Cable insulation degradation degree evaluation method based on frequency domain feature analysis

Abstract

The invention relates to a cable insulation degradation degree evaluation method based on frequency domain characteristic analysis, and belongs to the field of cable insulation degradation detection.The cable insulation degradation degree evaluation method comprises the steps that cable operation environment data and laboratory accelerated aging data are collected, and a low-frequency impedance characteristic data set with water tree growth stage labels is generated; extracting a dielectric response spectrogram and analyzing a relaxation polarization peak value; when the peak value exceeds a threshold value, extracting harmonic component characteristics and loss factor distribution characteristics by using wavelet transform and principal component analysis; constructing a support vector regression model to predict the length, density and growth stage of the water tree; for middle and later water trees, microstructure evolution characteristic parameters are calculated, and insulation penetration time is predicted; the degradation probability is analyzed through Monte Carlo simulation, and a risk distribution curve is generated; and optimizing the cable replacement time by adopting dynamic planning, and outputting an evaluation report for quantifying the degradation degree and maintenance suggestions. According to the invention, accurate evaluation and prediction of cable insulation degradation are realized, and an important guarantee is provided for safe operation of a power system.

Description

Technical Field

[0001] This invention relates to a method for assessing the degree of cable insulation degradation based on frequency domain feature analysis, belonging to the field of cable insulation degradation detection. Background Technology

[0002] Power cables are the lifeblood of energy transmission networks, and their insulation performance directly affects the safe and stable operation of the power grid. However, during long-term operation, the cable insulation layer gradually ages, especially in humid environments. Water molecules, under the influence of an electric field, penetrate the insulation material, forming a dendritic microstructure known as "water tree." This water tree structure gradually expands over time, eventually leading to insulation breakdown and failure, causing serious power supply accidents. In particular, as the service life of cables increases, the insulation material gradually deteriorates under the influence of electric fields, moisture, and mechanical stress, with the growth of water tree structures being a major factor leading to insulation failure.

[0003] Existing insulation degradation assessment methods mostly rely on time-domain testing or partial discharge detection. These methods are not sensitive enough in detecting early water tree growth and are difficult to quantify the degree of degradation. Especially in complex operating environments, the measurement results are easily affected by noise and cannot accurately reflect the microscopic changes in the insulation material.

[0004] Frequency domain characteristic analysis methods have attracted attention in recent years, but their application still faces key challenges. The expansion of water dendrites leads to a decrease in low-frequency impedance characteristics, a phenomenon closely related to charge migration and polarization mechanisms within dielectric materials. Due to the decrease in low-frequency impedance, the relaxation polarization peaks in the frequency domain response become wider and flatter, further leading to an increase in harmonic components in the dielectric response spectrum, reflecting the complex evolution of the microstructure within the insulating material. Nevertheless, the frequency distribution characteristics of the insulation loss factor, an important indicator characterizing the growth stage of water trees, have not been fully elucidated, particularly lacking systematic research on the dynamic correlation between different growth stages.

[0005] Therefore, how to establish a degradation assessment model based on frequency domain characteristics by analyzing the correspondence between the frequency distribution characteristics of insulation loss factor and the water tree growth stage, so as to accurately predict the insulation penetration and breakdown time, has become a key issue in the study of cable insulation degradation. Summary of the Invention

[0006] Based on the problems described in the background, the problem to be solved by the present invention is to provide a method for evaluating the degree of cable insulation degradation based on frequency domain feature analysis, so as to solve the problems mentioned above.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating the degree of cable insulation degradation based on frequency domain feature analysis, comprising the following steps:

[0008] (1) Collect the electric field strength and moisture concentration in the cable operating environment, and combine the water tree growth data from the accelerated aging experiment in the laboratory to generate low-frequency impedance characteristic data with water tree growth stage markings, and obtain the initial impedance characteristic distribution.

[0009] (2) Based on the initial impedance characteristic distribution, extract the dielectric response spectrum of the cable insulation material, calculate the width and height of the relaxation polarization peak through the dielectric response spectrum, and obtain the width and height of the relaxation polarization peak.

[0010] (3) If the relaxation polarization peak width and height exceed the preset peak threshold, the dielectric response spectrum is decomposed in the frequency domain to extract the frequency and amplitude of the harmonic component changes and obtain the harmonic component feature set.

[0011] (4) Principal component analysis was used to reduce the dimensionality of the frequency and amplitude of the harmonic component changes, and the frequency distribution principal components of the insulation loss factor were extracted to obtain the loss factor distribution characteristics.

[0012] (5) Based on the loss factor distribution characteristics, and using a support vector regressor, input the width and height of the relaxation polarization peak and the harmonic component feature set, and output the length and density of the water tree to predict the water tree growth stage, which includes the initial, middle and late stages.

[0013] (6) If the growth stage prediction result is in the middle and late stage of water tree growth, calculate the characteristic parameters of microstructure evolution, predict the insulation penetration time, and obtain the penetration time estimate.

[0014] (7) Analyze the degradation probability of different cable insulation intervals by the penetration time estimate, generate degradation risk distribution curves based on the degradation probability of different cable insulation intervals, determine the probability interval of the remaining service life of the cable, generate an insulation degradation assessment report based on frequency domain characteristics, and output quantitative degradation degree and maintenance recommendations.

[0015] Preferably, step (1) includes the following steps:

[0016] (1.1) Perform frequency scanning on the cable insulation layer to obtain the first original impedance amplitude data and the first phase angle data;

[0017] (1.2) Perform wavelet transform processing on the first original impedance amplitude data and the first phase angle data to obtain the first impedance characteristic value dataset;

[0018] (1.3) Accelerated aging test was conducted on the cable insulation layer within the temperature gradient change range, and the second impedance characteristic value dataset and water tree length data were recorded by a low frequency impedance analyzer.

[0019] (1.4) For the first impedance feature value dataset and the second impedance feature value dataset, a random forest algorithm is used to establish a feature mapping relationship. If the feature mapping relationship is within a preset threshold range, it is determined that the cable insulation layer is in the water tree germination stage. The impedance feature values ​​are classified into water tree growth stages by support vector machine to obtain the initial impedance characteristic distribution.

[0020] Preferably, step (2) includes the following steps:

[0021] (2.1) Receive the impedance data of the cable insulation layer, and perform frequency domain decomposition through Fourier transform based on the impedance data to obtain the original dielectric response spectrum;

[0022] (2.2) Wavelet denoising is performed on the original dielectric response spectrum, and the denoised dielectric response spectrum is obtained by extracting the dielectric loss factor and dielectric loss angle.

[0023] (2.3) Extract the relaxation time constant distribution from the noise reduction dielectric response spectrum, obtain the frequency domain response envelope through Hilbert transform, and generate the relaxation time spectrum;

[0024] (2.4) If the relaxation time spectrum has local maxima, then fit the local maxima to obtain the relaxation polarization peak profile curve, calculate the half-width of the relaxation polarization peak profile curve, and generate the relaxation polarization peak width and height.

[0025] Preferably, step (3) includes the following steps:

[0026] (3.1) Based on the relaxation polarization peak width and peak height in the dielectric response spectrum, a preset peak threshold is used for judgment. If the peak width exceeds the preset width threshold or the peak height exceeds the preset height threshold, an abnormal response spectrum is obtained.

[0027] (3.2) Perform four-level frequency domain decomposition on the abnormal response spectrum and obtain a multi-scale frequency domain spectrum through orthogonal transformation;

[0028] (3.3) Extract the frequency band signal from the multi-scale frequency domain spectrum, calculate the instantaneous frequency and amplitude using Hilbert transform, and obtain the frequency modulation feature sequence;

[0029] (3.4) Random forest classification is performed on the frequency modulation feature sequence to extract harmonic component parameters and generate an initial harmonic feature set. The initial harmonic feature set is smoothed by radial basis kernel support vector regression to obtain the harmonic component feature set.

[0030] Preferably, step (4) includes the following steps:

[0031] (4.1) Construct a loss factor frequency response matrix based on the harmonic frequency parameters and amplitude parameters, and obtain a standardized loss matrix by subtracting the mean and dividing the standard deviation.

[0032] (4.2) Calculate the covariance matrix for the standardized loss matrix, obtain the eigenvalues ​​and eigenvectors, select the number of principal components according to the cumulative contribution rate of the eigenvalues ​​exceeding the preset threshold, and obtain the dimension-reduced loss data through orthogonal transformation;

[0033] (4.3) A five-layer feedforward neural network is used to process the dimensionality reduction loss data. The number of hidden layer neurons in the neural network is twice the input dimension, and a loss feature vector is generated.

[0034] (4.4) The loss feature vector is nonlinearly mapped by a radial basis function support vector regressor. The main loss features are selected based on the contribution rate of the feature components to the total variance exceeding a preset threshold, and the loss factor distribution features are generated.

[0035] Preferably, step (5) includes the following steps:

[0036] (5.1) Based on the loss factor distribution characteristics, relaxation polarization peak width parameter and peak height parameter and harmonic characteristic set, the characteristics are standardized by subtracting the mean and dividing the standard deviation to obtain the standardized characteristic matrix;

[0037] (5.2) For the standardized feature matrix, principal component analysis is used to perform dimensionality reduction processing to obtain the feature components whose cumulative contribution rate exceeds a preset threshold, and thus obtain the dimensionality-reduced feature dataset.

[0038] (5.3) Based on the water tree length and water tree density values ​​in the dimensionality reduction feature dataset, the normalized target value is obtained by the maximum and minimum value normalization process. The normalized target value is divided into three categories by the k-means clustering algorithm to obtain the mid-term lower limit threshold, the mid-term upper limit threshold, and the late-term lower limit threshold.

[0039] (5.4) For the dimensionality reduction feature dataset, a radial basis function support vector regressor is used to establish a mapping relationship between the dimensionality reduction feature data and the water tree parameters. If the water tree length value is less than the lower threshold of the middle stage, it is determined to be in the early stage. If the water tree length value is in the middle stage threshold range, it is determined to be in the middle stage. If the water tree length value is greater than the lower threshold of the late stage, it is determined to be in the late stage.

[0040] Preferably, step (6) includes the following steps:

[0041] (6.1) Judge the prediction results of the growth stage. If the prediction results are in the range from the upper limit of the mid-term threshold to the late-term threshold, the feature extraction module is used to obtain the water tree structure feature sequence.

[0042] (6.2) Perform Hilbert transform processing on the water tree structure feature sequence, and obtain the water tree evolution feature sequence in the preset frequency band through the instantaneous amplitude and phase information extractor;

[0043] (6.3) Perform a preset number of wavelet decompositions on the water tree evolution feature sequence, and obtain multi-scale feature sequences through trend component and detail component extractors;

[0044] (6.4) The multi-scale feature sequence is fitted with a hyperbolic function, and an evolution rate parameter to breakdown time mapping function is established by a random forest regressor. The insulation penetration time estimate is obtained according to the time mapping function.

[0045] Preferably, step (7) includes the following steps:

[0046] (7.1) A normal distribution random number generator is used to obtain the sampling sequence of the insulation interval, and the initial probability distribution data is generated based on the sampling sequence;

[0047] (7.2) Based on the initial probability distribution data, a probability density function is constructed using the kernel density estimation method to obtain a continuous probability curve;

[0048] (7.3) Perform convolution operation on the continuous probability curve to obtain the degradation risk curve, and divide the remaining lifetime interval data according to the degradation risk curve;

[0049] (7.4) Calculate the degradation score for the remaining life interval data, perform random forest classification on the frequency domain feature data, set feature weight vector, calculate the comprehensive degradation score, generate maintenance level data based on the degradation score, obtain maintenance suggestions through the maintenance level data, and generate an insulation degradation assessment report based on frequency domain features.

[0050] The beneficial effects of this invention are:

[0051] 1. By collecting data on cable operating environment and accelerated aging in the laboratory, a low-frequency impedance characteristic dataset with annotations of water tree growth stages is generated. Combined with machine learning algorithms, the growth stages are effectively quantified, significantly improving the assessment accuracy and providing a scientific basis for cable maintenance.

[0052] 2. Based on the dielectric response spectrum generated from the initial impedance characteristic distribution, combined with wavelet denoising, Hilbert transform and convolutional neural network, the relaxation polarization peak width and height data are accurately extracted, which can accurately reflect the degradation state of the cable insulation layer. Compared with traditional methods, it shows stronger robustness in dealing with complex noise environments. Furthermore, the combination of convolutional neural network and double Gaussian fitting significantly improves the accuracy of feature extraction, laying a solid foundation for degradation assessment.

[0053] 3. When the relaxation polarization peak parameter is abnormal, wavelet frequency domain decomposition and Hilbert transform are used to extract the harmonic component feature set. By combining machine learning algorithms, the accuracy and reliability of harmonic feature extraction are improved. Especially in complex operating environments, it can effectively distinguish between normal and abnormal states.

[0054] 4. By using principal component analysis and neural networks to process harmonic features, the frequency distribution features of the loss factor are extracted. This fully leverages the advantages of dimensionality reduction and nonlinear mapping, significantly improving the efficiency and accuracy of feature extraction, and enabling more precise capture of the loss change trend caused by aging.

[0055] 5. A frequency domain feature model based on support vector regression was constructed, which integrates multi-dimensional frequency domain features to predict the length and density of water trees, and then determines the growth stage of water trees. The prediction results are highly consistent with the actual dissection and detection, which verifies the high accuracy of the model and provides technical support for optimizing cable maintenance strategies.

[0056] 6. For mid-to-late stage water trees, the penetration time estimate generated by frequency domain feature analysis, wavelet multi-scale decomposition and regression modeling can accurately reflect the dynamic characteristics of microstructure evolution, significantly improving the reliability and practicality of the prediction.

[0057] 7. Based on the penetration time estimate, the degradation probability distribution is generated through Monte Carlo simulation. Combined with dynamic programming, the optimal replacement time is determined, and a comprehensive assessment report containing quantitative degradation degree, risk level and maintenance recommendations is generated. This provides quantitative risk assessment and optimization decision support, significantly improving the scientific and economical nature of cable maintenance. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the method flow of the present invention; Detailed Implementation

[0059] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0060] Example 1

[0061] like Figure 1 As shown, this invention provides a method for evaluating the degree of cable insulation degradation based on frequency domain feature analysis, comprising the following steps:

[0062] (1) Collect the electric field strength and moisture concentration in the cable operating environment, and combine the water tree growth data from the accelerated aging experiment in the laboratory to generate low-frequency impedance characteristic data with water tree growth stage markings, and obtain the initial impedance characteristic distribution.

[0063] Step (1) includes the following steps:

[0064] (1.1) Collect electric field strength parameters and moisture concentration index around the cable according to the preset sampling interval, and perform frequency scanning on the cable insulation layer through low frequency impedance analysis to obtain the first original impedance amplitude data and the first phase angle data;

[0065] (1.2) Based on the first original impedance amplitude data and the first phase angle data, perform noise reduction and smoothing processing, and obtain the first impedance characteristic value dataset through wavelet transform;

[0066] (1.3) Accelerated aging test was conducted on the dielectric loss factor, polarization coefficient and conductivity parameters of the cable insulation layer within the temperature gradient range. The second impedance characteristic value dataset and water tree length data were recorded by a low frequency impedance analyzer.

[0067] The second impedance feature value dataset is generated by extracting features from impedance data obtained from accelerated aging experiments, and includes impedance magnitude, loss tangent, and frequency response curve feature values.

[0068] (1.4) For the first impedance feature value dataset and the second impedance feature value dataset, a random forest algorithm is used to establish a mapping relationship between impedance feature values ​​and water tree length. If the feature mapping relationship is within a preset threshold range, it is determined that the cable insulation layer is in the water tree germination stage. The impedance feature values ​​are classified into water tree growth stages by support vector machine to obtain the initial impedance characteristic distribution.

[0069] Specifically, based on the range of impedance modulus and the trend of loss tangent, the growth of water trees is divided into three stages: germination, expansion, and penetration. Cross-validation is used to verify the accuracy of the classification results, generating a low-frequency impedance characteristic dataset labeled with the water tree growth stages.

[0070] In step (1), the evaluation process begins with data acquisition to ensure the comprehensiveness and reliability of the input data. The acquisition equipment includes an electric field strength sensor, a humidity concentration detector, and a low-frequency impedance analyzer, which are used to acquire operating environment parameters and insulation layer impedance characteristics, respectively. The specific acquisition method can be determined according to the actual scenario.

[0071] In step (1.1), data is collected using an electric field strength sensor and a moisture concentration detector.

[0072] In step (1.2), the original impedance data needs to be preprocessed to eliminate noise interference. The denoising process uses the db3 wavelet basis function, and the number of decomposition layers is set to 4. A smooth impedance curve is generated by reconstruction. In feature extraction, statistical features such as energy concentration, standard deviation and kurtosis coefficient are selected to form the first impedance characteristic dataset.

[0073] In step (1.3), the accelerated aging experiment is carried out in a temperature gradient range of 60 to 90 degrees Celsius. The water tree length and impedance characteristics are recorded every 6 hours to generate a second impedance characteristic dataset, which includes impedance modulus, loss tangent and frequency response curve characteristics.

[0074] In step (1.4), the random forest algorithm uses 500 decision trees. The input features include impedance modulus ratio, loss angle change rate, and frequency response feature value. The mapping relationship threshold is set based on experimental data statistics. The support vector machine uses the radial basis kernel function with a kernel parameter of 0.1. The stage classification is based on the impedance modulus reduction (no more than 20% in the germination stage, 20% to 50% in the expansion stage, and more than 50% in the penetration stage). The generated low-frequency impedance characteristic dataset contains water tree growth stage annotations, providing a reliable basis for subsequent analysis.

[0075] In step (1) of this invention, an impedance characteristic dataset reflecting the growth stage of water trees is constructed through multi-source data acquisition and feature extraction. Compared with traditional methods, this method exhibits higher sensitivity in early water tree detection and effectively quantifies the growth stage through machine learning algorithms, significantly improving the evaluation accuracy and providing a scientific basis for cable maintenance.

[0076] (2) Based on the initial impedance characteristic distribution, extract the dielectric response spectrum of the cable insulation material, calculate the width and height of the relaxation polarization peak through the dielectric response spectrum, and obtain the width and height of the relaxation polarization peak.

[0077] Step (2) includes the following steps:

[0078] (2.1) Receive the impedance data of the cable insulation layer, and perform frequency domain decomposition through Fourier transform based on the impedance data to obtain the original dielectric response spectrum;

[0079] The original dielectric response spectrum was generated by sampling at logarithmic uniform intervals.

[0080] (2.2) Perform wavelet denoising processing based on the original dielectric response spectrum. Set the wavelet decomposition layer to 3 layers. Obtain the denoised dielectric response spectrum by extracting the dielectric loss factor and dielectric loss angle.

[0081] (2.3) Extract the relaxation time constant distribution from the noise reduction dielectric response spectrum, obtain the frequency domain response envelope through Hilbert transform, and generate the relaxation time spectrum;

[0082] (2.4) If the relaxation time spectrum has local maxima, then fit the local maxima to obtain the relaxation polarization peak profile curve, calculate the half-width of the relaxation polarization peak profile curve, and generate the relaxation polarization peak width and height.

[0083] Specifically, a set of relaxed polarization feature points is generated by extracting local maxima using a convolutional neural network. A double Gaussian distribution function is then used to fit the feature point set, and a Gaussian function variance threshold is set to obtain the relaxed polarization peak profile curve. Based on the relaxed polarization peak profile curve, the peak half-width is calculated using cubic spline interpolation, and the amplitude of the maximum peak point is recorded as the peak height. Support vector regression is then applied to smooth the peak parameters, generating relaxed polarization peak width and height data.

[0084] In step (2), the generation and analysis of dielectric response spectrum is the core step in evaluating cable insulation degradation. Frequency domain feature extraction reflects the microstructural changes caused by water tree growth. The initial impedance characteristic distribution provides a data basis for subsequent spectrum analysis. The analysis process combines multiple signal processing techniques to ensure the accuracy and reliability of feature extraction. The specific implementation method can be adjusted according to actual measurement needs.

[0085] In step (2.1), impedance data acquisition is performed using a low-frequency impedance analyzer with a frequency range set from 0.001 Hz to 1000 Hz. 128 sampling points are selected at logarithmic intervals to capture the polarization characteristics of the low-frequency band. The impedance data is converted into the frequency domain through fast Fourier transform to generate the original dielectric response spectrum, which includes impedance magnitude and phase angle information.

[0086] In step (2.2), in order to eliminate the influence of measurement noise on the spectrum, the db4 wavelet basis function is used for denoising. The number of decomposition layers is set to 3 to balance noise suppression and feature preservation. After denoising, the signal-to-noise ratio is significantly improved. The denoised dielectric response spectrum is generated by extracting the dielectric loss factor and dielectric loss angle, providing high-quality data for subsequent feature analysis.

[0087] In step (2.3), the relaxation time constant reflects the dynamic characteristics of the internal polarization process of the insulating material and is a key indicator for judging the growth stage of the water tree. By analyzing the noise reduction dielectric response spectrum, the relaxation time constant distribution is extracted, and the frequency domain response envelope is generated by Hilbert transform to construct the relaxation time spectrum. During the transformation process, a Gaussian window function with a bandwidth of 10 Hz is used to smooth the envelope curve.

[0088] In step (2.4), in order to accurately identify the local maxima in the spectrum, a five-layer convolutional neural network is used. The input is data of 251 frequency points. The network outputs 15 feature points in the main peak region and 8 feature points in the secondary peak region. These feature points constitute a polarization feature set, which provides a basis for subsequent peak fitting.

[0089] Furthermore, fitting local maxima is the core step in extracting the features of relaxed polarization peaks. A double Gaussian distribution function is used to fit the feature point set, with the variance of the primary peak Gaussian function set to 0.8 and the secondary peak set to 0.4. The root mean square error of the fit is controlled below 0.01, generating a smooth peak profile curve. Simultaneously, to accurately calculate the peak parameters, a cubic spline interpolation algorithm is used with an interpolation node interval of 0.1 Hz to calculate the half-width at half-maximum and maximum amplitude of the profile curve. These parameters are smoothed using support vector regression with a radial basis function kernel function and a kernel parameter of 0.1. After smoothing, the parameter change rate is less than 5%, ensuring data stability.

[0090] Meanwhile, the width and height of the relaxation polarization peak are directly related to the microstructural evolution of water tree growth. Long-term operation of cables leads to increased polarization loss due to water tree aging, manifested as a significant increase in peak width and a decrease in height. By continuously monitoring the trends of these parameters, early signs of degradation can be detected in a timely manner. The generated peak characteristic data includes width, height, and corresponding frequency information, providing reliable input for subsequent prediction of water tree growth stages. This step does not impose excessive limitations on the specific monitoring frequency, which can be adjusted by technicians according to the actual scenario.

[0091] In step (2) of this invention, the generated relaxation polarization peak data can accurately reflect the degradation state of the cable insulation layer through multi-level frequency domain analysis and feature extraction. Compared with traditional methods, this method shows stronger robustness in handling complex noise environments, and the combination of convolutional neural network and double Gaussian fitting significantly improves the accuracy of feature extraction, laying a solid foundation for degradation assessment.

[0092] (3) If the relaxation polarization peak width and height exceed the preset peak threshold, the dielectric response spectrum is decomposed in the frequency domain to extract the frequency and amplitude of the harmonic component changes and obtain the harmonic component feature set to quantify the degree of insulation degradation.

[0093] Step (3) includes the following steps:

[0094] (3.1) Based on the relaxation polarization peak width and peak height in the dielectric response spectrum, a preset peak threshold is used for judgment. If the peak width exceeds the preset width threshold or the peak height exceeds the preset height threshold, an abnormal response spectrum is obtained.

[0095] (3.2) Perform four-level frequency domain decomposition on the abnormal response spectrum and obtain a multi-scale frequency domain spectrum through orthogonal transformation; the four-level frequency domain decomposition is performed using the db4 wavelet basis function, and a multi-scale frequency domain spectrum is generated by obtaining the low-frequency approximation component and the high-frequency detail component.

[0096] (3.3) Extract the frequency band signal from the multi-scale frequency domain spectrum, calculate the instantaneous frequency and amplitude using Hilbert transform, and obtain the frequency modulation feature sequence;

[0097] The frequency band interval is set to octave, and the amplitude and center frequency of each frequency band signal are recorded.

[0098] (3.4) Random forest classification is performed on the frequency modulation feature sequence to extract harmonic component parameters and generate an initial harmonic feature set. The initial harmonic feature set is smoothed by radial basis kernel support vector regression to obtain the harmonic component feature set.

[0099] Specifically, radial basis function (RBF) kernel support vector regression is used to smooth the initial harmonic feature set, eliminate abrupt changes and noise interference, and the smoothed harmonic feature set is extracted using singular value decomposition. The feature vector corresponding to the largest singular value is selected as the harmonic component feature set.

[0100] In step (3), the relaxation polarization peak is the core indicator of dielectric polarization characteristics. Its abnormal changes reflect the degradation of insulation performance caused by water tree growth. When the peak parameter exceeds the normal range, it is necessary to further extract harmonic components through frequency domain analysis to capture the nonlinear characteristic changes in the degradation process, ensure the accuracy and robustness of feature extraction, and provide reliable data support for subsequent degradation assessment.

[0101] In step (3.1), the normal range of relaxation polarization peak value is determined based on a large number of healthy cable samples.

[0102] In step (3.2), the anomalous spectrum needs to be further decomposed to extract potential harmonic features. A four-level frequency domain decomposition is performed using the db4 wavelet basis function, which performs excellently in time-frequency localization and can effectively separate signals in different frequency bands. Through orthogonal transformation, a multi-scale frequency domain spectrum is generated, containing low-frequency approximate components and high-frequency detail components, laying the foundation for frequency band analysis.

[0103] In step (3.3), the frequency band signal extraction adopts a 1 / 3 octave band division scheme to adapt to the nonlinear characteristics of the medium response. The instantaneous frequency and amplitude of each frequency band signal are calculated by Hilbert transform to generate a frequency modulation feature sequence, revealing the frequency modulation characteristics caused by water tree aging. This periodic change is closely related to the microstructural evolution in the early stage of water tree growth, providing an important basis for the identification of the deterioration stage.

[0104] In step (3.4), the frequency modulation feature sequence is processed by a random forest classifier containing 500 decision trees. The input features include instantaneous frequency, amplitude envelope, and frequency modulation depth. To eliminate abrupt changes and noise interference in the sequence, radial basis function (RBF) kernel support vector regression is used for smoothing. Furthermore, features are extracted through singular value decomposition, and the feature vector corresponding to the first singular value is selected as the harmonic component feature set, effectively summarizing the main harmonic variation patterns.

[0105] In step (3) of this invention, the aging process of cable insulation material can be dynamically tracked by continuously monitoring the changes in the harmonic component feature set. This method improves the accuracy and reliability of harmonic feature extraction by combining wavelet transform, Hilbert transform and machine learning algorithms. Especially in complex operating environments, it can effectively distinguish between normal and abnormal states, providing a scientific basis for the formulation of cable maintenance strategies.

[0106] (4) Principal component analysis was used to reduce the dimensionality of the frequency and amplitude of the harmonic component changes, and the frequency distribution principal components of the insulation loss factor were extracted to obtain the loss factor distribution characteristics.

[0107] Step (4) includes the following steps:

[0108] (4.1) Construct a loss factor frequency response matrix based on the harmonic frequency parameters and amplitude parameters, and obtain a standardized loss matrix by subtracting the mean and dividing the standard deviation.

[0109] (4.2) Calculate the covariance matrix for the standardized loss matrix, obtain the eigenvalues ​​and eigenvectors, select the number of principal components according to the cumulative contribution rate of the eigenvalues ​​exceeding the preset threshold, and obtain the dimension-reduced loss data through orthogonal transformation;

[0110] A dimensionality reduction transformation matrix is ​​constructed based on the number of principal components selected, and an orthogonal transformation is performed on the standardized loss matrix.

[0111] (4.3) A five-layer feedforward neural network is used to process the dimensionality reduction loss data. The number of hidden layer neurons in the neural network is twice the input dimension, and a loss feature vector is generated.

[0112] (4.4) The loss feature vector is nonlinearly mapped through the radial basis kernel function support vector regressor to generate an initial feature space. The main loss features are selected based on the contribution rate of the feature components to the total variance exceeding a preset threshold. The main loss features are reconstructed in the frequency domain to generate loss factor distribution features.

[0113] In step (4), the wave component feature set contains the frequency response information of the medium loss. By reducing dimensions and extracting features, the data dimension can be effectively compressed and key loss characteristics can be preserved. This step constructs the frequency distribution characteristics of the loss factor through the combination of principal component analysis, neural network processing and support vector regression, providing accurate input for subsequent prediction of the water tree growth stage.

[0114] In step (4.1), the loss factor frequency response matrix is ​​generated through a harmonic component feature set, covering multiple frequency points within the frequency range of 0.001 to 1000 Hz. To eliminate dimensional differences and uneven data distribution, standardization is performed by subtracting the mean and dividing the standard deviation, generating a standardized loss matrix with a mean of 0 and a standard deviation of 1. Standardization not only improves the numerical stability of subsequent analyses but also enhances the comparability between features, laying the foundation for covariance matrix calculation.

[0115] In step (4.2), the covariance matrix is ​​used to analyze the correlation between features in the standardized loss matrix. Eigenvalues ​​and eigenvectors are extracted through eigenvalue decomposition. Eigenvalues ​​reflect the contribution of each principal component to the total variance, and eigenvectors define the dimensionality reduction direction. Through orthogonal transformation, the 256-dimensional standardized loss matrix is ​​reduced to 5 dimensions, generating dimensionality-reduced loss data. The first principal component mainly reflects the fundamental frequency loss characteristics, showing a logarithmic decreasing trend with frequency; the second principal component captures the influence of the third harmonic, forming local peaks in the mid-frequency band; the third principal component is related to higher harmonics, with slower attenuation in the high-frequency band. This dimensionality reduction method effectively preserves the frequency response characteristics of the loss factor while reducing computational complexity.

[0116] In step (4.3), a five-layer feedforward neural network is used to further mine the nonlinear features in the dimensionality-reduced loss data. The number of hidden layer neurons is set to twice the input dimension to balance model complexity and fitting ability. During training, the loss feature vectors of normal samples exhibit high clustering.

[0117] In step (4.4), the generated loss feature vector is nonlinearly mapped using a radial basis function support vector regressor with a kernel parameter set to 0.1, mapping it to an 8-dimensional feature space. The mapping results show that the projection points of normal samples are concentrated near the origin of the feature space, while those of aging samples diffuse outwards, exhibiting a significant clustering effect. Based on the contribution rate of each feature component to the total variance, three main feature components with a cumulative contribution rate exceeding 85% are selected to characterize the overall level, frequency dependence, and nonlinearity of the loss spectrum, respectively. The frequency distribution characteristics of the loss factor are generated through frequency domain reconstruction, revealing the aging characteristics.

[0118] In step (4) of this invention, the generation of the frequency distribution characteristics of the loss factor fully utilizes the advantages of dimensionality reduction and nonlinear mapping, significantly improving the efficiency and accuracy of feature extraction. Compared with traditional methods, this method, through the combination of principal component analysis and neural networks, can more accurately capture the trend of loss changes caused by aging, providing a reliable basis for the dynamic monitoring of cable insulation status.

[0119] (5) Based on the loss factor distribution characteristics, and using a support vector regressor, input the width and height of the relaxation polarization peak and the harmonic component feature set, and output the length and density of the water tree to predict the water tree growth stage, which includes the initial, middle and late stages.

[0120] Step (5) includes the following steps:

[0121] (5.1) Based on the loss factor distribution characteristics, relaxation polarization peak width parameter and peak height parameter and harmonic characteristic set, the characteristics are standardized by subtracting the mean and dividing the standard deviation to obtain the standardized characteristic matrix;

[0122] (5.2) For the standardized feature matrix, principal component analysis is used to perform dimensionality reduction processing to obtain the feature components whose cumulative contribution rate exceeds a preset threshold, and thus obtain the dimensionality-reduced feature dataset.

[0123] (5.3) Based on the water tree length and water tree density values ​​in the dimensionality reduction feature dataset, the normalized target value is obtained by the maximum and minimum value normalization process. The normalized target value is divided into three categories by the k-means clustering algorithm to obtain the mid-term lower limit threshold, the mid-term upper limit threshold, and the late-term lower limit threshold.

[0124] (5.4) For the dimensionality reduction feature dataset, a radial basis function support vector regressor is used to establish a mapping relationship between the dimensionality reduction feature data and the water tree parameters. If the water tree length value is less than the lower threshold of the middle stage, it is determined to be in the early stage. If the water tree length value is in the middle stage threshold range, it is determined to be in the middle stage. If the water tree length value is greater than the lower threshold of the late stage, it is determined to be in the late stage. The loss factor distribution characteristics reflect the aging degree of the water tree, while the relaxation polarization peak and harmonic characteristics reflect the medium response characteristics.

[0125] Specifically, a radial basis function kernel support vector regressor is constructed, a grid search parameter range is set, the optimal kernel parameters are selected through five-fold cross-validation, and then the support vector regressor is trained based on the optimal kernel parameters.

[0126] In step (5), the prediction of the water tree growth stage is a key link in the assessment of cable insulation degradation. By constructing a support vector regression model by integrating multi-dimensional frequency domain features, the feature data can be accurately mapped to the water tree parameters, providing a scientific basis for cable maintenance.

[0127] In step (5.1), the input features include the loss factor distribution, the width and height of the relaxation polarization peak, and the harmonic component features, which respectively reflect the dielectric loss, polarization characteristics, and nonlinear response. To unify the dimensions and eliminate differences in data distribution, the features are standardized by subtracting the mean and dividing the standard deviation, generating a standardized feature matrix with a mean of 0 and a standard deviation of 1.

[0128] In step (5.2), principal component analysis is applied to reduce the dimensionality of the standardized feature matrix, calculate the covariance matrix and extract eigenvalues ​​and eigenvectors, reducing the original 35-dimensional features to 5-dimensional features. This significantly reduces redundancy while retaining core information, which not only improves computational efficiency but also enhances the model's robustness to noise, providing high-quality input for subsequent modeling.

[0129] In step (5.3), the original water tree length ranges from 0.2 to 2.5 mm, and the density ranges from 50 to 500 trees per cubic centimeter. These are mapped to the 0-1 interval through maximum-minimum value normalization to eliminate the influence of dimensions. The k-means clustering algorithm is used to divide the normalized water tree length and density values ​​into three categories, generating stage thresholds. The clustering process uses Euclidean distance as a metric, iteratively optimizing the cluster centers. The final thresholds are: mid-term lower limit threshold of 0.8 mm length and 150 trees per cubic centimeter; mid-term upper limit threshold of 1.5 mm length and 300 trees per cubic centimeter; and late-term lower limit threshold of 2 mm length and 400 trees per cubic centimeter. These thresholds have clear physical meaning, reflecting the evolution of the water tree from germination to expansion and then to completion. In the initial stage, the water tree length is short and the density is low, indicating that deterioration has just begun; in the mid-term, the water tree expands significantly and the density increases; in the late-term, the water tree is close to completion, and the density increases significantly. The clustering results were verified by the silhouette coefficient, with an average coefficient of 0.85, indicating that the division has high discriminative power.

[0130] In step (5.4), the support vector regression model uses a radial basis function kernel function. The kernel parameters and penalty factor are optimized through grid search. The candidate kernel parameters range from 0.01 to 10, and the penalty factor ranges from 1 to 1000. Five-fold cross-validation results show that the root mean square error (RMSE) is minimized (approximately 0.05) when the kernel parameter is 0.1 and the penalty factor is 100, indicating excellent model generalization performance. The trained model maps dimensionality-reduced feature data to water tree length and density values. This mapping accurately reflects the nonlinear correlation between degradation features and water tree parameters. Based on the water tree length values ​​output by the model, the growth stage is determined by threshold comparison: lengths less than 0.8 mm are considered the initial stage, 0.8 to 1.5 mm the intermediate stage, and greater than 2 mm the late stage. For example, in a sample of cross-linked polyethylene cables that have been in operation for 10 years, the predicted length of samples in the low-voltage section is approximately 0.6 mm, indicating the initial stage (35%); the length of samples in the frequently fluctuating voltage section is around 1.2 mm, indicating the intermediate stage (45%); and the length of samples in the high-voltage impact section exceeds 2 mm, indicating the late stage (20%). The predicted results matched the actual anatomical examination with a 90% accuracy, verifying the high precision of the model.

[0131] In step (5) of this invention, by continuously monitoring the predicted results of water tree length and density, the aging process of cable insulation can be dynamically evaluated, which improves the reliability of the prediction of water tree growth stage. Especially in complex operating environments, it can effectively distinguish different degrees of degradation, providing technical support for optimizing cable maintenance strategies.

[0132] (6) If the growth stage prediction result is in the middle and late stage of water tree growth, calculate the characteristic parameters of microstructure evolution, predict the insulation penetration time, and obtain the penetration time estimate.

[0133] Step (6) includes the following steps:

[0134] (6.1) Judge the prediction results of the growth stage. If the prediction results are in the range from the upper limit of the mid-term threshold to the late-term threshold, the feature extraction module is used to obtain the water tree structure feature sequence.

[0135] (6.2) Perform Hilbert transform processing on the water tree structure feature sequence, and obtain the water tree evolution feature sequence in the preset frequency band through the instantaneous amplitude and phase information extractor;

[0136] (6.3) Perform a preset number of wavelet decompositions on the water tree evolution feature sequence, and obtain multi-scale feature sequences through trend component and detail component extractors;

[0137] (6.4) The multi-scale feature sequence is fitted with a hyperbolic function, and an evolution rate parameter to breakdown time mapping function is established through a random forest regressor. The insulation penetration time estimate is obtained based on the time mapping function.

[0138] During fitting, initial values ​​for fitting parameters are set, evolution rate parameters are calculated, historical breakdown data are reconstructed according to time series, periodic features are extracted using the maximum entropy method, and a breakdown feature sequence is generated. Furthermore, the initial penetration time is calculated based on the time mapping function, and the prediction results are optimized using a cubic exponential smoothing method. The smoothed prediction results are then calibrated using a sliding time window method to obtain an estimated insulation penetration time.

[0139] In step (6), the mid-to-late stage of water tree growth indicates that insulation degradation has entered a critical phase, and the rapid evolution of the microstructure significantly increases the risk of breakdown. This step quantifies the evolution characteristics of the water tree and predicts the insulation penetration time through frequency domain feature analysis, signal processing, and regression modeling, providing a scientific basis for optimizing the timing of cable replacement.

[0140] In step (6.1), the determination of the water tree growth stage is based on the previous prediction results. The upper limit of the middle stage corresponds to a water tree length of 1.5 mm and a loss factor of approximately 0.015. The threshold of the later stage corresponds to a length of 2.0 mm and a loss factor exceeding 0.02. If the prediction results fall within this range, it indicates that the water tree has entered a rapid development stage, and further analysis of microstructural changes is required. A feature extraction module is used to process the frequency domain features, generating a water tree structural feature sequence, which includes amplitude, phase, and frequency distribution information. For this sequence, Hilbert transform is applied to extract instantaneous amplitude and phase features in the 0.001 to 1000 Hz frequency band, generating a water tree evolution feature sequence.

[0141] In step (6.2), the Hilbert transform constructs an envelope by analyzing the signal, capturing nonlinear dynamic characteristics. Compared with the traditional Fourier transform, it is more suitable for analyzing abrupt signals, providing high-quality input for subsequent multi-scale feature extraction.

[0142] In step (6.3), the water tree evolution feature sequence contains long-term trend and short-term fluctuation information, requiring multi-scale analysis to separate features at different time scales. A four-level wavelet decomposition using the db4 wavelet basis function is employed due to its excellent ability to characterize abrupt changes. The first level of decomposition extracts high-frequency detail components, reflecting short-term fluctuations; subsequent levels progressively extract low-frequency trend components, reflecting long-term evolution. After four levels of decomposition, the generated multi-scale feature sequence contains trend components and detail components from each level, expanding the dimension from 128 points in the original sequence to a 256-dimensional feature vector containing both trends and details. This multi-scale analysis not only preserves the integrity of the evolutionary features but also highlights the nonlinear growth characteristics of the mid-to-late stages, laying the foundation for fitting and regression modeling.

[0143] In step (6.4), the hyperbolic function is used to fit the multi-scale feature sequence because it can effectively describe the nonlinear acceleration characteristics of water tree growth. The fitting parameters include asymptotic values, growth rate, and lag factor, which are optimized using the least squares criterion. The fitted evolution rate parameters show that the water tree growth rate in the middle and late stages is about 3 to 5 times that in the early stages. Combining historical breakdown data, periodic features are extracted through maximum entropy spectrum analysis, with the analysis window length set to 128 points and the order being 1 / 3 of the sample size. The results show that the breakdown time distribution of a certain cable exhibits a bimodal characteristic, with the main peak at 8000 hours and the secondary peak at 12000 hours, reflecting the influence of environmental stress. The random forest regressor uses 500 decision trees, with input features including evolution rate, cumulative loss, and environmental stress, and the output being breakdown time. The training results show that the correlation coefficient between evolution rate and breakdown time reaches -0.85, indicating a strong correlation. To improve prediction stability, a triple exponential smoothing method is used to optimize the initial prediction results, with smoothing coefficients of 0.7, 0.2, and 0.1, respectively, to reduce fluctuation amplitude. The predicted values ​​are further calibrated using a 24-hour sliding time window, with samples within the window allocated according to exponential weights.

[0144] In step (6) of this invention, the generated penetration time estimate can accurately reflect the dynamic characteristics of the water tree evolution in the middle and late stages through multi-scale feature extraction and regression modeling. Compared with traditional methods, this method comprehensively considers the microstructural characteristics and the statistical regularity of historical data, significantly improving the reliability and practicality of the prediction, and providing technical support for the dynamic management of cable insulation status.

[0145] (7) Analyze the degradation probability of different cable insulation intervals by the penetration time estimate, generate degradation risk distribution curves based on the degradation probability of different cable insulation intervals, determine the probability interval of the remaining service life of the cable, generate an insulation degradation assessment report based on frequency domain characteristics, and output quantitative degradation degree and maintenance recommendations.

[0146] Step (7) includes the following steps:

[0147] (7.1) Establish a Monte Carlo random sampling sequence based on the penetration time estimate, use a normal distribution random number generator to obtain the insulation interval sampling sequence, and generate initial probability distribution data based on the sampling sequence;

[0148] (7.2) Based on the initial probability distribution data, a probability density function is constructed using the kernel density estimation method to obtain a continuous probability curve;

[0149] (7.3) Perform convolution operation on the continuous probability curve to obtain the degradation risk curve, divide the remaining lifetime interval data according to the degradation risk curve, and generate lifetime probability distribution.

[0150] (7.4) Calculate the degradation score for the remaining life interval data, perform random forest classification on the frequency domain feature data, set feature weight vector, calculate the comprehensive degradation score, generate maintenance level data based on the degradation score, obtain maintenance suggestions through the maintenance level data, and generate an insulation degradation assessment report based on frequency domain features.

[0151] Specifically, a time optimization function is constructed using dynamic programming, a state transition cost matrix is ​​set, and the optimal replacement time point is solved using a forward recursion algorithm. Random forest classification is performed on the frequency domain feature data, a feature weight vector is set, a comprehensive degradation score is calculated, a grading standard is established based on the degradation score, four maintenance levels are divided, a maintenance work list is generated, and the lifetime probability distribution, optimal replacement time point, and maintenance level information are integrated to generate an insulation degradation assessment report.

[0152] In step (7), the penetration time estimate reflects the remaining life of the cable insulation. Probabilistic analysis and optimization algorithms are needed to quantify the degradation risk and formulate maintenance strategies. This step utilizes Monte Carlo simulation to generate a degradation probability distribution, combines it with dynamic programming to determine the optimal replacement time, and generates a comprehensive assessment report, providing data-driven support for cable operation management. Compared to traditional methods, this method significantly improves the accuracy and practicality of risk assessment through probabilistic modeling and multi-feature fusion.

[0153] In step (7.1), the Monte Carlo simulation simulates the degradation process under different operating conditions through random sampling, generating a probability distribution of penetration time. A normal distribution random number generator is used.

[0154] In step (7.2), a probability density function is constructed for the sampled sequence using the kernel density estimation method. A Gaussian kernel function is selected, and the bandwidth is set to 0.1 times the time standard deviation, i.e., 40 hours, to balance the smoothness of the curve with local features. If the bandwidth is too small, the curve may contain too much noise; if it is too large, local details will be lost. This probability density curve provides a continuous probability distribution basis for subsequent risk analysis, enhancing the reliability of the degradation time prediction.

[0155] In step (7.3), the generation of the degradation risk curve aims to transform the probability density into a cumulative risk distribution, reflecting the increasing trend of degradation probability over time. A sliding window is applied to the probability density curve, with the window length set to 0.05 times the total duration. A smooth degradation risk curve is generated through convolution operations. The risk curve exhibits a stepped upward characteristic. The probability distribution of the remaining lifespan intervals is calculated through piecewise integration. These interval divisions are based on preset risk thresholds, providing a quantitative basis for risk level assessment. Compared to directly using probability density, the risk curve more intuitively reflects the degradation dynamics over time, providing a clear reference for maintenance decisions.

[0156] In step (7.4), dynamic programming is used to determine the optimal cable replacement time, taking into account equipment costs, maintenance costs, and degradation risk costs. A three-dimensional state transition matrix is ​​constructed, with state variables including time, degradation level, and maintenance actions. The cost function covers replacement costs, power outage losses, and risk costs. The optimal replacement time is determined by solving the forward recursion algorithm. Frequency domain feature evaluation is based on 15 parameters, including loss factor distribution, harmonic characteristics, and polarization characteristics. A comprehensive degradation score is calculated using a random forest classifier.

[0157] The classifier uses the information gain criterion to allocate feature weights: loss factor with a weight of 0.3, harmonic features with a weight of 0.25, polarization features with a weight of 0.2, and other features with a combined weight of 0.25. The scoring uses a percentage system, dividing maintenance into four levels: above 90 points is normal, 75-90 points is mild degradation, 60-75 points is moderate degradation, and below 60 points is severe degradation. Based on the scores, maintenance level data is generated, and a work list is created, including replacement priorities, monitoring frequency, and maintenance plans.

[0158] In step (7) of this invention, an insulation degradation assessment report based on frequency domain characteristics is generated by integrating the degradation risk curve, remaining life probability range, and maintenance recommendations. This method, through the combination of Monte Carlo simulation and dynamic programming, provides quantitative risk assessment and optimization decision support, significantly improving the scientific and economic aspects of cable maintenance, and is particularly suitable for long-term high-voltage cable systems.

Claims

1. A method for evaluating a degree of deterioration of cable insulation based on frequency domain feature analysis, characterized by, The method comprises the following steps: (1) Collecting the electric field intensity and moisture concentration in the cable operation environment, combining the water tree growth data of the laboratory accelerated aging experiment, generating low-frequency impedance characteristic data containing water tree growth stage annotation, and obtaining initial impedance characteristic distribution; (2) According to the initial impedance characteristic distribution, the dielectric response spectrum of the cable insulation material is extracted, the width and height of the relaxation polarization peak are calculated through the dielectric response spectrum, and the width and height of the relaxation polarization peak are obtained; (3) If the width and height of the relaxation polarization peak exceed the preset peak threshold, the dielectric response spectrum is decomposed in the frequency domain, the frequency and amplitude of the harmonic component change are extracted, and a harmonic component feature set is obtained; (4) The principal component analysis method is used to reduce the dimension of the frequency and amplitude of the harmonic component change, and the frequency distribution principal component of the insulation loss factor is extracted, and the loss factor distribution feature is obtained; (5) According to the loss factor distribution feature, based on the support vector regressor, the width and height of the relaxation polarization peak and the harmonic component feature set are input, and the length and density of the water tree are output, so as to predict the water tree growth stage, and the stage includes the initial stage, the middle stage and the later stage; (6) If the growth stage prediction result is in the middle and later stages of the water tree growth, the characteristic parameters of the microstructure evolution are calculated, the insulation penetration time is predicted, and the penetration time estimation value is obtained; (7) The penetration time estimation value is analyzed to analyze the degradation probability of different cable insulation intervals, a degradation risk distribution curve is generated according to the degradation probability of different cable insulation intervals, the probability interval of the remaining service life of the cable is determined, an insulation degradation evaluation report based on the frequency domain feature is generated, and the quantitative degradation degree and maintenance suggestion are output.

2. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (1) comprises the following steps: (1.1) Frequency scanning is performed on the cable insulation layer to obtain first original impedance amplitude data and first phase angle data; (1.2) Wavelet transform processing is performed according to the first original impedance amplitude data and the first phase angle data to obtain a first impedance characteristic value data set; (1.3) Accelerated aging experiment is carried out on the cable insulation layer in the temperature gradient change interval, and a second impedance characteristic value data set and a water tree length data are recorded by a low-frequency impedance analyzer; (1.4) A random forest algorithm is used to establish a feature mapping relationship for the first impedance characteristic value data set and the second impedance characteristic value data set, and if the feature mapping relationship is within a preset threshold interval, it is determined that the cable insulation layer is in the water tree initiation period, and a support vector machine is used to classify the impedance characteristic value according to the water tree growth stage, and an initial impedance characteristic distribution is obtained.

3. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (2) comprises the following steps: (2.1) Receiving cable insulation layer impedance data, performing frequency domain decomposition on the impedance data by Fourier transform to obtain an original dielectric response spectrum; (2.2) According to the original dielectric response spectrum, wavelet denoising processing is performed, and a denoising dielectric response spectrum is obtained by extracting a dielectric loss factor and a dielectric loss angle; (2.3) The relaxation time constant distribution is extracted for the denoising dielectric response spectrum, the frequency domain response envelope is obtained by Hilbert transform, and a relaxation time spectrum is generated. (2.4) If the local maximum point exists in the relaxation time spectrum, fitting the local maximum point to obtain a relaxation polarization peak profile curve, calculating the half-width of the relaxation polarization peak profile curve, and generating the relaxation polarization peak width and height.

4. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (3) comprises the following steps: (3.1) According to the relaxation polarization peak width and peak height in the dielectric response spectrum, judging against the preset peak threshold value, if the peak width exceeds the preset width threshold value or the peak height exceeds the preset height threshold value, an abnormal response spectrum is obtained; (3.2) Four-layer frequency domain decomposition is performed on the abnormal response spectrum, and a multi-scale frequency domain spectrum is obtained through orthogonal transformation; (3.3) Extracting a frequency band signal from the multi-scale frequency domain spectrum, calculating the instantaneous frequency and amplitude by using Hilbert transform to obtain a frequency modulation feature sequence; (3.4) Random forest classification is performed on the frequency modulation feature sequence, and the harmonic component parameters are extracted to generate an initial harmonic feature set, and the initial harmonic feature set is smoothed by using a radial basis kernel support vector regression to obtain a harmonic component feature set.

5. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (4) comprises the following steps: (4.1) Constructing a loss factor frequency response matrix according to the harmonic frequency parameters and amplitude parameters, and obtaining a normalized loss matrix by using the mean subtraction and standard deviation division method; (4.2) Calculating the covariance matrix of the normalized loss matrix to obtain eigenvalues and eigenvectors, selecting the number of principal components according to the cumulative contribution rate of the eigenvalues exceeding a preset threshold value, and obtaining reduced loss data through orthogonal transformation; (4.3) Processing the reduced loss data by using a five-layer feedforward neural network, and generating a loss feature vector, wherein the number of neural network hidden layer neurons is twice the input dimension; (4.4) Nonlinear mapping of the loss feature vector is performed by using a radial basis kernel function support vector regressor, and the main loss features are selected according to the feature component contribution rate to the total variance exceeding a preset threshold value, and a loss factor distribution feature is generated.

6. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (5) comprises the following steps: (5.1) According to the loss factor distribution feature, the relaxation polarization peak width parameter and the peak height parameter, and the harmonic feature set, the feature standardization is performed by using the mean subtraction and standard deviation division method to obtain a normalized feature matrix; (5.2) The principal component analysis method is used for dimension reduction processing on the normalized feature matrix to obtain feature components with a cumulative contribution rate exceeding a preset threshold value, and a reduced feature data set is obtained; (5.3) According to the water tree length value and the water tree density value in the reduced feature data set, the normalized target value is obtained by using the maximum and minimum value normalization processing, the k-means clustering algorithm is used for three-class division of the normalized target value, and the medium-term lower threshold value and the medium-term upper threshold value and the late-term lower threshold value are obtained; (5.4) The radial basis kernel function support vector regressor is used to establish the mapping relationship between the reduced feature data and the water tree parameters, if the water tree length value is less than the medium-term lower threshold value, it is determined as the initial stage, if the water tree length value is located in the medium-term threshold interval, it is determined as the medium-term, and if the water tree length value is greater than the late-term lower threshold value, it is determined as the late stage.

7. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (6) comprises the following steps: (6.1) judging the growth stage prediction result, if the prediction result is in the interval from the upper limit of the medium stage threshold to the late stage threshold, using a feature extraction module to obtain a water tree structure feature sequence; (6.2) performing Hilbert transform processing on the water tree structure feature sequence, and using an instantaneous amplitude and phase information extractor to obtain a water tree evolution feature sequence in a preset frequency band; (6.3) performing preset layer number wavelet decomposition on the water tree evolution feature sequence, and using a trend component and detail component extractor to obtain a multi-scale feature sequence; (6.4) fitting the multi-scale feature sequence using a hyperbolic function, establishing an evolution rate parameter to breakdown time mapping function through a random forest regressor, and obtaining an insulation penetration time estimation value according to the time mapping function.

8. The method for evaluating the degree of insulation deterioration of a cable based on frequency domain feature analysis according to claim 1, characterized by, The step (7) comprises the following steps: (7.1) using a normal distribution random number generator to obtain an insulation interval sampling sequence, and generating initial probability distribution data according to the sampling sequence; (7.2) constructing a probability density function through a kernel density estimation method according to the initial probability distribution data, and obtaining a continuous probability curve; (7.3) performing convolution operation on the continuous probability curve, obtaining a degradation risk curve, and dividing residual life interval data according to the degradation risk curve; (7.4) calculating a degradation degree score for the residual life interval data, performing random forest classification on frequency domain feature data, setting a feature weight vector, calculating a comprehensive degradation degree score, generating maintenance level data according to the degradation degree score, obtaining maintenance suggestions through the maintenance level data, and generating an insulation degradation evaluation report based on frequency domain features.

Images