Method for monitoring the state of the power grid of an intelligent warning kiosk

Abstract

This invention relates to the field of power pipeline monitoring technology, specifically a method for monitoring the status of power pipelines using intelligent warning grounding boxes. The method includes: deploying intelligent warning grounding boxes at key nodes of the power pipeline network and establishing a monitoring network; simultaneously acquiring multi-source status information such as pipeline structure vibration, groundwater level and water accumulation, and cable sheath grounding current during routine monitoring; employing an improved graph convolutional neural network algorithm that dynamically constructs an adjacency matrix based on the pipeline network's physical topology to extract features and perform spatiotemporal correlation analysis on the multi-source information, generating a comprehensive status feature vector; inputting this vector into a status assessment model to complete anomaly identification and risk assessment, and generating tiered response instructions; the intelligent warning grounding boxes executing local responses and uploading data to the cloud, where the digital model of the pipeline network is updated and a 3D situation map and asset records are generated. This method improves the accuracy and response efficiency of pipeline status monitoring, achieving synergy between monitoring, response, and management.

Description

Technical Field

[0001] This invention relates to the field of power grid monitoring technology, and in particular to a method for monitoring the status of power grids using intelligent warning grounding boxes. Background Technology

[0002] As the core infrastructure for power transmission, the safe and stable operation of power pipelines is directly related to the reliability of power supply. Currently, the industry generally adopts the method of deploying monitoring equipment along the pipeline to monitor the operating status of underground pipelines. Conventional monitoring equipment mainly collects single-type monitoring data. Some equipment can simultaneously acquire multi-source monitoring information, but most of them use traditional data analysis algorithms for data processing.

[0003] Traditional data analysis algorithms often use fixed adjacency matrices, which cannot adapt to the complex and ever-changing physical topology of power grids. This results in inaccurate feature extraction from multi-source monitoring information, biases in spatiotemporal correlation analysis, and difficulty in accurately capturing the true operating status of grid nodes. Furthermore, conventional monitoring equipment can only perform data acquisition and uploading functions, lacking local response capabilities. The cloud management platform cannot update the digital model of the grid in real time based on uploaded data, making it difficult to form a closed loop of monitoring, response, and updating. This leads to delays in handling abnormal events in the grid and the inability to visualize asset status in real time.

[0004] There is a need for a power pipeline network status monitoring solution that can adapt to the physical topology of the pipeline network, achieve accurate analysis of multi-source data, and complete local response and cloud-based collaborative updates. This solution addresses the problems of poor adaptability and insufficient data processing accuracy of conventional monitoring algorithms, as well as insufficient linkage between monitoring equipment and the cloud platform and the lack of status visualization. This invention dynamically constructs an adjacency matrix of a graph convolutional neural network by integrating the physical properties of cables and the real-time status correlation between nodes. This allows the graph structure to adapt to the actual operating state of the pipeline network, improving the accuracy of feature extraction. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for monitoring the status of power grids using an intelligent warning ground box.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for monitoring the status of power grids using an intelligent warning grounding box, comprising: Deploy intelligent warning boxes at key nodes along the power grid to complete equipment power supply and network initialization, and establish a monitoring network; During the routine monitoring cycle, the intelligent warning box synchronously acquires multi-source status monitoring information of the underground pipe network. The multi-source status monitoring information includes pipe network structure vibration information, groundwater level and water accumulation information, and cable sheath grounding current information. Based on the improved graph convolutional neural network algorithm, feature extraction and spatiotemporal correlation analysis are performed on the multi-source state monitoring information to generate a comprehensive state feature vector of the pipeline network nodes. The improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network. The comprehensive status feature vector of the pipeline node is input into the status assessment model to identify abnormal event types and determine risk levels, and generate event discrimination results and corresponding graded response instructions. Based on the hierarchical response instructions, the intelligent warning box is controlled to perform local response actions and upload event data and device status to the cloud management platform through a preset communication protocol; Based on the uploaded event data, the digital model status of the corresponding pipeline section is updated on the cloud management platform, generating a visualized 3D situation map and asset status record.

[0007] As a further aspect of the present invention, the improved graph convolutional neural network algorithm is used to perform feature extraction and spatiotemporal correlation analysis on the multi-source state monitoring information to generate a comprehensive state feature vector of the pipeline network nodes, including: Based on the deployment location of the intelligent warning ground box and the cable connection relationship, a pipeline topology model is constructed with the ground box as the node and the cable connection path as the edge; Initial features in the time domain, frequency domain, and spatial distribution dimensions are extracted from the vibration information of the pipeline structure, groundwater level and water accumulation information, and cable sheath grounding current information to form an initial set of node features for each ground box node. The pipeline topology model is transformed into an adjacency matrix, and the dynamic connection weight of the adjacent edges is calculated based on the physical distance between the grounding nodes, the cable type and the laying environment to generate a dynamic adjacency matrix. The dynamic adjacency matrix and the initial set of node features are input into the improved graph convolutional neural network algorithm. The improved graph convolutional neural network algorithm updates the feature representation of each node through a multi-level information transmission and aggregation mechanism. During each information aggregation process, the status information of the relevant grounding nodes within a specified number of hops on the physically adjacent grounding nodes and the cable connection path is aggregated. After iterating through the information transfer and aggregation a specified number of times, the updated state feature vector of each ground node is output. This vector integrates the correlation information of the local and neighboring areas and serves as the comprehensive state feature vector of the pipeline node.

[0008] As a further aspect of the present invention, the improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network. The working principle of the improved graph convolutional neural network algorithm based on the physical topology of the pipeline network includes: Define the initial physical connection diagram of the pipeline network, with the location of each smart warning box as a node and the cable connection path as an edge; Obtain the attribute information of the cable connection path, including cable type, laying depth, and current cable load rate, and calculate the initial connection weight of each edge; The system acquires the current monitoring information uploaded by each ground cell node in real time, and calculates the spatiotemporal correlation of the states between ground cell nodes based on the monitoring information. The spatiotemporal correlation includes vibration propagation consistency, water level change synchronization and leakage current correlation. The initial connection weights based on physical connections are weighted and fused with the spatiotemporal correlations based on real-time states to generate dynamic edge weights that reflect the strength of physical connections and state correlations at the current moment. Based on the dynamic edge weights between all nodes, the dynamic adjacency matrix is ​​constructed and updated in real time.

[0009] As a further aspect of the present invention, the comprehensive state feature vector of the pipeline node is input into the state assessment model to identify abnormal event types and determine risk levels, including: The comprehensive state feature vector of the network node is input into an evaluation model containing multiple parallel sub-networks; One of the subnetworks analyzes the dimensions representing vibration information in the feature vectors, and by comparing them with a historical normal vibration pattern library, it identifies disturbance events of the types of construction disturbance, geological settlement, and vehicle rolling, and outputs the disturbance level. Another sub-network analyzes the dimensions representing water immersion information in the feature vector, and combines the rate of change and duration of water immersion to identify water immersion events of the types of slow seepage, gushing water, and internal flooding, and outputs the water immersion level. Another subnetwork analyzes the dimensions representing leakage information in the feature vector, and combines the leakage current amplitude, fluctuation characteristics and harmonic content to identify electrical events of insulation aging, partial discharge and external breakdown types, and outputs the electrical risk level. The disturbance level, water immersion level, and electrical risk level are input into a comprehensive decision logic. The comprehensive decision logic determines whether the final event type is a single event or a coupled event based on preset coupled risk rules, and determines a comprehensive risk level.

[0010] As a further aspect of the present invention, based on the hierarchical response command, the intelligent warning box is controlled to perform a local response action, and the event data and device status are uploaded to the cloud management platform via a preset communication protocol, including: When the graded response instruction is a level one warning instruction, the control intelligent warning box will store the recorded data containing the event discrimination result into the local cache, and package and upload it to the cloud management platform in the next preset regular communication window period; When the graded response command is a level 2 alarm command, the high-power communication unit of the intelligent warning box is immediately activated, a high-priority communication link is established with the cloud management platform, alarm information is pushed, and the device's built-in warning light is activated to perform a slow flashing mode. When the graded response instruction is a level three emergency reporting instruction, the communication channel is immediately seized, and an emergency reporting process containing real-time streaming data, precise device location information and event judgment results is initiated to the cloud management platform. At the same time, the high-frequency flashing of the warning light and the sounding of the buzzer are activated. After the event data is uploaded, the smart warning box will receive confirmation instructions or new control instructions from the cloud management platform, and adjust its subsequent monitoring or response behavior according to the control instructions.

[0011] As a further aspect of the present invention, the step of updating the digital model status of the corresponding pipeline section on the cloud management platform based on the uploaded event data, and generating a visualized three-dimensional situation map and asset status record, includes: The cloud management platform receives event data and device status information corresponding to the graded response instructions uploaded from the smart warning box, and parses out the location, type, grade and timestamp of the event. In the 3D digital model of the power network maintained by the platform, the network section and equipment node corresponding to the location of the event are located. Based on the analyzed event type and level, update the status attributes of the pipeline section corresponding to the event location in the 3D digital model, and render the pipeline section corresponding to the event location with a highlighted color, wherein the color maps the comprehensive risk level; In the three-dimensional digital model, the diffusion process of state influence from the event occurrence node to adjacent nodes is dynamically displayed along the cable path in the form of flowing particles or pulsating light effects. All data related to the abnormal events corresponding to the currently received event data and equipment status information, including original monitoring information fragments, event judgment results, and response records, are bound to the pipeline asset code and electronic identity of the smart warning box where the event occurred, and stored as a complete operation and maintenance record in the asset lifecycle database.

[0012] As a further aspect of the present invention, based on the deployment location of the intelligent warning junction box and the cable connection relationship, a pipeline topology model is constructed with the junction box as nodes and the cable connection path as edges, including the following steps: Obtain the as-built drawings or digital design drawings of the power network in the target monitoring area from the cloud management platform, and extract the preset installation location coordinates of all smart warning ground boxes and the preset cable laying path information from the drawings; After the smart warning ground box completes physical installation and power supply initialization, the actual geographical coordinates of each ground box are obtained through the global satellite navigation system positioning module, and the device is bound to the preset installation location coordinates through the device's unique identification code to complete the digital registration of the deployment location; Based on the preset laying path information of the cable, determine whether there is a direct cable connection between any two smart warning boxes, and define the two boxes with a direct cable connection as adjacent nodes. In the pipeline topology model, each smart warning box that has completed digital registration is a graph node, and the actual cable laying path between any two adjacent nodes is an undirected edge or a directed edge, wherein the direction of the edge is defined according to the power transmission direction of the cable or the preset monitoring information flow direction. The attributes of each node in the pipeline topology model are initialized to its corresponding unique device identifier and actual geographical coordinates, and the attributes of each edge are initialized to the identifier, type and length of the cable it represents.

[0013] As a further aspect of the present invention, the dynamic adjacency matrix and the initial set of node features are input into the improved graph convolutional neural network algorithm. The improved graph convolutional neural network algorithm updates the feature representation of each node through a multi-level information transmission and aggregation mechanism, including: The dynamic adjacency matrix and the initial set of node features for each node are input into the first hidden layer of the improved graph convolutional neural network; In the first hidden layer, for the current target node, its first-order physical neighbor node set is determined according to the dynamic adjacency matrix, and the initial set of node features of all nodes in the first-order physical neighbor node set is aggregated to generate the first-level aggregated features. The initial set of node features of the target node and the first-level aggregated features are subjected to nonlinear transformation and feature fusion to output the intermediate feature representation of the target node in the first hidden layer; The intermediate feature representations of all nodes output from the first hidden layer, along with the dynamic adjacency matrix, are input into the second hidden layer of the improved graph convolutional neural network. In the second hidden layer, for the target node, its extended set of neighboring nodes is determined according to the dynamic adjacency matrix. The extended set of neighboring nodes includes its first-order physical neighboring nodes and related nodes that can be reached within a preset number of topological hops through cable connection paths. The intermediate feature representations of all nodes in the extended set of neighboring nodes are aggregated to generate a second-level aggregated feature. The intermediate feature representations and second-level aggregated features of the target node from the first hidden layer are subjected to nonlinear transformation and feature fusion to output the updated state feature vector of the target node in the second hidden layer. The multi-level aggregation and fusion process is iteratively executed until the preset network depth is reached, and the final output state feature vector of each ground node is used as the comprehensive state feature vector of the pipeline node.

[0014] As a further aspect of the present invention, the step of obtaining the attribute information of the cable connection path, including cable type, laying depth, and current cable load rate, and calculating the initial connection weight of each edge, includes: From the power grid asset database, query the cable type, insulation material and design life corresponding to the cable connection path, and assign a basic weight coefficient that reflects the importance of its physical connection to different cable types; Acquire cable laying depth data, compare the laying depth with the preset standard depth, and calculate a depth correction coefficient that reflects the differences in the laying environment based on the degree of depth deviation. The system receives cable load rate data from the power system in real time and calculates a load dynamic coefficient that reflects the real-time operating pressure based on the ratio of the current load rate to the rated load. The initial connection weight of the cable connection path is obtained by weighting the basic weight coefficient, the depth correction coefficient and the load dynamic coefficient, wherein the basic weight coefficient has the dominant weight.

[0015] As a further aspect of the present invention, the comprehensive decision logic determines whether the final event type is a single event or a coupled event based on preset coupling risk rules. The generation of the coupling risk rules includes: Frequently co-occurring anomalous event combinations are extracted from the historical event database, including concurrent vibration and water immersion, and concurrent water immersion and electrical leakage. For each frequently co-occurring combination of anomalous events, we analyze their sequence of occurrence, time intervals, and spatial relationships to define a coupled event pattern. For each defined coupled event pattern, a risk superposition coefficient is set, which is greater than the simple sum of the risk levels of individual events; Establish a decision tree for coupled events. The input of the decision tree is the disturbance level, water immersion level, electrical risk level and their spatiotemporal relationship output by each sub-network. The leaf nodes of the decision tree correspond to specific single event types or coupled event types. The coupled event discrimination decision tree is integrated with the risk superposition coefficient to form the preset coupled risk rule.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: An improved graph convolutional neural network algorithm, which dynamically constructs an adjacency matrix based on the physical topology of a power grid, is used to extract features and perform spatiotemporal correlation analysis on multi-source state monitoring information, consisting of grid structure vibration information, groundwater level and water accumulation information, and cable sheath grounding current information, generating a comprehensive state feature vector for grid nodes. This algorithm overcomes the limitation of traditional fixed adjacency matrices, which cannot adapt to complex grid topologies, by precisely matching the dynamic adjacency matrix with the grid's physical topology. This makes the feature extraction of multi-source monitoring information more closely aligned with the actual operating scenario of the grid, and the spatiotemporal correlation analysis more targeted. It effectively avoids data processing deviations caused by poor topology adaptability in traditional algorithms, resulting in more comprehensive and accurate feature capture of grid node states.

[0017] Using intelligent warning ground boxes deployed at key nodes along the power grid as edge terminals, after acquiring event identification results and tiered response instructions, the system executes local response actions and uploads event data and equipment status to the cloud management platform via a preset communication protocol. The cloud management platform updates the digital model status of the corresponding grid section based on the uploaded data, generating a visualized 3D situation map and asset status records. This collaborative linkage model between edge terminals and the cloud platform changes the conventional monitoring equipment's limitation of only being able to collect and upload data without local response capabilities. It enables rapid local response to abnormal events, shortening the event handling cycle. Simultaneously, the real-time updates of the cloud-based digital model and the generation of the 3D situation map fill the gaps in conventional monitoring, such as the inability to visualize grid status in real time and the lack of dynamic records of asset status, achieving seamless integration of monitoring data, local response, and cloud updates. Attached Figure Description

[0018] Figure 1 A flowchart of the power grid status monitoring method for the intelligent warning ground box according to the present invention; Figure 2 The flowchart illustrates the process of dynamically constructing an adjacency matrix based on the physical topology of the pipeline network in the improved graph convolutional neural network algorithm. Figure 3 This is a flowchart illustrating the process of inputting a comprehensive state feature vector into a state assessment model for anomaly event identification and risk level determination. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 The power grid status monitoring method for the intelligent warning ground box of the present invention includes the following steps in its overall implementation scheme: Intelligent warning grounding boxes are deployed at key nodes along the power grid to initialize the power supply and network connection of the equipment, thereby establishing a distributed monitoring network. During routine monitoring cycles, the deployed intelligent warning grounding boxes synchronously collect multi-source status monitoring information of the underground pipeline network. This information mainly includes pipeline structure vibration information, groundwater level and water accumulation information, and cable sheath grounding current information. Based on an improved graph convolutional neural network algorithm, feature extraction and spatiotemporal correlation analysis are performed on the collected multi-source status monitoring information. This algorithm dynamically constructs an adjacency matrix based on the actual physical topology of the pipeline network, ultimately generating a comprehensive status feature vector for each pipeline node. This comprehensive status feature vector is input into a preset status assessment model to identify abnormal event types and determine risk levels, outputting event discrimination results and corresponding graded response commands. Based on the generated graded response commands, the intelligent warning grounding boxes that trigger the commands are controlled to execute corresponding local response actions, while simultaneously uploading event data and equipment status to the cloud management platform through a preset communication protocol. After receiving the uploaded event data, the cloud management platform updates the digital model status of the corresponding pipeline section within the platform and generates a visualized 3D situation map and asset status record, completing the closed loop from on-site perception to cloud management.

[0022] In one embodiment of the present invention, a pipeline topology model is constructed based on the deployment location of the smart warning ground boxes and the cable connection relationship, with the ground boxes as nodes and the cable connection paths as edges. This process includes: obtaining the as-built drawings or digital design drawings of the power pipeline network for the target monitoring area from a cloud management platform, and extracting the preset installation location coordinates of all smart warning ground boxes and the preset cable laying path information from the drawings. After the smart warning ground boxes complete physical installation and power supply initialization, the actual geographical coordinates of each ground box are obtained through its built-in global satellite navigation system positioning module, and bound to the preset installation location coordinates extracted from the drawings using the device's unique identifier, thus completing the digital registration of the deployment location. Based on the preset cable laying path information, it is determined whether there is a direct cable connection relationship between any two smart warning ground boxes, and the two ground boxes with a direct cable connection relationship are defined as adjacent nodes. In the pipeline topology model, each smart warning ground box that has completed digital registration is a graph node, and the actual cable laying path between any two adjacent nodes is an undirected or directed edge, where the direction of the edge is defined according to the power transmission direction of the cable or the preset monitoring information flow direction. The attributes of each node in the pipeline topology model are initialized to its corresponding unique device identifier and actual geographical coordinates, and the attributes of each edge are initialized to the identifier, type and length of the cable it represents.

[0023] After constructing the pipeline network topology model, initial features in the time domain, frequency domain, and spatial distribution dimensions are extracted from pipeline structure vibration information, groundwater level and water accumulation information, and cable sheath grounding current information, forming an initial set of node features for each grounding junction node. The pipeline network topology model is then transformed into an adjacency matrix, and based on the physical distance between junction nodes, cable type, and laying environment, the dynamic connection weights of adjacent edges are calculated to generate a dynamic adjacency matrix. The dynamic adjacency matrix and the initial set of node features are input into an improved graph convolutional neural network algorithm. This algorithm updates the feature representation of each node through a multi-level information transmission and aggregation mechanism. During each information aggregation process, state information from physically adjacent junction nodes and related junction nodes within a specified number of hops on the cable connection path is aggregated. After iterating through information transmission and aggregation a specified number of times, the updated state feature vector of each junction node is output. This vector integrates local and neighboring region correlation information and serves as the comprehensive state feature vector of the pipeline network node.

[0024] In practical implementation, a pipeline topology model is constructed based on the deployment location of the smart warning ground boxes and the cable connection relationship, with the ground boxes as nodes and the cable connection paths as edges. Taking a cable trench section containing four smart warning ground boxes as an example, the implementation process is as follows: The digital design drawings of the power network for this section are obtained from the cloud management platform. The preset installation location coordinates of the four smart warning ground boxes are extracted from the drawings, and the coordinate information is stored in latitude and longitude format. Simultaneously, the three preset cable laying paths connecting these coordinate points are extracted, and the path information includes cable identifiers and inflection point coordinate sequences. After the smart warning ground boxes complete physical installation and power supply initialization, each smart warning ground box obtains its actual geographical coordinates through its built-in global satellite navigation system positioning module. For example, the actual coordinates of smart warning ground box A are (X1, Y1), and the actual coordinates of smart warning ground box B are (X2, Y2). The actual coordinates are bound to the corresponding preset installation location coordinates extracted from the drawings using the device's unique identifier, completing the digital registration of the deployment location. The registration information is recorded in the device management list on the cloud management platform. Based on the preset cable laying path information, it can be determined that there is a direct cable connection between smart warning box A and smart warning box B, another direct cable connection between smart warning box B and smart warning box C, and a third direct cable connection between smart warning box C and smart warning box D. There is no direct cable connection between smart warning box A and smart warning box C. Based on this logic, the two boxes with a direct cable connection are defined as adjacent nodes. In the pipeline topology model, each digitally registered smart warning box is considered a graph node, with four nodes in total. The actual cable laying path between any two adjacent nodes is defined as an undirected edge. In specific implementations, the direction of the edge is defined according to the preset flow direction from the upstream box to the downstream box based on monitoring information. For example, the edge from smart warning box A to smart warning box B is defined as a directed edge. Initialize the attributes of each node in the pipeline topology model to its corresponding unique device identifier and actual geographic coordinates. For example, the unique device identifier is “E-BOX-001” and the actual geographic coordinates are (X1, Y1). Initialize the attributes of each edge to the identifier, type and length of the cable it represents. For example, the cable identifier is “CABLE-AB”, the cable type is “YJV22-8.7 / 10kV”, and the cable length is obtained by calculating the distance of the path inflection point coordinates, for example, 125.7 meters.

[0025] In some embodiments, after the pipeline network topology model is constructed, features are extracted from synchronously acquired multi-source status monitoring information to form an initial set of node features. For pipeline network structural vibration information, root mean square value and peak factor features are extracted from the time domain dimension, and 1 / 3 octave band energy features are extracted from the frequency domain dimension. For groundwater level and water accumulation information, current water level value and water level change rate features are extracted. For cable sheath grounding current information, fundamental effective value and third harmonic content features are extracted. All extracted features together constitute the initial set of node features for each grounding box node. The pipeline network topology model is transformed into an adjacency matrix, where the rows and columns of the matrix correspond to four intelligent warning grounding box nodes, and the values ​​of the matrix elements represent the connection relationships between nodes. Based on the physical distance between grounding box nodes, cable type, and laying environment, the dynamic connection weight of adjacent edges is calculated, where the physical distance is calculated based on node coordinates, the cable type is obtained from the asset database query, and the laying environment is determined based on construction records, generating a dynamic adjacency matrix. In a specific implementation, the calculation of dynamic connection weight can be expressed by a formula, for example, the dynamic connection weight of edge e. for: in: This indicates the physical distance between two connected smart warning ground box nodes. This indicates the importance coefficient determined based on the cable type. This indicates the environmental coefficient based on the environmental assessment of the laying environment. The harmonic coefficient is the coefficient of harmonicity. The distance decay factor is used. The dynamic adjacency matrix and the initial set of node features are input into an improved graph convolutional neural network algorithm. This algorithm performs multi-level information transfer and aggregation through a structure containing two hidden layers. During the information aggregation process in the first hidden layer, for the target node smart warning box B, its first-order physical neighbor set is determined based on the dynamic adjacency matrix. This set includes smart warning box A and smart warning box C. The initial sets of node features for smart warning box A and smart warning box C are aggregated to generate the first-level aggregated features. Nonlinear transformation and feature fusion are performed on the initial set of node features for smart warning box B and the first-level aggregated features to output the intermediate feature representation of smart warning box B in the first hidden layer. In the second hidden layer, for the target node Smart Warning Box B, its extended set of neighboring nodes is determined based on the dynamic adjacency matrix. The preset topology hop count is 2. The extended set of neighboring nodes includes its first-order physical neighbors Smart Warning Box A and Smart Warning Box C, as well as Smart Warning Box D connected through Smart Warning Box C. The intermediate feature representations of Smart Warning Box A, Smart Warning Box C, and Smart Warning Box D are aggregated to generate a second-level aggregated feature. A nonlinear transformation and feature fusion are performed on the intermediate feature representation of Smart Warning Box B from the first hidden layer and the second-level aggregated feature, outputting the updated state feature vector of Smart Warning Box B in the second hidden layer. The improved graph convolutional neural network algorithm iteratively performs two information transfers and aggregations, finally outputting the updated state feature vector of each box node. The state feature vectors of Smart Warning Box A, Smart Warning Box B, Smart Warning Box C, and Smart Warning Box D are collectively used as the comprehensive state feature vector of the network node for subsequent state evaluation.

[0026] In one embodiment of the present invention, the improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network, see reference. Figure 2Its working principle includes defining an initial physical connection diagram of the pipeline network using the location of each intelligent warning box as a node and the cable connection path as an edge. It acquires the attribute information of the cable connection path, including cable type, laying depth, and current cable load rate, and calculates the initial connection weight of each edge. Specifically, this process involves querying the cable model, insulation material, and design life corresponding to the cable connection path from the power pipeline network asset database, assigning a basic weight coefficient reflecting the importance of its physical connection to different cable types. It acquires cable laying depth data, compares the laying depth with a preset standard depth, and calculates a depth correction coefficient reflecting differences in the laying environment based on the degree of depth deviation. It receives cable load rate data from the power system in real time, and calculates a load dynamic coefficient reflecting real-time operating pressure based on the ratio of the current load rate to the rated load. Finally, it performs a weighted product of the basic weight coefficient, the depth correction coefficient, and the load dynamic coefficient to obtain the initial connection weight of the cable connection path, where the basic weight coefficient has the dominant weight. After obtaining the initial connection weights, the current monitoring information uploaded by each geocell node is acquired in real time. The spatiotemporal correlation of the states between geocell nodes is calculated based on this monitoring information. This spatiotemporal correlation includes vibration propagation consistency, water level change synchronization, and leakage current correlation. The initial connection weights based on physical connections are weighted and fused with the spatiotemporal correlations based on real-time states to generate dynamic edge weights reflecting the strength of physical connections and state correlations at the current moment. Based on these dynamic edge weights among all nodes, the dynamic adjacency matrix is ​​constructed and updated in real time.

[0027] In practical implementation, the improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network. The first step in its working principle is to define an initial physical connection graph of the pipeline network using the location of each smart warning box as a node and the cable connection path as an edge. For example, in a simple triangular topology containing smart warning box nodes P, Q, and R, node P is connected to node Q by cable L1, node Q is connected to node R by cable L2, and node P is connected to node R by cable L3. The attribute information of the cable connection paths is obtained to calculate the initial connection weight of each edge. For cable L1, its cable type is YJV22-8.7 / 10kV, its insulation material is cross-linked polyethylene, and its design life is 30 years, as determined from the power pipeline network asset database. Based on preset mapping rules, a basic weight coefficient reflecting the importance of its physical connection is assigned to this type of cable, denoted as [missing information - likely a typo]. The laying depth data for cable L1 is 1.2 meters, while the preset standard depth is 1.5 meters. A depth correction factor reflecting differences in the laying environment is calculated based on the degree of depth deviation. The calculation formula is as follows: ,get The system receives cable load rate data from the power monitoring system in real time. The current load rate of cable L1 is 75% of its rated load. Based on the ratio of the current load rate to the rated load, a load dynamic coefficient reflecting the real-time operating pressure is calculated. The calculation formula is as follows: ,get The basic weighting coefficient of cable L1. Depth correction factor With load dynamic coefficient The weighted product is calculated using the following formula: Where: the exponents 0.6, 0.2, and 0.2 represent the weights of each coefficient, with the basic weight coefficient having the dominant weight, to calculate the initial connection weight of cable connection path L1. Repeat the above process for cables L2 and L3 to calculate their initial connection weights. and .

[0028] In some embodiments, after calculating the initial connection weights of each edge, the current monitoring information uploaded by each ground cell node is obtained in real time, and the spatiotemporal correlation of the states between ground cell nodes is calculated based on the monitoring information. For example, within a time window T, node P and node Q synchronously upload vibration acceleration sequences, and the cross-correlation coefficient of the two sequences is calculated to obtain the vibration propagation consistency. Nodes P and Q synchronously uploaded water level value sequences. The similarity of the changing trends of the two sequences was calculated to determine the synchronicity of water level changes. Nodes P and Q synchronously uploaded the fundamental RMS value sequence of the grounding current. The covariance of the two sequences was calculated to obtain the correlation of the leakage current. To ensure consistent vibration propagation Synchronization of water level changes Correlation with leakage current By performing synthesis, the spatiotemporal correlation between node P and node Q based on real-time state is obtained. The calculation formula is: ,in These are the weight coefficients for each item, and their sum is 1. The initial connection weights are based on physical connections. Spatiotemporal correlation based on real-time state Weighted fusion is performed to generate dynamic edge weights that reflect the strength of physical connectivity and state association at the current moment. The calculation formula is: in: It is a fusion factor used to balance the influence of historical physical properties and real-time state. Based on the dynamic edge weights between all nodes, a dynamic adjacency matrix is ​​constructed and updated in real time. For a network containing nodes P, Q, and R, the element in the i-th row and j-th column of the dynamic adjacency matrix A is... That is, the dynamic edge weights between corresponding node pairs. If there is no direct cable connection between two nodes, the corresponding dynamic edge weight is 0. The dynamic adjacency matrix A will change with the input of new monitoring information and spatiotemporal correlation. It is periodically updated through recalculation.

[0029] Optionally, when acquiring cable laying depth data, if a cable segment is laid in a conduit, the depth correction coefficient calculation must also consider the impact of conduit type on heat dissipation and mechanical protection, introducing a conduit adjustment factor based on the standard depth comparison. Optionally, when calculating the load dynamic coefficient, if the cable is in an overload state (i.e., the current load rate exceeds 100% of the rated load), the load dynamic coefficient calculation formula will introduce a nonlinear penalty term, significantly increasing the load dynamic coefficient value to highlight the risk. It can be understood that vibration propagation consistency calculation in spatiotemporal correlation can select the energy of the vibration signal within a specific frequency band for cross-correlation calculation to more accurately capture the propagation characteristics of events such as construction vibration. It can be understood that the fusion factor... The value can be adjusted according to different stages of pipeline operation. In the early stage of new pipeline operation, a higher value can be set to rely more on physical connection attributes. After the pipeline operation is stable, the value can be appropriately reduced to enhance the weight of the impact of real-time status.

[0030] In one embodiment of the present invention, see [reference] Figure 3 The comprehensive state feature vector of the pipeline network nodes is input into a state assessment model to identify abnormal event types and determine risk levels. This state assessment model contains multiple parallel sub-networks. One sub-network analyzes the dimension representing vibration information in the feature vector, and by comparing it with a historical normal vibration pattern database, identifies disturbance events of the types of construction disturbance, geological settlement, and vehicle crushing, and outputs the disturbance level. Another sub-network analyzes the dimension representing water immersion information in the feature vector, and, combined with the rate of water immersion change and duration, identifies water immersion events of the types of slow seepage, inrush, and internal flooding, and outputs the water immersion level. Yet another sub-network analyzes the dimension representing leakage current information in the feature vector, and, combined with the leakage current amplitude, fluctuation characteristics, and harmonic content, identifies electrical events of the types of insulation aging, partial discharge, and external breakdown, and outputs the electrical risk level. The disturbance level, water immersion level, and electrical risk level are input into a comprehensive decision logic. The comprehensive decision logic determines the final event type as a single event or a coupled event based on preset coupling risk rules, and determines a comprehensive risk level.

[0031] The generation of the preset coupling risk rules includes: mining frequently co-occurring anomalous event combinations from a historical event database. For example, an association rule mining algorithm (such as the Apriori algorithm) is used to analyze event sequences in the historical event database to identify event combinations whose support and confidence both exceed preset thresholds. These anomalous event combinations include concurrent vibration and water immersion, and concurrent water immersion and electrical leakage. For each frequently co-occurring anomalous event combination, the order of occurrence, time interval, and spatial relationship are analyzed to define a coupling event pattern. A risk superposition coefficient is set for each defined coupling event pattern, which is greater than the simple sum of the risk levels of individual events. A coupling event discrimination decision tree is established. The input of the decision tree is the disturbance level, water immersion level, electrical risk level, and their spatiotemporal relationship output by each sub-network. The leaf nodes of the decision tree correspond to specific single event types or coupling event types. The coupling event discrimination decision tree and the risk superposition coefficient are integrated to form the preset coupling risk rules.

[0032] In practical implementation, the comprehensive state feature vector of the pipeline network nodes is input into the state assessment model for abnormal event type identification and risk level determination. The state assessment model contains multiple parallel sub-networks. The comprehensive state feature vector is a multi-dimensional vector output from an improved graph convolutional neural network algorithm, for example, with a dimension of 128. For a comprehensive state feature vector from a smart warning box node X, one of the parallel sub-networks analyzes the dimensions representing vibration information in the feature vector. The vibration information dimensions correspond to the 1st to 30th dimensions of the feature vector. This sub-network is a trained classification and regression model. By comparing the input feature subset with a historical normal vibration pattern library, which stores vibration feature benchmarks under geologically stable conditions without external disturbances, the sub-network calculates the deviation between the input features and the benchmarks. When the deviation exceeds a threshold, the specific type is identified, for example, the output identification result is "construction disturbance", and a quantified disturbance level is output, such as level 2. Another parallel sub-network analyzes the dimensions representing water immersion information in the feature vector, corresponding to dimensions 31 to 60 of the feature vector. This sub-network combines the water immersion change rate and duration for analysis. The water immersion change rate is extracted from the features, and the duration is obtained from the monitoring time-series information. The sub-network identifies the specific type, such as outputting "slow seepage," and outputs a quantified water immersion level, such as level 1. Yet another parallel sub-network analyzes the dimensions representing leakage current information in the feature vector, corresponding to dimensions 61 to 128 of the feature vector. This sub-network combines the leakage current amplitude, fluctuation characteristics, and harmonic content for analysis, identifying the specific type, such as outputting "insulation aging," and outputting a quantified electrical risk level, such as level 1. Disturbance level 2, water immersion level 1, and electrical risk level 1 are input into a comprehensive decision logic, which makes a judgment based on preset coupled risk rules.

[0033] In some embodiments, the process of generating the preset coupling risk rules includes mining frequently co-occurring combinations of abnormal events from a historical event database, which records all abnormal event records and their spatiotemporal information reported by smart warning boxes over the past year. Through association rule analysis, frequently co-occurring combinations of abnormal events are identified, such as the combination of "construction disturbance" and "slow seepage" occurring sequentially within 24 hours at spatially proximate nodes, and the combination of "water inrush" and "insulation aging" occurring at the same node. For each frequently co-occurring combination of abnormal events, the order of occurrence, time interval, and spatial relationship are analyzed. For example, if "slow seepage" occurs at a downstream node of the same cable trench section within 3 hours after "construction disturbance," a coupling event pattern named "construction-seepage coupling" is defined. For each defined coupling event pattern, a risk superposition coefficient is set, which is greater than the simple sum of the risk levels of individual events. For example, the risk superposition coefficient κ for the "construction-seepage coupling" pattern is set to 1.5. A decision tree for identifying coupled events is established. The input to the decision tree consists of the disturbance level, water immersion level, electrical risk level, and their spatiotemporal relationships from each sub-network. The spatiotemporal relationships include the time difference between event occurrences and the topological distance between nodes. The leaf nodes of the decision tree correspond to specific single event types or coupled event types; for example, one leaf node might represent "single construction disturbance," and another "construction-seepage coupling event." The coupled event decision tree is integrated with a risk superposition coefficient to form a preset coupled risk rule, which is applied in the comprehensive decision logic. Based on the preset coupled risk rule, the comprehensive decision logic determines whether the final event type is a single event or a coupled event and assigns a comprehensive risk level. For example, when the input is disturbance level 2, water immersion level 1, and the spatiotemporal relationship satisfies the "construction-seepage coupling" pattern, the decision tree identifies it as a coupled event, and the comprehensive risk level is determined. The calculation formula is: in: Disturbance level 2, It is classified as flood immersion level 1. With a risk factor of 1.5, the calculation yields... A table of frequent items mined from a historical event database, see Table 1: Table 1: Frequent Co-occurrence Event Combinations Mined from the Historical Event Database Optionally, when constructing the decision tree for coupled events, in addition to using the rank values, the original confidence probability output by each sub-network is also used as one of the input features to handle discrimination cases with ambiguous boundaries. Optionally, the risk superposition coefficient κ can be dynamically adjusted based on the actual maintenance costs and downtime caused by the coupled event patterns in historical records, assigning higher κ values ​​to patterns with more severe consequences. It is understandable that when mining frequent itemsets from the historical event database, minimum support and confidence thresholds need to be set to avoid generating invalid rules from incidentally co-occurring events. It is understandable that the topological distance in spatiotemporal relationships can be calculated based on the shortest path hop count in the pipeline network topology model, rather than purely physical straight-line distance, which better reflects the propagation path of influences within the pipeline network.

[0034] In one embodiment of the present invention, based on the graded response command, the intelligent warning base box is controlled to perform local response actions and upload event data and device status to the cloud management platform through a preset communication protocol. The specific process includes: when the graded response command is a Level 1 warning command, the intelligent warning base box stores the recorded data containing the event judgment result in a local cache, and packages and uploads it to the cloud management platform during the next preset regular communication window. When the graded response command is a Level 2 alarm command, the high-power communication unit of the intelligent warning base box is immediately activated, a high-priority communication link with the cloud management platform is established, alarm information is pushed, and the device's built-in warning light is activated to execute a slow flashing mode. When the graded response command is a Level 3 emergency reporting command, the communication channel is immediately seized, and an emergency reporting process containing real-time streaming data, precise device location information, and event judgment results is initiated to the cloud management platform, while simultaneously activating the high-frequency flashing of the warning light and the sounding of the buzzer. After completing the event data upload, the intelligent warning base box will receive a confirmation command or a new control command from the cloud management platform and adjust subsequent monitoring or response behaviors according to the control command.

[0035] Based on the uploaded event data, the digital model status of the corresponding pipeline section is updated on the cloud management platform, generating a visualized 3D situation map and asset status record. Specifically, the cloud management platform receives event data and equipment status information corresponding to the graded response instructions uploaded from the smart warning terminal, and parses the event location, type, level, and timestamp. In the 3D digital model of the power pipeline network maintained by the platform, the pipeline section and equipment node corresponding to the event location are located. Based on the parsed event type and level, the status attributes of the pipeline section corresponding to the event location in the 3D digital model are updated, and the pipeline section corresponding to the event location is rendered with a highlighted color, where the color maps to the comprehensive risk level. In the 3D digital model, the diffusion process of the state impact from the event location terminal node to adjacent nodes is dynamically displayed along the cable path in the form of flowing particles or pulsating light effects. All data related to the abnormal event corresponding to the currently received event data and equipment status information, including original monitoring information fragments, event judgment results, and response records, are bound to the pipeline asset code of the incident and the electronic identity of the smart warning terminal, and stored as a complete operation and maintenance record in the asset lifecycle database.

[0036] In practical implementation, the intelligent warning ground box is controlled to perform local response actions and upload event data according to the graded response instructions. For example, if intelligent warning ground boxes E, F, and G are deployed in a pipeline section, the state assessment model will generate graded response instructions after analyzing the monitoring data of intelligent warning ground box E. The graded response instructions include the event discrimination result "construction disturbance - level 2" and the instruction level. When the graded response instruction is a level 1 warning instruction, the intelligent warning ground box E is controlled to store the recorded data containing the event discrimination result in the local flash memory cache. The recorded data includes timestamps, vibration characteristic segments, and event types. It is then packaged and uploaded to the cloud management platform in the next preset regular communication window period. The regular communication window period is opened once every 6 hours, and the upload method is low-power narrowband IoT communication. When the tiered response command is a Level 2 alarm command, the high-power communication unit of the intelligent warning base box F is immediately activated. The high-power communication unit is a 4G communication module, which establishes a high-priority communication link with the cloud management platform. The communication link uses a dedicated APN access line and pushes alarm information including the event location, type "water inrush - Level 3", and preliminary assessment results. At the same time, the warning light built into the intelligent warning base box F is activated to execute a slow flashing mode with a flashing frequency of 1 Hz. When the tiered response command is a Level 3 emergency reporting command, the intelligent warning base box G immediately seizes the communication channel and initiates an emergency reporting process to the cloud management platform using the highest transmission power. The reported data includes real-time vibration and current data for the last 5 minutes, precise device positioning information provided by the global satellite navigation system positioning module, and the event judgment result "external breakdown - Level 5". Simultaneously, the high-frequency strong light flashing mode of the warning light and the buzzer sounding are activated, with a flashing frequency of 10 Hz and a buzzer sound pressure level of 90 decibels. After the event data upload is completed, the smart warning box will receive confirmation instructions or new control instructions from the cloud management platform. For example, smart warning box F receives a confirmation instruction that "instruction received", and smart warning box G receives a new control instruction to "increase the monitoring frequency to 1 time / second". The smart warning box adjusts its subsequent monitoring or response behavior according to the control instructions.

[0037] In some embodiments, the digital model status of the corresponding pipeline segment is updated on the cloud management platform based on the uploaded event data. The cloud management platform receives event data and equipment status information corresponding to the level 2 alarm command uploaded from the intelligent warning box F, and parses the event location as coordinates (Xf, Yf), type as "water inrush", level as "3", and timestamp as "2023-10-26 14:30:05". In the three-dimensional digital model of the power pipeline network maintained by the platform, the pipeline segment and equipment node corresponding to the event location are located according to the coordinates (Xf, Yf). The segment is identified as "Tunnel-Segment-07F" in the model, and the equipment node is identified as "E-BOX-F". Based on the analyzed event type "water inrush" and level "3", the status attribute of the "Tunnel-Segment-07F" pipeline section in the 3D digital model is updated, changing the status from "normal" to "water immersion - level 3". The "Tunnel-Segment-07F" pipeline section is rendered with a highlighted color, mapping the overall risk level, with level 3 corresponding to yellow highlighting. In the 3D digital model, the diffusion process of the state impact from the event-occurring smart warning box F to adjacent smart warning boxes E and G is dynamically displayed along the cable path in the form of flowing particles or pulsating light effects. The particle flow speed simulates the water diffusion speed. All data related to the abnormal event corresponding to the currently received event data and equipment status information, including original monitoring information fragments, event judgment results, and response records, are bound to the pipeline asset code "ASSET-2023-07F" where the event occurred and the electronic identity "E-BOX-F-ID" of smart warning box F, and stored as a complete operation and maintenance record in the asset lifecycle database. A hierarchical response command execution rule is shown in Table 2: Table 2: Execution Rules for Hierarchical Response Commands Optionally, under a level 2 alarm command, if a high-priority communication link fails to be established, the intelligent warning ground box can reconnect according to a preset retry algorithm, with a retry interval of [time missing]. The calculation formula is: in: This is the base interval time. This is the current number of retries. This is the maximum permissible interval time. Optionally, when the cloud management platform displays the diffusion process of state impact in the 3D digital model, the diffusion speed and range can be parameterized according to the event level and type. For example, the diffusion speed of a water inrush event can be configured to 0.1 m / s. It can be understood that when maintenance records are stored in the asset's full lifecycle database, the records will be associated with the asset's complete set of historical maintenance and inspection records to form a complete asset archive. It can be understood that after event data is uploaded, new control commands issued by the cloud management platform could be to adjust the sensor sampling frequency of the smart warning box or its adjacent boxes.

[0038] In one embodiment of the present invention, the dynamic adjacency matrix and the initial set of node features are input into the improved graph convolutional neural network algorithm. The algorithm updates the feature representation of each node through a multi-level information transfer and aggregation mechanism. This process includes: inputting the dynamic adjacency matrix and the initial set of node features for each node into the first hidden layer of the improved graph convolutional neural network. In the first hidden layer, for the current target node, its first-order physical neighbor node set is determined according to the dynamic adjacency matrix, and the initial set of node features of all nodes in the first-order physical neighbor node set is aggregated to generate a first-level aggregated feature. The initial set of node features of the target node and the first-level aggregated feature are subjected to nonlinear transformation and feature fusion to output the intermediate feature representation of the target node in the first hidden layer. The intermediate feature representations of all nodes output from the first hidden layer, together with the dynamic adjacency matrix, are input into the second hidden layer of the improved graph convolutional neural network. In the second hidden layer, for the target node, its extended set of neighboring nodes is determined according to the dynamic adjacency matrix. This extended set includes its first-order physical neighbors and related nodes reachable within a preset topology hop count via cable connections. The intermediate feature representations of all nodes in the extended set are aggregated to generate a second-level aggregated feature. A nonlinear transformation and feature fusion are performed on the intermediate feature representation of the target node from the first hidden layer and the second-level aggregated feature, outputting the updated state feature vector of the target node in the second hidden layer. This multi-level aggregation and fusion process is iteratively executed until a preset network depth is reached. The output state feature vector of each ground-level node is used as the comprehensive state feature vector of the network node.

[0039] In practical implementation, the dynamic adjacency matrix and the initial set of node features are input into an improved graph convolutional neural network algorithm. This algorithm updates the feature representation of each node through a multi-level information transmission and aggregation mechanism. Taking a small pipeline network topology containing smart warning box nodes M, N, O, and P as an example, the dynamic adjacency matrix is ​​a 4x4 matrix, and the initial set of node features contains four feature vectors. Each feature vector is a floating-point array of length 20, representing the initial features extracted from multi-source monitoring information. The dynamic adjacency matrix and the initial set of node features for each node are input into the first hidden layer of the improved graph convolutional neural network. In the first hidden layer, for the current target node, smart warning box node N, its first-order physical neighbor node set is determined according to the dynamic adjacency matrix. The dynamic adjacency matrix shows that smart warning box node N is directly connected to smart warning box nodes M and O; therefore, its first-order physical neighbor node set is {smart warning box node M, smart warning box node O}. An initial set of node features from all nodes in the first-order physical neighbor set is aggregated. The aggregation operation is a weighted average of the feature vectors, with the weights determined by the corresponding edge weights in the dynamic adjacency matrix, generating a first-level aggregated feature. This first-level aggregated feature is a vector with the same dimension as the initial set of node features. A nonlinear transformation and feature fusion are then performed on the initial set of node features and the first-level aggregated feature of the target node, the smart warning box node N. This outputs an intermediate feature representation of the smart warning box node N in the first hidden layer. The intermediate feature representation is a new feature vector of length 32. The nonlinear transformation is implemented through an activation function, and the feature fusion combines the two input features using a learnable weight matrix. The intermediate feature representations of all nodes output from the first hidden layer, along with the dynamic adjacency matrix, are then input into the second hidden layer of the improved graph convolutional neural network.

[0040] In some embodiments, in the second hidden layer, for the target node smart warning box node N, its extended neighboring node set is determined according to the dynamic adjacency matrix. The preset topology hop count is 2. The extended neighboring node set includes the first-order physical neighbors of smart warning box node N: smart warning box node M and smart warning box node O, as well as related nodes reachable within 2 hops via cable connections, specifically smart warning box node P connected through smart warning box node O. Therefore, the extended neighboring node set is {smart warning box node M, smart warning box node O, smart warning box node P}. The intermediate feature representations of all nodes in the extended neighboring node set are aggregated. The aggregation operation can be represented as: in: It is a second-level aggregation feature. It is the edge weight between node N and node j in the dynamic adjacency matrix. The intermediate feature representation of node j in the first hidden layer is used to generate the second-level aggregated feature. This is the intermediate feature representation of the target node N from the first hidden layer. Second-level aggregation features Perform nonlinear transformation and feature fusion to output the updated state feature vector of the intelligent warning box node N in the second hidden layer. The fusion formula is: in: It is an activation function. It is the trainable weight matrix of the second hidden layer. This is a bias term, and CONCAT represents a vector concatenation operation. The process iteratively executes multi-level aggregation and fusion until a preset network depth is reached. The preset network depth is 3, meaning there is a third hidden layer after the second hidden layer. The final output state feature vector of each ground box node serves as the comprehensive state feature vector of the network nodes. For example, the final comprehensive state feature vector of the intelligent warning ground box node N is a vector of length 64.

[0041] In some embodiments, the preset topology hop count can be adjusted based on the diameter of the pipeline topology or specific monitoring needs. For pipeline areas with complex structures and many branches, the preset topology hop count can be set to 3 to capture a wider range of correlations. Optionally, when performing feature aggregation in the first or second hidden layer, in addition to using weighted averaging, an attention mechanism can be used to calculate the weights of features of neighboring nodes, making the aggregation more focused on neighboring nodes with abnormal states or higher correlations. Optionally, the activation function used in the nonlinear transformation can be the ReLU function or the LeakyReLU function, used to introduce nonlinearity and alleviate the gradient vanishing problem. It is understood that the network depth setting needs to strike a balance between model expressive power and excessive smoothing; excessive depth may lead to excessive propagation of information from distant, irrelevant nodes. It is understood that each time information is aggregated, the dynamic adjacency matrix provides a real-time measure of the connection strength between nodes, which allows the weights of information transmission to be dynamically adjusted according to the pipeline's operating state, rather than remaining fixed.

[0042] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for monitoring the condition of the power grid of a smart warning ground box, characterized by, The method includes the following steps: Deploy intelligent warning boxes at key nodes along the power grid to complete equipment power supply and network initialization, and establish a monitoring network; During the routine monitoring cycle, the intelligent warning box synchronously acquires multi-source status monitoring information of the underground pipe network. The multi-source status monitoring information includes pipe network structure vibration information, groundwater level and water accumulation information, and cable sheath grounding current information. Based on the improved graph convolutional neural network algorithm, feature extraction and spatiotemporal correlation analysis are performed on the multi-source state monitoring information to generate a comprehensive state feature vector of the pipeline network nodes. The improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network. The comprehensive status feature vector of the pipeline node is input into the status assessment model to identify abnormal event types and determine risk levels, and generate event discrimination results and corresponding graded response instructions. Based on the hierarchical response instructions, the intelligent warning box is controlled to perform local response actions and upload event data and device status to the cloud management platform through a preset communication protocol; Based on the uploaded event data, the digital model status of the corresponding pipeline section is updated on the cloud management platform, generating a visualized 3D situation map and asset status record.

2. The method of claim 1, wherein the method further comprises: The improved graph convolutional neural network algorithm extracts features and performs spatiotemporal correlation analysis on the multi-source status monitoring information to generate a comprehensive status feature vector for the pipeline network nodes, including: Based on the deployment location of the intelligent warning ground box and the cable connection relationship, a pipeline topology model is constructed with the ground box as the node and the cable connection path as the edge; Initial features in the time domain, frequency domain, and spatial distribution dimensions are extracted from the vibration information of the pipeline structure, groundwater level and water accumulation information, and cable sheath grounding current information to form an initial set of node features for each ground box node. The pipeline topology model is transformed into an adjacency matrix, and the dynamic connection weight of the adjacent edges is calculated based on the physical distance between the grounding nodes, the cable type and the laying environment to generate a dynamic adjacency matrix. The dynamic adjacency matrix and the initial set of node features are input into the improved graph convolutional neural network algorithm. The improved graph convolutional neural network algorithm updates the feature representation of each node through a multi-level information transmission and aggregation mechanism. During each information aggregation process, the status information of the relevant grounding nodes within a specified number of hops on the physically adjacent grounding nodes and the cable connection path is aggregated. After iterating through the information transfer and aggregation a specified number of times, the updated state feature vector of each ground node is output. This vector integrates the correlation information of the local and neighboring areas and serves as the comprehensive state feature vector of the pipeline node.

3. The method of claim 2, wherein the method further comprises: The improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network, and adopts a method that integrates the correlation between physical attributes and real-time status. Its working principle includes the power pipeline network status monitoring method for intelligent warning boxes as described in claim 2, characterized in that the improved graph convolutional neural network algorithm dynamically constructs an adjacency matrix based on the physical topology of the pipeline network, and its working principle includes: Define the initial physical connection diagram of the pipeline network, with the location of each smart warning box as a node and the cable connection path as an edge; Obtain the attribute information of the cable connection path, including cable type, laying depth, and current cable load rate, and calculate the initial connection weight of each edge; The system acquires the current monitoring information uploaded by each ground cell node in real time, and calculates the spatiotemporal correlation of the states between ground cell nodes based on the monitoring information. The spatiotemporal correlation includes vibration propagation consistency, water level change synchronization and leakage current correlation. The initial connection weights based on physical connections are weighted and fused with the spatiotemporal correlations based on real-time states to generate dynamic edge weights that reflect the strength of physical connections and state correlations at the current moment. Based on the dynamic edge weights between all nodes, the dynamic adjacency matrix is ​​constructed and updated in real time.

4. The method of claim 1, wherein the method further comprises: The comprehensive state feature vector of the pipeline node is input into the state assessment model to identify abnormal event types and determine risk levels, including: The comprehensive state feature vector of the network node is input into an evaluation model containing multiple parallel sub-networks; One of the subnetworks analyzes the dimensions representing vibration information in the feature vectors, and by comparing them with a historical normal vibration pattern library, it identifies disturbance events of the types of construction disturbance, geological settlement, and vehicle rolling, and outputs the disturbance level. Another sub-network analyzes the dimensions representing water immersion information in the feature vector, and combines the rate of change and duration of water immersion to identify water immersion events of the types of slow seepage, gushing water, and internal flooding, and outputs the water immersion level. Another subnetwork analyzes the dimensions representing leakage information in the feature vector, and combines the leakage current amplitude, fluctuation characteristics and harmonic content to identify electrical events of insulation aging, partial discharge and external breakdown types, and outputs the electrical risk level. The disturbance level, water immersion level, and electrical risk level are input into a comprehensive decision logic. The comprehensive decision logic determines whether the final event type is a single event or a coupled event based on preset coupled risk rules, and determines a comprehensive risk level.

5. The method of claim 4, wherein the method further comprises: Based on the tiered response instructions, the intelligent warning box is controlled to perform local response actions and upload event data and device status to the cloud management platform via a preset communication protocol, including: When the graded response instruction is a level one warning instruction, the control intelligent warning box will store the recorded data containing the event discrimination result into the local cache, and package and upload it to the cloud management platform in the next preset regular communication window period; When the graded response command is a level 2 alarm command, the high-power communication unit of the intelligent warning box is immediately activated, a high-priority communication link is established with the cloud management platform, alarm information is pushed, and the device's built-in warning light is activated to perform a slow flashing mode. When the graded response instruction is a level three emergency reporting instruction, the communication channel is immediately seized, and an emergency reporting process containing real-time streaming data, precise device location information and event judgment results is initiated to the cloud management platform. At the same time, the high-frequency flashing of the warning light and the sounding of the buzzer are activated. After the event data is uploaded, the smart warning box will receive confirmation instructions or new control instructions from the cloud management platform, and adjust its subsequent monitoring or response behavior according to the control instructions.

6. The method of claim 5, wherein the method further comprises: Based on the uploaded event data, the digital model status of the corresponding pipeline section is updated on the cloud management platform, generating a visualized 3D situation map and asset status records, including: The cloud management platform receives event data and device status information corresponding to the graded response instructions uploaded from the smart warning box, and parses out the location, type, grade and timestamp of the event. In the 3D digital model of the power network maintained by the platform, the network section and equipment node corresponding to the location of the event are located. Based on the analyzed event type and level, update the status attributes of the pipeline section corresponding to the event location in the 3D digital model, and render the pipeline section corresponding to the event location with a highlighted color, wherein the color maps the comprehensive risk level; In the three-dimensional digital model, the diffusion process of state influence from the event occurrence node to adjacent nodes is dynamically displayed along the cable path in the form of flowing particles or pulsating light effects. All data related to the abnormal events corresponding to the currently received event data and equipment status information, including original monitoring information fragments, event judgment results, and response records, are bound to the pipeline asset code and electronic identity of the smart warning box where the event occurred, and stored as a complete operation and maintenance record in the asset lifecycle database.

7. The method of claim 2, wherein the method further comprises: Based on the deployment location of the intelligent warning junction box and its cable connection relationship, a pipeline topology model is constructed with the junction box as nodes and the cable connection path as edges, including the following steps: Obtain the as-built drawings or digital design drawings of the power network in the target monitoring area from the cloud management platform, and extract the preset installation location coordinates of all smart warning ground boxes and the preset cable laying path information from the drawings; After the smart warning ground box completes physical installation and power supply initialization, the actual geographical coordinates of each ground box are obtained through the global satellite navigation system positioning module, and the device is bound to the preset installation location coordinates through the device's unique identification code to complete the digital registration of the deployment location; Based on the preset laying path information of the cable, determine whether there is a direct cable connection between any two smart warning boxes, and define the two boxes with a direct cable connection as adjacent nodes. In the pipeline topology model, each smart warning box that has completed digital registration is a graph node, and the actual cable laying path between any two adjacent nodes is an undirected edge or a directed edge, wherein the direction of the edge is defined according to the power transmission direction of the cable or the preset monitoring information flow direction. The attributes of each node in the pipeline topology model are initialized to its corresponding unique device identifier and actual geographical coordinates, and the attributes of each edge are initialized to the identifier, type and length of the cable it represents.

8. The method for monitoring the status of power grid using an intelligent warning grounding box according to claim 2, characterized in that, The dynamic adjacency matrix and the initial set of node features are input into the improved graph convolutional neural network algorithm. The improved graph convolutional neural network algorithm updates the feature representation of each node through a multi-level information transfer and aggregation mechanism, including: The dynamic adjacency matrix and the initial set of node features for each node are input into the first hidden layer of the improved graph convolutional neural network; In the first hidden layer, for the current target node, its first-order physical neighbor node set is determined according to the dynamic adjacency matrix, and the initial set of node features of all nodes in the first-order physical neighbor node set is aggregated to generate the first-level aggregated features. The initial set of node features of the target node and the first-level aggregated features are subjected to nonlinear transformation and feature fusion to output the intermediate feature representation of the target node in the first hidden layer; The intermediate feature representations of all nodes output from the first hidden layer, along with the dynamic adjacency matrix, are input into the second hidden layer of the improved graph convolutional neural network. In the second hidden layer, for the target node, its extended set of neighboring nodes is determined according to the dynamic adjacency matrix. The extended set of neighboring nodes includes its first-order physical neighboring nodes and related nodes that can be reached within a preset number of topological hops through cable connection paths. The intermediate feature representations of all nodes in the extended set of neighboring nodes are aggregated to generate a second-level aggregated feature. The intermediate feature representations and second-level aggregated features of the target node from the first hidden layer are subjected to nonlinear transformation and feature fusion to output the updated state feature vector of the target node in the second hidden layer. The multi-level aggregation and fusion process is iteratively executed until the preset network depth is reached, and the final output state feature vector of each ground node is used as the comprehensive state feature vector of the pipeline node.

9. The method for monitoring the status of power grid using an intelligent warning grounding box according to claim 3, characterized in that, The acquisition of cable connection path attribute information, including cable type, laying depth, and current cable load rate, and the calculation of initial connection weight for each edge, includes: From the power grid asset database, query the cable type, insulation material and design life corresponding to the cable connection path, and assign a basic weight coefficient that reflects the importance of its physical connection to different cable types; Acquire cable laying depth data, compare the laying depth with the preset standard depth, and calculate a depth correction coefficient that reflects the differences in the laying environment based on the degree of depth deviation. The system receives cable load rate data from the power system in real time and calculates a load dynamic coefficient that reflects the real-time operating pressure based on the ratio of the current load rate to the rated load. The initial connection weight of the cable connection path is obtained by weighting the basic weight coefficient, the depth correction coefficient and the load dynamic coefficient, wherein the basic weight coefficient has the dominant weight.

10. The method for monitoring the status of power grid using an intelligent warning grounding box according to claim 4, characterized in that, The comprehensive decision logic determines whether the final event type is a single event or a coupled event based on preset coupling risk rules. The generation of the coupling risk rules includes: Frequently occurring anomalous event combinations are extracted from the historical event database, including combinations of vibration and water immersion, and water immersion and electrical leakage. For each frequently co-occurring combination of abnormal events, we analyze their order of occurrence, time intervals, and spatial relationships to define a coupled event pattern. For each defined coupled event pattern, a risk superposition coefficient is set, which is greater than the simple sum of the risk levels of individual events; Establish a decision tree for coupled events. The input of the decision tree is the disturbance level, water immersion level, electrical risk level and their spatiotemporal relationship output by each sub-network. The leaf nodes of the decision tree correspond to specific single event types or coupled event types. The coupled event discrimination decision tree is integrated with the risk superposition coefficient to form the preset coupled risk rule.

Images