A systematic monitoring method for electrical safety performance of a cable

Abstract

The application discloses a kind of systematic monitoring methods of cable electrical safety performance, it is related to electrical safety field, the application includes collection sensor data to edge computing node pre-processing data output as real-time calculation temperature data, local point data, resistance capacitance data and vibration data;Data is sent to hidden danger classification model and fault location model in cloud, model output early warning information to operation and maintenance engineering processing;Sensor data transmission process is discriminated according to data frequency, when discrimination is greater than 1kHz, load 1D-CNN arc detection model, when discrimination is less than 1kHz load LightGBM feature classification model, periodically load 24h sensor data less than 1kHz into LSTM-Attention gradual analysis model;The relevant data of local discharge generated by 1D-CNN arc detection model discrimination is sent to LSTM-Attention gradual analysis model for evaluating the local discharge trend of global cable.The application uses classification model output as prior knowledge of positioning model, which can improve positioning accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electrical safety, in particular to a systematic monitoring method for electrical safety performance of a cable. BACKGROUND

[0002] At present, the electrical safety management of cables still faces severe challenges. Cable insulation damage, aging and peeling, mechanical damage, uneven construction quality, and the flooding of counterfeit and inferior products are relatively prone to occur, and relevant electrical fire accidents occur frequently. In addition, with the continuous development of urbanization, influenced by factors such as dense cable laying, non-standard cable hidden in the market, and insufficient maintenance of old lines, the electrical safety risk is further increased. However, the application of new technologies has not been popularized, resulting in hidden dangers that are difficult to be discovered in time. In addition, the blurring of responsibility subjects (such as no clear responsible party for the maintenance of lines behind residential electricity meters) further exacerbates the management blind spot, and a systematic solution is urgently needed.

[0003] At present, several cable electrical safety prevention measures cannot form a systematic monitoring and positioning method, and how to systematically judge the specific type and location is a problem that needs to be solved. SUMMARY

[0004] The present application provides a systematic monitoring method for electrical safety performance of a cable, which solves the problems of the prior art.

[0005] In a first aspect, the present application provides a systematic monitoring method for electrical safety performance of a cable, comprising:

[0006] Collecting sensor data to an edge computing node to preprocess data output as real-time calculation temperature data, partial discharge data, resistance capacitance data, and vibration data;

[0007] Sending the data to a hidden danger classification model and a fault positioning model in the cloud, and outputting warning information from the model to operation and maintenance engineering processing;

[0008] Among them, the sensor data sending process is judged according to the data frequency, when the judgment is greater than 1kHz, the 1D-CNN arc detection model is loaded, when the judgment is less than 1kHz, the LightGBM feature classification model is loaded, and the 24h sensor data less than 1kHz is periodically loaded into the LSTM-Attention gradual analysis model;

[0009] Among them, the relevant data of the 1D-CNN arc detection model judgment producing partial discharge is sent to the LSTM-Attention for training, and the trained LSTM-Attention is used to evaluate the partial discharge trend of the global cable.

[0010] Further, the fault location model comprises: a cable network graph structure model, sensors are set as sensor nodes, each node is set with node features of temperature, current, humidity, partial discharge amount, cable sections are set between sensor nodes, each cable section and sensor node are numbered, and cable section features are set as cable length, cross-sectional area, laying depth, topological distance, etc., an adjacency matrix of the cable section and the sensor node is constructed based on Gaussian kernel weighting for performing message passing, a message passing mechanism is to splice and link cable section features and adjacent sensor node features to form a message function, the message function is updated in real time based on the maximum value and total value of the node position-based fault rate, the message function is transmitted between different levels of sensor nodes, the numbering of the same sensor node and cable section under different branches is totaled as a level, the branch relationship of multiple levels is distinguished from large to small to form different connection layers, the message function is transmitted and updated between all global fully connected layers, each connection layer is provided with a fault probability, and the fault probability is determined by output data of the hidden danger classification model.

[0011] Further, the fault probability is the maximum value of the fault probabilities of multiple sensor nodes under the same connection layer, and the numbering position data of the sensor nodes of the level under the connection layer is located based on the fault probability.

[0012] Further, the sensor data used for training the hidden danger classification model further comprises fault data generated by implementing finite element simulation based on COMSOL Multiphysics software, and the generation steps are as follows:

[0013] Step 1: geometric construction of cable layered structure: conductor, insulating layer, shielding layer, sheath and joint component;

[0014] Step 2: setting materials and corresponding attributes of the cable layered structure;

[0015] Step 3: constructing coupling equations of electric field, thermal field and structural mechanics field of the cable layered structure;

[0016] Step 4: importing fault scenarios to obtain simulation implementation of insulation aging, joint loosening, overload operation and moisture intrusion;

[0017] Step 5: setting boundary conditions corresponding to the electric field, thermal field and structural mechanics field respectively, applying rated current / voltage at the conductor end in the electric field, convective heat dissipation on the outer surface of the cable in the thermal field, forced air cooling at the joint, and setting fixed cable end displacement in the structural mechanics field;

[0018] Step 6: generating fault data for supplementing training of the hidden danger classification model.

[0019] Further, the collection sensor data to edge computing node pre-process data output is real-time calculation temperature data, partial discharge data, resistance capacitance data and vibration data, wherein the data pre-processing for the LSTM-Attention gradual change analysis model specifically includes:

[0020] The temperature data is normalized by a sliding window, and the processing window is a multiple of 24h, and the multiple is at least 1 times;

[0021] The correlation coefficient of the current data and the temperature is calculated;

[0022] The partial discharge data is extracted to pulse count / amplitude mean.

[0023] Further, the LSTM-Attention gradual change analysis model accesses the preprocessed data and encodes the time-dependent, and after adding the Attention mechanism to focus on the key period, the fault location model is combined to classify and output the fault hidden danger and correspond to the location.

[0024] Further, the LSTM-Attention gradual change analysis model accesses the preprocessed data and encodes the time-dependent, and after adding the Attention mechanism to focus on the key period, the fault location model is combined to classify and output the fault hidden danger and correspond to the location.

[0025] Further, wherein the 1D-CNN arc detection model performs the process of hidden danger identification as follows:

[0026] The 1D-CNN arc detection model accesses the preprocessed data types including partial discharge data, vibration signal and temperature mutation;

[0027] The pre-processing method of the partial discharge data is 200MHz sampling, band-pass filtering pulse alignment, and the input dimension is 2000 points / pulse;

[0028] The pre-processing method of the vibration signal is STFT time-frequency diagram to 1D vector, and the input dimension is 512 points / frequency axis;

[0029] The processing method of the temperature mutation signal is difference processing, and the input dimension is 120 points / 10-minute window;

[0030] The processed data is substituted into the 1D-CNN arc detection model for multi-scale convolution feature extraction, and a feature fusion strategy is performed on the data types related to multi-sensor data, specifically including fusing the features of the PD pulse and the vibration signal, and outputting the arc, vibration position and duration of the cable joint by the fault positioning model.

[0031] Further, the LightGBM feature classification model is used to load sensor data less than 1 kHz for monitoring the aging / wet abnormal state of the cable, and the corresponding data is labeled with the aging / wet abnormal state label, specifically including:

[0032] First, the timestamps of different sensor data are aligned and normalized to the same interval to eliminate the dimension, and after preliminary labeling by an unsupervised method, the original sensor data types and corresponding basic statistical features are constructed, the basic statistical features including mean, variance, maximum value and minimum value; the advanced statistical features include: time window features: statistical quantities of different time windows (such as 1 hour, 6 hours and 24 hours); correlation features: such as the covariance of temperature and humidity, and the correlation coefficient of current and temperature; trend features: such as the slope of linear regression (reflecting the temperature rise trend); periodic features: such as the mean value at the same time of the day (eliminating the influence of day and night); physical model features: constructed according to the physical properties of the cable, such as: insulation resistance decline rate: (current resistance - initial resistance) / running time, and capacitance change rate: (current capacitance - initial capacitance) / running time;

[0033] The importance of the features is set, and then the model is trained according to the normal state, aging state and wet state three state labels, and the data is labeled.

[0034] Further, the GNN aggregation model is also used to embed the sensor nodes into vectors, the processed high-frequency data is loaded into the 1D-CNN arc detection model for analysis, and the processed low-frequency data is loaded into the LightGBM feature classification model for analysis.

[0035] The systematic monitoring method for the electrical safety performance of the cable provided by the application uses the classification model output as prior knowledge of the positioning model, which can improve the positioning accuracy;

[0036] The application calculates the temperature gradient, current mutation and other basic features in real time, and can trigger emergency power-off within 5ms;

[0037] The application uses finite element simulation to generate fault data to accurately control the fault type, position and degree, expand the identification range of the model, and reduce the false alarm rate. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0039] Figure 1 GNN aggregation data classification flowchart provided for embodiment 1 of the application.

[0040] Figure 2 Cascade-fusion flowchart involving LightGBM feature classification model, 1D-CNN arc detection model and LSTM-Attention gradual analysis model provided for embodiment 1 of the application. DETAILED DESCRIPTION

[0041] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to indicate the same or similar elements, unless otherwise represented. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the application.

[0042] First, the terms involved in the application are explained:

[0043] Classification of main hidden dangers related to cable electrical safety:

[0044] 1. Exposed live parts:

[0045] Hidden danger: cable insulation layer damage, aging cracking, improper joint treatment (such as insulation tape falling off, joint box damage) leading to direct exposure of the conductor.

[0046] Risk: direct contact of personnel with exposed conductors, resulting in electric shock accidents.

[0047] Common scenarios: old lines, construction damage, animal gnawing, mechanical force damage.

[0048] 2. Insulation failure:

[0049] Hidden danger: aging: long-term use, high temperature, ultraviolet radiation leading to degradation of insulation material performance (brittle, cracking, resistance reduction). Inferior materials: use of cables or insulation materials that do not meet standards. Environmental erosion: moisture, corrosive gases / liquids, oil pollution eroding insulation. Overvoltage: lightning or operating overvoltage breaking down insulation.

[0050] Risk: insulation failure makes the originally uncharged metal sheath, armor layer or adjacent conductor live, or causes short circuit between conductors to produce arc, increasing the risk of indirect electric shock or arc burn.

[0051] 3. Mechanical damage:

[0052] Hidden danger: Construction excavation, heavy pressure, sharp instrument scratch, excessive bending, vehicle crushing, animal gnawing, etc. cause the outer sheath or insulation layer of the cable to be damaged.

[0053] Risk: Directly cause the conductor to be exposed or the insulation performance to be reduced, and cause electric shock or short circuit.

[0054] 4. Wet or water accumulation environment:

[0055] Hidden danger: Water enters the cable trench, pipe, joint box, or the cable is in a wet environment for a long time.

[0056] Risk: Moisture reduces the insulation resistance, which may cause the leakage current to increase, and even form a conductive path along the wet surface, increasing the risk of indirect electric shock (especially for metal sheathed or armored cables). It is extremely dangerous to bring a charged cable into water.

[0057] 5. Overheating:

[0058] Hidden danger: Overload: The current exceeds the safe load current of the cable. Poor heat dissipation: Dense cable laying, too thin pipe, too thick buried depth, covered by thermal insulation materials, etc. cause heat accumulation. Poor contact: loose connection of joints and terminals, increased resistance and heat generation.

[0059] Risk: High temperature accelerates insulation aging and failure, and even ignites surrounding combustible materials to cause fire. Insulation failure increases the risk of electric shock.

[0060] 6. Poor / missing grounding or equipotential bonding:

[0061] Hidden danger: Metal cable sheath, armored layer, cable bridge not reliably grounded or with too large grounding resistance. The exposed conductive parts of electrical equipment are not reliably connected to the protective conductor (PE). Special places (such as bathrooms, swimming pools) are not locally equipotential bonded.

[0062] Risk: When the equipment insulation fails and the shell is charged, the fault current cannot be effectively introduced into the ground, and the contact voltage may reach a dangerous level, causing serious indirect electric shock accidents.

[0063] 7. Failure or absence of protection devices:

[0064] Hidden danger: No residual current operated protective device (RCD / leakage protector) is installed or is not reasonably installed. The circuit breaker or fuse is not properly selected or fails to timely cut off the power supply in case of short circuit or overload.

[0065] Risk: In case of leakage, short circuit, or overload, the fault circuit cannot be timely cut off, and the dangerous state continues to exist.

[0066] 8. Improper design, installation, and maintenance:

[0067] Hidden danger: wrong cable selection (voltage grade, current-carrying capacity, environmental adaptability). Improper laying method (such as too small bending radius, excessive stretching, insufficient distance from heat source). Poor joint manufacturing process. Lack of regular inspection, testing and maintenance.

[0068] Risk: long-term safety hazards are buried, and the probability of various faults is increased.

[0069] 9. Insufficient safety distance:

[0070] Hidden danger: insufficient distance between cable and accessible area, or insufficient electrical clearance and creepage distance from non-insulated live parts (especially at high voltage).

[0071] Risk: accidental touch or through tools, metal objects, or in high humidity, contaminated environment, air breakdown or surface flashover.

[0072] The present application classifies the above 9 types according to characteristic types, classifies the characteristic indexes corresponding to the characteristic classification, forms the early prediction-precise classification-space positioning trinity protection of the original system threshold upgrade to hidden danger, and corrects the improper behavior of workers by using installation, construction and maintenance specifications for the rest of the safety distance, design, installation and maintenance, to prevent electric shock accidents.

[0073] The specific application scenario of the present application is the cable electrical safety protection scenario in industrial, public environment and home scenarios.

[0074] The present application provides a systematic monitoring method for the electrical safety performance of cables, which aims to solve the above technical problems of the prior art.

[0075] The technical solutions of the present application and how the present application solves the above technical problems will be described in detail below. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0076] Example 1:

[0077] I. Deploy sensors and edge computing gateways:

[0078] The hidden danger classification model is loaded with physical sensor types, including high-precision Hall sensors for monitoring current and voltage, CT coils for monitoring load fluctuations and harmonic distortion, distributed optical fibers (DTS) for monitoring temperature, wireless temperature measurement tags for monitoring joints and cable bending at intervals of ≤5 meters, capacitive humidity sensors for monitoring humidity, which are deployed in cable trenches / wells and joint boxes (insulation resistance drops when humidity > 80%), high-frequency HFCT / UHF sensors for monitoring partial discharge, which are used to detect discharge pulses ≥5pC (early signs of insulation deterioration), and MEMS accelerometers for monitoring vibration data, which are deployed above the direct-buried section of the cable and the bridge laying section. An edge computing gateway is also deployed for edge computing and data processing.

[0079] II. Simulate the generation of fault data with labels;

[0080] Before collecting sensor data to the edge computing node and preprocessing data output as real-time calculation temperature data, partial discharge data, resistance and capacitance data, and vibration data, the present application does not meet the data requirements for setting up the hidden danger classification model of the system due to the relatively small number of real cable fault samples. The present application uses an accurate mathematical model of the cable constructed in a virtual environment to simulate the response of the cable under the action of multiple physical fields such as electricity, heat, and mechanical force by applying different fault conditions (such as insulation damage, overload, and moisture). The fault type, location, and degree can be accurately controlled, thereby generating a large amount of fault data with labels. The principles and steps of generating fault data based on COMSOL Multiphysics software finite element simulation are as follows:

[0081] Step 1: Construct the layered structure of the cable: conductor, insulation layer, shielding layer, sheath, and joint components.

[0082] Step 2: Set the materials and corresponding properties of the cable layered structure, including setting the electrical conductivity and thermal expansion coefficient of the copper conductor, the thermal conductivity, dielectric constant, and loss tangent of the XLPE insulation layer, and the nonlinear thermal conductivity of the silicone rubber joint.

[0083] Step 3: Construct the coupling equations of the electric field, thermal field, and structural mechanics field.

[0084] The electric field is used to solve the current continuity equation, the thermal field is used to solve the heat conduction equation, and the structural mechanics field is used to solve the thermal expansion deformation equation.

[0085] Step 4: Import the fault scenarios, obtain simulation implementations of four different fault types, the simulation implementation of insulation aging is to reduce the insulation layer conductivity / increase tan delta value, or add air gap defects; the simulation implementation of joint loosening is to increase the pressure contact resistance; the simulation implementation of overload operation is to increase the conductor current density to 150%-200% of the design value, and the simulation implementation of moisture intrusion is to add moisture at the insulation layer defect.

[0086] Step 5: Set boundary conditions corresponding to the electrical field, thermal field and structural mechanics field respectively, apply rated current / voltage at the conductor end in the electrical field, convective heat dissipation on the outer surface of the cable in the thermal field, forced air cooling at the joint, and set fixed cable end displacement constraints in the structural mechanics field;

[0087] Step 6: Generate a usable machine learning dataset for training the hidden danger classification model;

[0088] It also includes mixing real data and obtained simulation data according to a certain proportion to train the hidden danger classification model.

[0089] Based on steps 1-6, the partial discharge simulation data instance is as follows:

[0090] Based on the above partial discharge (PD) simulation generated data;

[0091] Modeling in COMSOL:

[0092] Set a bubble defect with a diameter of 0.1mm in the insulation layer;

[0093] Apply a 10kV AC voltage;

[0094] Output results:

[0095] Discharge pulse waveform (amplitude, phase, frequency);

[0096] Spatial electromagnetic field distribution;

[0097] # Generate pulse sequence samples

[0098] pulses = []

[0099] for phase in [30°, 90°, 270°]: # Discharge phase in power frequency period

[0100] pulse = simulate_pd(voltage=10e3, defect_size=0.1e-3, phase)

[0101] pulses.append(pulse)

[0102] III. Collecting real-time sensor data

[0103] The real-time sensor data is collected to the edge computing node to preprocess the data, specifically including:

[0104] The edge computing gateway processes to obtain the following feature types:

[0105] The feature types corresponding to joint overheating: temperature rise rate (Delta T / Delta t), three-phase temperature difference imbalance, thermal image hot spot diffusion gradient;

[0106] The feature types corresponding to moisture intrusion: humidity-insulation resistance correlation coefficient, capacitance increment ratio;

[0107] The feature types corresponding to insulation aging: leakage current harmonic component, dielectric loss tangent (tan delta), partial discharge amplitude / frequency;

[0108] The feature types corresponding to mechanical damage: vibration spectrum energy entropy, abnormal shock wave detection data.

[0109] IV. Cloud model receives data output warning information

[0110] The preprocessed data is output to the cloud platform analysis engine for processing, which loads the hidden danger classification model and fault positioning model, and outputs the warning information to the warning dashboard. The warning information is processed by the operation and maintenance work order system according to the warning dashboard, specifically including:

[0111] Among them, the hidden danger classification model constructed by the application corresponds to different data application scenarios, wherein the LightGBM feature classification model is suitable for medium-scale data, and the medium-scale data is specifically 10^4 sample order of magnitude. The advantage of using the LightGBM feature classification model is that the selected features have high visualization and strong interpretability.

[0112] The 1D-CNN arc detection model is used to detect partial discharge pulse waveform and vibration spectrum data, corresponding to partial discharge data caused by insulation aging and moisture intrusion and vibration spectrum data caused by mechanical damage. The 1D-CNN arc detection model can automatically extract the spatiotemporal features in the data.

[0113] The LSTM-Attention gradual change analysis model is used to monitor temperature and current data in long time series, and can capture gradual change hazards based on long-term memory dependence;

[0114] The specific analysis of each model in the cloud is as follows:

[0115] In the LSTM-Attention gradual change analysis model of the application, the key event focusing on Attention is focused: the weight is assigned to the LSTM output of each time step, and the scoring function is set, which will automatically amplify the weight of abnormal inflection points such as temperature mutation + humidity increase in the model, weaken the smooth period data, facilitate intuitive display, and high readability. The Attention weight of each monitoring point is mapped to the cable network graph structure model for summary, and the output is the located sensor node and abnormal period.

[0116] Regarding the 1D-CNN arc detection model, the input data preprocessing is as follows: the preprocessing method of partial discharge data is 200MHz sampling, band-pass filtering pulse alignment, and the input dimension is 2000 points / pulse; the preprocessing method of vibration signal is STFT time-frequency diagram to 1D vector, and the input dimension is 512 points / frequency axis; the processing method of temperature mutation signal is difference processing T t -T t-1 , and the input dimension is 120 points / 10 minutes window;

[0117] As Figure 2 shown, the cascade-fusion involving the LightGBM feature classification model, the 1D-CNN arc detection model and the LSTM-Attention gradual change analysis model is implemented in the cloud as follows:

[0118] When the sampling frequency is greater than 1kHz, the original waveform data is sent to the 1D-CNN arc detection model;

[0119] When the sampling frequency is less than 1kHz, the feature vector is extracted and sent to the LightGBM feature classification model;

[0120] A 24-hour data buffer area is established, and the LSTM-Attention gradual change analysis is triggered periodically;

[0121] The special processing of high-frequency data is as follows:

[0122] After the 1D-CNN model recognizes the partial discharge feature:

[0123] Real-time alarm is triggered;

[0124] The discharge feature data is uploaded to the cloud LSTM training queue;

[0125] The partial discharge data is used to train the global trend model to improve the prediction ability of the system for slow discharge development;

[0126] The process of line-of-sight positioning of the fault positioning model based on graph neural network is as follows:

[0127] 1. Cable network modeling:

[0128] Sensors are abstracted as nodes, with features including: temperature, current, humidity, partial discharge;

[0129] Cable segments are abstracted as edges, with features including: length, cross-sectional area, laying depth, topological distance;

[0130] Hierarchical division according to physical topology:

[0131] Trunk lines are at the highest level (Level 0);

[0132] Branch lines are downgraded in turn (Level 1-N);

[0133] Each node records its own level and parent node;

[0134] 2. Hierarchical message passing:

[0135] Message generation: nodes fuse their own features and connected cable features to form message packets;

[0136] Intra-level transmission: nodes of the same level exchange current state information;

[0137] Cross-level transmission:

[0138] Sub-nodes upload aggregated information to parent nodes;

[0139] Parent nodes broadcast global state downward;

[0140] Fault probability calculation:

[0141] Each round of transmission updates the node state vector;

[0142] The final output is the fault probability value of each node;

[0143] 3. Positioning decision process:

[0144] Hierarchical fault probability = maximum value of all node probabilities in that level;

[0145] After determining the highest probability level, in that level:

[0146] Filter fault nodes exceeding the threshold value;

[0147] Sort by probability value to generate candidate positions;

[0148] Output includes: fault level, node ID list, position coordinates, probability distribution.

[0149] The cascade pipeline hybrid model of the LightGBM feature classification model and the 1D-CNN arc detection model cooperates to realize the label on the data;

[0150] It also loads a GNN model to process multiple sensor nodes, and aggregates and diverts data to the LightGBM feature classification model and the 1D-CNN arc detection model;

[0151] The data processing procedure for multiple feature dimensions in the 1D-CNN arc detection model is as follows:

[0152] Perform multi-scale convolutional feature extraction;

[0153] Execute feature fusion strategy:

[0154] When two feature dimensions are involved:

[0155] The PD pulse feature pd_branch and the vibration signal feature vib_branch are processed separately, and then the two types of features are fused using a formula:

[0156] merged = Concatenate()([Flatten()(pd_branch), vib_branch])

[0157] Based on the time series dimension, after feature classification, locate the data and find the probability of anomalies at each time point;

[0158] When three feature dimensions are involved, the procedure is as follows:

[0159] import torch

[0160] import torch.nn as nn

[0161] class CableFaultCNN(nn.Module):

[0162] def __init__(self, in_channels=3): # 3 types of sensors

[0163] super().__init__()

[0164] # Branch 1: Large kernel captures long-term features

[0165] self.branch1 = nn.Sequential(

[0166] nn.Conv1d(in_channels, 32, kernel_size=51, stride=3,padding=25),

[0167] nn.BatchNorm1d(32),

[0168] nn.ReLU(),

[0169] nn.MaxPool1d(3, stride=2) )

[0171] # Branch 2: Medium kernel captures local features

[0172] self.branch2 = nn.Sequential(

[0173] nn.Conv1d(in_channels, 64, kernel_size=25, padding=12),

[0174] nn.BatchNorm1d(64),

[0175] nn.ReLU(),

[0176] nn.MaxPool1d(3, stride=2) )

[0178] # Branch 3: Small kernel captures details

[0179] self.branch3 = nn.Sequential(

[0180] nn.Conv1d(in_channels, 64, kernel_size=9, padding=4),

[0181] nn.BatchNorm1d(64),

[0182] nn.ReLU() )

[0184] # Feature fusion

[0185] self.fusion = nn.Sequential(

[0186] nn.Conv1d(32+64+64, 128, kernel_size=3, padding=1),

[0187] nn.ReLU(),

[0188] nn.AdaptiveMaxPool1d(1) )

[0190] self.classifier = nn.Linear(128, 4) # 4 classes of hazards

[0191] def forward(self, x):

[0192] b1 = self.branch1(x)

[0193] b2 = self.branch2(x)

[0194] b3 = self.branch3(x)

[0195] # unify time dimension

[0196] b1 = nn.functional.interpolate(b1, size=b3.size(2), mode='linear')

[0197] b2 = nn.functional.interpolate(b2, size=b3.size(2), mode='linear')

[0198] fused = torch.cat([b1, b2, b3],dim=1)

[0199] fused = self.fusion(fused).squeeze(2)

[0200] return self.classifier(fused)

[0201] The process of processing the label for the cascaded pipe mixed model of the LightGBM feature classification model and the 1D-CNN arc detection model is as follows:

[0202] According to the data frequency classification of the original sensor sampling data, the corresponding LightGBM feature classification model is used to load the sensor data less than 1 kHz for monitoring the aging / wet abnormal state of the cable, and the corresponding data is labeled with the aging / wet abnormal state label, which specifically includes:

[0203] First, the timestamps of different sensor data are aligned and normalized to the same interval to eliminate the dimension, and after preliminary labeling by an unsupervised method, the original sensor data type and the corresponding basic statistical features are constructed, the basic statistical features include mean, variance, maximum value, minimum value; The advanced statistical features include: time window features; correlation features; trend features; periodicity features; physical model features, set the importance of the features, then train the model according to the normal state, aging state and wet state three state labels, and label the data.

[0204] The processing procedure of GNN aggregated data for data aggregation of three sensor nodes is as follows:

[0205] As shown in the GNN aggregated data classification flowchart, the sensor node embedding vector is loaded into the 1D-CNN arc detection model for analysis, and the processed low-frequency data is loaded into the LightGBM feature classification model for analysis. Figure 1

[0206] In engineering, the input data is:

[0207] Low frequency: joint temperature (1 Hz), environmental humidity (0.2 Hz)

[0208] High frequency: ultrasonic signal (200 kHz)

[0209] Processing procedure:

[0210] LightGBM branch:

[0211] Features: 10 min temperature rise rate = 0.8℃ / min, humidity change rate > 5% / min

[0212] Output: Moisture intrusion risk (probability 87%)

[0213] 1D-CNN branch:

[0214] # Ultrasonic signal feature extraction

[0215] cnn_feat = model_cnn(ultrasonic_wave) # Output: [0.02, 0.91, 0.03] → Partial discharge

[0216] Wherein, cnn_feat is CNN Features, indicating the feature vector extracted by CNN, in the present application, it means the feature tensor output by the 1D-CNN model, representing the fault mode in the ultrasonic signal, model_cnn is CNNModel, in the present application, it is a 1D convolutional neural network model (used for processing ultrasonic waveforms) that has been trained, and the ultrasonic_wave in the parentheses is Ultrasonic Wave Signal, indicating the ultrasonic signal, in the present application, it means the original ultrasonic waveform data (time domain or frequency domain representation) collected from the sensor;

[0217] Output: Partial discharge (probability 91%)

[0218] Decision fusion: Pfault =1−(1− PGBM )×(1− PCNN ​)= 1 - 0.13 x 0.09 = 98.8%;

[0219] Pfault P (at least one fault exists) = 1 - P (no fault exists), PCNN P (partial discharge) = 1D-CNN model output probability of partial discharge; PGBM P (moisture intrusion) = LightGBM model output probability of moisture intrusion;

[0220] Conclusion: Joint exists moisture + discharge composite fault.

[0221] The performance comparison experiment results are as follows:

[0222] Model Accuracy Insulation aging F1 Arc detection F1 Inference delay Single LightGBM 89.2% 0.93 0.62 2 ms Single 1D-CNN 91.7% 0.71 0.95 8 ms Cascade hybrid 95.3% 0.94 0.96 5 ms Feature fusion hybrid 96.8% 0.96 0.97 10 ms

[0223] 2. Special processing of high-frequency data;

[0224] After the 1D-CNN model identifies the partial discharge feature:

[0225] Trigger real-time alarm;

[0226] Upload discharge feature data to cloud LSTM training queue;

[0227] Partial discharge data is used to train global trend model to improve the prediction ability of the system for slow discharge development;

[0228] II. Hazard classification model collaborative work

[0229] 1. Multi-model division mechanism

[0230] 1D-CNN model: Focus on processing high-frequency arc features, identify microsecond-level transient faults;

[0231] LightGBM model: Analyze low-frequency features (temperature rise trend, leakage current harmonics, etc.), judge insulation aging and other gradual problems;

[0232] LSTM-Attention model:

[0233] Periodic analysis of 24-hour data window;

[0234] Capture slow-growing hazards (such as insulation deterioration caused by moisture intrusion);

[0235] Receive partial discharge data identified by CNN, continuously optimize discharge trend prediction;

[0236] 2. Model result fusion

[0237] Each model outputs a hazard type label with confidence;

[0238] Dynamic weighted fusion strategy:

[0239] Arcing faults give higher weights to CNN results;

[0240] Gradual faults focus on LSTM analysis results;

[0241] Final output contains: hazard type, severity level, development trend prediction.

[0242] III. Graph neural network positioning implementation

[0243] 1. Cable network modeling:

[0244] Abstract sensors as nodes, features include: temperature, current, humidity, partial discharge;

[0245] Abstract cable segments as edges, features include: length, cross-sectional area, laying depth, topological distance;

[0246] Divide levels according to physical topology:

[0247] Main line is high level (Level 0);

[0248] Branch lines are downgraded in turn (Level 1-N);

[0249] Each node records its level and parent node;

[0250] 2. Hierarchical message passing:

[0251] Message generation: nodes fuse their own features and connected cable features to form message packets;

[0252] Intra-level transmission: nodes in the same level exchange current state information;

[0253] Cross-level transmission:

[0254] Sub-nodes upload aggregated information to parent nodes;

[0255] Parent nodes broadcast global state downward;

[0256] Fault probability calculation:

[0257] Update node state vector after each round of transmission;

[0258] Final output of each node fault probability value;

[0259] 3. Positioning decision process:

[0260] Level fault probability = maximum value of all node probabilities in that level;

[0261] After determining the highest probability level, in that level:

[0262] Filter fault nodes exceeding the threshold;

[0263] Ranking candidate locations by probability values;

[0264] Output contains: fault level, node ID list, location coordinates, probability distribution.

[0265] IV. System Synergetic Workflow

[0266] 1. Real-time monitoring loop:

[0267] Edge nodes continuously process sensor data streams;

[0268] High-frequency arc data triggers real-time alarms (response time <10ms);

[0269] Hazard classification results update graph node features in real time;

[0270] 2. Periodic analysis loop;

[0271] LSTM global trend assessment is started every 24 hours;

[0272] Fusion of recent partial discharge data updates prediction model;

[0273] Output cable network health heat map;

[0274] 3. Positioning trigger mechanism:

[0275] When the hazard classification confidence exceeds the threshold:

[0276] Freeze the current graph state;

[0277] Perform hierarchical message passing (3-5 rounds);

[0278] Generate a positioning report and trigger a work order;

[0279] Major faults start multi-level joint positioning.

[0280] Case effect:

[0281] A commercial complex cable intelligent monitoring system:

[0282] Cable laying method: cable trench, bridge, vertical shaft, hidden

[0283] Main hazards: insulation aging, joint overheating, moisture intrusion

[0284] Typical scenario: power distribution room outgoing line

[0285] Sensor deployment scheme:

[0286] Laying method Sensor type Installation density Monitoring focus Cable trench Distributed optical fiber temperature measurement Continuous monitoring Local overheating point Cable bridge Wireless temperature measurement tag Every 5 meters Poor heat dissipation area Buried pipeline Fault current indicator Branch node Insulation deterioration

[0287] The disposal process of the system after systematic deployment of the hazards:

[0288] LSTM warning: humidity > 85% RH for 7 consecutive days

[0289] Network positioning: distribution room → merchant cable trench section

[0290] On-site processing: found that the drain pipe was broken → insulation was restored after repair

[0291] Implementation comparison effect:

[0292] Indicator Commercial property system Traditional method Hidden danger identification accuracy 97.8% <85% Positioning accuracy ±0.5m >5m Response time 15s > 30 min Electrical fire reduction rate 76% - Maintenance cost reduction 57% -

[0293] After the application of the present application to the commercial complex, the effect is obvious, the cable life is extended by 40%, and for commercial properties, management innovation is achieved by using predictive maintenance instead of planned maintenance.

[0294] The modules described as separate components can or can not be physically separate, and the components shown as modules can or can not be physical modules, i.e. can be located in one place, or can be distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0295] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods or systems. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

[0296] It should also be noted that the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, product or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, product or equipment. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of another identical element in the process, method, product or equipment including the element.

[0297] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

[0298] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

[0299] It is to be understood that the application is not limited to the precise structures hereinabove described and shown in the drawings, for purposes of illustration and education only, and that variations and changes can be made by persons skilled in the art in the implementation of the application without departing from the scope thereof. The scope of the application is limited only by the claims that follow.

Claims

1. A systematic monitoring method of the electrical safety performance of a cable, characterized in that, The method comprises the following steps: Collecting sensor data to the edge computing node to preprocess data output as real-time calculation temperature data, partial discharge data, resistance capacitance data and vibration data; Sending data to hidden danger classification model and fault location model in the cloud, and outputting warning information to operation and maintenance engineering processing; Wherein, the sensor data sending process is judged according to the data frequency, the sensor data greater than 1kHz is loaded into 1D-CNN arc detection model, the sensor data less than 1kHz and the data cycle not greater than 24h is loaded into LightGBM feature classification model, and the sensor data less than 1kHz and the data cycle greater than 24h is periodically loaded into LSTM-Attention gradual analysis model; The LSTM-Attention gradual analysis model accesses the preprocessed data and encodes the time sequence dependence, focuses on the key period after adding the attention mechanism, and cooperates with the fault location model to output the fault hidden danger classification and the corresponding location; Further comprising the following steps: After the LSTM-Attention gradual analysis model accesses the pretraining fault data generated by the finite element simulation, real sensor data and fault data generated by the finite element simulation are mixed in proportion as a training set for training, the training result of the LSTM-Attention gradual analysis model is corrected, and the proportion of the fault data generated by the finite element simulation is reduced after training and correction, until the proportion of the fault data generated by the finite element simulation is reduced to 0, and the training meets the iterative convergence condition, then the training of the LSTM-Attention gradual analysis model is completed; The process of hidden danger resolution of 1D-CNN arc detection model is as follows: 1D-CNN arc detection model accesses the preprocessed data types including partial discharge data, vibration signal and temperature mutation; The processed data is substituted into 1D-CNN arc detection model for multi-scale convolution feature extraction, and the data types involving multi-sensor data are executed feature fusion strategy, specifically including executing fusion of PD pulse and vibration signal features, cooperating with fault location model to output arc, vibration position and duration of cable joint; The LightGBM feature classification model is used to load sensor data less than 1kHz to monitor the aging / wet abnormal state of cable, and the corresponding data is labeled with aging / wet abnormal state label, specifically including: Firstly, the timestamps of different sensor data are aligned and normalized to the same interval to eliminate dimension, unsupervised method is used for preliminary labeling, original sensor data types and corresponding basic statistical features are constructed, basic statistical features include mean, variance, maximum value and minimum value; High-level statistical features include: time window features; Correlation features; Trend features; Periodic features; Physical model features, set the importance of features, then train the model according to normal state, aging state and wet state three state labels, and label the data; The 1D-CNN arc detection model discriminates the data related to the partial discharge and sends the data to the LSTM-Attention gradual analysis model for training, and the trained LSTM-Attention gradual analysis model is used to evaluate the partial discharge trend of the global cable.

2. The method of claim 1, wherein the method further comprises: The fault positioning model comprises: a cable network graph structure model, sensors are set as sensor nodes, each node is set with node features of temperature, current, humidity, and partial discharge amount, cable sections are set between the sensor nodes, each cable section and sensor node are numbered, and the cable section features are set as cable length, cross-sectional area, laying depth, and topological distance; a Gaussian kernel weighting-based adjacent matrix of the cable section and the sensor node is constructed for message passing; the message passing mechanism is to splice and link the cable section features and the features of adjacent sensor nodes to form a message function, the message function is updated in real time based on the maximum and total values of the node position-based failure rate, the message function is transmitted between different levels of sensor nodes, the same sensor node and cable section under different branches are numbered and totaled as a level, the branch relationship of multiple levels is distinguished from large to small to form different connection layers, and the message function is transmitted and updated between the global full connection layers; each connection layer is provided with a failure probability, and the failure probability is determined by the output data of the hidden danger classification model.

3. The systematic monitoring method of the electrical safety performance of a cable according to claim 2, characterized in that, The failure probability is the maximum value of the failure probabilities of multiple sensor nodes under the same connection layer, and the number position data of the sensor nodes in the level under the connection layer are positioned based on the failure probability.

4. The systematic monitoring method of the electrical safety performance of a cable according to claim 1, characterized in that, In the method, The sensor data used to train the hidden danger classification model further comprises fault data generated by implementing finite element simulation based on COMSOL Multiphysics software, and the generation steps are as follows: Step 1: geometrically constructing a cable layered structure: conductor, insulating layer, shielding layer, sheath, and joint component; Step 2: setting the materials and corresponding properties of the cable layered structure; Step 3: constructing the coupling equations of the electric field, thermal field, and structural mechanics field of the cable layered structure; Step 4: importing the fault scenarios to obtain simulation implementation of insulation aging, joint loosening, overload operation, and moisture intrusion; Step 5: setting boundary conditions for the electric field, thermal field, and structural mechanics field respectively, applying rated current / voltage at the conductor end of the electric field, convective heat dissipation on the outer surface of the cable in the thermal field, forced air cooling at the joint, and setting fixed cable end displacement constraints in the structural mechanics field; Step 6: generating supplementary fault data for training the hidden danger classification model.

5. The method of claim 1, wherein the method further comprises: The method further comprises embedding the sensor nodes into vectors by using a GNN aggregation model, loading the processed high-frequency data into the 1D-CNN arc detection model for analysis, and loading the processed low-frequency data into the LightGBM feature classification model for analysis.

Images