An on-line monitoring method for cable insulation layer

Abstract

The present invention relates to the technical field of power cables, and specifically includes an on-line monitoring method for cable insulation layers, which includes: the cable monitoring system performs harmonic monitoring through multiple monitoring circuits, constructs an edge topology network to analyze the circuit load, collects harmonic component characteristics, the cloud center node trains a load harmonic processing model and sends it to the edge node to process harmonic signals, evaluates cable insulation defects and outputs indicators, solves the harmonic interference caused by the cable circuit load and propagated through the cable, has an adverse effect on the insulation layer, and limits the accuracy of locating cable insulation defect points. The technical problem is solved by setting an external monitoring circuit at the cable interface, adopting a sub-node strategy to conduct a detailed analysis of the circuit load for each edge node, effectively eliminating harmonics caused by non-insulation layer defects, performing multi-scale feature extraction for harmonic changes related to insulation defects, finely capturing the influence of load changes on harmonic signals, and enhancing the technical effect of the recognition ability of insulation layer defects.

Description

Technical Field

[0001] The present invention relates to the technical field of power cables, and specifically relates to an on-line monitoring method for cable insulation layers. Background Art

[0002] With the continuous expansion and complexity of the scale of power systems, as an important part of power transmission, the safe and stable operation of cables is crucial. The aging and damage of cable insulation layers are one of the main reasons for power system failures. If not discovered and processed in time, power outages may occur.

[0003] Conventional cable insulation detection methods, such as judging the operating state of the cable insulation layer by the changes in the current and voltage flowing through the cable. However, in the scenario of long-distance cables, due to the complexity of cable lines and the uncertainty of interference factors, the changes in the current and voltage flowing through the cable are interfered, and the detection accuracy is severely limited.

[0004] In summary, in the prior art, there are technical problems that harmonic interference caused by cable circuit loads propagates through the cable, has an adverse impact on the insulation layer, and limits the accuracy of locating cable insulation defect points. Summary of the Invention

[0005] This application provides an on-line monitoring method for cable insulation layers, aiming to solve the technical problems in the prior art that harmonic interference caused by cable circuit loads propagates through the cable, has an adverse impact on the insulation layer, and limits the accuracy of locating cable insulation defect points.

[0006] In view of the above problems, the technical solution of this application is as follows:

[0007] This application provides an on-line monitoring method for cable insulation layers, wherein the method includes: obtaining a plurality of monitoring circuits connected to a cable monitoring system, each monitoring circuit is used to monitor harmonics of the cable and output a plurality of harmonic signals;

[0008] Regarding the plurality of monitoring circuits as edge nodes and the cable monitoring system as a cloud center node to generate an edge topology network;

[0009] Using a plurality of edge nodes in the edge topology network to perform load analysis on the plurality of monitoring circuits to obtain a plurality of circuit load information, including the number of circuit elements and the circuit load sizes corresponding to the plurality of circuit elements;

[0010] Collecting harmonic component characteristics according to the plurality of circuit load information and outputting load-harmonic signal characteristics;

[0011] The cloud center node in the edge topology network performs federated training on the load-harmonic signal characteristics and outputs a load harmonic processing model;

[0012] Download the load harmonic processing model to multiple edge nodes, and use the load harmonic processing model to perform harmonic processing on the multiple harmonic signals to obtain processed harmonic signals;

[0013] Perform cable insulation defect assessment on the processed harmonic signals, and output cable insulation defect assessment indicators.

[0014] In summary, one or more technical solutions provided in this application solve the technical problems that harmonic interference caused by the load of the cable circuit is transmitted through the cable, which has an adverse effect on the insulating layer and limits the accuracy of locating the cable insulation defect point. It realizes the comprehensive synchronous monitoring of a widely distributed cable system by setting an external monitoring circuit at the cable interface, adopting a sub-node strategy, constructing an edge topology network, carefully analyzing the circuit load through each edge node, effectively eliminating harmonics caused by non-insulating layer defects, extracting multi-scale features for harmonic changes related to insulation defects, finely capturing the influence of load changes on harmonic signals, and enhancing the technical effect of the ability to identify insulating layer defects. Description of the Drawings

[0015] Figure 1 It is a schematic flowchart of a method for on-line monitoring of a cable insulating layer provided by this application;

[0016] Figure 2 It is a schematic flowchart of a process for obtaining a load harmonic processing model in a method for on-line monitoring of a cable insulating layer provided by this application. Detailed Embodiments

[0017] The following will describe this application in detail with reference to the drawings. As Figure 1 shown, this application provides a method for on-line monitoring of a cable insulating layer, where the method includes:

[0018] S1: Obtain a plurality of monitoring circuits connected to the cable monitoring system, and each monitoring circuit is used to perform harmonic monitoring on the cable and output a plurality of harmonic signals;

[0019] S2: Use the plurality of monitoring circuits as edge nodes, and use the cable monitoring system as a cloud center node to generate an edge topology network;

[0020] Judging the operating state of the cable insulation layer by the change of current and voltage flowing through the cable. Specifically, the megohmmeter method, that is, judging the quality of the insulation layer by measuring the insulation resistance value of the cable, is usually carried out under the power-off state and is applicable to detecting the overall insulation condition of the cable; the high-voltage flashover method, that is, applying a high voltage higher than the working voltage to the cable to induce and locate the weaknesses or existing fault points of the insulation layer, is applicable to detecting the withstand voltage strength of the insulation layer; the bridge method, that is, detecting the defects of the insulation layer by comparing the resistance value or capacitance value of the cable, is applicable to finding local damages inside the cable.

[0021] In the actual use process, it is found that the harmonic interference caused by the cable circuit load will indeed limit the accuracy of locating the cable insulation defect points. Specifically, when the cable circuit load is heavy, high-order harmonics may be generated. High-order harmonics not only affect the normal operation of the power system, but also propagate through the cable and have an adverse effect on the insulation layer. The existence of high-order harmonics will cause the local temperature of the cable insulation layer to rise, accelerate its aging process, and then trigger insulation faults.

[0022] Through long-term monitoring and research, it is found that on the one hand, long-distance cable lines cross different geological and climatic environments, and the changes in environmental factors will cause changes in the physical and chemical properties of the cable insulation layer, which in turn affect the measurement results of current and voltage. For example, temperature changes will cause the expansion and contraction of cable materials, resulting in changes in the resistivity of the insulation layer; humidity changes will cause the insulation layer to be damp, reducing its insulation performance. On the other hand, there are various interference factors in long-distance cable lines, such as electromagnetic interference, noise interference, etc. The interference factors will affect the transmission and measurement of current and voltage signals, resulting in inaccurate detection results. Especially in a complex power grid environment, the electromagnetic interference of adjacent lines may have a serious impact on the insulation detection of long-distance cables.

[0023] Based on this, the present application ensures the comprehensive and synchronous monitoring of a widely distributed cable system by setting an external monitoring circuit at the cable interface and adopting a sub-node strategy. The constructed edge topology network improves the efficiency of data processing, and through the detailed analysis of the circuit load by the edge nodes, the harmonics caused by non-insulation layer defects are effectively eliminated, improving the accuracy and real-time performance of monitoring data. Specifically, for different sections of the cable, multiple monitoring circuits are designed and installed. The multiple monitoring circuits are designed to be able to access the cable interface to ensure the real-time monitoring of the electrical performance of the cable, especially the capture of harmonics. Among them, each monitoring circuit includes necessary sensors, signal conditioning modules, and preliminary data processing units to directly collect harmonic signals on the cable.

[0024] Build an edge computing architecture. Among them, each installed monitoring circuit is defined as an edge node. The edge node has a certain data processing ability and can initially analyze the collected harmonic signals, reduce the amount of data transmitted to the cloud center, and improve the response speed. Further, set one or more cloud center nodes for centralized processing, which are responsible for coordinating edge nodes, performing advanced analysis, model training and optimization, and the management of the overall monitoring system. Then, establish an edge topology network. Specifically, ensure that each edge node can maintain stable data communication with the cloud center node through wired or wireless means, involving the use of industrial routers, 4G / 5G cellular networks or private network communication technologies. Further, configure network communication protocols, including data transmission protocols, security protocols, etc., to ensure the safe and efficient exchange of data between edge nodes and the cloud center. The on-line real-time monitoring of the cable insulation layer is realized, providing support for the safe operation of the power system.

[0025] S3: Use multiple edge nodes in the edge topology network to perform load analysis on the multiple monitoring circuits, and obtain multiple circuit load information, including the number of circuit components and the circuit load sizes corresponding to the multiple circuit components.

[0026] S4: Collect harmonic component characteristics according to the multiple circuit load information, and output load-harmonic signal characteristics.

[0027] Use multiple edge nodes in the edge topology network to perform load analysis on the multiple monitoring circuits, and obtain multiple circuit load information. Specifically, ensure that all monitoring circuits are correctly installed and connected to the cable, and can continuously collect current and voltage signals in the cable, especially harmonic components. In addition, establish a reliable data communication protocol between the edge node and the cloud center to ensure that monitoring data can be uploaded to the edge node for processing in real time or regularly. Further, the edge node receives real-time current and voltage data from each monitoring circuit. The real-time current and voltage data received by each monitoring circuit contains the load status information of the circuit. Then, by analyzing the circuit design diagram or historical data, determine the types and quantities of circuit components at each monitoring point, such as transformers, capacitors, resistors, etc.; according to the current and voltage signals, combined with the impedance information of the components, calculate the real-time load size of each circuit component.

[0028] Collect harmonic component features based on the multiple circuit load information and output load-harmonic signal features. Specifically, use digital signal processing techniques (such as Fast Fourier Transform FFT) to separate the harmonic components from the original signal, remove the fundamental wave and other non-harmonic components, and only retain the harmonic signals related to the load. Further, according to the circuit load information, convert the harmonic signals at each monitoring point into feature vectors, where the feature vectors include the amplitude, phase, frequency components, etc. of specific harmonic orders; combine the number of circuit elements and the load size to label the harmonic features and form load-harmonic signal features, that is, each feature vector is associated with a specific circuit load state. Among them, each edge node independently processes the data of the monitoring circuit it is connected to. After completing the load analysis and feature extraction, summarize all the load-harmonic signal features of the corresponding node, and then dynamically and real-time grasp the load state of the cable and the resulting harmonic changes, providing a scientific basis for the evaluation of the health state of the cable insulation layer.

[0029] S5: The cloud center node in the edge topology network performs federated training on the load-harmonic signal features and outputs a load harmonic processing model.

[0030] S6: Download the load harmonic processing model to multiple edge nodes, and use the load harmonic processing model to perform harmonic processing on the multiple harmonic signals to obtain the processed harmonic signals.

[0031] S7: Perform cable insulation defect evaluation on the processed harmonic signals and output cable insulation defect evaluation indicators.

[0032] At the cloud center node, initialize a global load harmonic processing model according to a predefined algorithm framework (such as deep learning, machine learning, etc.). The load harmonic processing model is used to learn the load-harmonic signal features provided by different edge nodes to distinguish and process the harmonic interference caused by circuit loads. Further, download the initialized global model to each edge node, and each edge node starts local training according to the load-harmonic signal features it collects. Among them, each edge node performs model training based on its own circuit load information and corresponding harmonic signal features, including operations such as feature extraction, classification, and anomaly detection on the signals; after a certain period of local training, the edge node updates and uploads the trained candidate model to the cloud center.

[0033] The cloud central node receives model updates from all edge nodes, and uses a federated learning algorithm (such as FedAvg, federated averaging method) to perform weighted averaging or fusion to generate a new global model, taking into account the diversity of edge nodes and data privacy protection; Repeat multiple rounds of federated training until the model converges or reaches a preset stop condition. Then, download the finally optimized load harmonic processing model to each edge node again. The edge node uses this model to process the received harmonic signals, eliminate the harmonics caused by the circuit load, and only retain the harmonic signals reflecting the insulation condition of the cable itself;

[0034] Deeply analyze the processed harmonic signals to extract features related to insulation layer defects, such as the increase in specific harmonic components, frequency changes, etc.; Based on the extracted features, apply a preset evaluation algorithm to quantitatively evaluate the insulation defects of the cable, generating cable insulation defect evaluation indicators, such as insulation aging degree, damage probability, etc. Further, send the cable insulation defect evaluation indicators to the integrated automation management system for corresponding maintenance or preventive measures. Further, adjust the model parameters or training strategies according to the feedback in actual applications to continuously optimize the performance of the monitoring system, ensure long-term effective monitoring of the health status of the cable insulation layer, and realize a closed-loop process from model training to cable insulation defect evaluation, improving the intelligent level of cable maintenance and the accuracy of fault prediction.

[0035] Furthermore, perform federated training on the load-harmonic signal features to output a load harmonic processing model, such as Figure 2 As shown, the method of this application includes:

[0036] Train and initialize a global model in the cloud central node of the edge topology network;

[0037] Download the initialized global model to multiple edge nodes to obtain the multiple circuit load information corresponding to the multiple edge nodes;

[0038] Train in the corresponding edge nodes according to the multiple circuit load information to obtain multiple candidate harmonic processing models after convergence;

[0039] Then upload the multiple candidate harmonic models to the cloud central node, and the cloud central node performs weighted aggregation on the multiple candidate harmonic models to obtain a load harmonic processing model.

[0040] The specific steps to implement the federated training process to output a load harmonic processing model are as follows:

[0041] Perform federated training on the load-harmonic signal characteristics and output a load harmonic processing model. Specifically, design a basic load harmonic processing model architecture on the cloud central node (the load harmonic processing model architecture can be a convolutional neural network (CNN), a recurrent neural network (RNN), or other deep learning structures suitable for time series signal processing). The load harmonic processing model architecture usually contains multiple layers to identify and separate the harmonic signals caused by circuit loads and the harmonic characteristics of the cable itself. Then, assign initial weights and bias values to the model, which is the starting point of federated training. The initialization method can be random or preset values based on certain prior knowledge;

[0042] The cloud central node pushes the initialized global model to each edge node. Among them, each edge node is responsible for monitoring a specific part of the cable network, so the received model copies are the same. Further, each edge node collects the corresponding circuit load information and harmonic signal data according to the monitored circuit it is connected to. Data preprocessing includes steps such as filtering, standardization, and normalization to ensure that the data quality meets the model training requirements;

[0043] Based on the respective collected circuit load information, each edge node trains the model locally. During the training process, the model parameters are fine-tuned according to the unique dataset of the edge node, aiming to make the model better adapt to the harmonic characteristics in the local environment. Further, each edge node determines whether its model has reached the convergence state by monitoring metrics such as the loss function or accuracy. The convergence criterion can be reaching a predetermined number of iterations, the loss value being lower than a threshold, or the performance on the validation set no longer improving. Then, when the model is locally trained and considered to have converged, the edge node uploads the candidate model it has trained to the cloud central node. The candidate model contains the updated model parameters, reflecting the harmonic characteristics of the cable section covered by the corresponding edge node;

[0044] After receiving the candidate models from all edge nodes, the cloud central node performs a weighted aggregation operation, aiming to comprehensively update the models of each edge node. Considering the differences in data volume and quality of different nodes, fuse the model parameters through weighted average or other strategies to form a more general and generalized load harmonic processing model. Further, the aggregated model will be tested on a small part of the validation set to ensure the improvement or at least the stability of the model performance. Then, the verified load harmonic processing model is downloaded to the edge node again for real-time processing of the monitored harmonic signals, eliminating the influence of circuit loads, and evaluating the health status of the cable insulation layer; According to the actual application effect and newly collected data, the cloud central node decides to conduct a new round of federated training to continuously iterate and optimize the model performance. The federated training mechanism enhances the generalization ability of the model, protects data privacy, and realizes efficient and distributed online monitoring of cable insulation layers.

[0045] Furthermore, the method of the present application further includes:

[0046] Construct a target function. When the convergence condition of the target function is satisfied, multiple candidate harmonic processing models after convergence are obtained. The expression of the target function is as follows: ;

[0047] Wherein, is the target function, and iterative updates are performed with as the target to output multiple candidate harmonic models, is the number of edge nodes, is the th weight parameter of the model corresponding to the th edge node, is the th edge node with parameter for the signal loss function generated by the harmonic signal

[0048] Construct a target function. Specifically, the expression of the target function is as follows: , wherein, is the target function corresponding to the entire federated learning system, aiming to minimize the total signal loss of all edge nodes. Iterative updates are performed with as the target to output multiple candidate harmonic models, is the number of edge nodes, is the th weight parameter of the model corresponding to the th edge node, reflecting the importance or contribution degree of the th edge node, is the th edge node with parameter for the signal loss function generated by the harmonic signal

[0049] Local training and loss calculation. Further, on each edge node, local harmonic signal data and initialized model parameters are used for training to optimize , that is, reduce the signal loss; for each edge node, calculate the signal loss under the current model parameters. According to the calculated loss function value, perform iterative updates on the model parameters using local gradient descent or other optimization algorithms to make approach the minimum value. Among them, each edge node uploads the updated model parameters or gradient information to the cloud center node;

[0050] Furthermore, the cloud center node receives the information uploaded by all edge nodes, combines the updates of all nodes using the weighted average method (or other aggregation rules), and updates the global model parameters; calculates the objective function value under the new global model parameters and compares it with the value of the previous round. If the change is less than a preset threshold or the preset number of iterations is reached, the model is considered to have converged; if the convergence condition is not met, continue iterative optimization until the convergence condition is satisfied. Then, once the model converges, the cloud center node distributes the final global model parameters back to each edge node. The final global model parameters represent the candidate harmonic processing model after convergence. The federated learning process based on the objective function effectively integrates the data and computing capabilities of multiple edge nodes for harmonic signal processing under different circuit load conditions.

[0051] Furthermore, to obtain multiple circuit load information, the method of the present application further includes:

[0052] Judge the multiple edge nodes according to the multiple circuit load information to obtain component similarity and load similarity;

[0053] Perform weighting according to the component similarity and load similarity to obtain a similarity weighting result;

[0054] Cluster the multiple edge nodes according to the similarity weighting result and output multiple types of edge nodes;

[0055] Perform federated learning on the multiple types of edge nodes respectively to obtain multiple types of load harmonic processing models.

[0056] To obtain multiple circuit load information, specifically, collect the real-time load information of the cable segments connected to each edge node through circuit monitoring, including the number of circuit components, the load size of each component, etc.; analyze the circuit component types and configurations monitored by each edge node to evaluate the similarity degree between each edge node. For example, compare based on attributes such as the model, specification, and function of the components; according to the collected load size data, calculate the load level of each edge node and evaluate the similarity of the load pattern, involving comparing indicators such as the fluctuation range, peak value, and mean value of the load. Further, assign appropriate weight parameters to the component similarity and load similarity, depending on which aspect is more important in the actual application. For example, if the component configuration has a greater impact on harmonics, its weight can be set higher;

[0057] Combined with the scoring of component similarity and load similarity, the comprehensive similarity score of each edge node is calculated through a weighted formula. Further, an appropriate clustering algorithm is selected, such as K-means, hierarchical clustering, or DBSCAN, etc., and the edge nodes are grouped according to the comprehensive similarity score; the selected clustering algorithm is run, and the edge nodes are assigned to different categories according to the similarity weighting result to form multiple categories of edge nodes. Then, an initial harmonic processing model framework is prepared for each category of edge nodes, or a global initial model is used as a starting point;

[0058] On the edge nodes of each category, local training of the model is carried out using the data of this category, reflecting the decentralized characteristics of federated learning, that is, the model is optimized on local data; the edge nodes of each category upload the trained model parameters or gradients to the cloud central node, and the cloud central node performs an aggregation operation on the model parameters, such as weighted average, to generate a global model for this category; multiple rounds of local training and cloud aggregation are carried out to further improve the model performance until the model converges or reaches a predetermined stopping condition;

[0059] After completing the federated learning, the obtained multiple categories of load harmonic processing models are deployed back to the corresponding edge nodes. The edge nodes of each category apply their exclusive models to process harmonic signals, so as to more accurately evaluate the cable insulation defects, realize the effective utilization of circuit load information in the online monitoring of the cable insulation layer, improve the pertinence and accuracy of the model, and further enhance the flexibility and adaptability of the system.

[0060] Furthermore, the method of this application also includes:

[0061] After obtaining the multiple categories of load harmonic processing models, use the multiple categories of load harmonic processing models to perform harmonic processing on multiple harmonic signals output by the multiple categories of edge nodes respectively;

[0062] Among them, each category of edge nodes in the multiple categories of edge nodes stores a corresponding load harmonic processing model, and each category of edge nodes is connected to the monitoring circuit terminal of its own category.

[0063] Ensure that the corresponding load harmonic processing models have been loaded in the local storage or memory of each category of edge nodes, which are optimized for the characteristics of their respective categories during the previous federated learning process; before formal application, perform functional verification on the models on each category of edge nodes to confirm that the models are loaded correctly and can respond and process data normally. Further, the monitoring circuits on each edge node continuously monitor the cable for harmonics and capture the harmonic signals in the cable in real time; the monitoring circuit terminal transmits the captured original harmonic signals to the edge nodes of its own category for processing.

[0064] When the edge node receives a harmonic signal, it immediately processes the signal using the deployed load harmonic processing model, including operations such as filtering, noise reduction, and feature extraction using the model to eliminate unnecessary harmonic components introduced by the circuit load. Further, according to the specific design of the model, it is necessary to dynamically adjust the processing parameters to adapt to the actual working state and environmental changes of the current cable. The processed harmonic signal is further analyzed to evaluate potential defects in the cable insulation layer. This typically involves analyzing the change trend of the harmonic signal, anomalies in the frequency components, etc.;

[0065] In the actual application process, it also includes that if signs of insulation defects are found, the edge node should promptly generate an alarm signal and report it to the monitoring center through the network for preventive measures or maintenance arrangements; regularly or as needed, monitor the processing effect of the model, evaluate the accuracy and stability of the model, and collect feedback data on the model's performance; according to the monitoring results and new monitoring data, fine-tune or update the model in a timely manner to maintain the best performance of the model. The updated model is redeployed to the edge node to ensure the efficiency and accuracy of the online monitoring of the cable insulation layer. At the same time, using the multi-class load harmonic processing models obtained by federated learning, customized processing strategies are provided for different types of circuit load environments to enhance the reliability of the system.

[0066] Furthermore, for harmonic component feature acquisition based on the multiple circuit load information and outputting load-harmonic signal features, the method of the present application includes:

[0067] Setting circuit load samples according to the multiple circuit load information;

[0068] Obtaining harmonic component samples based on the circuit load samples;

[0069] Inputting the harmonic component samples into a feature convolution network for multi-scale feature convolution and outputting multi-scale harmonic signal features;

[0070] Analyzing the change characteristics of the multi-scale harmonic signal features under the circuit load samples and outputting load-harmonic signal features, where the load-harmonic signal features are the change characteristics of the harmonic component signals affected by load changes.

[0071] For harmonic component feature acquisition based on the multiple circuit load information and outputting load-harmonic signal features, specifically, collect circuit operation data at different time periods or working conditions from the monitoring system, including key parameters such as current, voltage, and power; according to the collected data, define different circuit load states, such as light load, full load, transient load, etc., and create a circuit load sample set based on these states. Each sample contains the electrical parameters of the circuit under specific load conditions;

[0072] For each circuit load sample, apply the Fast Fourier Transform (FFT) or other spectral analysis methods to convert from the time-domain signal to the frequency domain and isolate the harmonic components. Then, associate each harmonic component with its corresponding circuit load state to form a labeled set of harmonic component samples for subsequent feature learning and analysis. Construct a feature convolutional network that includes multiple convolutional layers, with each convolutional layer using convolutional kernels of different scales to capture the features of the harmonic signal at different time scales or frequency bandwidths. Input the harmonic component samples into the feature convolutional network, where each sample undergoes multi-layer convolutional processing in the network, and each layer focuses on different feature scales of the signal.

[0073] The output of the network is the multi-scale harmonic signal features, which reflect the intensity, distribution, and change trend of the harmonic signal at different frequency bands and time scales. By comparing the changes in harmonic features under different load states, identify specific patterns of the impact of load changes on the harmonic signal, such as the phenomenon that certain harmonic components increase significantly with the increase in load. Further, integrate the analysis results into the load-harmonic signal features, including determining which harmonic components are most sensitive to load changes and the laws of change with the load state. Then, determine the load-harmonic signal features to provide a basis for subsequent cable insulation layer health assessment, realizing the fine analysis of harmonic signals in a complex circuit environment, detailing the harmonic components most affected by the load and their change characteristics, and the features indicating potential insulation problems, providing data support for on-line monitoring of the cable insulation layer.

[0074] Furthermore, input the harmonic component samples into the feature convolutional network for multi-scale feature convolution, and the output is the multi-scale harmonic signal features. The method of this application includes:

[0075] Set multi-scale convolutional kernels and use the multi-scale convolutional kernels to train the feature convolutional network.

[0076] Input the harmonic component samples into the trained feature convolutional network, and the convolutional features of the feature convolutional network include frequency characteristics, amplitude characteristics, and wavelength characteristics.

[0077] Perform multi-scale feature convolution on the harmonic component samples according to the frequency characteristics, amplitude characteristics, and wavelength characteristics, and output the multi-scale harmonic signal features.

[0078] Input the harmonic component samples into the feature convolution network for multi-scale feature convolution, and output multi-scale harmonic signal features. Specifically, collect the circuit operation data at different time periods or working conditions from the monitoring system, including key parameters such as current, voltage, and power. Define different circuit load states according to the collected data, such as light load, full load, transient load, etc., and create a circuit load sample set according to the circuit load state. Each sample contains the electrical parameters of the circuit under specific load conditions.

[0079] For each circuit load sample, apply the Fast Fourier Transform (FFT) or other spectrum analysis methods to convert from the time-domain signal to the frequency domain, and separate the harmonic components. Then, associate each harmonic component with its corresponding circuit load state to form a labeled harmonic component sample set for subsequent feature learning and analysis. Further, construct a feature convolution network. The feature convolution network contains multiple convolution layers, and each convolution layer uses convolution kernels of different scales to capture the features of the harmonic signal at different time scales or frequency bandwidths. Input the harmonic component samples into the feature convolution network. Each sample undergoes multi-layer convolution processing by the network, and each layer focuses on different feature scales of the signal. The output is multi-scale harmonic signal features, which reflect the intensity, distribution, and change trend of the harmonic signal at different frequency bands and time scales. By comparing the changes in harmonic features under different load states, identify specific patterns of the impact of load changes on the harmonic signal, such as the phenomenon that certain harmonic components increase significantly with the increase of the load.

[0080] Integrate the analysis results into load-harmonic signal features, including determining which harmonic components are most sensitive to load changes and the law of how the harmonic components change with the load state. Organize and report the load-harmonic signal features, determine the harmonic components with the greatest impact on the load and their change characteristics, and the features indicate potential insulation problems, providing a basis for the subsequent health assessment of the cable insulation layer, realizing the fine analysis of the harmonic signal in a complex circuit environment, and providing support for the on-line monitoring of the cable insulation layer.

[0081] Furthermore, use the load harmonic processing model to perform harmonic processing on the multiple harmonic signals. The method of this application includes:

[0082] Connect the load harmonic processing model to a low-pass filter;

[0083] Obtain the harmonic processing parameters output by the load harmonic processing model, and the low-pass filter performs harmonic processing on the multiple harmonic signals with the harmonic processing parameters to obtain the processed harmonic signals.

[0084] Harmonically process the multiple harmonic signals by using the load harmonic processing model. Specifically, ensure that the load harmonic processing model has been federally trained and downloaded to the edge node or the centralized processing unit. The load harmonic processing model is used to identify and process the unnecessary harmonics caused by the circuit load. Configure a digital low-pass filter, which will work in cooperation with the load harmonic processing model. The cut-off frequency of the low-pass filter needs to be determined according to the frequency characteristics monitored by the cable insulation layer to ensure that only signals below this frequency are allowed to pass through, thereby effectively removing high-frequency noise and load-related harmonics.

[0085] Extract the required harmonic processing parameters from the load harmonic processing model. The harmonic processing parameters include the cut-off frequency, attenuation coefficient, etc., which are crucial for guiding how the low-pass filter accurately filters out the circuit load harmonics. Further, apply the extracted parameters to the low-pass filter. Through software programming or hardware configuration, ensure that the performance of the filter matches the output parameters of the model to achieve the best harmonic rejection effect. Then, send the multiple harmonic signals output by the monitoring circuit into the integrated system. The multiple harmonic signals include all types of harmonics, including those generated by cable insulation defects and those generated by the circuit load.

[0086] Use the load harmonic processing model to conduct preliminary analysis and preprocessing on the input harmonic signals, which involves signal feature extraction, classification, or pre-filtering, to provide more accurate guidance for the subsequent filtering steps. Further, based on the parameters provided by the model, the low-pass filter processes the signals to remove the harmonic components related to the circuit load and only retains the harmonic signals related to the cable insulation defects. The processed harmonic signals are output. At this time, most of the load influence in the signals has been eliminated, which is conducive to more accurately evaluating the cable insulation condition in the subsequent analysis steps. Then, use the processed harmonic signals to evaluate the cable insulation defects, which involves analyzing the amplitude and frequency changes of the harmonics or comparing them with known defect patterns. According to the evaluation results and the actual monitoring effects, continuously feedback and adjust the model parameters and filter configuration to improve the monitoring accuracy and efficiency. Implement the complete process from model parameter application to signal processing and then to the final insulation defect evaluation to ensure the effectiveness and accuracy of the on-line monitoring of the cable insulation layer.

[0087] In summary, the beneficial effects of the embodiments of this application are as follows:

[0088] Adopt the sub-node monitoring technology, combined with the edge-cloud computing architecture, use the edge computing node to conduct load analysis and feature extraction on the harmonic signals, combined with the federated learning technology, effectively eliminate the interference caused by the load and environmental factors, focus on analyzing the harmonic signals directly related to the health status of the insulation layer, realize the real-time and synchronous monitoring of long-distance cables, and improve the monitoring accuracy and efficiency.

[0089] Through a multi-scale feature convolutional network, the frequency, amplitude, and wavelength characteristics of harmonic signals are deeply analyzed, and edge nodes are clustered and customized model training is carried out, enabling the monitoring system to adapt to different types of loads and circuit configurations, and improving the sensitivity and accuracy of insulation layer defect identification.

[0090] 3. By setting circuit load samples according to multiple circuit load information; obtaining harmonic component samples based on the circuit load samples; inputting the harmonic component samples into a feature convolutional network for multi-scale feature convolution to output multi-scale harmonic signal features; and analyzing the change characteristics of the multi-scale harmonic signal features under the circuit load samples to output load-harmonic signal features, where the load-harmonic signal features are the change characteristics of the harmonic component signals affected by load changes. The fine analysis of harmonic signals in a complex circuit environment is realized, the harmonic components most affected by the load and their change characteristics are described in detail, and the characteristics indicate potential insulation problems, providing data support for on-line monitoring of cable insulation layers.

[0091] In summary, any step can be stored as computer instructions or programs in an unrestricted computer memory and can be called and recognized by an unrestricted computer processor, without further limitation here.

[0092] Furthermore, the above technical solutions only reflect the preferred technical solutions of the technical solutions of the embodiments of the present application. Some changes that those skilled in the art may make to some parts thereof all reflect the principles of the new type of the embodiments of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application.

Claims

1. A method for online monitoring of cable insulation layer, characterized in that: The method comprises: Acquire multiple monitoring circuits connected to the cable monitoring system, each monitoring circuit is used to perform harmonic monitoring on the cable and output multiple harmonic signals; Using the multiple monitoring circuits as edge nodes and the cable monitoring system as a cloud center node to generate an edge topology network; Utilize multiple edge nodes in the edge topology network to perform load analysis on the multiple monitoring circuits to obtain multiple circuit load information. Specifically, collect real-time load information of the cable segments connected to each edge node through the monitoring circuit, including the number of circuit components and the load size of each component; Harmonic component characteristics are collected according to the multiple circuit load information, and load-harmonic signal characteristics are output. The fundamental wave and other non-harmonic components are eliminated using digital signal processing technology, and only harmonic signals related to the load are retained. According to the circuit load information, the harmonic signal of each monitoring point is converted into a feature vector, and the feature vector includes the amplitude, phase, and frequency components of a specific harmonic order; the feature vector is marked in combination with the number of circuit elements and the load size to form a load-harmonic signal feature; The cloud center node in the edge topology network performs federated training on the load-harmonic signal features and outputs a load harmonic processing model; Downloading the load harmonic processing model to multiple edge nodes, and performing harmonic processing on the multiple harmonic signals using the load harmonic processing model to obtain processed harmonic signals; Evaluate the cable insulation defects on the processed harmonic signals and output the cable insulation defect evaluation index; The load-harmonic signal feature is federated trained to output a load harmonic processing model, and the method includes: Training and initializing a global model in a cloud center node of the edge topology network; Downloading the initialized global model to multiple edge nodes, and obtaining the multiple circuit load information corresponding to the multiple edge nodes; Performing training in corresponding edge nodes according to the plurality of circuit load information to obtain a plurality of converged candidate harmonic processing models; The multiple candidate harmonic processing models are then uploaded to the cloud center node, and the cloud center node performs weighted aggregation on the multiple candidate harmonic processing models to obtain a load harmonic processing model.

2. The method according to claim 1, characterized in that An objective function is constructed. When the convergence condition of the objective function is met, multiple candidate harmonic processing models after convergence are obtained. The expression of the objective function is as follows: ； in, As the objective function, Iteratively update and output multiple candidate harmonic processing models for the target, is the number of edge nodes, The kth edge node corresponds to the weight parameter of the model, is the kth edge node pair harmonic signal with parameter w The resulting signal loss function.

3. The method according to claim 1, characterized in that Acquiring multiple circuit load information, the method further includes: Judging the plurality of edge nodes according to the plurality of circuit load information, and obtaining component similarity and load similarity; Perform weighting according to the component similarity and the load similarity to obtain a similarity weighted result; Clustering the plurality of edge nodes according to the similarity weighted result, and outputting a plurality of types of edge nodes; Federated learning is performed on the multiple types of edge nodes respectively to obtain multiple types of load harmonic processing models.

4. The method according to claim 3, characterized in that After obtaining multiple types of load harmonic processing models, the multiple types of load harmonic processing models are used to perform harmonic processing on multiple harmonic signals output by the multiple types of edge nodes respectively; Each type of edge node among the multiple types of edge nodes stores a corresponding load harmonic processing model, and each type of edge node is connected to a monitoring circuit terminal of the type to which it belongs.

5. The method according to claim 1, characterized in that Harmonic component characteristics are collected according to the plurality of circuit load information, and load-harmonic signal characteristics are output, the method comprising: According to the plurality of circuit load information, setting a circuit load sample; Acquire harmonic component samples based on the circuit load samples; Inputting the harmonic component samples into a feature convolution network for multi-scale feature convolution, and outputting multi-scale harmonic signal features; The change characteristics of the multi-scale harmonic signal characteristics under the circuit load sample are analyzed, and a load-harmonic signal characteristic is output, wherein the load-harmonic signal characteristic is a change characteristic of the harmonic component signal affected by the load change.

6. The method according to claim 5, characterized in that Inputting the harmonic component samples into a feature convolution network for multi-scale feature convolution, and outputting multi-scale harmonic signal features, the method includes: Setting a multi-scale convolution kernel, and using the multi-scale convolution kernel to train a feature convolution network; Inputting the harmonic component samples into a trained feature convolutional network, wherein the convolutional features of the feature convolutional network include frequency characteristics, amplitude characteristics and wavelength characteristics; The harmonic component samples are subjected to multi-scale feature convolution according to the frequency characteristics, amplitude characteristics and wavelength characteristics, and multi-scale harmonic signal characteristics are output.

7. The method according to claim 1, characterized in that The load harmonic processing model is used to perform harmonic processing on the multiple harmonic signals, and the method includes: Connecting the load harmonic processing model to a low-pass filter; The harmonic processing parameters output by the load harmonic processing model are obtained, and the low-pass filter performs harmonic processing on the multiple harmonic signals based on the harmonic processing parameters to obtain processed harmonic signals.

Images