A high-voltage bushing long-time operation reliability accelerated electrical life test method

Abstract

The application provides an accelerated electrical life test method for long-time operation reliability of high-voltage bushings, and belongs to the technical field of high-voltage bushing experiments. The application extracts a discharge sequence complexity index through wavelet packet decomposition noise reduction and a symbolic dynamics complexity algorithm, extracts an aging characteristic component through independent component analysis dimension reduction processing, inputs an ultra-sparse time domain coupling identification model to predict an electrical tree branch growth rate and a Weibull distribution shape parameter, adopts Monte Carlo random simulation to generate a breakdown time sample and perform Weibull distribution fitting to calculate an acceleration factor, optimizes test conditions through an adaptive stress regulation function, and finally adopts a Bayesian inference framework to establish an acceleration factor posterior distribution and generate a posterior sample through a Markov chain Monte Carlo sampling method to calculate a confidence interval of a long-term life prediction value, thereby solving the problem that when extrapolating a long-term life, an acceleration factor statistical error is magnified, leading to difficulty in accurately evaluating prediction confidence.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage bushing experimental technology, and specifically relates to an accelerated electrical life test method for the long-term operational reliability of high-voltage bushings. Background Technology

[0002] Long-term reliability assessment of high-voltage bushings typically employs accelerated electrical life (AEW) testing. This method shortens the test cycle by applying voltage and temperature stresses exceeding rated values, and then extrapolates the lifespan under normal operating conditions using an acceleration factor. Traditional AEW testing relies primarily on the statistical characteristics of partial discharge amplitude and the mean dielectric loss angle as aging condition monitoring parameters. The acceleration factor is calculated by fitting breakdown time data using a Weibull distribution, and the results are extrapolated to long-term lifespan using point estimation methods. However, in current high-voltage bushing AEW testing, due to limited sample sizes and significant differences between accelerated stress conditions and actual operating conditions, the acceleration factor calculated using traditional methods often contains substantial statistical uncertainty. This uncertainty is amplified when extrapolating to long-term lifespan, resulting in excessively wide confidence intervals for life predictions or even an inability to provide an effective confidence assessment. In other words, existing technologies suffer from a technical problem where the statistical error of the acceleration factor is amplified when extrapolating long-term lifespan using AEW testing, making accurate assessment of prediction confidence difficult. Summary of the Invention

[0003] In view of this, the present invention provides an accelerated electrical life test method for the long-term operational reliability of high-voltage bushings, which can solve the technical problem in the prior art where the statistical error of the acceleration factor is amplified when extrapolating the long-term life in the accelerated electrical life test, making it difficult to accurately assess the prediction confidence.

[0004] This invention is implemented as follows: It provides an accelerated electrical life test method for the long-term operational reliability of high-voltage bushings. A partial discharge ultrasonic sensor array and a dielectric loss meter are arranged on the surface of the high-voltage bushing to simultaneously acquire partial discharge pulse sequences and dielectric loss angle time sequences. The partial discharge pulse sequences are subjected to wavelet packet decomposition for noise reduction to extract singular values ​​and reconstruct signals. The complexity index of the discharge sequence is calculated using a symbolic dynamics complexity algorithm. The dielectric loss angle time sequence is subjected to independent component analysis for dimensionality reduction to extract aging characteristic components. An ultra-sparse temporal coupling identification model is input, outputting predicted values ​​of electrical tree growth rate and Weibull distribution shape parameters. A Monte Carlo random simulation method is used to generate electrical tree growth path samples and calculate the breakdown time. The breakdown time is fitted with a Weibull distribution to calculate the acceleration factor and its confidence interval. Voltage stress and temperature stress are optimized using an adaptive stress control function. A Bayesian inference framework is used to establish the posterior distribution of the acceleration factor, and a Markov chain Monte Carlo sampling method is used to generate posterior samples of the acceleration factor to calculate the confidence interval of the long-term lifetime prediction value.

[0005] Specifically, the wavelet packet decomposition noise reduction process involves performing a five-level wavelet packet decomposition on the partial discharge pulse sequence, selecting the db4 wavelet basis function, constructing a Hankel matrix for the wavelet coefficients of each frequency band, performing singular value decomposition, retaining the singular vectors corresponding to the first eight to twelve singular values, reconstructing the wavelet coefficients through the retained singular vectors, and obtaining the singular value reconstructed signal through inverse wavelet packet transformation.

[0006] Specifically, the symbol dynamics complexity algorithm performs amplitude normalization processing on the singular value reconstruction signal, sets a symbolization threshold, encodes pulses with normalized amplitude greater than the symbolization threshold as symbol 1, and encodes pulses with normalized amplitude less than or equal to the symbolization threshold as symbol 0 to form a binary symbol string, and uses the Lempel-Ziv algorithm to calculate the complexity of the binary symbol string.

[0007] Specifically, the independent component analysis dimensionality reduction process involves centering and whitening the dielectric loss angle time series, using a fast independent component analysis algorithm to iteratively optimize the separation matrix to maximize the non-Gaussianity of the separated signal, setting the number of independent components to three to correspond to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism respectively, calculating the component loading matrix, and extracting the three independent components with the largest absolute load values ​​as aging feature components.

[0008] Specifically, the ultra-sparse temporal coupling identification model involves an input layer receiving the discharge sequence complexity index and aging feature components, a second hidden layer introducing an ultra-sparse temporal coupling mechanism to control information transmission between neurons through a dynamic connection matrix, the sparsity of which is determined by the discharge sequence complexity index, the connection weights being modulated by a temporal coupling strength function, and an output layer outputting the predicted value of electrical tree growth rate and the shape parameters of the Weibull distribution.

[0009] Specifically, the sparsity calculation of the dynamic connection matrix is ​​based on the normalized relationship between the discharge sequence complexity index and the maximum and minimum values ​​of the discharge sequence complexity index in the training set. The sparsity value ranges from 0.65 to 0.95.

[0010] Specifically, the time-domain coupling strength function is the product of the coupling weight and the periodic modulation term, and the coupling frequency is determined by the rate of change of the dielectric loss angle time series, ranging from 0.01 to 0.1 Hz.

[0011] Specifically, the Monte Carlo random simulation method generates electric tree growth path samples by establishing a three-dimensional coordinate system and setting the electric tree growth step size. The growth direction of each step is determined by the local electric field intensity gradient and random perturbation. The random perturbation follows a normal distribution, and the standard deviation of the normal distribution is determined by the fractal dimension. In each growth step, the electric field intensity at the current position is calculated. If the electric field intensity exceeds the material breakdown field strength threshold, it is determined to be a breakdown.

[0012] Specifically, the fractal dimension is calculated by projecting the electric tree growth path sample onto a two-dimensional plane using box counting, covering the electric tree growth path sample with squares of different sizes, and statistically analyzing the logarithmic relationship between the number of squares required to cover the electric tree growth path sample and the square size. The slope of the logarithmic relationship is the fractal dimension.

[0013] Specifically, the Weibull distribution fitting involves sorting the breakdown times from smallest to largest, using the maximum likelihood estimation method to calculate the shape and scale parameters of the Weibull distribution, using the fitted shape and scale parameters of the Weibull distribution to calculate the acceleration factor, and using the Bootstrap resampling method to evaluate the uncertainty of the estimated shape and scale parameters of the Weibull distribution and calculate the 95% confidence interval of the acceleration factor.

[0014] Specifically, the adaptive stress control function extracts the acceleration factor and acceleration factor confidence interval obtained by fitting the Weibull distribution, calculates the ratio of the acceleration factor confidence interval width to the acceleration factor as the uncertainty index, calculates the ratio of the initial voltage stress to the rated voltage as the voltage ratio, calculates the stress adjustment coefficient based on the uncertainty index and the voltage ratio, and adjusts the new voltage stress and new temperature stress respectively when the uncertainty index is in different intervals.

[0015] Specifically, the Bayesian inference framework establishes the posterior distribution of the acceleration factor by setting the prior distribution of the acceleration factor as a log-normal distribution, constructing a likelihood function based on the fitting results of the Weibull distribution and experimental observation data, and calculating the posterior distribution of the acceleration factor using the Bayesian formula.

[0016] Specifically, the Markov chain Monte Carlo sampling method uses the Metropolis-Hastings algorithm to generate acceleration factor posterior samples from the acceleration factor posterior distribution. The sampling chain length is 10,000 to 50,000, and the first 1,000 to 5,000 acceleration factor posterior samples are discarded as pre-burning samples.

[0017] Specifically, the long-term lifetime prediction value is obtained by dividing each sample value of the posterior sample of the acceleration factor by the characteristic lifetime under accelerated test conditions to obtain the characteristic lifetime under normal operation conditions. The 2.5% quantile and 97.5% quantile of the posterior sample of the acceleration factor are used as the 95% confidence interval of the long-term lifetime prediction value.

[0018] The partial discharge ultrasonic sensor array comprises 8 to 16 sensor units arranged in a circular array on the outer surface of the high-voltage bushing, with a spacing of 50 to 100 mm between adjacent sensor units. The dielectric loss meter uses the principle of capacitive voltage division to calculate the dielectric loss tangent by measuring the phase difference between voltage and current.

[0019] Specifically, the coupling frequency calculation of the ultra-sparse time-domain coupling identification model is the product of the coupling frequency of 0.01 plus the difference between the dielectric loss angle time series at adjacent sampling points and the time interval between adjacent sampling points.

[0020] This invention establishes an ultra-sparse temporal-domain coupled identification model to predict the growth rate of electrical trees and combines this with Monte Carlo random simulation to generate a large number of breakdown time samples. A Bayesian inference framework is used to construct the posterior distribution of the acceleration factor, and Markov chain Monte Carlo sampling is employed to generate posterior samples of the acceleration factor, thus solving the problem that traditional point estimation methods cannot quantify the statistical error of the acceleration factor. The ultra-sparse temporal-domain coupled identification model dynamically adjusts the sparsity of neural network connections and the temporal coupling strength function, embedding the polarization relaxation physical process of insulating materials under alternating electric fields into the model architecture. This makes the model prediction not only dependent on the statistical characteristics of the data but also follow the physical mechanism of electrical tree growth, significantly improving the accuracy of electrical tree growth rate prediction. The breakdown time samples generated by Monte Carlo random simulation based on this prediction result are more consistent with actual statistical laws. The Bayesian inference framework combines the prior information of the acceleration factor with experimental observation data to calculate the posterior distribution. The Markov chain Monte Carlo sampling method generates a large number of samples from this posterior distribution to fully characterize the uncertainty of the acceleration factor and transfer it to the calculation of the confidence interval of the long-term lifetime prediction value. In summary, this invention solves the technical problem mentioned in the background art where the statistical error of the acceleration factor is amplified when extrapolating long-term lifetime in accelerated electrical lifetime tests, making it difficult to accurately assess the prediction confidence. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a waveform diagram of a partial discharge pulse sequence.

[0023] Figure 3 This is a time series curve showing the change in the dielectric loss angle.

[0024] Figure 4 The graph shows the relationship between the discharge sequence complexity index and aging time.

[0025] Figure 5 This is a probability distribution diagram of long-term lifetime predictions. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0027] like Figure 1The diagram shows a flowchart of an accelerated electrical life test method for long-term operational reliability of high-voltage bushings provided by this invention. This method includes the following steps: S1. A partial discharge ultrasonic sensor array and a dielectric loss meter are uniformly arranged on the surface of the high-voltage bushing, wherein the frequency response range of the partial discharge ultrasonic sensor array is 20 to 500 kHz. S2. Install the high-voltage bushing on the accelerated aging test platform, apply an initial voltage stress of 1.5 to 2.0 times the rated voltage, apply an initial temperature stress of 60 to 90°C, and simultaneously collect partial discharge pulse sequences and dielectric loss angle time sequences, recording the sampling frequency at 10MHz; S3. Perform wavelet packet decomposition noise reduction on the partial discharge pulse sequence acquired in step S2, extract singular value reconstructed signal, calculate discharge sequence complexity index through symbol dynamics complexity algorithm, and perform independent component analysis dimensionality reduction on the dielectric loss angle time series acquired in step S2 to extract aging characteristic components. S4. Input the discharge sequence complexity index and aging feature components extracted in step S3 into the ultra-sparse temporal coupling identification model. The ultra-sparse temporal coupling identification model outputs the predicted value of electric tree growth rate and the shape parameter of Weibull distribution. S5. Based on the predicted electrical tree growth rate output in step S4, generate 1,000 to 5,000 electrical tree growth path samples using the Monte Carlo random simulation method, and calculate the breakdown time of each electrical tree growth path sample by combining the fractal dimension. S6. Fit the breakdown time obtained in step S5 to a Weibull distribution, calculate the acceleration factor and the acceleration factor confidence interval. If the ratio of the acceleration factor confidence interval width to the acceleration factor is greater than 0.3, recalculate the voltage stress and temperature stress through the adaptive stress control function and return to step S2. S7. If the ratio of the confidence interval width of the acceleration factor to the acceleration factor in step S6 is less than or equal to 0.3, the posterior distribution of the acceleration factor is established using the Bayesian inference framework. 10,000 to 50,000 posterior samples of the acceleration factor are generated using the Markov chain Monte Carlo sampling method, and the 95% confidence interval of the long-term lifetime prediction is calculated.

[0028] The wavelet packet decomposition and noise reduction process specifically includes: performing 5-level wavelet packet decomposition on the partial discharge pulse sequence acquired in step S2, selecting the db4 wavelet basis function, and obtaining 32 frequency bands after decomposition; constructing a Hankel matrix for the wavelet coefficients of each frequency band, wherein the dimension of the Hankel matrix is ​​50×50; performing singular value decomposition on the Hankel matrix, retaining the singular vectors corresponding to the first 8 to 12 singular values; reconstructing the wavelet coefficients through the retained singular vectors, and obtaining the singular value reconstructed signal through inverse wavelet packet transformation.

[0029] The measurement accuracy of the dielectric loss meter is as follows: class.

[0030] The wavelet packet decomposition effectively separates discharge signals and noise components of different frequency bands by simultaneously subdividing the partial discharge pulse sequence in both the time and frequency domains. The singular value decomposition extracts the main features of the partial discharge pulse sequence through matrix rank reduction, suppressing the influence of random noise and electromagnetic interference. The wavelet packet decomposition denoising process can accurately extract weak partial discharge signals in strong noise environments, avoiding misjudgments caused by noise fluctuations in traditional thresholding methods. The singular value reconstruction signal retains the time-frequency characteristics of the real discharge signal, providing high signal-to-noise ratio input data for subsequent discharge type identification and aging status assessment.

[0031] The specific steps of the symbolic dynamics complexity algorithm include: performing amplitude normalization processing on the singular value reconstruction signal obtained in step S3, and the formula for amplitude normalization processing is as follows: The The normalized amplitude of the i-th pulse, the The original amplitude unit of the i-th pulse is mV. The maximum amplitude in the singular value reconstructed signal is in mV. The minimum amplitude unit in the singular value reconstructed signal is mV; the symbolization threshold is set to 0.5, pulses with a normalized amplitude greater than 0.5 are encoded as symbol 1, and pulses with a normalized amplitude less than or equal to 0.5 are encoded as symbol 0, forming a binary symbol string; the frequency of occurrence of all different symbol patterns with lengths of 2 to 6 in the binary symbol string is counted; the complexity of the binary symbol string is calculated using the Lempel-Ziv algorithm. The Lempel-Ziv algorithm scans the binary symbol string from left to right, and the complexity counter is incremented by 1 whenever a new symbol pattern appears. The final discharge sequence complexity index is the ratio of the complexity counter value to the length of the binary symbol string.

[0032] The symbol pattern refers to a continuous combination of symbols in a binary symbol string. The symbol dynamics complexity algorithm converts continuous discharge amplitude signals into discrete binary symbol strings. By statistically analyzing the diversity of symbol patterns, it quantifies the randomness of the discharge process. In the early stages of aging, the discharge is dominated by fixed defects and exhibits a periodic symbol pattern, thus the discharge sequence complexity index is low. As breakdown approaches, the randomness increases due to the superposition of multiple discharge sources, resulting in a significant increase in the discharge sequence complexity index. The dynamic change of the discharge sequence complexity index reflects the evolution stage of the microstructure of the insulating material, providing a sensitive feature parameter for identifying the aging state. The symbol dynamics complexity algorithm, through pattern statistics using the Lempel-Ziv algorithm, can capture the long-range correlation and non-stationary characteristics in the discharge sequence, and compared with traditional statistical features, it can more accurately characterize the evolution of discharge behavior during electrical tree growth.

[0033] The independent component analysis (ICA) dimensionality reduction steps specifically include: centering the dielectric loss angle time series collected in step S2 by subtracting the mean of the dielectric loss angle time series from the dielectric loss angle value at each time step; performing a whitening transformation on the centered dielectric loss angle time series, which eliminates the correlation between data at different times through eigenvalue decomposition of the covariance matrix; iteratively optimizing the separation matrix using the FastICA algorithm, where the iteration process of the FastICA algorithm maximizes the non-Gaussianity of the separation signal, which is measured by the negative entropy approximation function; setting the number of independent components to three, which correspond to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively; calculating the component loading matrix, which represents the contribution weight of the data at each time step in the dielectric loss angle time series to the three independent components; and extracting the three independent components with the largest absolute values ​​of the loadings in the component loading matrix as aging feature components.

[0034] The independent component analysis (ICA) dimensionality reduction process is based on the statistical independence assumption of source signals. It decomposes the observed parameter of dielectric loss angle time series into uncorrelated independent sources. The dielectric loss angle time series is affected by multiple aging mechanisms such as hydrolysis aging, pyrolysis aging, and oxidation aging, forming aliased signals. By maximizing the non-Gaussianity criterion, the independent contribution of each aging mechanism is separated. The component loading matrix reveals the activity level of each aging mechanism at different measurement times, realizing the projection transformation from the high-dimensional time series space to the low-dimensional intrinsic aging space. The ICA dimensionality reduction process eliminates redundant information in the dielectric loss angle time series, extracts aging feature components that reflect the essence of insulation degradation, and significantly improves the accuracy and computational efficiency of the ultra-sparse time-domain coupled identification model in identifying aging states. Compared with the principal component analysis method, the FastICA algorithm can separate nonlinear aliased aging mechanism signals, which is more consistent with the physical nature of the multiple aging processes of insulating materials.

[0035] The specific structure of the ultra-sparse temporal coupling identification model is as follows: the input layer receives four parameters extracted in step S3: the discharge sequence complexity index and the aging feature components; the first hidden layer contains 64 neurons, and the activation function of the first hidden layer is the hyperbolic tangent function; the second hidden layer contains 32 neurons, and the second hidden layer introduces an ultra-sparse temporal coupling mechanism. The ultra-sparse temporal coupling mechanism controls the information transmission between neurons through a dynamic connection matrix. The sparsity of the dynamic connection matrix is ​​determined by the discharge sequence complexity index extracted in step S3, and the formula for calculating the sparsity of the dynamic connection matrix is ​​as follows: The sparsity of the dynamic connection matrix, s, ranges from 0.65 to 0.95, and C is the discharge sequence complexity index extracted in step S3. To maximize the complexity exponent of the training set of discharge sequences, the To minimize the exponential complexity of the training-focused firing sequence; each neuron in the second hidden layer establishes a connection only with neurons in the first hidden layer whose sparsity ratio corresponds to that of the dynamic connection matrix, and the weights of these connections are modulated by a temporal coupling strength function, which is described as follows: The The coupling weights between the i-th neuron and the j-th neuron at time t are... Based on the weights, the The coupling frequency is determined by the rate of change of the dielectric loss angle time series acquired in step S2, ranging from 0.01 to 0.1 Hz, where T is the normalized time constant with a value of 100 s; the third hidden layer contains 16 neurons, and the activation function of the third hidden layer is the ReLU function; the output layer contains 2 neurons, and the 2 neurons of the output layer output the predicted value of the electric tree growth rate in units of μm / h and the shape parameter of the Weibull distribution, respectively.

[0036] The steps for establishing the training dataset of the ultra-sparse temporal coupling identification model specifically include: collecting 200 to 500 sets of accelerated aging test data of high-voltage bushings under different voltage and temperature stress conditions; extracting the discharge sequence complexity index and aging feature components from each set of accelerated aging test data of high-voltage bushings as input samples for the training dataset; measuring the actual electrical tree growth rate through microscopic section observation and electrical tree imaging technology, and recording the breakdown time as the output label of the training dataset; and dividing the training dataset into a training subset, a validation subset, and a test subset in a ratio of 7:2:1.

[0037] The specific steps for training the ultra-sparse temporal coupling identification model include: updating parameters using the Adam optimizer, with an initial learning rate of 0.001 that decays to 0.95 times the original value every 50 epochs; the loss function is a weighted sum of mean squared error and physical constraint terms, with the physical constraint terms penalizing predictions that violate the monotonicity of electrical tree growth; evaluating the performance of the ultra-sparse temporal coupling identification model on a validation subset every 10 epochs during training, and stopping training early if the validation subset loss does not decrease for 20 consecutive epochs; and selecting the model parameters with the minimum validation subset loss as the training result of the ultra-sparse temporal coupling identification model.

[0038] The ultra-sparse temporal coupling mechanism dynamically adjusts the sparsity of the dynamic connection matrix within the ultra-sparse temporal coupling identification model. This allows the model to adaptively select information transmission paths based on the discharge sequence complexity exponent. In the early stages of aging, when discharge regularity is strong, high-sparse connections are used to reduce parameter redundancy. As discharge randomness increases near breakdown, the sparsity of the dynamic connection matrix is ​​reduced to capture richer feature interactions. The temporal coupling strength function introduces a periodic modulation term to simulate the polarization relaxation process of insulating materials under alternating electric fields. The coupling frequency is correlated with the rate of change of the dielectric loss angle over time. The intrinsic connection between macroscopic electrical parameters and microscopic molecular motion is revealed. The ultrasparse temporal coupling mechanism embeds the temporal evolution law of physical processes into the ultrasparse temporal coupling identification model architecture, so that the prediction of the ultrasparse temporal coupling identification model not only depends on the statistical characteristics of data but also follows the physical mechanism of electric tree growth. This significantly enhances the generalization ability and extrapolation accuracy of the ultrasparse temporal coupling identification model under accelerated stress condition changes. At the same time, the sparse connection of the dynamic connection matrix greatly reduces the computational complexity and overfitting risk, enabling the ultrasparse temporal coupling identification model to accurately identify the electric tree growth law and reliably predict the breakdown time distribution under limited experimental data.

[0039] The calculation steps of the adaptive stress regulation function specifically include: extracting the acceleration factor and acceleration factor confidence interval obtained from the Weibull distribution fitting in step S6; calculating the ratio of the acceleration factor confidence interval width to the acceleration factor, denoted as the uncertainty index; calculating the ratio of the initial voltage stress applied in step S2 to the rated voltage, denoted as the voltage ratio; and calculating the stress regulation coefficient based on the uncertainty index and the voltage ratio. The formula for calculating the stress regulation coefficient is as follows: The The stress adjustment coefficient, the As an uncertainty indicator, the For voltage multiplier; when the uncertainty index When ∈ [0.3, 0.5), the new voltage stress is the initial voltage stress applied in step S2 multiplied by the stress adjustment coefficient, and the new temperature stress remains unchanged compared to the initial temperature stress applied in step S2; when the uncertainty index When the voltage index is ∈ [0.5, 0.7), the new voltage stress is the initial voltage stress applied in step S2 multiplied by the stress adjustment coefficient, and the new temperature stress is the initial temperature stress applied in step S2 plus 5-10°C; when the uncertainty index is... When the voltage is ∈ [0.7, 1.0], the new voltage stress is the initial voltage stress applied in step S2 multiplied by 1.2 times the stress adjustment coefficient, and the new temperature stress is the initial temperature stress applied in step S2 plus 10-15°C; when the uncertainty index When the voltage is greater than 1.0, the new voltage stress is 2.5 to 3.0 times the rated voltage, and the new temperature stress is 90 to 100℃.

[0040] The specific steps of generating electrical tree growth path samples using the Monte Carlo random simulation method include: establishing a three-dimensional coordinate system, with the origin of the three-dimensional coordinate system set at the initial defect location of the electrical tree; setting the electrical tree growth step size to 0.1–0.5 μm, with the growth direction of each step determined by the local electric field intensity gradient and random perturbation; the local electric field intensity gradient being calculated through finite element simulation, with a mesh size of 1 μm; and the random perturbation following a normal distribution, the standard deviation of which is determined by the fractal dimension, and the formula for calculating the standard deviation of the normal distribution is as follows: rad, the Let be the standard deviation of the normal distribution, and the... The fractal dimension is used. In each growth step, the electric field strength at the current position is calculated. If the electric field strength exceeds the material breakdown field strength threshold of 25-35 kV / mm, it is determined to be a breakdown. The current growth step number is recorded and multiplied by the electric tree growth step length to obtain the electric tree length. The above process is repeated to generate 1000-5000 electric tree growth path samples. The electric tree length of each electric tree growth path sample is divided by the electric tree growth rate prediction value output in step S4 and converted into a breakdown time unit of h.

[0041] The fractal dimension is calculated using box counting. Electrical tree growth path samples are projected onto a two-dimensional plane, and squares of different sizes are used to cover the samples. The logarithmic relationship between the number of squares required to cover the samples and the square size is calculated; the slope of this logarithmic relationship is the fractal dimension, which ranges from 1.6 to 1.9. The Monte Carlo stochastic simulation method overcomes the strong randomness of electrical tree growth due to microscopic defects and electric field fluctuations by randomly sampling a large number of electrical tree growth path samples. The fractal dimension quantifies the spatial complexity of the electrical tree growth path samples, and the standard deviation of the normal distribution, determined based on the fractal dimension, ensures that random disturbances conform to the statistical laws of actual electrical tree growth. Compared to deterministic models, the Monte Carlo stochastic simulation method significantly improves the reliability of lifetime prediction by obtaining the probability distribution of breakdown time, providing a statistical sample basis for subsequent Weibull distribution fitting.

[0042] The specific steps of fitting the Weibull distribution include: sorting the breakdown times obtained in step S5 using the Monte Carlo random simulation method from smallest to largest; calculating the shape and scale parameters of the Weibull distribution using the maximum likelihood estimation method, where the shape parameter reflects the trend of failure rate over time and the scale parameter characterizes the characteristic lifetime; calculating the acceleration factor using the fitted shape and scale parameters of the Weibull distribution, where the acceleration factor is the ratio of the characteristic lifetime under accelerated testing conditions to the characteristic lifetime under normal operating conditions; and evaluating the uncertainty of the estimated shape and scale parameters of the Weibull distribution using the Bootstrap resampling method, where the Bootstrap resampling method involves 500 to 1000 resampling cycles, and calculating the 95% confidence interval of the acceleration factor as the acceleration factor confidence interval.

[0043] The accelerated test conditions refer to the initial voltage stress and initial temperature stress applied in step S2, and the normal operating conditions refer to the working state of the high-voltage bushing under rated voltage and ambient temperature, wherein the ambient temperature is 15-35℃.

[0044] The steps of establishing the posterior distribution of the acceleration factor within the Bayesian inference framework specifically include: setting the prior distribution of the acceleration factor calculated in step S6 as a log-normal distribution, wherein the parameters of the log-normal distribution are determined based on historical experimental data and engineering experience; constructing a likelihood function based on the Weibull distribution fitting results and experimental observation data from step S6; calculating the posterior distribution of the acceleration factor using the Bayesian formula, wherein the posterior distribution of the acceleration factor is proportional to the product of the log-normal distribution and the likelihood function; and generating posterior samples of the acceleration factor from the posterior distribution using a Markov chain Monte Carlo sampling method. The sampling algorithm used is the Metropolis-Hastings algorithm. The sampling chain length of the Markov chain Monte Carlo sampling method is 10,000 to 50,000. The first 1,000 to 5,000 acceleration factor posterior samples are discarded as pre-burn-in samples. The long-term lifetime prediction value is calculated based on the acceleration factor posterior samples. The long-term lifetime prediction value is the characteristic lifetime under normal operating conditions, which is obtained by dividing each sample value of the acceleration factor posterior samples by the characteristic lifetime under accelerated test conditions. The 2.5% quantile and 97.5% quantile of the acceleration factor posterior samples are used as the 95% confidence interval of the long-term lifetime prediction value.

[0045] The Markov chain Monte Carlo sampling method constructs a Markov chain whose stationary distribution converges to the posterior distribution of the acceleration factor, thereby achieving numerical sampling of complex probability distributions. The Metropolis-Hastings algorithm generates candidate samples in each iteration and decides whether to accept the candidate samples based on the acceptance probability, which is determined by the ratio of posterior probability densities. The Markov chain Monte Carlo sampling method samples from non-standard distributions without calculating a normalization constant. The discarding of pre-burning samples ensures that the sampling chain has fully converged to a stationary distribution. The posterior sample of the acceleration factor fully characterizes the uncertainty of the acceleration factor and transmits the statistical error of the acceleration factor to the long-term lifetime prediction value. The 95% confidence interval of the long-term lifetime prediction value quantitatively assesses the credibility of the extrapolation results, providing a risk assessment basis for high-voltage bushing operation and maintenance decisions. Compared with the point estimation method, it provides more comprehensive uncertainty quantification information. The Bayesian inference framework effectively solves the problem that the statistical error of the acceleration factor is amplified when extrapolating long-term lifetime from short-term test data, making it difficult to assess the confidence level.

[0046] The partial discharge ultrasonic sensor array comprises 8 to 16 sensor units, which are arranged in a circular array on the outer surface of the high-voltage bushing, with a spacing of 50 to 100 mm between adjacent sensor units. The dielectric loss meter employs the capacitive voltage divider principle, calculating the dielectric loss tangent by measuring the phase difference between voltage and current. The accelerated aging test platform includes a high-voltage power supply temperature control box and a data acquisition system. The high-voltage power supply output voltage ranges from 0 to 100 kV, and the temperature control box has a temperature adjustment range from room temperature to 150°C with a temperature fluctuation of less than ±2°C.

[0047] The discharge sequence complexity index ranges from 0.1 to 0.9. The aging characteristic component is the time-domain average of three independent components, which correspond to the contributions of the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively. The physical meaning of the predicted electrical tree growth rate is the average extension length of the electrical tree per unit time.

[0048] The material breakdown field strength threshold is determined according to the type of insulating material. For cross-linked polyethylene, the material breakdown field strength threshold is 28 to 32 kV / mm, and for silicone rubber, the material breakdown field strength threshold is 25 to 30 kV / mm.

[0049] Furthermore, the number of wavelet packet decomposition layers in step S3 is determined based on the sampling frequency and discharge pulse main frequency band recorded in step S2. When the sampling frequency is 10MHz, 5-layer decomposition is selected, and when the sampling frequency is 20MHz, 6-layer decomposition is selected.

[0050] Furthermore, the formula for calculating the coupling frequency of the ultra-sparse temporal coupling identification model in step S4 is as follows: / Hz, the The coupling frequency is in Hz. The difference between adjacent sampling points of the dielectric loss angle time series collected in step S2, wherein The time interval between adjacent sampling points is in seconds.

[0051] Furthermore, in step S6, the width of the acceleration factor confidence interval is the upper limit of the acceleration factor confidence interval minus the lower limit of the acceleration factor confidence interval.

[0052] The present invention also provides a computer-based method for forming an accelerated electrical life test system for long-term operational reliability of high-voltage bushings. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run on the computer, they execute the above-described method.

[0053] The specific implementation methods of the above steps are described in detail below.

[0054] The specific implementation of step S1 involves uniformly arranging 8 to 16 partial discharge ultrasonic sensor units in a circumferential array on the outer surface of the high-voltage bushing. The spacing between adjacent sensor units is controlled within the range of 50 to 100 mm. The frequency response range of the sensor units is set to 20 to 500 kHz to capture the ultrasonic signals generated by partial discharge. Simultaneously, a dielectric loss meter is installed on the high-voltage bushing to monitor the change in dielectric loss angle in real time. The dielectric loss meter uses the principle of capacitive voltage division to calculate the tangent of the dielectric loss angle by measuring the phase difference between voltage and current, achieving a measurement accuracy of one ten-thousandth. This sensor arrangement scheme can comprehensively monitor the partial discharge activity and overall electrical performance degradation process of the internal insulation material of the high-voltage bushing, providing high-precision multi-source monitoring data for subsequent accelerated aging tests.

[0055] The specific implementation of step S2 involves fixing the high-voltage bushing onto an accelerated aging test platform equipped with a high-voltage power supply and a temperature control box. The output voltage range of the high-voltage power supply is 0 to 100 kV, and the temperature control box has a temperature regulation range of room temperature to 150 degrees Celsius with temperature fluctuations controlled within ±2 degrees Celsius. The applied initial voltage stress is 1.5 to 2.0 times the rated voltage of the high-voltage bushing to accelerate the internal electric field distortion and electrical tree germination process of the insulation material. The applied initial temperature stress is 60 to 90 degrees Celsius to promote the thermo-oxidative aging and molecular chain breakage of the insulation material. The data acquisition system synchronously records the pulse sequence output by the partial discharge ultrasonic sensor array and the dielectric loss angle time sequence measured by the dielectric loss meter at a sampling frequency of 10 MHz. This step significantly shortens the test cycle by applying voltage and temperature stresses higher than the normal operating level, allowing the insulation material to rapidly undergo the same aging mechanism as long-term operation under controllable conditions.

[0056] The specific implementation of wavelet packet decomposition and noise reduction in step S3 is as follows: the acquired partial discharge pulse sequence is decomposed into 5 layers of wavelet packets, and the db4 wavelet basis function is selected. After decomposition, wavelet coefficients of 32 frequency bands are obtained. A Hankel matrix with a dimension of 50 by 50 is constructed for the wavelet coefficients of each frequency band. Singular value decomposition is performed on the Hankel matrix, and the singular vectors corresponding to the first 8 to 12 singular values ​​are retained. The wavelet coefficients are reconstructed through the retained singular vectors, and inverse wavelet packet transform is performed to obtain the singular value reconstructed signal. This process utilizes the characteristic of wavelet packet decomposition to effectively separate the discharge signal and noise components of different frequency bands by simultaneously subdividing the signal in the time and frequency domains. Singular value decomposition extracts the main features of the partial discharge pulse sequence and suppresses random noise and electromagnetic interference through matrix rank reduction operation, so that weak partial discharge signals can be accurately extracted in strong noise environment and the time-frequency characteristics of the real discharge signal can be preserved.

[0057] The specific implementation of the symbol dynamics complexity algorithm in step S3 involves amplitude normalization of the singular value reconstruction signal. The normalization process subtracts the minimum amplitude from the original amplitude of each pulse and divides by the difference between the maximum and minimum amplitudes. A symbolization threshold of 0.5 is set, and pulses with a normalized amplitude greater than 0.5 are encoded as symbol 1, while pulses with a normalized amplitude less than or equal to 0.5 are encoded as symbol 0, forming a binary symbol string. The frequency of occurrence of all different symbol patterns of length 2 to 6 in the binary symbol string is counted. The Lempel-Ziv algorithm is used to scan the binary symbol string from left to right and... When a new symbol pattern appears, the complexity counter is incremented by 1. The final discharge sequence complexity index is the ratio of the complexity counter value to the length of the binary symbol string. This algorithm converts continuous discharge amplitude signals into discrete binary symbol strings and quantifies the randomness of the discharge process by statistically analyzing the diversity of symbol patterns. In the early stage of aging, the discharge is dominated by fixed defects and exhibits a periodic symbol pattern, so the complexity index is low. When approaching breakdown, the randomness of multi-source discharges is enhanced by their superposition, so the complexity index increases significantly. This dynamic change law reflects the evolution stage of the microstructure of the insulating material and provides sensitive characteristic parameters for identifying the aging state.

[0058] The specific implementation of the independent component analysis dimensionality reduction process in step S3 involves centering the acquired dielectric loss angle time series, i.e., subtracting the mean of the entire time series from the dielectric loss angle value at each time step. The centered data is then subjected to a whitening transformation. The correlation between data at different times is eliminated through eigenvalue decomposition of the covariance matrix. The FastICA algorithm is used to iteratively optimize the separation matrix to maximize the non-Gaussianity of the separated signals, and the non-Gaussianity is measured using a negative entropy approximation function. Three independent components are set, corresponding to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively. The component loading matrix is ​​calculated to characterize the contribution weight of each time step in the dielectric loss angle time series to the three independent components. The three independent components with the largest absolute values ​​of the loading in the component loading matrix are extracted as aging feature components. This method, based on the statistical independence assumption of the source signals, decomposes the aliased signal formed by the combined action of multiple aging mechanisms into mutually uncorrelated independent sources. By maximizing the non-Gaussianity criterion, the independent contribution of each aging mechanism is separated, revealing the activity level of each aging mechanism at different measurement times. This achieves a projection transformation from a high-dimensional time series space to a low-dimensional intrinsic aging space and eliminates redundant information.

[0059] The specific implementation of step S4 involves constructing an ultra-sparse temporal coupling identification model. The input layer receives four parameters: the discharge sequence complexity index and three aging feature components. The first hidden layer contains 64 neurons and uses the hyperbolic tangent function as the activation function. The second hidden layer contains 32 neurons and introduces an ultra-sparse temporal coupling mechanism. This mechanism controls the information transmission between neurons through a dynamic connection matrix. The sparsity of the dynamic connection matrix is ​​determined by the discharge sequence complexity index. The calculation method is 0.95 minus 0.3 multiplied by the difference between the discharge sequence complexity index and the minimum value in the training set, divided by the difference between the maximum and minimum values ​​in the training set. The sparsity value ranges from 0.65 to 0.9. 5. In the second hidden layer, each neuron establishes a connection only with neurons in the first hidden layer whose sparsity ratio is equal to that of the neurons in the first hidden layer. The connection weights are modulated by a temporal coupling strength function, which is the result of multiplying the base weights by 1 and adding 0.2 times the value of a sine function. The independent variable of the sine function is twice pi multiplied by the coupling frequency multiplied by the time interval and divided by the normalized time constant. The coupling frequency is determined by the rate of change of the dielectric loss angle time series and ranges from 0.01 to 0.1 Hz. The normalized time constant is 100 seconds. The third hidden layer contains 16 neurons and uses the ReLU function as the activation function. The output layer contains 2 neurons that output the predicted value of the electrical tree growth rate and the shape parameter of the Weibull distribution, respectively. The model training dataset is established by collecting 200 to 500 sets of accelerated aging test data under different stress conditions. The discharge sequence complexity index and aging feature components of each set of data are extracted as input samples. The actual electrical tree growth rate is measured by microscopic section observation and electrical tree imaging technology, and the breakdown time is recorded as the output label. The dataset is divided into training subset, validation subset, and test subset in a ratio of 7:2:1. Model training employs the Adam optimizer for parameter updates, with an initial learning rate of 0.001 that decays to 0.95 times its original value every 50 epochs. The loss function is a weighted sum of mean squared error and physical constraint terms, which penalize predictions that violate the monotonicity of electrical tree growth. During training, model performance is evaluated on a validation subset every 10 epochs. If the loss on the validation subset does not decrease for 20 consecutive epochs, training is stopped early, and the model parameters with the minimum validation subset loss are selected as the final training result. This ultra-sparse temporal coupling mechanism dynamically adjusts the sparsity of the model's internal connections, enabling the model to adaptively select information transmission paths based on discharge complexity. The temporal coupling strength function introduces a periodic modulation term to simulate the polarization relaxation process of insulating materials under alternating electric fields, embedding the temporal evolution of the physical process into the model architecture. This ensures that predictions not only rely on statistical data characteristics but also follow the physical mechanism of electrical tree growth.

[0060] The specific implementation of step S5 involves generating 1000 to 5000 electrical tree growth path samples using the Monte Carlo random simulation method, establishing a three-dimensional coordinate system with the origin set at the initial defect location of the electrical tree, setting the electrical tree growth step size to 0.1 to 0.5 micrometers, and determining the growth direction of each step by a combination of the local electric field intensity gradient and random perturbation. The local electric field intensity gradient is calculated using finite element simulation with a mesh size of 1 micrometer, and the random perturbation follows a normal distribution with its standard deviation determined by the fractal dimension by multiplying by 0.1. The electric field strength at the current position is calculated in each growth step by subtracting 1.5 from the fractal dimension. If the electric field strength exceeds the material breakdown field strength threshold, breakdown is determined. The breakdown field strength threshold for cross-linked polyethylene is 28 to 32 kV / mm, and for silicone rubber, it is 25 to 30 kV / mm. The current growth step number is recorded and multiplied by the electrical tree growth step size to obtain the electrical tree length. The electrical tree length of each electrical tree growth path sample is divided by the predicted electrical tree growth rate output in step S4 to convert it into breakdown time. The fractal dimension is calculated using the box counting method. The electrical tree growth path samples are projected onto a two-dimensional plane and covered with squares of different sizes. The logarithmic relationship between the number of squares required for coverage and the square size is calculated. The slope of this logarithmic relationship is the fractal dimension, ranging from 1.6 to 1.9. This method overcomes the strong randomness of electric tree growth caused by microscopic defects and electric field fluctuations by using a large number of random samples. It quantifies the spatial complexity of the electric tree growth path by fractal dimension and determines the standard deviation of random perturbation based on fractal dimension so that the perturbation conforms to the statistical law of actual electric tree growth.

[0061] The specific implementation of step S6 involves sorting the breakdown times obtained in step S5 from smallest to largest, and using the maximum likelihood estimation method to calculate the shape and scale parameters of the Weibull distribution. The shape parameter reflects the trend of failure rate over time, while the scale parameter characterizes the characteristic lifetime. The acceleration factor is calculated using the fitted shape and scale parameters. The acceleration factor is the ratio of the characteristic lifetime under accelerated testing conditions to the characteristic lifetime under normal operating conditions. The accelerated testing conditions refer to the initial voltage stress and initial temperature stress applied in step S2, and the normal operating conditions refer to the working state of the high-voltage bushing at rated voltage and ambient temperature of 15 to 35 degrees Celsius. The uncertainty of the shape and scale parameter estimates is evaluated using the Bootstrap resampling method, with 500 to 1000 resampling cycles, and the 95% confidence interval of the acceleration factor is calculated. If the ratio of the acceleration factor confidence interval width to the acceleration factor is greater than 0.3, the voltage stress and temperature stress are recalculated using an adaptive stress control function. The calculation process of the adaptive stress regulation function first extracts the acceleration factor and acceleration factor confidence interval obtained by fitting the Weibull distribution. The ratio of the acceleration factor confidence interval width to the acceleration factor is calculated as an uncertainty index. The ratio of the initial voltage stress to the rated voltage is calculated as the voltage ratio. The stress regulation coefficient is calculated based on the uncertainty index and the voltage ratio. The stress regulation coefficient is equal to 1 plus 0.5 multiplied by the uncertainty index minus 0.3, divided by 0.7, multiplied by 2.5, minus the voltage ratio, and divided by 0.5. When the uncertainty index is in the range of 0.3 to 0.5, the new voltage stress is the initial voltage stress multiplied by the stress adjustment coefficient, while the new temperature stress remains unchanged from the initial temperature stress. When the uncertainty index is in the range of 0.5 to 0.7, the new voltage stress is the initial voltage stress multiplied by the stress adjustment coefficient, while the new temperature stress is the initial temperature stress plus 5 to 10 degrees Celsius. When the uncertainty index is in the range of 0.7 to 1.0, the new voltage stress is the initial voltage stress multiplied by 1.2 times the stress adjustment coefficient, while the new temperature stress is the initial temperature stress plus 10 to 15 degrees Celsius. When the uncertainty index is greater than 1.0, the new voltage stress is 2.5 to 3.0 times the rated voltage, while the new temperature stress is 90 to 100 degrees Celsius.

[0062] The specific implementation of step S7 involves establishing the posterior distribution of the acceleration factor using a Bayesian inference framework. The prior distribution of the acceleration factor calculated in step S6 is set to a log-normal distribution, with the log-normal distribution parameters determined based on historical experimental data and engineering experience. A likelihood function is constructed based on the Weibull distribution fitting results and experimental observation data from step S6. The posterior distribution of the acceleration factor is calculated using the Bayesian formula, and the posterior distribution is proportional to the product of the prior distribution and the likelihood function. A Markov chain Monte Carlo sampling method is used to generate posterior samples of the acceleration factor from the posterior distribution. The sampling algorithm is the Metropolis-Hastings algorithm, with a sampling chain length of 10,000 to 50,000. The first 1,000 to 5,000 posterior samples are discarded as pre-burning samples. The long-term lifetime prediction value is calculated based on the posterior samples. The long-term lifetime prediction value is the characteristic lifetime under normal operating conditions obtained by dividing each sample value of the posterior sample by the characteristic lifetime under accelerated testing conditions. The 2.5% quantile and 97.5% quantile of the posterior samples are used as the 95% confidence interval of the long-term lifetime prediction value. The Markov chain Monte Carlo sampling method constructs a Markov chain whose stationary distribution converges to the posterior distribution of the acceleration factor, thus achieving numerical sampling of complex probability distributions. The Metropolis-Hastings algorithm generates candidate samples in each iteration and decides whether to accept a candidate sample based on the acceptance probability determined by the ratio of the posterior probability density. This method can sample from non-standard distributions without calculating a normalization constant. Discarding samples during the pre-burning period ensures that the sampling chain has fully converged to a stationary distribution. The posterior sample of the acceleration factor fully characterizes the uncertainty of the acceleration factor and propagates the statistical error to the long-term lifetime prediction value. The 95% confidence interval of the long-term lifetime prediction value quantitatively assesses the credibility of the extrapolation results, providing a risk assessment basis for high-voltage bushing operation and maintenance decisions.

[0063] It should be noted that the key technical ideas of this invention include the following aspects. First, the ultra-sparse temporal coupling identification model embeds the physical aging mechanism of insulating materials into a neural network architecture through dynamic sparse connections and temporal coupling strength functions. This enables the model to adaptively adjust the information transmission path according to the discharge complexity and simulate the polarization relaxation process under the action of alternating electric fields. Compared with traditional black-box neural network models, this significantly improves the generalization ability and extrapolation accuracy when accelerated stress conditions change. At the same time, sparse connections reduce computational complexity and overfitting risk, allowing the model to accurately identify the growth law of electrical trees under limited experimental data. Second, the symbolic dynamics complexity algorithm combined with independent component analysis (ICA) for dimensionality reduction extracts aging-sensitive features from partial discharge pulse sequences and dielectric loss angle time series. Symbolic dynamics quantifies the random evolution process of discharge through statistical symbol pattern diversity to capture the characteristics of electrical tree growth stages. ICA separates the aliased multi-mechanism aging signals into independent contributions such as hydrolysis, pyrolysis, and oxidation, eliminating redundant information. Compared with traditional statistical feature extraction methods, this more accurately characterizes the microstructure evolution of insulating materials and improves the accuracy and sensitivity of aging state identification. Third, the Bayesian inference framework, combined with Markov chain Monte Carlo sampling, establishes a posterior distribution of the acceleration factor to achieve comprehensive quantification of uncertainty. This method constructs a posterior distribution by fusing prior knowledge and experimental observation data, and generates a large number of posterior samples through Monte Carlo sampling, transferring the statistical error of the acceleration factor to the long-term lifetime prediction value. Compared with traditional point estimation methods, it provides a confidence interval to quantitatively assess the credibility of the extrapolation results, effectively solving the problem that statistical errors are amplified when extrapolating long-term lifetimes from short-term experimental data, making it difficult to assess confidence. The synergistic effect of these three technical approaches lies in the ultra-sparse temporal coupled identification model, which achieves a deep fusion of physical mechanisms and data-driven approaches based on high-quality feature parameters extracted by symbolic dynamics and independent component analysis, accurately predicting the electric tree growth rate and Weibull distribution parameters. The Bayesian inference framework quantifies the uncertainty of the identification model's output and optimizes experimental conditions through adaptive stress control. The three form a complete closed loop from feature extraction to mechanism identification to uncertainty quantification, significantly improving the prediction accuracy and reliability of accelerated lifetime experiments. Compared with existing technologies, it has significant advantages in experimental efficiency and lifetime prediction confidence.

[0064] It should be noted that this invention also solves the following technical problem: the inaccurate extraction of discharge features in existing accelerated electrical lifetime tests due to strong noise interference with partial discharge signals. This invention effectively separates the discharge signal and noise components of different frequency bands by simultaneously subdividing the partial discharge pulse sequence in both the time and frequency domains through wavelet packet decomposition. A Hankel matrix is ​​constructed for the wavelet coefficients of each frequency band, and singular value decomposition is performed to retain the singular vectors corresponding to the first eight to twelve singular values. The main features of the partial discharge pulse sequence are extracted through matrix rank reduction, suppressing the influence of random noise and electromagnetic interference. The singular value reconstructed signal retains the time-frequency characteristics of the true discharge signal, providing high signal-to-noise ratio input data for subsequent symbolic dynamics complexity algorithms. This invention also solves the technical problem that existing experimental methods struggle to accurately identify the contribution of each of the multiple aging mechanisms in insulating materials. This invention uses independent component analysis (ICA) for dimensionality reduction. Based on the statistical independence assumption of source signals, it decomposes the dielectric loss angle time series into uncorrelated independent sources. It employs a fast ICA algorithm to iteratively optimize the separation matrix to maximize the non-Gaussianity of the separated signals. The number of independent components is set to three, corresponding to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively. The component loading matrix is ​​calculated to reveal the activity level of each aging mechanism at different measurement times. This achieves a projection transformation from a high-dimensional time series space to a low-dimensional intrinsic aging space. Compared with the principal component analysis method, it can separate nonlinear aliased aging mechanism signals, which is more consistent with the physical nature of the multiple aging processes of insulating materials.

[0065] Specifically, the principle of this invention is: the fundamental principle that can solve the problem of statistical error amplification of acceleration factors lies in combining physical-driven modeling with probabilistic inference to achieve effective transmission and quantification of uncertainty. The ultra-sparse temporal coupling identification model, through adaptive adjustment of the sparsity of the dynamic connection matrix and periodic modulation of the temporal coupling strength function, correlates the evolution of discharge sequence complexity and the rate of change of dielectric loss angle over time with the physical process of electrical tree growth. This enables the model to accurately identify the growth rate of electrical trees and output the shape parameters of the Weibull distribution under limited experimental data. Monte Carlo stochastic simulation, based on this physical driving prediction result, maps the randomness of the microscopic growth mechanism to the probability distribution of macroscopic breakdown time through extensive path sampling. Weibull distribution fitting extracts the acceleration factor and its confidence interval from this probability distribution. The Bayesian inference framework treats the acceleration factor as a random variable rather than a deterministic value. It integrates historical experience information through prior distribution, integrates current experimental observation data through likelihood function, and achieves optimal fusion of the two types of information through posterior distribution calculation. The Markov chain Monte Carlo sampling method efficiently extracts samples from the complex posterior distribution, completely preserving the statistical error information of the acceleration factor. This error is transferred to the long-term lifetime prediction value through division and the confidence interval is calculated, realizing a full-chain quantitative evaluation from experimental data uncertainty to the confidence of extrapolation results.

[0066] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S1 and S2 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0067] The specific implementation of wavelet packet decomposition and noise reduction in step S3 is as follows: the partial discharge pulse sequence acquired in step S2 is decomposed into 5 layers of wavelet packets, the db4 wavelet basis function is selected, and the number of frequency bands after decomposition is 32. A Hankel matrix is ​​constructed for the wavelet coefficients of each frequency band. The dimension of the Hankel matrix is ​​50×50. Singular value decomposition is performed on the Hankel matrix, and the singular vectors corresponding to the first 8 to 12 singular values ​​are retained. The wavelet coefficients are reconstructed through the retained singular vectors, and the inverse wavelet packet transform is used to obtain the singular value reconstructed signal.

[0068] The specific implementation of the symbolic dynamics complexity algorithm in step S3 is to perform amplitude normalization processing on the singular value reconstruction signal obtained in step S3. The formula for amplitude normalization processing is as follows: ; In the formula, Let be the normalized amplitude of the i-th pulse, where i ranges from 1 to N, and N is the total number of pulses in the singular value reconstructed signal. The original amplitude of the i-th pulse is in mV. The maximum amplitude in the singular value reconstructed signal is expressed in mV. The minimum amplitude unit in the singular value reconstructed signal is mV. A symbolization threshold of 0.5 is set. Pulses with a normalized amplitude greater than 0.5 are encoded as symbol 1, and pulses with a normalized amplitude less than or equal to 0.5 are encoded as symbol 0, forming a binary symbol string. The frequency of occurrence of all different symbol patterns of length 2 to 6 in the binary symbol string is counted. The complexity of the binary symbol string is calculated using the Lempel-Ziv algorithm. The Lempel-Ziv algorithm scans the binary symbol string from left to right, incrementing the complexity counter by 1 each time a new symbol pattern appears. The final discharge sequence complexity index is the ratio of the complexity counter value to the length of the binary symbol string.

[0069] The specific implementation of the independent component analysis dimensionality reduction process in step S3 is as follows: the dielectric loss angle time series collected in step S2 is centered by subtracting the mean of the dielectric loss angle time series from the dielectric loss angle value at each time step. The centered dielectric loss angle time series is then subjected to a whitening transformation. This whitening transformation eliminates the correlation between data at different times through eigenvalue decomposition of the covariance matrix. The separation matrix is ​​iteratively optimized using the FastICA algorithm. The iteration process of the FastICA algorithm maximizes the non-Gaussianity of the separated signal. The non-Gaussianity is measured by the negative entropy approximation function. The number of independent components is set to 3, which correspond to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively. The component loading matrix is ​​calculated, which represents the contribution weight of the data at each time step in the dielectric loss angle time series to the 3 independent components. The 3 independent components with the largest absolute value of the loading in the component loading matrix are extracted as aging feature components.

[0070] The specific implementation of the ultra-sparse temporal coupling identification model in step S4 is as follows: the input layer receives four parameters extracted in step S3, namely the discharge sequence complexity index and aging feature components; the first hidden layer contains 64 neurons, and the activation function of the first hidden layer is the hyperbolic tangent function; the second hidden layer contains 32 neurons, and the sparsity calculation formula of the dynamic connection matrix of the second hidden layer is expressed as follows: ; In the formula, s is the sparsity of the dynamic connection matrix, ranging from 0.65 to 0.95, and C is the discharge sequence complexity index extracted in step S3. To maximize the complexity exponent of the training set of discharge sequences. To minimize the exponential complexity of the training-focused firing sequence, each neuron in the second hidden layer establishes a connection only with neurons in the first hidden layer whose sparsity proportions correspond to the dynamic connection matrix. The weights of these connections are modulated by a temporal coupling strength function, which is described as follows: ; In the formula, Let be the coupling weight between the i-th neuron and the j-th neuron at time t, where i ranges from 1 to 32 and j ranges from 1 to 64, corresponding to the neuron numbers in the second and first hidden layers, respectively. Based on the weights, The coupling frequency is determined by the rate of change of the dielectric loss angle time series acquired in step S2, ranging from 0.01 to 0.1 Hz. T is the normalized time constant with a value of 100 s, and t is the current time in seconds. The third hidden layer contains 16 neurons, and the activation function of the third hidden layer is the ReLU function. The output layer contains 2 neurons, and the 2 neurons of the output layer output the predicted value of the electric tree growth rate in μm / h and the shape parameter of the Weibull distribution, respectively.

[0071] The specific implementation of generating electrical tree growth path samples using the Monte Carlo random simulation method in step S5 is as follows: A three-dimensional coordinate system is established, with its origin set at the initial defect location of the electrical tree. The growth step size is set to 0.1 to 0.5 μm. The growth direction at each step is determined by the local electric field intensity gradient and random perturbation. The random perturbation follows a normal distribution, and the formula for calculating the standard deviation of the normal distribution is as follows: ; In the formula, The standard deviation of a normal distribution is expressed in rad. The fractal dimension is 0.1, expressed in rad, to ensure consistency between the rad units on the right and left sides of the equation. The fractal dimension ranges from 1.6 to 1.9 and is calculated using box counting, with the formula as follows: ; In the formula, For fractal dimension, The required side length for covering the electric tree growth path sample is The number of squares, The grid scale unit is μm. In each growth step, the electric field strength at the current position is calculated. If the electric field strength exceeds the material breakdown field strength threshold of 25 to 35 kV / mm, it is determined to be a breakdown. The current growth step number is recorded and multiplied by the electric tree growth step size to obtain the electric tree length. The above process is repeated to generate 1000 to 5000 electric tree growth path samples. The electric tree length of each electric tree growth path sample is divided by the predicted electric tree growth rate value output in step S4 and converted into a breakdown time unit of h.

[0072] The specific implementation of the Weibull distribution fitting in step S6 is as follows: The breakdown times obtained in step S5 using the Monte Carlo random simulation method are sorted from smallest to largest. The shape and scale parameters of the Weibull distribution are calculated using the maximum likelihood estimation method. The shape parameter reflects the failure rate over time, and the scale parameter characterizes the characteristic lifetime. An acceleration factor is calculated using the fitted shape and scale parameters of the Weibull distribution. The acceleration factor is the ratio of the characteristic lifetime under accelerated testing conditions to the characteristic lifetime under normal operating conditions. The uncertainty of the estimated shape and scale parameters of the Weibull distribution is evaluated using the Bootstrap resampling method. The number of resampling iterations in the Bootstrap resampling method is between 500 and 1000. The 95% confidence interval of the acceleration factor is calculated as the acceleration factor confidence interval. The formula for calculating the width of the acceleration factor confidence interval is as follows: ; In the formula, To determine the confidence interval width of the acceleration factor, This represents the upper limit of the confidence interval for the acceleration factor. This represents the lower limit of the confidence interval for the acceleration factor.

[0073] The specific implementation of the adaptive stress control function in step S6 is as follows: The acceleration factor and its confidence interval obtained from the Weibull distribution fitting in step S6 are extracted. The ratio of the confidence interval width to the acceleration factor is calculated and denoted as the uncertainty index. The formula for calculating the uncertainty index is as follows: ; In the formula, As an indicator of uncertainty, To determine the confidence interval width of the acceleration factor, This represents the average value of the acceleration factor. The ratio of the initial voltage stress applied in calculation step S2 to the rated voltage is denoted as the voltage multiplier, and the formula for calculating the voltage multiplier is as follows: ; In the formula, This is the voltage multiplier. The initial voltage stress applied in step S2 is in kV. The rated voltage is in kV. The formula for calculating the stress adjustment coefficient is as follows: ; In the formula, This is the stress adjustment coefficient. As an indicator of uncertainty, This refers to the voltage multiplier. When the uncertainty index... When ∈ [0.3, 0.5), the new voltage stress is the initial voltage stress applied in step S2 multiplied by the stress adjustment coefficient, and the new temperature stress remains unchanged at the initial temperature stress applied in step S2. When the uncertainty index... When ∈ [0.5, 0.7), the new voltage stress is the initial voltage stress applied in step S2 multiplied by the stress adjustment coefficient, and the new temperature stress is the initial temperature stress applied in step S2 plus 5 to 10°C. When the uncertainty index... When the voltage is ∈ [0.7, 1.0], the new voltage stress is the initial voltage stress applied in step S2 multiplied by 1.2 times the stress adjustment coefficient, and the new temperature stress is the initial temperature stress applied in step S2 plus 10 to 15°C. When the uncertainty index... When the voltage is greater than 1.0, the new voltage stress is 2.5 to 3.0 times the rated voltage, and the new temperature stress is 90 to 100℃.

[0074] The specific implementation of establishing the posterior distribution of the acceleration factor using the Bayesian inference framework in step S7 is as follows: The prior distribution of the acceleration factor calculated in step S6 is set to a log-normal distribution. The parameters of the log-normal distribution are determined based on historical experimental data and engineering experience. A likelihood function is constructed. Based on the Weibull distribution fitting results and experimental observation data in step S6, the posterior distribution of the acceleration factor is calculated using the Bayesian formula. The posterior distribution of the acceleration factor is proportional to the product of the log-normal distribution and the likelihood function. A Markov chain Monte Carlo sampling method is used to generate posterior samples of the acceleration factor from the posterior distribution. The sampling algorithm used in this method is the Metropolis-Hastings algorithm. The sampling chain length of the Markov chain Monte Carlo sampling method is 10,000 to 50,000. The first 1,000 to 5,000 acceleration factor posterior samples are discarded as pre-burn-in samples. The long-term lifetime prediction value is calculated based on the acceleration factor posterior samples. The long-term lifetime prediction value is the characteristic lifetime under normal operating conditions, obtained by dividing each sample value of the acceleration factor posterior samples by the characteristic lifetime under accelerated test conditions. The 2.5% quantile and 97.5% quantile of the acceleration factor posterior samples are used as the 95% confidence interval of the long-term lifetime prediction value.

[0075] The formula for calculating the coupling frequency is expressed as follows: ; In the formula, The coupling frequency is in Hz. The difference between adjacent sampling points in the time series of dielectric loss angle collected in step S2. The time interval between adjacent sampling points is in seconds (s), 0.01 is in Hz, and 0.09 is in [missing unit]. This makes the unit on the right side of the equation, Hz, consistent with the unit on the left side.

[0076] It should be noted that the variables involved in this invention are explained in detail in Table 1.

[0077] Table 1. Variable Explanation Table

[0078] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To address the problem of excessively long reliability assessment cycles for high-voltage bushings, technicians used the method of this invention to conduct accelerated electrical life tests on a batch of 110 kV cross-linked polyethylene high-voltage bushings. The test object was an oil-paper insulated composite bushing with a rated voltage of 110 kV and a rated current of 1200 amps. This batch of bushings was planned for use in the high-voltage side connection of the main transformer in a substation, with an operating ambient temperature range of -20 to 40 degrees Celsius. Technicians first uniformly arranged 12 partial discharge ultrasonic sensor units in a circumferential array on the outer surface of the high-voltage bushing. The frequency response range of the sensor units was 20 to 500 kHz, and the spacing between adjacent sensor units was 75 mm. Simultaneously, a dielectric loss meter was installed to monitor changes in the dielectric loss angle. The measurement accuracy of the dielectric loss meter was [missing information]. class.

[0079] Technicians installed the high-voltage bushing on an accelerated aging test platform equipped with a high-voltage power supply with an output voltage range of 0 to 100 kV and a temperature control box with a temperature regulation range of room temperature to 150 degrees Celsius. Based on the bushing's rated voltage of 110 kV, the technicians set the initial voltage stress to 198 kV (1.8 times the rated voltage) and the initial temperature stress to 75 degrees Celsius. The data acquisition system synchronously recorded the partial discharge pulse sequence and dielectric loss angle time series at a sampling frequency of 10 MHz for a continuous test duration of 480 hours. Figure 2 As shown, the amplitude of the partial discharge pulse was small and the distribution was relatively regular in the early stage of the experiment. As the test time increased, the pulse amplitude gradually increased and the frequency of occurrence increased.

[0080] Technicians performed wavelet packet decomposition denoising on the acquired partial discharge pulse sequence, selecting the db4 wavelet basis function for a 5-level decomposition to obtain wavelet coefficients for 32 frequency bands. A 50x50 Hankel matrix was constructed for each frequency band's wavelet coefficients, and singular value decomposition was performed. The singular vectors corresponding to the first 10 singular values ​​were retained for signal reconstruction. The reconstructed signal underwent amplitude normalization, and a symbolization threshold of 0.5 was set to convert the normalized amplitude into a binary symbol string. The Lempel-Ziv algorithm was used to calculate the discharge sequence complexity index. Experimental data showed that in the initial 0-100 hours of the experiment, the discharge sequence complexity index remained in the range of 0.15 to 0.22. In the middle stage (100-300 hours), the complexity index gradually increased to the range of 0.35 to 0.48. Near the breakdown point (300-480 hours), the complexity index rapidly climbed to the range of 0.68 to 0.82. Figure 3 As shown, the dielectric loss angle time series was 0.0028 to 0.0035 in the early stage of the test, and continued to increase to 0.0089 to 0.0124 as the aging process progressed.

[0081] Technicians performed independent component analysis (ICA) to reduce the dimensionality of the dielectric loss angle time series. After centering and whitening transformations, the FastICA algorithm was used to extract three independent components, corresponding to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism, respectively. Component loading matrix analysis showed that in the early stage of the experiment, the contribution of the hydrolysis aging mechanism was 0.52, the contribution of the pyrolysis aging mechanism was 0.31, and the contribution of the oxidation aging mechanism was 0.17. In the middle and later stages of the experiment, the contribution of the pyrolysis aging mechanism increased to 0.45, while the contribution of the oxidation aging mechanism increased to 0.38, reflecting the dynamic evolution of different aging mechanisms with temperature and electric field stress. Technicians input the discharge sequence complexity index and the three aging feature components into a pre-trained ultra-sparse temporal coupling identification model. The model input layer received four parameters. The first hidden layer contained 64 neurons using the hyperbolic tangent activation function, and the second hidden layer contained 32 neurons and introduced an ultra-sparse temporal coupling mechanism. The sparsity of the dynamic connection matrix was adaptively adjusted according to the discharge sequence complexity index, with a sparsity of 0.88 in the early stage of the experiment and decreasing to 0.71 in the later stage. The coupling frequency of the time-domain coupling strength function is determined by the rate of change of the dielectric loss angle over time; the calculated coupling frequency is 0.068 Hz. For example... Figure 4 As shown, as the test progresses, the discharge sequence complexity index and aging time exhibit a non-linear upward trend, and this change pattern is highly consistent with the physical process of electrical tree growth.

[0082] The predicted growth rate of electrical trees from the ultra-sparse temporal-domain coupled identification model is 2.34 micrometers per hour, with a Weibull distribution shape parameter of 3.87. Technicians used Monte Carlo random simulation to generate 3000 electrical tree growth path samples, established a three-dimensional coordinate system with the origin set at the initial defect location of the electrical tree, and set the growth step size to 0.3 micrometers. In each growth step, the growth direction is determined by both the local electric field intensity gradient and random perturbation. The local electric field intensity gradient was calculated using finite element simulation with a mesh size of 1 micrometer, and the random perturbation followed a normal distribution with its standard deviation determined by the fractal dimension. The fractal dimension calculated using the box counting method was 1.76, and the standard deviation of the random perturbation was determined to be 0.026 radians. For cross-linked polyethylene material, the breakdown field strength threshold was set to 30 kV / mm; when the electric field strength at a certain location during electrical tree growth exceeded this threshold, breakdown was considered complete. In the experiment, 3,000 samples of electric tree growth paths exhibited a typical dendritic fractal structure, with the main path extending along the direction of the maximum electric field intensity gradient, while the branch paths showed a certain degree of divergence due to random perturbations.

[0083] Technicians analyzed the breakdown time distribution of 3000 samples, ranging from 178 to 562 hours. The breakdown time was fitted to a Weibull distribution using the maximum likelihood estimation method, yielding a shape parameter of 3.92 and a scale parameter of 385 hours. When calculating the acceleration factor, normal operating conditions were set at a rated voltage of 110 kV and an ambient temperature of 25°C, while accelerated testing conditions were set at a voltage stress of 198 kV and a temperature stress of 75°C, resulting in an acceleration factor of 47.6. The uncertainty of parameter estimation was assessed through 800 resampling operations using the Bootstrap resampling method, yielding a 95% confidence interval for the acceleration factor ranging from 41.3 to 54.8. The confidence interval width of the acceleration factor was 13.5, and the ratio of the acceleration factor to the acceleration factor itself was 0.28, which is less than the threshold of 0.3, satisfying the conditions for continuing lifetime prediction.

[0084] Technicians used a Bayesian inference framework to establish the posterior distribution of the acceleration factor, setting the prior distribution to a log-normal distribution, with parameters determined based on historical test data of similar casings. A likelihood function was constructed based on the Weibull distribution fitting results and experimentally observed breakdown time data, and the posterior distribution was calculated using Bayes' theorem. The Metropolis-Hastings algorithm was used for Markov chain Monte Carlo sampling, with a sampling chain length of 30,000. The first 3,000 samples were discarded as pre-burn-in samples, ultimately obtaining 27,000 posterior samples of the acceleration factor. The long-term lifetime prediction calculated based on the posterior samples was 18,300 hours, corresponding to approximately 2.1 years of service life under normal operating conditions. The 2.5% and 97.5% quantiles of the posterior samples were statistically analyzed, yielding a 95% confidence interval of 15,900 to 21,200 hours for the long-term lifetime prediction. Figure 5 As shown, the probability distribution of long-term life predictions exhibits a right-skewed characteristic, and the relatively moderate confidence interval width reflects the reliability of the prediction results.

[0085] The advancements of this invention compared to traditional accelerated life testing methods are mainly reflected in the following aspects. Traditional methods typically employ fixed stress levels and extrapolate long-term lifetimes based on simple statistical models. This fails to adequately consider the dynamic evolution of discharge behavior and electrical parameters during the aging process of insulating materials, leading to significant uncertainty in prediction results and a lack of quantitative assessment of extrapolation reliability. This invention, through symbolic dynamics complexity algorithms and independent component analysis, extracts sensitive features reflecting microscopic aging mechanisms from multi-source monitoring data, overcoming the limitation of traditional methods that rely solely on macroscopic electrical parameters and cannot accurately characterize insulation degradation stages. The ultra-sparse temporal coupling identification model embeds the temporal evolution of physical processes into a neural network architecture. By simulating the polarization relaxation process of insulating materials under alternating electric fields through dynamic sparse connections and temporal coupling strength functions, the model prediction not only relies on statistical data characteristics but also follows the physical mechanism of electrical tree growth, significantly improving its generalization ability to different stress conditions. The Monte Carlo stochastic simulation method combines fractal dimension to quantify the spatial complexity of electric tree growth and obtains the probability distribution of breakdown time through extensive random sampling, overcoming the limitation of deterministic models in describing the strong randomness caused by microscopic defects and electric field fluctuations. A Bayesian inference framework integrates prior knowledge and experimental observation data to construct the posterior distribution of the acceleration factor. Markov chain Monte Carlo sampling achieves comprehensive quantification of uncertainty, transferring the statistical error of the acceleration factor to the long-term lifetime prediction and providing a confidence interval to quantitatively assess the reliability of the extrapolation results. This effectively solves the problem of statistical error amplification leading to difficulty in assessing confidence levels when extrapolating long-term lifetimes from short-term experimental data using traditional point estimation methods. An adaptive stress control mechanism dynamically optimizes experimental conditions based on the width of the acceleration factor confidence interval, avoiding the risks of low experimental efficiency or altered failure mechanisms caused by excessive stress levels in traditional methods. This ensures that accelerated testing maximizes experimental efficiency while maintaining the same aging mechanism as actual operation.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An accelerated electrical life test method for the long-term operational reliability of high-voltage bushings, characterized in that, A partial discharge ultrasonic sensor array and a dielectric loss meter are deployed on the surface of the high-voltage bushing to simultaneously acquire partial discharge pulse sequences and dielectric loss angle time series. Wavelet packet decomposition is used to denoise the partial discharge pulse sequences to extract singular values ​​and reconstruct signals. The discharge sequence complexity index is calculated using a symbolic dynamics complexity algorithm. Independent component analysis is used to reduce the dimensionality of the dielectric loss angle time series to extract aging characteristic components. The ultra-sparse temporal coupling identification model is input to output the predicted value of electrical tree growth rate and the shape parameters of Weibull distribution. The Monte Carlo random simulation method is used to generate electrical tree growth path samples and calculate the breakdown time. The Weibull distribution is used to fit the breakdown time to calculate the acceleration factor and the confidence interval of the acceleration factor. The voltage stress and temperature stress are optimized by an adaptive stress control function. A Bayesian inference framework is used to establish the posterior distribution of the acceleration factor, and the posterior samples of the acceleration factor are generated by the Markov chain Monte Carlo sampling method to calculate the confidence interval of the long-term lifetime prediction value.

2. The method according to claim 1, characterized in that, The wavelet packet decomposition denoising process specifically involves performing a five-level wavelet packet decomposition on the partial discharge pulse sequence, selecting the db4 wavelet basis function, constructing a Hankel matrix for the wavelet coefficients of each frequency band, and performing singular value decomposition to retain the singular vectors corresponding to the first eight to twelve singular values. The wavelet coefficients are then reconstructed using the retained singular vectors, and the inverse wavelet packet transform is performed to obtain the singular value reconstructed signal.

3. The method according to claim 2, characterized in that, The symbol dynamics complexity algorithm specifically involves normalizing the amplitude of the singular value reconstructed signal, setting a symbolization threshold, encoding pulses with normalized amplitudes greater than the threshold as symbol 1, and encoding pulses with normalized amplitudes less than or equal to the threshold as symbol 0 to form a binary symbol string, and using the Lempel-Ziv algorithm to calculate the complexity of the binary symbol string.

4. The method according to claim 3, characterized in that, The independent component analysis dimensionality reduction process specifically involves centering and whitening the dielectric loss angle time series, using a fast independent component analysis algorithm to iteratively optimize the separation matrix to maximize the non-Gaussianity of the separated signals, setting the number of independent components to three to correspond to the hydrolysis aging mechanism, the pyrolysis aging mechanism, and the oxidation aging mechanism respectively, calculating the component loading matrix, and extracting the three independent components with the largest absolute load values ​​as aging feature components.

5. The method according to claim 4, characterized in that, The ultra-sparse temporal coupling identification model specifically involves the input layer receiving the discharge sequence complexity index and aging feature components, the second hidden layer introducing an ultra-sparse temporal coupling mechanism to control the information transmission between neurons through a dynamic connection matrix, the sparsity of which is determined by the discharge sequence complexity index, the connection weights being modulated by a temporal coupling strength function, and the output layer outputting the predicted value of electrical tree growth rate and the shape parameters of the Weibull distribution.

6. The method according to claim 5, characterized in that, The sparsity calculation of the dynamic connection matrix is ​​specifically based on the normalized relationship between the discharge sequence complexity index and the maximum and minimum values ​​of the discharge sequence complexity index in the training set. The sparsity value ranges from 0.65 to 0.

95.

7. The method according to claim 6, characterized in that, The time-domain coupling strength function is specifically the product of the coupling weight and the periodic modulation term, and the coupling frequency is determined by the rate of change of the dielectric loss angle time series, ranging from 0.01 to 0.1 Hz.

8. The method according to claim 7, characterized in that, The Monte Carlo random simulation method generates electric tree growth path samples. Specifically, a three-dimensional coordinate system is established and the electric tree growth step size is set. The growth direction of each step is determined by the local electric field intensity gradient and random perturbation. The random perturbation follows a normal distribution, and the standard deviation of the normal distribution is determined by the fractal dimension. In each growth step, the electric field intensity at the current position is calculated. If the electric field intensity exceeds the material breakdown field strength threshold, it is determined to be a breakdown.

9. The method according to claim 8, characterized in that, The fractal dimension is specifically determined by projecting electrical tree growth path samples onto a two-dimensional plane using box counting, covering the electrical tree growth path samples with squares of different sizes, and statistically analyzing the logarithmic relationship between the number of squares required to cover the electrical tree growth path samples and the square size. The slope of this logarithmic relationship is the fractal dimension.

10. The method according to claim 9, characterized in that, The Weibull distribution fitting process involves sorting the breakdown times from smallest to largest, using the maximum likelihood estimation method to calculate the shape and scale parameters of the Weibull distribution, using the fitted shape and scale parameters to calculate the acceleration factor, and using the Bootstrap resampling method to evaluate the uncertainty of the estimated shape and scale parameters of the Weibull distribution and calculate the 95% confidence interval of the acceleration factor.

Images