A cable insulation life prediction method and system based on LSTM accelerated aging mapping

Abstract

The application discloses a cable insulation life prediction method and system based on LSTM accelerated aging mapping, and relates to the technical field of submarine cable insulation life prediction. The existing method has defects such as insufficient modeling of multiple stresses and nonlinearity, separation of laboratory and field data, and difficulty in small sample modeling, and it is difficult to accurately predict the insulation life of submarine cables. The application comprises data collection: collecting insulation electrical performance, physicochemical performance and mechanical performance parameters under accelerated aging conditions; data preprocessing: homodirectional processing of inverse indicators, improved box plot denoising, time-space K nearest neighbor algorithm for filling missing data and normalization; LSTM model construction: embedding the differential equation of dielectric constant as a physical constraint and introducing an exponential weight; model training and evaluation: adopting 5-fold cross-validation, and quantifying the prediction accuracy through mean square error and decision coefficient. The technical scheme improves the prediction accuracy, reduces the risk of failure caused by insulation aging, reduces the maintenance cost, and improves the reliability of the marine energy transmission system.

Description

Technical Field

[0001] The present invention relates to the field of submarine cable analysis technology, and in particular to a cable insulation life prediction method and system based on LSTM accelerated aging mapping. Background Art

[0002] As the core infrastructure for marine energy transmission, high-voltage submarine cables must serve for long periods in a multi-physics coupling environment characterized by high voltage, high current, high humidity, high salinity, and seawater pressure. The reliability of their insulation systems directly determines the safety of grid operation. Failures caused by insulation aging account for as much as 68% (CIGRE report). Therefore, accurate prediction of insulation life is crucial for submarine cable operation and maintenance decisions.

[0003] Currently, submarine cable operation and maintenance primarily relies on regular inspections and maintenance strategies (such as offline dielectric loss testing and partial discharge testing), but real-time life prediction is impossible and maintenance costs are high. Methods used to achieve real-time life prediction include acceleration factor methods and experimental methods, but these methods have the following problems:

[0004] 1. Disadvantages of the traditional acceleration factor method: It relies on models based on linear assumptions, such as the Arrhenius equation and the inverse power law, and cannot characterize nonlinear aging under the synergistic action of multiple stresses. Thermal aging models such as the Arrhenius equation only consider a single temperature stress, while in actual operation, electric field distortion (such as space charge accumulation) and mechanical vibration (wave impact) will accelerate material degradation.

[0005] 2. Laboratory-field data gap: The distribution of accelerated aging data and natural aging data is very different. There are significant differences between laboratory accelerated aging data and real-world environments (IEEE Std 1017 points out that the error of traditional accelerated tests can be as high as 40-60%). The lack of an effective mapping mechanism means that the evaluation results may not accurately reflect the actual aging behavior of the material.

[0006] 3. Too few actual data samples: The Weibull distribution requires a large amount of failure data to support it, but the actual number of failure samples for high-voltage cables is small (0.3 times per 100 kilometers per year); the acquisition cycle for natural aging data is long (10-20 years).

[0007] 4. The challenge of multi-physics coupling: Various stresses often coexist and are coupled to each other. For example, the presence of humidity exacerbates the damage caused by electric fields to insulation. Effects such as electric field distortion, thermal oxidation aging, and mechanical fatigue are intertwined, making it difficult to quantify a single factor. In reality, the various stress effects are not linearly independent, and simple superposition can result in significant errors. Summary of the Invention

[0008] The technical problem to be solved and the technical task to be addressed by this invention are to improve and enhance existing technical solutions by providing a cable insulation life prediction method and system based on LSTM accelerated aging mapping, thereby achieving the goal of real-time and accurate prediction of cable insulation life. To this end, this invention adopts the following technical solutions.

[0009] A cable insulation life prediction method based on LSTM accelerated aging mapping includes the following steps:

[0010] 1) Data collection: Under the set accelerated aging experimental conditions, key performance parameters at different aging time points are collected, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters;

[0011] 2) Data preprocessing: Inverse indicators are oriented in the same direction; outlier detection is performed using an improved boxplot method, and the IQR coefficient is dynamically adjusted to identify and remove outliers; missing data are filled using the spatiotemporal K-nearest neighbor algorithm; and data are compressed to the [0, 1] interval using min-max normalization.

[0012] 3) LSTM model construction: The input layer receives preprocessed time series data; the dielectric constant differential equation is embedded in the loss function as a soft constraint; the loss function introduces exponential weights to make the model more focused on the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable;

[0013] 4) Model training and optimization: Split the data into training and test sets, use 5-fold cross-validation to optimize hyperparameters, monitor the mean squared error of the validation set to prevent overfitting, and obtain a trained LSTM model.

[0014] 5) Life prediction and evaluation: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and determination coefficient.

[0015] This technical solution simultaneously collects three types of parameters: electrical, physical, chemical, and mechanical. It quantifies the synergistic degradation effects of electric field, temperature, and mechanical stress, overcoming the one-sidedness of traditional single-parameter approaches and providing comprehensive input for nonlinear modeling. This solution employs an adaptive data preprocessing mechanism to align inverse indicators, thereby unifying the direction of aging trends. An improved boxplot method dynamically adjusts the IQR to accommodate non-Gaussian distributed data, exploits spatiotemporal correlations to correct missing values, and normalizes data to eliminate dimensional differences. The IQR (Interquartile Range) refers to the interquartile range. Anomaly detection and missing value filling address noise interference in small sample data (such as experimental error and local data loss), improving data quality and making subsequent LSTM modeling more stable and reliable. The LSTM model is enhanced with physical constraints. Physical equation constraints enhance model interpretability, avoid purely data-driven overfitting, and address the challenge of multi-stress nonlinear coupling. Dynamic weighting enables the model to focus on critical aging stages, sensitively capturing lifespan mutations, and improving prediction timeliness. 5-fold cross-validation optimizes hyperparameters to fully utilize limited data; monitors the validation set MSE and performs early stopping (termination if there is no continuous improvement) to suppress overfitting; addresses the lack of generalization ability caused by differences in laboratory and field data distribution, ensuring that the model maintains high accuracy under small sample conditions.

[0016] As a preferred technical means: in step 1), the insulation electrical performance parameters include dielectric constant; the physical and chemical performance parameters include carbonyl index, crystallinity, thermal decomposition temperature and melting temperature; the mechanical performance parameters include elastic modulus; the dielectric constant is collected by a broadband dielectric spectrometer; the crystallinity, thermal decomposition temperature and melting temperature are collected by a synchronous thermal analyzer; the carbonyl index is collected by scanning with an attenuated total reflection infrared spectrometer.

[0017] A broadband dielectric spectrometer collects dielectric responses at different frequencies, directly reflecting the internal polarization behavior of insulating materials. It sensitively captures microstructural changes caused by early aging and provides frequency-domain characteristics for dielectric loss modeling. Elastic modulus quantifies a material's ability to resist elastic deformation. Combined with tensile strength, it forms a dual-dimensional "stiffness-strength" assessment system, accurately characterizing the degree of molecular chain damage under mechanical stress (such as wave impact) and directly correlating this with the structural reliability of submarine cable installation and operation. A simultaneous thermal analyzer measures crystallinity, melting temperature, and thermal decomposition temperature in a single test, establishing a physical correlation between "crystallization state change → thermal stability degradation." This avoids sample differences and experimental errors caused by multiple transfers during separate testing, improving the accuracy and efficiency of aging mechanism analysis. Attenuated total reflection infrared spectroscopy directly reflects the progress of thermal oxidative aging, correlates it with insulation degradation rates, and enables non-destructive testing of chemical aging without damaging the sample section. It can also continuously track oxidation levels at multiple time points within the same sample.

[0018] As a preferred technical means: the inverse indicator adopts the following formula to be unidirectional:

[0019] ; wherein, is the original inverse index value, is the positive index value after homogenization;

[0020] The abnormal value determination condition of the improved box plot method is: if the data point satisfies X < Q1-1.5*IQR or X > Q3+1.5*IQR, the data is considered to be an abnormal value, wherein ; Q1 is the first quartile, and Q3 is the third quartile;

[0021] The formula for filling missing data based on the space-time K-nearest neighbor algorithm is:

[0022] ;

[0023] In the formula, z i is the spatial position coordinate of the i-th neighbor, z t is the spatial position coordinate of the target point, σ is a spatial attenuation coefficient, and k is the number of neighbors, is the observation value of the t-i-th point in the time dimension;

[0024] The normalization formula is:

[0025]

[0026] wherein X min and X max are the minimum and maximum values in the data set, respectively.

[0027] Preprocessing provides high consistency, low noise, and complete and regular input data for subsequent prediction models, and is the cornerstone of ensuring the accuracy of life prediction. Converting the inverse index to a positive index eliminates the trend conflict of multiple parameters; improving the convergence efficiency of the model: avoiding training shock caused by the chaotic direction of the input signal of LSTM, and accelerating the convergence; retaining the physical meaning, the converted is still monotonically negatively correlated with the degree of material degradation, without distorting the original aging law. The traditional box plot assumes normal distribution, while aging data is mostly skewed (such as Weibull distribution), and the dynamic IQR coefficient can flexibly adjust the threshold boundary; it is beneficial to improve the noise removal rate; at the same time, it avoids mistakenly deleting mutation data in the later aging period. The space-time K-nearest neighbor missing value filling method fuses the correlation of space-time two dimensions, avoids data missing caused by local sensor failure, and fuses the spatial position and time continuity. Normalization eliminates the difference in dimensions, avoids the model biasing towards large numerical parameters caused by the difference in orders of magnitude; and accelerates the training of LSTM and stabilizes the gradient update.

[0028] As a preferred technical means: in step 3), the architecture of the LSTM network model adopts a bidirectional LSTM layer, the bidirectional structure is jointly modeled through forward and backward propagation; and the differential equation of the dielectric constant is embedded into the loss function as a soft constraint:

[0029]

[0030] Where γ is the dielectric constant; is the activation energy; R is the ideal gas constant; T is the absolute temperature; is the electric field intensity; n is the material constant; is the correction term for LSTM output.

[0031] The bidirectional LSTM layer combines forward and back-propagation modeling to learn historical aging patterns along the chronological order and capture future degradation signs in reverse chronological order. This helps capture aging inflection points and improves the sensitivity of identifying sudden changes in insulation life. Joint modeling reduces the underreporting rate of key nodes and reduces the modeling error associated with the slow and rapid changes in the dielectric constant. The physical equation for dielectric constant variation is embedded as a regularization term in the loss function to fit the nonlinear effects of electric field distortion (space charge accumulation), coupling temperature and electrochemical degradation, and learning the synergistic effects of vibration, temperature, and electric field. The physical term dominates the macroscopic variation trend of the dielectric constant, while the correction term focuses on compensating for local nonlinear perturbations, thereby reducing the model's generalization error under small sample sizes. By integrating physical constraints with data-driven approaches, the model retains the advantages of LSTM in handling complex nonlinearities while embedding the interpretable physical mechanism of dielectric constant variation, fundamentally solving the industry challenge of multi-stress coupling modeling.

[0032] As a preferred technical approach: the exponential weight introduced in the loss function is: ;

[0033] Where α is the weight adjustment coefficient, t is the current time point, and T is the total aging time.

[0034] Sudden changes in performance parameters (such as a sudden drop in crystallinity and a surge in dielectric constant) in the later stages of aging (e.g., nearing end-of-life) are crucial for operational and maintenance decision-making. Exponential weighting increases exponentially with time t, automatically assigning higher weight to later data during model training, prioritizing the nonlinear degradation characteristics of late aging. This prevents early data from dominating training and ignoring critical mutation signals. Through exponential weighting, the model can optimize its predictive capabilities for high-risk phases, reducing the risk of sudden failures caused by delayed predictions. Through a nonlinear weighting mechanism, exponential weighting forces the model to focus on steep declines (such as exponential decay or step changes) in the time-performance curve, enabling a more accurate fit to accelerated aging patterns under multiple stress couplings (e.g., sudden performance drops caused by the synergistic effects of electric field and temperature). Exponential weighting guides the model to exploit these high-order features, improving its ability to model non-stationary time series data, such as performance fluctuations under time-varying stresses. The exponential weight mechanism uses a "time-importance" exponential mapping to automatically focus on key information without relying on additional labeled data. It solves the inherent defect of the traditional model of "treating all data equally" in a low-cost computing manner, making the deep learning model more in line with the physical law of submarine cable insulation aging, which is "slow in the early stage and sudden in the later stage."

[0035] As a preferred technical measure: When monitoring the validation set mean square error in step 4), terminate training if there is no improvement over multiple epochs.

[0036] When the model consistently reduces loss on the training set but the MSE on the validation set no longer improves, it indicates that the model is beginning to learn the noise in the training data or the random fluctuations of specific samples (i.e., overfitting). The early stopping mechanism monitors the performance of the validation set and forces training to terminate before overfitting occurs, ensuring that the model retains its predictive power for unseen data (such as real-time on-site monitoring data). Early stopping, by terminating training after the MSE on the validation set stabilizes, can better adapt to the differences in data distribution across multiple scenarios. Combined with the "5-fold cross-validation" strategy, the early stopping method independently optimizes the training termination point in each cross-validation fold, avoiding hyperparameter evaluation bias caused by a fixed number of epochs. For example, datasets in different folds may converge at different rates due to differences in aging samples. The early stopping method dynamically determines the optimal training duration for each set of data, improving the accuracy of hyperparameter optimization.

[0037] As a preferred technical means: in step 5), the mean square error formula is:

[0038]

[0039] The coefficient of determination formula is:

[0040]

[0041] Where N is the number of test samples, is the actual lifespan value, is the predicted lifespan value, is the mean of the actual life span values.

[0042] The mean square error formula and the coefficient of determination formula provide quantifiable risk indicators. Using industry-standard mathematical tools, they transform "life prediction accuracy" into a calculable, comparable, and traceable quantitative metric. This not only meets the technical verification requirements of deep learning models but also aligns with the practical requirements of submarine cable operations and maintenance for "result reliability and scientific decision-making." It serves as a critical bridge between algorithm development and engineering application. These two indicators jointly identify model flaws, drive algorithm iteration, and provide a quantifiable, diagnosable, and iterative model verification tool, providing a closed-loop optimization basis for high-reliability cable life prediction systems.

[0043] Another technical solution of the present invention is to provide a cable insulation life prediction system using a cable insulation life prediction method based on LSTM accelerated aging mapping, the cable insulation life prediction system comprising:

[0044] The data acquisition module is used to collect key performance parameters at different aging time points under the set accelerated aging experimental conditions, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters; insulation electrical performance parameters include dielectric constant; physical and chemical performance parameters include crystallinity, melting temperature, carbonyl index, and thermal decomposition temperature; mechanical performance parameters include elastic modulus;

[0045] Data preprocessing module, including:

[0046] The isotropic unit is used to perform isotropic processing on inverse indices including carbonyl index and dielectric constant;

[0047] Anomaly detection unit, used to detect outliers using the improved box plot method, dynamically adjust the IQR coefficient, and identify and remove outliers;

[0048] Missing value filling unit, used to fill missing data based on the spatiotemporal K-nearest neighbor algorithm;

[0049] Normalization unit, used to compress data to the interval [0, 1];

[0050] LSTM prediction module, including:

[0051] Bidirectional LSTM layer, which receives preprocessed time series data;

[0052] Physical constraint layer, embedding the dielectric constant differential equation in the loss function;

[0053] The loss function optimization unit introduces exponential weights to make the model pay more attention to the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable;

[0054] The model training and optimization module divides the data into training and test sets, uses 5-fold cross-validation to optimize hyperparameters, and monitors the mean square error of the validation set to prevent overfitting.

[0055] Life prediction and evaluation module: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and determination coefficient.

[0056] From data acquisition and preprocessing to model construction and evaluation, this method systematically addresses the shortcomings of traditional cable life prediction methods, particularly in multi-physics modeling, small sample size processing, and engineering practicality. It offers significant advantages, providing reliable technical support for the accurate assessment of high-voltage cable insulation condition. Specifically, it simultaneously collects electrical, physical, chemical, and mechanical properties to comprehensively capture the multi-physics coupling effects of the insulation aging process, providing richer characteristic information for life prediction. Accelerated aging experimental conditions, such as high temperature and high pressure, can be set to accelerate data acquisition, shortening the data acquisition cycle and addressing the time-consuming acquisition of natural aging data while maintaining physical relevance to the actual aging process. Mathematical transformations are performed on inverse indicators to unify their direction of change, avoiding model training bias caused by differences in their physical meaning. An improved boxplot method, through adaptive adjustment of the IQR coefficient, more accurately identifies outliers in non-Gaussian distributions, making it particularly suitable for data characterized by sudden changes in performance during late aging. By combining time series and spatial location information to fill missing values, the method leverages the temporal and spatial correlations of the data, resulting in a more accurate representation of the physical laws of cable aging than traditional interpolation methods. Bidirectional LSTM layers simultaneously capture the forward and backward dependencies of time series data, effectively modeling nonlinear degradation characteristics (such as performance mutation points) during the aging process. Compared to unidirectional LSTMs, they offer stronger time series modeling capabilities. The dielectric constant differential equation is incorporated into the loss function as physical prior knowledge, ensuring that model predictions conform to the physical laws of material aging, enhancing model interpretability and engineering credibility. A weighting factor increases the training weight of late-stage data, forcing the model to focus on learning mutation characteristics in the later stages of aging, significantly improving end-of-life prediction accuracy. Model stability is verified through multiple rounds of data partitioning, reducing the risk of overfitting caused by random data partitioning and ensuring model generalization. Real-time monitoring of the validation set mean squared error (MSE) automatically terminates training when no performance improvement is achieved, avoiding overtraining and improving model training efficiency and prediction reliability. Combining the MSE and coefficient of determination, the model's accuracy is comprehensively quantified from two dimensions: absolute error and relative goodness of fit, providing clear targets for model optimization. Through physical constraints and data preprocessing, the model can adapt to complex aging environments characterized by the coupling of multiple factors, such as electric fields, temperature, and mechanical stress, making it particularly suitable for harsh operating conditions such as high-voltage submarine cables. The spatiotemporal K-nearest neighbor algorithm and exponential weighting mechanism effectively alleviate the scarcity of high-voltage cable fault samples, reducing reliance on large-scale failure data and better aligning with engineering practices. Based on the time series prediction characteristics of LSTM, the system can process online monitoring data in real time, dynamically update remaining life predictions, and support timely adjustments to operation and maintenance decisions.

[0057] As a preferred technical means: the data acquisition module includes:

[0058] The dielectric constant acquisition unit uses a broadband dielectric spectrometer to collect data. After each group of samples completes multiple valid tests, the statistical distribution method is used for data processing.

[0059] The crystallinity, thermal decomposition temperature and melting temperature collection unit is collected by a synchronous thermal analyzer. The sample is heated at a constant rate under a protective atmosphere. The thermal decomposition temperature is analyzed by obtaining thermogravimetric curves at different aging times. The crystallinity and melting temperature are analyzed by switching the atmosphere in stages.

[0060] The carbonyl index acquisition unit is collected by attenuated total reflection infrared spectrometer, which uses a set reflection crystal and incident angle to obtain spectral data through scanning and calculation;

[0061] The elastic modulus acquisition unit is collected through a universal material testing machine. The universal material testing machine stretches the sample and records the maximum tensile force and the change in the marked distance at fracture.

[0062] After each sample group completes multiple valid tests, statistical analysis using a two-parameter Weibull or normal distribution is performed to reduce the impact of random errors, ensuring that the dielectric constant data reflects the probabilistic distribution characteristics of the material's overall dielectric properties and meets the requirements of reliability assessment. During the physical simulation of the protective atmosphere and staged heating, argon can be used initially to simulate the oxygen-free environment of submarine cable laying, accurately measuring crystallinity and melting temperature and avoiding oxidation interference with the crystalline state. Later, an oxygen atmosphere can be switched to simulate aerobic thermal aging in service. Thermal decomposition temperatures are obtained from thermogravimetric curves to quantify the impact of oxidative degradation on thermal stability. The simultaneous thermal analyzer's staged atmosphere control and broadband dielectric spectrum frequency scanning simulate the multi-stage aging process experienced in actual service, enabling accelerated test data to reflect natural aging patterns through physical mechanisms, addressing the data distortion problem associated with traditional accelerated testing. Nondestructive testing using carbonyl index acquisition directly obtains infrared spectra of the material's surface oxide layer, eliminating the need for sample sectioning and enabling continuous tracking of the same sample at different aging time points, thus avoiding the destruction of the aging gradient structure by traditional sample preparation.

[0063] As a preferred technical means: the physical constraint layer embeds the dielectric constant differential equation as follows:

[0064]

[0065] Where γ is the dielectric constant; is the activation energy; R is the ideal gas constant; T is the absolute temperature; is the electric field intensity; n is the material constant; is the correction term for LSTM output.

[0066] Beneficial effects:

[0067] 1. A life prediction method based on the fusion of LSTM and physical constraints is adopted. By embedding the dielectric constant differential equation, the model interpretability is enhanced, the problem of multi-stress nonlinear coupling is effectively solved, and the prediction accuracy is significantly improved.

[0068] 2. Introducing an exponentially weighted loss function to strengthen the weight of late-stage aging data significantly improves the model's predictive sensitivity to lifespan mutation stages, significantly reducing prediction errors compared to traditional models.

[0069] 3. Based on the mapping relationship of the traditional Arrhenius equation, the correction term of LSTM output is introduced , compensate for the influence of multiple stress coupling effects and improve the accuracy of life assessment.

[0070] 4. A missing value filling algorithm based on spatiotemporal proximity, combined with a dynamic weight adjustment mechanism, effectively improves the robustness and reliability of small sample data modeling.

[0071] 5. By defining the reciprocal conversion method and normalization process of the inverse index, the dimensional differences of multi-physics field data are eliminated to ensure the consistency and stability of the model input.

[0072] 6. The overall architecture design of the model lightweight, edge deployment and real-time prediction system improves the efficiency of project implementation and long-term operation and maintenance adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0074] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0075] Example 1:

[0076] like Figure 1 As shown, the present invention includes the steps of:

[0077] S1: Data collection: Under the set accelerated aging experimental conditions, key performance parameters at different aging time points are collected, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters;

[0078] S2: Data preprocessing: Inverse indicators are processed in the same direction; outlier detection is performed using the improved box plot method, the IQR coefficient is dynamically adjusted, and outliers are identified and removed; missing data are filled based on the spatiotemporal K-nearest neighbor algorithm; and data are compressed to the [0, 1] interval through minimum-maximum normalization.

[0079] S3: LSTM model construction: The input layer receives preprocessed time series data; the dielectric constant differential equation is embedded in the loss function as a soft constraint; the loss function introduces exponential weights to make the model pay more attention to the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable;

[0080] S4: Model training and optimization: Divide the data into training and test sets, use 5-fold cross-validation to optimize hyperparameters; monitor the mean square error of the validation set to prevent overfitting; and obtain a trained LSTM model;

[0081] S5: Life prediction and evaluation: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and determination coefficient.

[0082] In this embodiment, cross-linked polyethylene (XLPE) cables are selected as the object for data collection, and accelerated aging experimental conditions (such as a high temperature environment of 110° C.) are set.

[0083] Elastic modulus is a key property of polymer materials, directly impacting their appropriate application. Materials were tested using a Zwick / Roell Z020 universal testing machine manufactured by Zwick, Germany. After a dumbbell specimen was placed in the machine's grips, the grips moved at a speed of 250 ± 50 mm / min. During the test, the maximum tensile force was measured and recorded. The distance between the two marked lines at the time of fracture was also measured on the same specimen.

[0084] The test reference standards IEC 60247-2014 and ASTM D924-15 were used. The dielectric loss of each XLPE was tested using a Novocontrol Concept 80 broadband and wide-temperature dielectric spectrometer produced in Germany. Before the test, the sample surface was wiped with anhydrous ethanol. After the sample was dry, its thickness was measured. Then, both sides of the sample were ion sputtered with gold for 60 seconds and a current of 20mA. The gold-sputtered sample was then placed between two copper electrodes with diameters of 20mm and 30mm respectively. The sample was fixed in the test device for measurement. The dielectric loss test can set the test temperature and the test frequency range is 1~10 6 Hz.

[0085] The chemical structure of the exposed surface of the sample was characterized by attenuated total reflection infrared spectrometer, with ZnSe as the reflective crystal, an incident angle of 45°, 32 scans, and a scanning range of 700-4000 cm -1 , with a resolution of 4cm -1 , to calculate the carbonyl index.

[0086] The crystallinity, melting temperature, and thermal decomposition temperature of the sample were analyzed using a TGA / DSC3+ simultaneous thermal analyzer. Under argon protection, the sample was heated from room temperature (30°C) to 200°C at a rate of 10°C / min. The atmosphere was then switched to oxygen and the temperature was continued at the same rate to 250°C. The simultaneous thermal analyzer combines DSC and TGA, allowing the simultaneous measurement of heat flow and weight change curves in a single test.

[0087] In step 2 of the method provided by the present invention, the elastic modulus, dielectric constant (50 Hz), carbonyl index, crystallinity, thermal decomposition temperature and melting temperature in the above step 1 are set as variables X1, X2, X3, X4, X5, X6 that change with aging time.

[0088] Since the dielectric constant (X2) and carbonyl index (X3) are inverse indices in the multi-physics field aging process of submarine cables, they need to be converted into the same direction using the following formula: .

[0089] Use the improved box plot method to detect outliers and dynamically adjust the IQR coefficient to adapt to non-Gaussian distribution data. First, calculate the first quartile (Q1) and the third quartile (Q3) of the data:

[0090]

[0091] If the data point X<Q1-1.5×IQR或X> Q3+1.5×IQR, the data is considered to be an outlier.

[0092] At the same time, based on the spatiotemporal K-Nearest Neighbors (KNN) algorithm, the missing data is filled with the mean of the nearest K observations:

[0093]

[0094] z i is the position coordinate, σ=0.1 is the spatial attenuation coefficient (cable axial position).

[0095] Finally, normalization is performed. In order to eliminate the influence of different dimensions and make all eigenvalues ​​within the same range, minimum-maximum normalization is used, as shown in the following formula:

[0096]

[0097] where X min and X max are the minimum and maximum values ​​in the dataset respectively; after normalization, all data are compressed to the interval [0, 1] to improve the stability of model training.

[0098] In step 3 of the method provided by the present invention, the LSTM network model is constructed using a bidirectional LSTM layer. The bidirectional structure is modeled by joint forward and backward propagation to improve the sensitivity to the aging inflection point. The dielectric constant differential equation is embedded as a soft constraint in the loss function:

[0099]

[0100] Where, γ: dielectric constant; Ea: activation energy; R: ideal gas constant; T: absolute temperature; : electric field intensity; : the correction term of LSTM output, used to compensate for the shortcomings of traditional models.

[0101] The data closer to the end of life is more critical to operation and maintenance decisions, so an exponential weight is introduced into the loss function to make the model pay more attention to the characteristics of the mutation in the later aging stage:

[0102]

[0103] In step 4 of the method provided in the application, the pretreated data is divided into 80% training set and 20% test set, and 5-fold cross-validation is used to optimize the hyperparameters. Monitor the mean square error (MSE) of the validation set. If there is no lower value for 10 consecutive epochs, terminate the training to prevent overfitting.

[0104] In step 5 of the method provided in the application, the training accuracy of the model is evaluated, and the indicators are calculated mean square error (MSE) and determination coefficient (R 2 ), the formulas are respectively:

[0105]

[0106] Among them, is the actual aging data, is the model prediction result.

[0107] The following specific test data and test curve results are used to demonstrate the effect of the application

[0108] According to the cable insulation life prediction method based on LSTM accelerated aging mapping provided in the application, a multi-physical field accelerated aging experiment is carried out on the insulation material of high-voltage submarine cable, and the changes of aging time and key performance parameters are recorded. The original data is shown in Table 1.

[0109] Table 1 Original data of key performance parameters

[0110]

[0111] In step 2 of the method provided in the application, in order to eliminate the influence of dimension in the original data, inverse index homogenization is carried out. Because X2 (dielectric constant) and X3 (carbonyl index) are inverse indexes, reciprocal conversion is adopted, and in order to avoid the risk of zero denominator, a small quantity is introduced to solve the numerical calculation anomaly:

[0112]

[0113] Standardization processing is carried out: using minimum-maximum normalization to compress the data to the interval [0, 1], the results are shown in Table 2.

[0114] Table 2 Data after standardization

[0115]

[0116] Based on their normalized rates of change, select the parameters that contribute most to lifespan prediction. Select X3 (carbonyl index), X4 (crystallinity), and X5 (thermal decomposition temperature) as the primary parameters; select X1 (elastic modulus) as the backup parameter. Then construct an LSTM network structure: the input layer has a 3-time step (24-hour window) and 3 features (X3, X4, X5). The bidirectional LSTM layer has 64 neurons and a tanh activation function. The physical constraint layer embeds the dielectric constant differential equation constraint. The output layer predicts the remaining lifespan percentage for the next 12 hours.

[0117] During training, the data was divided into the first 6 training groups and the last 2 testing groups. The training was terminated if the validation loss did not decrease after 10 consecutive epochs. The prediction results of the LSTM model are shown in Table 3:

[0118] Table 3 Calculated life prediction results

[0119]

[0120] In step 5 of the method provided by the present invention, the mean square error and the coefficient of determination are calculated to evaluate the accuracy of the model, and the results are compared with those obtained by traditional linear regression to compare the RMSE and R 2 , the specific evaluation results are shown in Table 4:

[0121] Table 4 Evaluation results

[0122]

[0123] The smaller the RMSE root mean square error, the better. 2 The closer to 1, the better. This shows that the prediction results are more accurate and better than the traditional linear regression method.

[0124] Example 2:

[0125] A cable insulation life prediction system based on LSTM accelerated aging mapping, including:

[0126] 1. Data acquisition module, used to collect key performance parameters at different aging time points under the set accelerated aging test conditions, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters; insulation electrical performance parameters include dielectric constant; physical and chemical performance parameters include crystallinity, melting temperature, carbonyl index, and thermal decomposition temperature; mechanical performance parameters include elastic modulus;

[0127] The data acquisition module includes:

[0128] The dielectric constant acquisition unit places the gold-sprayed sample between two copper electrodes with diameters of 20 mm and 30 mm respectively, and fixes the sample in the test device for measurement;

[0129] Crystallinity and melting temperature acquisition unit: the sample is heated at a constant rate under a protective atmosphere, and the atmosphere is switched in stages for analysis;

[0130] The carbonyl index acquisition unit is measured by attenuated total reflection infrared spectrometer, using a set reflection crystal and incident angle, and calculating by scanning to obtain spectral data;

[0131] Thermal decomposition temperature acquisition unit: the sample is heated at a constant rate under a protective atmosphere to obtain thermogravimetric curves at different aging times;

[0132] The elastic modulus acquisition unit uses a universal material testing machine to stretch the specimen and record the maximum tensile force and the change in the marked distance at fracture.

[0133] 2. Data preprocessing module, including:

[0134] Isotropic unit, used to perform isotropic processing on inverse indices including dielectric constant and carbonyl index;

[0135] Anomaly detection unit, used to detect outliers using the improved box plot method, dynamically adjust the IQR coefficient, and identify and remove outliers;

[0136] Missing value filling unit, used to fill missing data based on the spatiotemporal K-nearest neighbor algorithm;

[0137] Normalization unit, used to compress data to the interval [0, 1];

[0138] 3. LSTM prediction module, including:

[0139] Bidirectional LSTM layer, which receives preprocessed time series data;

[0140] Physical constraint layer, embedding the dielectric constant differential equation in the loss function;

[0141] The loss function optimization unit introduces exponential weights to make the model pay more attention to the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable;

[0142] Fourth, the model training and optimization module divides the data into training and test sets, uses 5-fold cross-validation to optimize hyperparameters, and monitors the mean square error of the validation set to prevent overfitting;

[0143] 5. Life prediction and evaluation module: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and determination coefficient.

[0144] This embodiment systematically addresses the shortcomings of traditional cable life prediction methods from data acquisition and preprocessing to model construction and evaluation. It has significant advantages in multi-physics field modeling, small sample processing and engineering practicality, and provides reliable technical support for the accurate evaluation of the insulation status of high-voltage cables.

[0145] It is understandable that the detailed functional implementation of the above modules can be found in the introduction of the aforementioned method embodiment, and no further details are given here.

[0146] The cable insulation life prediction method and system based on LSTM accelerated aging mapping shown above is a specific embodiment of the present invention, which has reflected the essential characteristics and progress of the present invention. It can be modified equivalently according to actual usage needs under the guidance of the present invention, and all are within the scope of protection of this solution.

Claims

1. A cable insulation life prediction method based on LSTM accelerated aging mapping, characterized in that The following steps are involved: 1) Data collection: Under the set accelerated aging experimental conditions, key performance parameters at different aging time points are collected, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters; 2) Data preprocessing: Inverse indicators are oriented in the same direction; outlier detection is performed using an improved boxplot method, and the IQR coefficient is dynamically adjusted to identify and remove outliers; missing data are filled using the spatiotemporal K-nearest neighbor algorithm; and data are compressed to the [0, 1] interval using min-max normalization. 3) LSTM model construction: The input layer receives preprocessed time series data; the dielectric constant differential equation is embedded in the loss function as a soft constraint; the loss function introduces exponential weights to make the model more focused on the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable; 4) Model training and optimization: Split the data into training and test sets, use 5-fold cross-validation to optimize hyperparameters, monitor the mean squared error of the validation set to prevent overfitting, and obtain a trained LSTM model. 5) Life prediction and evaluation: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and coefficient of determination; In step 3), the LSTM network model uses a bidirectional LSTM layer. The bidirectional structure is jointly modeled through forward and back propagation. The dielectric constant differential equation is embedded in the loss function as a soft constraint: Where γ is the dielectric constant; is the activation energy; R is the ideal gas constant; T is the absolute temperature; is the electric field intensity; n is the material constant; is the correction term for LSTM output.

2. A cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1, characterized in that: In step 1), the insulation electrical performance parameters include dielectric constant; the physical and chemical performance parameters include carbonyl index, crystallinity, thermal decomposition temperature and melting temperature; the mechanical performance parameters include elastic modulus; the dielectric constant is collected by a broadband dielectric spectrometer; The crystallinity, thermal decomposition temperature and melting temperature were collected by synchronous thermal analyzer; the carbonyl index was obtained by scanning with attenuated total reflection infrared spectrometer.

3. The cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1 is characterized in that: The inverse indicator is converted into the same direction using the following formula: , where is the original inverse index value, is the positive indicator value after homologation; The outlier determination condition of the improved box plot method is: if a data point satisfies X < Q1 - 1.5 × IQR or X > Q3 + 1.5 × IQR, then the data is considered an outlier, where ; Q1 is the first quartile and Q3 is the third quartile; The formula for filling missing data based on the spatiotemporal K-nearest neighbor algorithm is: ； Where z i is the spatial coordinate of the i-th neighbor, z t is the spatial coordinate of the target point, σ is the spatial attenuation coefficient, k is the number of neighbors, is the observation value of the ti-th point in the time dimension; The normalization formula is: where X min and X max are the minimum and maximum values ​​in the data set, respectively.

4. The cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1 is characterized in that: The exponential weight introduced in the loss function is: ; Where α is the weight adjustment coefficient, t is the current time point, and S is the total aging time.

5. The cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1 is characterized in that: In step 4), when monitoring the mean square error of the validation set, if there is no improvement over multiple epochs, the training is terminated.

6. The cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1 is characterized in that: In step 5), the mean square error formula is: The coefficient of determination formula is: Where N is the number of test samples, is the actual lifespan value, is the predicted lifespan value, is the mean of the actual life span values.

7. A cable insulation life prediction system using the cable insulation life prediction method based on LSTM accelerated aging mapping according to claim 1, characterized in that include: The data acquisition module is used to collect key performance parameters at different aging time points under the set accelerated aging experimental conditions, including insulation electrical performance parameters, physical and chemical performance parameters, and mechanical performance parameters; insulation electrical performance parameters include dielectric constant; physical and chemical performance parameters include carbonyl index, crystallinity, thermal decomposition temperature, and melting temperature; Mechanical performance parameters include elastic modulus; Data preprocessing module, including: Isotropic unit, used to perform isotropic processing on inverse indices including dielectric constant and carbonyl index; Anomaly detection unit, used to detect outliers using the improved box plot method, dynamically adjust the IQR coefficient, and identify and remove outliers; Missing value filling unit, used to fill missing data based on the spatiotemporal K-nearest neighbor algorithm; Normalization unit, used to compress data to the interval [0, 1]; LSTM prediction module, including: Bidirectional LSTM layer, which receives preprocessed time series data; Physical constraint layer, embedding the dielectric constant differential equation in the loss function; The loss function optimization unit introduces exponential weights to make the model pay more attention to the characteristics of late aging mutations; the output layer predicts the remaining life percentage of the cable; The model training and optimization module divides the data into training and test sets, uses 5-fold cross-validation to optimize hyperparameters, and monitors the mean square error of the validation set to prevent overfitting. Life prediction and evaluation module: Use the trained model to predict the remaining life of the cable, and evaluate the prediction accuracy through mean square error and determination coefficient.

8. The cable insulation life prediction system according to claim 7, characterized in that: The data acquisition module includes: The dielectric constant acquisition unit uses a broadband dielectric spectrometer to collect data. After each group of samples completes multiple valid tests, the statistical distribution method is used for data processing. The crystallinity, thermal decomposition temperature and melting temperature collection unit is collected by a synchronous thermal analyzer. The sample is heated at a constant rate under a protective atmosphere. The thermal decomposition temperature is analyzed by obtaining thermogravimetric curves at different aging times. The crystallinity and melting temperature are analyzed by switching the atmosphere in stages. The carbonyl index acquisition unit is collected by attenuated total reflection infrared spectrometer, which uses a set reflection crystal and incident angle to obtain spectral data through scanning and calculation; The elastic modulus acquisition unit is collected through a universal material testing machine. The universal material testing machine stretches the sample and records the maximum tensile force and the change in the marked distance at fracture.

9. The cable insulation life prediction system according to claim 7, characterized in that: The physical constraint layer embeds the dielectric constant differential equation as follows: Where γ is the dielectric constant; is the activation energy; R is the ideal gas constant; T is the absolute temperature; is the electric field intensity; n is the material constant; is the correction term for LSTM output.

Images