12 m / s. According to specifications, the parameters are normalized to a normalized value of 0.95 for temperature, 0.95 for humidity, and 0.6 for wind speed. If the environmental weighting coefficients for temperature, humidity, and wind speed are 0.3, 0.5, and 0.2 respectively, then the environmental coupling weighting coefficient is 0.3 multiplied by 0.95, plus 0.5 multiplied by 0.95, plus 0.2 multiplied by 0.6, resulting in 0.895. Taking the base threshold B as 0.7, the dynamic threshold range D is equal to the square root of 0.7 multiplied by the natural logarithm e raised to the power of 0.895, hence D is 0.7 multiplied by 1.565, resulting in 1.096. Combining historical data, the correction factor K for humidity-related faults in this area is found to be 0.88, ultimately generating an adaptive evaluation threshold of 0.964. The dynamic adjustment evaluation criteria fully consider the impact of extreme humidity on cable insulation, which can identify abnormal trends in dielectric loss earlier than fixed thresholds, avoid the misjudgment problem of traditional methods under special climatic conditions, and can be adapted to the differentiated operating conditions of cables in different regions by correcting historical data.

[0110] Specifically, when using the sliding window clustering algorithm to classify dynamically coupled features in real time, the time window length Δt is first determined to be 10 seconds and the sliding step size δ. t The interval is 2 seconds to ensure continuous monitoring coverage without omissions. A feature vector F is extracted from the dynamic coupling feature sequence for each window; the feature vector contains the main peak frequency f of the vibrational energy spectrum. p Displacement correlation coefficient CI, vibration energy entropy E v Three key indicators, including f p CI is obtained through power spectrum peak detection, and E is calculated based on the aforementioned calculations. v The formula is derived from the Shannon entropy formula, which describes the energy proportion at each frequency point of the vibration signal. The feature vector is input into a sliding window clustering model for real-time classification, and the cluster centers are updated as follows:

[0111] C k =L·F t +(1-L)·C k-1 ,

[0112] In the formula C k For the updated k-th class center, F t Let L be the feature vector of the current window, and L be the forgetting factor set to 0.2. Experimental calibration is used to ensure the model remains sensitive to sudden changes while avoiding excessive fluctuations. If the Euclidean distance between the new window features and the nearest class center is less than a threshold, it is classified into an existing category; otherwise, a new category is generated. After initial classification, a joint determination is made using an adaptive evaluation threshold. The final level is determined by the product of the initial state level and the deviation from the threshold. If the product exceeds a preset risk coefficient, an upgrade determination is triggered; otherwise, the original classification result is maintained.

[0113] By integrating time series analysis and a dynamic threshold coordination mechanism, a sliding window is used to capture the state evolution pattern, and material performance boundaries are combined to achieve dual verification. The technical effect is that the evaluation results not only reflect the clustering characteristics of real-time monitoring data but also dynamically adapt to changes in the operating environment. Especially when cables undergo gradual degradation, the model can accurately identify the turning point from quantitative to qualitative change through continuous window data accumulation. Compared to a single classification model, this dual-judgment mechanism effectively reduces false alarms caused by sudden interference, while enhancing the detection sensitivity of hidden defects and ensuring the consistency of the physical meaning of the evaluation results under different operating conditions.

[0114] Specifically, the system periodically retrieves all cable operating status level records from the most recent assessment period, including the output results of the preset dynamic assessment model corresponding to each timestamp. Simultaneously, it acquires the fault event times and types recorded in actual maintenance work orders within the same time period, forming a fault tagging sequence. The difference value CY is defined as the ratio of the number of model misjudgments to the total number of assessments, where the number of misjudgments includes two scenarios: the number of times N is missed when the model outputs a normal level but an actual fault exists. miss And the number N of false alarms where the model alarms but there is no actual fault. false When calculating, take CY = (N) miss +N false ) / (N total +1), adding 1 to the denominator is to prevent division by zero errors.

[0115] The weight distribution of the dynamic coupling features is adjusted in reverse based on the difference value to generate an updated dynamic evaluation model;

[0116] Specifically, each dynamic coupling feature (such as the main peak of the vibration energy spectrum, displacement correlation index, etc.) has an initial weight w. i This is stored in the model configuration file. The weight adjustment magnitude Δw is generated based on the difference values. i =λ·CY·(P c -P e ) / P c Where λ is the preset learning rate constant, P c P represents the average contribution of this feature to the correctly evaluated samples. e The standard deviation of the feature contribution in the erroneous samples. The adjusted new weights w′ i =Δw i ·(1+Δw i The weights are then normalized to ensure the total weight of all features is 1. If the adjusted weight of a feature exceeds the reasonable range [0.05, 0.5], a manual review process is triggered.

[0117] When a new sensor is added, the input dimension of the dynamic evaluation model is expanded and the classification boundary is retrained.

[0118] Specifically, when a new sensor (such as a distributed temperature sensor) is added to the system, its data type is first analyzed and a mapping relationship with existing dynamic coupling features is established. The original vibration-displacement correlation matrix is ​​expanded from N×N dimensions to (N+1)×(N+1) dimensions. The expanded matrix adds rows and columns to describe the coupling relationship between the new sensor features and other features. When retraining the classification boundary, an incremental learning algorithm is used. The original model parameters are retained as initial values, and the new sensor data and historical datasets are mixed in a 7:3 ratio for iterative training until the classification accuracy of the new samples is higher than a preset threshold.

[0119] For example, after three months of continuous operation, the dynamic evaluation model of an underground cable output a cumulative rating result 420 times. During the same period, two partial discharge faults (both captured by the model) and five false alarm events were actually recorded. The calculated difference value was 0.0119. Assuming the initial vibration energy weight is 0.3, the average contribution P of this feature in the correct samples is... c =0.85, Standard deviation of error sample contribution P e =0.12. Taking the learning rate λ = 0.1, the adjustment increment is calculated to be 0.00099. The adjusted weights are updated to 0.3 × (1 + 0.00099) = 0.3003, and fine-tuned after normalization. When a current phase sensor is added, the original model expands from 12 vibration-related features to 13. The newly added phase fluctuation feature achieves 95.2% classification accuracy through two weeks of incremental learning. The difference-driven weight adjustment mechanism enables the model to autonomously optimize feature importance allocation, reducing the false alarm rate caused by environmental interference. Simultaneously, it maintains the consistency of state assessment when adding a new sensing dimension, avoiding performance drops due to system upgrades. Through the elastic expansion of dynamically coupled features, it can flexibly adapt to equipment iterations in cable monitoring networks, ensuring that the assessment system continuously tracks emerging fault modes during line aging.

[0120] Based on the structural loosening warning signal, a three-dimensional stress distribution simulation diagram of the cable is generated; combined with the wind speed and direction parameters in the environmental data, a preset offset of the loosening physical location is determined; the preset offset is compared with a preset safety margin threshold, and a maintenance priority suggestion is output.

[0121] Specifically, when a structural loosening warning signal is triggered, the material parameters E and cross-sectional area of ​​the corresponding cable segment are retrieved, where E represents the elastic modulus, obtained from the cable model parameter library. The vibration acceleration time history curve a(t) and displacement time history curve d(t) from the denoised multi-source feature data are used as input loads to construct a three-dimensional cable model containing the physical locations of loosening in the finite element analysis software. The stress field distribution σ(x,y,z) at each node is calculated using the Newton-Raphson iteration method, with the specific equation σ=E·J+ρ·a(t), where J is the strain obtained from the displacement differential d'(t), and ρ is the cable linear density. The wind speed and direction parameters are decomposed into a lateral component W. x and longitudinal component W y This is superimposed on the displacement excitation component, where W x =W·cosθ, W y =W·sinθ, where θ is the real-time angle between the wind direction and the cable axis. In the simulation model, the dynamic calculation step size is set to 0.01 seconds, and the x-axis of the loosened zone node is recorded in real time. i displacement vector D i (t)=[dx i ,dy i ,dz i The maximum root mean square displacement D within the preset time window is taken. max And compare it with the critical threshold D determined by the material properties. c Comparison. Critical threshold D c Determined by cable safety operation specifications. The margin factor η = D is defined. c / (D max ·k s ), where k s The environmental safety factor is determined by the wind speed level. Three priority ranges are set: a margin coefficient η less than or equal to 1.1 indicates emergency repair; a margin coefficient η between 1.1 and 1.5 indicates priority monitoring; and a margin coefficient η greater than or equal to 1.5 indicates routine tracking. The system automatically generates recommended handling measures and links them to the equipment ledger, forming a maintenance work order data package.

[0122] For example, a loosening warning was triggered at the connection point of a 110kV cable tower. Simulation calculations showed a linear density of 2.8 kg / m and an elastic modulus of 12 GPa. The real-time wind speed was 15 m / s, and the wind direction formed a 30° angle with the cable axis. This was decomposed into W... x =12.99m / s, W y = 7.5 m / s. The maximum offset D was measured within an 8-second time window. m Given ax = 8.3 mm, calculate the critical threshold D. c = 91.8mm. Take k s=1.2 (corresponding to level 8 wind force), the calculated margin coefficient η = 91.8 / (8.3×1.2) = 9.2. The system determines that it belongs to the routine tracking level and generates a re-inspection work order for next month. The degree of mechanical condition deterioration of the loosened part is accurately quantified through three-dimensional stress simulation, and the offset assessment after comprehensively considering the dynamic influence of wind load effectively avoids the one-sidedness of simple displacement monitoring. The margin coefficient calculated dynamically based on material properties provides a physical basis for maintenance decisions, which can prevent the waste of resources caused by over-maintenance and avoid missing high-risk defects. It is particularly suitable for scenarios where mountain transmission lines are affected by complex meteorological conditions.

[0123] Nanosecond-level synchronization timestamps are applied to vibration and displacement sensors using a GPS / BeiDou high-precision clock module. Based on a three-dimensional cable spatial coordinate model, the measurement vectors of heterogeneous sensors are converted into a unified spatial reference system to eliminate the interference of data spatiotemporal misalignment on feature analysis. Wavelet packet decomposition combined with environmental parameter regression compensation strategy is used to perform frequency band threshold denoising on the vibration signal. For fiber optic grating displacement data, a wind speed-displacement response model is introduced to calculate the environmental disturbance compensation amount, ensuring that the mechanical body response is preserved during feature extraction. Through nonparametric correlation analysis of vibration energy spectrum entropy and displacement time-domain statistics, a vibration-... The displacement-coupled time-varying feature matrix quantitatively characterizes the mechanical coupling state of the cable. A sliding discrimination window is set to divide the continuous monitoring data into two failure modes: local defects and structural loosening. The threshold baseline is calculated in real time based on the weight allocation of environmental parameters. A threshold correction function is established by combining the historical performance curve of the equipment and the feedback of maintenance records, so that the alarm threshold can be dynamically adjusted with the aging process of the cable and environmental stress. The sliding window agglomerative hierarchical clustering (HAC) is used to classify dynamic features in real time. The state level is determined by combining Bayesian decision theory and physical safety margin model, which solves the problem of misjudgment under complex working conditions by a single criterion.

[0124] By employing spatiotemporal calibration and multi-source data fusion technologies, the spatiotemporal reference deviation of heterogeneous sensors is eliminated, achieving precise synchronization and collaborative analysis of multidimensional data. Combined with frequency band noise suppression and environmental coupling compensation algorithms, the influence of external interference on monitoring signals is effectively eliminated, significantly improving the authenticity of mechanical feature extraction. Utilizing a vibration displacement dynamic correlation matrix construction and pattern division mechanism, the coupled evolution law of cable operating conditions is deeply explored, enabling differentiated identification of early loosening and local defects. Adaptive threshold dynamic generation and a dual-judgment model are adopted to overcome the limitations of traditional fixed thresholds, enhancing the robustness of state assessment in complex environments. Simultaneously, the system parameters are endowed with dynamic optimization and incremental expansion capabilities, ensuring that the assessment model continuously adapts to the evolution of operating conditions throughout the cable's entire life cycle, forming a system that combines forward-looking early warning with precise analysis. Based on the same inventive concept, such as... Fig. 2 As shown, the present invention also provides a multi-source data-driven comprehensive evaluation system for cable operating status, the system comprising:

[0125] The multi-dimensional synchronous acquisition module is used to acquire vibration data output by preset high-frequency micro-vibration sensors, displacement data output by preset fiber optic micro-displacement sensors, and temperature, humidity and wind speed data from preset environmental sensors at key cable nodes.

[0126] The spatiotemporal calibration module is used to perform spatiotemporal calibration on the vibration data, displacement data and environmental data to generate multi-source synchronous data;

[0127] The noise suppression module is used to perform noise suppression processing on the multi-source synchronous data to generate denoised multi-source feature data.

[0128] The dynamic coupling analysis module is used to construct a vibration-displacement correlation matrix based on the denoised multi-source feature data and generate dynamic coupling features.

[0129] An adaptive threshold module is used to generate an adaptive evaluation threshold based on the coupling coefficient between the dynamic coupling characteristics and the environmental data.

[0130] The status assessment module is used to input the dynamic coupling characteristics and adaptive assessment threshold into a preset dynamic assessment model and output the cable operating status level.

[0131] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0132] The operating principle and advantages of this invention are as follows: This technical solution has significant advantages in multi-source data collaborative processing. Through a hierarchical distributed sensor network architecture, it achieves comprehensive synchronous acquisition of vibration, displacement, and environmental parameters. Combined with the real-time upload capability of 5G edge nodes, it solves the problem of data silos in traditional monitoring. The spatiotemporal dual-dimensional adaptive calibration algorithm innovatively introduces dual-reference time anchoring and hierarchical grid mapping, controlling the time synchronization error within ±3-8μs. Spatial heterogeneity errors are significantly eliminated through Gaussian kernel function smoothing, improving the fusion accuracy of multi-source data under a unified spatiotemporal framework by more than 40%. The dynamic time bending-sliding window fusion algorithm achieves precise alignment of environmental data and physical quantity data through the optimal time matching path, taking into account the sampling frequency difference. Combined with the trend consistency verification mechanism, it effectively avoids feature misjudgment caused by data misalignment, providing highly reliable basic data for subsequent analysis.

[0133] In terms of the accuracy of feature extraction and state assessment, the composite strategy adopted by the solution demonstrates outstanding performance. The differentiated denoising algorithm designed for different data characteristics improves the signal-to-noise ratio of vibration signals by ≥12-18dB, controls the displacement data reconstruction error within 0.01-0.03mm, and improves the smoothness of environmental data by 30-50%, thus eliminating noise interference for feature extraction. The multi-domain feature set combined with Relief-F dimensionality reduction technology significantly reduces the data dimension while retaining 85-95% of the contribution rate features, thereby enhancing the model processing efficiency. The construction of dynamically coupled feature vectors, through the spatiotemporal correlation matrix and pattern partitioning mechanism, achieves accurate identification of local defects and structural loosening patterns. Combined with the attention mechanism of the improved deep learning fusion model, the assessment accuracy is stabilized at 95-98%, and the F1 score is ≥0.93-0.97. Compared with the traditional single feature assessment method, the fault identification sensitivity is improved by more than 35%.

[0134] The dynamic adaptability and closed-loop optimization mechanism of the solution ensure the reliability of long-term assessment. The environmental factor dynamic coupling model quantifies the influence weights of temperature, humidity, and wind speed on cable status through partial least squares regression and analytic hierarchy process. Combined with the dynamic threshold adjustment mechanism, the assessment threshold can be corrected in real time with environmental changes and equipment aging, avoiding the misjudgment problem of fixed thresholds under extreme conditions. The closed-loop optimization mechanism uses online incremental learning and sliding window updates to fine-tune the model every 300-800 sets of new data. With quarterly threshold calibration, the assessment accuracy is maintained at over 93-97% in the long term. In addition, the system supports dimensional expansion of new sensors and retraining of classification boundaries, maintaining the ability to identify emerging fault modes throughout the entire cable life cycle, significantly reducing operation and maintenance costs and extending the effective monitoring cycle of equipment.

[0135] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A multi-source data driven cable running state comprehensive evaluation method, characterized in that, Comprise the following steps: S1: Using hierarchical distributed sensor network architecture, deploying three types of core sensors at the key nodes of the cable: high-frequency micro-vibration sensors select piezoelectric sensors, built-in temperature compensation module, synchronous acquisition of axial and radial vibration acceleration data; fiber Bragg grating micro-displacement sensor uses wavelength demodulation technology, through the distributed fiber layout to realize 360° circumferential displacement monitoring, generate displacement vector data; environmental sensor array integrates temperature, humidity, wind speed, air pressure and corrosion gas sensors, builds environmental parameter space-time distribution field, all sensors realize real-time data upload through 5G edge nodes, uses timestamp encryption mechanism to ensure the traceability of original data, S2: Propose a time-space two-dimensional adaptive calibration algorithm, time calibration takes the optical frequency clock of fiber Bragg grator sensor as the reference, uses sliding window cross-correlation correction method for vibration sensor data, calculates the time offset Δt of adjacent data, generates a dynamic compensation curve through quadratic polynomial fitting, improves the timestamp correction accuracy of vibration data; Space calibration establishes a three-dimensional coordinate model of the cable, uses spatial interpolation mapping method for sensor data at different positions, maps discrete monitoring point data to grid elements based on Delaunay triangulation algorithm, eliminates spatial heterogeneity through Gaussian kernel function smoothing processing, and the calibrated data passes the space-time consistency test before entering the next step; S3: Design a composite denoising strategy for different types of data characteristics, vibration data uses an improved variational mode decomposition-wavelet threshold joint algorithm, first determines the optimal decomposition layer number K and penalty factor α of VMD through particle swarm optimization algorithm, separates the modal components, and uses an improved soft and hard threshold function to process the noisy components to improve the signal-to-noise ratio; displacement data uses Kalman filter-empirical mode decomposition fusion algorithm, first removes high-frequency random noise through Kalman filter, then decomposes the residual signal through EEMD, and removes the noisy intrinsic mode function; environmental data uses a sliding average-anomaly detection joint algorithm, repairs abnormal data using local weighted regression algorithm to improve data smoothness; S4: Construct a multi-domain feature set, time-domain features extract peak factor, kurtosis, pulse index and other statistics of vibration signals, and mean, variance, maximum deviation and other indicators of displacement data; frequency domain features obtain main frequency distribution and spectral entropy of vibration signals through fast Fourier transform, and extract energy proportion of frequency band through wavelet packet decomposition; time-frequency domain features generate time-frequency matrix of vibration signals through short-time Fourier transform, extract singular value entropy and energy matrix of the matrix, reduce features through Relief-F algorithm, and retain core features; S5: Propose a dynamic updating mechanism for time-space correlation matrix, set a time window, construct a cross-correlation coefficient matrix R(i,j) of vibration acceleration and displacement, where i is the vibration feature dimension and j is the displacement feature dimension; Introduce time decay factor θ to iteratively update the historical matrix data, enhance the weight proportion of the latest data; extract the main eigenvalue through singular value decomposition of the matrix, construct a dynamic coupling feature vector, and reflect the dynamic correlation strength of vibration and displacement; S6: Design the dynamic coupling model of environmental factors, use the partial least squares regression algorithm to calculate the correlation coefficient of dynamic coupling characteristics and each environmental parameter, and establish the coupling coefficient matrix; introduce the environmental influence weight factor, determine the weight proportion of each environmental factor by the analytic hierarchy process; calculate the comprehensive coupling coefficient to quantify the influence degree of environmental factors on the cable state; S7: Construct a dynamic threshold adjustment mechanism, generate an initial threshold interval based on historical normal state data using kernel density estimation algorithm; Combine the real-time environmental coupling coefficient to dynamically correct the threshold using piecewise function; introduce the device aging coefficient to modify the threshold annually, ensuring that the threshold dynamically adapts to the device life cycle; S8: Use an improved deep learning fusion model, with a bidirectional long short-term memory network as the main body, the input layer receives dynamic coupling characteristics, the hidden layer is set to multiple layers, and the attention mechanism is added to strengthen the key feature weight; The training optimization uses a hybrid optimization algorithm to optimize the model parameters, and the early stopping method is used to prevent overfitting; cross-validation is used for model verification to ensure model accuracy; S9: Design a five-level evaluation system, level 1 is that the dynamic coupling characteristics are within the low threshold range, with no abnormal fluctuations; level 2 is that the characteristics are within the lower threshold range, with slight fluctuations; level 3 is that the characteristics are within the medium threshold range, which needs attention; level 4 is that the characteristics are within the higher threshold range, suggesting maintenance; level 5 is that the characteristics exceed the threshold range, with an emergency warning, the model outputs the level at the same time generates a feature contribution report, clearly identifying the main factors leading to state degradation; S10: Establish a closed-loop optimization mechanism, periodically collect the deviation between actual maintenance data and evaluation results, and calculate the correction coefficient; Use online incremental learning algorithm to fine-tune the model with new data, update the training set through sliding window; periodically calibrate the threshold generation mechanism, optimize the coupling coefficient calculation model combined with environmental change trend, and maintain the evaluation accuracy.

2. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The S2 specifically further includes using a double-reference time anchoring method for vibration data and displacement data, taking the optical frequency clock of the fiber Bragg grating sensor as the first-level reference, and introducing the timing signal of the 5G base station as the second-level reference; calculate the time offset Δt of adjacent data through the sliding window cross-correlation algorithm, generate a dynamic correction curve combined with the temperature drift compensation model, and real-time calibrate the vibration data timestamp; the calibrated data needs to pass the time consistency verification to generate time synchronization data with confidence markers; Based on the preset cable three-dimensional coordinate model, the layered grid mapping algorithm is adopted, the uniform grid is divided for the straight cable, the encryption grid is adopted for the complex structure such as the bending section and the joint, the spatial topological relationship is constructed through the Delaunay triangulation, when the discrete monitoring point coordinates of the time synchronization data are converted into the grid node coordinates, the spatial attenuation factor is introduced to modify the data in the edge area, and the spatial heterogeneity error is eliminated; for the environmental data and the space-time calibrated data, the dynamic time warping-sliding window fusion algorithm is adopted, the optimal time matching path is calculated through the DTW algorithm, then the sampling frequency of the vibration data is taken as the benchmark to interpolate or downsample the environmental data, the trend consistency check is introduced in the alignment process, the local realignment mechanism is triggered when the change trend deviation exceeds the threshold, and finally the multi-source synchronous data containing the space-time coordinates, the data value and the quality score are generated, and the reliability basis is provided for the subsequent processing.

3. The multi-source data driven cable operation state comprehensive evaluation method according to claim 2, characterized in that, The S2 also includes noise suppression processing on the multi-source synchronous data to generate denoised multi-source feature data, specifically including using an adaptive frequency band optimization division method on the vibration data and the displacement data, combining the frequency spectrum feature library of the cable operation history data, performing clustering analysis on the signal frequency spectrum after fast Fourier transform through an improved spectral clustering algorithm, automatically identifying the key frequency band and the noise dominant frequency band related to the cable fault, and realizing accurate division of 3-6 dynamic frequency bands; different denoising strategies are adopted for the characteristics of each frequency band, the improved variational mode decomposition algorithm is introduced to optimize the decomposition parameters to separate the effective signal and noise in the key frequency band, and the wavelet packet threshold denoising threshold is dynamically adjusted according to the frequency band energy in the noise dominant frequency band, and the frequency band decomposition signal is reconstructed; A dynamic compensation model of environmental interference is constructed, a three-dimensional interference coefficient matrix is established based on temperature, humidity and wind speed data, the time sequence association law of environmental parameters and noise energy is learned through a long short-term memory network, and the corresponding environmental interference compensation amount of each frequency band is output in real time, the temperature interference adopts segmented linear compensation, the humidity interference introduces a lag correction factor, and the wind speed interference adopts a turbulence intensity correction model; the double-stage interference deduction is performed on the frequency band decomposition signal, the environmental interference compensation amount is deducted first, then the change rate of adjacent sampling points is checked through signal smoothness, and residual impulse noise is removed when the change rate is less than a threshold, finally the denoised multi-source feature data is generated, and the noise suppression efficiency index of each frequency band is attached, the effective signal retention rate is greater than or equal to 90%, and the noise removal rate is greater than or equal to 85%.

4. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The dynamic coupling feature vector is constructed in the S5, specifically including the following steps: 1) Feature extraction, from the denoised multi-source feature data, the improved time-frequency domain fusion extraction method is used to extract the vibration energy spectrum feature and the displacement time domain waveform feature. For the vibration energy spectrum feature, the time-frequency matrix is obtained by short-time Fourier transform of the vibration signal, the energy proportion of each frequency unit is calculated (formula: E_k = ∫(t1 to t2) |S(k,t)|2dt / ∫(t1 to t2)∑|S(k,t)|2dt, where S(k,t) is the time-frequency amplitude at k frequency, t1 to t2 is the analysis period), and the top 5 frequency units and their corresponding energy change rates are selected to form the vibration energy spectrum feature vector E = [E1, E1', E2, E2',..., E5, E5']; 2) Correlation index calculation, the sliding window weighted correlation algorithm is used to calculate the correlation index in the preset period, the time window length is set to T, and the sliding step is T / 5. In each window, the following steps are used to calculate: 3) Calculate the Pearson correlation coefficient matrix R of the vibration energy spectrum feature vector E and the displacement time domain waveform feature vector D, where R(i,j) = Cov(E_i,D_j) / (σ_Ei·σ_Dj), Cov is the covariance, and σ is the standard deviation; 4) Introduce feature importance weight ω (obtained by training historical fault data with random forest algorithm, ω_i represents the weight of the i-th vibration feature, and ω_j represents the weight of the j-th displacement feature); 5) Calculate the comprehensive correlation index C = ∑(i=1 to 10)∑(j=1 to 5)ω_i·ω_j·|R(i,j)| in the window, and obtain the correlation index sequence {C1, C2,..., Cn} in the preset period; 6) Dynamic coupling feature generation: based on the correlation index sequence, a multi-scale coupling feature vector is constructed, the correlation index sequence is decomposed by wavelet packet, three scale decomposition coefficients are obtained, the energy values E_scale1, E_scale2, E_scale3 of each scale are calculated, and the mean μ_C, standard deviation σ_C and maximum fluctuation ΔC_max of the correlation index are combined to generate the dynamic coupling feature vector V = [μ_C, σ_C, ΔC_max, E_scale1, E_scale2, E_scale3]; 7) Pattern division, the adaptive threshold clustering method is used to divide the dynamic coupling feature into local defect pattern and structure loose pattern according to the time window; 8) Based on historical data, the center feature vectors V_defect (local defect) and V_loose (structure loose) of the two types of patterns are obtained by K-means clustering; Compute the Euclidean distance between the current window dynamic coupling feature vector V and the two class center vectors: d_defect = ||V-V_defect||, d_loose = ||V-V_loose||; 9) Introduce a dynamic discrimination threshold θ = 0.5*(d_defect_avg + d_loose_avg), where d_defect_avg and d_loose_avg are the historical average distances; 10) When d_defect < θ, determine as local defect mode; When d_loose < θ, determine as structure loose mode; When neither is satisfied, mark as mixed mode and trigger secondary verification to increase window length and recalculate.

5. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The step of generating an initial threshold interval based on historical normal state data in S7 includes the following steps: 1) Environment coupling weight coefficient calculation: using an improved analytic hierarchy process-entropy weight fusion algorithm to calculate the environment coupling weight coefficients of temperature, humidity and wind speed data; 2) Judgment matrix construction: according to the significance of temperature, humidity and wind speed on the cable state in historical data, a 3*3 judgment matrix A is established, where A(i,j) represents the influence degree ratio of the i-th environmental parameter on the j-th (1-9 scale method); 3) Analytic hierarchy process weight calculation: the maximum eigenvalue λmax and the corresponding eigenvector of the judgment matrix are solved by eigenvalue method, and the subjective weight ω_sub = [ω_T, ω_H, ω_W] (temperature, humidity and wind speed weight) is obtained after consistency test (CR < 0.1); 4) Entropy weight method weight calculation: standardizing the environmental parameter data (x_ij' = (x_ij-minx_j) / (maxx_j-minx_j)), calculating the information entropy H_j = -k∑x_ij'lnx_ij'(k = 1 / lnn, n is the sample number), and obtaining the objective weight ω_obj = [ω_T', ω_H', ω_W'], where ω_j' = (1-H_j) / ∑(1-H_j); 5) Weight fusion: introducing a dynamic balance coefficient α (dynamically adjusted according to environmental stability, 0.3 ≤ α ≤ 0.7), calculating the comprehensive coupling weight coefficient ω = [αω_T+(1-α)ω_T',αω_H+(1-α)ω_H',αω_W+(1-α)ω_W'], and ensuring that the weight reflects both subjective experience and data objective characteristics; 6) Nonlinear superposition to generate dynamic threshold range: using a hyperbolic tangent function nonlinear superposition model instead of simple addition; 7) Set the basic threshold Th_base to the upper limit of the 95% confidence interval of the historical normal data determined by kernel density estimation; 8) Calculate the environmental influence correction factor f(ω,T,H,W) = tanh(ω_T·T_norm+ω_H·H_norm+ω_W·W_norm), where T_norm, H_norm and W_norm are the standardized values of temperature, humidity and wind speed (0-1 range), and the tanh function ensures that the correction factor changes smoothly between 0 and 1. 9) Dynamic threshold range generation, upper limit Th_up = Th_base x (1 + k1 f (ω, T, H, W)), lower limit Th_low = Th_base x (1 - k2 f (ω, T, H, W)), where k1, k2 are nonlinear adjustment coefficients (k1 = 0.2-0.5, k2 = 0.1-0.3, set according to device type), through the model to realize the nonlinear mapping of environmental influence, avoid the threshold mutation caused by linear superposition; 10) History fault data correction threshold upper and lower limit, using fault density weighted correction algorithm. Build fault-environment correlation matrix: classify the historical fault data according to temperature, humidity and wind speed interval, and count the fault frequency density ρ(i,j,k) (the number of faults in the i temperature interval, j humidity interval, k wind speed interval / total sample number); Calculate correction coefficient: calculate correction coefficient Δ_up = ∑ρ(i,j,k)×δ_up(i,j,k), Δ_low = ∑ρ(i,j,k)×δ_low(i,j,k) for dynamic threshold upper and lower limit, respectively, where δ_up(i,j,k) is the upper limit correction amount corresponding to the fault in this interval (take the average of the difference between the characteristic value at fault and the original upper limit), δ_low(i,j,k) is the lower limit correction amount; Corrected threshold: the upper limit of the final adaptive evaluation threshold Th_up' = Th_up + Δ_up, the lower limit Th_low' = Th_low - Δ_low, ensure that the threshold is more stringent in high fault risk environment interval and more relaxed in low risk interval.

6. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The S10 adopts a sliding window clustering algorithm to classify the dynamic coupling features in real time, generating a preliminary state level; The preliminary state level and the adaptive evaluation threshold are jointly determined to generate a cable operation state level.

7. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The method further comprises: obtaining the difference value of the cable operation state level and the actual fault record in a preset period; According to the difference value, the weight distribution of the dynamic coupling feature is adjusted reversely to generate an updated dynamic evaluation model; When a new sensor is accessed, the input dimension of the dynamic evaluation model is expanded and the classification boundary is retrained.

8. The multi-source data driven cable operation state comprehensive evaluation method according to claim 4, characterized in that, The determination of the structure loosening mode includes identifying the low-frequency main peak amplitude of the vibration energy spectrum and the displacement time domain increment; When the low-frequency main peak amplitude exceeds a preset proportion of the displacement time domain increment, a structure loosening early warning signal is triggered.

9. The multi-source data driven cable operation state comprehensive evaluation method according to claim 1, characterized in that, The method further comprises: generating a cable three-dimensional stress distribution simulation diagram according to the structure loosening early warning signal, combining the wind speed and direction parameters in the environmental data, and calibrating the preset size of the offset of the loosening physical position. Compare the preset size of the offset with the preset safety margin threshold to output a maintenance priority suggestion.

10. A multi-source data driven cable operation state comprehensive evaluation system, adopting the multi-source data driven cable operation state comprehensive evaluation method of any one of claims 1-9, characterized in that, The system comprises: A multi-dimensional synchronous acquisition module for acquiring vibration data output by a preset high-frequency micro-vibration sensor of a cable key node, displacement data output by a preset fiber grating micro-displacement sensor, and temperature, humidity and wind speed data of a preset environmental sensor; A space-time calibration module for space-time calibration of the vibration data, displacement data and environmental data to generate multi-source synchronous data; The noise suppression module is configured to perform noise suppression processing on the multi-source synchronous data to generate multi-source feature data after noise suppression. The dynamic coupling analysis module is configured to construct a vibration-displacement correlation matrix based on the multi-source feature data after noise suppression to generate dynamic coupling features. The adaptive threshold module is configured to generate an adaptive evaluation threshold according to a coupling coefficient of the dynamic coupling features and environmental data. The state evaluation module is configured to input the dynamic coupling features and the adaptive evaluation threshold into a preset dynamic evaluation model to output a cable operation state level.

Images