Building electrical safety protection system and method thereof

Abstract

The invention discloses a building electrical safety protection method, which comprises the following steps of performing real-time data acquisition on a building electrical system to obtain multi-modal time sequence data; constructing the multi-modal time sequence data into graph structure data with a node-edge topological relation, and encoding the multi-modal original observation data corresponding to each node into a multi-modal initial feature vector of the node; graph neural network feature extraction is carried out on the graph structure data, and a global feature vector sequence used for representing the operation state of the electrical system is obtained in combination with an attention mechanism; inputting the global feature vector sequence into a pre-constructed time sequence prediction model to perform operation state prediction, and judging whether potential abnormality exists or not; and when the residual error exceeds a preset threshold value, fault backtracking positioning is carried out on the key node according to the attention weight, and an electrical fault point is generated in combination with node characteristics. According to the invention, the fault response speed and the emergency disposal efficiency can be effectively improved, and the safety accident rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of electrical safety protection technology, and in particular to a building electrical safety protection system and method. Background Technology

[0002] With the acceleration of urbanization and the popularization of intelligent buildings, the stability and safety of building electrical systems, as an important component of building safety, are receiving increasing attention. Building electrical systems typically include multiple subsystems such as power supply and distribution, lighting, air conditioning, and elevators. These systems have complex structures, numerous nodes, and variable operating states, making them highly susceptible to various factors such as load fluctuations, equipment aging, and environmental changes, which can lead to potential faults or even electrical safety accidents.

[0003] Currently, building electrical safety protection largely relies on manual inspections and alarm mechanisms based on traditional rules. These methods suffer from poor detection timeliness, low data utilization, and a lack of predictive judgment of potential hazards, making them ill-suited to the "early warning, rapid response, and high precision" safety management requirements of modern intelligent buildings. Furthermore, sensors in building electrical systems are increasingly being deployed in multiple modes, collecting various types of time-series data such as voltage, current, power, temperature, and humidity. Therefore, effectively modeling and fusing this multimodal time-series data has become a key technical challenge for the intelligent upgrading of building electrical safety protection.

[0004] In recent years, Graph Neural Networks (GNNs) have attracted widespread attention due to their ability to handle data with non-Euclidean structures. Building electrical systems inherently possess topological network characteristics, with clear connections, geographical proximity, and functional coupling between different nodes (such as electrical cabinets, distribution boxes, and terminal equipment). Therefore, abstracting building electrical systems as graph structures and utilizing GNNs to model their states and extract features has become an effective way to improve the accuracy of fault detection and risk prediction in building electrical systems.

[0005] However, existing GNN-based electrical fault detection methods generally lack the modeling of time dynamics and the introduction of attention mechanisms, making it difficult to fully capture the heterogeneous information interaction characteristics between nodes. They also lack the ability to detect potential anomalies and locate faults from the perspective of residual analysis, making it difficult to quickly and accurately locate fault points after a fault occurs. This affects the efficiency of fault repair and the rapid recovery of electrical systems, which may lead to an expansion of the fault's impact range, causing greater economic losses and safety risks. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a building electrical safety protection method, comprising the following steps: Real-time data acquisition of building electrical systems to obtain multimodal time series data; The multimodal time series data is constructed into a graph structure with node-edge topological relationships, and the original multimodal observation data corresponding to each node is encoded into the multimodal initial feature vector of the node; The graph structure data is subjected to graph neural network feature extraction, and an attention mechanism is used to obtain a global feature vector sequence for characterizing the operating state of the electrical system; The global feature vector sequence is input into a pre-built time series prediction model to predict the running status, and the presence of potential anomalies is determined based on the residual sequence between the predicted values ​​and the actual observed values. When the residual exceeds a preset threshold, the key nodes are backtracked and located according to the attention weight, and the electrical fault point and its risk level assessment result are generated by combining the node characteristics.

[0008] As a preferred embodiment of the building electrical safety protection method of the present invention, the method includes the following steps: constructing the multimodal time series data into a graph structure with node-edge topological relationships, and encoding the multimodal original observation data corresponding to each node into the multimodal initial feature vector of the node. The locations where sensors are installed in the building electrical system are modeled as nodes in a graph, with each node corresponding to a multimodal raw observation data sequence for a specific location; Based on the electrical connections, physical proximity, and functional coupling between nodes, edges are established between nodes to form graph structure data with node-edge topology. Assign an edge weight to each edge to represent the similarity, transitivity, or dependency between nodes; The multimodal raw observation data of each node are preprocessed and fused, and the fused multimodal features are encoded into the initial multimodal feature vector of the node for subsequent feature extraction operations of the graph neural network.

[0009] As a preferred embodiment of the building electrical safety protection method of the present invention, the method includes the following steps: extracting graph neural network features from the graph structure data and obtaining a global feature vector sequence for characterizing the operating state of the electrical system by combining an attention mechanism: Construct the graph neural network model structure and define the input and output dimensions; In each layer of the graph neural network computation, a multi-head attention mechanism is introduced to assign different attention weights to the features of each neighboring node in order to dynamically capture the heterogeneity of the information interaction intensity between nodes at the current time. The intermediate features output from each layer of the graph neural network are concatenated or pooled to form a high-order semantic vector representation of each node, so as to preserve the multi-level contextual information of the node in the graph structure. A global readout operation is performed on the high-order semantic vector representations of all nodes in the entire graph to generate a global feature vector sequence of uniform dimension, which is used to characterize the overall operating status and potential risk trends of the current building electrical system.

[0010] As a preferred embodiment of the building electrical safety protection method of the present invention, the step of inputting the global feature vector sequence into a pre-constructed time-series prediction model for operational status prediction, and determining whether there are potential anomalies based on the residual sequence between the predicted values ​​and the actual observed values, includes the following steps: Build a time series prediction model and train it using historical samples; Input the global feature vector sequence into the time series prediction model to generate a sequence of system state prediction values ​​for multiple future time periods; The predicted value sequence is compared with the corresponding actual observation value sequence in the original multimodal observation data, and the residual sequence is calculated. Determine whether the residual exceeds the preset anomaly detection threshold. If any value in the residual sequence is greater than the anomaly detection threshold, then it is determined that there is a potential abnormal event in the current electrical system.

[0011] As a preferred embodiment of the building electrical safety protection method of the present invention, the step of performing fault backtracking and localization on key nodes based on attention weights, and generating electrical fault points and their risk level assessment results in combination with node characteristics, includes the following steps: Temporal backtracking analysis is performed based on the attention mechanism of graph neural networks to locate key nodes in the graph structure that lead to abnormal events. Analyze the evolution trajectory of multimodal historical features of key nodes to trace the source and determine the type of anomaly. The comprehensive risk score is calculated by combining multi-dimensional time-series backtracking information of key nodes, and the risk assessment result is generated.

[0012] As a preferred embodiment of the building electrical safety protection method of the present invention, the comprehensive risk score is obtained by weighted fusion of multi-dimensional time-series backtracking information, which includes residual amplitude, attention weight, node importance level and topological centrality index. The residual magnitude refers to the residual value extracted from the residual sequence corresponding to the key node as an important measure of the degree of abnormal deviation. The larger the residual magnitude, the greater the difference between the current state of the node and the normal value predicted by the model, and the higher the probability of failure. The attention weight refers to the average attention weight of the current node at the moment of an anomaly, obtained from the attention distribution matrix in the multi-head attention layer. This value reflects the degree of contribution of the node to the propagation of information in the whole graph. The higher the weight, the greater its impact on the overall state of the system. The node importance level refers to the pre-classification of each node in the building electrical system according to its functional criticality. This classification can be determined based on electrical drawings or expert annotations, and is used to reflect the degree of impact of the node on the stability of the power system. Topological centrality is a metric that calculates the centrality of each node in a graph based on the constructed graph structure data. It measures the degree to which a node is central in the information flow or fault propagation path.

[0013] As a preferred embodiment of the building electrical safety protection method of the present invention, the formula for calculating the comprehensive risk score is as follows:

[0014] in: Indicates the first The comprehensive risk score of each node, with a value range of [value range missing]. ; express Normalized residual magnitude values ​​for each node; Indicates the first Normalized attention weights for each node; Indicates the first Normalized values ​​of the importance level of each node; Indicates the first The normalized value of the centrality index of each node; , , , The weighting coefficients for each indicator satisfy... This is used to reflect the degree of impact of each indicator on overall risk.

[0015] As a preferred embodiment of the building electrical safety protection method of the present invention, wherein: according to the... The risk level is divided into four categories: safe, low-risk, medium-risk, and high-risk. The classification rules and emergency response strategies for each level are as follows: when When the node is in a normal state, no additional action is required; only routine periodic monitoring is performed. when When the node status deviates slightly from the expected level, it is classified as a low-risk level. The emergency strategy is as follows: If the node status deviates slightly from the expected level, it is recommended to increase the monitoring frequency of the node or adjust the local equipment operating parameters to reduce potential risks. when When a node shows a clear abnormal trend, a fault warning mechanism needs to be implemented, and maintenance personnel should be dispatched for on-site inspection. when When a node is classified as high-risk, the emergency response strategy is as follows: if a node has a serious electrical fault or potential safety threat, the emergency response mechanism must be triggered immediately, including fault isolation, power outage protection, linkage alarm and emergency handling procedures.

[0016] As a preferred embodiment of the building electrical safety protection method of the present invention, in which: when classifying the risk level and formulating the emergency strategy, the real-time operating status context information of the building electrical system is further combined with a multi-condition joint decision-making mechanism for comprehensive judgment; The context information includes: the current building's load type and load level, peak electricity consumption period labels, the safety level labels of the floor and area, and the historical maintenance records and inspection cycles of relevant nodes; In the aforementioned multi-condition joint decision-making mechanism, a context information vector is constructed based on the above context information, and it is fused with the comprehensive risk score value of each node to obtain a context-aware fused risk score value. The integrated risk score is compared with the preset risk level classification range, and the risk level of the node is dynamically reassessed based on the comparison results.

[0017] This application also provides a building electrical safety protection system, including: The multimodal data acquisition module is used to acquire real-time data from various sensor nodes in the building electrical system. The graph structure construction and node feature encoding module is used to construct the collected multimodal time series data into a graph data with a node-edge topology according to the physical connection relationship, geographical proximity and functional coupling of the building electrical system; and to perform feature fusion and encoding on the multimodal observation data corresponding to each node to generate the initial multimodal feature vector of the node for graph neural network processing. The graph neural network analysis module configures the graph neural network structure and embeds a multi-head attention mechanism to extract features from the constructed graph structure, generate a high-order semantic representation of each node, and generate a global feature vector sequence through a full graph readout operation to characterize the overall operating status of the building electrical system. The time series prediction and anomaly detection module includes a trained time series prediction model, which receives a global feature vector sequence generated by a graph neural network to predict the system's operating state at multiple future time points; and calculates and analyzes the residuals between the predicted values ​​and the actual observed values ​​to determine whether there are any abnormal states. The key node tracing and risk scoring module backtracks key abnormal nodes based on the attention mechanism, extracts their multi-dimensional temporal features, residual amplitude, attention weight, node importance level, and topological centrality index; and calculates a comprehensive risk score through weighted fusion to classify risk levels. The risk level classification and emergency response strategy module is used to classify safety levels, low risk levels, medium risk levels and high risk levels according to the magnitude of the integrated risk score, and to trigger corresponding emergency strategies according to the corresponding level.

[0018] The beneficial effects of this invention are: 1. This invention performs graph structure modeling on multimodal time series data collected from building electrical systems and encodes the original observation data into multimodal initial feature vectors of nodes. This enables the system to fully express the physical connection relationships and functional coupling characteristics between electrical nodes while maintaining the rich semantic information of the original data, thereby enhancing the model's ability to model the overall operating state of the electrical system.

[0019] 2. This invention utilizes graph neural networks to extract deep features from the constructed graph structure data and introduces an attention mechanism to enhance the feature perception capability of key nodes and key moments, thereby generating a global feature vector sequence and inputting it into a time-series prediction model for operational status prediction. This mechanism effectively overcomes the problem of insufficient modeling of node heterogeneity in traditional algorithms, enabling early prediction of potential electrical anomalies and achieving proactive identification of electrical hazards.

[0020] 3. This invention dynamically analyzes the residual sequence between predicted and actual observed values. When an anomaly is found to exceed a preset threshold, it can use an attention mechanism to deduce key influencing nodes and further integrate the feature information of the nodes to generate electrical fault points and their risk level assessment results. This can effectively improve fault response speed and emergency response efficiency, and reduce the incidence of safety accidents and system recovery costs. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Fig. 1 This is a flowchart of a building electrical safety protection method according to the present invention.

[0022] Fig. 2 This is a schematic diagram of the potential anomaly judgment process of a building electrical safety protection method according to the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0027] Example 1 Reference Figs. 1-2The first embodiment of the present invention provides a building electrical safety protection method, comprising the following steps: S1: Real-time data acquisition of building electrical systems to obtain multimodal time series data.

[0028] Specifically, in use, multimodal sensor devices are deployed at multiple key locations in the building's electrical system. These multimodal sensors include infrared thermal imaging sensors, ultrasonic detection sensors, electromagnetic interference sensors, and ambient gas sensors, which are used to collect multimodal time-series data in real time, including infrared thermal imaging, ultrasonic waves, electromagnetic interference, and ambient gas.

[0029] S2: Construct multimodal time series data into graph structure data with node-edge topological relationships, and encode the original multimodal observation data corresponding to each node into the multimodal initial feature vector of the node.

[0030] Specifically, the multimodal time series data is constructed into a graph structure with node-edge topological relationships, and the original multimodal observation data corresponding to each node is encoded into the node's initial multimodal feature vector, including the following steps: S21: Model the locations of sensors in the building electrical system as nodes in a graph, such as key electrical units like distribution boxes, cable interfaces, motors, control switches, and lighting circuits. Each node corresponds to a multimodal raw observation data sequence for a specific location.

[0031] S22: Based on the electrical connection relationship, physical proximity and functional coupling relationship between nodes, establish edges between nodes to form graph structure data with node-edge topology.

[0032] It should be noted that electrical connections, such as cables connecting a transformer to an electricity meter and then to the consumer terminal; physical proximity in terms of spatial location, such as floor wiring or equipment distances being less than a certain threshold; and functional coupling relationships, such as a set of fans and electrically controlled doors operating in tandem, are considered coupled devices.

[0033] S23: Assign edge weights to each edge to represent the similarity, transitivity, or dependency between nodes.

[0034] Specifically, edge weights are determined based on at least one of the following: physical distance between nodes, electrical load coupling strength, or node feature similarity.

[0035] It should be noted that in building electrical systems, physical distance often affects the transmission range and speed of electrical events (such as arcing, overload, and overheating). Physically adjacent electrical devices often have interconnected states; therefore, using the physical distance between nodes as the basis for setting edge weights helps to characterize the potential coupling brought about by adjacency.

[0036] Some electrical devices share power through the same circuit or power supply network, and their load changes may affect each other. For example, an increase in the load of one transformer may cause multiple downstream switchgear to overheat. Therefore, edge weights calculated based on load coupling strength can more realistically reflect the conduction relationships and risk propagation paths within the system.

[0037] If the multimodal observation data of two nodes are highly similar in time series or frequency domain characteristics, it usually means that they are in similar operating states or affected by the same disturbance source. This similarity can be measured by methods such as Euclidean distance, cosine similarity, and DTW. Using this similarity to set edge weights helps graph neural networks more accurately capture the dynamic evolution of the system and the trend of local anomaly propagation.

[0038] S24: Preprocess and fuse the multimodal raw observation data of each node, including normalization and feature extraction of data such as infrared thermal imaging, ultrasonic detection, electromagnetic interference, and ambient gas; encode the fused multimodal features into the initial multimodal feature vector of the node for subsequent feature extraction operations of the graph neural network.

[0039] S3: Extract features from graph structure data using graph neural networks, and combine this with an attention mechanism to obtain a global feature vector sequence that characterizes the operating state of the electrical system.

[0040] Specifically, graph neural network feature extraction is performed on graph-structured data, and an attention mechanism is used to obtain a global feature vector sequence that characterizes the operating state of the electrical system. This includes the following steps: S31: Construct the graph neural network model structure and define the input and output dimensions.

[0041] Specifically, based on the graph structure data constructed in step S2, a model structure containing multi-layer graph neural networks (such as GAT or GIN) is built. The input of each layer of the graph neural network is the multimodal initial feature vector of the node and its adjacency matrix, and the output is the node feature representation after fusing the information of neighboring nodes. The ability of nodes to perceive long-distance structural information is gradually enhanced by stacking multiple layers.

[0042] S32: In each layer of the graph neural network computation, a multi-head attention mechanism is introduced to assign different attention weights to the features of each neighboring node in order to dynamically capture the heterogeneity of the information interaction intensity between nodes at the current moment, thereby improving the model's ability to distinguish abnormal signal propagation paths.

[0043] S33: Concatenate or pool the intermediate features output from each layer of the graph neural network to form a high-order semantic vector representation for each node, so as to preserve the multi-level contextual information of the node in the graph structure and enhance the overall semantic modeling effect.

[0044] S34: Perform a global readout operation on the high-order semantic vector representations of all nodes in the entire graph. For example, through global average pooling or graph-level attention readout functions, generate a global feature vector sequence of uniform dimension to characterize the overall operating status and potential risk trends of the current building electrical system. Use this global feature vector sequence as input features for subsequent state prediction, providing the time-series prediction model with dynamic evolution clues that integrate structural information and node semantics, thereby achieving highly accurate prediction of the future operating status of the system.

[0045] S4: Input the global feature vector sequence into the pre-built time series prediction model to predict the running status, and determine whether there are potential anomalies based on the residual sequence between the predicted values ​​and the actual observed values.

[0046] Specifically, the global feature vector sequence is input into a pre-built time-series prediction model to predict the running state, and the presence of potential anomalies is determined based on the residual sequence between the predicted and actual observed values, including the following steps: S41: Construct a time series prediction model and train it using historical samples.

[0047] The time series prediction model uses a deep learning model suitable for capturing the evolution patterns of time series, such as LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), or Transformer time series model. The global feature vector sequence obtained in step S3 is used as input samples, combined with historical operating status label data of the building electrical system, for supervised learning training to optimize prediction accuracy.

[0048] S42: Input the global feature vector sequence extracted by the graph neural network at the current time step into the time series prediction model to generate a sequence of system state prediction values ​​for multiple future time steps.

[0049] For example, predicting the changing trends of electrical parameters or risk indicator values ​​over the next n time steps and outputting a sequence of predicted values. .

[0050] S43: Compare the predicted value sequence with the corresponding actual observation value sequence in the original multimodal observation data, and calculate the residual sequence.

[0051] For example, let the actual observation sequence be... Then the residual sequence is .

[0052] S44: Determine whether the residual exceeds the preset anomaly detection threshold.

[0053] It should be noted that the anomaly detection threshold is determined by a statistical model, such as 3. The rule is determined, and if any value in the residual sequence is greater than the anomaly detection threshold, step S5 is executed.

[0054] S5: When the residual exceeds the preset threshold, the key nodes are backtracked and located according to the attention weight, and the electrical fault point and its risk level assessment result are generated by combining the node characteristics.

[0055] Specifically, the process involves tracing and locating key nodes based on attention weights, and generating electrical fault points and their risk level assessments by combining node characteristics. This includes the following steps: S51: Perform time-series backtracking analysis based on graph neural network attention mechanism to locate key nodes in the graph structure that lead to abnormal events.

[0056] Furthermore, the evolution trajectory of attention weights for each node in the multi-head attention layer of the graph neural network is extracted within the current time and historical time windows, constructing a time series graph of node influence contribution over time. Combined with the time of residual anomaly occurrence, a source analysis is performed on nodes whose attention weights continuously increase or abruptly change during historical periods, and the set of nodes that have a key impact on global prediction bias is marked as the key node set. This set can be obtained by taking the top k nodes by attention weight, reflecting the core role of nodes in the generation and propagation of anomalies.

[0057] S52: Analyze the evolution trajectory of multimodal historical features of key nodes to trace the source and determine the type of anomaly.

[0058] In practice, multimodal characteristic time-series data (including infrared thermal imaging temperature rise, voltage and current abrupt changes, electromagnetic interference fluctuations, etc.) of each key node are extracted within a certain time window before the anomaly. Combined with the type of electrical components corresponding to the node and the preset fault feature library, typical anomaly evolution modes (such as overload heating, poor contact, arc discharge, etc.) are compared to trace the potential causes and evolution paths of the anomaly, thereby improving the understanding of the causal chain of anomalies within the system.

[0059] S53: Calculate the comprehensive risk score by combining the multi-dimensional time-series backtracking information of key nodes, and generate the risk assessment results.

[0060] Specifically, by combining multi-dimensional time-series backtracking information of key nodes at the moment of anomaly occurrence, a comprehensive risk score is calculated for each key node to determine the risk level of the electrical fault. The comprehensive risk score is obtained by weighted fusion of the following multi-dimensional time-series backtracking information, which includes: Residual magnitude: Based on the residual sequence obtained in step S4, the residual values ​​corresponding to key nodes are extracted as an important measure of the degree of abnormal deviation. The larger the residual magnitude, the greater the difference between the current state of the node and the normal value predicted by the model, and the higher the probability of failure.

[0061] Attention Weight: Based on the attention distribution matrix extracted from the multi-head attention layer in step S51, obtain the average attention weight of the current node at the moment the anomaly occurs. This value reflects the degree of contribution of the node to the propagation of information throughout the graph; the higher the weight, the greater its impact on the overall state of the system.

[0062] Node Importance Level: Each node in the building's electrical system is pre-classified according to its functional criticality. For example, transformers, main switches, and main cable nodes are considered more important than ordinary lighting circuits. This classification can be based on electrical drawings or expert annotations and reflects the degree of impact of each node on the stability of the power system.

[0063] Topological centrality metrics: Based on the graph structure constructed in step S2, calculate the centrality metric (such as betweenness centrality, degree centrality, or eigenvector centrality) of each node in the graph. This metric measures the centrality of the node in the information flow or fault propagation path. Higher centrality indicates that the node is more likely to become a central node for fault propagation.

[0064] After normalizing the above indicators, the comprehensive risk score of the node is obtained using the following formula:

[0065] in: Indicates the first The comprehensive risk score of each node, with a value range of [value range missing]. This is used to characterize the security level of the node's current state.

[0066] express Normalized residual magnitude values ​​for each node; Indicates the first Normalized attention weights for each node; Indicates the first Normalized values ​​of the importance level of each node; Indicates the first The normalized value of the centrality index of each node; , , , The weighting coefficients for each indicator satisfy... This is used to reflect the degree of impact of each indicator on overall risk.

[0067] according to The risk level is divided into four categories: safe, low-risk, medium-risk, and high-risk. The classification rules and emergency response strategies for each level are as follows: when When the node is in a normal state, no additional action is required; only routine periodic monitoring is performed. when When the node status deviates slightly from the expected level, it is classified as a low-risk level. The emergency strategy is as follows: If the node status deviates slightly from the expected level, it is recommended to increase the monitoring frequency of the node or adjust the local equipment operating parameters to reduce potential risks. when When a node shows a clear abnormal trend, a fault warning mechanism needs to be implemented, and maintenance personnel should be dispatched for on-site inspection. when When a node is classified as high-risk, the emergency response strategy is as follows: if a node has a serious electrical fault or potential safety threat, the emergency response mechanism must be triggered immediately, including fault isolation, power outage protection, linkage alarm and emergency handling procedures.

[0068] It should be noted that, , , Set them to 0.25, 0.5, and 0.75 respectively.

[0069] Furthermore, based on the established criteria, when classifying risk levels and formulating emergency strategies, the real-time operating status context information of the building electrical system should be further combined with a multi-condition joint decision-making mechanism for comprehensive judgment. Contextual information includes: the current building's load type and load level, peak electricity consumption period labels, the safety level labels of the floor and area, and the historical maintenance records and inspection cycles of relevant nodes; In the multi-condition joint decision-making mechanism, a context information vector is constructed based on the above context information, and it is fused with the comprehensive risk score value of each node to obtain a fused risk score value. The integrated risk score is compared with the preset risk level classification range, and the risk level of the node is dynamically reassessed based on the comparison results to improve the accuracy of risk identification and the pertinence of emergency response.

[0070] It should be noted that the formula for calculating the integrated risk score is as follows:

[0071] in: Represents a node The fusion risk score represents the comprehensive risk score recalculated under context-aware conditions, which is used for the final risk level classification judgment. Indicates the first The comprehensive risk score of each node is calculated solely based on the node's physical state and operational indicators (such as voltage, current, temperature rise, etc.), without any contextual information. This is the weighting coefficient used in the fusion formula to control the basic risk score in the total score. ; These are the weighting coefficients used in the fusion formula to control the weighting of contextual information in the overall score. and ; Represents a node The Each contextual information feature value (such as load level, maintenance cycle, etc.) is totaled. One contextual feature; No. The importance weighting factor of each context feature is used to characterize the contribution of different context dimensions to the comprehensive score. The value can be set according to expert experience; for example, the weight of the key functional area label is greater than the inspection cycle. The dimension of the context information vector represents the number of context features collected by the system. Represents the context vector The weighted sum is used to quantitatively express the impact of context on risk scores.

[0072] In this application, the integrated risk scoring formula is used to construct a context information vector based on the operating status and safety context information of the building electrical system. And combined with the corresponding importance weight factors A weighted fusion is performed. Two control weights are introduced into the fusion score. , This is used to dynamically balance the influence of basic risk indicators and context-aware information, enabling accurate re-determination of node risk levels.

[0073] Example as follows: Assumption: Node number is Its basic comprehensive risk score The results have been calculated using historical failure rates, real-time current, voltage fluctuation rates, and other characteristics, and then normalized. .

[0074] Context information dimension set to ,include: Load level The normalized value is 0.9 (for heavy load), indicating that 1 represents the highest load and 0 represents the lowest. Floor / Area Risk Label The normalized value is 0.85 (for high-risk floors), indicating that 1 represents the highest risk area; inspection cycle status. The normalized value is 0.6 (about to expire), indicating that 1 means it has not been maintained for a long time.

[0075] The weight coefficients for each dimension are set as follows: , ,

[0076] The context-weighted score is then:

[0077] Set fusion weights: (Original rating percentage) (Context vector percentage)

[0078] The final fusion-based risk score for this node is then... If a medium-risk level is triggered, the corresponding fault warning mechanism will be implemented, and maintenance personnel will be dispatched for on-site inspection. If necessary, preventive shutdown or load adjustment operations can be performed.

[0079] In summary, this invention performs graph structure modeling on multimodal time-series data collected from building electrical systems and encodes the original observation data into multimodal initial feature vectors for nodes. This enables the system to comprehensively express the physical connections and functional coupling characteristics between electrical nodes while preserving the rich semantic information of the original data, thus enhancing the model's ability to model the overall operating state of the electrical system. This invention utilizes graph neural networks to perform deep feature extraction on the constructed graph structure data and introduces an attention mechanism to enhance the feature perception ability of key nodes and key moments, thereby generating a global feature vector sequence and inputting it into the time-series prediction model for operating state prediction. This mechanism effectively overcomes the problem of insufficient modeling of node heterogeneity in traditional algorithms, enabling early prediction of potential electrical anomalies and achieving proactive identification of electrical hazards. This invention dynamically analyzes the residual sequence between predicted and actual observation values. When an anomaly exceeds a preset threshold, the attention mechanism can be used to deduce key influencing nodes, and further integrate the node's feature information to generate electrical fault points and their risk level assessment results. This effectively improves fault response speed and emergency handling efficiency, and reduces the incidence of safety accidents and system recovery costs.

[0080] Example 2, an embodiment of the present invention, provides a building electrical safety protection system, including: a multimodal data acquisition module for real-time data acquisition from various sensor nodes in the building electrical system; a graph structure construction and node feature encoding module for constructing the acquired multimodal time series data into a node-edge topology graph data according to the physical connection relationship, geographical proximity, and functional coupling of the building electrical system; and performing feature fusion and encoding on the multimodal observation data corresponding to each node to generate an initial multimodal feature vector for graph neural network processing; and a graph neural network analysis module, configuring a graph neural network structure and embedding a multi-head attention mechanism for feature extraction from the constructed graph structure, generating a high-order semantic representation of each node, and generating a global feature vector sequence through a full graph readout operation. The system characterizes the overall operating status of the building's electrical system; the time-series prediction and anomaly detection module includes a trained time-series prediction model that receives a global feature vector sequence generated by a graph neural network to predict the system's operating status at multiple future moments; it also calculates and analyzes the residuals between the predicted and actual observed values ​​to determine if any abnormal states exist; the key node tracing and risk scoring module uses an attention mechanism to backtrack key abnormal nodes, extracting their multi-dimensional time-series features, residual amplitude, attention weight, node importance level, and topological centrality index; and it calculates a comprehensive risk score through weighted fusion to classify risk levels; the risk level classification and emergency response strategy module is used to classify safety levels, low-risk levels, medium-risk levels, and high-risk levels based on the magnitude of the fused risk score, and triggers corresponding emergency strategies according to the corresponding level.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method of electrical safety protection in buildings, characterized in that, The method comprises the following steps: Real-time data acquisition is performed on the building electrical system to obtain multi-modal time series data; The multi-modal time series data is constructed into graph structure data with node-edge topological relationship, and the multi-modal original observation data corresponding to each node is encoded into the multi-modal initial feature vector of the node; Graph neural network feature extraction is performed on the graph structure data, and a global feature vector sequence for representing the running state of the electrical system is obtained by combining an attention mechanism; The global feature vector sequence is input into a pre-constructed time series prediction model for running state prediction, and whether there is a potential anomaly is determined according to the residual sequence between the predicted value and the actual observation value; When the residual exceeds a preset threshold, the key nodes are located by fault backtracking according to the attention weight, and the electrical fault point and the risk level evaluation result are generated in combination with the node features.

2. The method for electrical safety protection of a building according to claim 1, wherein: The multi-modal time series data is constructed into graph structure data with node-edge topological relationship, and the multi-modal original observation data corresponding to each node is encoded into the multi-modal initial feature vector of the node, which comprises the following steps: Each position provided with a sensor in the building electrical system is modeled as a node of a graph, and each node corresponds to a multi-modal original observation data sequence of a specific position; Based on the electrical connection relationship between the nodes, the physical proximity and functional coupling relationship of the spatial geographical position, the edges between the nodes are established to form graph structure data with node-edge topological relationship; An edge weight is set for each edge to represent the similarity, transmission strength or dependence degree between the nodes; The multi-modal original observation data of each node is preprocessed and fused, and the fused multi-modal features are encoded into the multi-modal initial feature vector of the node for subsequent graph neural network feature extraction operation.

3. The method for electrical safety protection of a building according to claim 1, wherein: The graph neural network feature extraction is performed on the graph structure data, and a global feature vector sequence for representing the running state of the electrical system is obtained by combining an attention mechanism, which comprises the following steps: A graph neural network model structure is constructed, and the input and output dimensions are defined; In each layer of graph neural network calculation, a multi-head attention mechanism is introduced to give different attention weights to the features of adjacent nodes to dynamically capture the heterogeneity of the information interaction strength between the nodes at the current time; The intermediate features output by each layer of the graph neural network are concatenated or pooled to form a high-order semantic vector representation of each node to retain the multi-level context information of the node in the graph structure; The high-order semantic vector representation of all nodes in the whole graph is read out globally to generate a global feature vector sequence of uniform dimension for representing the overall running state and potential risk trend of the current building electrical system.

4. The method for electrical safety protection of a building of claim 1, wherein: The global feature vector sequence is input into a pre-constructed time series prediction model for running state prediction, and whether there is a potential anomaly is determined according to the residual sequence between the predicted value and the actual observation value, which comprises the following steps: A time series prediction model is constructed and trained with historical samples; The global feature vector sequence is input into the time series prediction model to generate a system state prediction value sequence for future multiple time points; comparing the predicted value sequence with a corresponding actual observation value sequence in the original multi-modal observation data to calculate a residual sequence; determining whether the residual exceeds a preset anomaly detection threshold, and if any value in the residual sequence is greater than the anomaly detection threshold, determining that the current electrical system has a potential abnormal event.

5. The method for electrical safety protection of a building according to claim 1, wherein: The fault backtracking positioning of the key node according to the attention weight, and the generation of the electrical fault point and the risk level evaluation result combining the node features, include the following steps: Performing time series backtracking analysis based on the graph neural network attention mechanism to locate the key node causing the abnormal event in the graph structure; Analyzing the multi-modal historical feature evolution trajectory of the key node to trace the abnormal cause type; Calculating a comprehensive risk score value combining the multi-dimensional time series backtracking information of the key node, and generating a risk evaluation result.

6. The method for electrical safety protection of a building according to claim 5, wherein: The comprehensive risk score value is obtained by weighted fusion of the multi-dimensional time series backtracking information, which includes residual amplitude, attention weight, node importance level and topological centrality index; The residual amplitude, i.e. according to the residual sequence, extracts the residual value corresponding to the key node as an important measure of abnormal deviation, the greater the residual amplitude, the greater the difference between the current state of the node and the normal value predicted by the model, and the higher the possibility of failure; The attention weight, i.e. according to the attention distribution matrix in the multi-head attention layer, obtains the average attention weight of the current node at the time of abnormal occurrence, which reflects the contribution degree of the node in the information propagation of the whole graph, and the higher the weight, the greater the influence of the node on the overall state of the system; The node importance level, i.e. the nodes in the building electrical system are divided into levels according to their functional criticality in advance, which can be determined based on electrical drawings or expert annotation to reflect the influence of the node on the stability of the power system; The topological centrality index, i.e. based on the constructed graph structure data, calculates the centrality index of each node in the graph, which is used to measure the core degree of the node in the information flow or fault propagation path.

7. The method for electrical safety protection of a building according to claim 6, wherein: The comprehensive risk score value calculation formula is as follows: ； Wherein: represents the comprehensive risk score value of the i-th node, the value range is ;​ representing a residual magnitude normalization value for the node; Indicates the first Normalized attention weights for each node; represents the importance level normalized value of the th node; Indicates the first The normalized value of the centrality index of each node; , , , is the weighted coefficient of each index, satisfying , for reflecting the influence degree of each index on the comprehensive risk.

8. The method for electrical safety protection of a building according to claim 7, wherein: According to the above The value range divides the risk level into four categories, namely, a safe level, a low-risk level, a medium-risk level, and a high-risk level. The division rules and emergency strategies of each level are as follows: When the security level is divided into three levels, the emergency strategy is: the node state is normal, no additional operation is needed, and only routine periodic monitoring is performed; When the node state deviates from the expectation slightly, the emergency strategy is: the node state deviates from the expectation slightly, it is suggested to increase the monitoring frequency of the node or adjust the local device operation parameter to reduce the potential hidden danger; When the risk level is medium, the emergency strategy is: there is an obvious abnormal trend in the node, the fault warning mechanism needs to be executed, and the operation and maintenance personnel need to be dispatched for on-site inspection. When the node is divided into a high-risk level, the emergency strategy is: there is a serious electrical failure or potential safety threat in the node, which needs to trigger the emergency response mechanism immediately, including fault isolation, power protection, linkage alarm and emergency handling process start.

9. The method for electrical safety protection of a building according to claim 8, wherein: When performing the risk level division and emergency strategy formulation, further combine the real-time running state context information of the building electrical system, and adopt a multi-condition joint decision mechanism for comprehensive judgment; The context information includes: the load type and load level of the current building, the power peak period label, the safety level label of the floor and area, and the historical maintenance record and inspection cycle of the related node; In the multi-condition joint decision mechanism, a context information vector is constructed based on the above context information, and is fused with the comprehensive risk score value of each node to obtain a fusion type risk score value under the context perception; Compare the fusion type risk score value with the preset risk level division range, and dynamically re-evaluate the risk level of the node according to the comparison result.

10. A building electrical safety protection system based on the building electrical safety protection method according to any one of claims 1-9, characterized in that: It includes, A multi-modal data acquisition module for real-time data acquisition of each sensor node in the building electrical system; The graph structure construction and node feature coding module is configured to construct the collected multi-modal time series data into graph data of a node-edge topology structure according to the physical connection relationship, geographical proximity and functional coupling of the building electrical system; and to perform feature fusion and coding on the multi-modal observation data corresponding to each node to generate an initial multi-modal feature vector of the node for graph neural network processing. The graph neural network analysis module is configured to embed a multi-head attention mechanism in a graph neural network structure, to perform feature extraction on the constructed graph structure, to generate a high-order semantic representation of each node, and to generate a global feature vector sequence through a full-graph reading operation to represent the overall operation state of the building electrical system. The time series prediction and anomaly detection module includes a trained time series prediction model, receives the global feature vector sequence generated by the graph neural network, predicts the system operation state at multiple future time points, and calculates and analyzes the residual between the predicted value and the actual observation value to determine whether there is an abnormal state. The key node tracing and risk scoring module is configured to trace key abnormal nodes based on an attention mechanism, extract multi-dimensional time series features, residual amplitude, attention weight, node importance level and topological centrality indicators, and calculate a comprehensive risk score value through weighted fusion to divide a risk level. The risk level classification and emergency response strategy module is configured to divide the safety level, low risk level, medium risk level and high risk level according to the size of the fused risk score value, and trigger the corresponding emergency strategy according to the corresponding level.

Images