Online monitoring method and system based on power transmission line

Abstract

The invention provides an online monitoring method and system based on a power transmission line, and relates to the technical field of power transmission line monitoring. According to the method, the heterogeneous sensing terminal, edge calculation, the graph neural network and Bayesian reasoning are combined, multi-source data acquisition, state identification and risk prediction are realized, the fault diagnosis accuracy and the risk early warning capability are improved, and the intelligent level and the safety guarantee capability of power transmission line operation are enhanced; the monitoring data is analyzed in real time through an edge calculation unit to generate a state label, a potential fault mode is recognized by combining graph neural network modeling space-time relevance, a risk factor library is further constructed, and a real-time fault probability graph is generated based on a Bayesian network. And dynamic identification and early warning of risk types such as wire strand breakage, icing overrun and mechanical fatigue can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission line monitoring, and in particular to a method and system based on online monitoring of power transmission lines. Background Art

[0002] Online monitoring technology for transmission lines is a key component of modern smart grids. Current mainstream monitoring methods rely on collecting data such as current, voltage, temperature, sag, and micro-meteorological conditions on conductors and uploading it to a backend monitoring center for centralized analysis. These systems typically deploy sensor nodes, wireless communication modules, a central server, and an alarm system to identify transmission line status and provide preliminary fault warnings.

[0003] With the advancement of AI, the Internet of Things, and edge computing technologies, transmission line monitoring systems are evolving toward "ubiquitous sensing + distributed processing + intelligent early warning + autonomous decision-making." Future monitoring systems must not only possess high-frequency, low-latency data acquisition and processing capabilities, but also enable rapid closed-loop response to abnormal events and the ability to combine environmental variables and historical operating data for trend forecasting. Furthermore, collaborative processing of multi-source, heterogeneous data and self-learning capabilities are key research areas.

[0004] Although a certain scale of monitoring systems have been deployed, they still face many problems: (1) The monitoring data type is single, making it impossible to implement multi-factor coupled judgment; (2) Sensor nodes operate independently, and the system lacks data coordination capabilities; (3) Fault identification relies on static rules and is difficult to adapt to different line environments; (4) Monitoring results cannot be effectively fed back for system optimization, resulting in insufficient fault identification accuracy and warning timeliness. Therefore, a complete online monitoring method that integrates data collaborative collection, edge self-identification, central intelligent reasoning, and system self-evolution is urgently needed. Summary of the Invention

[0005] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a method and system based on online monitoring of transmission lines, which can improve the intelligent monitoring level and safety assurance capability of transmission line operation.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for online monitoring of a transmission line, comprising:

[0008] Heterogeneous sensor terminals are deployed at multiple tower nodes along the transmission line, and monitoring data is collected using the heterogeneous sensor terminals; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is equipped with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module;

[0009] An edge computing unit is configured in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label;

[0010] Based on a graph neural network, a global fusion calculation is performed on the final state labels and the monitoring data of the plurality of nodes to identify potential failure modes and evolution trends, and obtain a fusion result;

[0011] The fusion results are subjected to dynamic risk modeling, a risk factor library is constructed, and a real-time fault probability map is generated based on a Bayesian network; the real-time fault probability map is used to classify and identify at least one of the following risk types: wire breakage, mechanical fatigue caused by dancing, excessive icing, abnormal grounding current, and external force interference.

[0012] Preferably, the image acquisition module triggers an image acquisition task at a set time interval or when the final state label changes; the image information includes conductor sag, tower structure status and ice cover contour feature map.

[0013] Preferably, an edge computing unit is configured in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary status label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final status label, including:

[0014] Using the local analysis sub-model in the edge computing unit, time series modeling and multi-dimensional feature extraction are performed on the monitoring data of the current node to generate a preliminary state label;

[0015] Collect monitoring data and preliminary status labels of a plurality of nodes adjacent to the current node within a set time window;

[0016] Based on the spatial topological relationship and time correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label.

[0017] Preferably, based on the spatial topological relationship and temporal correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label, including:

[0018] Constructing a local graph structure subgraph including the current node and adjacent nodes; the nodes of the local graph structure subgraph represent the locations of transmission towers, and the edges represent the physical connection relationships between the towers;

[0019] Based on the local graph structure subgraph, using the formula Calculate the probability distribution of the final state label of the current node; wherein, is the final state label of node i; is the initial state label of the adjacent node j; is the set of adjacent nodes of node i; α ij is the attention weight of node j to node i, which is calculated based on the spatial topology and state similarity; W t is a trainable weight matrix for modeling temporal correlation; W s is the trainable weight matrix used to model the static monitoring data of the current node; X i is the original monitoring data feature vector of node i; σ(·) is the activation function used to normalize the output label; the calculation formula of the attention weight is: in, is the cosine similarity between state labels, indicating the consistency of node states; d ij is the geographical distance between node i and node j, which represents the geometric relationship of spatial topology; is the spatial similarity weight based on the Gaussian kernel function, σ represents the scale parameter affecting the distance; γ and λ are the preset weight adjustment factors of state similarity and spatial similarity.

[0020] Preferably, a global fusion calculation is performed on the final state labels and monitoring data of the plurality of nodes based on a graph neural network to identify potential failure modes and evolution trends, and obtain a fusion result, including:

[0021] All the nodes and topological connection relationships are constructed into a graph structure and input into the graph neural network; the input feature vector of each node is obtained by splicing the final state label and the monitoring data;

[0022] Based on the multi-layer graph attention network structure, the propagation and update formula is used to iteratively update node features; the propagation and update formula is: Among them, the feature vector of node i at the lth layer, W (l) is the weight matrix of the lth layer, is the attention weight of node j to node i in layer l;

[0023] In the final layer output of the graph neural network, the global pooling operation and the evolution trend identification module are combined to output a fusion result containing potential fault modes and evolution trends.

[0024] Preferably, in the final layer output of the graph neural network, a global pooling operation is combined with a temporal convolutional network to output a fusion result containing potential failure modes and evolution trends, including:

[0025] In a plurality of preset continuous time periods, a graph-level representation vector of each time point is obtained based on the graph neural network, and a historical time-series graph-level representation vector sequence arranged in chronological order is formed;

[0026] The historical time series graph-level representation vector sequence is used as input and fed into a temporal convolutional network;

[0027] Using the temporal convolutional network to learn the historical time series graph-level representation vector sequence to obtain a trend feature vector;

[0028] Performing feature concatenation or weighted fusion on the graph-level representation vector at the current moment and the trend feature vector to construct a fused representation vector;

[0029] The fusion representation vector is input into a trained classification and prediction module, and a potential failure mode label and a corresponding evolution trend score are output to form the fusion result.

[0030] Preferably, the fusion representation vector is input into a trained classification and prediction module, and the potential failure mode label and the corresponding evolution trend score are output to form the fusion result, including:

[0031] Inputting the fused representation vector into the fault classification subnetwork in the classification and prediction module; the fault classification subnetwork includes at least one fully connected layer and one softmax layer, which is used to output the probability distribution of multiple potential fault types;

[0032] Synchronously inputting the fused representation vector into the risk trend prediction subnetwork in the classification and prediction module; the risk trend prediction subnetwork includes at least one fully connected regression layer for outputting a fault evolution risk score;

[0033] The output results of the classification subnetwork and the prediction subnetwork are combined as the fusion result; the fusion result is used to characterize the current state of the transmission line and the future evolution trend.

[0034] Preferably, the fusion results are subjected to risk dynamic modeling, a risk factor library is constructed, and a real-time failure probability map is generated based on a Bayesian network, including:

[0035] Extract physical quantity characteristics and statistical indicators related to fault types such as conductor breakage, mechanical fatigue, excessive icing, abnormal ground current, and external force interference based on preset known fault cases and historical monitoring data;

[0036] A risk factor library is constructed based on the physical quantity characteristics and the statistical indicators; the risk factors include tension change rate, tilt angle change amplitude, temperature and humidity combined fluctuation index, breeze speed anomaly value, current fluctuation frequency and image icing index;

[0037] Performing feature attribution analysis on the fusion results, using correlation analysis and information gain methods to screen out feature dimensions related to the risk factors and use them as key features in the fusion results;

[0038] Mapping the key features to corresponding risk factor dimensions to construct a risk factor value vector at the current moment, which is used to characterize the multidimensional risk situation under the current transmission line operation state;

[0039] Based on the historical monitoring data, known fault labels and engineering knowledge, a structural learning method is used to establish the dependency relationship between the risk factors and generate the topological structure and conditional probability table of the Bayesian network;

[0040] The risk factor value vector at the current moment is input into the Bayesian network model as evidence, and the posterior probability of each fault type is calculated using the inference algorithm, and a real-time fault probability map is output;

[0041] Based on the maximum a posteriori probability results of various risk events in the real-time fault probability map, risk type classification and identification are performed for status warning and risk intervention decision-making of the transmission line.

[0042] A system based on online monitoring of transmission lines, comprising:

[0043] A data acquisition unit is configured to deploy heterogeneous sensor terminals at multiple tower nodes along the transmission line and collect monitoring data using the heterogeneous sensor terminals; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is provided with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module;

[0044] a label determination unit configured to configure an edge computing unit in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label;

[0045] a fusion unit, configured to perform a global fusion calculation on the final state labels and the monitoring data of the plurality of nodes based on a graph neural network, identify potential failure modes and evolution trends, and obtain a fusion result;

[0046] A fault detection unit is used to perform risk dynamic modeling on the fusion results, construct a risk factor library, and generate a real-time fault probability map based on a Bayesian network; the real-time fault probability map is used to classify and identify at least one of the following risk types: wire breakage, mechanical fatigue caused by dancing, excessive icing, abnormal grounding current, and external force interference.

[0047] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] By deploying heterogeneous sensor terminals along the transmission lines and configuring edge computing units, the present invention can realize real-time collection and on-site preliminary analysis of multi-source monitoring data, significantly improving the real-time performance and response speed of data processing; combining with graph neural networks to perform global fusion reasoning on the status of each node, enhancing the ability to identify potential fault modes and evolution trends under complex spatiotemporal relationships; further, by constructing a risk factor library and Bayesian network, dynamic modeling and classification prediction of fault probability are realized, with high fault diagnosis accuracy and risk warning foresight, which overall improves the intelligent monitoring level and safety assurance capability of transmission line operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0051] Figure 2 A flowchart of label determination provided by an embodiment of the present invention;

[0052] Figure 3 A flowchart for determining the fusion result provided in an embodiment of the present invention;

[0053] Figure 4 The following is a schematic diagram of the system structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] The purpose of the present invention is to provide a method and system based on online monitoring of transmission lines, which can improve the intelligent monitoring level and safety assurance capability of transmission line operation.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a method based on online monitoring of transmission lines, comprising:

[0058] Step 100: Deploy heterogeneous sensor terminals at multiple tower nodes along the transmission line and use the heterogeneous sensor terminals to collect monitoring data; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is equipped with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module;

[0059] Step 200: Configure an edge computing unit on the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node based on a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with adjacent monitoring data to obtain a final state label;

[0060] Step 300: Perform global fusion calculation on the final state labels and monitoring data of multiple nodes based on the graph neural network to identify potential failure modes and evolution trends and obtain fusion results;

[0061] Step 400: Perform risk dynamic modeling on the fusion results, build a risk factor library, and generate a real-time fault probability map based on the Bayesian network; the real-time fault probability map is used to classify and identify at least one of the following risk types: wire breakage, mechanical fatigue caused by dancing, excessive icing, abnormal ground current, and external force interference.

[0062] Preferably, the image acquisition module triggers an image acquisition task at a set time interval or when the final state label changes; the image information includes conductor sag, tower structure status and ice cover contour feature map.

[0063] In one embodiment of the present invention, heterogeneous sensor terminals are deployed at multiple tower nodes along a transmission line to provide real-time online monitoring of the transmission line's operating status. Each heterogeneous sensor terminal is an independent monitoring unit, installed at a tower node. Its structure includes, but is not limited to, a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module. The tension sensor is used to obtain real-time changes in conductor tension; the inclination sensor is used to monitor the inclination of the tower structure; the temperature and humidity sensor and micro-anemometer are used to collect environmental meteorological data; the voltage and current sensor is used to sense the conductor's operating electrical parameters; and the image acquisition module is used to capture on-site image information to assist in status identification and diagnosis.

[0064] The image acquisition module is connected to the node status perception logic and is used to periodically start the image acquisition task at a set time interval (such as every 30 minutes), or automatically trigger the image acquisition action when a change in the status label of the node is detected (such as a sudden change in tension, a sudden change in inclination, or a sharp increase in ice thickness) to capture visual information of key events. The collected image information includes but is not limited to: conductor sag images, deformation images of tower structures, and ice contour morphology images. This image information will be input into the subsequent edge computing unit and graph neural network analysis process as monitoring data together with the physical quantity data, thereby enhancing the robustness of state judgment and the accuracy of fault identification at the image semantic level.

[0065] Preferably, if Figure 2 As shown, an edge computing unit is configured in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label, including:

[0066] Using the local analysis sub-model in the edge computing unit, time series modeling and multi-dimensional feature extraction are performed on the monitoring data of the current node to generate a preliminary state label;

[0067] Collect monitoring data and preliminary status labels of a plurality of nodes adjacent to the current node within a set time window;

[0068] Based on the spatial topological relationship and time correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label.

[0069] Optionally, in a preferred embodiment of the present invention, an edge computing unit is integrated into the heterogeneous sensor terminal deployed at the tower node to process the collected monitoring data on site, reduce the transmission pressure of the original data and improve the response efficiency of state recognition. Specifically, a local analysis sub-model is embedded in the edge computing unit. The model is pre-trained and loaded into the local storage, and has the ability to extract multi-dimensional features and perform sequence modeling for the tension, current, voltage, inclination, environmental quantities and image features of the current node. The model performs state trend analysis based on continuous monitoring data within a period of time (for example, the past 5 minutes), combines a convolutional-recurrent network structure (such as CNN-LSTM) to extract time series dynamic features, and outputs preliminary state labels, such as "normal", "abnormal tension" or "icing warning" and other label results.

[0070] Furthermore, to improve the spatial robustness of state judgment, the edge computing unit synchronously collects monitoring data and preliminary state labels of multiple nodes adjacent to the current node within a set time window (e.g., within 60 seconds), and constructs a topological map between local nodes based on the system's preset spatial topological structure information (e.g., tower number, geographic coordinates, and electrical connection relationships). On this basis, the collaborative judgment sub-model in the edge computing unit uses the data of the current node and neighboring nodes within this time period as joint input. Through the attention mechanism and graph convolution calculation (e.g., the GAT module based on the local graph structure), it models the temporal dependency and state similarity between nodes, completes cross-compensation and fusion reasoning of state information, and outputs the final state label of the current node. This significantly enhances the robustness of edge-side state judgment and multi-source fusion capabilities, providing a reliable foundation for subsequent global analysis of graph neural networks.

[0071] Preferably, based on the spatial topological relationship and temporal correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label, including:

[0072] Constructing a local graph structure subgraph including the current node and adjacent nodes; the nodes of the local graph structure subgraph represent the locations of transmission towers, and the edges represent the physical connection relationships between the towers;

[0073] Based on the local graph structure subgraph, using the formula Calculate the probability distribution of the final state label of the current node; wherein, is the final state label of node i; is the initial state label of the adjacent node j; is the set of adjacent nodes of node i; α ij is the attention weight of node j to node i, which is calculated based on the spatial topology and state similarity; W t is a trainable weight matrix for modeling temporal correlation; W s is the trainable weight matrix used to model the static monitoring data of the current node; X i is the original monitoring data feature vector of node i; σ(·) is the activation function used to normalize the output label; the calculation formula of the attention weight is: in, is the cosine similarity between state labels, indicating the consistency of node states; d ij is the geographical distance between node i and node j, which represents the geometric relationship of spatial topology; is the spatial similarity weight based on the Gaussian kernel function, σ represents the scale parameter affecting the distance; γ and λ are the preset weight adjustment factors of state similarity and spatial similarity.

[0074] Optionally, to achieve high-precision, multi-source collaborative judgment of node status, a collaborative judgment sub-model is constructed in the edge computing unit, specifically for generating a more robust, context-aware final state label based on the monitoring information of the current node and its neighboring nodes. Specifically, a local graph structure subgraph is constructed based on the current node and several neighboring nodes. The nodes in this subgraph represent the locations of the tower nodes, and the edges represent the physical connection relationships between the towers. Graph neural reasoning is performed on this subgraph, modeling the spatial structure and state interaction between neighboring nodes as an information propagation process. The consistency between state labels is measured using cosine similarity.

[0075] Through the above calculation process, the model can integrate the current node timing characteristics, neighborhood state distribution and spatial geometric relationships at the edge computing layer, thereby accurately generating the final state label. It has good context perception and real-time performance, which is significantly better than the traditional static judgment mechanism, reflecting the technological breakthrough and creativity of the present invention in the state reasoning method.

[0076] Exemplarily, the monitoring data of the current node and its neighboring nodes are collected in real time by the edge computing unit, the spatial distance parameter is calculated based on the geographical location information of the tower, the similarity between the state labels is determined by the cosine similarity calculated from the feature vectors between the preliminary state labels of each node, the spatial scale parameter is obtained by fitting the spatial influence range of historical monitoring data, and the weighted coefficient of state and spatial similarity is automatically learned through a model optimization process based on the training set. The weight parameters involved in the modeling process are trained on the terminal side or the central server through the graph neural network, and the corresponding model version is deployed on the edge node to achieve efficient reasoning. The above parameters work together to ensure that this embodiment can consider both real-time and consistency of spatial structure and state evolution during local reasoning.

[0077] Preferably, if Figure 3 As shown, based on the graph neural network, the final state labels of multiple nodes and the monitoring data are globally fused and calculated to identify potential failure modes and evolution trends, and obtain fusion results, including:

[0078] All the nodes and topological connection relationships are constructed into a graph structure and input into the graph neural network; the input feature vector of each node is obtained by splicing the final state label and the monitoring data;

[0079] Based on the multi-layer graph attention network structure, the propagation and update formula is used to iteratively update node features; the propagation and update formula is: Among them, the feature vector of node i at the lth layer, W (l) is the weight matrix of the lth layer, is the attention weight of node j to node i in layer l;

[0080] In the final layer output of the graph neural network, the global pooling operation and the evolution trend identification module are combined to output a fusion result containing potential fault modes and evolution trends.

[0081] Specifically, in order to realize the joint analysis of the multi-node status information of the transmission line, all the tower nodes and their physical connection relationships are constructed into a graph structure and input into the graph neural network model for processing. Specifically, the features contained in each node are composed of the final state label of the node and its monitoring data, and the connection relationship between the nodes is established based on the spatial topology between the towers. Through the processing of multi-layer graph neural networks, this embodiment can extract the high-order relationship between node features layer by layer, and dynamically adjust the influence weights between each node, so as to realize the contextual understanding and comprehensive judgment of the node status. In this process, the network will automatically learn which neighboring nodes have more representative information, so as to more reasonably integrate multi-source data and improve the comprehensive perception of the line operation status.

[0082] Furthermore, in order to extract the overall state and evolution trend of the transmission line, all node feature information output by the last layer of the graph neural network is aggregated to form a unified global representation vector. This global representation is then input into the trend identification module, which uses a time modeling mechanism to learn the evolution path of historical time series information and output the development trend of potential faults in the current stage. Finally, this embodiment outputs a fusion result containing potential fault types and evolution risk scores based on the fusion judgment of node states and trends. This method not only realizes the collaborative modeling of spatial structure and temporal evolution, but also improves the accuracy and foresight of fault identification, and has significant engineering practical value and model innovation.

[0083] Preferably, in the final layer output of the graph neural network, a global pooling operation is combined with a temporal convolutional network to output a fusion result containing potential failure modes and evolution trends, including:

[0084] In a plurality of preset continuous time periods, a graph-level representation vector of each time point is obtained based on the graph neural network, and a historical time-series graph-level representation vector sequence arranged in chronological order is formed;

[0085] The historical time series graph-level representation vector sequence is used as input and fed into a temporal convolutional network;

[0086] Using the temporal convolutional network to learn the historical time series graph-level representation vector sequence to obtain a trend feature vector;

[0087] Performing feature concatenation or weighted fusion on the graph-level representation vector at the current moment and the trend feature vector to construct a fused representation vector;

[0088] The fusion representation vector is input into a trained classification and prediction module, and a potential failure mode label and a corresponding evolution trend score are output to form the fusion result.

[0089] In a further preferred embodiment of the present invention, in order to achieve unified analysis and deep integration of multi-node status information of the entire transmission line, this embodiment constructs a global fault identification model based on a graph neural network. First, all tower nodes deployed with heterogeneous sensor terminals and their actual physical connection relationships are abstracted into a graph structure. The nodes in the graph correspond to transmission towers, and the edges correspond to the physical topological relationships of the line connections. For each node, its input feature vector is composed of the final state label output by the edge computing unit and the monitoring data corresponding to the node (including tension, current, voltage, inclination, humidity, image features, etc.), which is spliced and combined to fully characterize the node's operating status and external environment characteristics at the current moment.

[0090] Based on this graph structure, a multi-layer graph attention neural network is constructed, and information propagation and feature updates between nodes are performed iteratively layer by layer. In each layer, a propagation update formula based on the attention mechanism is adopted to achieve weighted convergence of the influence of different neighbor node features, thereby enhancing the sensitivity to fault-related areas. The output of the final layer of the graph neural network is a high-level semantic representation of each node. This embodiment further introduces a global pooling operation on this basis to extract the graph-level representation vector and input the vector into the evolution trend recognition module. This module can use a temporal convolutional network or a gated recurrent unit structure to model the timing graph data, thereby effectively identifying potential fault modes and their evolution trends over time. The final output fusion result not only has global spatial consistency, but also reflects the dynamic characteristics of state evolution, which significantly improves the recognition ability of this embodiment for complex fault chains, and has high accuracy and real-time performance.

[0091] Preferably, the fusion representation vector is input into a trained classification and prediction module, and the potential failure mode label and the corresponding evolution trend score are output to form the fusion result, including:

[0092] Inputting the fused representation vector into the fault classification subnetwork in the classification and prediction module; the fault classification subnetwork includes at least one fully connected layer and one softmax layer, which is used to output the probability distribution of multiple potential fault types;

[0093] Synchronously inputting the fused representation vector into the risk trend prediction subnetwork in the classification and prediction module; the risk trend prediction subnetwork includes at least one fully connected regression layer for outputting a fault evolution risk score;

[0094] The output results of the classification subnetwork and the prediction subnetwork are combined as the fusion result; the fusion result is used to characterize the current state of the transmission line and the future evolution trend.

[0095] In a further preferred embodiment of the present invention, in order to enhance the intelligent analysis and prediction capabilities of the operating status of complex transmission lines, after completing the global feature fusion of the graph neural network, this embodiment inputs the obtained fusion representation vector into the pre-trained classification and prediction module to perform the identification of potential fault modes and quantitative evaluation of evolution trends. The classification and prediction module includes two parallel sub-networks, which are respectively used to realize the discrimination of multiple categories of fault types and the continuous prediction of risk evolution trends. First, the fusion representation vector is input into the fault classification sub-network, which is composed of several fully connected layers and finally connected to a Softmax layer to realize the multi-classification probability output of typical fault modes (such as broken wires, excessive ice coverage, dancing fatigue, etc.), reflecting the possibility of different risk types under the current state.

[0096] At the same time, the fusion representation vector is synchronously input to the risk trend prediction subnetwork, which uses at least one fully connected regression layer structure to quantitatively predict the risk development trend implied by the current monitoring data, and outputs a continuous value evolution score, which is used to measure the severity or acceleration of the fault mode evolution over time. The output results of the above two subnetworks together constitute the fusion result, which not only characterizes the state category of the transmission line at the current moment, but also provides a quantifiable future evolution trend assessment. This structural design of fusion classification and trend prediction breaks through the traditional technical limitations of only performing state identification, realizes the organic combination of static state judgment and dynamic trend deduction, and significantly improves the perception ability and predictive foresight of complex operating scenarios of this embodiment, reflecting strong creativity and engineering practical value.

[0097] Preferably, the fusion results are subjected to risk dynamic modeling, a risk factor library is constructed, and a real-time failure probability map is generated based on a Bayesian network, including:

[0098] Extract physical quantity characteristics and statistical indicators related to fault types such as conductor breakage, mechanical fatigue, excessive icing, abnormal ground current, and external force interference based on preset known fault cases and historical monitoring data;

[0099] A risk factor library is constructed based on the physical quantity characteristics and the statistical indicators; the risk factors include tension change rate, tilt angle change amplitude, temperature and humidity combined fluctuation index, breeze speed anomaly value, current fluctuation frequency and image icing index;

[0100] Performing feature attribution analysis on the fusion results, using correlation analysis and information gain methods to screen out feature dimensions related to the risk factors and use them as key features in the fusion results;

[0101] Mapping the key features to corresponding risk factor dimensions to construct a risk factor value vector at the current moment, which is used to characterize the multidimensional risk situation under the current transmission line operation state;

[0102] Based on the historical monitoring data, known fault labels and engineering knowledge, a structural learning method is used to establish the dependency relationship between the risk factors and generate the topological structure and conditional probability table of the Bayesian network;

[0103] The risk factor value vector at the current moment is input into the Bayesian network model as evidence, and the posterior probability of each fault type is calculated using the inference algorithm, and a real-time fault probability map is output;

[0104] Based on the maximum a posteriori probability results of various risk events in the real-time fault probability map, risk type classification and identification are performed for status warning and risk intervention decision-making of the transmission line.

[0105] Furthermore, in order to achieve real-time risk perception and classification identification under the operation status of transmission lines, this embodiment first mines monitoring features and statistical indicators that are highly correlated with typical fault events based on a large amount of historical monitoring data and known fault cases. These fault types include but are not limited to conductor breakage, mechanical fatigue, excessive icing, abnormal grounding current, and external force interference. Through feature engineering, a group of key physical quantity indicators with strong characterization capabilities and high measurability, such as tension change rate, inclination angle change amplitude, temperature and humidity combined fluctuation index, micro-wind speed anomaly, current fluctuation frequency, and image icing index, are extracted and constructed as a risk factor library. These risk factors provide a structured input basis for subsequent modeling.

[0106] On this basis, this embodiment performs feature attribution analysis on the fusion results output by the graph neural network. Using statistical correlation analysis and information gain evaluation methods, it screens out feature dimensions that have a significant mapping relationship with the constructed risk factors and defines them as key features in the fusion results. This step ensures that the information extracted from the fusion representation is physically interpretable and can accurately reflect the source of the risk factors for the occurrence of faults. Subsequently, the system maps these key features to the risk factor space and generates a risk factor value vector that characterizes the current state, which is used to express the risk distribution state of the current transmission line in multiple dimensions.

[0107] To model the potential causal and conditional dependencies between risk factors, this invention introduces Bayesian network technology. Specifically, this embodiment utilizes historical monitoring data and labeled fault type samples, combined with on-site expert knowledge, to automatically learn the dependency paths between risk factors through a structural learning algorithm. This constructs the topological structure of the Bayesian network, and further utilizes a parameter learning algorithm to generate a conditional probability table. This Bayesian model possesses strong uncertainty reasoning capabilities, effectively addressing the complex relationships between multi-source monitoring data and providing real-time probabilistic prediction support.

[0108] Finally, this embodiment inputs the risk factor value vector obtained at the current moment as evidence into the Bayesian network, and uses variational reasoning or exact reasoning algorithms to calculate the posterior probability distribution of various risk events (such as broken wires, excessive ice coverage, etc.), and outputs a real-time fault probability map. Based on the maximum posterior probability results of each risk event in the map, this embodiment completes the classification and identification of risk types, and provides a quantitative basis for early risk warning and scheduling intervention for transmission lines. Compared with traditional rule judgment methods, this method has significant advantages such as dynamic adaptability, probabilistic output, and traceable logical chains, reflecting its innovation and engineering practicality in the identification of complex transmission system states.

[0109] Corresponding to the above method, this embodiment provides a system based on online monitoring of transmission lines, including:

[0110] A data acquisition unit is configured to deploy heterogeneous sensor terminals at multiple tower nodes along the transmission line and collect monitoring data using the heterogeneous sensor terminals; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is provided with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module;

[0111] a label determination unit configured to configure an edge computing unit in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label;

[0112] a fusion unit, configured to perform a global fusion calculation on the final state labels and the monitoring data of the plurality of nodes based on a graph neural network, identify potential failure modes and evolution trends, and obtain a fusion result;

[0113] A fault detection unit is used to perform risk dynamic modeling on the fusion results, construct a risk factor library, and generate a real-time fault probability map based on a Bayesian network; the real-time fault probability map is used to classify and identify at least one of the following risk types: wire breakage, mechanical fatigue caused by dancing, excessive icing, abnormal grounding current, and external force interference.

[0114] The beneficial effects of the present invention are as follows:

[0115] This paper proposes an intelligent analysis method based on online transmission line monitoring. By integrating technologies such as edge computing, graph neural networks, and Bayesian reasoning, it enables efficient perception of multi-source heterogeneous data, intelligent state identification, and real-time prediction of fault risks. This method improves the responsiveness of traditional transmission line monitoring systems to abnormal conditions under complex operating conditions, reduces the burden of manual inspections, and offers excellent real-time and automated capabilities, significantly enhancing the intelligence of transmission line operations and maintenance.

[0116] This invention implements real-time preprocessing and preliminary status analysis of monitoring data by deploying heterogeneous sensor terminals and edge computing units at each tower node. This distributed intelligent architecture effectively alleviates computing pressure on central nodes and improves the overall system's responsiveness and scalability. Furthermore, through edge collaborative reasoning, a certain degree of spatiotemporal information fusion and judgment can be performed locally, providing high-quality preliminary label data for subsequent graph model analysis.

[0117] During the global analysis phase, the present invention constructs a graph neural network to perform feature fusion and inference calculations on the status labels and monitoring data of multiple nodes, identifying potential failure modes and evolution trends, and outputting the fusion results. Combining a trend identification module with a classification and prediction architecture, the system can accurately diagnose the current state and predict future evolution risks, providing a decision-making basis for optimizing equipment operation and maintenance strategies, with excellent predictive capabilities and interpretability.

[0118] Furthermore, the present invention constructs a risk factor library based on the fusion results and introduces a Bayesian network for risk modeling and reasoning, outputting a real-time fault probability map. This map quantifies the probability of occurrence of various risk events, shifting risk identification from empirical judgment to data-driven analysis. By integrating statistical learning with domain knowledge in its modeling strategy, the system possesses powerful generalization and adaptability, effectively supporting intelligent monitoring and risk warning throughout the entire life cycle of transmission lines.

[0119] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0120] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method based on online monitoring of transmission lines, characterized in that: include: Heterogeneous sensor terminals are deployed at multiple tower nodes along the transmission line, and monitoring data is collected using the heterogeneous sensor terminals; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is equipped with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module; An edge computing unit is configured in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label; Based on a graph neural network, a global fusion calculation is performed on the final state labels and the monitoring data of the plurality of nodes to identify potential failure modes and evolution trends, and obtain a fusion result; Performing risk dynamic modeling on the fusion results, building a risk factor library, and generating a real-time failure probability map based on a Bayesian network; The real-time fault probability map is used to classify and identify at least one of the following risk types: conductor breakage, mechanical fatigue caused by dancing, excessive icing, abnormal grounding current, and external force interference.

2. The method for online monitoring of power transmission lines according to claim 1, characterized in that: The image acquisition module triggers an image acquisition task at a set time interval or when the final state label changes; the image information includes conductor sag, tower structure status and ice cover contour feature map.

3. The method for online monitoring of power transmission lines according to claim 1, characterized in that: An edge computing unit is configured in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label, including: Using the local analysis sub-model in the edge computing unit, time series modeling and multi-dimensional feature extraction are performed on the monitoring data of the current node to generate a preliminary state label; Collect monitoring data and preliminary status labels of a plurality of nodes adjacent to the current node within a set time window; Based on the spatial topological relationship and time correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label.

4. The method for online monitoring of power transmission lines according to claim 3, characterized in that: Based on the spatial topological relationship and temporal correlation between the nodes, the collaborative judgment sub-model is used to perform joint reasoning on the state of the current node to generate the final state label, including: Constructing a local graph structure subgraph including the current node and adjacent nodes; the nodes of the local graph structure subgraph represent the locations of transmission towers, and the edges represent the physical connection relationships between the towers; Based on the local graph structure subgraph, using the formula Calculate the probability distribution of the final state label of the current node; wherein, is the final state label of node i; is the initial state label of the adjacent node j; is the set of adjacent nodes of node i; α ij is the attention weight of node j to node i, which is calculated based on the spatial topology and state similarity; W t is a trainable weight matrix for modeling temporal correlation; W s is the trainable weight matrix used to model the static monitoring data of the current node; X i is the original monitoring data feature vector of node i; σ(·) is the activation function used to normalize the output label; the calculation formula of the attention weight is: in, is the cosine similarity between state labels, indicating the consistency of node states; d ij is the geographical distance between node i and node j, which represents the geometric relationship of spatial topology; is the spatial similarity weight based on the Gaussian kernel function, σ represents the scale parameter affecting the distance; γ and λ are the preset weight adjustment factors of state similarity and spatial similarity.

5. The method for online monitoring of power transmission lines according to claim 4, characterized in that: Based on the graph neural network, a global fusion calculation is performed on the final state labels and the monitoring data of the multiple nodes to identify potential failure modes and evolution trends, and obtain a fusion result, including: All the nodes and topological connection relationships are constructed into a graph structure and input into the graph neural network; the input feature vector of each node is obtained by splicing the final state label and the monitoring data; Based on the multi-layer graph attention network structure, the propagation and update formula is used to iteratively update the node features; the propagation and update formula is: Among them, the feature vector of node i at the lth layer, W (l) is the weight matrix of the lth layer, is the attention weight of node j to node i in layer l; In the final layer output of the graph neural network, the global pooling operation and the evolution trend identification module are combined to output a fusion result containing potential fault modes and evolution trends.

6. The method for online monitoring of power transmission lines according to claim 5, characterized in that: In the final layer output of the graph neural network, the global pooling operation is combined with the temporal convolutional network to output a fusion result containing potential failure modes and evolution trends, including: In a plurality of preset continuous time periods, a graph-level representation vector of each time point is obtained based on the graph neural network, and a historical time-series graph-level representation vector sequence arranged in chronological order is formed; The historical time series graph-level representation vector sequence is used as input and fed into a temporal convolutional network; Using the temporal convolutional network to learn the historical time series graph-level representation vector sequence to obtain a trend feature vector; Performing feature concatenation or weighted fusion on the graph-level representation vector at the current moment and the trend feature vector to construct a fused representation vector; The fusion representation vector is input into a trained classification and prediction module, and a potential failure mode label and a corresponding evolution trend score are output to form the fusion result.

7. The method for online monitoring of power transmission lines according to claim 6, characterized in that: The fusion representation vector is input into the trained classification and prediction module, and the potential failure mode label and the corresponding evolution trend score are output to form the fusion result, including: Inputting the fused representation vector into the fault classification subnetwork in the classification and prediction module; the fault classification subnetwork includes at least one fully connected layer and one softmax layer, which is used to output the probability distribution of multiple potential fault types; Synchronously inputting the fused representation vector into the risk trend prediction subnetwork in the classification and prediction module; the risk trend prediction subnetwork includes at least one fully connected regression layer for outputting a fault evolution risk score; The output results of the classification subnetwork and the prediction subnetwork are combined as the fusion result; the fusion result is used to characterize the current state of the transmission line and the future evolution trend.

8. The method for online monitoring of power transmission lines according to claim 1, characterized in that: Perform risk dynamic modeling on the fusion results, build a risk factor library, and generate a real-time failure probability map based on the Bayesian network, including: Extract physical quantity characteristics and statistical indicators related to fault types such as conductor breakage, mechanical fatigue, excessive icing, abnormal ground current, and external force interference based on preset known fault cases and historical monitoring data; A risk factor library is constructed based on the physical quantity characteristics and the statistical indicators; the risk factors include tension change rate, tilt angle change amplitude, temperature and humidity combined fluctuation index, breeze speed anomaly value, current fluctuation frequency and image icing index; Performing feature attribution analysis on the fusion results, using correlation analysis and information gain methods to screen out feature dimensions related to the risk factors and use them as key features in the fusion results; Mapping the key features to corresponding risk factor dimensions to construct a risk factor value vector at the current moment, which is used to characterize the multidimensional risk situation under the current transmission line operation state; Based on the historical monitoring data, known fault labels and engineering knowledge, a structural learning method is used to establish the dependency relationship between the risk factors and generate the topological structure and conditional probability table of the Bayesian network; The risk factor value vector at the current moment is input into the Bayesian network model as evidence, and the posterior probability of each fault type is calculated using the inference algorithm, and a real-time fault probability map is output; Based on the maximum a posteriori probability results of various risk events in the real-time fault probability map, risk type classification and identification are performed for status warning and risk intervention decision-making of the transmission line.

9. A system based on online monitoring of transmission lines, characterized in that: include: A data acquisition unit is configured to deploy heterogeneous sensor terminals at multiple tower nodes along the transmission line and collect monitoring data using the heterogeneous sensor terminals; the monitoring data includes raw physical quantity data and image information; the heterogeneous sensor terminals include multiple nodes; each node is provided with a tension sensor, an inclination sensor, a temperature and humidity sensor, a micro-anemometer, a voltage and current sensor, and an image acquisition module; a label determination unit configured to configure an edge computing unit in the heterogeneous sensor terminal to perform real-time analysis on the monitoring data of the current node according to a preset model in the edge computing unit, generate a preliminary state label, and perform spatiotemporal collaborative judgment with the adjacent monitoring data to obtain a final state label; a fusion unit, configured to perform a global fusion calculation on the final state labels and the monitoring data of the plurality of nodes based on a graph neural network, identify potential failure modes and evolution trends, and obtain a fusion result; A fault detection unit is used to perform risk dynamic modeling on the fusion results, build a risk factor library, and generate a real-time fault probability map based on a Bayesian network; The real-time fault probability map is used to classify and identify at least one of the following risk types: conductor breakage, mechanical fatigue caused by dancing, excessive icing, abnormal grounding current, and external force interference.

Images