State evaluation and life prediction method for multi-source data fusion of high-voltage cable

Abstract

The invention discloses a state evaluation and life prediction method for multi-source data fusion of a high-voltage cable. According to the method, a full-process technical scheme of acquisition, preprocessing, fusion, evaluation and prediction is constructed. In a multi-source data acquisition stage, full-dimensional data support is formed. In the data preprocessing stage, abnormal value mixed detection, missing value supplementation of physical rule constraints, data desensitization and feature preliminary screening are integrated. In the multi-source data fusion stage, strategies of feature-level fusion, federal learning and decision-level fusion are adopted. In the state evaluation stage, fuzzy comprehensive evaluation of dynamic weight and Bayesian correction is adopted, the state is accurately judged, and weak links are positioned. And in the life prediction stage, an improved prediction model is constructed, and dynamic correction and digital twinborn verification are combined to realize accurate prediction of the residual life. Accurate judgment of the state of the high-voltage cable and dynamic prediction of the service life can be achieved, and the lean management level of a power system on the high-voltage cable is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of high-voltage cable monitoring and maintenance technology in power systems, specifically to a method for condition assessment and life prediction of high-voltage cables through multi-source data fusion. Background Technology

[0002] In modern power systems, high-voltage cables serve as the core carriers of electrical energy transmission, and their operational status directly determines the security and stability of the power supply. As power grids age, cables are susceptible to insulation degradation, partial discharge, and micro-defect propagation due to long-term load cycles, environmental corrosion, and material aging. Failure to promptly and accurately assess their condition and predict their remaining lifespan can lead to major faults such as insulation breakdown and line outages, causing economic losses and social impact. Current high-voltage cable condition assessment and lifespan prediction technologies still have several shortcomings, with the main problems as follows:

[0003] (1) Incomplete data collection dimensions and weak supporting foundation. Existing technologies mostly focus on the collection of data on cable operating status (such as temperature and current) and macroscopic environment (such as temperature and humidity), ignoring some key information: First, the microscopic state data of cable materials, such as the porosity of insulation materials, crack length, and dielectric constant change rate, which directly reflect the essential characteristics of material aging; Second, digital twin benchmark data, such as the three-dimensional geometric model of the cable, geological data of the laying path, and information on the distribution of surrounding structures. Without such data, it is difficult to establish a mapping relationship between physical entities and digital mirrors, and it is impossible to provide accurate benchmark references for evaluation and prediction, which leads to deviations in the depiction of the actual operating status of the cable.

[0004] (2) Data preprocessing methods are inadequate, resulting in insufficient data quality and security. Existing technologies often employ a single algorithm for outlier detection, which is insufficient to identify hidden atypical anomalies in nonlinear data such as partial discharge and ultrasonic signals. Furthermore, the completion of missing values ​​does not fully incorporate the physical laws governing cable operation. In addition, with the increasing demands for power data privacy protection, traditional preprocessing procedures do not include data desensitization.

[0005] (3) Many fusion technologies are limited to single-level fusion, such as feature-level or decision-level fusion. On the one hand, feature extraction relies on a single model and cannot simultaneously capture the time trend of operating status data, the spatial distribution of environmental data, and the cross-dimensional correlation features of material data. On the other hand, in the scenario of multiple substation data dispersion, traditional fusion needs to collect the original data of each station, without introducing privacy computing technology, and the handling of conflict evidence in decision-level fusion is simple, making it difficult to guarantee the robustness of the decision results.

[0006] In summary, existing technologies are insufficient for accurately determining cable condition and dynamically and safely predicting lifespan, failing to meet the demands of modern power systems for lean management of high-voltage cables. A comprehensive technical solution integrating multi-dimensional data and optimized algorithm architecture is needed. Summary of the Invention

[0007] To address the problems existing in current high-voltage cable condition assessment and life prediction technologies across data acquisition, preprocessing, fusion, assessment, and prediction stages, this invention proposes a multi-source data fusion method for high-voltage cable condition assessment and life prediction, comprising the following steps:

[0008] S1 multi-source data acquisition establishes a data support system covering operational status, environmental impact, historical data and maintenance, material microstructure, and digital twin benchmark data;

[0009] S2 data preprocessing involves cleaning, standardizing, spatiotemporal alignment, data anonymization, and preliminary feature screening of the collected multi-source data;

[0010] S3 employs a multi-source, multi-level strategy involving feature-level fusion, federated learning, and decision-level fusion to mine deep correlation information among multi-source data.

[0011] S4 cable condition assessment, based on an extended assessment index system, uses dynamic combination weighting and Bayesian correction fuzzy comprehensive assessment to determine cable condition and locate weak points.

[0012] The S5 cable life prediction method involves screening key influencing factors, improving the prediction model training, dynamic correction, and digital twin verification, combined with life grading and early warning, to predict the remaining life of the cable.

[0013] Preferably, S1 includes:

[0014] S11 Operation status data acquisition: Distributed acoustic sensing modules and electrical pulse ultrasonic partial discharge positioning units are set up, and fiber optic sensing modules are deployed along the cable body to collect temperature, strain and scattering signals, while simultaneously collecting power grid harmonics and voltage sag and swell data.

[0015] S12 Environmental Impact Data Acquisition: Monitoring stations are set up in outdoor direct-buried cable sections, tunnels, near heat sources, and crossing sections. These stations integrate temperature and humidity, soil moisture content, atmospheric corrosion, and vibration sensors to collect light and rainfall data. Invalid data is filtered out through edge computing.

[0016] S13 Historical and Operational Data Acquisition: Retrieve cable factory parameters, commissioning fault records, maintenance data, material formulas, and production process parameters; Connect with the digital twin platform to collect three-dimensional geometric models, geological data of the laying path, and surrounding structures.

[0017] Microscopic data of S14 material were acquired by periodically testing with an ultrasonic flaw detector and a dielectric response tester to obtain data such as porosity, crack length, and dielectric constant change rate.

[0018] Preferably, S2 includes:

[0019] S21 data cleaning employs a hybrid algorithm of box plot, isolated forest, and wavelet packet transform to detect outliers. Combined with the physical laws of cable operation, continuous missing values ​​are filled in by linear interpolation and LSTM prediction, while discrete missing values ​​are filled in by mode imputation and neighborhood verification.

[0020] S22 data standardization: continuous data is standardized using Z-score, nonlinear data is standardized after Box-Cox transformation, and material micro data is standardized using maximum and minimum. Differential privacy technology is used to desensitize sensitive data.

[0021] S23 spatiotemporal alignment employs time synchronization and edge node clock calibration to control timestamp errors and establish a three-dimensional correlation mapping between monitoring points, cable sections, and digital twin nodes;

[0022] S24 initial feature screening: features with absolute correlation coefficients exceeding the threshold are screened using Pearson correlation coefficient and mutual information entropy. Principal component analysis is used to reduce the dimensionality of redundant data, and principal components with prominent cumulative contribution rates are retained.

[0023] Preferably, S3 includes:

[0024] S31 Feature-level fusion constructs a CNN-LSTM-Transformer hybrid feature extraction model to extract the temporal trend of running data, spatial features of environmental partial discharge data, and cross-dimensional correlation features of micro twin data, respectively. It combines adversarial generative networks to enhance features and then fuses them through attention-weighted fusion.

[0025] S32 Federation Feature Fusion adopts a horizontal federated learning architecture. Each substation trains a local model and uploads parameters. The federated coordination node aggregates parameters using a federated averaging algorithm and protects parameter security through a secret sharing protocol.

[0026] The S33 decision-level fusion adopts the DS evidence theory to classify cable status into four levels: normal, attention, abnormal, and severe abnormal. It optimizes probability allocation and handles conflict evidence according to the conflict coefficient. It evaluates reliability through confidence entropy and outputs decision results.

[0027] Preferably, S4 includes:

[0028] S41 Construct an evaluation index system, which includes insulation, thermal stability, mechanical properties, environmental adaptability, material microstructure, and digital twin matching degree;

[0029] S42 determines the weight of indicators by using the analytic hierarchy process and entropy weight method, combined with dynamic adjustment factors, and adjusts the proportion of subjective and objective weights according to the initial, stable and aging operation stages of the cable.

[0030] S43 constructs a multi-level fuzzy evaluation model, performs fuzzy synthesis based on the trapezoidal fuzzy membership function, combines Bayesian correction to reduce uncertainty, and outputs the state level and a list of weak links.

[0031] Preferably, S5 includes:

[0032] S51 uses correlation analysis, grey relational analysis, and causal inference to identify key factors affecting lifespan;

[0033] S52 constructs an improved Transformer model with a causal attention mechanism, using sine and cosine position encoding and time decay factor optimization, and trains it with the AdamW optimizer and Huber loss function;

[0034] S53 dynamic correction updates data in real time and fine-tunes the model periodically. Emergency correction is triggered when the cable condition changes. Combined with digital twin simulation verification, it outputs the remaining lifespan and confidence interval.

[0035] S54 provides tiered early warnings based on remaining lifetime and displays the contribution of key factors in relation to these factors.

[0036] Preferably, in S21, the outlier detection uses wavelet packet transform to perform multi-level decomposition of the nonlinear data of partial discharge and ultrasonic signals to extract effective signal components; continuous data missing value completion combines the physical laws of cable operation, and optimizes the results through linear interpolation, LSTM prediction and trend constraints; discrete data missing values ​​are ensured to match the actual scenario through mode filling and neighborhood similarity verification.

[0037] Preferably, in the CNN-LSTM-Transformer hybrid feature extraction model in S31, the CNN adopts a ResNet18 architecture with 64 3×3 convolutional kernels, the LSTM network contains 3 hidden layers with 128 neurons per layer, and the Transformer is configured with 8 multi-head attention heads; the feature-level fusion also includes using an adversarial generative network to generate feature samples under simulated fault conditions to expand feature diversity; the attention-weighted fusion mechanism assigns dynamic weights to the spatial features extracted by the CNN, the temporal features extracted by the LSTM, and the cross-dimensional features extracted by the Transformer;

[0038] Preferably, in step S43, the fuzzy comprehensive evaluation describes the correspondence between index data and evaluation level through a trapezoidal fuzzy membership function. After multi-level sequential fuzzy synthesis, the synthesis result is corrected using Bayes' theorem to reduce evaluation uncertainty. The final output results include the scores of each level of index, the status level, and a list of weak links including the degree of exceedance of specific indexes.

[0039] Preferably, the key factors affecting lifespan include insulation performance decay rate, cumulative maximum conductor temperature, total duration of ambient temperature exceeding the standard, average atmospheric corrosion rate, porosity growth rate, number of historical fault repairs, and average deviation of digital twin matching degree.

[0040] The present invention has the following beneficial effects:

[0041] (1) This method increases the acquisition of cable material microscopic data and digital twin benchmark data: material microscopic data is obtained through ultrasonic flaw detection and dielectric response testing to obtain parameters such as porosity and dielectric constant change rate; digital twin benchmark data constructs the mapping relationship between physical entities and digital mirrors, providing accurate benchmark references. It can completely depict the real state of the cable from the microscopic material state to the macroscopic operating scenario, avoiding evaluation bias caused by insufficient data dimensions, and laying a data foundation for subsequent accurate analysis.

[0042] (2) This method uses a hybrid algorithm of box plot, isolated forest and wavelet packet transform to handle outliers; missing value handling is combined with the optimization results of the physical laws of cable operation to ensure that the supplemented data matches the actual scene; the data desensitization process uses differential privacy technology to protect sensitive information; the preliminary feature screening compresses the data dimensionality through correlation and mutual information entropy analysis. The preprocessing scheme significantly improves the data quality while taking into account privacy and processing efficiency.

[0043] (3) This method adopts a multi-level fusion strategy of feature-level fusion, federated learning, and decision-level fusion. Feature-level fusion can simultaneously capture the temporal trend of operational data, the spatial distribution of environmental data, and the cross-dimensional correlation features of material data. The federated learning stage solves the data privacy and ownership disputes in multi-substation scenarios through local model training and parameter aggregation. Decision-level fusion uses dynamic correction of conflict evidence and confidence entropy reliability assessment. The multi-level fusion system fully explores the deep correlation of multi-source data.

[0044] (4) This method expands the evaluation index system to include two categories of indicators: material microstructure and digital twin matching degree. The index weights are assigned using a combination of analytic hierarchy process (AHP), entropy weighting, and dynamic adjustment factors. A Bayesian correction mechanism is introduced into the comprehensive evaluation to reduce evaluation uncertainty. This method improves the predictive ability of long-term time series data by jointly screening key factors through correlation, grey relational analysis, and causal inference, helping to avoid fault risks in advance and extend the effective service life of cables.

[0045] In summary, this invention enables accurate assessment of the condition of high-voltage cables and dynamic, safe prediction of their lifespan, which can effectively improve the level of lean management of high-voltage cables in power systems. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the method steps of an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] This invention provides a method for condition assessment and lifetime prediction of high-voltage cables through multi-source data fusion. Specifically targeting 110kV cross-linked polyethylene high-voltage cables, it integrates the complementary characteristics of multi-source heterogeneous data, the advantages of deep learning feature extraction, and privacy-preserving computation technologies to construct a complete technical solution encompassing data acquisition, preprocessing, fusion, assessment, and prediction. By expanding the multi-dimensional data acquisition range, optimizing the algorithm model architecture, and enhancing dynamic adaptation capabilities, it achieves accurate cable condition determination and dynamic lifetime prediction while simultaneously ensuring data privacy and real-time prediction. Figure 1 As shown, it includes the following steps:

[0050] S1 Multi-source Data Acquisition. Building upon existing data collection on operational status, environmental impact, and historical maintenance, this system adds material microscopic data and digital twin baseline data collection. It employs a distributed monitoring, fixed-site, and digital twin linkage model to establish a comprehensive data support system.

[0051] S11 Operational Status Data Acquisition. The synchronous acquisition network configuration has been optimized, and a distributed acoustic sensing module and a partial discharge ultrasonic positioning unit have been added. Fiber optic sensing modules are deployed with a monitoring point every 30 meters along the cable body, synchronously acquiring conductor temperature, sheath temperature, strain data, and fiber Rayleigh scattering signals. The sampling frequency has been increased to 2Hz, and data accuracy is controlled within ±0.05℃. The partial discharge sensor adopts a dual-mode design of electrical pulse and ultrasound, installed at cable joints, terminations, and critical sections. The electrical pulse sampling frequency is maintained at 10MHz, and the ultrasonic signal sampling frequency is set to 2MHz, synchronously capturing discharge pulse signals, discharge quantity, phase information, and ultrasonic propagation time. The online monitoring terminal has added harmonic content and voltage sag / boost data acquisition functions, shortening the data update cycle to 30 seconds, comprehensively capturing the impact of power grid fluctuations on the cable.

[0052] S12 Environmental Impact Data Acquisition. The monitoring dimensions and deployment range of environmental monitoring stations have been expanded, with stations installed in outdoor direct-buried cable sections, tunnel sections, sections near heat sources, and crossing sections. Each monitoring station integrates temperature and humidity sensors, soil moisture sensors, light intensity sensors, atmospheric corrosion sensors, and vibration sensors. The monitoring range covers ambient temperature -40℃ to 85℃, relative humidity 0% to 100%, soil moisture content 0% to 50%, atmospheric corrosion corresponding to SO2 concentration 0 to 50 ppb, and vibration acceleration 0 to 10 m / s². Simultaneously, data on illumination duration and rainfall intensity are collected. The sampling frequency has been optimized to once every 2 minutes, and invalid environmental data is filtered in real time through edge computing nodes to reduce transmission pressure.

[0053] S13 Historical and Operational Data Acquisition. Upgrade the structured historical database, retrieving original parameters such as factory insulation resistance and withstand voltage test results, commissioning years, annual fault records, and periodic inspection records. Add cable material formulation data, production process parameters (including cross-linking temperature and cooling rate), and original partial discharge waveform data from each maintenance. Simultaneously, connect to the cable digital twin model platform to collect the cable's three-dimensional geometric model, laying path geological exploration data, and surrounding structure distribution data, constructing a mapping relationship between physical entities and digital mirror data to provide a benchmark reference for subsequent assessment and prediction.

[0054] S14 Material Microscopic Data Acquisition.

[0055] A microscopic condition monitoring system for cable materials was established, employing ultrasonic flaw detectors and dielectric response testers to periodically inspect the cable body and joints. The ultrasonic flaw detector emits ultrasonic waves at a frequency of 1MHz to collect data on microscopic defects such as porosity and crack length within the material. The dielectric response tester uses a dielectric spectrum test within a frequency range of 0.1Hz to 1kHz to obtain parameters such as polarization current and dielectric constant change rate of the insulating material. Sampling is set to be performed every 6 months to supplement microscopic characterization data of material aging.

[0056] S2 Data Preprocessing. Building upon traditional cleaning, standardization, and spatiotemporal alignment, this process adds data anonymization and preliminary feature screening. It employs adaptive algorithms, edge computing, and a collaborative solution for privacy protection to improve data quality, security, and subsequent processing efficiency.

[0057] S21 Data Cleaning. The outlier detection and missing value imputation strategy was optimized using a hybrid algorithm of box plots, isolated forests, and wavelet packet transform. First, obvious outliers exceeding 1.5 times the interquartile range were removed using box plots. Then, the isolated forest algorithm was used to identify hidden atypical anomalies. Finally, wavelet packet transform was used to perform a 5-level decomposition of nonlinear data such as partial discharge and ultrasonic signals to extract effective signal components and suppress random noise and interference pulses. For continuous data missing values, linear interpolation, LSTM prediction, and trend constraint methods were used, combined with the physical laws of cable operation to optimize the imputation results. For discrete data missing values, mode imputation and neighborhood similarity verification were used to ensure that the imputed data matched the actual operating scenario.

[0058] S22 Data Standardization. The standardization process was improved, employing differentiated processing schemes for different data types. Continuous data was standardized using Z-score standardization. Nonlinear data such as partial discharge and insulation resistance underwent Box-Cox transformation before standardization, making the data distribution closer to a normal distribution. Material microscopic data was standardized using maximum-minimum standardization, mapping to the 0-1 interval while preserving the relative differences in the data. Data anonymization was performed. For sensitive data involving substation locations and maintenance personnel information, differential privacy techniques were used to add Gaussian noise, with the noise intensity controlled within 5% of the original data variance, balancing data privacy and analytical accuracy.

[0059] S23 Spatiotemporal Alignment. The spatiotemporal synchronization mechanism has been upgraded, employing GPS and BeiDou dual-mode timing and an edge node clock calibration scheme to control the timestamp error of all collected data within 0.5ms. For spatial alignment, based on the GIS coordinates of cable monitoring points and the spatial index of the digital twin model, a three-dimensional correlation mapping is established between monitoring points, cable sections, and digital twin nodes. Environmental data, operational status data, and material microscopic data are precisely matched to the corresponding cable sections according to spatial coordinates. Temporal consistency verification is performed, checking the trend continuity of data from the same monitoring point at different time periods, eliminating contradictory temporal data, and forming a four-dimensional data matrix of time, space, data, and digital twin nodes.

[0060] S24 Feature preliminary screening.

[0061] A preliminary feature screening step is performed to reduce the impact of redundant data on subsequent fusion. A joint algorithm using Pearson correlation coefficient and mutual information entropy is employed to calculate the correlation between each data dimension and the cable status, retaining feature dimensions with an absolute correlation coefficient ≥ 0.4. For highly redundant data, principal component analysis is used for dimensionality reduction, retaining principal components with a cumulative contribution rate ≥ 90%, compressing the four-dimensional data matrix into a low-dimensional effective feature set, thus improving the efficiency of subsequent fusion.

[0062] S3 Multi-Source Data Fusion. It adopts a multi-level fusion strategy of feature-level fusion, federated learning, and decision-level fusion. It introduces a federated learning framework to protect data privacy, optimizes feature extraction and decision fusion algorithms, and fully explores the deep correlation information of multi-source data.

[0063] S31 Feature-level Fusion. A hybrid CNN-LSTM-Transformer feature extraction model is constructed to enhance multi-dimensional feature capture capabilities. Long-term time sequences from operational state data are input into an LSTM network, and temporal trend features are extracted through three hidden layers with 128 neurons per layer. Partial discharge waveform data, ultrasonic signal atlases, and spatial distribution data of environmental data are input into a CNN network, and a ResNet18 architecture with 64 3×3 convolutional kernels is used to extract local spatial features. Material microscopic data and digital twin benchmark data are input into a Transformer network, and cross-dimensional correlation features are extracted through eight multi-head attention heads. A feature enhancement module is implemented, using a generative adversarial network to generate feature samples under simulated fault conditions, expanding feature diversity. Through an attention-weighted fusion mechanism, importance weights for the three types of features are automatically assigned, with the weights dynamically controlled within a range of 0.1 to 0.5, outputting a fused high-dimensional feature vector.

[0064] S32 Federated Feature Fusion. A federated learning stage is set up, suitable for scenarios with dispersed data from multiple substations. A horizontal federated learning architecture is adopted, with each substation acting as a local node. Based on step S31, each substation trains its local feature extraction model, uploading only model parameters, not the raw data. At the federated coordination node, a federated averaging algorithm is used to aggregate the parameters of each local model, iteratively updating the global feature fusion model. A secret sharing protocol is used to divide the data into 5 data shards to protect the security of model parameter transmission and prevent parameter leakage. During federated training, the model convergence condition is set to global loss value fluctuation ≤ 0.0001, and the maximum number of iterations is 300 rounds, ensuring that the global model fuses features from multiple sites without compromising local data privacy.

[0065] S33 Decision-Level Fusion. The DS evidence theory is optimized and improved to enhance the accuracy and robustness of decision fusion. Cable status is divided into four decision levels: normal, attention, abnormal, and severe abnormal. Based on the federated fusion feature vectors, basic probability assignment functions are constructed for four feature categories: operational status, environmental impact, material microstructure, and historical maintenance. Fuzzy membership functions are introduced to optimize probability assignment accuracy. For conflicting evidence, a conflict intensity judgment and dynamic weight adjustment strategy is adopted: when the evidence conflict coefficient is ≥0.7, the conflicting evidence is corrected using a weighted average method; when the conflict coefficient is <0.7, the DS synthesis rule is directly applied. Decision reliability assessment is performed, calculating the confidence entropy for each decision level. If the confidence entropy is ≤0.3, the decision result is output; otherwise, the feature refusion process is triggered, and the final fused decision result and a reliability score from 0 to 10 are output.

[0066] S4 Cable Condition Assessment. The assessment index system and model were optimized, and a dynamic weight adjustment and Bayesian correction mechanism were introduced to achieve accurate assessment combining quantitative and qualitative methods, and to identify weak points in the cable's condition.

[0067] S41 Evaluation Index System Construction. The evaluation index system is expanded by adding two new categories: material microstructure and digital twin matching degree, based on existing indices such as insulation performance, thermal stability, mechanical performance, and environmental adaptability. Insulation performance indices include dielectric response characteristic parameters and ultrasonic positioning accuracy for partial discharge; thermal stability indices include temperature increment due to harmonic losses; mechanical performance indices include cumulative vibration fatigue value; environmental adaptability indices include duration of atmospheric corrosion exceeding standards; material microstructure indices include porosity, crack length growth rate, and dielectric constant change rate; and digital twin matching degree indices include the deviation value and duration of the deviation between actual operating data and digital twin simulation data. Each index is divided into four threshold ranges—excellent, good, qualified, and unqualified—based on cable operating standards and material aging patterns, forming a six-level, twenty-four-item evaluation index system.

[0068] S42 Indicator Weight Determination. An improved combined weighting method is adopted, using the analytic hierarchy process (AHP), entropy weighting, and dynamic adjustment factors to determine indicator weights. Several senior experts with extensive experience in the power industry can be invited to score using the AHP, constructing a judgment matrix and calculating subjective weights; objective weights are calculated based on the information entropy of fused data. Dynamic adjustment factors are implemented, setting different adjustment coefficients according to the cable's operational stage: for the initial operation period of 1-3 years after commissioning, the subjective weight is 0.7 and the objective weight is 0.3; for the stable operation period of 3-15 years, the subjective weight is 0.6 and the objective weight is 0.4; for the aging period of over 15 years, the subjective weight is 0.5 and the objective weight is 0.5. By balancing expert experience and objective data patterns through dynamic adjustment factors, the weight allocation aligns with the characteristics of the cable's entire life cycle.

[0069] S43 Fuzzy Comprehensive Assessment. A three-level fuzzy comprehensive assessment model is constructed to improve assessment accuracy. Six-level indicators are used as primary assessment factors, various subordinate specific indicators as secondary assessment factors, and monitoring data for each specific indicator as tertiary assessment factors. A trapezoidal fuzzy membership function is established based on the indicator threshold range to accurately describe the correspondence between indicator data and evaluation levels. Fuzzy synthesis is performed sequentially at the tertiary, secondary, and primary levels. Weighted average method and Bayesian correction are used during the synthesis process. Bayes' theorem is used to correct the fuzzy synthesis results, reducing assessment uncertainty. The final cable status level is determined according to the maximum membership principle. The scores of each level's indicators, status levels, and a list of weak points are output, clearly defining specific indicators and the degree of exceedance, providing accurate basis for operation and maintenance decisions.

[0070] S5 cable life prediction. Causal inference and digital twin dynamic mapping technologies are introduced to optimize the screening and prediction model of life-influencing factors. A full-process prediction system is constructed, encompassing factor screening, model training, dynamic correction, and twin verification, thereby improving the accuracy and real-time performance of remaining life prediction.

[0071] S51 Lifespan Influencing Factor Screening. A combined screening method of correlation analysis, grey relational analysis, and causal inference was used to accurately identify key influencing factors. First, candidate factors with a correlation degree ≥ 0.6 were screened from the fused data through correlation and grey relational analysis. Then, a Bayesian network was introduced to construct a causal relationship graph between factors to eliminate spurious correlation factors. Finally, the key influencing factors were determined to include insulation performance decay rate, cumulative maximum conductor temperature, total duration of ambient temperature exceeding the standard, mean atmospheric corrosion rate, porosity growth rate, number of historical fault repairs, and mean deviation of digital twin matching degree, totaling 7 core factors, reducing the complexity of the prediction model.

[0072] S52 Prediction Model Training. An improved Transformer and causal attention mechanism prediction model was constructed. Using time-series data of seven selected key factors as input and the actual remaining cable life as output, the model architecture includes a six-layer encoder and a two-layer decoder, equipped with eight multi-head attention heads. A causal attention mechanism was implemented, introducing causal constraints into the Transformer's multi-head attention calculation, considering only the impact of historical time-series data on future lifespan, avoiding reverse causal interference. The position encoding optimization module combines sine-cosine position encoding with a time decay factor to improve the model's ability to capture long-term time-series data. The AdamW optimizer and Huber loss function were used, and the model was trained iteratively for 800 epochs. The model convergence condition was set as validation set loss ≤ 0.0008 to ensure prediction accuracy.

[0073] S53 Dynamic Correction. A dynamic correction mechanism is established, incorporating real-time data feedback, status linkage, and digital twin verification. Newly acquired data is updated hourly and continuously input into the trained prediction model, with model parameters fine-tuned monthly. If the cable condition assessment level changes, an emergency correction process is immediately triggered, introducing a condition correction coefficient to adjust the prediction results. The correction coefficient ranges from 0.8 to 1.2, adjusted according to the magnitude of the condition change. A digital twin verification step is included, inputting the prediction results into the cable's digital twin model to simulate the cable's operating state under the predicted lifespan. The mean square error is used to calculate the deviation between the simulation results and the actual assessment state. If the deviation is ≥15%, the model parameters are re-optimized based on the digital twin simulation data, ultimately outputting the remaining lifespan prediction value and a 95% confidence interval.

[0074] S54 Lifespan Grading Early Warning. A lifespan grading early warning function is set up, classifying warning levels based on the predicted remaining lifespan: Green warning (remaining lifespan ≥ 8 years, no special handling required, routine inspections performed); Yellow warning (5 years ≤ remaining lifespan < 8 years, enhanced monitoring of key indicators); Orange warning (2 years ≤ remaining lifespan < 5 years, development of preventative maintenance plans); Red warning (remaining lifespan < 2 years, immediate replacement or major overhaul arranged). Along with outputting the warning level, the contribution of key influencing factors is displayed, clarifying the core reasons for lifespan degradation.

[0075] The high-voltage cable multi-source data fusion state assessment and lifetime prediction method of this invention improves data support, processing efficiency, fusion accuracy, assessment completeness and prediction practicality. It can accurately assess the state and predict the lifetime of high-voltage cables in power grid operation, reduce fault outage losses and ensure the safe and stable operation of power grid.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for condition assessment and lifetime prediction of high-voltage cables using multi-source data fusion, characterized in that, Includes the following steps: S1 multi-source data acquisition establishes a data support system covering operational status, environmental impact, historical data and maintenance, material microstructure, and digital twin benchmark data; S2 data preprocessing involves cleaning, standardizing, spatiotemporal alignment, data anonymization, and preliminary feature screening of the collected multi-source data; S3 employs a multi-source, multi-level strategy involving feature-level fusion, federated learning, and decision-level fusion to mine deep correlation information among multi-source data. S4 cable condition assessment, based on an extended assessment index system, uses dynamic combination weighting and Bayesian correction fuzzy comprehensive assessment to determine cable condition and locate weak points. The S5 cable life prediction method involves screening key influencing factors, improving the prediction model training, dynamic correction, and digital twin verification, combined with life grading and early warning, to predict the remaining life of the cable.

2. The method according to claim 1, characterized in that, S1 includes: S11 Operation status data acquisition: Distributed acoustic sensing modules and electrical pulse ultrasonic partial discharge positioning units are set up, and fiber optic sensing modules are deployed along the cable body to collect temperature, strain and scattering signals, while simultaneously collecting power grid harmonics and voltage sag and swell data. S12 Environmental Impact Data Acquisition: Monitoring stations are set up in outdoor direct-buried cable sections, tunnels, near heat sources, and crossing sections. These stations integrate temperature and humidity, soil moisture content, atmospheric corrosion, and vibration sensors to collect light and rainfall data. Invalid data is filtered out through edge computing. S13 Historical and Operational Data Acquisition: Retrieve cable factory parameters, commissioning fault records, maintenance data, material formulas, and production process parameters; Connect with the digital twin platform to collect three-dimensional geometric models, geological data of the laying path, and surrounding structures. Microscopic data of S14 material were acquired by periodically testing with an ultrasonic flaw detector and a dielectric response tester to obtain data such as porosity, crack length, and dielectric constant change rate.

3. The method according to claim 1, characterized in that, S2 includes: S21 data cleaning employs a hybrid algorithm of box plot, isolated forest, and wavelet packet transform to detect outliers. Combined with the physical laws of cable operation, continuous missing values ​​are filled in by linear interpolation and LSTM prediction, while discrete missing values ​​are filled in by mode imputation and neighborhood verification. S22 data standardization: continuous data is standardized using Z-score, nonlinear data is standardized after Box-Cox transformation, and material micro data is standardized using maximum and minimum. Differential privacy technology is used to desensitize sensitive data. S23 spatiotemporal alignment employs time synchronization and edge node clock calibration to control timestamp errors and establish a three-dimensional correlation mapping between monitoring points, cable sections, and digital twin nodes; S24 initial feature screening: features with absolute correlation coefficients exceeding the threshold are screened using Pearson correlation coefficient and mutual information entropy. Principal component analysis is used to reduce the dimensionality of redundant data, and principal components with prominent cumulative contribution rates are retained.

4. The method according to claim 1, characterized in that, S3 includes: S31 Feature-level fusion constructs a CNN-LSTM-Transformer hybrid feature extraction model to extract the temporal trend of running data, spatial features of environmental partial discharge data, and cross-dimensional correlation features of micro twin data, respectively. It combines adversarial generative networks to enhance features and then fuses them through attention-weighted fusion. S32 Federation Feature Fusion adopts a horizontal federated learning architecture. Each substation trains a local model and uploads parameters. The federated coordination node aggregates parameters using a federated averaging algorithm and protects parameter security through a secret sharing protocol. The S33 decision-level fusion adopts the DS evidence theory to classify cable status into four levels: normal, attention, abnormal, and severe abnormal. It optimizes probability allocation and handles conflict evidence according to the conflict coefficient. It evaluates reliability through confidence entropy and outputs decision results.

5. The method according to claim 1, characterized in that, S4 includes: S41 Construct an evaluation index system, which includes insulation, thermal stability, mechanical properties, environmental adaptability, material microstructure, and digital twin matching degree; S42 determines the weight of indicators by using the analytic hierarchy process and entropy weight method, combined with dynamic adjustment factors, and adjusts the proportion of subjective and objective weights according to the initial, stable and aging operation stages of the cable. S43 constructs a multi-level fuzzy evaluation model, performs fuzzy synthesis based on the trapezoidal fuzzy membership function, combines Bayesian correction to reduce uncertainty, and outputs the state level and a list of weak links.

6. The method according to claim 1, characterized in that, S5 includes: S51 uses correlation analysis, grey relational analysis, and causal inference to identify key factors affecting lifespan; S52 constructs an improved Transformer model with a causal attention mechanism, using sine and cosine position encoding and time decay factor optimization, and trains it with the AdamW optimizer and Huber loss function; S53 dynamic correction updates data in real time and fine-tunes the model periodically. Emergency correction is triggered when the cable condition changes. Combined with digital twin simulation verification, it outputs the remaining lifespan and confidence interval. S54 provides tiered early warnings based on remaining lifetime and displays the contribution of key factors in relation to these factors.

7. The method according to claim 3, characterized in that, In S21, outlier detection uses wavelet packet transform to perform multi-level decomposition of nonlinear data of partial discharge and ultrasonic signals to extract effective signal components; continuous data missing value completion combines the physical laws of cable operation and optimizes the results through linear interpolation, LSTM prediction and trend constraint. For discrete data, missing values ​​are filled using the mode and neighborhood similarity verification to ensure they match the actual scenario.

8. The method according to claim 4, characterized in that, The CNN-LSTM-Transformer hybrid feature extraction model in S31 uses a ResNet18 architecture with 64 3×3 convolutional kernels, the LSTM network has 3 hidden layers with 128 neurons per layer, and the Transformer has 8 multi-head attention heads. The feature-level fusion also includes using an adversarial generative network to generate feature samples under simulated fault conditions to expand feature diversity. The attention-weighted fusion mechanism assigns dynamic weights to the spatial features extracted by the CNN, the temporal features extracted by the LSTM, and the cross-dimensional features extracted by the Transformer.

9. The method according to claim 5, characterized in that, In S43, the fuzzy comprehensive evaluation describes the correspondence between index data and evaluation level through trapezoidal fuzzy membership function. After multi-level sequential fuzzy synthesis, Bayes' theorem is used to correct the synthesis result to reduce evaluation uncertainty. The final output results include the index scores of each level, the status level, and a list of weak links including the degree of exceedance of specific indicators.

10. The method according to claim 5, characterized in that, The key factors affecting lifespan include insulation performance decay rate, cumulative maximum conductor temperature, total duration of ambient temperature exceeding the standard, average atmospheric corrosion rate, porosity growth rate, number of historical fault repairs, and average deviation of digital twin matching degree.

Images