High-voltage cable buffer layer ablation detection method and system based on waveform atlas

Abstract

The invention relates to a high-voltage cable buffer layer ablation detection method and system based on a waveform atlas, and the method comprises the following steps: S1, obtaining multi-modal data of a high-voltage cable, and carrying out the preprocessing of the multi-modal data; s2, extracting waveform time-frequency domain features, vibration features and infrared features, and integrating the features into a multi-modal feature vector training set; s3, constructing an intelligent analysis model based on the multi-modal deep learning network, and training based on the multi-modal feature vector training set to obtain a trained intelligent analysis model; s4, multi-modal data collected in real time are input into the trained intelligent analysis model, and ablation results including whether ablation exists or not, the ablation position and the severity level are obtained; and S5, based on the output of the intelligent analysis model, in combination with a multi-mode signal collected in real time, visualization is performed on an ablation result through waveform superposition and three-dimensional modeling, and according to the severity level, a grading alarm mechanism is triggered, and a real-time notification is pushed. According to the invention, the efficiency, precision and reliability of high-voltage cable fault detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of high-voltage cable fault detection, and particularly to a method and system for detecting ablation of the buffer layer of high-voltage cables based on waveform spectra. Background Art

[0002] High-voltage cables are key infrastructure in modern power transmission networks, and their safe and reliable operation directly affects the stability and efficiency of the power system. However, during long-term operation, due to factors such as overvoltage, environmental corrosion, temperature and humidity changes, and cable aging, the buffer layer of the cable (a key protective structure between the insulating layer and the outer sheath) may experience ablation. This ablation not only reduces the insulation performance of the cable but may also lead to serious consequences such as cable breakdown, short circuit, and high-temperature faults, threatening the safety of power transmission. Therefore, early detection and diagnosis of high-voltage cable buffer layer ablation are crucial for the operation and maintenance of the power system.

[0003] Traditional cable detection methods, such as manual inspections, off-line tests, and single-modal reflected wave analysis, have problems such as low detection accuracy, long time consumption, and difficulty in real-time monitoring. Especially when the fault locations are unevenly distributed or the environmental noise interference is significant, the detection effect may drop significantly. Summary of the Invention

[0004] To solve the above problems, the purpose of the present invention is to provide a method and system for detecting ablation of the buffer layer of high-voltage cables based on waveform spectra, significantly improving the efficiency, accuracy, and reliability of high-voltage cable fault detection.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for detecting ablation of the buffer layer of high-voltage cables based on waveform spectra, comprising the following steps:

[0007] S1: Obtain multimodal data of the high-voltage cable, including cable reflection waveform signals, infrared signals, and vibration change signals, and preprocess them;

[0008] S2: Based on the preprocessed multimodal data, extract waveform time-frequency domain features, vibration features, and infrared features, and integrate the extracted time-frequency domain features, vibration features, and infrared features into a multimodal feature vector training set;

[0009] S3: Construct an intelligent analysis model based on a multimodal deep learning network, and train it based on the multimodal feature vector training set to obtain a trained intelligent analysis model;

[0010] S4: Input the real-time collected multimodal data into the trained intelligent analysis model to obtain ablation results, including whether ablation exists, ablation location, and severity level;

[0011] S5: Based on the output of the intelligent analysis model, combined with the multi-modal signals collected in real time, visualize the ablation results through waveform superposition and three-dimensional modeling, and trigger a hierarchical alarm mechanism according to the severity level to push real-time notifications.

[0012] Further, S1 is specifically as follows:

[0013] Inject spike pulse signals into the cable through low-frequency and high-frequency pulse signal generators to obtain cable reflection waveform signals; at the ablation position of the buffer layer, collect infrared signals through additional infrared detection equipment; at the ablation point, collect vibration change signals through vibration sensors;

[0014] Align all signals through a unified time reference to construct a data set containing normal and various ablation conditions.

[0015] Further, the time-frequency domain features include extracting time-domain and frequency-domain features reflecting ablation characteristics from the cable reflection waveform signals, specifically as follows;

[0016] Calculate the amplitude and morphological changes of the cable reflection waveform signal to obtain the reflected wave amplitude change A diff :

[0017] A diff =|A burn -A normal |;

[0018] Wherein, A burn is the waveform amplitude of the ablation point, and A normal is the amplitude of the normal waveform;

[0019] Use the normalized Euclidean distance to measure the difference between the ablation point and the normal point waveforms to obtain the waveform distortion coefficient C dist :

[0020]

[0021] Where x burn,i and x normal,i are the amplitude points of the ablation waveform and the standard waveform respectively, and n is the number of sampling points;

[0022] Obtain the gradient change of the cable reflection waveform signal:

[0023]

[0024] Perform Fourier transform on the cable reflection waveform signal, transfer the time-domain signal to the frequency domain, and analyze the frequency components and energy characteristics, including the main peak frequency f peak 、high-frequency component energy ratio P high 、

[0025]

[0026] Among them, X(f) is the spectrum of the cable reflection waveform signal, and f is the frequency; f t is the threshold frequency in the high-frequency band;

[0027] The cable reflection waveform signal is decomposed into multi-resolution signals by using wavelet decomposition, and the local feature W of the signal in different frequency bands is extracted Q ;

[0028] The extracted cable reflection waveform signal features are integrated into a feature vector F wave :

[0029] F wave = [A diff , C dist , G wave , f peak , P high , W Q .

[0030] Furthermore, the extracted time-frequency domain features, vibration features, and infrared features are integrated into a multi-modal feature vector training set, specifically as follows:

[0031] Vibration features are extracted according to the vibration change signals collected by vibration sensors, including the root mean square value R vib , the main vibration frequency f vib , the peak factor, and the total energy E of the vibration signal vib , and the vibration feature vector is expressed as:

[0032] F vib = [R vib , C peak , f vib , E vib ;

[0033] The temperature features in the infrared signal are extracted to reflect the abnormal temperature distribution of the ablation point, including the hot spot temperature T max , the average temperature T avg , the temperature difference between the hot spot area and the surrounding environment T grad , and the hot spot area ratio R calculated based on infrared image segmentation hot , then the infrared feature vector is expressed as:

[0034] F IR = [T max , T avg , T grad , R hot ;

[0035] The time-frequency domain feature vector F wave , the vibration feature vector F vib and the infrared feature vector F IRConcatenate and integrate to obtain the final multi-modal feature vector F multi ;

[0036] Construct a multi-modal feature training set D by vectorizing the features of the sample data train :

[0037]

[0038] wherein is the multi-modal feature vector of the i-th sample, and Y i is the corresponding annotation

[0039] Furthermore, the multi-modal deep learning network includes a multi-modal key feature extractor, a feature fusion module, and an output layer; the multi-modal feature extractor designs multiple parallel feature extraction branches to process the three feature vectors of waveform, vibration, and infrared respectively, and uses a convolutional neural network, a long short-term memory network, and a fully connected layer to extract key features respectively; the feature fusion module fuses the key features based on the attention mechanism and outputs the fused features; the output layer adopts multi-task prediction, including fault classification, location regression, and severity assessment

[0040] Furthermore, the multi-modal feature extractor includes a convolutional neural network, a long short-term memory network, and a fully connected layer branch to process the three feature vectors of waveform, vibration, and infrared respectively, specifically as follows

[0041] The convolutional neural network branch, according to the waveform feature Fwave, extracts the enhanced waveform feature F′ based on the spatio-temporal local feature extraction of the MSC-CNN and SE modules wave ;

[0042] The long short-term memory network branch, based on the vibration feature Fvib, uses LSTM to capture the time series feature F′ of the vibration data vib ;

[0043] The fully connected layer branch, for the infrared thermal imaging data FIR, uses a multi-layer fully connected network to extract the high-dimensional temperature feature F′ IR

[0044] Furthermore, the enhanced waveform feature extracted based on the spatio-temporal local feature extraction of the MSC-CNN and SE modules is specifically as follows

[0045] Based on the waveform feature F wave , use multiple convolution kernels k of different sizes to extract multi-scale features

[0046]

[0047] where is the weight of the l-th convolution kernel; * represents the convolution operation ​is the bias term; σ is the activation function;

[0048] Concatenate the multi-scale features extracted by different convolutional kernels together:

[0049]

[0050] Perform global average pooling on each channel of F msc to compress the time dimension and obtain the channel description vector z c :

[0051]

[0052] where T' is the time step after convolution and pooling, and C' is the number of channels after concatenation;

[0053] Input z c into a two-layer fully connected network to calculate the channel weight α c :

[0054] Weight each channel of F msc to obtain the enhanced waveform feature F′ wave :

[0055] F′ wave = α c ·F msc (t, c).

[0056] Furthermore, the feature fusion module combines the features of each modality and designs an attention-weighted fusion module to highlight the key features and achieve modality complementarity, specifically as follows:

[0057] Concatenate the features of each modality into a joint representation:

[0058] F′ multi = Concat(F′ wave , F′ vib , F′ IR )

[0059] Input the joint representation into an attention network to learn the importance weights α wave , α vib , α IR :

[0060] α i′ = Softmax(W a ·F multi + b a ), i′ ∈ {wave, vib, IR};

[0061] Calculate the fused features using a weighted strategy, and during the fusion process, enhance the mutual information through cross-modal attention, calculate the correlation between different modal features, and output the enhanced interaction features:

[0062]

[0063] Among them, F fused (wave, vib), F fused (wave, IR), F fused (vib, IR) are the interaction features after enhancement of waveform, vibration, and infrared respectively; W q , W k , W v are the projection matrices of query, key, and value respectively; is the normalization factor, and d is the modal dimension (i.e., $C);

[0064] Finally, fuse the features of all modalities:

[0065] F final = F fused (wave, vib) + F fused (wave, IR) + F fused (vib, IR).

[0066] Furthermore, for the output layer, design multi-task outputs, including a classification task, a location regression task, and a severity regression task. Specifically:

[0067] The classification task uses a Softmax classifier to determine whether there is an ablation fault in the current data:

[0068] P class = Softmax(W class ·F final + b class );

[0069] Among them, W class and b class are the weight and bias of the classification task respectively;

[0070] For the location regression task, use a fully connected regression output module to predict the ablation location L pos :

[0071] L pos = W pos ·F final + b pos ;

[0072] Among them, W pos and b posThey are the weight and bias for the position regression task respectively;

[0073] For the severity regression task, a fully connected regression network is used to output the ablation severity score Ssev:

[0074] S sev = W sev ·F final + b sev ;

[0075] where W sev and b sev are the weight and bias for the severity regression task respectively;

[0076] And a shared branch Fshared is introduced, which is connected to the independent fully connected layers of the classification task, position regression task and severity regression task respectively to reduce redundant feature learning.

[0077] A high-voltage cable buffer layer ablation detection system based on waveform atlas, comprising a data acquisition unit, a feature extraction unit, an intelligent analysis unit and a visualization unit:

[0078] The data acquisition unit acquires multimodal data of the high-voltage cable, including cable reflection waveform signals, infrared signals, and vibration change signals, and performs preprocessing;

[0079] The feature extraction unit extracts waveform time-frequency domain features, vibration features and infrared features based on the preprocessed multimodal data, and integrates the extracted time-frequency domain features, vibration features and infrared features into a multimodal feature vector training set;

[0080] The intelligent analysis unit inputs the trained intelligent analysis model according to the real-time acquired multimodal data to obtain ablation results, including whether ablation exists, ablation location and severity level (mild, moderate, severe);

[0081] The visualization unit visualizes the ablation results based on the output of the intelligent analysis unit, combined with the real-time acquired multimodal signals, through waveform superposition and three-dimensional modeling, and triggers a hierarchical alarm mechanism according to the severity level to push real-time notifications.

[0082] The present invention has the following beneficial effects:

[0083] 1. The present invention can not only achieve centimeter-level fault location and severity level assessment, but also trigger a hierarchical alarm mechanism and provide maintenance suggestions, thus significantly improving the efficiency, accuracy and reliability of high-voltage cable fault detection

[0084] 2. The present invention comprehensively utilizes cable reflection waveform signals, infrared signals, and vibration change signals. Through the collaborative analysis of multi-modal data, it covers various manifestations of ablation faults (time domain, frequency domain, temperature anomaly, vibration characteristics). Different modal data provides complementary information, improving the detection accuracy and robustness of the model for ablation faults;

[0085] 3. The time-frequency domain feature extraction method of the present invention can capture the changes of waveforms at different time scales, comprehensively characterize the waveform characteristics of ablation. The multi-scale features extracted by MSC-CNN and the channel weighting mechanism of the SE module can enhance key features and improve the accuracy of time-frequency domain analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0087] The following further describes the present invention in detail with reference to the drawings and specific embodiments:

[0088] Refer to Figure 1 In this embodiment, a method for detecting ablation of the buffer layer of high-voltage cables based on waveform maps is provided, including the following steps:

[0089] S1: Obtain multi-modal data of high-voltage cables, including cable reflection waveform signals, infrared signals, and vibration change signals, and perform preprocessing;

[0090] S2: Based on the preprocessed multi-modal data, extract waveform time-frequency domain features, vibration features, and infrared features, and integrate the extracted time-frequency domain features, vibration features, and infrared features into a multi-modal feature vector training set;

[0091] S3: Construct an intelligent analysis model based on a multi-modal deep learning network, and train it based on the multi-modal feature vector training set to obtain a trained intelligent analysis model;

[0092] S4: Input the real-time collected multi-modal data into the trained intelligent analysis model to obtain ablation results, including whether ablation exists, ablation location, and severity level (mild, moderate, severe);

[0093] S5: Based on the output of the intelligent analysis model, combined with the real-time collected multi-modal signals, visualize the ablation results through waveform superposition and three-dimensional modeling, and trigger a hierarchical alarm mechanism according to the severity level to push real-time notifications.

[0094] In this embodiment, S1 is specifically:

[0095] Inject a spike pulse signal into the cable through a low-frequency and high-frequency pulse signal generator to obtain the cable reflection waveform signal; at the ablation position of the buffer layer, collect the infrared signal through an additional infrared detection device; at the ablation point, collect the vibration change signal through a vibration sensor;

[0096] Align all signals with a unified time reference to construct a data set containing normal and various ablation conditions (such as reference waveforms, defect waveforms, infrared anomalies, and vibration data).

[0097] In this embodiment, the time-frequency domain features include the time-domain and frequency-domain features reflecting the ablation characteristics extracted from the cable reflection waveform signal (such as the TDR waveform), specifically as follows;

[0098] Calculate the amplitude and morphological changes of the cable reflection waveform signal to obtain the reflection wave amplitude change A diff :

[0099] A diff =|A burn -A normal |;

[0100] Among them, A burn is the waveform amplitude of the ablation point, and A normal is the amplitude of the normal waveform;

[0101] Use the normalized Euclidean distance to measure the difference between the ablation point and the normal point waveforms to obtain the waveform distortion coefficient C dist :

[0102]

[0103] where x burn,i and x normal,i are the amplitude points of the ablation waveform and the standard waveform respectively, and n is the number of sampling points;

[0104] Obtain the gradient change of the cable reflection waveform signal:

[0105]

[0106] Perform Fourier transform on the cable reflection waveform signal to transfer the time-domain signal to the frequency domain, and analyze the frequency components and energy characteristics, including the main peak frequency f peak , the high-frequency component energy ratio P high ,

[0107]

[0108]

[0109] Among them, X(f) is the spectrum of the cable reflection waveform signal, and f is the frequency; f tis the threshold frequency in the high-frequency band;

[0110] Use wavelet decomposition to decompose the cable reflection waveform signal into multi-resolution signals, and extract the local features W of the signal in different frequency bands Q ;

[0111] Integrate the extracted cable reflection waveform signal features into a feature vector F wave :

[0112] F wave =[A diff ,C dist ,G wave ,f peak ,P high ,W Q .

[0113] In this embodiment, the extracted time-frequency domain features, vibration features, and infrared features are integrated into a multi-modal feature vector training set, specifically as follows:

[0114] Extract vibration features according to the vibration change signals collected by vibration sensors, including root mean square value R vib , main vibration frequency f vib , peak factor, and total vibration signal energy E vib , and the vibration feature vector is expressed as:

[0115] F vib =[R vib ,C peak ,f vib ,E vib ;

[0116] Extract the temperature features in the infrared signal to reflect the abnormal temperature distribution of the ablation point, including hot spot temperature T max , average temperature T avg , temperature difference between the hot spot area and the surrounding environment T grad , and the proportion R of the hot spot area calculated based on infrared image segmentation hot , then the infrared feature vector is expressed as:

[0117] F IR =[T max ,T avg ,T grad ,R hot ;

[0118] Concatenate and integrate the time-frequency domain feature vector F wave , vibration feature vector F vib , and infrared feature vector F IR into the final multi-modal feature vector F multi ;

[0119] Construct a multi-modal feature training set D by vectorizing the features of the sample data train :

[0120]

[0121] wherein, is the multi-modal feature vector of the i-th sample, and Y i is the corresponding annotation (ablation location and severity level).

[0122] In this embodiment, the multi-modal deep learning network includes a multi-modal key feature extractor, a feature fusion module, and an output layer; the multi-modal feature extractor designs multiple parallel feature extraction branches to process the waveform, vibration, and infrared feature vectors respectively, and uses a convolutional neural network, a long short-term memory network, and a fully connected layer to extract key features respectively; the feature fusion module fuses the key features based on the attention mechanism and outputs the fused features; the output layer uses multi-task prediction, including fault classification, location regression, and severity assessment.

[0123] In this embodiment, the multi-modal feature extractor includes a convolutional neural network, a long short-term memory network, and a fully connected layer branch, which process the waveform, vibration, and infrared feature vectors respectively, as follows:

[0124] The convolutional neural network branch, according to the waveform feature Fwave, extracts the enhanced waveform feature F′ based on the spatio-temporal local feature extraction of the MSC-CNN and SE modules wave ; uses the CNN convolutional layer to extract the spatio-temporal local features of the waveform, and adds multi-scale convolution (MSC-CNN); different-sized convolutional kernels kk extract the multi-scale features of the waveform (capturing short-term and long-term dependencies): a channel attention mechanism (SE module) is added after the convolutional layer to learn the weighted weights of important channels;

[0125] The long short-term memory network branch, based on the vibration feature Fvib, uses LSTM to capture the time series feature F′ of the vibration data vib ;

[0126] The fully connected layer branch uses a multi-layer fully connected network to extract the high-dimensional temperature feature F′ for the infrared thermal imaging data FIR IR .

[0127] In this embodiment, the enhanced waveform feature extracted based on the spatio-temporal local feature extraction of the MSC-CNN and SE modules is as follows:[[]]

[0128] Based on the waveform feature F wave , use multiple different-sized convolutional kernels k to extract multi-scale features:

[0129]

[0130] where is the weight of the l-th layer convolutional kernel; * represents the convolution operation; is the bias term; σ is the activation function;

[0131] Concatenate the multi-scale features extracted by different convolutional kernels together:

[0132]

[0133] Perform global average pooling on each channel of F msc to compress the time dimension and obtain the channel description vector z c :

[0134]

[0135] where T' is the time step after convolution and pooling, and C' is the number of channels after concatenation;

[0136] Input z c into a two-layer fully connected network to calculate the channel weight α c :

[0137] Weight each channel of F msc to obtain the enhanced waveform feature F′ wave :

[0138] F′ wave =α c ·F msc (t, c).

[0139] In this embodiment, the feature fusion module combines the features of each modality and designs an attention weighted fusion module to highlight the key features and achieve modality complementarity, as follows:

[0140] Concatenate the features of each modality into a joint representation:

[0141] F′ multi =Concat(F′ wave , F′ vib , F′ IR )

[0142] Input the joint representation into an attention network to learn the importance weights α wave , α vib , α IR :

[0143] α i′ =Softmax(W a ·F multi +b a ), i′ ∈ {wave, vib, IR};

[0144] Calculate the fused features using a weighted strategy, and during the fusion process, enhance the mutual information through cross-modal attention, calculate the correlation between different modal features, and output the enhanced interaction features:

[0145]

[0146] Among them, F fused (wave, vib), F fused (wave, IR), F fused (vib, IR) are the interaction features after enhancement of waveform, vibration, and infrared respectively; W q , W k , W v are the projection matrices of Query, Key, and Value respectively; is the normalization factor, and d is the modal dimension (i.e., $C);

[0147] Finally, fuse the features of all modalities:

[0148] F final = F fused (wave, vib) + F fused (wave, IR) + F fused (vib, IR).

[0149] In this embodiment, for the output layer, multi-task output is designed, including a classification task, a location regression task, and a severity regression task. Specifically:

[0150] The classification task uses a Softmax classifier to determine whether there is an ablation fault in the current data:

[0151] P class = Softmax(W class ·F final + b class );

[0152] Among them, W class and b class are the weight and bias of the classification task respectively;

[0153] For the location regression task, use a fully connected regression output module to predict the ablation location L pos :

[0154] L pos = W pos ·F final + b pos ;

[0155] Among them, W posand b pos are the weight and bias of the position regression task respectively;

[0156] For the severity regression task, a fully connected regression network is used to output the ablation severity score Ssev:

[0157] S sev = W sev ·F final + b sev ;

[0158] where W sev and b sev are the weight and bias of the severity regression task respectively;

[0159] And a shared branch Fshared is introduced, and Fshared is connected to the independent fully connected layers of the classification task, position regression task and severity regression task respectively to reduce redundant feature learning.

[0160] In this embodiment, a high-voltage cable buffer layer ablation detection system based on waveform atlas is also provided, including a data acquisition unit, a feature extraction unit, an intelligent analysis unit and a visualization unit:

[0161] The data acquisition unit acquires multimodal data of the high-voltage cable, including cable reflection waveform signals, infrared signals, and vibration change signals, and performs preprocessing;

[0162] The feature extraction unit extracts waveform time-frequency domain features, vibration features and infrared features based on the preprocessed multimodal data, and integrates the extracted time-frequency domain features, vibration features and infrared features into a multimodal feature vector training set;

[0163] The intelligent analysis unit inputs the trained intelligent analysis model according to the real-time acquired multimodal data to obtain ablation results, including whether ablation exists, ablation position and severity level (mild, moderate, severe);

[0164] The visualization unit visualizes the ablation results based on the output of the intelligent analysis unit, combines with the real-time acquired multimodal signals, through waveform superposition and three-dimensional modeling, and triggers a hierarchical alarm mechanism according to the severity level to push real-time notifications

[0165] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0167] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0168] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0169] As mentioned above, it is only a preferred embodiment of the present invention, and it is not intended to limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas, characterized in that, It includes the following steps: S1: Obtain multimodal data of high-voltage cables, including cable reflection waveform signals, infrared signals, and vibration change signals, and perform preprocessing; S2: Based on the preprocessed multimodal data, extract waveform time-frequency domain features, vibration features, and infrared features, and integrate the extracted time-frequency domain features, vibration features, and infrared features into a multimodal feature vector training set; S3: Construct an intelligent analysis model based on a multimodal deep learning network, and train it based on the multimodal feature vector training set to obtain a trained intelligent analysis model; S4: Input the real-time collected multimodal data into the trained intelligent analysis model to obtain ablation results, including whether ablation exists, ablation location, and severity level; S5: Based on the output of the intelligent analysis model, combined with the real-time collected multimodal signals, visualize the ablation results through waveform superposition and 3D modeling, and trigger a hierarchical alarm mechanism according to the severity level to push real-time notifications.

2. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 1, characterized in that The specific content of S1 is as follows: Inject spike pulse signals into the cable through a low-frequency and high-frequency pulse signal generator to obtain cable reflection waveform signals; at the ablation position of the buffer layer, collect infrared signals through an additional infrared detection device; at the ablation point, collect vibration change signals through a vibration sensor; Align all signals through a unified time reference to construct a dataset containing normal and various ablation conditions.

3. The method for detecting ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 2, wherein, The time-frequency domain features include time-domain and frequency-domain features reflecting ablation characteristics extracted from the cable reflection waveform signals, specifically as follows; Calculate the amplitude and morphological changes of the cable reflection waveform signal to obtain the reflection wave amplitude change A diff : A diff = |A burn -A normal |; Among them, A burn is the waveform amplitude of the ablation point, and A normal is the amplitude of the normal waveform; The normalized Euclidean distance is used to measure the difference between the waveforms of the ablation points and the normal points, and the waveform distortion coefficient C is obtained dist : where x burn,i and x normal,i are the amplitude points of the ablation waveform and the standard waveform respectively, and n is the number of sampling points; Obtain the gradient change of the cable reflection waveform signal: Perform Fourier transform on the cable reflection waveform signal, migrate the time-domain signal to the frequency domain, and analyze the frequency components and energy characteristics, including the main peak frequency f peak , the energy ratio P of the high-frequency components high , where X(f) is the spectrum of the cable reflection waveform signal, f is the frequency; f t is the threshold frequency of the high-frequency band; The cable reflection waveform signal is decomposed into multi-resolution signals using wavelet decomposition, and the local feature W of the signal in different frequency bands is extracted Q ; Integrate the extracted cable reflection waveform signal features into a feature vector F wave : F wave = [A diff , C dist , G wave , f peak , P high , W Q .

4. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 3, wherein The integration of the extracted time-frequency domain features, vibration features, and infrared features into a multimodal feature vector training set is specifically as follows: Extract vibration features according to the vibration change signals collected by vibration sensors, including the root mean square value R vib , the main vibration frequency f vib , the peak factor, and the total energy E of the vibration signal vib , and the vibration feature vector is expressed as: F vib = [R vib , C peak , f vib , E vib ; Extract the temperature characteristics in the infrared signal to reflect the abnormal temperature distribution of the ablation point, including the hot spot temperature T max , average temperature T avg , temperature difference between the hot spot area and the surrounding environment T grad , proportion of the hot spot area R calculated based on the infrared image segmentation hot , then the infrared feature vector is expressed as: F IR = [T max , T avg , T grad , R hot ; The time-frequency domain feature vector F wave , the vibration feature vector F vib and the infrared feature vector F IR are spliced and integrated into the final multi-modal feature vector F multi ; Construct a multi-modal feature training set D by vectorizing the features of the sample data train : Among them, is the multi-modal feature vector of the i-th sample, and Y i is the corresponding annotation.

5. The method for detecting ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 1, characterized in that, The multimodal deep learning network includes a multimodal key feature extractor, a feature fusion module, and an output layer; the multimodal feature extractor designs multiple parallel feature extraction branches to process waveform, vibration, and infrared feature vectors respectively, and uses a convolutional neural network, a long short-term memory network, and a fully connected layer to extract key features; the feature fusion module fuses the key features based on the attention mechanism and outputs the fused features; the output layer uses multitask prediction, including fault classification, location regression, and severity assessment.

6. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 5, wherein The multimodal feature extractor includes a convolutional neural network, a long short-term memory network, and a fully connected layer branch to process waveform, vibration, and infrared feature vectors respectively, specifically as follows: The convolutional neural network branch extracts the enhanced waveform feature F′ based on the spatio-temporal local feature of MSC-CNN and SE module according to the waveform feature Fwave wave ; The long short-term memory network branch captures the time series feature F' of the vibration data using LSTM based on the vibration feature Fvib vib ; For the fully connected layer branch, for the infrared thermal imaging data FIR, a multi-layer fully connected network is used to extract high-dimensional temperature features F′ IR .

7. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 6, wherein The waveform features enhanced by spatio-temporal local feature extraction based on MSC-CNN and SE modules are specifically as follows: Based on waveform feature F wave , multiple convolutional kernels k of different sizes are used to extract multi-scale features: where is the weight of the convolutional kernel in the l-th layer; * represents the convolution operation; is the bias term; σ is the activation function; Stitch together the multi-scale features extracted by different convolutional kernels: Perform global average pooling on each channel of F msc to compress the time dimension and obtain the channel description vector z c : Among them, T' is the time step after convolution and pooling, and C' is the number of channels after stitching; Input z c into a two-layer fully connected network to calculate the channel weights α c : Weight each channel of F msc to obtain the enhanced waveform feature F′ wave : F′ wave = α c ·F msc (t, c).

8. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 6, wherein The feature fusion module combines the features of each modality and designs an attention-weighted fusion module to highlight key features and achieve modality complementarity, specifically as follows: Stitch the features of each modality into a joint representation: F′ multi = Concat(F′ wave , F′ vib , F′ IR ) Input the joint representation into an attention network to learn the importance weights α for each feature modality wave , α vib , α IR : α i′ = Softmax(W a ·F multi + b a ), i′ ∈ {wave, vib, IR}; Use a weighted strategy to calculate the fused features, and during the fusion process, enhance the mutual information through cross-modal attention, calculate the correlation between different modality features, and output the enhanced interaction features: Among them, F fused (wave, vib), F fused (wave, IR), F fused (vib, IR) are the interaction features after waveform, vibration, and infrared enhancement respectively; W q , W k , W v are the projection matrices of Query, Key, and Value respectively; is the normalization factor, and d is the modal dimension (i.e., $C); Finally, fuse the features of all modalities: F final = F fused (wave, vib) + F fused (wave, IR) + F fused (vib, IR).

9. The method for detecting the ablation of the buffer layer of a high-voltage cable based on a waveform atlas according to claim 5, characterized in that, The output layer is designed for multi-task output, including a classification task, a location regression task, and a severity regression task. Specifically: The classification task uses a Softmax classifier to determine whether there is an ablation fault in the current data: P class = Softmax(W class ·F final + b class ); Among them, W class and b class are the weight and bias of the classification task respectively; For the position regression task, a fully connected regression output module is used to predict the ablation position L pos : L pos = W pos · F final + b pos ; Among them, W pos and b pos are the weight and bias of the position regression task respectively; For the severity regression task, a fully connected regression network is used to output the ablation severity score Ssev: S sev = W sev · F final + b sev ; Among them, W sev and b sev are the weight and bias of the severity regression task, respectively; And a shared branch Fshared is introduced, which is connected to the independent fully connected layers of the classification task, the location regression task, and the severity regression task respectively to reduce redundant feature learning.

10. A high-voltage cable buffer layer ablation detection system based on waveform maps, characterized in that, It includes a data acquisition unit, a feature extraction unit, an intelligent analysis unit, and a visualization unit: The data acquisition unit obtains multi-modal data of high-voltage cables, including cable reflection waveform signals, infrared signals, and vibration change signals, and performs preprocessing; The feature extraction unit extracts waveform time-frequency domain features, vibration features, and infrared features based on the preprocessed multi-modal data, and integrates the extracted time-frequency domain features, vibration features, and infrared features into a multi-modal feature vector training set; The intelligent analysis unit inputs the real-time acquired multi-modal data into the trained intelligent analysis model to obtain ablation results, including whether ablation exists, ablation location, and severity level; The visualization unit visualizes the ablation results based on the output of the intelligent analysis unit, combined with the real-time acquired multi-modal signals, through waveform superposition and 3D modeling, and triggers a hierarchical alarm mechanism according to the severity level to push real-time notifications.

Images