Crosslinked cable health state evaluation system and method

Abstract

The invention relates to the technical field of crosslinked cables, and discloses a crosslinked cable health state evaluation system and method, and the system obtains an original PD signal flow through an HFCT sensor, and then carries out the high-frequency sampling to extract an original PD pulse sequence. For the sequences, an empirical mode decomposition method is adopted for denoising, so that the accuracy of subsequent analysis data is ensured. Thirdly, generating a PRPD map, extracting statistical moment features, shape description features and texture features from the PRPD map, and forming a multi-dimensional feature set; in order to further optimize feature information, the features are processed by using a principal component analysis method, and key features are extracted. And finally, inputting the data fused with the multiple key features into an evaluation module based on a support vector machine model to realize accurate evaluation of the health state of the crosslinked cable. Therefore, the problems of noise interference and single characteristic in the traditional detection means are effectively solved, and the accuracy and reliability of cable health state evaluation are improved.

Description

Technical Field

[0001] This application relates to the technical field of cross-linked cables, and more specifically, to a cross-linked cable health status assessment system and method. Background Art

[0002] In modern power systems, cross-linked cables, as one of the key power transmission and distribution equipment, their operating status directly affects the safety and reliability of power supply. With the continuous expansion of the power grid scale and the progress of technology, the accurate assessment of the health status of cross-linked cables has become particularly important. However, traditional detection methods such as off-line testing and regular maintenance are not only time-consuming and laborious, but also often unable to detect potential problems in a timely manner, resulting in handling after a failure occurs, bringing great risks to the stable operation of the power system.

[0003] In the prior art, partial discharge (PD) detection, as an effective on-line monitoring means, has been widely used in the early diagnosis of cable insulation defects. However, traditional PD analysis methods have many deficiencies. For example, the original PD signals directly obtained from sensors usually contain a large amount of noise, which seriously affects the accuracy of subsequent feature extraction; in addition, the PD signals corresponding to different types of cable faults have high complexity and diversity, making it difficult for a single feature or a simple combination of features to comprehensively reflect the actual health status of the cable. Especially in the face of complex working conditions, traditional methods are prone to misjudgment or missed judgment.

[0004] Therefore, an optimized cross-linked cable health status assessment scheme is expected. Summary of the Invention

[0005] To solve the above technical problems, this application is proposed. The embodiments of this application provide a cross-linked cable health status assessment system and method, which overcome the problems of noise interference and single feature in traditional detection means, and improve the accuracy and reliability of cable health status assessment.

[0006] In a first aspect, a method for evaluating the health state of a cross-linked cable is provided, including: obtaining the original PD signal stream collected by an HFCT sensor; performing high-frequency sampling on the original PD signal stream to obtain an original PD pulse sequence; denoising the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; generating a PRPD map based on the denoised PD pulse sequence; extracting statistical moment features, shape description features, and texture features from the PRPD map to obtain PRPD map statistical moment features, PRPD map shape description features, and PRPD map texture features; performing principal component analysis on the PRPD map statistical moment features, the PRPD map shape description features, and the PRPD map texture features to obtain PRPD map statistical moment principal component features, PRPD map shape description principal component features, and PRPD map texture principal component features; fusing the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features to obtain multi-dimensional fused PD features; inputting the multi-dimensional fused PD features into a cross-linked cable health state evaluation module based on an SVM model to obtain a health state evaluation result.

[0007] In combination with the first aspect, in a possible implementation, the statistical moment features include average phase, phase standard deviation, phase skewness, phase kurtosis, average amplitude, amplitude standard deviation, amplitude skewness, and amplitude kurtosis.

[0008] In combination with the first aspect, in a possible implementation, the PRPD map shape description features include Hu invariant moments, contour length, area, circularity, aspect ratio, convexity, eccentricity, Fourier descriptors, phase symmetry index, number of peaks / clusters, peak positions, peak intensity / density, and fractal dimension.

[0009] In combination with the first aspect, in a possible implementation, the PRPD map texture features include energy / second moment of the angle, contrast, correlation, variance, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, and difference entropy.

[0010] In combination with the first aspect, in a possible implementation manner, the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD spectrogram are fused to obtain multi-dimensional fused PD features, including: respectively inputting the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD spectrogram into a feature self-attention analysis network to obtain a first self-attention fusion weight, a second self-attention fusion weight, and a third self-attention fusion weight; based on the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight, performing significance adjustment on the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD spectrogram to obtain significance-modulated statistical moment principal component features of the PRPD spectrogram, significance-modulated shape description principal component features of the PRPD spectrogram, and significance-modulated texture principal component features of the PRPD spectrogram; and concatenating the significance-modulated statistical moment principal component features, the significance-modulated shape description principal component features, and the significance-modulated texture principal component features of the PRPD spectrogram to obtain the multi-dimensional fused PD features.

[0011] In combination with the first aspect, in a possible implementation manner, respectively inputting the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD spectrogram into a feature self-attention analysis network to obtain a first self-attention fusion weight, a second self-attention fusion weight, and a third self-attention fusion weight, including: ; where is a pre-trained reference weight vector, represents the transpose of a vector, represents matrix multiplication, represents the th feature among the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD spectrogram, represents the corresponding weight matrix, represents the corresponding feature length, represents an activation function, represents the th weight among the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight, where the pre-trained reference weight vector and the weight matrix are obtained through training.

[0012] In combination with the first aspect, in a possible implementation manner, it further includes: correcting the weight matrix, including: calculating third-order probability distribution basis functions based on orthogonal bases for the weight matrix to obtain a first-order correction matrix, a second-order correction matrix, and a third-order correction matrix; performing low-order activation fusion based on high-order projection stability on the first-order correction matrix, the second-order correction matrix, and the third-order correction matrix to obtain a fusion correction matrix; and performing dot multiplication correction on the weight matrix with the fusion correction matrix to obtain a corrected weight matrix.

[0013] In the second aspect, a cross-linked cable health status evaluation system is provided for performing the above-mentioned cross-linked cable health status evaluation method, and is characterized by including: a signal flow acquisition module for acquiring an original PD signal flow collected by an HFCT sensor; a high-frequency sampling module for performing high-frequency sampling on the original PD signal flow to obtain an original PD pulse sequence; a pulse sequence denoising module for denoising the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; a spectrogram generation module for generating a PRPD spectrogram based on the denoised PD pulse sequence; a feature extraction module for extracting statistical moment features, shape description features, and texture features from the PRPD spectrogram to obtain PRPD spectrogram statistical moment features, PRPD spectrogram shape description features, and PRPD spectrogram texture features; a principal component analysis module for performing principal component analysis on the PRPD spectrogram statistical moment features, the PRPD spectrogram shape description features, and the PRPD spectrogram texture features to obtain PRPD spectrogram statistical moment principal component features, PRPD spectrogram shape description principal component features, and PRPD spectrogram texture principal component features; a multi-dimensional feature fusion module for fusing the PRPD spectrogram statistical moment principal component features, the PRPD spectrogram shape description principal component features, and the PRPD spectrogram texture principal component features to obtain multi-dimensional fusion PD features; and an evaluation result generation module for inputting the multi-dimensional fusion PD features into a cross-linked cable health status evaluation module based on an SVM model to obtain a health status evaluation result.

[0014] Compared with the prior art, the cross-linked cable health status evaluation system and method provided by the present application obtain the original PD signal stream through an HFCT sensor, and then perform high-frequency sampling to extract the original PD pulse sequence. For these sequences, the empirical mode decomposition method is used for denoising to ensure the accuracy of the subsequent analysis data. Then, a PRPD map is generated, and statistical moment features, shape description features, and texture features are extracted from it to form a multi-dimensional feature set. To further optimize the feature information, the principal component analysis method is used to process the above features and extract the key features. Finally, the data integrating various key features is input into the evaluation module based on the SVM model to achieve the accurate evaluation of the cross-linked cable health status. In this way, the problems of noise interference and single feature in traditional detection means are effectively overcome, and the accuracy and reliability of cable health status evaluation are improved. Through this method, not only the cable fault diagnosis ability is enhanced, but also a solid guarantee is provided for the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other objects, features, and advantages of the present application will become more obvious. The accompanying drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0016] Figure 1 FIG. illustrates a schematic flow chart of a cross-linked cable health status evaluation method according to an embodiment of the present application.

[0017] Figure 2 FIG. illustrates a schematic diagram of data flow of a cross-linked cable health status evaluation method according to an embodiment of the present application.

[0018] Figure 3 FIG. illustrates a schematic flow chart of S7 in a cross-linked cable health status evaluation method according to an embodiment of the present application.

[0019] Figure 4 FIG. illustrates a schematic block diagram of a cross-linked cable health status evaluation system according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] Hereinafter, exemplary embodiments of the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0021] Figure 1 FIG. illustrates a schematic flow chart of a cross-linked cable health status evaluation method according to an embodiment of the present application.Figure 2 The figure shows a schematic diagram of data flow of a cross-linked cable health status evaluation method according to an embodiment of the present application. As Figure 1 and Figure 2 shown, the present application provides a cross-linked cable health status evaluation method, including: S1, obtaining an original PD (partial discharge) signal stream collected by an HFCT sensor; S2, performing high-frequency sampling on the original PD signal stream to obtain an original PD pulse sequence; S3, denoising the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; S4, generating a PRPD (phase-resolved partial discharge) map based on the denoised PD pulse sequence; S5, extracting statistical moment features, shape description features, and texture features from the PRPD map to obtain PRPD map statistical moment features, PRPD map shape description features, and PRPD map texture features; S6, performing principal component analysis on the PRPD map statistical moment features, the PRPD map shape description features, and the PRPD map texture features to obtain PRPD map statistical moment principal component features, PRPD map shape description principal component features, and PRPD map texture principal component features; S7, fusing the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features to obtain multi-dimensional fusion PD features; S8, inputting the multi-dimensional fusion PD features into a cross-linked cable health status evaluation module based on an SVM (support vector machine) model to obtain a health status evaluation result.

[0022] Specifically, in step S1, an original PD signal stream collected by an HFCT sensor is obtained. It should be understood that the partial discharge phenomenon is an important sign of the aging or damage of power cable insulation materials. When there are defects inside the cable, such as air gaps, cracks, or other insulation weaknesses, partial discharges may occur under high voltage. These discharge activities will release weak but characteristic electromagnetic signals. By detecting and analyzing these signals, the health status of the cable can be evaluated. The HFCT sensor has become an ideal choice for monitoring partial discharge signals because it can efficiently capture current changes in the high-frequency range. At the same time, the HFCT sensor has a non-invasive installation method, which means it can be installed and data collected without affecting the normal operation of the cable. This is especially important for continuous online monitoring because it allows real-time monitoring of the cable's health status without interrupting the power supply service, timely discovery of potential problems, and taking preventive measures to avoid the occurrence of sudden failures.

[0023] In a specific embodiment, the HFCT sensor is installed on the ground wire loop at the cable grounding end or the intermediate joint, or can be placed at a suitable position outside the cable sheath to ensure that it can effectively sense the current mutation caused by internal discharge of the cable. During installation, it is necessary to ensure that the magnetic core of the sensor completely surrounds the current-carrying conductor to form a closed magnetic circuit, and maintain the mechanical stability and electromagnetic shielding effectiveness of the connection between the sensor and the cable to reduce external interference.

[0024] Specifically, in step S2, high-frequency sampling is performed on the original PD signal stream to obtain an original PD pulse sequence. It should be understood that the partial discharge signal is essentially a high-frequency current pulse with an extremely short duration, and its frequency components cover a wide range, usually reaching dozens of megahertz, or even exceeding one hundred megahertz. If the sampling frequency is insufficient, it is easy to cause signal aliasing and loss of spectral information, resulting in distortion of pulse characteristics and seriously affecting subsequent signal discrimination and health status assessment. Therefore, the use of high-frequency sampling technology can capture the transient changes of the signal with sufficient delicate time resolution to ensure the true restoration of the discharge signal and the integrity of its characteristics.

[0025] Specifically, the sampling frequency should be reasonably selected according to the bandwidth characteristics of the partial discharge signal to meet the Nyquist sampling theorem, that is, the sampling frequency should be at least twice the highest frequency of the signal. For example, for a partial discharge signal with a highest frequency component of up to 100 MHz, the sampling frequency is usually set at 200 MHz and above. The core part of the sampling device is a high-speed analog-to-digital converter (ADC), and its resolution and sampling rate directly determine the quality of the sampled data. A high-resolution ADC can capture tiny voltage changes, and high-speed sampling ensures the complete capture of continuous pulses to form an accurate digital signal sequence.

[0026] Before sampling, the original analog signal needs to undergo analog preprocessing, including using a band-pass filter to filter out low-frequency interferences such as power frequency and its harmonics, and at the same time suppressing useless frequencies outside the detection frequency band. This preprocessing not only improves the signal-to-noise ratio but also reduces the complexity of subsequent digital signal processing. In actual operation, the filter parameters are designed to retain the effective frequency band of the partial discharge signal, such as 30 MHz to 90 MHz, and effectively filter 50 Hz power frequency and its harmonics. The filtered analog signal is sent to the ADC unit and converted into a discrete digital signal under the drive of a high-speed sampling model, and the output is a set of amplitude sampling points arranged in chronological order, constituting the original PD pulse sequence.

[0027] Specifically, in step S3, the original PD pulse sequence is denoised based on empirical mode decomposition to obtain a denoised PD pulse sequence. It should be understood that partial discharge signals are a typical type of non-linear and non-stationary high-frequency transient impact signals, with complex and variable characteristics, weak pulse amplitudes and extremely short durations, and are extremely easily masked by environmental noise, power frequency interference, inherent noise of instruments and equipment, and various random electromagnetic interferences, resulting in a significant increase in the complexity of signal identification and diagnosis. Traditional linear filtering methods are difficult to effectively distinguish the overlapping frequency bands of signals and noise, and often result in the loss and distortion of signal characteristics due to fixed filtering parameters. In this context, empirical mode decomposition technology, due to its data-driven and adaptive signal decomposition mechanism, has become an ideal tool for processing and removing noise interference in partial discharge pulse sequences.

[0028] Specifically, the empirical mode decomposition method is based on time series data, without the need for prior model information, and automatically analyzes the internal time-frequency structure of the signal. It can decompose a complex signal into a set of intrinsic mode functions (IMFs), and each IMF corresponds to the oscillation mode of the signal at a specific scale. These IMFs are arranged in descending order of frequency. High-frequency IMFs mostly represent details and noise components, while low-frequency IMFs correspond to more stable and application-related signal characteristics. This process is based on local extreme points, and by constructing upper and lower envelope lines, calculating the local mean, and performing repeated "screening" to ensure that each IMF meets specific conditions. The core advantages of empirical mode decomposition lie in its adaptive characteristics and multi-level decomposition method, enabling it to effectively handle the non-steady characteristics and multi-scale complexity of partial discharge signals.

[0029] In a specific embodiment, the process of implementing EMD denoising on the original PD pulse sequence includes: starting from the collected high-frequency sampled numerical signal, first identifying all local maximum and minimum points in the signal sequence, and respectively generating the upper and lower envelope lines of the signal through interpolation methods. Subsequently, calculate the mean curve of the envelope lines, subtract the mean from the original signal to obtain a candidate IMF. Through repeated iterative "screening" processes until the definition conditions of the IMF are met, that is, the number of local extreme points and zero-crossing points is close or the same, and the mean curve is smooth enough. The first IMF is the oscillation component with the highest frequency in the signal, that is, usually containing more random noise. After peeling off the first IMF, repeat the above process for the remaining signal, and successively obtain subsequent IMFs until the remaining signal is a monotonic or constant function.

[0030] Specifically, in step S4, a PRPD pattern is generated based on the denoised PD pulse sequence. It should be understood that since the occurrence of the partial discharge phenomenon is directly related to the phase of the AC voltage borne by the cable, each pulse not only has amplitude information but also precisely corresponds to a certain determined phase angle of the AC voltage. The PRPD pattern maps the occurrence phase of each partial discharge pulse and the signal intensity carried by its amplitude to a two-dimensional plane, with the phase as the abscissa and the amplitude as the ordinate, and establishes a statistical histogram or scatter plot form for all the collected pulses to form a graphical phase-amplitude distribution map. This pattern reveals the law of the change of partial discharge with the voltage cycle, showing the concentrated distribution of discharge activities in specific voltage intervals and the typical pulse characteristics of different discharge types.

[0031] In a specific embodiment, generating a PRPD pattern based on the denoised PD pulse sequence includes: extracting two sets of information, namely phase and amplitude, from the preprocessed data, dividing the subsequent pulse events into several equally spaced phase intervals according to the voltage phase (commonly, for example, dividing 0° to 360° into 180 2° intervals), and counting the number of pulse events with amplitude distribution in each interval, so as to reflect the density and intensity of the discharge activity. At the specific algorithm implementation level, the system traverses the denoised PD pulse sequence, reads the phase angle of each pulse point, and locates it to the corresponding phase interval; at the same time, collects the amplitude data and performs maximum truncation or normalization processing to eliminate the influence of abnormal extreme values on the statistics. Subsequently, according to the specified amplitude resolution, the amplitude is discretized into multiple amplitude levels, and the pulse count in each combination of phase interval and amplitude level is counted. This two-dimensional histogram matrix is the essential form of the PRPD pattern, usually presented as a grayscale image or a color image, and each pixel value represents the number of discharge pulses under a specific phase-amplitude combination.

[0032] Specifically, in step S5, statistical moment features, shape description features, and texture features are extracted from the PRPD pattern to obtain PRPD pattern statistical moment features, PRPD pattern shape description features, and PRPD pattern texture features. It should be understood that the PRPD pattern itself is a statistical distribution expression form of the partial discharge signal in the two-dimensional space of the grid voltage phase and discharge amplitude, and is a comprehensive manifestation of the cable partial discharge characteristics in terms of time-frequency distribution, pulse intensity, and spatial form. Since there are various types of insulation defects caused by partial discharge and the discharge activities show complex and irregular spatial distributions and dynamic changes, a single feature dimension is difficult to cover all the key information. Combining statistical features, morphological descriptions, and texture analysis can deeply reveal the microscopic dynamics and macroscopic laws of the partial discharge signal from different angles, comprehensively reflect the cable health status, and significantly improve the accuracy and robustness of diagnosis.

[0033] Specifically, the statistical moment features mainly calculate the statistics of the phase and amplitude distributions of the PRPD pattern, objectively reflecting the average performance and fluctuation degree of partial discharge within a cycle. In one embodiment, the statistical moment features include average phase, phase standard deviation, phase skewness, phase kurtosis, average amplitude, amplitude standard deviation, amplitude skewness, and amplitude kurtosis. Among them, the average phase and average amplitude reveal the concentration characteristics of discharge activities, the phase standard deviation describes the data dispersion degree, the phase skewness reflects the skewness of the distribution, the phase kurtosis reveals the sharpness of the distribution peak state, and the amplitude kurtosis represents the tail distribution characteristics.

[0034] Specifically, the shape description features analyze the spatial geometric characteristics and topological structure of the PRPD pattern graph to reveal the morphological characteristics of discharge activities in the two-dimensional phase-amplitude space. In one embodiment, the PRPD pattern shape description features include Hu invariant moments, contour length, area, circularity, aspect ratio, convexity, eccentricity, Fourier descriptors, phase symmetry index, number of peaks / clusters, peak positions, peak intensity / density, and fractal dimension. Among them, the Hu invariant moments ensure that the features are invariant to geometric transformations such as rotation, scaling, and translation, guaranteeing the generalization ability of the model. The contour length, area, etc. quantify the spatial range and morphological size of the discharge patches. Indicators such as circularity, convexity, and eccentricity reflect the symmetry and complexity of the graph, and the Fourier descriptors analyze the frequency domain components of the graph edge contour, emphasizing periodicity and structural regularity. Peak and cluster attributes such as number, position, and intensity reveal the pulse aggregation trend of partial discharge in a specific voltage phase segment. Such shape features can well capture the distribution characteristics and dynamic behavior of discharge, make up for the limitations of simple statistical features, and provide spatial dimension support for fault mode recognition.

[0035] Specifically, the texture features focus on the internal structure and local correlation of the spatial gray distribution of the PRPD pattern, and mine the spatial consistency, contrast, and complexity of the image through methods such as the gray-level co-occurrence matrix. In one embodiment, the PRPD pattern texture features include energy / second moment of the angle, contrast, correlation, variance, inverse difference, sum average, sum variance, sum entropy, entropy, difference variance, and difference entropy. Among them, indicators such as energy, contrast, and correlation measure the thickness, light and dark changes of the pattern texture and the spatial relationship between pixels; variance and inverse difference reflect the discreteness and uniformity of the texture; entropy-like indicators measure the complexity and information content of the image, and difference entropy and difference variance further capture the local gray change law. These texture features reflect the spatial organization law of partial discharge in time phase and amplitude and the random fluctuation of discharge patterns, helping to identify the random or regular components of lesion signals, and are a beneficial supplement to morphological and statistical features.

[0036] In a specific embodiment, statistical moment features, shape description features, and texture features are extracted from the PRPD map to obtain PRPD map statistical moment features, PRPD map shape description features, and PRPD map texture features, including: First, the PRPD map generated after denoising needs to be subjected to format normalization processing, including image size unification, gray level normalization, and necessary preprocessing filtering. Subsequently, mathematical statistical methods are used to calculate the statistical moment features. Generally, the PRPD data is subjected to probability density estimation according to the phase and amplitude intervals, and then the mean value, variance, skewness coefficient, and kurtosis coefficient are calculated one by one based on the definition of statistics. For the shape description features, first, binary processing is performed on the PRPD map image to identify the connected regions of the main discharge pulse regions, the contours are extracted, the area and perimeter are obtained through boundary tracing and region analysis, and then the geometric invariant moment theory and Fourier transform are used to analyze the complex morphology and structure, and the spatial characteristics and symmetry of the graph are measured. The peak and cluster features depend on the detection algorithm for local extreme points. The obviously concentrated pulse peaks are divided into clusters, and information such as the number of clusters, positions, and intensities is counted. The extraction of texture features depends on the calculation method based on the gray level co-occurrence matrix. According to the joint distribution of the statistical gray values of the pixel pairs in the image, indicators such as energy, entropy, and contrast are calculated in combination with different direction and distance parameters to quantify the texture details and structural complexity of the image.

[0037] Specifically, in step S6, principal component analysis is performed on the PRPD map statistical moment features, the PRPD map shape description features, and the PRPD map texture features to obtain PRPD map statistical moment principal component features, PRPD map shape description principal component features, and PRPD map texture principal component features. It should be understood that as a two-dimensional statistical display form of the partial discharge signal, the PRPD map realizes multi-angle characterization of the complex physical mechanism and geometric characteristics of the partial discharge through in-depth feature extraction in three dimensions of its statistical moment, shape description, and texture. The statistical moment features reflect the basic statistical attributes of the pulse distribution, such as the mean value, skewness, kurtosis, etc. The shape description features reveal the geometric shape and topological structure of the discharge signal in the spatial distribution, including invariant moments, contour features, etc. The texture features focus on the detailed texture information in the image such as the gray level co-occurrence matrix, reflecting the spatial correlation and complexity of the partial discharge map. However, these feature sets often contain a large number of dimensions, and there are linear correlations and information redundancies between the features, resulting in a large amount of redundant information in the feature space. In such a high-dimensional space, if directly used for subsequent classification and evaluation, it will not only increase the computational burden, slow down the calculation speed, but also may cause the risk of overfitting of the model and reduce the generalization ability.

[0038] Therefore, to solve the problem of feature dimensionality disaster and improve the expression efficiency of the feature space, it is a reasonable choice to perform principal component analysis on the three types of features respectively. As a classic linear dimensionality reduction technique, principal component analysis uses linear algebra methods to find the orthogonal vectors with the largest variance information in the original data, that is, the principal components, so as to achieve the optimal expression of data in the low-dimensional space. PCA can project the data onto a newly constructed uncorrelated coordinate system, making the principal components independent of each other, eliminating the multicollinearity of the original features, reducing the dimensions while retaining most of the information, effectively suppressing the influence of noise and irrelevant variables, and enhancing the discriminability of the features. At the same time, performing PCA on the statistical moment features, shape description features, and texture features separately can avoid the confusion caused by differences in scale and distribution among different feature categories, enabling each feature dimension to maximize its expression efficiency.

[0039] In a specific embodiment, principal component analysis is performed on the PRPD map statistical moment features, the PRPD map shape description features, and the PRPD map texture features to obtain the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features, including: First, an observation sample matrix is constructed for each of the three types of extracted features, where each row is an observation sample and the columns correspond to different feature indicators. Since the dimensions of the features are different, to ensure the effectiveness of the analysis, it is necessary to standardize the data. This step usually centers around the mean of each feature column, performs a mean subtraction operation, and normalizes by the standard deviation to ensure zero mean and unit variance of the data, avoiding bias in the subsequent calculation of the covariance matrix caused by feature scale differences.

[0040] After that, the covariance matrix is calculated based on the standardized data. The covariance matrix reflects the linear dependence relationship between features and is the mathematical basis for PCA. It is a symmetric real matrix, ensuring the existence of a set of orthogonal eigenvectors and corresponding real eigenvalues. The covariance matrix is calculated using eigenvalue decomposition or singular value decomposition methods to obtain the eigenvalues and their corresponding eigenvectors. The magnitude of the eigenvalue represents the scale of the data variance in the direction of the corresponding eigenvector and represents the information contribution in that direction. By selecting the eigenvector with the largest eigenvalue as the principal component basis vector, a new orthogonal coordinate system is constructed.

[0041] Specifically, in step S7, the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern are fused to obtain multi-dimensional fused PD features. It should be understood that the partial discharge signal is essentially a complex non-linear and non-stationary electrical phenomenon. Its manifestation in the PRPD pattern not only includes the overall statistical distribution trend, but also includes the morphological information of spatial geometry and the texture features with rich details. A single type of feature cannot comprehensively capture all signal laws. The statistical features reveal the global statistical characteristics of amplitude and phase, the shape description features reflect the geometric and topological structures of the pulse distribution in two-dimensional space, and the texture features deeply explore the gray-scale spatial correlation and microscopic change laws of the image. Fusing the three principal component features helps to construct a strong expression space across scales, angles, and dimensions, fundamentally enhancing the stability and discrimination of fault features, thereby improving the accuracy and reliability of cable health status assessment.

[0042] In one embodiment, as Figure 3 shown, in step S7, fusing the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern to obtain multi-dimensional fused PD features includes: S71, respectively inputting the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern into the feature self-attention analysis network to obtain the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight; S72, based on the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight, performing significance adjustment on the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern to obtain the significance-modulated statistical moment principal component features of the PRPD pattern, the significance-modulated shape description principal component features of the PRPD pattern, and the significance-modulated texture principal component features of the PRPD pattern; S73, cascading the significance-modulated statistical moment principal component features of the PRPD pattern, the significance-modulated shape description principal component features of the PRPD pattern, and the significance-modulated texture principal component features of the PRPD pattern to obtain the multi-dimensional fused PD features.

[0043] Specifically, considering that the PRPD pattern of partial discharge in cross-linked cables reveals information in multiple characteristic dimensions, although these characteristics respectively capture different attributes of the partial discharge signal, there are complex non-linear relationships and interdependencies among them. Statistical moment characteristics reflect the global distribution and basic statistical laws of the signal, shape description characteristics reveal the geometric structure information of the discharge image, and texture characteristics reflect the subtle gray-scale and spatial correlation changes. Since there may be overlaps in probability distributions and high-order complex couplings among these characteristics during the fusion process, if simply combined, it will limit the effectiveness of feature utilization and even lead to information redundancy and interference. The feature self-attention analysis network inputs the statistical moment principal components, shape description principal components, and texture principal components of the PRPD pattern respectively, and adaptively generates their respective self-attention fusion weights, which helps to dynamically quantify and adjust the contribution degree of each characteristic to the final fused feature expression. This approach enables the network to flexibly capture the intrinsic significance and importance differences of each characteristic, avoiding the effective information of some characteristics being masked by strong correlations or common noises, thereby enhancing the discriminability and stability of the fused features.

[0044] In one embodiment, inputting the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern into the feature self-attention analysis network respectively to obtain the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight, including: ; where is the pre-trained reference weight vector, represents the transpose of the vector, represents matrix multiplication, represents the th feature among the statistical moment principal component features of the PRPD pattern, the shape description principal component features of the PRPD pattern, and the texture principal component features of the PRPD pattern, represents the corresponding weight matrix, represents the corresponding feature length, represents the activation function, represents the th weight among the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight. Here, it should be understood by those skilled in the art that the above pre-trained reference weight vector and the corresponding weight matrix are obtained by continuous adjustment and optimization during the training process, that is, based on a large number of labeled sample data, the weight parameters are continuously adjusted through the backpropagation algorithm, combined with regularization constraints to avoid overfitting.

[0045] Specifically, when obtaining the self-attention fusion weights for the statistical moment principal component features, the shape description principal component features, and the texture principal component features of the PRPD map, considering that the statistical moment features, shape description features, and texture features of the PRPD map have their respective probability distribution reconstructions in the fusion space. Therefore, when fusing features, the non-linear associations between the high-order cumulant features of the PRPD map itself will form high-order dependence relationships, thus affecting the decoupling of the corresponding weight matrices and increasing the convergence difficulty of the self-attention fusion weights. In a preferred example, the weight matrix is corrected.

[0046] Specifically, correcting the weight matrix includes: First, for the corresponding weight matrix , perform the calculation of the third-order probability distribution basis function based on the Hermite orthogonal basis to obtain the first-order correction matrix, the second-order correction matrix, and the third-order correction matrix, expressed as: ; where represents the corresponding first-order correction matrix, represents the corresponding second-order correction matrix, represents the corresponding third-order correction matrix, represents the transpose of the vector, represents the element-wise multiplication, represents the matrix multiplication, represents the element-wise subtraction.

[0047] That is, for the probability distribution space of the statistical moment feature vector , the shape description feature vector , and the texture feature vector jointly composed, where represents the set of real numbers, represent the lengths of the statistical moment feature vector, the shape description feature vector, and the texture feature vector respectively, perform the expansion of the orthogonal basis probability density function based on the Hermite polynomial coefficients, and then perform the low-order activation fusion based on the high-order projection stability for the first-order correction matrix, the second-order correction matrix, and the third-order correction matrix to obtain the fusion correction matrix, expressed as: ; where represents the corresponding fusion correction matrix.

[0048] Thus, by taking the high-order projection as the upper bound of the change speed of the weight matrix distribution, the low-order immobility is limited, thereby enhancing the quantization degree of the probability density expansion smoothness as the order increases.

[0049] Finally, the weight matrix is dot-multiplied and corrected with the fusion correction matrix to obtain the corresponding corrected weight matrix, expressed as: ; where represents the corresponding corrected weight matrix.

[0050] In this way, the mixed-order mode smoothing fusion under the decoupling of the high-order dependence relationship of the corrected weight matrix is realized, and the corrected weight matrix is used as the weight matrix in the inference stage to calculate the self-attention fusion weight, thereby enhancing the convergence effect of the self-attention fusion weight and improving the calculation of the self-attention fusion weight .

[0051] Specifically, in step S8, the multi-dimensional fusion PD features are input into the cross-linked cable health status evaluation module based on the SVM model to obtain the health status evaluation result. It should be understood that the multi-dimensional fusion PD features map the deep information of the partial discharge signal in multiple feature dimensions such as statistical moments, shape descriptions, and textures, presenting the comprehensive performance of the internal defects of the cable. The SVM-based evaluation module makes full use of this rich feature information, combines its excellent classification performance and good generalization ability, and realizes the accurate discrimination of the cross-linked cable status.

[0052] Specifically, the multi-dimensional fusion features involve the statistical moment principal component features, the shape description principal component features, and the texture principal component features. The three respectively explain the distribution of the partial discharge pulses, the geometric space morphology, and the texture space features from different angles. The feature vector formed after fusion has both global tendency and local details, showing a higher diagnostic information carrying capacity. A single feature is difficult to fully correspond to the complex nature of the partial discharge signal, and the high-dimensional feature space constructed by the fusion features helps to reveal the internal differences between various health conditions. The SVM model, as a supervised machine learning algorithm, focuses on finding the optimal hyperplane that can maximize the distance between classes, and is very suitable for dealing with practical problems containing high-dimensional fusion features and complex class boundaries, so that cable samples of different health levels are clearly separated in the feature space.

[0053] Here, those skilled in the art should be aware that the SVM model needs to be trained before inference. During the training process, the SVM determines the hyperplane parameters that can maximize the class margin by minimizing the structural risk function. For linearly separable problems, the model directly constructs a linear separation surface using the training samples; while the partial discharge characteristic data of cross-linked cables often exhibits a non-linear distribution. The typical implementation strategy is to introduce the kernel function technique to map the data to a high-dimensional feature space to achieve linear separability. Commonly used kernel functions include the Radial Basis Function (RBF), polynomial kernel, and Sigmoid kernel, etc. The choice of the kernel function is determined by the characteristics of the actual data distribution. The kernel parameters and the penalty factor C are determined by hyperparameter tuning methods such as cross-validation and grid search to achieve the balance of the optimal generalization ability and classification accuracy. The training algorithm generally adopts the Sequential Minimal Optimization (SMO) method to efficiently solve the support vectors and decision boundaries for quadratic programming problems. During the iterative process, if a sample contributes little to the model, it is not included in the support vector set, greatly saving computational resources. The support vector machine defines the important feature space of the decision boundary by finding the support vectors closest to the hyperplane, ensuring that the model has good generalization and robust performance.

[0054] After training, the model contains a set of support vectors, kernel function parameters, and decision function coefficients, and can effectively classify the multi-dimensional fusion PD feature vectors of unknown samples. The health status evaluation module uses this SVM model to predict the multi-dimensional fusion PD features obtained in real-time or offline, and outputs the health status category of the cable. This result provides a clear diagnostic judgment for users and assists in equipment maintenance decisions.

[0055] In summary, the cross-linked cable health status evaluation method provided by this application obtains the original PD signal stream through the HFCT sensor, and then performs high-frequency sampling to extract the original PD pulse sequence. For these sequences, the empirical mode decomposition method is used for denoising to ensure the accuracy of the subsequent analysis data. Then, the PRPD pattern is generated, and the statistical moment features, shape description features, and texture features are extracted from it to form a multi-dimensional feature set. To further optimize the feature information, the principal component analysis method is used to process the above features and extract the key features. Finally, the data integrating a variety of key features is input into the evaluation module based on the support vector machine model to achieve the accurate evaluation of the cross-linked cable health status. This effectively overcomes the problems of noise interference and single feature in traditional detection means, and improves the accuracy and reliability of cable health status evaluation.

[0056] This application also provides a cross-linked cable health status evaluation system, as Figure 4As shown, the cross-linked cable health status evaluation system 400 includes: a signal flow acquisition module 410 for acquiring the original PD signal flow collected by HFCT sensors; a high-frequency sampling module 420 for performing high-frequency sampling on the original PD signal flow to obtain an original PD pulse sequence; a pulse sequence denoising module 430 for denoising the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; a spectrogram generation module 440 for generating a PRPD spectrogram based on the denoised PD pulse sequence; a feature extraction module 450 for extracting statistical moment features, shape description features, and texture features from the PRPD spectrogram to obtain PRPD spectrogram statistical moment features, PRPD spectrogram shape description features, and PRPD spectrogram texture features; a principal component analysis module 460 for performing principal component analysis on the PRPD spectrogram statistical moment features, the PRPD spectrogram shape description features, and the PRPD spectrogram texture features to obtain PRPD spectrogram statistical moment principal component features, PRPD spectrogram shape description principal component features, and PRPD spectrogram texture principal component features; a multi-dimensional feature fusion module 470 for fusing the PRPD spectrogram statistical moment principal component features, the PRPD spectrogram shape description principal component features, and the PRPD spectrogram texture principal component features to obtain multi-dimensional fusion PD features; and an evaluation result generation module 480 for inputting the multi-dimensional fusion PD features into a cross-linked cable health status evaluation module based on an SVM model to obtain a health status evaluation result.

[0057] The basic principles of the present application have been described in conjunction with specific embodiments. However, it should be noted that the advantages, advantages, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, advantages, effects, etc. are essential for each embodiment of the present application. In addition, the above-disclosed specific details are only for the purpose of illustration and facilitating understanding, and not for limitation. The above details do not limit the present application to necessarily adopt the above specific details to implement.

[0058] The block diagrams of the devices, apparatuses, equipment, and systems involved in the present application are only illustrative examples and do not intend to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open-ended words, meaning "including but not limited to", and can be used interchangeably with each other. The word "or" and "and" used herein refer to the word "and / or", and can be used interchangeably with each other unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to", and can be used interchangeably with each other.

[0059] It should also be noted that in the devices, equipment and methods of the present application, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations shall be regarded as equivalent solutions of the present application.

[0060] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0061] The above description has been presented for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the form disclosed herein. Although several example aspects and embodiments have been discussed above, those skilled in the art will recognize some of their variations, modifications, alterations, additions, and subcombinations.

Claims

1. A method for evaluating the health status of a cross-linked cable, characterized in that, Including: Obtaining the original PD signal stream collected by the HFCT sensor; Performing high-frequency sampling on the original PD signal stream to obtain an original PD pulse sequence; Denosing the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; Generating a PRPD map based on the denoised PD pulse sequence; extracting statistical moment features, shape description features, and texture features from the PRPD map to obtain PRPD map statistical moment features, PRPD map shape description features, and PRPD map texture features; Performing principal component analysis on the PRPD map statistical moment features, the PRPD map shape description features, and the PRPD map texture features to obtain PRPD map statistical moment principal component features, PRPD map shape description principal component features, and PRPD map texture principal component features; Fusing the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features to obtain multi-dimensional fusion PD features; Inputting the multi-dimensional fusion PD features into a cross-linked cable health status evaluation module based on an SVM model to obtain a health status evaluation result.

2. The cross-linked cable health state evaluation method according to claim 1, wherein, The statistical moment features include average phase, phase standard deviation, phase skewness, phase kurtosis, average amplitude, amplitude standard deviation, amplitude skewness, and amplitude kurtosis.

3. The cross-linked cable health state evaluation method according to claim 1, characterized in that The PRPD map shape description features include Hu invariant moments, contour length, area, circularity, aspect ratio, convexity, eccentricity, Fourier descriptors, phase symmetry index, number of peaks / clusters, peak positions, peak intensity / density, and fractal dimension.

4. The cross-linked cable health state evaluation method according to claim 1, wherein, The PRPD map texture features include energy / second-order moment of the angle, contrast, correlation, variance, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, and difference entropy.

5. The cross-linked cable health state evaluation method according to claim 1, characterized in that Fusing the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features to obtain multi-dimensional fusion PD features, including: respectively inputting the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features into a feature self-attention analysis network to obtain a first self-attention fusion weight, a second self-attention fusion weight, and a third self-attention fusion weight; performing significance adjustment on the PRPD map statistical moment principal component features, the PRPD map shape description principal component features, and the PRPD map texture principal component features based on the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight to obtain significance-modulated PRPD map statistical moment principal component features, significance-modulated PRPD map shape description principal component features, and significance-modulated PRPD map texture principal component features; cascading the significance-modulated PRPD map statistical moment principal component features, the significance-modulated PRPD map shape description principal component features, and the significance-modulated PRPD map texture principal component features to obtain the multi-dimensional fusion PD features.

6. The cross-linked cable health state evaluation method according to claim 5, wherein, Input the statistical moment principal component features, shape description principal component features, and texture principal component features of the PRPD map into the feature self-attention analysis network respectively to obtain the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight, including: ; where is the pre-trained reference weight vector, represents the transpose of the vector, represents matrix multiplication, represents the th feature among the statistical moment principal component features, shape description principal component features, and texture principal component features of the PRPD map, represents the corresponding weight matrix, represents the corresponding feature length, represents the activation function, represents the th weight among the first self-attention fusion weight, the second self-attention fusion weight, and the third self-attention fusion weight.

7. The cross-linked cable health state evaluation method according to claim 6, characterized in that It also includes correcting the weight matrix, including: calculating a third-order probability distribution basis function based on an orthogonal basis for the weight matrix to obtain a first-order correction matrix, a second-order correction matrix, and a third-order correction matrix; performing low-order activation fusion based on high-order projection stability on the first-order correction matrix, the second-order correction matrix, and the third-order correction matrix to obtain a fusion correction matrix; and performing dot product correction on the weight matrix with the fusion correction matrix to obtain a corrected weight matrix.

8. A cross-linked cable health status evaluation system for implementing the cross-linked cable health status evaluation method according to any one of claims 1-7, characterized in that, It includes: a signal flow acquisition module for acquiring an original PD signal flow collected by an HFCT sensor; a high-frequency sampling module for performing high-frequency sampling on the original PD signal flow to obtain an original PD pulse sequence; a pulse sequence denoising module for denoising the original PD pulse sequence based on empirical mode decomposition to obtain a denoised PD pulse sequence; a spectrogram generation module for generating a PRPD spectrogram based on the denoised PD pulse sequence; a feature extraction module for extracting statistical moment features, shape description features, and texture features from the PRPD spectrogram to obtain PRPD spectrogram statistical moment features, PRPD spectrogram shape description features, and PRPD spectrogram texture features; a principal component analysis module for performing principal component analysis on the PRPD spectrogram statistical moment features, the PRPD spectrogram shape description features, and the PRPD spectrogram texture features to obtain PRPD spectrogram statistical moment principal component features, PRPD spectrogram shape description principal component features, and PRPD spectrogram texture principal component features; a multi-dimensional feature fusion module for fusing the PRPD spectrogram statistical moment principal component features, the PRPD spectrogram shape description principal component features, and the PRPD spectrogram texture principal component features to obtain multi-dimensional fusion PD features; an evaluation result generation module for inputting the multi-dimensional fusion PD features into a cross-linked cable health status evaluation module based on an SVM model to obtain a health status evaluation result.

Images