A method and system for detecting electrical performance of a power cable

Abstract

The application relates to the technical field of power equipment state monitoring, and specifically discloses a power cable electric performance detection method and system, wherein wideband electromagnetic signals, mechanical vibration signals and temperature signals at cable monitoring points are synchronously collected; wavelet coherence coefficients among the signals are calculated through continuous wavelet transformation, and a multi-modal coupling tensor fusing amplitude and cross-modal time-frequency correlation characteristics is constructed; time slicing and high-order singular value decomposition are performed on the tensor to obtain a time-varying core tensor sequence, which is mapped into a high-dimensional manifold curve, a system state fingerprint is generated by calculating the local curvature distribution and topological invariant of the curve; and the fingerprint is input into a pre-trained defect prediction model to directly output the incubation probability and evolution stage of defects; the application breaks through the limitation of traditional methods that are not sensitive to early weak defects, and realizes early warning and accurate diagnosis of cable insulation latent defects.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a method and system for testing the electrical performance of power cables. Background Technology

[0002] As a critical carrier of electrical energy transmission, the early deterioration of the insulation of power cables (such as the emergence of electrical trees and water trees) is a major cause of power outages. Currently, the testing of cable electrical performance mainly relies on traditional methods such as partial discharge detection and dielectric loss measurement. These technologies are generally based on amplitude or phase analysis of a single type of electrical signal (such as pulse current or ultra-high frequency electromagnetic waves) and determine the insulation status through preset thresholds. However, early microscopic defects in cable insulation only generate weak signals with extremely low amplitude, broad spectrum, and no strong correlation with the power frequency period during their initiation stage. These signals are easily drowned out by on-site electromagnetic noise and equipment background noise, resulting in serious detection blind spots in existing monitoring systems.

[0003] Existing technologies lack the ability to effectively extract and analyze weak co-evolution patterns across physical modes from monitoring signals with strong noise backgrounds and multiple heterogeneous sources, which characterize the incubation process of early microscopic defects (such as submicroscopic electrical trees) in cable insulation, and to achieve automatic and accurate identification of their defect types and development stages. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for testing the electrical performance of power cables, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for testing the electrical performance of power cables includes the following steps:

[0007] S1: Synchronously acquire multi-source heterogeneous time-series signals at the power cable monitoring point, including broadband electromagnetic signals, mechanical vibration signals and temperature signals;

[0008] S2: Perform cross-modal coupling analysis on multi-source heterogeneous time-series signals, extract time-frequency domain correlation features between pairs of broadband electromagnetic signals, mechanical vibration signals and temperature signals, and generate multi-modal coupling tensors based on time-frequency domain correlation features;

[0009] S3: Perform tensor manifold expansion on the multimodal coupling tensor to reconstruct a nonlinear manifold structure that can characterize the state of multi-physics interaction, and calculate the set of local curvature distribution and topological invariants of the nonlinear manifold structure in the current high-dimensional space as the system state fingerprint.

[0010] S4: Input the system state fingerprint into the pre-trained defect prediction model. The defect prediction model is configured to directly output the prediction results of the type of micro-defects that are being gestated in the insulation medium of the power cable based on the evolution sequence of the local curvature distribution and topological invariants in the system state fingerprint. The prediction results include the probability of defect gestation and the determination of the evolution stage.

[0011] As a further aspect of the present invention: S2 specifically includes:

[0012] Continuous wavelet transform is performed on each of the broadband electromagnetic signal, mechanical vibration signal and temperature signal to obtain the time spectrum of the wavelet coefficients of each signal;

[0013] For any two different types of signals, calculate the wavelet coherence coefficients of the corresponding wavelet coefficients at multiple preset scales to obtain the signal pair coherence spectrum composed of wavelet coherence coefficients, and obtain the set of coherence coefficients between all signal pairs.

[0014] The coherence coefficients of all signal pairs are arranged and combined into a symmetric coupling matrix according to the signal type.

[0015] The coupling matrix is ​​externally productted with the original time-domain waveforms of each signal in the time dimension to generate a fourth-order tensor, which serves as the multimodal coupling tensor.

[0016] As a further aspect of the present invention: the process of obtaining the wavelet coefficients and their corresponding spectrum for each signal specifically includes:

[0017] For broadband electromagnetic signals, continuous wavelet transform is performed using wavelet basis functions with impulse oscillation attenuation characteristics;

[0018] For mechanical vibration signals, continuous wavelet transform is performed using symmetric wavelet basis functions with zero phase shift characteristics;

[0019] For temperature signals, the local instantaneous frequencies of the temperature signal sequence are first calculated, and the analysis scale range of the continuous wavelet transform is adjusted based on the local instantaneous frequencies before the transform is performed.

[0020] The transformed wavelet coefficient sequences are reconstructed into three-dimensional time-spectrum structures containing time, scale, and amplitude dimensions, which are used as the time-spectrum of wavelet coefficients.

[0021] As a further aspect of the present invention: the process of acquiring the system status fingerprint is as follows:

[0022] Tensor fibers are expanded along the time modes of the multimodal coupling tensor to obtain a series of high-order tensor slices. High-order singular value decomposition is performed on each tensor slice to extract core tensors and form a time-varying core tensor sequence.

[0023] Each core tensor in the time-varying core tensor sequence is mapped to a point on an abstract manifold, and the entire sequence is reconstructed into a discrete curve embedded in a high-dimensional space by calculating the geodesic connectivity between adjacent points, serving as a nonlinear manifold structure.

[0024] The rate of change of the angle between the tangent space and the normal space fitted by the local neighborhood point set at each point on the discrete curve is calculated to obtain the local curvature distribution. At the same time, the connectivity and hole number features of the discrete curve as a whole under continuous deformation are extracted as topological invariants.

[0025] The local curvature distribution and the topological invariant are spliced ​​and normalized in chronological order to form a system state fingerprint.

[0026] As a further aspect of the present invention: the extraction of core tensors to construct a time-varying core tensor sequence specifically includes:

[0027] Based on the preset time window length and sliding step size, slide slices are performed along the time dimension of the multimodal coupling tensor to obtain a series of tensor slices with fixed time spans;

[0028] For each tensor slice, the decomposition rank estimate of each mode is pre-set based on the physical meaning of different non-temporal modes of the tensor slice, and this is used as a constraint condition.

[0029] Under the constraint of the decomposition rank estimation of each mode, a higher-order singular value decomposition is performed on each tensor slice, and only the number of orthogonal components determined by the decomposition rank estimation are retained to calculate the dimensionality-reduced core tensor.

[0030] Arrange the dimensionality-reduced core tensors corresponding to all tensor slices in chronological order according to the time windows corresponding to the core tensors to form a time-varying core tensor sequence.

[0031] As a further aspect of the present invention: the calculation of the dimensionality-reduced core tensor specifically includes:

[0032] A preliminary decomposition is performed on each non-temporal mode matrix of the tensor slice to obtain the initial set of orthogonal components corresponding to each mode;

[0033] Calculate the energy concentration of each component in the initial orthogonal component set, and simultaneously calculate the intermodal coupling degree of the corresponding components between different modes;

[0034] Based on the energy concentration and the intermodal coupling degree, all components are comprehensively evaluated, and initial orthogonal components that meet the preset comprehensive evaluation threshold are selected to form a set of components to be refined.

[0035] Using the set of components to be refined as the initial value for iteration, iterative refinement is performed under the constraint of decomposition rank estimation until convergence. The final refined orthogonal components are truncated according to the number of decomposition rank estimates, and the core tensor after dimensionality reduction is calculated.

[0036] As a further aspect of the present invention: S4 specifically includes:

[0037] The local curvature distribution and topological invariants contained in the system state fingerprint are input into a temporal attention network in chronological order. The temporal attention network generates time-varying temporal attention weights through self-correlation calculation.

[0038] By utilizing temporal attention weights to weight and focus on the temporal features of the input and fuse them with context, the key evolutionary features most relevant to the defect gestation process are extracted.

[0039] Key evolutionary features are input into a sequence-to-sequence mapping network. The mapping network maps continuous temporal features into a discrete state transition path and simultaneously calculates the matching confidence of the state transition path with each pattern in a predefined library of multiple typical defect evolutionary patterns.

[0040] The highest matching confidence is output as the fertility probability, and the pattern stage corresponding to the end of the state transition path is output as the evolution stage determination result.

[0041] As a further aspect of the present invention: the calculation process of the matching confidence is as follows:

[0042] The key evolutionary feature sequences are encoded point by point using an encoding network, and the feature vector at each time point is converted into a high-dimensional state code;

[0043] The high-dimensional state-coded sequence is dynamically time-warped and matched with each reference path in the typical defect evolution pattern library to calculate the minimum cumulative matching cost from the start point of the sequence to the current point.

[0044] Based on the minimum cumulative matching cost, and with the imposition of physical evolution constraints prohibiting path backtracking and state transitions, a path with the lowest global matching cost is determined through backtracking search, which serves as the discrete state transition path.

[0045] The matching confidence is calculated by combining the final normalized distance between the path with the lowest global matching cost and the corresponding reference path, as well as the proportion of unmatched feature points during path alignment.

[0046] A power cable electrical performance testing system, comprising:

[0047] A multi-source signal synchronous acquisition module is used to synchronously acquire multi-source heterogeneous time-series signals at power cable monitoring points. The multi-source heterogeneous time-series signals include broadband electromagnetic signals, mechanical vibration signals, and temperature signals.

[0048] The cross-modal coupling feature extraction module is used to perform cross-modal coupling analysis on multi-source heterogeneous time-series signals, extract the time-frequency domain correlation features between pairs of broadband electromagnetic signals, mechanical vibration signals and temperature signals, and generate multi-modal coupling tensors based on the time-frequency domain correlation features;

[0049] The tensor manifold analysis and fingerprint generation module is used to expand the multimodal coupled tensor into a nonlinear manifold structure that can characterize the state of multi-physics interaction. It also calculates the set of local curvature distribution and topological invariants of the nonlinear manifold structure in the current high-dimensional space as the system state fingerprint.

[0050] The defect prediction and assessment module is used to input the system state fingerprint into the pre-trained defect prediction model. The defect prediction model is configured to directly output the prediction results of the type of micro-defects that are being gestated in the insulation medium of the power cable based on the evolution sequence of the local curvature distribution and topological invariants in the system state fingerprint. The prediction results include the probability of defect gestation and the determination of the evolution stage.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention simultaneously collects and fuses three heterogeneous physical signals: electromagnetic, vibration, and temperature. By calculating their time-frequency domain coherence, a coupling tensor is constructed, and manifold learning is further used to extract their essential geometric and topological features (such as curvature distribution). This method can keenly capture the cooperative change patterns between multiple physical fields caused by early defects. Even if the change of a single signal does not reach the traditional threshold, the deviation of its comprehensive pattern can be effectively detected, thereby advancing defect identification from "passive response to obvious faults" to the "proactive early warning of potential risks" stage, greatly enhancing the preventive maintenance capability of the power grid.

[0053] (2) This invention constructs a pattern library containing various typical defect evolution paths, and uses an attention network to adaptively focus on key temporal features most relevant to defect development. Then, a dynamic time warping algorithm is used to rigorously compare and map real-time data with the pattern library. This method not only quantifies the probability of defect existence (probability of gestation) in the form of matching confidence, but also clearly indicates the most likely defect type and its current specific development stage (e.g., early budding stage). This provides maintenance personnel with clear and quantitative decision-making basis, realizing a leap from "qualitative alarm" to "quantitative diagnosis and prognosis," effectively avoiding misjudgment and over-maintenance, and optimizing the allocation of maintenance resources. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 As shown, the present invention is a method for testing the electrical performance of power cables, comprising the following steps:

[0059] S1: Synchronously acquire multi-source heterogeneous time-series signals at the power cable monitoring point, including broadband electromagnetic signals, mechanical vibration signals and temperature signals;

[0060] S2: Perform cross-modal coupling analysis on multi-source heterogeneous time-series signals, extract time-frequency domain correlation features between pairs of broadband electromagnetic signals, mechanical vibration signals and temperature signals, and generate multi-modal coupling tensors based on time-frequency domain correlation features;

[0061] S3: Perform tensor manifold expansion on the multimodal coupling tensor to reconstruct a nonlinear manifold structure that can characterize the state of multi-physics interaction, and calculate the set of local curvature distribution and topological invariants of the nonlinear manifold structure in the current high-dimensional space as the system state fingerprint.

[0062] S4: Input the system state fingerprint into the pre-trained defect prediction model. The defect prediction model is configured to directly output the prediction results of the type of micro-defects that are being gestated in the insulation medium of the power cable based on the evolution sequence of the local curvature distribution and topological invariants in the system state fingerprint. The prediction results include the probability of defect gestation and the determination of the evolution stage.

[0063] In S1, the acquisition of broadband electromagnetic signals is achieved by clamping a high-frequency current transformer onto the cable grounding wire or shielding layer. The operating frequency band of this transformer covers a preset wide frequency range, and it can continuously sense and record transient pulse currents and their high-frequency oscillation waveforms related to partial discharge and electrical treeing in the cable, forming a broadband electromagnetic signal sequence.

[0064] The acquisition of mechanical vibration signals is achieved by fixing a microelectromechanical accelerometer (MEMS) sensor to the cable connector housing or a nearby supporting structure. Based on its preset frequency response range, the sensor continuously measures minute vibrations of the housing caused by internal discharge, mechanical loosening, or electrostriction, and converts the vibration acceleration into a corresponding voltage signal output, forming a mechanical vibration signal sequence.

[0065] Temperature signals are acquired through a distributed fiber optic temperature sensing system laid on the surface of the cable joint or through resistance temperature detectors (RTDs) placed at key temperature measurement points. This sensing system continuously measures the spatial distribution of temperature on the cable joint surface or the temperature value at specific points at preset sampling intervals, converting the physical quantity of temperature into a continuous electrical signal, forming a temperature signal sequence.

[0066] In S2, firstly, continuous wavelet transforms are performed on the broadband electromagnetic signal, mechanical vibration signal, and temperature signal to obtain a detailed representation of each signal in the time-frequency domain. For the broadband electromagnetic signal, a wavelet basis function exhibiting a rapidly decaying oscillation pattern in the time domain is selected for the transform, such as a certain type of wavelet simulating a decaying oscillation waveform. This type of wavelet basis function has a stronger matching and capturing ability for transient pulse current waveforms. For the mechanical vibration signal, a wavelet basis function with strict symmetry and whose time-domain waveform is symmetrical about the center point is selected for the transform, such as a certain symmetrical wavelet. This type of basis function can ensure that no phase distortion is introduced during the transform process, thereby accurately maintaining the timing information of the vibration event. For the relatively slowly changing temperature signal, its time series is first estimated locally in real time. Specifically, within a sliding time window, the instantaneous frequency estimate is obtained by detecting the rate of change of the signal phase with time. Then, based on this instantaneous frequency estimate, the analysis scale range used in the continuous wavelet transform is dynamically adjusted so that the center of the scale range is inversely proportional to the instantaneous frequency, thereby achieving adaptive matching of the analysis bandwidth to the rate of change of the temperature signal. After completing the transformation of the above three signals, the resulting wavelet coefficient sequences are organized according to three dimensions: time point, analysis scale, and coefficient amplitude, and constructed into three independent three-dimensional arrays. Each array is a time spectrum of wavelet coefficients of a set of signals.

[0067] Secondly, based on the obtained three sets of wavelet coefficient time-frequency spectra, the time-frequency domain correlation strength between any two different physical signals is calculated. For any two specified signals, such as broadband electromagnetic signals and mechanical vibration signals, a pair of wavelet coefficients at the same time point and the same analysis scale is extracted. The calculation process of the wavelet coherence coefficient is as follows: First, the average value of the product of this pair of wavelet coefficients at multiple consecutive time points is calculated to obtain a smoothed estimate of the cross-spectral density; then, the average value of the squared amplitudes of the wavelet coefficients of each of the two signals at the same time point is calculated to obtain a smoothed estimate of their respective autospectral densities; finally, the squared amplitude of the smoothed cross-spectral density estimate is divided by the product of the smoothed autospectral density estimates of the two signals, and the square root of the result is the wavelet coherence coefficient value near that time point at that analysis scale. This coefficient value is between 0 and 1; the closer the value is to 1, the stronger the linear correlation between the two signals in that time-frequency region. Repeating this calculation at multiple preset analysis scales yields a two-dimensional spectrum with time and scale as two-dimensional coordinates and coherence coefficient values ​​as intensity, called the signal pair coherence spectrum. The above calculations are performed on all possible signal combinations (broadband electromagnetic-mechanical vibration, broadband electromagnetic-temperature, mechanical vibration-temperature) to obtain three signal pair coherence spectra, which are the sets of coherence coefficients between all signal pairs.

[0068] Next, the calculated sets of coherence coefficients are organized into a comprehensive coupling matrix. This matrix is ​​a three-dimensional array. Its first and second dimensions represent the type indices of the two signals involved in the calculation, for example, the first index represents a broadband electromagnetic signal, and the second index represents a mechanical vibration signal. Since the coherence coefficient calculation does not distinguish between orders (the coherence between signal A and signal B is equal to the coherence between signal B and signal A), the matrix is ​​symmetric about its main diagonal. The third dimension of the matrix contains the wavelet coherence coefficient sequence of the corresponding signal pair at all preset analysis scales. Through this arrangement, the coupling matrix systematically characterizes the time-varying coupling strength relationship between any two physical signals at multiple frequency band scales.

[0069] Finally, the coupling matrix and the original time-domain signal waveforms are used to synthesize the final multimodal coupling tensor. Specifically, at each identical time point, a slice of the coupling matrix (a two-dimensional matrix whose elements are the coherence coefficients of each signal pair at that time and scale) is extracted, along with a three-dimensional vector composed of the original sampled amplitudes of the three signals at that time. Then, an operation called outer product is performed, whereby the three-dimensional vector is multiplied sequentially by each element of the two-dimensional matrix, and the results are arranged in a specific order. This operation expands a three-dimensional vector and a two-dimensional matrix into a five-dimensional array block. This operation is repeated for all time points, and all the generated five-dimensional array blocks are arranged in chronological order, ultimately forming a fourth-order tensor. The four dimensions of this tensor correspond to: time point, signal type one, signal type two, and analysis scale. This fourth-order tensor is the multimodal coupling tensor, which simultaneously encapsulates the amplitude information of the original signals, the time-frequency coupling strength information between signals, and the time evolution information.

[0070] In S3, firstly, the multimodal coupling tensor undergoes temporal sequential decomposition. A time window length is set, for example, containing 256 consecutive sampling time points, and a sliding step size is set, for example, 64 time points. Along the temporal dimension of the tensor, the time window is moved with this sliding step size, sequentially extracting multiple tensor blocks of fixed time lengths, called tensor slices. Each tensor slice retains all signal modes and scale coupling information within that time period. For each tensor slice, its core features need to be extracted. For this purpose, higher-order singular value decomposition (SVD) is introduced. Before performing the decomposition, a decomposition rank estimate needs to be pre-set for each non-temporal mode of the tensor slice (e.g., signal type one, signal type two, and analysis scale). This estimate is pre-determined based on the physical meaning of each mode; for example, the rank of the mode reflecting signal type coupling can be set to 3, and the rank of the mode reflecting scale correlation can be set to 5. Its function is to constrain the number of components in each direction after decomposition, ensuring that the extracted features have a clear physical dimension and are not excessively redundant.

[0071] Under the constraints of the set rank estimation for each mode decomposition, a higher-order singular value decomposition (SVD) is performed on each tensor slice to compute a dimensionality-reduced core tensor. This process begins with a preliminary SVD of each non-temporal mode matrix of the tensor slice, obtaining the initial set of left singular vectors corresponding to each mode, i.e., the initial set of orthogonal components. Subsequently, the energy concentration of each component is calculated, defined as the proportion of the squared singular values ​​of that component to the sum of the squared singular values ​​of all components in that mode. Simultaneously, the intermodal coupling degree between corresponding components from different modes needs to be calculated. For components from two different modes, their intermodal coupling degree is... Defined as the ability of these two components to jointly explain the variance of the original tensor slice data. Specifically, let the variance come from the modal... The Each component is , from modality The Each component is Then the coupling degree It can be calculated using the following formula: ;

[0072] in, Represents the current tensor slice, symbol and They represent the tensor along the mode, respectively. and modality Modular multiplication, Representing vectors transpose, This represents the Frobenius norm of the computed tensor (i.e., the square root of the sum of the squares of all its elements). This formula calculates the proportion of signal energy retained after simultaneously projecting a slice of the original tensor onto the subspace spanned by these two components. This represents the Frobenius norm of the computed tensor (i.e., the square root of the sum of the squares of all its elements). This formula calculates the proportion of signal energy retained after simultaneously projecting a slice of the original tensor onto the subspace spanned by these two components. This value can be a weighted sum of its energy concentration and the average of its coupling degree across all relevant modes. All comprehensive evaluation values... Components greater than a threshold of 0.75 are selected to form a set of components to be refined. Using this set as the starting point for iterative optimization, under the hard constraint of decomposition rank estimation, alternating least squares method is used for iterative refinement until the change in the objective function (i.e., the Frobenius norm error between the reconstructed tensor and the original tensor slice) is less than a preset minimum threshold (e.g., ...). When the condition is met, convergence is considered achieved. The final refined modal orthogonal components are truncated according to the preset decomposition rank estimate, retaining only the first few components. The main components ( (For the decomposition rank estimation of the corresponding mode), the dimensionality-reduced core tensor is calculated through tensor shrinking operation. subscript Identify the time window to which this core tensor belongs. Represent the dimensionality-reduced core tensors for all time windows. Arranged in chronological order according to their corresponding time windows, among which, The total number of time windows constitutes the time-varying core tensor sequence.

[0073] Subsequently, each core tensor in the aforementioned time-varying core tensor sequence is considered as a point in a high-dimensional space. Specifically, all elements of each core tensor are arranged into a high-dimensional vector in a specific order. Connecting all points in the sequence forms a discrete curve embedded in the high-dimensional space, which serves as a nonlinear manifold structure characterizing the evolution of the system state. The connections between points are not simple straight lines, but rather geodesic connections between adjacent points are calculated. This is achieved by calculating the shortest path in the local tangent space formed by the adjacent points and their local neighborhoods.

[0074] Next, the local geometric and topological features of this discrete curve manifold structure are calculated. For each point on the curve, the preceding and following points are selected. Neighboring points (e.g.) A local neighborhood point set is formed. Based on this neighborhood point set, the tangent space (stretched by the directions of the first few principal components) and the normal space (the complement space orthogonal to the tangent space) at this point are fitted using principal component analysis. Local curvature is characterized by calculating the rate of change of the tangent space direction along the curve at this point, specifically estimated by the rate of change of the angle between the principal directions of the tangent spaces of adjacent points. Furthermore, topological invariants describing the overall topology of this discrete curve are extracted, mainly its connectivity (ensuring the curve is a single continuous branch) and the number of "holes" (zero in the case of this one-dimensional curve, used to verify the simplicity of the structure). These topological features remain unchanged under continuous shape deformation.

[0075] Finally, the calculated local curvature values ​​at all time points are arranged in chronological order to form a curvature distribution sequence, and then concatenated with the topological invariants extracted from the overall curve (such as connectivity markers and hole counts) to form an original feature vector. This original feature vector is then normalized, for example using a minimum-maximum normalization method, so that the values ​​of all its elements fall between 0 and 1. The normalized feature vector constitutes the final system state fingerprint, which comprehensively reflects the geometric and topological evolution characteristics of the cable's internal state under the influence of multiphysics, and is used for subsequent defect prediction.

[0076] In S4, the core of the defect prediction model is a combination of a temporal attention network and a sequence-to-sequence mapping network. First, the local curvature distribution sequence and the topological invariant sequence contained in the system state fingerprint are concatenated in strict temporal order to form a complete temporal feature vector sequence characterizing the evolution of the system state. This temporal feature vector sequence is input into the temporal attention network. This network generates a set of attention weights that dynamically change over time through a self-correlation computation mechanism. Specifically, for each feature vector at each time point in the input sequence, the network calculates its correlation score with all other feature vectors at all other time points in the sequence (including itself). The correlation score is typically calculated based on vector dot product operations and scaling, essentially measuring the similarity or dependence between system state features at different times. Then, a non-linear normalization function (e.g., the Softmax function) is applied to these raw correlation scores, transforming them into a set of weight values ​​that sum to 1, i.e., the temporal attention weights. The magnitude of these weights directly reflects the importance of the features at the corresponding time point to the final judgment task.

[0077] Subsequently, using the temporal attention weights generated in the previous step, the original temporal feature vector sequence is weighted, focused, and fused with context to extract the key evolutionary features most relevant to the defect gestation process. Specifically, the original feature vector at each time point is multiplied by its corresponding attention weight, and then summed with the weighted feature vectors from all other time points in the sequence. This process is equivalent to filtering and summarizing historical and current information based on importance, thereby generating a new feature vector sequence that incorporates global contextual information. Each vector in this new sequence is called a key evolutionary feature, which more centrally reflects the pattern information related to the dynamics of potential defect development.

[0078] Next, the obtained key evolutionary feature sequence is input into a sequence-to-sequence mapping network. This network is designed to map continuous, dense key evolutionary feature time series into a discrete state transition path. Simultaneously, it needs to calculate the matching confidence level between this path and each reference path in a predefined library of multiple typical defect evolution patterns. This pattern library stores typical state evolution sequences from the inception to development of different types of micro-defects (such as electrical trees, water trees, cavities, etc.) summarized through historical case analysis, experimental data, or theoretical models. Each reference path consists of a series of discrete state nodes connected in developmental order.

[0079] The calculation process for the matching confidence is as follows. First, a coding network is used to encode the key evolutionary feature sequence point by point. This coding network typically consists of several fully connected layers or one-dimensional convolutional layers, and its function is to transform the key evolutionary feature vector at each time point into a higher-dimensional, more discriminative representation space. The output is called the high-dimensional state code. Then, a dynamic time warping algorithm is used to match the obtained high-dimensional state code sequence with a reference path in the pattern library. Dynamic time warping works by constructing a cumulative cost matrix. The rows of this matrix correspond to each point in the high-dimensional state code sequence, and the columns correspond to each state node in the reference path. The value of each element in the matrix represents the local distance (such as Euclidean distance) between a point in the coding sequence and a node in the reference path when aligning them. The algorithm uses dynamic programming to calculate the minimum cumulative matching cost to reach each position, starting from the beginning of the matrix. This cost is defined as the minimum sum of the local distances of all possible alignment paths to that position. During the calculation, constraints determined by the physical evolution of cable defects were applied: prohibition of path backtracking (i.e., time cannot be reversed) and prohibition of state transitions (i.e., the development process is usually continuous and intermediate states cannot be skipped). After filling the entire cumulative cost matrix, an alignment path that minimizes the global cumulative matching cost is found by backtracking from the endpoint. This path is determined as the discrete state transition path, indicating the development trajectory most likely to be followed by the observation sequence on the reference path in the pattern library. Finally, the matching confidence is calculated by combining two factors: first, the final normalization distance corresponding to the path with the lowest global matching cost (i.e., the value at the endpoint of the cumulative cost matrix), the smaller the distance, the higher the confidence; second, the proportion of points in the observation sequence or reference path that failed to match successfully during the alignment process, the lower the proportion, the higher the confidence. A specific synthesis method is to normalize the final normalization distance and multiply it by (1 minus the proportion of unmatched points) to obtain the final matching confidence value.

[0080] Finally, the method iterates through all pre-stored typical defect evolution patterns in the pattern library, repeating the matching process described above to obtain the matching confidence score corresponding to each reference path. From all calculated matching confidence scores, the one with the highest value is selected, and this value is directly output as the "defect gestation probability" for this prediction. Simultaneously, the type of the reference path corresponding to this highest confidence score is identified as the most likely defect type; and the stage marked on the corresponding reference path based on the endpoint reached by the determined state transition path (e.g., "early germination stage," "accelerated development stage," or "critical failure stage") is output as the "evolution stage determination result." Through these steps, the method completes the automatic mapping and decision-making from system state fingerprints to specific defect prediction results.

[0081] Please see Figure 2As shown, a power cable electrical performance testing system includes:

[0082] A multi-source signal synchronous acquisition module is used to synchronously acquire multi-source heterogeneous time-series signals at power cable monitoring points. The multi-source heterogeneous time-series signals include broadband electromagnetic signals, mechanical vibration signals, and temperature signals.

[0083] The cross-modal coupling feature extraction module is used to perform cross-modal coupling analysis on multi-source heterogeneous time-series signals, extract the time-frequency domain correlation features between pairs of broadband electromagnetic signals, mechanical vibration signals and temperature signals, and generate multi-modal coupling tensors based on the time-frequency domain correlation features;

[0084] The tensor manifold analysis and fingerprint generation module is used to expand the multimodal coupled tensor into a nonlinear manifold structure that can characterize the state of multi-physics interaction. It also calculates the set of local curvature distribution and topological invariants of the nonlinear manifold structure in the current high-dimensional space as the system state fingerprint.

[0085] The defect prediction and assessment module is used to input the system state fingerprint into the pre-trained defect prediction model. The defect prediction model is configured to directly output the prediction results of the type of micro-defects that are being gestated in the insulation medium of the power cable based on the evolution sequence of the local curvature distribution and topological invariants in the system state fingerprint. The prediction results include the probability of defect gestation and the determination of the evolution stage.

[0086] The working principle of this invention is as follows: First, broadband electromagnetic signals, mechanical vibration signals, and temperature signals are simultaneously acquired at cable monitoring points. Next, by performing continuous wavelet transforms on these signals and calculating the wavelet coherence coefficients between pairs of signals, a multimodal coupled tensor is constructed, integrating the original amplitude and cross-modal time-frequency coupling relationship. Then, time slicing and higher-order singular value decomposition are performed on this tensor to obtain a time-varying core tensor sequence, which is then mapped to a discrete manifold curve in a high-dimensional space. By calculating the local curvature distribution and topological invariants of this curve, a system state fingerprint characterizing the comprehensive operating state of the cable is generated. Finally, this fingerprint is input into a defect prediction model composed of a time attention network and a sequence-to-sequence mapping network. The model uses a dynamic time warping algorithm to match it with a pre-set defect evolution pattern library, directly outputting the defect gestation probability and evolution stage determination results.

[0087] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting the electrical performance of a power cable, characterized in that, The method comprises the following steps: S1: synchronously collecting multi-source heterogeneous time sequence signals at a power cable monitoring point, the multi-source heterogeneous time sequence signals comprising wideband electromagnetic signals, mechanical vibration signals and temperature signals; S2: performing cross-modal coupling analysis on the multi-source heterogeneous time sequence signals, extracting time-frequency domain correlation features between the wideband electromagnetic signals, the mechanical vibration signals and the temperature signals, and generating a multi-modal coupling tensor based on the time-frequency domain correlation features; S3: performing tensor manifold unfolding on the multi-modal coupling tensor, reconstructing into a nonlinear manifold structure capable of representing a multi-physical field cooperative action state, and calculating a local curvature distribution and a set of topological invariants of the nonlinear manifold structure in a current high-dimensional space as a system state fingerprint, specifically comprising: performing tensor fiber unfolding along a time mode of the multi-modal coupling tensor to obtain a series of high-order tensor slices, and performing high-order singular value decomposition on each tensor slice to extract core tensors to form a time-varying core tensor sequence; mapping each core tensor in the time-varying core tensor sequence to a point on an abstract manifold, and reconstructing the entire sequence into a discrete curve embedded in a high-dimensional space by calculating the geodesic connection between adjacent points as the nonlinear manifold structure; calculating the change rate of the included angle between the tangent space and the normal space based on the local neighborhood point set at each point on the discrete curve to obtain the local curvature distribution, and simultaneously extracting the connectivity and the number of holes that remain unchanged under continuous deformation of the entire discrete curve as the topological invariants; splicing and normalizing the local curvature distribution and the topological invariants in time sequence to form the system state fingerprint; the extracting core tensors to form a time-varying core tensor sequence specifically comprises: slidingly slicing along the time dimension of the multi-modal coupling tensor according to a preset time window length and a sliding step to obtain a series of tensor slices with a fixed time span; for each tensor slice, pre-setting the decomposition rank estimation of each mode according to the physical meaning of different non-time modes of the tensor slice, taking the decomposition rank estimation as a constraint condition; performing high-order singular value decomposition on each tensor slice under the constraint of the decomposition rank estimation of each mode, and only retaining the orthogonal components determined by the decomposition rank estimation to calculate the core tensors after dimension reduction; arranging the core tensors after dimension reduction corresponding to all tensor slices in the order of the time windows corresponding to the core tensors to form the time-varying core tensor sequence; the calculating the core tensors after dimension reduction specifically comprises: performing preliminary decomposition on each non-time mode matrix of the tensor slice to obtain an initial orthogonal component set corresponding to each mode; Calculate the energy concentration of each component in the initial orthogonal component set, and simultaneously calculate the intermodal coupling degree of corresponding components between different modes. The calculation process for the intermodal coupling degree is as follows: Let the energy concentration of the components originating from the mode be... The Each component is , from modality The Each component is Then the intermodal coupling degree It can be calculated using the following formula: ; wherein, denotes the current tensor slice, the notation and denotes the tensor multiplication along the mode and the mode , denotes the transpose of the vector , denotes the computation of the Frobenius norm of a tensor, i.e. the square root of the sum of the squares of all elements; comprehensively evaluating all components according to the energy concentration degree and the inter-modal coupling degree to select the initial orthogonal components satisfying a preset comprehensive evaluation threshold to form a to-be-refined component set; taking the to-be-refined component set as an iterative initial value, performing iterative refinement under the constraint of the decomposition rank estimation until convergence, truncating the refined orthogonal components finally obtained according to the number of decomposition rank estimations, and calculating the core tensors after dimension reduction; S4: inputting the system state fingerprint into a pre-trained defect prediction model, the defect prediction model being configured to directly output a prediction result of a type of micro-defect incubating in the power cable insulating medium according to the local curvature distribution and the evolution sequence of the topological invariant in the system state fingerprint, the prediction result including an incubation probability and an evolution stage determination of the defect.

2. The method of claim 1, wherein, The S2 specifically comprises: performing continuous wavelet transform on each of the broadband electromagnetic signal, the mechanical vibration signal and the temperature signal respectively to obtain a wavelet coefficient time-frequency spectrum corresponding to each signal; for any two different types of signals, calculating wavelet coherence coefficients of the corresponding wavelet coefficient time-frequency spectra on multiple preset scales to obtain a signal pair coherence spectrum composed of wavelet coherence coefficients, and obtaining a coherence coefficient set between all signal pairs; arranging and combining the coherence coefficient sets of all signal pairs according to signal types into a symmetric coupling relationship matrix; performing outer product operation on the coupling relationship matrix and the original time-domain waveform of each signal in the time dimension to generate a fourth-order tensor as a multi-modal coupling tensor.

3. A method of detecting the electrical performance of a power cable according to claim 2, characterized in that, The wavelet coefficient time-frequency spectrum corresponding to each signal is obtained by specifically comprising: for the broadband electromagnetic signal, using a wavelet basis function with shock oscillation attenuation characteristics to perform continuous wavelet transform; for the mechanical vibration signal, using a symmetric wavelet basis function with zero phase shift characteristics to perform continuous wavelet transform; for the temperature signal, first calculating the local instantaneous frequency of the temperature signal sequence, and then adjusting the analysis scale range of the continuous wavelet transform based on the local instantaneous frequency, and then performing the transform; transforming each wavelet coefficient sequence into a three-dimensional time-frequency spectrum structure containing time dimension, scale dimension and amplitude dimension, and taking the three-dimensional time-frequency spectrum structure as the wavelet coefficient time-frequency spectrum.

4. The method of claim 1, wherein, The S4 specifically comprises: inputting the local curvature distribution and the topological invariant contained in the system state fingerprint into a time attention network in time sequence, and the time attention network generates time attention weights that dynamically change over time through self-association calculation; using the time attention weights to weight, focus and contextually fuse the input time sequence features to extract key evolution features most relevant to the defect incubation process; inputting the key evolution features into a sequence-to-sequence mapping network, the mapping network maps the continuous time sequence features into a discrete state transition path, and simultaneously calculates the matching confidence of each mode in a pre-defined library of multiple typical defect evolution modes; outputting the highest matching confidence as the incubation probability, and outputting the mode stage corresponding to the end point of the state transition path as the evolution stage determination result.

5. A method of detecting the electrical performance of a power cable according to claim 4, characterized in that, The matching confidence calculation process is as follows: using an encoding network to point-by-point encode the key evolution feature sequence to convert the feature vector at each time point into a high-dimensional state code; performing dynamic time warping matching between the high-dimensional state code sequence and each reference path in the library of typical defect evolution modes to calculate the minimum cumulative matching cost from the start point of the sequence to the current point; A global path with the lowest matching cost is determined by backtracking search based on the minimum accumulated matching cost, and physical evolution constraints of prohibiting path backtracking and state jumping are applied; The matching confidence is calculated based on the final regularized distance between the global path with the lowest matching cost and the corresponding reference path, and the proportion of unmatched feature points in the path alignment process.

6. A power cable electrical performance detection system characterized by, A method for detecting the electrical performance of a power cable according to any one of claims 1-5, comprising: A multi-source signal synchronous acquisition module is configured to synchronously acquire multi-source heterogeneous time series signals at the monitoring point of the power cable, wherein the multi-source heterogeneous time series signals include broadband electromagnetic signals, mechanical vibration signals, and temperature signals; A cross-modal coupling feature extraction module is configured to perform cross-modal coupling analysis on the multi-source heterogeneous time series signals, extract time-frequency domain correlation features between the broadband electromagnetic signals, the mechanical vibration signals, and the temperature signals, and generate a multi-modal coupling tensor based on the time-frequency domain correlation features; A tensor manifold analysis and fingerprint generation module is configured to perform tensor manifold unfolding on the multi-modal coupling tensor, reconstruct a nonlinear manifold structure capable of representing the cooperative action state of multiple physical fields, and calculate a local curvature distribution and a set of topological invariants of the nonlinear manifold structure in the current high-dimensional space as a system state fingerprint; A defect prediction and evaluation module is configured to input the system state fingerprint into a pre-trained defect prediction model, wherein the defect prediction model is configured to directly output a prediction result of a type of micro-defect incubating in the insulating medium of the power cable according to an evolution sequence of the local curvature distribution and the topological invariants in the system state fingerprint, and the prediction result includes an incubation probability and an evolution stage determination of the defect.

Images