Early fire early warning method, device, equipment and medium

Abstract

The invention relates to an early fire early warning method, device and equipment and a medium. The method comprises the following steps: acquiring local multi-modal observation data and state confidence of each monitoring point, determining the monitoring point of which the local multi-modal observation data meets a preset exception triggering requirement as an exception monitoring point, and selecting the monitoring points to form a neighborhood reference monitoring point set of the exception monitoring point; updating the static monitoring point network topological graph based on local multi-mode observation data to obtain a dynamic local monitoring point network topological graph; based on the dynamic local monitoring point network topological graph, analyzing space-time correlation characteristics among the monitoring points to obtain a state embedding vector; and based on the state embedding vector and the state confidence, quantifying the fault probabilities of the abnormal monitoring points and the monitoring points, obtaining a fault probability vector, fusing the fault probability vector of each monitoring point, obtaining a fusion credibility vector, and generating fire early warning information. By adopting the method, the timeliness and accuracy of early fire early warning can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fire early warning, and particularly relates to an early fire warning method, device, equipment and medium. BACKGROUND

[0002] With the development of fire early warning technology, an early fire warning system based on discrete sensors appears. The traditional system usually adopts a centralized architecture, and its core logic relies on a single sensor to independently compare the physical signals such as smoke concentration, temperature, carbon monoxide concentration and the like collected by the sensor with a preset static threshold value.

[0003] In the traditional technology, the system processing flow is as follows: each sensor node continuously monitors the local environmental parameters, and once any reading exceeds the independently set threshold value, it is determined to be abnormal. The node triggers a local sound and light alarm, or uploads the original data to a central server through a wireless network. The central server aggregates all data from the entire network, makes a secondary judgment through simple logic rules (such as multiple adjacent nodes alarming at the same time) or additional centralized algorithms, and finally makes a global decision on whether to issue a fire alarm by the central node.

[0004] However, the current way has obvious defects: the judgment mechanism based on a single threshold value cannot effectively distinguish the source of abnormal readings. The sensor reading exceeding the threshold value may be caused by a real fire (especially a slow-developing smoldering fire), or by a sensor fault caused by aging, drift, dust accumulation, or by temporary environmental interference such as kitchen fume, equipment heat source, steam, and the like. The three types of events may appear similar in the time series data of a single sensor, and the traditional system cannot distinguish them, resulting in a dilemma between "high false alarm rate" and "high risk of missed alarm", which is the core pain point of the "fire-fault-disturbance" ternary confusion problem, leading to insufficient timeliness and accuracy of early fire warning. SUMMARY

[0005] Therefore, it is necessary to provide an early fire warning method, device, equipment and medium capable of improving the timeliness and accuracy of early fire warning in view of the above technical problems.

[0006] In a first aspect, the application provides an early fire warning method, comprising:

[0007] acquiring local multi-modal observation data and state confidence of each monitoring point, the local multi-modal observation data comprising a collection time label and a collection location label, and the state confidence comprising a health degree confidence and a local abnormality confidence;

[0008] determining a monitoring point whose local multi-modal observation data meets a preset abnormal trigger requirement as an abnormal monitoring point, and selecting a neighborhood reference monitoring point set of the abnormal monitoring point based on the collection location label;

[0009] updating the preset static monitoring point network topology graph based on the abnormal monitoring point, the set of neighborhood reference monitoring points and the local multi-modal observation data to obtain a dynamic local monitoring point network topology graph;

[0010] analyzing the spatio-temporal correlation features between the monitoring points in the abnormal monitoring point and the set of neighborhood reference monitoring points based on the dynamic local monitoring point network topology graph to obtain a state embedding vector;

[0011] quantifying the fault probability of the abnormal monitoring point and the monitoring points based on the state embedding vector and the state confidence to obtain a fault probability vector; the fault probability vector includes a fire probability component, a monitoring point fault component and an environmental interference probability component;

[0012] fusing the fault probability vectors of the monitoring points to obtain a fusion trust degree vector, and generating fire warning information based on the fusion trust degree vector.

[0013] Further, the static monitoring point network topology graph includes graph nodes, connection edges and edge weights of the connection edges, the graph nodes are the monitoring points, the connection edges are used to represent the connection relationship between the monitoring points, and the edge weights are the proximity degrees between the monitoring points;

[0014] updating the preset static monitoring point network topology graph based on the abnormal monitoring point, the set of neighborhood reference monitoring points and the local multi-modal observation data to obtain a dynamic local monitoring point network topology graph, including:

[0015] extracting the graph nodes and the connection edges from the static monitoring point network topology graph based on the abnormal monitoring point and the set of neighborhood reference monitoring points to form an initial local subgraph;

[0016] normalizing the local multi-modal observation data for each monitoring point in the abnormal monitoring point and the set of neighborhood reference monitoring points to obtain a standardized observation feature vector of each monitoring point;

[0017] adding the standardized observation feature vector as a graph node feature to the graph node corresponding to the abnormal monitoring point or the monitoring point in the initial local subgraph to obtain a dynamic local graph for each monitoring point in the abnormal monitoring point and the set of neighborhood reference monitoring points;

[0018] selecting the local multi-modal observation data with a collection time label meeting the preset monitoring window parameter requirement to form a time sequence feature sequence for each monitoring point in the abnormal monitoring point and the set of neighborhood reference monitoring points, and obtaining the dynamic local monitoring point network topology graph based on the time sequence feature sequence and the dynamic local graph.

[0019] Further, analyzing the spatio-temporal correlation features between the monitoring points in the abnormal monitoring point and the set of neighborhood reference monitoring points based on the dynamic local monitoring point network topology graph to obtain a state embedding vector, including:

[0020] The steps S1 to S5 are repeated until a preset maximum number of iterations is reached, obtaining the state embedding vector of the anomaly monitoring point and the state embedding vector of each monitoring point;

[0021] The steps S1 to S5 are as follows:

[0022] S1, based on a preset linear transformation weight matrix, linearly transforming the state embedding vector to obtain a hidden feature vector of a graph node; wherein when the first S1 step is executed, the state embedding vector is a normalized observation feature vector;

[0023] S2, based on the dynamic local monitoring point network topology graph and the hidden feature vector, calculating the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph;

[0024] S3, for each graph node in the dynamic local monitoring point network topology graph, based on the attention weight, weighting and aggregating the hidden feature vectors of the graph nodes connected to the graph node through the connected edges to obtain a spatial aggregation feature vector;

[0025] S4, for each graph node in the dynamic local monitoring point network topology graph, based on the convolutional neural network, processing the time dynamic feature vector to obtain an updated time dynamic feature vector; when the first S4 step is executed, the time dynamic feature vector is a time series feature sequence;

[0026] S5, for each graph node in the dynamic local monitoring point network topology graph, based on a preset nonlinear activation function, processing the spatial aggregation feature vector, the time dynamic feature vector and the state embedding vector to obtain an updated state embedding vector; wherein when the first S5 step is executed, the state embedding vector is a preset initial state embedding vector.

[0027] Further, based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph is calculated, including:

[0028] Based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph is calculated, and the form of the attention weight between the graph nodes is:

[0029]

[0030] wherein, is the index of any graph node in the dynamic local monitoring point network topology graph, is the hidden feature vector of the graph node connected to the graph node any one of the graph nodes connected by the connection edge, is an initial attention weight between the graph node and the graph node , is an initial attention weight between the graph node and the graph node , is another graph node connected to the graph node in the dynamic local monitoring point network topology graph through a connection edge, is a set of graph nodes connected to the graph node in the dynamic local monitoring point network topology graph through connection edges, is an initial attention weight between the graph node and the graph node ,

[0031] wherein the initial attention weight between the graph node and the graph node is calculated by the following formula:

[0032]

[0033] wherein, is an initial attention weight between the graph node and the graph node , is a nonlinear activation function, is a hidden feature vector of the graph node , is a hidden feature vector of the graph node , is a proximity between the graph node and the graph node , is an attention vector parameter.

[0034] Further, based on the state embedding vector and the state confidence, the failure probability of the abnormal monitoring point and the monitoring point is quantified to obtain a failure probability vector, including:

[0035] For each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set, the state embedding vector is input into a preset basic probability distribution generation network to obtain an initial basic probability distribution; the initial basic probability distribution includes a fire hypothesis initial probability value, a monitoring point failure initial probability value, an environmental disturbance initial probability value and other interference probability values; the basic probability distribution generation network is built based on a multilayer perception mechanism;

[0036] For each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set, the health degree confidence is multiplied by the local abnormal confidence to obtain a comprehensive confidence factor;​

[0037] adjusting the initial basic probability assignment based on the comprehensive confidence factor to obtain a weighted basic probability assignment for each of the abnormal monitoring point and the neighbor reference monitoring point set; the weighted basic probability assignment comprises a fire probability component, a monitoring point failure component and an environmental interference probability component;

[0038] adjusting the other interference probability value based on the comprehensive confidence factor to obtain a calibrated uncertainty probability value, and obtaining a failure probability vector based on the calibrated uncertainty probability value and the weighted basic probability assignment.

[0039] Further, fusing the failure probability vectors of the monitoring points to obtain a fusion confidence vector, and generating fire warning information based on the fusion confidence vector, comprising:

[0040] recursively fusing the failure probability vector of the abnormal monitoring point with the failure probability vectors of the monitoring points in the neighbor reference monitoring point set to obtain a fusion confidence vector; the fusion confidence vector comprises a fusion fire probability component, a fusion monitoring point failure probability component and a fusion environmental interference probability component;

[0041] comparing the fusion monitoring point failure probability component and the fusion environmental interference probability component to obtain a comparison result, and determining a failure confidence component according to the comparison result;

[0042] calculating the difference between the fusion fire probability component and the failure confidence component to obtain an interference confidence difference value;

[0043] when the fusion fire probability component is greater than a preset first threshold value and the interference confidence difference value is greater than a preset second threshold value, generating fire warning information based on the local multi-modal observation data of the abnormal monitoring point and the fusion confidence vector.

[0044] In a second aspect, the present application further provides an early fire warning device, comprising:

[0045] a data acquisition module for acquiring local multi-modal observation data of each monitoring point and state confidence, wherein the local multi-modal observation data comprises a collection time label and a collection location label, and the state confidence comprises a health confidence and a local abnormal confidence;

[0046] an abnormal point analysis module for determining a monitoring point as an abnormal monitoring point when the local multi-modal observation data of the monitoring point meets a preset abnormal trigger requirement, and selecting a monitoring point set as a neighbor reference monitoring point set of the abnormal monitoring point based on the collection location label;

[0047] The connection relationship updating module is configured to update the preset static monitoring point network topology graph based on the abnormal monitoring point, the set of neighborhood reference monitoring points, and the local multi-modal observation data, and obtain a dynamic local monitoring point network topology graph.

[0048] The time sequence feature analysis module is configured to analyze the spatiotemporal correlation features between the abnormal monitoring point and the monitoring points in the set of neighborhood reference monitoring points based on the dynamic local monitoring point network topology graph, and obtain a state embedding vector.

[0049] The fault probability analysis module is configured to quantify the fault probabilities of the abnormal monitoring point and the monitoring points based on the state embedding vector and the state confidence, and obtain a fault probability vector; the fault probability vector includes a fire probability component, a monitoring point fault component, and an environmental interference probability component.

[0050] The early warning information generation module is configured to fuse the fault probability vectors of the monitoring points to obtain a fusion trust degree vector, and generate fire early warning information based on the fusion trust degree vector.

[0051] In a third aspect, the present application further provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements any early fire warning method of the first aspect of the present application when executing the computer program.

[0052] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement any early fire warning method of the first aspect of the present application.

[0053] The early fire warning method, device, equipment and medium described above, by acquiring local multi-modal observation data and state confidence of each monitoring point, the local multi-modal observation data includes acquisition time label and acquisition location label, and the state confidence includes health degree confidence and local anomaly confidence; the monitoring point whose local multi-modal observation data meets the preset anomaly trigger requirement is determined as an anomaly monitoring point, and based on the acquisition location label, a neighborhood reference monitoring point set composed of anomaly monitoring points is selected; based on the anomaly monitoring point, the neighborhood reference monitoring point set and the local multi-modal observation data, the preset static monitoring point network topology graph is updated to obtain a dynamic local monitoring point network topology graph; based on the dynamic local monitoring point network topology graph, the spatio-temporal correlation characteristics between each monitoring point in the anomaly monitoring point and the neighborhood reference monitoring point set are analyzed to obtain a state embedding vector; based on the state embedding vector and the state confidence, the failure probability of the anomaly monitoring point and the monitoring point is quantified to obtain a failure probability vector; the failure probability vector includes a fire probability component, a monitoring point failure component and an environmental interference probability component; the failure probability vectors of each monitoring point are fused to obtain a fusion trust degree vector, and based on the fusion trust degree vector, fire warning information is generated. The traditional single threshold judgment is upgraded to a ternary classification decision based on the spatio-temporal pattern recognition of swarm intelligence, which effectively distinguishes real fire, sensor failure and environmental interference in complex environment, improves the accuracy and reliability of early fire warning, and reduces the dependence on the center node and communication load through local collaborative calculation. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 A flowchart of an early fire warning method provided by an embodiment of the present application;

[0056] Figure 2 A structural diagram of an early fire warning device provided by an embodiment of the present application;

[0057] Figure 3 A structural diagram of a computer device of an early fire warning method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0059] In one embodiment, as shown in Figure 1 Fig. 1, a method for early fire warning is provided. The embodiment is exemplified by the method applied to a terminal. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction of the terminal and the server. In the embodiment, the method includes the following S101-S106, wherein:

[0060] S101, obtaining local multi-modal observation data and state confidence of each monitoring point, the local multi-modal observation data including acquisition time label and acquisition location label, and the state confidence including health confidence and local anomaly confidence.

[0061] Specifically, the terminal obtains the local multi-modal observation data and the state confidence of each monitoring point. For any monitoring point , the mathematical form of the local multi-modal observation data is , is the acquisition time label of the local multi-modal observation data, used to uniquely identify the time when the data is generated; is the acquisition location label, used to represent the observation point of data acquisition and associate with a specific physical location; is the total number of acquisition time labels, is a temperature parameter, in degrees Celsius (℃), collected by a thermocouple or a digital temperature sensor; is a smoke concentration parameter, in percentage concentration (%obs / m), collected by a photoelectric or ionization smoke sensor; is a state confidence carbon monoxide concentration parameter, in parts per million (ppm), collected by an electrochemical sensor. The state confidence is an index for evaluating the reliability of the data of the monitoring point, including health confidence and local anomaly confidence . The health confidence is a long-term statistic, with an initial value of 1. As the monitoring point operates, it is updated periodically with the consistency of its historical readings or neighbor readings. The lower the value, the more likely the monitoring point is to age or fail. The local anomaly confidence is a short-term calculation, calculated based on the deviation of the current observation value from the dynamic baseline (such as the exponentially weighted moving average) maintained by the monitoring point. The higher the value, the more abnormal the current reading.

[0062] S102, determining a monitoring point whose local multi-modal observation data meets a preset anomaly trigger requirement as an anomaly monitoring point, and selecting a neighbor reference monitoring point set of the anomaly monitoring point based on the acquisition location label.

[0063] Specifically, the preset anomaly trigger requirement is defined as the value of any parameter in the local multimodal observation data exceeding its corresponding preset dynamic threshold. The terminal iterates through the local multimodal observation data of all monitoring points. When the value of any parameter in the local multimodal observation data of any monitoring point exceeds its corresponding dynamic threshold, the monitoring point is determined to be an abnormal monitoring point. Based on the collection location label of the abnormal monitoring point, the terminal queries and selects all other monitoring points whose physical distance from the abnormal monitoring point is less than or equal to a preset influence radius from the preset monitoring point location database, forming a neighborhood reference monitoring point set for the abnormal monitoring point. The preset dynamic threshold is calculated based on the historical local multimodal observation data of each monitoring point. For any parameter in the local multimodal observation data of any monitoring point (which can be a temperature parameter, smoke concentration parameter, or carbon monoxide concentration parameter), the terminal obtains the historical mean and standard deviation of that parameter for the monitoring point and calculates it using the following formula: The dynamic threshold of this parameter at the monitoring point. It is the index of any monitoring point. It is an index to any parameter, such as temperature, smoke concentration, or carbon monoxide concentration. It is a monitoring point Parameters obtained Historical average, It is a monitoring point Parameters obtained Historical standard deviation It is a monitoring point Medium parameters The dynamic threshold. The preset monitoring point location database can be compiled based on the data collection location tags of each monitoring point and the actual physical records between monitoring points during actual work. The preset influence radius can be set according to actual work requirements.

[0064] S103. Based on the abnormal monitoring points, the neighboring reference monitoring point set, and the local multimodal observation data, update the preset static monitoring point network topology map to obtain the dynamic local monitoring point network topology map.

[0065] Specifically, the preset static monitoring point network topology is a graph data structure used to represent the connection relationships of all sensor nodes in the entire monitoring area, and its mathematical form is: Among them, vertex set Each graph node in Corresponding to a unique monitoring point And record the collection location labels. Edge set Each connecting edge in Used to characterize monitoring points and There are pre-defined connection relationships between them, based on the proximity of communication links or physical locations during deployment. Weight matrix elements in It is a connecting edge The edge weights are used to quantify the proximity between two monitoring points, specifically using the following formula: , calculated to be. Where It is a monitoring point and The physical distance between them can be obtained from the aforementioned preset monitoring point location database. Based on this preset static monitoring point network topology and the identified abnormal monitoring points and their neighboring reference monitoring point sets, the terminal updates the static monitoring point network topology to obtain a dynamic local monitoring point network topology. The dynamic local monitoring point network topology is mathematically represented as... Among them, vertex set Each graph node in A unique monitoring point in the set of corresponding anomaly monitoring points or neighboring reference monitoring points. edge set Each connecting edge in Used to characterize monitoring points and There are pre-defined connection relationships between them, based on the proximity of communication links or physical locations during deployment. Weight matrix elements in It is a connecting edge The edge weights are used to quantify the proximity between two monitoring points, and each graph node also has graph node features, which are obtained by normalizing the local multimodal observation data (especially temperature, smoke, and carbon monoxide parameters) of each monitoring point.

[0066] S104. Based on the dynamic local monitoring point network topology, the spatiotemporal correlation characteristics between abnormal monitoring points and monitoring points in the neighborhood reference monitoring point set are analyzed to obtain the state embedding vector.

[0067] Specifically, the terminal takes the dynamic local monitoring point network topology graph obtained in the last step as input, and processes it using a graph neural network model capable of capturing spatial correlation and temporal evolution. Through multiple rounds of iterative calculations, the model interacts and fuses the original observation features of each graph node (i.e., monitoring point) with the features of its neighbor nodes, while analyzing the temporal variation pattern of its own observation value. In this process, the model learns and highlights abnormal patterns that exhibit gradient diffusion in space and continuous evolution in time (which are usually characteristics of real fires), while weakening isolated, abrupt, or random abnormal patterns (which may be characteristics of faults or local disturbances). After model processing, each monitoring point is transformed (embedded) into a fixed-length, high-dimensional numerical vector, i.e., a state embedding vector. This vector is a deep feature representation that encodes the abnormal state of the monitoring point itself, the context state information of its neighbors, and the temporal dynamics of the state. It is a key high-order feature extracted from the original data for subsequent root cause differentiation. The mathematical form of the state embedding vector is , where is the state embedding vector of monitoring point, and is the state embedding vector of monitoring point.

[0068] S105, based on the state embedding vector and the state confidence, quantifying the fault probability of the abnormal monitoring point and the monitoring points to obtain a fault probability vector; the fault probability vector includes a fire probability component, a monitoring point fault component, and an environmental disturbance probability component.

[0069] Specifically, the terminal obtains the state embedding vector and the state confidence of each monitoring point, analyzes each type of fault for each monitoring point, and obtains a fault probability vector including a fire probability component, a monitoring point fault component, and an environmental disturbance probability component. The mathematical form of the fault probability vector is , where is the fault probability vector of monitoring point, and is the fault probability vector of monitoring point. is the fire probability component, used to represent the probability of monitoring point detecting a fire; is the monitoring point fault component, used to represent the probability of monitoring point having a fault; is the environmental disturbance probability component, used to represent the probability of monitoring point being misjudged by environmental disturbance.

[0070] ​​​​​​​S106, fuse the fault probability vectors of each monitoring point to obtain a fusion trust degree vector, and generate fire warning information based on the fusion trust degree vector.

[0071] Specifically, the terminal fuses the fault probability vectors of the abnormal monitoring point and all its neighboring reference monitoring points to obtain a fusion trust degree vector, and the mathematical form of the fusion trust degree vector is , is a fusion fire probability component, used to represent the overall trust degree after fusion for the existence of fire; is a fusion monitoring point fault probability component, used to represent the overall trust degree after fusion for the existence of monitoring point fault; is a fusion environmental interference probability component, used to represent the overall trust degree after fusion for the existence of environmental interference leading to misjudgment. The terminal compares the fusion result with a preset judgment threshold to obtain a comparison result, and integrates the local multi-modal observation data of the abnormal monitoring point and the fusion trust degree vector according to the comparison result to obtain fire warning information. The preset judgment threshold is used to determine the situation of the fusion fire probability component, the fusion monitoring point fault probability component and the fusion environmental interference probability component when there is fire, which can be set according to actual work.

[0072] The early fire warning method provided in this embodiment can accurately locate the abnormal monitoring point from all network monitoring points based on a dynamic threshold and construct a local reference neighbor set thereof; and can construct a dynamic local network graph rich in space-time information by combining a static network topology with dynamic multi-modal observation data; can also use a graph neural network model to perform deep analysis on the local network to extract state embedding vectors of each monitoring point that can comprehensively reflect the node's own state, spatial context and time sequence evolution; can also combine the health and abnormal confidence of the sensor to convert high-order features into fault probability vectors that quantitatively represent the possibility of different fault sources; and can fuse the probability evidence from multiple monitoring points to form a fusion trust degree of group consensus, and make intelligent decisions based on the fusion trust degree to generate fire warning information. The traditional single threshold judgment is upgraded to a ternary classification decision based on group intelligent space-time pattern recognition, which effectively distinguishes real fire, sensor fault and environmental interference in a complex environment, improves the accuracy and reliability of early fire warning, and reduces the dependence on the center node and communication load through local collaborative calculation.

[0073] In one of the embodiments, the static monitoring point network topology graph includes graph nodes, connection edges and edge weights of the connection edges, the graph nodes are each monitoring point, the connection edges are used to represent the connection relationship between each monitoring point, and the edge weight is the proximity between each monitoring point;

[0074] Based on anomaly monitoring points, neighboring reference monitoring point sets, and local multimodal observation data, the preset static monitoring point network topology is updated to obtain a dynamic local monitoring point network topology, including:

[0075] S201. Based on the abnormal monitoring points and the set of neighboring reference monitoring points, graph nodes and connecting edges are extracted from the static monitoring point network topology graph to form an initial local subgraph.

[0076] Specifically, the terminal has acquired a preset static monitoring point network topology map. For the identified anomaly monitoring points: , and its neighborhood reference monitoring point set are The terminal uses these monitoring points in the vertex set. Find the corresponding graph node set Subsequently, in the border town Find all connections in the middle. An edge connecting any two nodes in a graph forms a subset of edges. Meanwhile, from the weight matrix Extract the edge weights corresponding to these edges. , The graph structure formed by the graph and its corresponding weights This is the initial local subgraph. This subgraph inherits information from the global static topology regarding the preset connection relationships and proximity between these monitoring points.

[0077] S202, for each monitoring point in the set of abnormal monitoring points and neighboring reference monitoring points, normalize the local multimodal observation data to obtain the standardized observation feature vector of each monitoring point.

[0078] Specifically, for each monitoring point in the initial local subgraph The terminal obtains the latest local multimodal observation vector based on its acquisition time label. The terminal uses preset global statistical parameters for Z-score normalization. These preset global statistical parameters include: global temperature mean. and standard deviation Global average smoke concentration and standard deviation Global average carbon monoxide concentration and standard deviation These parameters are obtained by collecting historical data from all monitoring points across the entire network during the initial deployment phase or long-term normal operation. The standardized calculation formula is: , , Standardized observation vector This is the standardized observation feature vector of the monitoring point.

[0079] S203, for each of the abnormal monitoring point and the neighbor reference monitoring point set, the standardized observation feature vector is taken as the graph node feature, added to the initial local subgraph, and the graph node corresponding to the abnormal monitoring point or the monitoring point is obtained. Dynamic local graph.

[0080] Specifically, the terminal takes the standardized observation feature vector of each monitoring point calculated in the last step as the dynamic feature data, and assigns it to the corresponding graph node in the initial local subgraph . Specifically, a graph node feature is created for each graph node , and the vector is stored in the attribute. After this operation, the initial local subgraph which only contains static connection information is injected with feature information reflecting the latest real-time state of each monitoring point, thereby transforming into a dynamic local graph containing both topology structure and node state features.

[0081] S204, for each of the abnormal monitoring point and the neighbor reference monitoring point set, the local multi-modal observation data with the collection time label meeting the preset monitoring window parameter requirement is selected to form a time sequence feature sequence, and based on the time sequence feature sequence and the dynamic local graph, a dynamic local monitoring point network topology graph is obtained.

[0082] Specifically, the preset monitoring window length parameter is used to represent the length of the data time traceability of the early fire warning, which can be set according to actual work. For each graph node in the dynamic local graph (corresponding to the monitoring point ), the terminal selects all local multi-modal observation data with the collection time label in the interval , is the preset monitoring window parameter, is any collection time label. The terminal arranges these data in chronological order, and also performs the normalization processing of the observation vector at each time point as described in S202, to obtain a standardized observation feature sequence with a length of : This sequence describes the dynamic change process of each physical quantity of the monitoring point in the recent period of time. The terminal associates the time sequence feature sequence as the time sequence attribute of the graph node to the corresponding graph node of the dynamic local graph . A dynamic local monitoring point network topology graph containing static connection topology, dynamic node instantaneous feature and dynamic node time sequence is obtained.

[0083] The early fire warning method provided by the embodiment ensures the relevance of the analysis range by accurately cutting out the local structure skeleton (initial local subgraph) related to the abnormal event from the global topology; then, the dimensional differences are eliminated by performing global statistical-based normalization processing on the multi-modal raw data, and a comparable standard feature vector is generated, thereby providing a unified input scale for subsequent model analysis; the instantaneous features after normalization are also assigned to the corresponding nodes in the graph, so that the static "point" has a dynamic "state"; and the time dimension information is introduced by associating each node with the time sequence composed of its recent historical data according to the preset monitoring window parameter. The graph structure thus constructed organically integrates the three types of key information, i.e., spatial connection relationship, node real-time state and state time evolution, to obtain a dynamic local monitoring point network topology graph. This lays a solid data foundation for subsequent deep mining of the spatiotemporal propagation and correlation pattern of abnormal events in the local network.

[0084] In one of the embodiments, based on the dynamic local monitoring point network topology graph, the spatiotemporal correlation features between the abnormal monitoring points and the neighbor reference monitoring points are analyzed to obtain state embedding vectors, including:

[0085] The steps S1 to S5 are repeated until a preset maximum number of iterations is reached to obtain the state embedding vector of the abnormal monitoring point and the state embedding vectors of the monitoring points.

[0086] Specifically, the terminal repeats the steps S1 to S5 for each graph node in the dynamic local monitoring point network topology graph until a preset maximum number of iterations is reached to obtain the state embedding vector of the abnormal monitoring point and the state embedding vectors of the monitoring points. The preset maximum number of iterations can be set according to the problem complexity and the demand for computing resources in actual work, and is set to 4 by default. The preset maximum number of iterations is not limited in the embodiment.

[0087] The steps S1 to S5 are as follows:

[0088] S1, performing linear transformation on the state embedding vector based on a preset linear transformation weight matrix to obtain a hidden feature vector of the graph node; wherein, when the step S1 is performed for the first time, the state embedding vector is the normalized observation feature vector.

[0089] Specifically, when performing the first cycle processing, for each graph node in the dynamic local monitoring point network topology graph , the terminal reads the graph node features, i.e., the normalized observation feature vector. Specifically, the terminal calls the preset linear transformation weight matrix . Wherein, is the dimension of the input feature vector, is the dimension of the hidden layer, is the index of the current cycle. Through the matrix multiplication operation: , the input features of each node are mapped to a new high-dimensional space to obtain the hidden feature vector of the node . For the subsequent cycle processing, for each graph node in the dynamic local monitoring point network topology graph, the terminal reads the updated state embedding vector obtained in the last cycle processing, and calls the preset linear transformation weight matrix , where is the dimension of the input feature vector, is the dimension of the hidden layer, is the index of the current cycle. Through the matrix multiplication operation: , the input features of each node are mapped to a new high-dimensional space to obtain the hidden feature vector of the node , is the updated state embedding vector of the node obtained in the last cycle processing. The preset linear transformation weight matrix is learned according to the model training process in actual work, by inputting the labeled historical training samples (i.e. the dynamic local monitoring point network topology graph and its true root label corresponding to the historical abnormal event) into the model. In the calculation of each round , all nodes share and use the same matrix for linear transformation, and according to the difference between the final output of the model (i.e. the prediction of the event root) and the true label, the loss function is calculated, and the gradient of the loss to all parameters in the model is calculated through the back propagation algorithm, and the gradient descent optimizer is used to optimize according to the calculated gradient.

[0090] S2, based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph is calculated.

[0091] Specifically, after obtaining the hidden feature vectors of all nodes , the attention weight between any two nodes connected by a connection edge and in the graph is calculated. The attention weight quantifies the importance of the state embedding vector update of the adjacent node to the center node in this cycle.

[0092] ​S3, for each graph node in the dynamic local monitoring point network topology graph, based on the attention weight, the hidden feature vectors of the graph nodes connected with the graph node through the connection edge are weighted and aggregated to obtain a spatial aggregation feature vector.

[0093] Specifically, for each graph node in the dynamic local monitoring point network topology graph , the terminal has obtained the attention weight between it and all neighbor nodes , and the hidden feature vectors of all neighbor nodes . Using the following formula: , spatial aggregation is performed to obtain the spatial aggregation feature vector of the graph node , wherein is the spatial aggregation feature vector of node obtained at the cycle round , which represents the summary of the states of all neighbor nodes in the local neighborhood around node , and the contribution of the neighbor is dynamically determined by the attention weight. If the reading of a certain neighbor node is highly correlated with the abnormal mode of the center node, the attention weight of the neighbor node may be higher, so as to dominate in the aggregation vector. This step makes the features of each node integrate the local spatial context information of the network. is the set of graph nodes connected with the graph node through the connection edge in the dynamic local monitoring point network topology graph.

[0094] S4, for each graph node in the dynamic local monitoring point network topology graph, based on the convolutional neural network, a time dynamic feature vector is processed to obtain an updated time dynamic feature vector; when the first S4 step is performed, the time dynamic feature vector is the time sequence feature sequence.

[0095] Specifically, when performing the first cycle processing, for each graph node in the dynamic local monitoring point network topology graph , the terminal reads its associated time sequence feature sequence , and a preset one-dimensional convolutional neural network is used to process the sequence. The one-dimensional convolutional neural network includes one or more preset one-dimensional convolution kernel parameters . The convolution kernel slides along the time dimension on the sequence to perform convolution operation: at each time step, the features in the local time period covered by the convolution kernel are weighted and combined, and the pattern features in the local time period are extracted through a nonlinear activation function. By stacking multiple convolution layers or using convolution kernels of different sizes, the one-dimensional convolutional neural network can capture dynamic patterns at different time scales from short-term fluctuations to long-term trends. After the one-dimensional convolutional neural network processing, the entire time sequence is encoded into a fixed-length vector , the time dynamic feature vector. For the subsequent cycle processing, for each graph node in the dynamic local monitoring point network topology , the terminal reads the updated time dynamic feature vector obtained by the last cycle processing, and processes it by using the preset one-dimensional convolutional neural network to obtain the updated time dynamic feature vector obtained by the current processing round of the graph node . The preset one-dimensional convolutional neural network is trained using the labeled historical training samples (i.e., the dynamic local monitoring point network topology corresponding to the historical abnormal event and the real root label).

[0096] S5, for each graph node in the dynamic local monitoring point network topology, the spatial aggregation feature vector, the time dynamic feature vector and the state embedding vector are processed based on the preset nonlinear activation function to obtain the updated state embedding vector; wherein when the first S5 step is executed, the state embedding vector is a preset initial state embedding vector.

[0097] Specifically, the preset initial state embedding vector is a zero vector. When performing the first cycle processing, for each graph node in the dynamic local monitoring point network topology , the terminal splices the spatial aggregation feature vector, the time dynamic feature vector and the preset initial state embedding vector of the node to obtain the joint feature vector of the current processing round, and inputs the joint feature vector into a preset fully connected layer and processes it through a nonlinear activation function ReLU: . and are the weight and bias parameters of the fully connected layer. The calculated is the updated state embedding vector of the graph node in the first cycle. It integrates the historical state of the node itself, the current spatial context and the time evolution mode, and is a higher-level representation than the input feature. For the subsequent cycle processing, for each graph node in the dynamic local monitoring point network topology , the terminal splices the spatial aggregation feature vector, the time dynamic feature vector and the initial state embedding vector updated in the last round of the node obtained in the current processing round to obtain the joint feature vector of the current processing round, and inputs the joint feature vector into a preset fully connected layer and processes it through a nonlinear activation function ReLU: . is the spatial aggregation feature vector of the node obtained in the cycle round , is the calculated time dynamic feature vector of the node obtained in the cycle round , is a graph node at the processing round . The preset full connection layer is trained using the labeled historical training samples, i.e., the dynamic local monitoring point network topology graph and the real root label corresponding to the historical abnormal event.

[0098] The early fire warning method provided in the embodiment realizes deep spatiotemporal feature analysis of the dynamic local monitoring point network topology graph by designing an iterative loop framework including S1 to S5 steps. Through multiple rounds of "feature transformation-information aggregation-fusion update", three-way information from space, time, and history is nonlinearly fused to generate a more powerful state embedding vector. After multiple iterations, information is transmitted and interacted between graph nodes multiple times, and the final state embedding vector of each node deeply encodes the complex correlation pattern of itself and its multi-hop neighbors at multiple time steps. It is crucial to distinguish fire (with spatiotemporal diffusion), sensor failure (usually spatially isolated), and environmental interference (possibly spatially local but temporally random), providing reliable high-order feature input data for subsequent accurate probability quantification.

[0099] In one of the embodiments, based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph is calculated, including:

[0100] S401, based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connected edge in the dynamic local monitoring point network topology graph is calculated, and the form of the attention weight between the graph nodes is:

[0101]

[0102] wherein, is the index of any graph node in the dynamic local monitoring point network topology graph, is any graph node connected to the graph node in the dynamic local monitoring point network topology graph through a connected edge, is the attention weight between the graph node and the graph node , is the initial attention weight between the graph node and the graph node , is another graph node connected to the graph node in the dynamic local monitoring point network topology graph through a connected edge, is another graph node connected to the graph node The set of connected graph nodes Graph nodes Graph Nodes Initial attention weights between;

[0103] Among them, graph nodes Graph Nodes Initial attention weights between It is calculated using the following formula:

[0104]

[0105] in, Graph nodes Graph Nodes Initial attention weights between them It is a non-linear activation function. Graph nodes The hidden feature vectors, Graph nodes The hidden feature vectors, Graph nodes Graph Nodes The degree of proximity between them These are the attention vector parameters.

[0106] Specifically, the terminal uses a formula to calculate the attention weight between any two graph nodes connected by a connecting edge in the dynamic local monitoring point network topology graph. Wherein, the graph node... Hidden feature vectors Graph nodes Hidden feature vectors This is obtained through the aforementioned S1. (Graph node) Graph Nodes Proximity between It is obtained by calculating the edge weights of the connecting edges in the dynamic local monitoring point network topology graph. (Nonlinear activation function) It is the LeakyReLU function. (Graph node) Graph Nodes Initial attention weights between Characterizes the situation where graph nodes are not considered. In the case of other neighbors, one of the neighbor nodes For graph nodes The fundamental importance of attention vector parameters. It is trained using pre-labeled historical training samples (i.e., the dynamic local monitoring point network topology map corresponding to historical anomalous events and their true root source labels). Graph nodes Graph Nodes Weight of intention between The numerical range of the value of the attention weight is [0, 1]. In the dynamic local monitoring point network topology graph, the graph nodes connected by the connection edges The set of graph nodes connected The set of graph nodes connected by the connection edges in the dynamic local monitoring point network topology graph The set of graph nodes connected by the connection edges in the dynamic local monitoring point network topology graph

[0107] The early fire warning method provided by the embodiment provides an interpretable and learnable dynamic weighting mechanism for spatial information aggregation in the graph neural network by using the above-defined attention weight calculation formula. When the model learns the spatiotemporal correlation pattern, it tends to pay more attention to node pairs that are related in features and adjacent in space, which is exactly consistent with the law of the diffusion of physical quantities such as heat and smoke along the spatial gradient in a real fire. The effective extraction of the spatiotemporal correlation features consistent with the physical law is realized, and different types of abnormal events are successfully distinguished.

[0108] In one of the embodiments, based on the state embedding vector and the state confidence, the failure probability of the abnormal monitoring point and the monitoring point is quantified to obtain a failure probability vector, including:

[0109] S501, for each monitoring point in the abnormal monitoring point and the neighbor reference monitoring point set, input the state embedding vector into a preset basic probability assignment generation network to obtain an initial basic probability assignment; the initial basic probability assignment includes a fire hypothesis initial probability value, a monitoring point failure initial probability value, an environmental interference initial probability value, and other interference probability values; the basic probability assignment generation network is built based on a multilayer perception mechanism.

[0110] Specifically, the preset basic probability assignment generation network is a lightweight neural network based on a multilayer perception mechanism architecture, and its structure is fixed after training. The dimension of the input layer of the basic probability assignment generation network is the same as that of the state embedding vector. The network internally contains one or more hidden layers, each layer is composed of a fully connected unit and a nonlinear activation function (such as ReLU), which is used to learn the complex nonlinear mapping from high-level spatiotemporal features to probability distribution. The output layer of the network is usually a fully connected layer with 4 neurons and using a Softmax activation function. For each monitoring point in the abnormal monitoring point and the neighbor reference monitoring point set, the terminal inputs the state embedding vector thereof into the preset basic probability assignment generation network, and the preset basic probability assignment generation network outputs a 4-dimensional vector . Wherein, is the index of any monitoring point, which can be the abnormal monitoring point or any monitoring point in the neighbor reference monitoring point set, is the initial basic probability assignment of the monitoring point . is the initial basic probability assignment of the monitoring point is the fire hypothesis initial probability value of the monitoring point, representing the initial probability mass that the preset basic probability assignment generating network preliminarily judges that the abnormal event belongs to the fire based on the state embedding vector of the monitoring point; is the monitoring point is the monitoring point fault initial probability value of the monitoring point, representing the initial probability mass that the preset basic probability assignment generating network preliminarily judges that the abnormal event belongs to the self-fault of the monitoring point based on the state embedding vector of the monitoring point; is the monitoring point is the environmental interference initial probability value of the monitoring point, representing the initial probability mass that the preset basic probability assignment generating network preliminarily judges that the abnormal event belongs to the environmental interference based on the state embedding vector of the monitoring point; is the monitoring point is the other interference probability value of the monitoring point, representing the initial probability mass that the preset basic probability assignment generating network cannot determine the specific category and assigns to "uncertain". The four values satisfy the probability sum of 1: The preset basic probability assignment generating network is obtained by training in a federated learning paradigm. In the federated learning framework, a central server initializes a global MLP model. Multiple participants (such as early warning systems of different buildings or different regions) train the received global model locally using their own historical data set (containing a large number of state embedding vector samples and their corresponding real event category labels), but the training data is always kept locally and not uploaded. Each participant only uploads the encrypted update (gradient) of the model parameters to the server. The server aggregates these updates to generate a new global model and distributes it again. After multiple iterations, a shared model with superior performance and protection of the privacy of each party is finally obtained as the preset basic probability assignment generating network.

[0111] S502, for each monitoring point in the set of abnormal monitoring points and neighborhood reference monitoring points, multiply the health confidence and the local abnormal confidence to obtain a comprehensive confidence factor.

[0112] Specifically, for each monitoring point in the set of abnormal monitoring points and neighborhood reference monitoring points, the terminal calculates the product of the health confidence and the local abnormal confidence: to obtain the comprehensive confidence factor of the monitoring point . Wherein, is the index of any monitoring point, which can be any monitoring point in the set of abnormal monitoring points or neighborhood reference monitoring points, is the health confidence of the monitoring point , is the local abnormal confidence of the monitoring point , is the local abnormal confidence of the monitoring point , is the local abnormal confidence of the monitoring point is the local abnormal confidence of the monitoring point The overall reliability factor. This factor comprehensively reflects both "whether the sensor's hardware is reliable" and "whether its current reading is truly abnormal." An abnormal reading ( (High) But the sensor itself may have a problem. The overall credibility of nodes with low (low) It will be reduced.

[0113] S503, for each monitoring point in the abnormal monitoring point and the neighboring reference monitoring point set, adjusts the initial basic probability allocation based on the comprehensive confidence factor to obtain a weighted basic probability allocation; the weighted basic probability allocation includes fire probability component, monitoring point failure component and environmental interference probability component.

[0114] Specifically, for each monitoring point in the abnormal monitoring point and the neighboring reference monitoring point set, the terminal uses the calculated comprehensive confidence factor to adjust the initial probability values ​​of the fire assumption, the monitoring point failure, and the environmental interference in its initial basic probability allocation. The specific operation is as follows: , , .in It is the index of any monitoring point, which can be an anomaly monitoring point or any monitoring point in the neighborhood reference monitoring point set. It is a monitoring point The initial probability value of the fire assumption. It is a monitoring point The initial probability value of the monitoring point failure. It is a monitoring point Initial probability value of environmental disturbance, It is a monitoring point The overall credibility factor It is a monitoring point The fire probability component, It is a monitoring point The fault component of the monitoring point It is a monitoring point The environmental interference probability component. This means that if the overall reliability of a monitoring point is low (e.g., poor sensor health), then even if the initial fire probability calculated by the network from its state embedding vector is high, this judgment "component" will be significantly weakened. The terminal integrates the fire probability component, monitoring point fault component, and environmental interference probability component of the same monitoring point to obtain the weighted basic probability allocation for that monitoring point.

[0115] S504, for each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set, adjusts other interference probability values ​​based on the comprehensive confidence factor to obtain the calibration uncertainty probability value, and obtains the fault probability vector based on the calibration uncertainty probability value and the weighted basic probability allocation.

[0116] Specifically, for each monitoring point in the set of abnormal monitoring points and neighborhood reference monitoring points, the terminal uses the following formula: According to the comprehensive credibility factor thereof, the other interference probability value is adjusted to obtain the calibrated uncertainty probability value of the monitoring point, and the calibrated uncertainty probability value of the same monitoring point and the weighted basic probability assignment are integrated to obtain the fault probability vector of the monitoring point. Wherein, is the index of any monitoring point, which can be any monitoring point in the set of abnormal monitoring points or neighborhood reference monitoring points, is the other interference probability value of the monitoring point is the comprehensive credibility factor of the monitoring point is the calibrated uncertainty probability value of the monitoring point

[0117] The early fire warning method provided in the embodiment realizes end-to-end conversion from feature space to probability space by directly mapping high-dimensional features representing the space-time context of nodes into preliminary probability judgments of three hypotheses through an intelligent network obtained through federated learning. The concept of comprehensive credibility factor is introduced, which combines the reliability of the sensor itself and the abnormality degree of the current reading as a unified adjustment coefficient. The coefficient is used to weight the preliminary judgment, so that the evidence from high-credibility nodes is strengthened, and the evidence from low-credibility nodes is weakened, and through recalibration of uncertainty (S504), the integrity of the probability vector is ensured, and the suspicion degree increased due to low sensor credibility is quantified into the uncertainty component, to obtain the fault probability vector of each monitoring point. The whole process deeply integrates data-driven artificial intelligence judgment and hardware-based reliability evaluation, generating a composite evidence body containing both intelligent inference and confidence evaluation, providing high-quality and high-information input data for subsequent group consensus decision-making based on evidence theory.

[0118] In one of the embodiments, the fault probability vectors of the monitoring points are fused to obtain a fusion trust degree vector, and based on the fusion trust degree vector, fire warning information is generated, including:

[0119] S601, the fault probability vector of the abnormal monitoring point is recursively fused with the fault probability vectors of the monitoring points in the set of neighborhood reference monitoring points to obtain a fusion trust degree vector; the fusion trust degree vector includes a fusion fire probability component, a fusion monitoring point fault probability component and a fusion environmental interference probability component.

[0120] Specifically, the terminal obtains the fault probability vector of the abnormal monitoring point and the fault probability vectors of the monitoring points in the set of neighborhood reference monitoring points, and combines them in order into a fault probability vector set:​​​ where is the failure probability vector of the abnormal monitoring point, and the rest is the failure probability vector of each monitoring point in the neighborhood reference monitoring point set. The terminal fuses the first two vectors and according to the Dempster combination rule, which is specifically designed to handle probability assignments with uncertainty. The fusion process calculates the joint support of two evidences for all possible propositions (single-point hypotheses such as and compound hypotheses such as , and mainly focuses on single-point hypotheses in this scheme), and specifically handles the conflicting parts between the two evidences (for example, one strongly supports fire and the other strongly supports failure), and resolves conflicts and enhances consensus through a normalization factor. After fusion, a new intermediate probability vector representing the joint opinion of the two nodes is generated. Then, the same pairwise fusion is performed on the intermediate vector and the next vector to obtain . This recursion is performed until all failure probability vectors are fused to obtain the final fused basic probability assignment vector . The terminal extracts the first three components from it to form the fused trust degree vector . Among them, is the fused trust degree vector, is the fused fire probability component, which is used to represent the overall trust degree of the fire hypothesis after comprehensive consideration of the evidence provided by the entire local sensor and its uncertainty; is the fused monitoring point failure probability component, which is used to represent the overall trust degree of the monitoring point failure hypothesis after comprehensive consideration of the evidence provided by the entire local sensor and its uncertainty;is the fusion environmental disturbance probability component, which is used to represent the overall trust in environmental disturbance after comprehensive consideration of the evidence provided by the entire local sensor and its uncertainty. Among them, the Dempster combination rule is a mathematical evidence synthesis method for dealing with uncertain and conflicting information, which is the core component of Dempster-Shafer evidence theory. This rule is used to fuse two or more independent "basic probability assignment" functions (each function represents the trust degree of different propositions assigned by each evidence source, including the quantification of uncertainty) into a new and comprehensive basic probability assignment function. This rule works in the following way: when fusing two evidences, it multiplies and sums up the evidence strengths supporting the same conclusion (or set of conclusions), thereby enhancing consensus; at the same time, it specially calculates the completely contradictory part between the two evidences (i.e. one supports A and the other supports B, and A and B have no intersection), and this part of the conflict is taken as a normalization factor. In the final fusion result, the consensus part is enhanced, while the conflict part is proportionally distributed to all non-conflicting propositions, or converted into overall uncertainty. In this way, the fused result not only reflects the common direction of multiple evidences, but also quantifies and handles the conflict between evidences in a mathematical way, thereby obtaining a more robust and reliable group trust degree evaluation than any single evidence.

[0121] S602, compare the fusion monitoring point failure probability component and the fusion environmental disturbance probability component to obtain a comparison result, and determine the failure trust degree component according to the comparison result.

[0122] Specifically, the terminal obtains the fusion trust degree vector After that, it is necessary to determine which of the two possibilities of non-fire is the more major disturbance source currently considered by the group, so as to clearly compare with the fire trust degree in the subsequent. The terminal compares the size of (fusion monitoring point failure probability component) and (fusion environmental disturbance probability component). The comparison result is a comparison result: or According to this comparison result, the terminal determines the failure trust degree component : if , let , which means that the group considers that the possibility of sensor failure is not lower than environmental disturbance; if , let , which means that the group considers that environmental disturbance is the more major non-fire reason. represents the trust degree of the group to the "most major non-fire abnormal reason".

[0123] S603, calculate the difference between the fusion fire probability component and the failure trust degree component to obtain the disturbance trust degree difference.

[0124] Specifically, the terminal calculates the fire trust degree in the fusion trust degree vector The difference between the main non-fire trust degree determined in the previous step The difference between the main non-fire trust degree determined in the previous step This interference trust degree difference is a key decision indicator, which quantifies the trust advantage of the "fire hypothesis" over the "most likely adverse competing hypothesis". A large positive difference indicates that the fire hypothesis is not only highly credible itself, but also significantly better than other possibilities; a small difference or negative value indicates that the advantage of the fire hypothesis is not obvious, or even less credible than other possibilities.

[0125] S604, when the fusion fire probability component is greater than the preset first threshold and the interference trust degree difference is greater than the preset second threshold, generating fire warning information based on the local multi-modal observation data of the abnormal monitoring point and the fusion trust degree vector.

[0126] Specifically, the preset first threshold is used to set the minimum fusion fire probability component requirement for the "fire hypothesis", ensuring that only when the trust degree for fire in the fusion group consensus reaches a sufficiently high basic level, the preliminary qualification for triggering the alarm is met, thereby filtering out those with weak or ambiguous overall evidence. The preset second threshold is used to set the interference trust degree difference requirement of the "fire hypothesis" relative to the main competing hypothesis, ensuring that the fire trust degree is not only high, but also significantly higher than the most likely non-fire reason (sensor failure or environmental interference), thereby effectively excluding those confusing scenarios where multiple possibilities coexist and are difficult to distinguish in decision-making, preventing false positives. The first threshold is set to 0.6 and the second threshold is set to 0.2 by default according to the historical false positive rate, false negative rate risk trade-off and actual work safety requirements. The terminal compares the fusion fire probability component with the preset first threshold and compares the interference trust degree difference with the preset second threshold. When the fusion fire probability component is greater than the preset first threshold and the interference trust degree difference is greater than the preset second threshold, the terminal converts the local multi-modal observation data of the abnormal monitoring point and the fusion trust degree vector according to the preset language organization rule to obtain structured fire warning information. The preset language organization rule can be set according to the length of the local multi-modal observation data and the format of the fire warning information in actual work.

[0127] The early fire warning method provided by the embodiment generates fire warning information through sequentially performing recursive evidence fusion, finding out the most important competitive hypothesis through comparison, calculating the difference to quantify the relative advantage of the fire hypothesis, and judging the double threshold condition. The problem of "three-element confusion" is fully solved. Through deep fusion and comparison of distributed evidence, a consensus conclusion much higher than the reliability of any single sensor is obtained, thereby improving the accuracy of early fire warning in complex environment and reducing false positives caused by sensor failure or local environmental interference.

[0128] In the early fire warning method, local multi-modal observation data and state confidence of each monitoring point are obtained, the local multi-modal observation data includes acquisition time label and acquisition location label, and the state confidence includes health degree confidence and local anomaly confidence; monitoring points whose local multi-modal observation data meet preset anomaly triggering requirements are determined as anomaly monitoring points, and based on the acquisition location label, a neighborhood reference monitoring point set of the anomaly monitoring points is selected; based on the anomaly monitoring points, the neighborhood reference monitoring point set and the local multi-modal observation data, a preset static monitoring point network topology graph is updated to obtain a dynamic local monitoring point network topology graph; based on the dynamic local monitoring point network topology graph, spatio-temporal correlation characteristics between each monitoring point in the anomaly monitoring points and the neighborhood reference monitoring point set are analyzed to obtain a state embedding vector; based on the state embedding vector and the state confidence, failure probabilities of the anomaly monitoring points and the monitoring points are quantified to obtain a failure probability vector; the failure probability vector includes a fire probability component, a monitoring point failure component and an environmental interference probability component; the failure probability vectors of the monitoring points are fused to obtain a fusion trust degree vector, and based on the fusion trust degree vector, fire warning information is generated. The traditional single threshold judgment is upgraded to a three-element classification decision based on group intelligence spatio-temporal pattern recognition, which effectively distinguishes real fire, sensor failure and environmental interference in complex environment, improves the accuracy and reliability of early fire warning, and reduces the dependence on the center node and communication load through local collaborative calculation.

[0129] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0130] Based on the same inventive concept, the embodiments of the present application also provide an early fire warning device for implementing the early fire warning method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more early fire warning device embodiments provided below can refer to the limitations of the early fire warning method described above, which will not be repeated here.

[0131] In one exemplary embodiment, as shown in Figure 2 An early fire warning device 200 is provided, comprising:

[0132] A data acquisition module 201 is configured to acquire local multi-modal observation data and state confidence of each monitoring point, wherein the local multi-modal observation data comprises a collection time label and a collection location label, and the state confidence comprises a health degree confidence and a local anomaly confidence;

[0133] An anomaly point analysis module 202 is configured to determine a monitoring point whose local multi-modal observation data meets a preset anomaly trigger requirement as an anomaly monitoring point, and select a neighborhood reference monitoring point set of the anomaly monitoring point based on the collection location label;

[0134] A connection relationship updating module 203 is configured to update a preset static monitoring point network topology graph based on the anomaly monitoring point, the neighborhood reference monitoring point set, and the local multi-modal observation data, to obtain a dynamic local monitoring point network topology graph;

[0135] A time sequence feature analysis module 204 is configured to analyze a spatio-temporal correlation feature between each monitoring point in the anomaly monitoring point and the neighborhood reference monitoring point set based on the dynamic local monitoring point network topology graph, to obtain a state embedding vector;

[0136] A fault probability analysis module 205 is configured to quantify a fault probability of the anomaly monitoring point and the monitoring point based on the state embedding vector and the state confidence, to obtain a fault probability vector; the fault probability vector comprises a fire probability component, a monitoring point fault component, and an environmental interference probability component;

[0137] An early warning information generation module 206 is configured to fuse the fault probability vectors of each monitoring point to obtain a fusion trust degree vector, and generate fire warning information based on the fusion trust degree vector.

[0138] Further, the static monitoring point network topology graph comprises graph nodes, connection edges, and edge weights of the connection edges, the graph nodes are each monitoring point, the connection edges are used to represent the connection relationship between each monitoring point, and the edge weights are the proximity degrees between each monitoring point;

[0139] The connection relationship updating module is further configured to:

[0140] extracting graph nodes and connecting edges from the static monitoring point network topology graph based on the abnormal monitoring point and the set of neighboring reference monitoring points to form an initial local subgraph;

[0141] normalizing the local multi-modal observation data for each monitoring point in the abnormal monitoring point and the set of neighboring reference monitoring points to obtain a standardized observation feature vector of each monitoring point;

[0142] for each monitoring point in the abnormal monitoring point and the set of neighboring reference monitoring points, adding the standardized observation feature vector as the graph node feature to the graph node corresponding to the abnormal monitoring point or the monitoring point in the initial local subgraph to obtain a dynamic local graph;

[0143] for each monitoring point in the abnormal monitoring point and the set of neighboring reference monitoring points, selecting local multi-modal observation data with acquisition time labels meeting preset monitoring window parameter requirements to form a time sequence feature sequence, and obtaining a dynamic local monitoring point network topology graph based on the time sequence feature sequence and the dynamic local graph.

[0144] Further, the time sequence feature analysis module is further used for:

[0145] repeating the steps S1 to S5 until a preset maximum number of cycles is reached to obtain the state embedding vector of the abnormal monitoring point and the state embedding vector of each monitoring point;

[0146] wherein, the steps S1 to S5 are:

[0147] S1, performing linear transformation on the state embedding vector based on a preset linear transformation weight matrix to obtain a hidden feature vector of the graph node; wherein, when the first S1 step is executed, the state embedding vector is the standardized observation feature vector;

[0148] S2, calculating the attention weight between the graph nodes at both ends of each connecting edge in the dynamic local monitoring point network topology graph based on the dynamic local monitoring point network topology graph and the hidden feature vector;

[0149] S3, for each graph node in the dynamic local monitoring point network topology graph, performing weighted aggregation on the hidden feature vectors of the graph nodes connected to the graph node through the connecting edges based on the attention weight to obtain a spatial aggregation feature vector;

[0150] S4, for each graph node in the dynamic local monitoring point network topology graph, processing the time dynamic feature vector based on a convolutional neural network to obtain an updated time dynamic feature vector; when the first S4 step is executed, the time dynamic feature vector is the time sequence feature sequence;

[0151] S5, for each graph node in the dynamic local monitoring point network topology graph, based on a preset nonlinear activation function, the spatial aggregation feature vector, the time dynamic feature vector and the state embedding vector are processed to obtain an updated state embedding vector; wherein when the first S5 step is executed, the state embedding vector is a preset initial state embedding vector.

[0152] Further, the time series feature analysis module is further used for:

[0153] Based on the dynamic local monitoring point network topology graph and the hidden feature vector, the attention weight between the graph nodes at both ends of each connection edge in the dynamic local monitoring point network topology graph is calculated, and the form of the attention weight between the graph nodes is:

[0154]

[0155] Wherein, is the index of any graph node in the dynamic local monitoring point network topology graph, is any graph node connected to the graph node through a connection edge in the dynamic local monitoring point network topology graph, is the attention weight between the graph node and the graph node , is the initial attention weight between the graph node and the graph node , is another graph node connected to the graph node through a connection edge in the dynamic local monitoring point network topology graph, is a set of graph nodes connected to the graph node through a connection edge in the dynamic local monitoring point network topology graph, is the initial attention weight between the graph node and the graph node ,

[0156] Wherein, the initial attention weight between the graph node and the graph node is calculated by the following formula:

[0157]

[0158] Wherein, is the initial attention weight between the graph node and the graph node , is a nonlinear activation function, is a hidden feature vector of the graph node , is a graph node​ a hidden feature vector, is a graph node and a proximity between the graph node and a graph node is an attention vector parameter.

[0159] Further, the fault probability analysis module is further configured to:

[0160] input the state embedding vector into a preset basic probability assignment generation network to obtain an initial basic probability assignment for each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set; the initial basic probability assignment includes a fire hypothesis initial probability value, a monitoring point fault initial probability value, an environmental interference initial probability value and other interference probability values; the basic probability assignment generation network is built based on a multilayer perception mechanism;

[0161] multiply the health confidence and the local anomaly confidence to obtain a comprehensive confidence factor for each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set;

[0162] adjust the initial basic probability assignment based on the comprehensive confidence factor to obtain a weighted basic probability assignment for each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set; the weighted basic probability assignment includes a fire probability component, a monitoring point fault component and an environmental interference probability component;

[0163] adjust the other interference probability values based on the comprehensive confidence factor to obtain a calibrated uncertainty probability value for each monitoring point in the abnormal monitoring point and the neighborhood reference monitoring point set, and obtain a fault probability vector based on the calibrated uncertainty probability value and the weighted basic probability assignment.

[0164] Further, the early warning information generation module is further configured to:

[0165] recursively fuse the fault probability vector of the abnormal monitoring point with the fault probability vectors of the monitoring points in the neighborhood reference monitoring point set to obtain a fusion trust degree vector; the fusion trust degree vector includes a fusion fire probability component, a fusion monitoring point fault probability component and a fusion environmental interference probability component;

[0166] compare the fusion monitoring point fault probability component and the fusion environmental interference probability component to obtain a comparison result, and determine a fault trust degree component according to the comparison result;

[0167] calculate a difference value between the fusion fire probability component and the fault trust degree component to obtain an interference trust degree difference value;

[0168] when the fusion fire probability component is greater than a preset first threshold value and the interference trust degree difference value is greater than a preset second threshold value, generate fire early warning information based on the local multi-modal observation data of the abnormal monitoring point and the fusion trust degree vector.

[0169] In one embodiment, as Figure 3 A computer device is provided, comprising:

[0170] At least one processor 301, and a memory 302 connected with the at least one processor 301 in communication: the memory stores application code executable by the at least one processor, and the application code is executed by the at least one processor to enable the at least one processor to perform the early fire warning method as described above.

[0171] The computer device can further comprise a monitoring point 303.

[0172] The processor 301, the memory 302 and the monitoring point 303 can be connected through a bus or other means, and the connection through the bus is taken as an example in the figure.

[0173] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps in the above method embodiments.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts are described with reference to the parts of the method embodiments. The above described device embodiments are merely illustrative, and the components described as separate components can be or can not be physically separated, and the components displayed as units can be or can not be physical units, i.e. they can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present disclosure according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0175] The above described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. An early fire warning method, characterized in that, The method includes: Acquire local multimodal observation data and state confidence scores for each monitoring point. The local multimodal observation data includes collection time labels and collection location labels. The state confidence scores include health confidence scores and local anomaly confidence scores. The monitoring points whose local multimodal observation data meet the preset anomaly triggering requirements are identified as anomaly monitoring points, and based on the collection location label, the monitoring points are selected to form a neighborhood reference monitoring point set of the anomaly monitoring points; Based on the anomaly monitoring points, the neighborhood reference monitoring point set, and the local multimodal observation data, the preset static monitoring point network topology is updated to obtain a dynamic local monitoring point network topology. Based on the dynamic local monitoring point network topology, the spatiotemporal correlation characteristics between the abnormal monitoring points and each monitoring point in the neighborhood reference monitoring point set are analyzed to obtain the state embedding vector. Based on the state embedding vector and the state confidence, the abnormal monitoring point and the failure probability of the monitoring point are quantified to obtain a failure probability vector; the failure probability vector includes a fire probability component, a monitoring point failure component and an environmental interference probability component. The fault probability vectors of each monitoring point are fused to obtain a fused trust vector, and fire early warning information is generated based on the fused trust vector.

2. The method according to claim 1, characterized in that, The static monitoring point network topology graph includes graph nodes, connecting edges, and edge weights of the connecting edges. The graph nodes are each of the monitoring points, the connecting edges are used to characterize the connection relationship between each of the monitoring points, and the edge weights are the proximity between each of the monitoring points. The process of updating the preset static monitoring point network topology based on the anomaly monitoring points, the neighborhood reference monitoring point set, and the local multimodal observation data to obtain a dynamic local monitoring point network topology includes: Based on the abnormal monitoring points and the neighborhood reference monitoring point set, the graph nodes and the connecting edges are extracted from the static monitoring point network topology graph to form an initial local subgraph; For each of the monitoring points in the set of abnormal monitoring points and the set of neighboring reference monitoring points, the local multimodal observation data is normalized to obtain the standardized observation feature vector of each monitoring point. For each of the anomaly monitoring points and the neighboring reference monitoring point set, the standardized observation feature vector is used as the graph node feature and added to the anomaly monitoring point or the graph node corresponding to the monitoring point in the initial local subgraph to obtain the dynamic local graph; For each of the anomaly monitoring points and the neighboring reference monitoring point set, the local multimodal observation data whose acquisition time label meets the preset monitoring window parameter requirements are selected to form a time-series feature sequence, and the dynamic local monitoring point network topology map is obtained based on the time-series feature sequence and the dynamic local graph.

3. The method according to claim 2, characterized in that, Based on the dynamic local monitoring point network topology, the spatiotemporal correlation features between the abnormal monitoring points and the monitoring points in the neighborhood reference monitoring point set are analyzed to obtain a state embedding vector, including: Repeat steps S1 to S5 until the preset maximum number of iterations is reached to obtain the state embedding vector of the anomaly monitoring point and the state embedding vector of each monitoring point; Steps S1 to S5 are as follows: S1, based on a preset linear transformation weight matrix, the state embedding vector is linearly transformed to obtain the hidden feature vector of the graph node; wherein, when the first step S1 is executed, the state embedding vector is the standardized observation feature vector; S2, based on the dynamic local monitoring point network topology graph and the hidden feature vector, calculate the attention weight between the graph nodes at both ends of each connecting edge in the dynamic local monitoring point network topology graph; S3, for each graph node in the dynamic local monitoring point network topology graph, based on the attention weight, the hidden feature vectors of the graph nodes connected to the graph nodes through the connecting edges are weighted and aggregated to obtain a spatial aggregated feature vector; S4, for each graph node in the dynamic local monitoring point network topology graph, the time dynamic feature vector is processed based on a convolutional neural network to obtain the updated time dynamic feature vector; when the first step S4 is executed, the time dynamic feature vector is the time-series feature sequence; S5, for each graph node in the dynamic local monitoring point network topology graph, the spatial aggregation feature vector, the temporal dynamic feature vector, and the state embedding vector are processed based on a preset nonlinear activation function to obtain the updated state embedding vector; wherein, when the first step S5 is executed, the state embedding vector is a preset initial state embedding vector.

4. The method according to claim 3, characterized in that, The calculation of attention weights between graph nodes at both ends of each connecting edge in the dynamic local monitoring point network topology graph, based on the dynamic local monitoring point network topology graph and the hidden feature vector, includes: Based on the dynamic local monitoring point network topology and the hidden feature vector, the attention weights between the graph nodes at both ends of each connecting edge in the dynamic local monitoring point network topology are calculated. The form of the attention weights between the graph nodes is: in, It is the index of any node in the dynamic local monitoring point network topology graph. It is a dynamic local monitoring point network topology graph that connects edges to graph nodes. Any connected graph node, Graph nodes Graph Nodes The weight of intention between them Graph nodes Graph Nodes Initial attention weights between them It is a dynamic local monitoring point network topology graph that connects edges to graph nodes. Another graph node connected, It is a dynamic local monitoring point network topology graph that connects edges to graph nodes. The set of connected graph nodes Graph nodes Graph Nodes Initial attention weights between; Among them, graph nodes Graph Nodes Initial attention weights between It is calculated using the following formula: in, Graph nodes Graph Nodes Initial attention weights between them It is a non-linear activation function. Graph nodes The hidden feature vectors, Graph nodes The hidden feature vectors, Graph nodes Graph Nodes The degree of proximity between them These are the attention vector parameters.

5. The method according to claim 1, characterized in that, The step of quantifying the anomaly monitoring point and the failure probability of the monitoring point based on the state embedding vector and the state confidence to obtain a failure probability vector includes: For each of the abnormal monitoring points and the neighboring reference monitoring point set, the state embedding vector is input into a preset basic probability allocation generation network to obtain an initial basic probability allocation; the initial basic probability allocation includes the initial probability value of the fire hypothesis, the initial probability value of the monitoring point failure, the initial probability value of environmental interference, and other interference probability values; the basic probability allocation generation network is constructed based on a multilayer perceptron. For each monitoring point in the set of anomaly monitoring points and the set of neighborhood reference monitoring points, the health confidence score is multiplied by the local anomaly confidence score to obtain a comprehensive confidence factor. For each of the monitoring points in the abnormal monitoring point and the neighborhood reference monitoring point set, the initial basic probability allocation is adjusted based on the comprehensive confidence factor to obtain a weighted basic probability allocation; the weighted basic probability allocation includes the fire probability component, the monitoring point fault component, and the environmental interference probability component. For each of the monitoring points in the abnormal monitoring point and the neighborhood reference monitoring point set, the other interference probability values ​​are adjusted based on the comprehensive confidence factor to obtain a calibration uncertainty probability value, and the fault probability vector is obtained based on the calibration uncertainty probability value and the weighted basic probability allocation.

6. The method according to claim 1, characterized in that, The method involves fusing the fault probability vectors of each monitoring point to obtain a fused trust vector, and generating fire early warning information based on the fused trust vector, including: The fault probability vector of the abnormal monitoring point is recursively fused with the fault probability vector of each monitoring point in the neighborhood reference monitoring point set to obtain the fused trust vector; the fused trust vector includes a fused fire probability component, a fused monitoring point fault probability component, and a fused environmental interference probability component. The fault probability component of the fused monitoring point and the fused environmental interference probability component are compared to obtain the comparison result, and the fault confidence component is determined based on the comparison result. The difference between the fused fire probability component and the fault confidence component is calculated to obtain the interference confidence difference. When the fused fire probability component is greater than a preset first threshold and the interference confidence difference is greater than a preset second threshold, the fire warning information is generated based on the local multimodal observation data of the abnormal monitoring point and the fused confidence vector.

7. An early fire warning device, characterized in that, The device includes: The data acquisition module is used to acquire local multimodal observation data and state confidence scores of each monitoring point. The local multimodal observation data includes collection time labels and collection location labels, and the state confidence scores include health confidence scores and local anomaly confidence scores. The anomaly analysis module is used to identify monitoring points in the local multimodal observation data that meet the preset anomaly triggering requirements as anomaly monitoring points, and to select the monitoring points to form a neighborhood reference monitoring point set of the anomaly monitoring points based on the collection location label. The connection relationship update module is used to update the preset static monitoring point network topology map based on the abnormal monitoring point, the neighborhood reference monitoring point set and the local multimodal observation data, so as to obtain a dynamic local monitoring point network topology map. The temporal feature analysis module is used to analyze the spatiotemporal correlation features between the abnormal monitoring points and each monitoring point in the neighborhood reference monitoring point set based on the dynamic local monitoring point network topology map, and obtain the state embedding vector. The fault probability analysis module is used to quantify the fault probability of the abnormal monitoring point and the monitoring point based on the state embedding vector and the state confidence, and obtain a fault probability vector; the fault probability vector includes a fire probability component, a monitoring point fault component and an environmental interference probability component. The early warning information generation module is used to fuse the fault probability vectors of each monitoring point to obtain a fused trust vector, and generate fire early warning information based on the fused trust vector.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images