Metering device fault early warning method and device, equipment and medium

Abstract

The invention provides a metering device fault early warning method and device, equipment and a medium, and belongs to the technical field of intelligent power grids, and the method comprises the steps: carrying out the risk grade marking of each metering device according to the multi-modal data and historical meteorological disasters corresponding to a plurality of metering devices, and determining an initial sample set corresponding to each metering device; for each metering device, executing the following early warning operations: determining a physical topological graph according to one metering device and the metering device associated with the metering device; determining time sequence data of the metering device according to the initial sample set, and determining a semantic object based on the time sequence data, the physical topological graph and the graph convolutional neural network so as to determine a semantic structure graph; and obtaining a fault diagnosis result and early warning information of the metering device through a multi-modal fusion model according to the physical topological graph and the semantic structure graph. According to the metering device fault early warning method and device, the equipment and the medium provided by the invention, the accuracy of fault diagnosis can be improved.

Description

A method, device, equipment, and medium for early warning of metering device failures Technical Field

[0001] This application belongs to the field of smart grid technology, and more specifically, relates to a method, device, equipment, and medium for early warning of metering device faults. Background Technology

[0002] With the accelerated advancement of energy transition and the construction of new power systems, smart metering devices (such as smart meters) have been deployed on a large scale. Simultaneously, intermittent and nonlinear diverse loads, such as energy storage systems and electric vehicle charging stations, are widely connected to the power grid. As a core link connecting power grid companies and users, enabling accurate trade settlement and lean power grid management, the real-time performance and accuracy of electricity metering data directly affect the fairness and impartiality of electricity billing and the safe and efficient operation of the power grid.

[0003] Currently, fault diagnosis and maintenance of power metering equipment mainly rely on the following methods: First, threshold alarms based on the equipment's own operating parameters (e.g., voltage, current, power consumption) are used to manually identify out-of-limit data through manually set rules; second, periodic manual on-site inspections and meter readings are used to visually inspect the equipment's appearance and wiring. These traditional methods have gradually revealed the following limitations in practical applications: First, the diagnostic dimensions are singular, ignoring the coupling effects of the environment. Most existing methods only focus on the electrical parameters of the metering device itself, failing to incorporate external risk factors such as its geographical location and the surrounding micro-meteorological environment (e.g., potential exposure to historical disasters like rainstorms, high temperatures, and typhoons) into a comprehensive analysis framework. This makes it difficult to detect latent faults induced by the environment in advance. Second, data correlation is fragmented. Existing methods treat each metering device as an independent unit, failing to effectively model and utilize the physical connections between metering devices formed through communication networks and power lines. They also lack in-depth mining of the patterns (semantic correlations) exhibited by user electricity consumption behavior in time series and group correlations, resulting in the underutilization of the complementary value and correlations between multi-source heterogeneous data. Finally, relying on fixed rules and manual intervention results in insufficient adaptability and real-time performance. Threshold-based judgments struggle to capture complex and dynamically changing fault modes, while manual inspection methods are costly and time-consuming, unable to provide real-time responses and early warnings to environmental changes and potential fault risks. Overall, these methods suffer from problems such as delayed fault identification and a high rate of false positives and false negatives.

[0004] Therefore, there is an urgent need to conduct research on methods for accurate identification and early warning of metering equipment faults based on multidimensional data analysis, so as to provide a guarantee for the efficient operation of the power system. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, equipment, and medium for early warning of metering device faults, which integrates physical topology maps and semantic topology maps to improve the accuracy of fault diagnosis, reduce the false negative rate, and accurately identify potential risks of metering devices.

[0006] A first aspect of this application provides a method for early warning of metering device malfunctions, including:

[0007] Acquire multimodal data corresponding to multiple metering devices, label the metering devices with risk levels based on the multimodal data and historical meteorological disasters, and determine the initial sample set corresponding to each metering device based on the risk level labeling results. The multimodal data includes geographical location, operating status electrical parameters, power consumption, and real-time images.

[0008] For each metering device, perform the following warning operation:

[0009] Based on the metering device and other metering devices that have communication and power transmission connections with it, determine the physical topology.

[0010] The timing data of the metering device is determined based on the initial sample set. Based on the timing data and physical topology of the metering device, semantic objects are determined through graph convolutional neural networks. Connection edges of semantic objects are generated according to preset rules. Semantic structure graph is determined based on the connection edges of semantic objects.

[0011] Based on the physical topology graph and semantic structure graph, and through a multimodal fusion model based on an attention mechanism, the physical topology graph and semantic structure graph are fused to obtain the fault diagnosis results and early warning information of the metering device.

[0012] In one embodiment, a physical topology map is determined based on the metering device and metering devices that have communication and power transmission associations with the metering device, including:

[0013] Determine the location of the nodes on the graph based on the coordinate information of the metering device;

[0014] Based on the metering devices and the metering devices associated with the communication links and transmission lines, determine the connection edges between each node.

[0015] Node attributes are determined based on the appearance of the metering device, coordinate information, and disaster labels.

[0016] Construct a physical topology graph containing spatial location, connection relationships, and electrical characteristics based on the node locations, connecting edges, and node attributes.

[0017] In one implementation, determining node attributes based on the appearance of the metering device, coordinate information, and disaster tags, further includes:

[0018] Feature extraction is performed on the real-shot images to obtain appearance state feature vectors, which are used to characterize the appearance state of the measuring device.

[0019] During the data acquisition phase, the coordinate information corresponding to the metering device is obtained;

[0020] Obtain historical disaster data;

[0021] The disaster label of the metering device is determined based on historical disaster data and the corresponding geographical coordinates of the historical disaster data;

[0022] Among them, node attributes are determined based on the appearance, coordinate information, and disaster labels of the metering device, including:

[0023] The node attributes of the metering device are determined based on the appearance status feature vector, the coordinate information of the metering device, and the disaster label of the metering device.

[0024] In one implementation, based on the time-series data and physical topology of the metering device, semantic objects are determined using a graph convolutional neural network, and connection edges of the semantic objects are generated according to preset rules. A semantic structure graph is then determined based on these connection edges, including:

[0025] The time series data of the metering device is preprocessed to obtain preprocessed time series data;

[0026] The preprocessed time series data is input into the physical topology graph to obtain a physical topology graph that incorporates dynamic features;

[0027] Semantic features are obtained by combining a physical topology graph with dynamic features and a graph neural convolutional network, and semantic objects are determined based on the semantic features.

[0028] The semantic object is generated by generating connection edges based on the semantic object, temporal proximity edges, and co-occurrence edges. Temporal proximity edges are used to represent events that are closely continuous in time and have causal or triggering relationships. Co-occurrence edges are used to represent events that occur in the same time period under common influence.

[0029] The edge set of the semantic graph is determined based on the connecting edges of the semantic objects. The edge set of the semantic graph represents the set of all connecting edges.

[0030] The node set of the semantic graph is determined based on the semantic object and semantic features;

[0031] The semantic structure graph is determined based on the edge set and node set of the semantic graph.

[0032] In one implementation, semantic features are obtained based on a physical topology graph that incorporates dynamic features and a graph neural convolutional network, including:

[0033] The physical topology graph incorporating dynamic features is input into the graph neural network. Based on the multiple convolutional layers of the graph neural network, the aggregated feature information between each node in the physical topology graph incorporating dynamic features and its neighboring nodes is determined.

[0034] The aggregated feature information is input into the fully connected layer to obtain the corresponding semantic features.

[0035] In one implementation, based on a physical topology graph and a semantic structure graph, and through a multimodal fusion model based on an attention mechanism, the physical topology graph and the semantic structure graph are fused to obtain fault diagnosis results and early warning information for the metering device, including:

[0036] A graph isomorphic network is used to extract feature vectors from the physical topology graph and the semantic structure graph to obtain physical feature vectors and semantic feature vectors.

[0037] The physical feature vector and the semantic feature vector are input into a multimodal fusion model with an attention mechanism to obtain the fused feature vector;

[0038] The fault diagnosis results and early warning information corresponding to the fused feature vectors are determined based on the fused feature vectors and the preset activation function.

[0039] In one implementation, it further includes:

[0040] Based on the fault diagnosis results and early warning information, a field verification message is sent to the terminal of the corresponding maintenance personnel so that the maintenance personnel can conduct on-site verification of the metering device.

[0041] Obtain the verification results and determine the warning level and the corresponding deployment and prevention measures based on the verification results.

[0042] A second aspect of this application provides a metering device fault early warning device, comprising:

[0043] The sample determination module is used to acquire multimodal data corresponding to multiple metering devices, and to mark the risk level of the metering devices based on the multimodal data and historical meteorological disasters. Based on the risk level marking results, the initial sample set corresponding to each metering device is determined. The multimodal data includes geographical location, operating status electrical parameters, power consumption and real-time images.

[0044] The physical topology determination module is used to determine the physical topology based on the metering device and other metering devices that have communication and power transmission connections with the metering device.

[0045] The semantic structure graph determination module is used to determine the time series data of the metering device based on the initial sample set, determine the semantic objects based on the time series data and physical topology graph of the metering device, and generate the connection edges of the semantic objects according to preset rules, and determine the semantic structure graph based on the connection edges of the semantic objects.

[0046] The early warning information determination module is used to obtain the fault diagnosis results and early warning information of the metering device by fusing the physical topology graph and semantic structure graph through a multimodal fusion model based on an attention mechanism.

[0047] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described metering device fault early warning method.

[0048] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described metering device fault warning method.

[0049] The beneficial effects of the metering device fault early warning method, device, equipment, and medium provided in this application are as follows: This application's embodiments acquire multimodal data such as the metering device's geographical location, operating status electrical parameters, electricity consumption, and real-time images, and label the risk level based on this multimodal data and historical meteorological disasters, thus determining a more representative initial sample set; by determining a physical topology map based on the metering device and metering devices with which it has communication and power transmission relationships, it can clearly present the physical connection and communication relationships between metering devices, helping to understand the operating environment of the metering device and discover potential faults; this application's embodiments determine the time-series data of the metering device based on a more representative initial sample set, and use a graph convolutional neural network to determine semantic objects and generate connection edges, thereby constructing a semantic structure graph to mine the hidden semantic relationships in the multimodal data, providing a more in-depth basis for the fault diagnosis of the metering device; this application's embodiments, based on an attention mechanism-based multimodal fusion model, fuse the physical topology map and the semantic structure graph, comprehensively considering the physical relationships and semantic features of the metering device, thereby obtaining more comprehensive and accurate fault diagnosis and early warning information. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 is a flowchart illustrating a metering device fault early warning method provided in an embodiment of this application;

[0052] Figure 2 is a flowchart illustrating a physical topology graph construction method provided in an embodiment of this application;

[0053] Figure 3 is a flowchart illustrating a semantic structure graph construction method provided in an embodiment of this application;

[0054] Figure 4 is a schematic diagram of the process of multimodal data fusion based on attention mechanism provided in an embodiment of this application;

[0055] Figure 5 is another flowchart illustrating a metering device fault early warning method provided in an embodiment of this application;

[0056] Figure 6 is a schematic diagram comparing the ablation methods corresponding to different fault diagnosis methods provided in an embodiment of this application;

[0057] Figure 7 is a structural block diagram of a metering device fault early warning device provided in an embodiment of this application;

[0058] Figure 8 is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0059] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0061] Please refer to Figure 1, which is a flowchart illustrating a metering device fault early warning method according to an embodiment of this application. The method may include:

[0062] S101: Acquire multimodal data corresponding to multiple metering devices, label the metering devices with risk levels based on the multimodal data and historical meteorological disasters, and determine the initial sample set corresponding to each metering device based on the risk level labeling results.

[0063] In this embodiment, the multimodal data includes geographical location, operating electrical parameters (e.g., current, voltage), power consumption, and real-world images (transformer).

[0064] In this embodiment, the metering device that performs fault early warning is regarded as the key metering device.

[0065] In this embodiment, multimodal data from electricity consumption information collection terminals and marketing business application terminals are collected and integrated. Combined with historical disaster and environmental data obtained from the network, the metering devices are labeled with risk levels. The collected sample data is then preprocessed to form an initial sample set. The electricity consumption information collection terminal can be an electricity consumption information collection system, and the marketing business application terminal can be a marketing business application system.

[0066] For example, this embodiment can combine BeiDou satellite positioning to obtain the geographical coordinates (e.g., latitude and longitude) of the metering device through the marketing system asset number (e.g., meter box ID). For metering devices in multi-story or underground spaces, modern engineering surveying technology (e.g., GPS-RTK) can be used to obtain their three-dimensional coordinate data, and three-dimensional positioning labels can be applied to key metering devices: Longitude, Latitude, and Altitude. Based on these three-dimensional positioning labels, the geographical location coordinate information of key metering devices can be tagged.

[0067] This embodiment acquires power data, including voltage, current, power consumption, and frequency, as well as non-power data such as metering device status and environmental data, based on an electricity information collection system. The physical appearance of the metering device is determined using high-definition real-shot images of the device obtained by automated inspection equipment or monitoring and capturing equipment; these real-shot images characterize the physical appearance of the metering device.

[0068] This embodiment can crawl abnormal environmental information of metering devices from the network, i.e., historical disaster data. Geocoding technology can be used to convert the addresses in the historical disaster data into geographic coordinates. The coordinate data of the metering device is then overlaid and analyzed with vector data of various disasters, and a proximity analysis is performed. Risk labels are then calculated, including disaster type, historical disaster frequency, distance to the disaster point, and disaster risk index. Disaster types include floods and fires; the historical disaster frequency refers to the number of times various disasters have occurred within a preset radius (e.g., 500 meters) of the coordinate data of the metering device in the past three years; the distance to the disaster point refers to the distance between the metering device and the nearest historical disaster point (e.g., the geographical area covered by a flood and the resulting power outage of nearby metering devices); the disaster risk index, as a comprehensive indicator, can be calculated using a weighted scoring formula: Risk Index = Σ(Disaster Type Weight × Disaster Intensity Weight × Distance Attenuation Coefficient).

[0069] In this embodiment, the disaster type weights can be set using an expert scoring method (Delphi method). For example, experts can independently compare the relative importance of various disaster types pairwise. All expert scores are collected and statistically analyzed (e.g., calculating the average). The analysis results are fed back to the experts for multiple rounds of anonymous discussion and revision until expert opinions converge. Finally, the weights of each disaster type are normalized. The disaster intensity weight can be set according to national or industry standards, dividing each disaster into several levels, and then assigning an exponentially increasing weight value to each level; generally, the higher the intensity, the greater the weight. The distance attenuation coefficient can be set based on the disaster impact. In this embodiment, the disaster impact decreases as the distance from the disaster source increases; therefore, an exponential attenuation coefficient e is set. -s d Where d is the actual distance from the assessment point to the disaster source; s is the attenuation constant. The larger the value of the attenuation constant, the faster the attenuation rate. The attenuation constant can be obtained by fitting historical impact range data or given by expert experience.

[0070] By calculating the aforementioned labels, risk labels such as "disaster type," "historical number of disasters," "distance from disaster point," and "disaster risk index" can be assigned to key metering devices in disaster-prone areas. Based on the various data from these metering devices, the data is preprocessed to generate an initial sample set.

[0071] S102: Determine the physical topology based on the metering device and the metering devices that have communication and power transmission connections with the metering device.

[0072] In this embodiment, the key metering device (the metering device currently performing fault early warning) is used as a node in the physical topology graph. Connection edges are established based on the communication and power transmission association between the current metering device and other metering devices. The physical topology graph structure is established using the appearance status of the metering device, device coordinate information, and device disaster label as node attributes.

[0073] In one implementation, a physical topology map is determined based on the current metering device and metering devices that have communication and power transmission associations with the current metering device, including:

[0074] Determine the location of the graph node based on the coordinate information of the current metering device;

[0075] Based on the current metering device and the metering devices associated with the communication links and transmission lines that exist between the current metering device, determine the connection edges between each node;

[0076] Determine node attributes based on the current appearance of the metering device, coordinate information, and disaster labels;

[0077] Construct a physical topology graph containing spatial location, connection relationships, and electrical characteristics based on the node locations, connecting edges, and node attributes.

[0078] In one implementation, determining node attributes based on the appearance of the metering device, coordinate information, and disaster tags, further includes:

[0079] Feature extraction is performed on the real-shot image to obtain an appearance state feature vector, which is used to characterize the current appearance state of the metering device.

[0080] During the data acquisition phase, obtain the coordinate information corresponding to the current metering device;

[0081] Obtain historical disaster data;

[0082] The disaster label of the metering device is determined based on historical disaster data and the corresponding geographical coordinates of the historical disaster data;

[0083] Among them, node attributes are determined based on the appearance, coordinate information, and disaster labels of the metering device, including:

[0084] The node attributes of the metering device are determined based on the appearance status feature vector, the coordinate information of the metering device, and the disaster label of the metering device.

[0085] In this embodiment, the coordinate information of the current metering device is used as a graph node in the physical topology map. The physical connection relationships between nodes based on communication link association and transmission line association are used as connection edges. The current appearance status, coordinate information, and disaster tags of the metering device are used as node attributes.

[0086] This embodiment extracts features from real-world images to obtain a high-dimensional feature vector, which characterizes the current appearance of the metering device. The coordinate data of each metering device acquired during the data acquisition phase is used as its location feature (coordinate information). The calculated disaster label is used as the disaster attribute feature of the current metering device. All node attribute features of the metering device are aggregated to form its attribute feature vector, which characterizes the state of the metering device.

[0087] This embodiment constructs a physical topology graph G that includes spatial location, connection relationships, and electrical characteristics based on the node positions, connecting edges, and node attributes described above. phy It can be represented as:

[0088] Among them, G phy The structure representing the physical topology graph; Represents the edge set of a physical topology graph. Denotes the i-th edge in the edge set; A set of nodes representing a physical topology graph. Represents node attributes, A i L i and D i These represent the appearance of the metering device, the device coordinate information, and the device disaster label, respectively; M represents the number of nodes in the physical topology graph; and Q represents the number of edges.

[0089] For example, referring to Figure 2, the coordinate information of the key metering device (the current metering device) is first obtained. This coordinate information is then used as nodes in the association graph, with each metering device corresponding to one node, the node's position determined by its coordinates. Next, based on the communication and power transmission associations between the current metering device and other metering devices, physical connections are established and represented as connecting edges in the association graph, thus reflecting the topological connections between devices. Then, node attributes are added to each node, including device appearance status (describing the device's physical appearance and operating status), device coordinate information (recording the device's spatial location), and device disaster label (identifying whether the metering device is affected by a disaster or the type and extent of the impact). Combining these nodes, connecting edges, and node attributes, the physical topology graph structure is constructed, forming a complete physical topology graph reflecting the spatial location, connection relationships, and status attributes of the key metering devices.

[0090] S103: Determine the time series data of the current metering device based on the initial sample set, determine the semantic objects based on the time series data and physical topology diagram of the current metering device, and generate the connection edges of the semantic objects according to preset rules, and determine the semantic structure graph based on the connection edges of the semantic objects.

[0091] In this embodiment, time-series data is used as input data for physical topology graph nodes. Semantic features are extracted using graph convolutional neural networks to identify and output semantic objects. At the same time, connection edges between semantic objects are generated based on temporal proximity edges and co-occurrence edges to form a semantic structure graph.

[0092] In one implementation, based on the time-series data and physical topology of the metering device, semantic objects are determined using a graph convolutional neural network, and connection edges of the semantic objects are generated according to preset rules. A semantic structure graph is then determined based on these connection edges, including:

[0093] The time series data of the metering device is preprocessed to obtain preprocessed time series data;

[0094] The preprocessed time series data is input into the physical topology graph to obtain a physical topology graph that incorporates dynamic features;

[0095] Semantic features are obtained by combining a physical topology graph with dynamic features and a graph neural convolutional network, and semantic objects are determined based on the semantic features.

[0096] The semantic object is generated by generating connection edges based on the semantic object, temporal proximity edges, and co-occurrence edges. Temporal proximity edges are used to represent events that are closely continuous in time and have causal or triggering relationships. Co-occurrence edges are used to represent events that occur in the same time period under common influence.

[0097] The edge set of the semantic graph is determined based on the connecting edges of the semantic objects. The edge set of the semantic graph represents the set of all connecting edges.

[0098] The node set of the semantic graph is determined based on the semantic object and semantic features;

[0099] The semantic structure graph is determined based on the edge set and node set of the semantic graph.

[0100] In one implementation, semantic features are obtained based on a physical topology graph that incorporates dynamic features and a graph neural convolutional network, including:

[0101] The physical topology graph incorporating dynamic features is input into the graph neural network. Based on the multiple convolutional layers of the graph neural network, the aggregated feature information between each node in the physical topology graph incorporating dynamic features and its neighboring nodes is determined.

[0102] The aggregated feature information is input into the fully connected layer to obtain the corresponding semantic features.

[0103] In this embodiment, time-series data within a preset time period (which can be adjusted according to the actual scenario) is used as input data for physical topology graph nodes; a graph convolutional neural network is used to extract semantic features from the time-series data and output semantic objects; temporal proximity edges and co-occurrence edges are used to generate connection edges of different types for semantic objects; and a semantic structure graph is determined through the above semantic features and connection edges. The time-series data in this embodiment may include voltage, current, power consumption, and frequency.

[0104] In this embodiment, a graph convolutional neural network is used to extract semantic features from time-series data. The specific steps for outputting semantic objects may include:

[0105] Based on the physical topology graph constructed above, the time-series data measurements of all graph nodes at each time point are taken as a graph signal. The graph convolutional neural network learns the smoothness and change patterns of the graph signal in space through neighbor aggregation, extracts multi-level semantic features, and uses the graph convolutional neural network to output continuous time-series data as discrete semantic objects with clear physical meaning, such as voltage over-limit, current phase loss, power reversal, load overcapacity, three-phase imbalance, small nighttime electricity consumption, and zero current anomaly.

[0106] In this embodiment, temporal proximity edges can capture events that are closely consecutive in time and may have a causal or triggering relationship. For example, for events A and B, if the difference between the end time of event A and the start time of event B is less than a preset time threshold ΔT (the preset time threshold can be set to 1 minute), then a connection edge from event A to event B is established, and this connection edge is the temporal proximity edge.

[0107] In this embodiment, co-occurrence edges are events that are influenced by a common cause and occur within the same time period. For example, a time window T is set (this time window can be set to 1 hour), and all event pairs that co-occur within this time window T are counted. If the number of times event pairs (A, B) co-occur exceeds the minimum support threshold, then a connection edge is established, and this connection edge is used as a co-occurrence edge.

[0108] This embodiment determines the semantic structure graph based on the above node set and edge set. This semantic structure graph can be represented as follows:

[0109] Among them, G sem Represents a semantic structure graph; A set of nodes representing a semantic graph; Represents the edge set of a semantic graph; Let N represent the i-th connecting edge; N represents the number of nodes; and P represents the number of edges.

[0110] For example, referring to Figure 3, the time-series data of the key metering device within a preset time period is first used as input. This time-series data corresponds to each graph node in the physical topology graph. Then, a graph convolutional neural network (GCN) is used to extract features from the input time-series data to obtain the semantic object corresponding to each node. This semantic object includes the spatiotemporal correlation and state semantic information in the time-series data. Next, based on the relationship between the semantic objects, two types of connection edges are generated, including temporal proximity edges (used to reflect the dependency relationship of the same node at different time steps) and co-occurrence edges (used to reflect the co-occurrence or mutual influence relationship of the states of different nodes in the same or similar time periods). Finally, by combining the semantic objects (nodes) and semantic connection edges, the semantic graph structure is constructed to form a semantic structure graph that can characterize the time-series behavior and associated semantics of the metering device.

[0111] S104: Based on the physical topology graph and semantic structure graph, and through a multimodal fusion model based on an attention mechanism, the physical topology graph and semantic structure graph are fused to obtain the fault diagnosis results and early warning information of the current metering device.

[0112] In this embodiment, by introducing a multimodal fusion module based on an attention mechanism, the physical topology graph and the semantic structure graph are fused to output the fault diagnosis results and early warning information of the current metering device.

[0113] In one implementation, based on a physical topology graph and a semantic structure graph, and through a multimodal fusion model based on an attention mechanism, the physical topology graph and the semantic structure graph are fused to obtain fault diagnosis results and early warning information for the metering device, including:

[0114] A graph isomorphic network is used to extract feature vectors from the physical topology graph and the semantic structure graph to obtain physical feature vectors and semantic feature vectors.

[0115] The physical feature vector and the semantic feature vector are input into a multimodal fusion model with an attention mechanism to obtain the fused feature vector;

[0116] The fault diagnosis results and early warning information corresponding to the fused feature vectors are determined based on the fused feature vectors and the preset activation function.

[0117] In this embodiment, a graph isomorphic network is used to extract multimodal feature vectors. Specifically, a physical topology graph is used. Taking a specific example, feature vector extraction is performed. Specifically, for node v in this physical topology graph... phy In this embodiment, the features of the node itself and the features of all its neighboring nodes can be aggregated using a summation aggregation operation. The aggregation formula can be:

[0118] Where AGG(·) represents the aggregation operation; k represents the k-th convolutional layer; N(v phy ) represents the central node v phy The set of adjacent nodes, Indicates the characteristics of the central node. This represents the characteristics of adjacent nodes.

[0119] In this embodiment, after feature aggregation is completed, the learnable parameter e is used to adjust the center node v. phy The fusion ratio between a feature and its neighboring features. Then, the aggregated node features are updated based on a multilayer perceptron:

[0120] in, This represents the updated node characteristics.

[0121] In this embodiment, the graph isomorphic network aggregates the final feature vectors of all nodes through a summation operation to obtain the global feature vector h of the physical topology graph. phy The global feature vector is: Then, a more refined and relevant physical topology graph feature vector is extracted using a multilayer perceptron with two fully connected layers. Similarly, the semantic structure graph can be processed through the above operations to obtain a global feature vector h. sem Extraction.

[0122] In this embodiment, after extracting the feature vector, the physical feature vector h is... phy and semantic feature vector h sem The features are concatenated to form a merged feature vector h. concat The merged feature vector can be represented as: h concat =Concat(h sem ,h phy This embodiment ensures that both physical topological and semantic information within the same feature space are preserved through a splicing operation.

[0123] This embodiment introduces an attention mechanism, enabling the multimodal fusion model to dynamically determine whether to rely more on physical topological information or semantic information during the decision-making process. This allows the model to adaptively adjust the importance of features based on different task objectives, thus focusing more on the information most helpful for the prediction task. This embodiment generates normalized attention weights through a fully connected layer combined with a sigmoid function, obtaining the physical feature attention score w. phy and semantic feature attention score w sem Among them, the physical feature attention score is: w sem =Sigmoid(FC) Layer (h sem ), where FC Layer This represents a fully connected layer (FC); the semantic feature attention score is: w phy =Sigmoid(FC) Layer (h phy This embodiment calculates the final fused feature vector h by weighting and summing the physical feature vector, semantic feature vector, and their corresponding attention weights. fus The fused feature vector is: h fus =w phy ·h phy +w sem ·h sem .

[0124] The weighted eigenvector h fusThe input is fed into a multilayer perceptron (MLP) for model training. Fully connected layers are used to integrate features and perform nonlinear transformations. ReLU activation function and Dropout layer are introduced to enhance the model's expressive power while preventing overfitting.

[0125] In this embodiment, the Softmax activation function is used as the last layer in the training of the multilayer perceptron (MLP). The activation function outputs the sample fault diagnosis result, that is, the probability that the fault of the metering device belongs to each fault category (e.g., "normal", "overload", "appearance damage"), thereby realizing the fault diagnosis and disaster early warning of the metering device.

[0126] In this embodiment, the Softmax activation function can be defined as: Among them, y i C represents the predicted output value for fault type diagnosis of the metering device, where C is the number of fault types.

[0127] For example, referring to Figure 4, firstly, based on the constructed physical topology graph and semantic structure graph, a graph isomorphic network is used to extract their respective multimodal features, resulting in physical feature vectors and semantic feature vectors. The extracted physical feature vectors and semantic feature vectors are concatenated along the feature dimension to form a merged feature vector, thus fusing physical connectivity and semantic association information. The concatenated feature vector is then input into a fusion module based on an attention mechanism to further weight and integrate and enhance features from different sources, resulting in a unified feature representation after fusion. The fused feature vector is then input into a multilayer perceptron for classification training, and the probability is normalized using the Softmax activation function in the last layer, ultimately outputting the fault diagnosis result of the corresponding metering device.

[0128] As can be seen from the above, this application proposes a method, device, equipment, and medium for early warning of metering device faults. It constructs a physical topology graph based on the communication relationships between metering devices, and simultaneously uses graph convolutional networks to extract deep feature associations from time-series data to generate a semantic structure graph. By combining the physical connections between metering devices with the temporal semantic features of the data, a comprehensive analysis and diagnosis of the causes of metering device faults can be achieved. This embodiment also proposes a multimodal data fusion method based on an attention mechanism, which can adaptively determine the contribution of physical topology features and semantic features to the diagnostic results and accurately focus on key information, thereby significantly improving the accuracy of fault diagnosis and the reliability of early warning, solving the problems of large errors and low efficiency in current metering device fault diagnosis. This embodiment, by constructing a physical topology graph and a semantic structure graph, deeply mines and integrates various types of metering device data, and adopts a feature fusion method based on an attention mechanism. This not only enhances the model's ability to process multimodal data but also improves the overall prediction accuracy and interpretability of the model, effectively improving the performance of the metering fault diagnosis model.

[0129] In one embodiment of this application, it further includes:

[0130] Based on the fault diagnosis results and early warning information, a field verification message is sent to the terminal of the corresponding maintenance personnel so that the maintenance personnel can conduct on-site verification of the current metering device.

[0131] Obtain the verification results and determine the warning level and the corresponding deployment and prevention measures based on the verification results.

[0132] In this embodiment, based on the fault diagnosis results and early warning information of the metering device obtained above, on-site verification information can be sent to the terminals of the corresponding maintenance personnel. This enables the maintenance dispatch personnel to conduct on-site investigations of the fault diagnosis information and to make advance deployments to prevent potential disaster risks based on the early warning level and fault type. Obtaining the results of the on-site investigations by the maintenance personnel and using them as key samples to update the initial sample set in a timely manner can further optimize the diagnostic model, improve identification accuracy, and reduce false alarms and missed alarms.

[0133] For example, referring to Figure 5, firstly, multimodal data of the metering device is acquired, including geographic information, power grid topology, electricity consumption data, historical disaster data, and actual images of the device. This multimodal data is then preprocessed, including cleaning, alignment, and formatting, to form an initial sample set. Based on the current metering device and the power transmission relationships or communication associations between other metering devices, a topology map reflecting the physical connection structure of the metering devices is constructed, forming a physical topology map structure. Based on the time-series operating data of the metering devices, a graph convolutional neural network is used to extract time-series features, and a semantic structure map is constructed by combining time-series proximity and co-occurrence relationships. The physical topology map and the semantic structure map are then input into a multimodal data fusion module based on an attention mechanism for feature extraction, fusion, and enhancement, ultimately outputting the metering fault diagnosis result. Based on the fault diagnosis results and early warning information, on-site verification information is sent to the terminals of the corresponding operation and maintenance personnel. This allows the operation and maintenance personnel to conduct on-site information verification based on the fault diagnosis results, carry out advance deployment and prevention of disaster risks, complete fault handling, and feed back the fault handling results to the initial sample set, forming a closed-loop process of diagnosis, handling, and feedback. This supports the continuous optimization of the model and the adjustment of operation and maintenance strategies.

[0134] For example, this embodiment uses the State Grid electricity consumption information collection system as the experimental platform, selecting monitoring data from various metering devices (e.g., single-phase smart meters, three-phase smart meters, instrument transformers, etc.) within a preset time range (the predicted time range in this embodiment can be from January 1, 2024 to December 31, 2024). This monitoring data includes eight major fault types and their normal states, such as "normal," "exterior damage," "suspected electricity theft," "overload," and "short circuit," totaling 285,420 valid data points. All valid data are used as the initial sample set. Of these, 228,336 cases can be selected as the training sample set, and 57,084 cases as the test set.

[0135] To verify the effectiveness of this embodiment, precision, recall, and F1 score can be used as performance evaluation metrics.

[0136] Precision represents the proportion of true positive samples out of all samples predicted as positive. It is calculated as follows: Wherein, TP represents the number of samples where the actual value of the metering device is normal and the predicted value is also normal, and FP represents the number of samples where the actual value of the metering device is faulty and the predicted value is normal.

[0137] Recall is the number of positive samples that are successfully predicted out of all true positive samples. It is calculated as follows: Where FN represents the number of samples where the actual value of the metering device is normal and the predicted value is faulty.

[0138] The F1 score is the harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model. It is calculated as follows:

[0139] This embodiment selects test set sample data and compares the method provided in this embodiment with three similar methods: cue learning, parallel dual-channel fault diagnosis, and space-time cross attention mechanism, so as to evaluate the fault diagnosis accuracy of this embodiment. The performance comparison results of the four methods are shown in Table 1.

[0140] Table 1 shows the performance comparison results of different methods.

[0141]

[0142] According to Table 1 above, under different evaluation indicators, the method provided in this embodiment outperforms other similar methods, demonstrating excellent fault diagnosis performance and enabling accurate fault analysis of metering devices under multiple environmental influencing factors. Experimental results show that the model constructed using the attention mechanism-based multimodal data fusion approach in this embodiment, by deeply mining the information of the metering device and fully utilizing multimodal data, achieves more accurate and reliable fault diagnosis and disaster early warning for metering devices.

[0143] This embodiment also included ablation experiments. The ablation comparisons were made with the method provided in this embodiment (Method 1), the method for diagnosing metering device faults without constructing a physical topology graph (Method 2), the method for diagnosing metering device faults without constructing a semantic graph (Method 3), and the method for multimodal fusion without using an attention-based fusion mechanism (Method 4). Specific experimental results are shown in Figure 6.

[0144] As shown in Figure 6, according to Method 2 (fault diagnosis without constructing a physical topology graph), the model lacks an explicit representation of the connections between devices, making it unable to obtain the topological relationships of fault factors. This makes it difficult to distinguish between root causes and derivative phenomena, resulting in a significant decrease in prediction accuracy and model robustness. According to Method 3 (fault diagnosis without constructing a semantic graph), the semantic graph contains the most crucial time-series data for fault diagnosis. The normality or abnormality of the time-series data can determine most fault sources. However, Method 3 does not construct a semantic graph, meaning it fails to obtain the most critical time-series data, leading to the most severe performance degradation. Method 4 (fault diagnosis without using attention-based fusion) also shows a significant performance decrease. The fact that all the above methods lead to varying degrees of performance degradation further verifies that the feature-level fusion method used in this embodiment can more effectively integrate heterogeneous information from both physical and semantic modalities, and its performance is superior to simple feature splicing. This embodiment constructs a physical topology graph and a semantic structure graph, deeply mines and integrates data from various metering devices, and adopts a feature fusion method based on an attention mechanism. This not only enhances the model's ability to process multimodal data, but also improves the model's overall prediction accuracy and interpretability, effectively improving the performance of the metering fault diagnosis model.

[0145] Corresponding to the metering device fault early warning method in the above embodiments, Figure 7 is a structural block diagram of a metering device fault early warning device provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiments of this application are shown. Referring to Figure 7, the metering device fault early warning device 70 includes: a sample determination module 71, a physical topology map determination module 72, a semantic structure map determination module 73, and an early warning information determination module 74.

[0146] Among them, the sample determination module 71 is used to acquire multimodal data corresponding to multiple metering devices, mark the risk level of the metering devices according to the multimodal data and historical meteorological disasters, and determine the initial sample set corresponding to each metering device according to the risk level marking results. The multimodal data includes geographical location, operating status electrical parameters, power consumption and real-shot images.

[0147] The physical topology determination module 72 is used to determine the physical topology based on the metering device and other metering devices that have communication and power transmission connections with the metering device.

[0148] The semantic structure graph determination module 73 is used to determine the time series data of the metering device based on the initial sample set, determine the semantic objects based on the time series data and physical topology graph of the metering device, and generate the connection edges of the semantic objects according to preset rules, and determine the semantic structure graph based on the connection edges of the semantic objects.

[0149] The early warning information determination module 74 is used to obtain the fault diagnosis results and early warning information of the metering device by fusing the physical topology graph and the semantic structure graph through a multimodal fusion model based on the attention mechanism.

[0150] In one embodiment of this application, when determining a physical topology map based on the metering device and metering devices that have communication and power transmission associations with the metering device, the physical topology map determination module 72 is specifically used for:

[0151] Determine the location of the nodes on the graph based on the coordinate information of the metering device;

[0152] Based on the metering devices and the metering devices associated with the communication links and transmission lines, determine the connection edges between each node.

[0153] Node attributes are determined based on the appearance of the metering device, coordinate information, and disaster labels.

[0154] Construct a physical topology graph containing spatial location, connection relationships, and electrical characteristics based on the node locations, connecting edges, and node attributes.

[0155] In one embodiment of this application, when determining node attributes based on the appearance of the metering device, coordinate information, and disaster tags, the physical topology map determination module 72 is specifically used for:

[0156] Feature extraction is performed on the real-shot images to obtain the appearance state feature vector, which is used to characterize the appearance state of the measuring device.

[0157] During the data acquisition phase, the coordinate information corresponding to the metering device is obtained;

[0158] Obtain historical disaster data;

[0159] The disaster label of the metering device is determined based on historical disaster data and the corresponding geographical coordinates of the historical disaster data;

[0160] The node attributes of the metering device are determined based on the appearance status feature vector, the coordinate information of the metering device, and the disaster label of the metering device.

[0161] In one embodiment of this application, when determining semantic objects based on the time-series data and physical topology of the metering device, and using a graph convolutional neural network, and generating connection edges of the semantic objects according to preset rules, and determining the semantic structure graph based on the connection edges of the semantic objects, the semantic structure graph determination module 73 is specifically used for:

[0162] The time series data of the metering device is preprocessed to obtain preprocessed time series data;

[0163] The preprocessed time series data is input into the physical topology graph to obtain a physical topology graph that incorporates dynamic features;

[0164] Semantic features are obtained by combining a physical topology graph with dynamic features and a graph neural convolutional network, and semantic objects are determined based on the semantic features.

[0165] The semantic object is generated by generating connection edges based on the semantic object, temporal proximity edges, and co-occurrence edges. Temporal proximity edges are used to represent events that are closely continuous in time and have causal or triggering relationships. Co-occurrence edges are used to represent events that occur in the same time period under common influence.

[0166] The edge set of the semantic graph is determined based on the connecting edges of the semantic objects. The edge set of the semantic graph represents the set of all connecting edges.

[0167] The node set of the semantic graph is determined based on the semantic object and semantic features;

[0168] The semantic structure graph is determined based on the edge set and node set of the semantic graph.

[0169] In one embodiment of this application, when obtaining semantic features based on a physical topology graph that incorporates dynamic features and a graph neural convolutional network, the semantic structure graph determination module 73 is specifically used for:

[0170] The physical topology graph incorporating dynamic features is input into the graph neural network. Based on the multiple convolutional layers of the graph neural network, the aggregated feature information between each node in the physical topology graph incorporating dynamic features and its neighboring nodes is determined.

[0171] The aggregated feature information is input into the fully connected layer to obtain the corresponding semantic features. In one embodiment of this application, when fusing the physical topology graph and the semantic structure graph using a multimodal fusion model based on an attention mechanism to obtain the fault diagnosis result and early warning information of the metering device, the early warning information determination module 74 is specifically used for:

[0172] A graph isomorphic network is used to extract feature vectors from the physical topology graph and the semantic structure graph to obtain physical feature vectors and semantic feature vectors.

[0173] The physical feature vector and the semantic feature vector are input into a multimodal fusion model with an attention mechanism to obtain the fused feature vector;

[0174] The fault diagnosis results and early warning information corresponding to the fused feature vectors are determined based on the fused feature vectors and the preset activation function.

[0175] In one embodiment of this application, the device further includes a deployment prevention module, which is used for:

[0176] Based on the fault diagnosis results and early warning information, a field verification message is sent to the terminal of the corresponding maintenance personnel so that the maintenance personnel can conduct on-site verification of the metering device.

[0177] Obtain the verification results and determine the warning level and the corresponding deployment and prevention measures based on the verification results.

[0178] Referring to Figure 8, which is a schematic block diagram of an electronic device provided in an embodiment of this application, the electronic device 800 in this embodiment, as shown in Figure 8, may include one or more processors 801, one or more input devices 802, one or more output devices 803, and one or more memories 804. The processors 801, input devices 802, output devices 803, and memories 804 communicate with each other via a communication bus 805. The memories 804 are used to store computer programs, which include program instructions. The processors 801 are used to execute the program instructions stored in the memories 804. The processors 801 are configured to invoke the program instructions to execute the functions of the modules in the above-described device embodiments, such as the sample determination module 71, physical topology map determination module 72, semantic structure map determination module 73, and early warning information determination module 74 shown in Figure 7.

[0179] It should be understood that, in the embodiments of this application, the processor 801 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0180] Input device 802 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 803 may include a display (LCD, etc.), a speaker, etc.

[0181] The memory 804 may include read-only memory and random access memory, and provides instructions and data to the processor 801. A portion of the memory 804 may also include non-volatile random access memory. For example, the memory 804 may also store initial sample sets, physical topology diagrams, timing data, semantic structure diagrams, and fault diagnosis results and early warning information.

[0182] In specific implementations, the processor 801, input device 802, and output device 803 described in the embodiments of this application can execute the implementation method described in the metering device fault early warning method provided in the embodiments of this application, or they can execute the implementation method of the electronic device described in the embodiments of this application, which will not be repeated here.

[0183] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0184] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0185] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0187] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces or units, or they may be electrical, mechanical, or other forms of connection.

[0188] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0189] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0190] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for early warning of faults in a metering device, characterized in that, include: Acquire multimodal data corresponding to multiple metering devices, and label the metering devices with risk levels based on the multimodal data and historical meteorological disasters. Determine the initial sample set corresponding to each metering device based on the risk level labeling results. The multimodal data includes geographical location, operating status electrical parameters, power consumption, and real-time images. For each metering device, perform the following early warning operation: determine the physical topology map based on the metering device and the metering devices that have communication and power transmission associations with it. The timing data of the metering device is determined based on the initial sample set. Based on the timing data of the metering device and the physical topology graph, semantic objects are determined through a graph convolutional neural network. Connection edges of the semantic objects are generated according to preset rules. A semantic structure graph is determined based on the connection edges of the semantic objects. The physical topology graph and the semantic structure graph are fused through a multimodal fusion model based on an attention mechanism to obtain the fault diagnosis results and early warning information of the metering device.

2. The method for early warning of metering device faults as described in claim 1, characterized in that, The step of determining the physical topology map based on the metering device and metering devices that have communication and power transmission associations with the metering device includes: determining the location of graph nodes based on the coordinate information of the metering device; determining the connecting edges between each node based on the metering device and metering devices that have communication link associations and power transmission line associations with the metering device; determining the node attributes based on the appearance, coordinate information, and disaster labels of the metering device; and constructing a physical topology map including spatial location, connection relationship, and electrical characteristics based on the graph node locations, the connecting edges, and the node attributes.

3. The method for early warning of metering device faults as described in claim 2, characterized in that, Before determining the node attributes based on the appearance status, coordinate information, and disaster label of the metering device, the method further includes: extracting features from the captured image to obtain an appearance status feature vector, which is used to characterize the appearance status of the metering device; acquiring the coordinate information corresponding to the metering device during the data acquisition phase; acquiring historical disaster data; and determining the disaster label of the metering device based on the historical disaster data and the geographical coordinates corresponding to the historical disaster data. Specifically, determining the node attributes based on the appearance status, coordinate information, and disaster label of the metering device includes: determining the node attributes of the metering device based on the appearance status feature vector, the coordinate information of the metering device, and the disaster label of the metering device.

4. The method for early warning of metering device faults as described in claim 1, characterized in that, The process of determining semantic objects based on the time-series data of the metering device and the physical topology graph, using a graph convolutional neural network, generating connection edges of the semantic objects according to preset rules, and determining a semantic structure graph based on the connection edges of the semantic objects includes: preprocessing the time-series data of the metering device to obtain preprocessed time-series data; inputting the preprocessed time-series data into the physical topology graph to obtain a physical topology graph incorporating dynamic features; obtaining semantic features based on the physical topology graph incorporating dynamic features and the graph convolutional neural network, and determining semantic objects based on the semantic features; generating connection edges of the semantic objects based on the semantic objects, temporally adjacent edges, and co-occurrence edges, where temporally adjacent edges represent events that are closely continuous in time and have a causal or triggering relationship, and co-occurrence edges represent events that occur within the same time period under common influence; determining the edge set of the semantic graph based on the connection edges of the semantic objects, where the edge set of the semantic graph represents the set of all connection edges; determining the node set of the semantic graph based on the semantic objects and the semantic features; and determining a semantic structure graph based on the edge set and the node set of the semantic graph.

5. A method for early warning of metering device faults as described in claim 4, characterized in that, The step of obtaining semantic features based on the physical topology graph with fused dynamic features and the graph neural convolutional network includes: inputting the physical topology graph with fused dynamic features into the graph neural network; determining the aggregated feature information between each node in the physical topology graph with fused dynamic features and its adjacent nodes based on the multiple convolutional layers of the graph neural network; and inputting the aggregated feature information into a fully connected layer to obtain the corresponding semantic features.

6. The method for early warning of metering device faults as described in claim 1, characterized in that, The step of fusing the physical topology graph and the semantic structure graph using an attention-based multimodal fusion model to obtain fault diagnosis results and early warning information for the metering device includes: extracting feature vectors from the physical topology graph and the semantic structure graph using a graph isomorphic network to obtain physical feature vectors and semantic feature vectors; inputting the physical feature vectors and the semantic feature vectors into the attention-based multimodal fusion model to obtain fused feature vectors; and determining the fault diagnosis results and early warning information corresponding to the fused feature vectors based on the fused feature vectors and a preset activation function.

7. A method for early warning of faults in a metering device as described in claim 6, characterized in that, Also includes: Based on the fault diagnosis results and early warning information, a field verification message is sent to the terminal of the corresponding maintenance personnel so that the maintenance personnel can conduct on-site verification of the metering device. Obtain the verification results, and determine the warning level and the corresponding deployment and prevention measures based on the verification results.

8. A fault early warning device for a metering device, characterized in that, include: The sample determination module is used to acquire multimodal data corresponding to multiple metering devices, mark the risk level of the metering devices according to the multimodal data and historical meteorological disasters, and determine the initial sample set corresponding to each metering device according to the risk level marking results. The multimodal data includes geographical location, operating status electrical parameters, power consumption and real-time images. The physical topology determination module is used to determine the physical topology based on the metering device and the metering devices that have communication and power transmission associations with the metering device. The semantic structure graph determination module is used to determine the time-series data of the metering device based on the initial sample set, determine semantic objects based on the time-series data of the metering device and the physical topology graph, and generate connection edges of the semantic objects according to preset rules, and determine the semantic structure graph based on the connection edges of the semantic objects; the warning information determination module is used to fuse the physical topology graph and the semantic structure graph based on the attention mechanism and the semantic structure graph to obtain the fault diagnosis results and warning information of the metering device.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images