A method for identifying defects and faults of a crosslinked polyethylene cable

Abstract

The present application belongs to the technical field of non-destructive testing and intelligent operation and maintenance of power systems, and particularly relates to a cross-linked polyethylene cable defect and fault identification method, which first acquires a steady-state input impedance sequence of an XLPE cable to be measured in a wide frequency band, and extracts corresponding amplitude spectrum and phase spectrum; then utilizes an image conversion operator to convert the one-dimensional amplitude spectrum and phase spectrum into two-dimensional feature maps respectively, and splices in the channel dimension to obtain an impedance multi-channel fusion map; a Transform-CNN double-branch feature extraction network is constructed, local spatial features of the map are extracted through the CNN branch, and global sequence dependent features of the map are extracted through the Transformer branch; finally, the local features and global features are weighted and fused through a cross-attention mechanism, and are sent into a fully connected classification layer to output the defect and fault type of the cable.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing and intelligent operation and maintenance technology of power systems, and specifically relates to a method for identifying defects and faults in cross-linked polyethylene cables. Background Technology

[0002] With the deepening of urban power grid modernization, cross-linked polyethylene (XLPE) cables have become the core of urban power supply networks. Due to their long-term operation in complex underground environments, these cables are subjected to the combined effects of electrical stress, thermal stress, and moisture erosion, making them highly susceptible to local insulation defects such as thermal aging, water trees, or electrical trees. If not detected in time, these defects can escalate into serious system faults such as short circuits, open circuits, or high-resistance flashovers, directly threatening the safe operation of the power grid. Currently, detection methods for XLPE cables mainly include partial discharge detection, dielectric loss analysis, and swept-frequency impedance spectroscopy (IS). Among these, swept-frequency impedance spectroscopy has received widespread attention in the field of cable defect diagnosis due to its ability to cover a wide frequency band and its high sensitivity to changes in cable distributed parameters. Therefore, this invention provides a method for identifying defects and faults in cross-linked polyethylene cables. This method deeply integrates cable physical mechanisms and possesses high-order spectral characterization capabilities and global-local feature collaborative capture capabilities. Such an intelligent identification method is of significant practical importance for achieving accurate early warning and fault location of XLPE cable defects. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying defects and faults in cross-linked polyethylene (XLPE) cables. This method is a multimodal and accurate identification method for early weak defects and system faults in XLPE cables based on the deep fusion of high-order steady-state impedance spectrum mapping and Transformer-CNN. It addresses the technical problems in existing XLPE cable condition assessment and fault diagnosis, such as low sensitivity in early weak defect identification, lack of physical interpretability of pure data-driven AI models, and insufficient mining of one-dimensional impedance sequence features.

[0004] This invention aims to introduce high-order mathematical coding operators to deeply explore the nonlinear spatial topology and dynamic transfer characteristics of cable impedance sequences in a wide frequency band, and to use a Transformer-CNN dual-branch neural network with physical mechanism constraints to achieve accurate classification, quantitative grading and spatial positioning of cable insulation defects and system faults.

[0005] The specific technical solution adopted by this invention is as follows: A method for identifying defects and faults in cross-linked polyethylene cables includes the following steps: Step S1: Broadband impedance spectrum data acquisition and normalization preprocessing; In step S1: obtain broadband steady-state input impedance spectrum data of the first end of the XLPE cable under test, and extract the frequency-amplitude spectrum sequence and the frequency-phase spectrum sequence. Step S2: Two-dimensional feature reconstruction of the amplitude spectrum based on the Gram angle field GASF operator; two-dimensional feature reconstruction of the phase spectrum based on the Markov transfer field MTF operator; In step S2: the amplitude spectrum sequence is converted into an impedance amplitude evolution spectrum, the phase spectrum sequence is converted into a phase dynamic transfer spectrum, and the impedance amplitude evolution spectrum and the phase dynamic transfer spectrum are spliced ​​together in the feature channel dimension to generate a multi-channel fused spectrum. The specific transformation process of the feature map in step S2 is as follows: The normalized amplitude spectrum sequence is mapped to the polar coordinate system through inverse cosine transformation. The angles and characteristics between different frequency points are calculated using Gram angle and field GASF operator to generate a two-dimensional spatial spectrum that reflects the correlation between amplitude and frequency evolution, which serves as the impedance amplitude evolution spectrum. The amplitude range of the phase spectrum sequence is divided into several state intervals, the state transition probability between each frequency point is calculated, and a two-dimensional image reflecting the phase transition law is generated using the Markov transition field MTF operator, which serves as the phase dynamic transition map.

[0006] Step S3: Dual-branch parallel feature extraction and cross-attention fusion; Step S3 includes: Step S31: Construct a dual-branch parallel feature extraction network composed of a convolutional neural network (CNN) and a Transformer; input the multi-channel fused map into the dual-branch parallel feature extraction network, use the CNN branch to extract the local spatial features of the map, and use the Transformer branch to extract the global sequence features of the map; In step S31, the specific processing procedure of the dual-branch parallel feature extraction network is as follows: CNN branch processing: The multi-channel fused spectrum undergoes multiple convolutions, normalization, and ReLU activation function operations to extract the local defect feature matrix F of the impedance spectrum within a specific frequency band. CNN ; Transformer branch processing: The multi-channel fused spectrum is divided into multiple image patches and flattened. To preserve the temporal information of impedance spectrum evolution with frequency, the image patches are encoded with position vectors. Subsequently, a multi-head self-attention mechanism is used to extract correlation features across the entire frequency band in parallel from multiple dimensions, realizing global feature modeling of complex fault modes. The cross-frequency band global correlation feature matrix F is extracted. Trans .

[0007] Step S32: Through the cross-attention mechanism, the extracted local spatial features and global sequence features are weighted and fused to obtain the final feature fusion tensor; In S32, the specific fusion calculation process of the cross-attention mechanism is as follows: Feature fusion is achieved using a cross-attention mechanism: local defect features extracted by a convolutional neural network (CNN) are mapped to a query matrix. The global correlation feature matrix F extracted by Transformer Trans Mapped to a key matrix AND-value matrix By calculating the correlation weights between local and global features, adaptive enhancement of weak defect signals can be achieved. Calculate the feature fusion tensor using the following formula. : ; in, represents the feature dimension of the key matrix.

[0008] Step S4: Defect and fault state classification based on fully connected and Softmax.

[0009] In step S4: the feature fusion tensor is input into the classifier softmax model, and the defect and fault identification results of the XLPE cable under test are output.

[0010] In step S4, the classifier softmax model includes a fully connected layer and a softmax output layer. The classifier softmax model uses the amplitude frequency domain correlation features extracted in the early stage to identify thermal aging defects, water tree defects, and shielding layer damage defects of the cable body insulation. It also uses the phase transfer law features to identify short circuit faults, open circuit faults, and high-resistance grounding faults. Finally, the softmax output layer outputs the comprehensive identification probability of each defect and fault.

[0011] Step S4 also includes an assessment of the severity of the defect: Calculate the feature fusion tensor The L2 norm is calculated, and the frequency offset of the first measured impedance resonance peak is extracted. The L2 norm distance and frequency offset are used as additional physical constraints input to the classifier to correct and output the final defect identification result. Calculate the defect severity index using the following formula. : ; in, This indicates the calculation of the L2 norm. To measure the resonant frequency, The standard resonant frequency of a good cable. and These are the preset weighting coefficients.

[0012] The technical effects achieved by this invention are as follows: This invention does not use neural networks as simple classifiers, but instead achieves high-precision correspondence on the map with image features such as Gram angle field GASF texture offset and Markov transition field MTF probability interval, resulting in diagnostic results with high physical traceability.

[0013] This invention establishes a mathematical correlation between deep learning features and cable physical defects by mapping the correlation between the offset of Gram angle field GASF characteristic defects and impedance frequency, and the correspondence between the probability interval of Markov transfer field MTF and the phase state evolution probability.

[0014] This invention employs a Transform-CNN multi-scale feature fusion mechanism to ensure that while the model focuses on local sudden failures, it can also eliminate environmental interference based on the global evolution trend, which greatly reduces the false alarm rate and enhances the system's noise resistance under complex operating conditions. Attached Figure Description

[0015] Figure 1 This is a block diagram of the detection system in this invention; Figure 2 This is a schematic diagram comparing the amplitude spectrum frequency band shift during thermal aging in this invention. Figure 3 This is a schematic diagram comparing the impedance of the water tree and the resonant peak in this invention; Figure 4 This is a schematic diagram comparing shielding damage and phase broadband fluctuations in this invention. Figure 5 This is a schematic diagram comparing the faults in this invention: sudden changes in specific resonant frequencies; Figure 6 This is a radar image reconstructed from the multi-defect index fusion feature in this invention. Detailed Implementation

[0016] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0017] like Figures 1-6 As shown, a method for identifying defects and faults in cross-linked polyethylene (XLPE) cables addresses the technical problems in existing technologies, such as the difficulty in extracting early weak defects (e.g., water trees, thermal aging) from XLPE cables and the high likelihood of misjudgment under complex high-frequency noise interference. The method specifically includes the following steps: Step S1: Broadband impedance spectrum data acquisition and normalization preprocessing; In step S1: obtain broadband steady-state input impedance spectrum data of the first end of the XLPE cable under test, and extract the frequency-amplitude spectrum sequence and the frequency-phase spectrum sequence. In step S1, during implementation, the impedance spectrum sequence of the XLPE cable under test is obtained over a wide bandwidth, and the amplitude spectrum sequence and phase spectrum sequence are separated. Since there is a significant dimensional difference between the amplitude spectrum and the phase spectrum, to eliminate the influence of dimensions and accelerate the convergence of the subsequent neural network, a maximum-minimum linear mapping function is used to normalize the two sets of sequences respectively. The formula is: ; in, This is the original impedance amplitude or phase sequence. and These are the minimum and maximum values ​​in the sequence, respectively. This is the normalized sequence after mapping to the interval between 0 and 1.

[0018] Step S2: Two-dimensional feature reconstruction of the amplitude spectrum based on the Gram angle field GASF operator; two-dimensional feature reconstruction of the phase spectrum based on the Markov transfer field MTF operator; In step S2: the amplitude spectrum sequence is converted into an impedance amplitude evolution spectrum, the phase spectrum sequence is converted into a phase dynamic transfer spectrum, and the impedance amplitude evolution spectrum and the phase dynamic transfer spectrum are spliced ​​together in the feature channel dimension to generate a multi-channel fused spectrum. The specific transformation process of the feature map in step S2 is as follows: The normalized amplitude spectrum sequence is mapped to the polar coordinate system through inverse cosine transformation. The angles and characteristics between different frequency points are calculated using Gram angle and field GASF operator to generate a two-dimensional spatial spectrum that reflects the correlation between amplitude and frequency evolution, which serves as the impedance amplitude evolution spectrum. The amplitude range of the phase spectrum sequence is divided into several state intervals, the state transition probability between each frequency point is calculated, and a two-dimensional image reflecting the phase transition law is generated using the Markov transition field MTF operator, which serves as the phase dynamic transition map.

[0019] In the specific implementation process of step S2: In the two-dimensional feature reconstruction of the amplitude spectrum based on the Gram angle field GASF operator: the normalized amplitude spectrum sequence is mapped to angles in polar coordinates through inverse cosine transform. ,Right now Subsequently, the angles and characteristics between each frequency point are calculated to generate a two-dimensional spatial map reflecting the correlation between amplitude and frequency evolution, serving as the first two-dimensional feature map. Its feature matrix... The calculation formula is: ; in, Indicates the first in the image Line 1 The pixel values ​​of the column. This step, through nonlinear transformation of trigonometric functions, artificially amplifies the minute amplitude defects that are difficult to detect in the one-dimensional sequence, transforming them into significant symmetric features in the two-dimensional image, effectively enhancing the representation ability of weak defects such as water trees.

[0020] In the two-dimensional feature reconstruction of the phase spectrum based on the Markov Transition Field (MTF) operator, the amplitude range of the normalized phase spectrum sequence is divided into several state intervals. The state transition probability of the signal at each frequency point in different intervals is calculated, generating a two-dimensional image reflecting the dynamic phase transition law, which serves as the second two-dimensional feature map. The calculation formula for the MTF operator is as follows: ; in, These are elements in the state transition probability matrix. and They represent the first division. The and the first Each state interval Indicates the first The state interval to which the signal belongs at each frequency point This indicates the state of the previous adjacent frequency point. This step effectively filters out high-frequency random noise in the phase spectrum and transforms complex phase change patterns into obvious image features, making it suitable for identifying polarity reversal features caused by short circuits or open circuits.

[0021] Step S3: Dual-branch parallel feature extraction and cross-attention fusion; Step S3 includes: Step S31: Construct a dual-branch parallel feature extraction network composed of a convolutional neural network (CNN) and a Transformer; input the multi-channel fused map into the dual-branch parallel feature extraction network, use the CNN branch to extract the local spatial features of the map, and use the Transformer branch to extract the global sequence features of the map; In step S31, the specific processing procedure of the dual-branch parallel feature extraction network is as follows: CNN branch processing: The multi-channel fused spectrum undergoes multiple convolutions, normalization, and ReLU activation function operations to extract the local defect feature matrix F of the impedance spectrum within a specific frequency band. CNN ; Transformer branch processing: The multi-channel fused spectrum is divided into multiple image patches and flattened. To preserve the temporal information of impedance spectrum evolution with frequency, the image patches are encoded with position vectors. Subsequently, a multi-head self-attention mechanism is used to extract correlation features across the entire frequency band from multiple dimensions in parallel, realizing global feature modeling of complex fault modes. The cross-frequency band global correlation feature matrix F is extracted. Trans .

[0022] Step S32: Through the cross-attention mechanism, the extracted local spatial features and global sequence features are weighted and fused to obtain the final feature fusion tensor; In S32, the specific fusion calculation process of the cross-attention mechanism is as follows: Feature fusion is achieved using a cross-attention mechanism: local defect features extracted by a convolutional neural network (CNN) are mapped to a query matrix. The global correlation feature matrix F extracted by Transformer Trans Mapped to a key matrix AND-value matrix By calculating the correlation weights between local and global features, adaptive enhancement of weak defect signals can be achieved. Calculate the feature fusion tensor using the following formula. : ; in, represents the feature dimension of the key matrix.

[0023] In the specific implementation of step S3: a Transform-CNN deep learning network model is constructed. First, the local spatial defect features of the first and second two-dimensional feature maps are extracted using CNN branches, and the ReLU nonlinear activation function is used during the sampling process. By breaking the linear accumulation of network layers, a high-dimensional nonlinear decision space is constructed.

[0024] Simultaneously, global sequence modeling is performed using the Transformer branch. The two-dimensional feature map is divided into multiple image patches and flattened. To preserve the temporal physical information of impedance spectrum evolution with frequency, position encoding is injected into the image patches. Subsequently, the patches are input into a multi-head self-attention module to extract global correlation features across frequency bands.

[0025] To address the issue of single local features being easily affected by interference, this invention employs a cross-attention mechanism for feature fusion. The local defect features extracted by the CNN are mapped to a query matrix. The global correlation feature matrix F extracted by TransformerTrans Mapped to a key matrix AND-value matrix Its scaling dot product attention formula is: ; in, The dimension of the input features is defined. By calculating the correlation weights between local suspicious points and global background patterns, the model can adaptively utilize global trend information to calibrate local abnormal signals and output a multi-channel fused feature map.

[0026] Step S4: Defect and fault state classification based on fully connected and Softmax.

[0027] In step S4: the feature fusion tensor is input into the classifier softmax model, and the defect and fault identification results of the XLPE cable under test are output.

[0028] In step S4, the classifier softmax model includes a fully connected layer and a softmax output layer. The classifier softmax model uses the amplitude frequency domain correlation features extracted in the early stage to identify thermal aging defects, water tree defects, and shielding layer damage defects of the cable body insulation. It also uses the phase transfer law features to identify short circuit faults, open circuit faults, and high-resistance grounding faults. Finally, the softmax output layer outputs the comprehensive identification probability of each defect and fault.

[0029] Step S4 also includes an assessment of the severity of the defect: Calculate the feature fusion tensor The L2 norm is calculated, and the frequency offset of the first measured impedance resonance peak is extracted. The L2 norm distance and frequency offset are used as additional physical constraints input to the classifier to correct and output the final defect identification result. Calculate the defect severity index using the following formula. : ; in, This indicates the calculation of the L2 norm. To measure the resonant frequency, The standard resonant frequency of a good cable. and These are the preset weighting coefficients.

[0030] In the specific implementation of step S4: the classifier model includes a fully connected layer and a softmax output layer. The fused features obtained in step S4 are input into this classifier, mapped to classification weights for each category, and the final probability distribution is output. The defect and fault identification results specifically include two categories: One category consists of weak early defects that characterize local insulation deterioration, including thermal aging defects in the cable body insulation, water tree defects in the insulation, and shielding layer damage defects. The second category consists of explicit faults that characterize overall transmission impairment, including short-circuit faults, open-circuit faults, and high-resistance grounding faults.

[0031] This invention also includes auxiliary verification based on the physical characteristics of the cable: specifically, to prevent misjudgments from purely data-driven models, physical features are introduced to assist in the discrimination. On one hand, the measured impedance sequence is calculated. Reconstructing the baseline sequence of the model L2 distance between As a model reliability evaluation index: ; On the other hand, the center frequency of the first resonance peak in the impedance spectrum is extracted, and its frequency offset relative to the reference state is calculated. The L2 norm distance With frequency offset As an additional physical constraint input to the classifier, it enables secondary verification of the AI ​​recognition results, further improving the noise resistance of diagnosing minute defects under complex working conditions.

[0032] In this invention, by Figure 2 As can be seen, the GASF operator introduced in this invention can amplify and transform the difficult-to-detect one-dimensional signal distortion into a two-dimensional spatial texture, effectively solving the problem that early weak defects are easily masked and missed.

[0033] In this invention, by Figure 3 As can be seen, this invention extracts the frequency shift of the resonant peak and constructs additional physical constraints by combining it with the L2 norm to address the impedance loss of the first resonant peak caused by water tree defects. This gives the model intuitive physical interpretability and significantly improves the diagnostic accuracy for minute defects.

[0034] In this invention, by Figure 4 It can be seen that damage to the shielding layer will cause severe broadband random phase fluctuations. This invention utilizes the state transition mechanism of the MTF operator to transform complex phase change patterns into obvious image features, intuitively demonstrating the algorithm's excellent noise resistance and robustness under complex electromagnetic interference.

[0035] In this invention, by Figure 5 It can be seen that obvious faults such as short circuits or open circuits manifest as drastic changes in resonant frequency. The Transformer branch of this invention, with its long-range dependency modeling capability across frequency bands, can accurately pinpoint abrupt changes in the global evolution, ensuring precise location and zero false negatives for severe system faults.

[0036] In this invention, by Figure 6 As can be seen, in response to the problem of overlapping classifications where multiple defect features are intertwined and easily confused, the Transform-CNN dual-branch cross-fusion architecture of this invention can effectively decouple multi-dimensional indicators such as L2 norm, frequency shift, amplitude loss and phase fluctuation, so that various defects present a spatial distribution with clear boundaries. It completely breaks through the bottleneck of easy misjudgment of single features and successfully achieves full-dimensional and high-precision identification from weak early defects to obvious malignant faults.

[0037] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for identifying defects and faults in cross-linked polyethylene cables, characterized in that: Includes the following steps: Step S1: Broadband impedance spectrum data acquisition and normalization preprocessing; Step S2: Two-dimensional feature reconstruction of the amplitude spectrum based on the Gram angle field GASF operator; two-dimensional feature reconstruction of the phase spectrum based on the Markov transfer field MTF operator; Step S3: Dual-branch parallel feature extraction and cross-attention fusion; Step S4: Classification of defects and fault states based on fully connected and Softmax.

2. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 1, characterized in that: In step S1: obtain broadband steady-state input impedance spectrum data of the first end of the XLPE cable under test, and extract the frequency-amplitude spectrum sequence and the frequency-phase spectrum sequence. In step S2: the amplitude spectrum sequence is converted into an impedance amplitude evolution spectrum, the phase spectrum sequence is converted into a phase dynamic transfer spectrum, and the impedance amplitude evolution spectrum and the phase dynamic transfer spectrum are spliced ​​together in the feature channel dimension to generate a multi-channel fused spectrum. Step S3 includes: Step S31: Construct a dual-branch parallel feature extraction network composed of a convolutional neural network (CNN) and a Transformer; input the multi-channel fused map into the dual-branch parallel feature extraction network, use the CNN branch to extract the local spatial features of the map, and use the Transformer branch to extract the global sequence features of the map; Step S32: Through the cross-attention mechanism, the extracted local spatial features and global sequence features are weighted and fused to obtain the final feature fusion tensor; In step S4: the feature fusion tensor is input into the classifier softmax model, and the defect and fault identification results of the XLPE cable under test are output.

3. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 2, characterized in that: The specific transformation process of the feature map in step S2 is as follows: The normalized amplitude spectrum sequence is mapped to the polar coordinate system through inverse cosine transformation. The angles and characteristics between different frequency points are calculated using Gram angle and field GASF operator to generate a two-dimensional spatial spectrum that reflects the correlation between amplitude and frequency evolution, which serves as the impedance amplitude evolution spectrum. The amplitude range of the phase spectrum sequence is divided into several state intervals, the state transition probability between each frequency point is calculated, and a two-dimensional image reflecting the phase transition law is generated using the Markov transition field MTF operator, which serves as the phase dynamic transition map.

4. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 3, characterized in that: In step S31, the specific processing procedure of the dual-branch parallel feature extraction network is as follows: CNN branch processing: The multi-channel fused spectrum undergoes multiple convolutions, normalization, and ReLU activation function operations to extract the local defect feature matrix F of the impedance spectrum within a specific frequency band. CNN ; Transformer branch processing: The multi-channel fused spectrum is divided into multiple image patches and flattened. To preserve the temporal information of impedance spectrum evolution with frequency, the image patches are encoded with position vectors. Then, a multi-head self-attention mechanism is used to extract the associated features in parallel from multiple dimensions across the entire frequency band, thereby realizing global feature modeling of complex fault modes. Extracting the global correlation feature matrix F across frequency bands Trans .

5. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 4, characterized in that: In step S32, the specific fusion calculation process of the cross-attention mechanism is as follows: Feature fusion is achieved using a cross-attention mechanism: local defect features extracted by a convolutional neural network (CNN) are mapped to a query matrix. The global correlation feature matrix F extracted by Transformer Trans Mapped to a key matrix AND-value matrix By calculating the correlation weights between local and global features, adaptive enhancement of weak defect signals is achieved. Calculate the feature fusion tensor using the following formula. : ； in, represents the feature dimension of the key matrix.

6. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 5, characterized in that: In step S4, the classifier softmax model includes a fully connected layer and a softmax output layer. The classifier softmax model uses the amplitude frequency domain correlation features extracted in the early stage to identify thermal aging defects, water tree defects, and shielding layer damage defects of the cable body insulation. It also uses the phase transfer law features to identify short circuit faults, open circuit faults, and high-resistance grounding faults. Finally, the softmax output layer outputs the comprehensive identification probability of each defect and fault.

7. The method for identifying defects and faults in cross-linked polyethylene cables according to claim 6, characterized in that: Step S4 also includes an assessment of the severity of the defect: Calculate the feature fusion tensor The L2 norm is calculated, and the frequency offset of the first measured impedance resonance peak is extracted. The L2 norm distance and frequency offset are used as additional physical constraints input to the classifier to correct and output the final defect identification result. Calculate the defect severity index using the following formula. : ； in, This indicates the calculation of the L2 norm. To measure the resonant frequency, The standard resonant frequency of a good cable. and These are the preset weighting coefficients.

Images