Emergency event response method

Abstract

The invention relates to an emergency event response method, and belongs to the technical field of emergency response, and the method comprises the steps: obtaining multi-mode emergency data of a collection end through a preset interface; inputting the multi-modal emergency data into a preset cross-modal fusion model for feature extraction, attention weight calculation and data fusion to obtain a global emergency situation tensor; inputting the global emergency situation tensor into a preset space-time diagram convolutional network coupling model, and performing feature extraction on the spatial features and the time features based on the spatial dimension and the time dimension to obtain a situation representation result; a Markov decision process model is constructed, an action vector is defined, a state vector is defined according to a situation representation result, an action decision is generated through a pre-selected strategy network, the Markov decision process model responds to an emergency event according to the action decision, a response result is obtained, and the real-time performance and accuracy of emergency response are remarkably improved. The problems of data fragmentation and decision lag in a traditional method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of emergency response, in particular to an emergency event response method. BACKGROUND

[0002] Emergencies such as natural disasters and accident disasters are sudden, destructive and chain-like, often causing significant casualties, property losses and social disorder in a short period of time. Traditional emergency response relies on manual experience and static rules, which is difficult to cope with the complex evolution of disasters, leading to delayed decision-making and misallocation of resources, further exacerbating disaster losses.

[0003] Currently, emergency response technology collects disaster data through sensors, GIS, social media, etc., but existing fusion technology cannot handle the challenges of modal differences and different spatio-temporal scales; or uses linear programming, heuristic rules or single AI model for resource scheduling, but relies on static parameters and cannot adapt to the dynamic changes of the disaster site.

[0004] In addition, traditional methods cannot achieve semantic-level fusion, leading to fragmented situational awareness, single models are difficult to capture the spatio-temporal evolution of disasters, prediction results are poor in real-time and cannot be explained, static optimization models cannot dynamically adjust strategies, leading to delayed or misallocated resource scheduling, AI-generated black-box decisions lack transparency and explainability, commanders cannot understand the logic or intervene, reducing emergency efficiency. SUMMARY

[0005] In view of the deficiencies in the related art, the purpose of the present application is to provide an emergency event response method to solve the technical problems that the prior art cannot achieve semantic-level fusion, leading to fragmented situational awareness, single models are difficult to capture the spatio-temporal evolution of disasters, prediction results are poor in real-time and cannot be explained, static optimization models cannot dynamically adjust strategies, leading to delayed or misallocated resource scheduling, AI-generated black-box decisions lack transparency and explainability, commanders cannot understand the logic or intervene, reducing emergency efficiency.

[0006] The present application provides an emergency event response method, comprising the following steps: Data acquisition step: acquiring multi-modal emergency data of the acquisition end through a preset interface; Feature fusion step: inputting the multi-modal emergency data into a preset cross-modal fusion model for feature extraction, attention weight calculation and data fusion to obtain a global emergency situation tensor; The situation characterization result output step: inputting the global emergency situation tensor into a preset space-time graph convolution network coupling model, extracting spatial features based on a spatial dimension to obtain a spatial feature vector, extracting time features based on a time dimension to obtain a time feature vector, and obtaining a situation characterization result according to the spatial feature vector and the time feature vector; The decision response step: constructing a Markov decision process model, defining an action vector, defining a state vector according to the situation characterization result, generating an action decision through a preselected policy network according to the state vector and the action vector, and responding to an emergency event according to the action decision through the Markov decision process model to obtain a response result.

[0007] The data acquisition step collects multi-modal emergency data through a preset interface, ensuring the diversity and availability of the data; the feature fusion step extracts features and calculates attention weights through a cross-modal fusion model, realizing semantic-level fusion of multi-modal data and improving data utilization efficiency; the situation characterization result output step captures space-time dependency through a space-time graph convolution network coupling model, enhancing the comprehensiveness and dynamics of situation awareness; and the decision response step trains a policy network based on a proximal policy optimization algorithm to generate an action decision to dynamically respond to an emergency event, thereby significantly improving the real-time performance and accuracy of emergency response and solving the problems of data fragmentation and decision lag in traditional methods.

[0008] In some embodiments of the application, the feature fusion step specifically includes: The feature extraction step: the multi-modal emergency data is input into a modal-specific encoder in the cross-modal fusion model for modal encoding and feature extraction, to obtain multi-modal features; The attention weighting calculation step: the cross-modal attention layer in the cross-modal fusion model is used to calculate cross-modal attention weights of the multi-modal features, to obtain a weight matrix, and the weight matrix and the multi-modal features are weighted and aggregated to obtain cross-modal interaction features; The data fusion step: the cross-modal interaction features and the multi-modal features are spliced or weighted and fused to obtain the global emergency situation tensor.

[0009] Modal encoding and feature extraction through a modal-specific encoder ensure that the features of different modal data can be fully mined; the cross-modal attention layer establishes semantic associations between different modal features, calculates attention weights, and enables the cross-modal fusion model to dynamically focus on key modal information; the weighted fusion step splices or weightedly fuses the cross-modal interaction features and the multi-modal features to generate a global emergency situation tensor, improves the precision of multi-modal data fusion, enhances the adaptability of the cross-modal fusion model to complex emergency scenarios, and provides a high-quality data basis for subsequent situation awareness and decision-making.

[0010] In some embodiments of the present application, the situation characterization result outputting step is specifically: The space feature extraction step: in the STGCN layer of the spatio-temporal graph convolution network coupling model, a graph structure is generated according to a spatial topology structure, and after a graph convolution operation is performed on the global emergency situation tensor, a space feature is extracted based on a space dimension to obtain a space feature vector; The time feature extraction step: a time position encoding is added at an output end of the STGCN layer, a space-time encoding is performed on the space feature vector according to the time position encoding, an association weight of a time step is calculated through a self-attention mechanism in the Transformer layer of the spatio-temporal graph convolution network coupling model to dynamically analyze a key time node, a time feature is extracted based on a time dimension through the Transformer layer to obtain a time feature vector.

[0011] Through the space feature extraction step and the time feature extraction step, the spatio-temporal dependence relationship of the emergency event is comprehensively captured in the spatio-temporal graph convolution network coupling model. In the space feature extraction step, the graph convolution operation and the space feature vector extraction operation are performed, which enhances the understanding of the geographical distribution feature by the spatio-temporal graph convolution network coupling model. Through the time position encoding and the self-attention mechanism, the key time node in the disaster evolution process is dynamically analyzed in the time feature extraction step, which improves the processing capability of the time sequence data by the spatio-temporal graph convolution network coupling model, so that the spatio-temporal graph convolution network coupling model can more accurately represent the evolution rule of the emergency situation, and provides a more reliable basis for decision-making.

[0012] In some embodiments of the present application, the space feature extraction step is specifically: A graph structure is generated according to a spatial topology structure, and the graph structure is represented by an adjacency matrix as:

[0013] wherein, is an adjacency weight between nodes and ; is a geographical distance between nodes and , and the nodes are geographical entities related to the emergency event; is a scale parameter for controlling weight decay.

[0014] The influence of the geographical distance between nodes on the weight is quantified through the mathematical expression of the adjacency matrix, and the calculation of the adjacency weight is based on an exponential decay function, so that nodes with a closer distance have a higher weight, thereby more accurately reflecting the spatial dependence relationship, enhancing the sensitivity of the spatio-temporal graph convolution network coupling model to geographical spatial features, and providing a scientific and reasonable input for the graph convolution operation, further improving the accuracy and efficiency of the spatial dimension modeling.

[0015] In some embodiments of the application, the calculation model of the time position encoding is:

[0016] wherein, is the time position encoding; is the time step; is the dimension index; is the total number of feature dimensions.

[0017] The time step and the dimension index are mapped into the position encoding through a sine function, so that the spatio-temporal graph convolution network coupling model can capture the sequential relationship and periodic characteristics in the time series, preserve the continuity of the time information, and enhance the modeling ability of the spatio-temporal graph convolution network coupling model for long-term dependence relationship, thereby providing a powerful mathematical tool for time dimension modeling, and improving the identification accuracy of the spatio-temporal graph convolution network coupling model for key time nodes in the disaster evolution process.

[0018] In some embodiments of the application, the emergency event response method further comprises: A decision assistance step: adding a real-time fusion data layer, a situation assessment layer and a decision instruction layer on the digital twin map platform, and displaying the global emergency situation tensor, the situation characterization result and the action vector, respectively.

[0019] By displaying the global emergency situation tensor, the situation characterization result and the action vector on the digital twin map platform in real time, an intuitive visualization tool is provided for the command personnel.

[0020] In some embodiments of the application, the decision response step specifically comprises: A reward function definition step: defining a reward function in the Markov decision process model, and calculating a total reward value according to the reward function; A decision evaluation step: training a policy network and an evaluation network through a proximal policy optimization algorithm, the policy network generates an action decision in combination with the reward function, and the evaluation network evaluates the action decision in combination with the reward function.

[0021] By introducing the reward function into the Markov decision process model, a scientific framework is provided for the training of the policy network and the evaluation network; The proximal policy optimization algorithm optimizes the policy network and the evaluation network, generates action decisions according to the policy network, evaluates the action decisions through the evaluation network, realizes dynamic response to emergency events, can adapt to dynamic changes in the disaster site, and also ensures the rationality and efficiency of the decisions through the guidance of the reward function, and significantly improves the intelligent level of emergency response.

[0022] In some embodiments of the application, the calculation model of the state vector is:

[0023] wherein, is a state vector; is a vector of situation representation results; is a real-time available resource state vector; is an environmental constraint vector.

[0024] By organically combining the vector of situation representation results, the real-time available resource state vector and the environmental constraint vector, the state characteristics of the emergency scene are comprehensively described, the perception ability of the Markov decision process model to the complex environment is enhanced, and rich information input is provided for the policy network, thereby improving the accuracy and adaptability of the decision, and laying a solid foundation for efficient execution of emergency response.

[0025] In some embodiments of the application, the calculation model of the total reward value is:

[0026] wherein, is a total reward value; is a success reward value; is a time penalty value; is a resource consumption penalty value; is a constraint violation penalty value; is a risk penalty value.

[0027] Through the calculation model of the total reward value, the success reward, the time penalty, the resource consumption penalty, the constraint violation penalty and the risk penalty are comprehensively considered, a comprehensive optimization target is provided for the training of the policy network, the key factors in the emergency response are balanced, and the generation of unreasonable decisions is avoided through the penalty mechanism, thereby ensuring the accuracy of the action vector generated by the policy network.

[0028] In some embodiments of the application, the decision response step further comprises: A model training step: according to historical cases and preset physical rules, an emergency response simulation environment is constructed, and in the emergency response simulation environment, the Markov decision process model is trained through the proximal policy optimization algorithm and the experience return pool.

[0029] By constructing an emergency response simulation environment and using an experience replay pool to train the Markov decision process model, efficient learning and improved generalization ability of the Markov decision process model are realized. The simulation environment constructed based on historical cases and preset physical rules provides rich and real training scenes; the experience replay pool mechanism improves data utilization efficiency and enhances the stability of training, accelerates the convergence of the policy network, and enhances the generalization ability of the Markov decision process model through simulation of diversified scenes, so that the Markov decision process model can adapt to various emergency events, and the problems of insufficient real scene training data and high trial and error cost are overcome. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, specific embodiments of the present application will be described in detail below with reference to the drawings. Those skilled in the art can obtain other drawings without making creative efforts based on these drawings. Figure 1 A flowchart of an emergency event response method provided by the embodiment of the present application; Figure 2 A flowchart of a feature fusion step S2 provided by the embodiment of the present application; Figure 3 A flowchart of a situation representation result output step S3 provided by the embodiment of the present application; Figure 4 A flowchart of a decision response step S4 provided by the embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be described and explained below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. Based on the embodiments provided by the present application, all other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application. It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be understood that the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices. In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other. The technical solutions of the application will be described in detail below with specific embodiments and the accompanying drawings.

[0032] As shown in the accompanying drawings, Figure 1 The application provides an emergency event response method, comprising the following steps: Data acquisition step S1: acquiring multi-modal emergency data of the collection end through a preset interface; optionally, after acquiring the multi-modal emergency data, the multi-modal emergency data is also preprocessed; The preprocessing operation of the multi-modal emergency data includes standardization, data cleaning and spatiotemporal stamp alignment operation; wherein, the data cleaning includes median filter denoising, KNN missing value processing; the spatiotemporal stamp alignment operation includes GIS coordinate interpolation based on.

[0033] Optionally, the collection end includes a physical sensor, a geographic information system database, a social media platform API, an emergency reporting system, a camera and a historical case library; The preset interface adopts MQTT protocol or RESTful protocol; the data in the physical sensor is acquired through the MQTT protocol, and the data in the geographic information system database or the emergency reporting system is acquired through the RESTful protocol; Feature fusion step S2: inputting the preprocessed multi-modal emergency data into a preset cross-modal fusion model for feature extraction, attention weight calculation and data fusion to obtain a global emergency situation tensor; Specifically, the cross-modal fusion model is based on the Transformer architecture, the number of Transformer layers, the number of heads and the dimension of the hidden layer of the cross-modal fusion model are defined, and the semantic level fusion of the multi-modal emergency data is realized; Situation representation result output step S3: inputting the global emergency situation tensor into a preset spatiotemporal graph convolution network coupling model, extracting spatial features based on the spatial dimension to obtain a spatial feature vector, extracting temporal features based on the time dimension to obtain a temporal feature vector, and obtaining a situation representation result according to the spatial feature vector and the temporal feature vector; Specifically, the situation representation result includes event type and confidence, influence range geofence, affected population or critical facility density heat map, key risk point atlas with risk level, and event multi-step evolution trend probability distribution; Decision response step S4: constructing a Markov decision process model, defining an action vector, defining a state vector according to the situation representation result, generating an action decision through a preselected policy network according to the state vector and the action vector, and responding to the emergency event according to the action decision through the Markov decision process model to obtain a response result.

[0034] The multi-modal emergency data is collected and preprocessed through the data acquisition step by using the preset interface, ensuring the diversity and availability of the data; the feature fusion step extracts features and calculates attention weights through the cross-modal fusion model, realizing semantic-level fusion of multi-modal data and improving data utilization efficiency; the situation representation result output step captures the spatio-temporal dependency relationship through the spatio-temporal graph convolution network coupling model, enhancing the comprehensiveness and dynamics of situation awareness; and the decision response step trains the strategy network based on the proximal policy optimization algorithm to generate action decisions to dynamically respond to emergency events, thereby significantly improving the real-time performance and accuracy of emergency response, and solving the problems of data fragmentation and decision lag in traditional methods.

[0035] In some embodiments, as shown in FIG. 2, the feature fusion step S2 is specifically: Figure 2 The feature extraction step S21: the multi-modal emergency data is input into the modal-specific encoder in the cross-modal fusion model for modal encoding and feature extraction, to obtain multi-modal features; Specifically, in the modal-specific encoder, the numerical sequence obtained through the physical sensor, the text sequence obtained through the social media platform AIP or the emergency reporting system, the image or video frame obtained through the camera, and the graph structure data obtained through the geographic information system database in the multi-modal emergency data are processed respectively; According to the different types of modal-specific encoders, the emergency data of different modalities is processed; for example, BERT processes the text sequence, ResNet processes the image, and GCN processes the graph structure data; The attention weighting calculation step S22: the cross-modal attention layer in the cross-modal fusion model is used to calculate the cross-modal attention weights of the multi-modal features, to obtain a weight matrix, and the weight matrix is used to weight and aggregate the multi-modal features, to obtain cross-modal interaction features; Specifically, through the cross-modal attention weight calculation, the selective focusing and interaction of key information are realized; for example, the "road collapse" described in the text sequence is associated with the image features of the corresponding position camera and the abnormal vibration data of the physical sensor on the road segment; The data fusion step S23: through the fusion mode combining Concatenation and FC Layer, the cross-modal interaction features and the multi-modal features are spliced or fused to obtain a unified, spatio-temporally aligned and semantically rich global emergency situation tensor.

[0036] ​The modal encoding and feature extraction are performed by the modal-specific encoder, so that the features of different modal data can be fully mined; the semantic association between the features of different modal is established by the cross-modal attention layer, and the attention weight is calculated, so that the cross-modal fusion model can dynamically focus on the key modal information, effectively solving the problem of semantic-level fusion of multi-source heterogeneous data, and significantly improving the comprehensiveness, accuracy and real-time of situation awareness; The weighted fusion step splices or weights the cross-modal interaction features and the multi-modal features to generate a global emergency situation tensor, thereby improving the precision of multi-modal data fusion and enhancing the adaptability of the cross-modal fusion model to complex emergency scenarios, providing a high-quality data basis for subsequent situation awareness and decision-making.

[0037] The cross-modal fusion model adopts a modal-specific encoder and combines a query anchor mechanism to describe the text sequence as a query benchmark, dynamically calculates the association weight of the image, sensor and GIS data relative to the text sequence through the cross-modal attention, realizes the feature alignment driven by disaster semantics, and solves the problem of time-space offset of multi-source data.

[0038] In some embodiments, as shown in Figure 3 The situation representation result output step S3 is specifically: The spatial feature extraction step S31: in the STGCN layer of the spatio-temporal graph convolution network coupling model, a graph structure is generated according to a spatial topology structure, a graph convolution operation is performed on the global emergency situation tensor, and then spatial features are extracted based on the spatial dimension to obtain a spatial feature vector; Specifically, the spatial topology structure can be a GIS road network or a geographic grid, for example, a graph structure is generated based on road continuity; In the STGCN layer, the number of graph convolution layers and parameters are defined; The spatial feature vector obtains the dependency relationship and proximity relationship of the spatial dimension, including the relationship that a disaster spreads along a traffic line and the influence between regions; The time feature extraction step S32: a time position encoding is added at the output end of the STGCN layer, the spatial feature vector is spatio-temporally encoded according to the time position encoding, the association weight of the time step is calculated through the self-attention mechanism in the Transformer layer of the spatio-temporal graph convolution network coupling model, the key time nodes are dynamically analyzed, the time features are extracted based on the time dimension through the Transformer layer, and a time feature vector is obtained; Specifically, the number of Transformer encoder layers and the number of heads are defined in the Transformer layer; The cross-entropy loss is used to optimize the event classification task, and the MSE mean square error loss is used to optimize the range regression or quantity regression task; The time dimension dependency is obtained by a time feature vector, which includes the stage characteristics, periodicity and other relationships of disaster development; Through the spatial feature extraction step and the time feature extraction step, the comprehensive capture of the spatio-temporal dependency of the emergency event is realized in the spatio-temporal graph convolution network coupling model. In the spatial feature extraction step, the graph convolution operation and the spatial feature vector extraction operation are performed, which enhances the understanding of the geographical distribution characteristics of the spatio-temporal graph convolution network coupling model; the time feature extraction step dynamically analyzes the key time nodes in the disaster evolution process through the time position coding and the self-attention mechanism, which improves the processing ability of the spatio-temporal graph convolution network coupling model for time series data, fully utilizes the spatial graph structure and the time dependency, realizes the high-precision and interpretable evaluation and prediction of the complex emergency situation, and provides a reliable basis for decision-making, so that the spatio-temporal graph convolution network coupling model can more accurately represent the evolution law of the emergency situation, and provides a more reliable basis for decision-making.

[0039] In some embodiments, the spatial feature extraction step S31 is specifically: According to the spatial topological structure, a graph structure is generated, and the graph structure is represented by an adjacency matrix as:

[0040] wherein, is an adjacency weight between the node and the node . is a geographical distance between the node and the node , and the node is a geographical entity related to the emergency event; is a scale parameter for controlling the weight decay, is 5km, which is used to simulate the spatial propagation decay of the disaster.

[0041] Specifically, the node is a geographical entity related to the emergency event, including a monitoring device, a key place, a disaster-related point, a dynamic target and an administrative region.

[0042] The influence of the geographical distance between the nodes on the weight is quantified through the mathematical expression of the adjacency matrix, and the calculation of the adjacency weight is based on an exponential decay function, so that the nodes with closer distance have higher weight, thereby more accurately reflecting the spatial dependency and enhancing the sensitivity of the spatio-temporal graph convolution network coupling model to the geographical spatial features. In addition, the scientific and reasonable input is provided for the graph convolution operation, which further improves the accuracy and efficiency of the spatial dimension modeling.

[0043] In some embodiments, a time position coding is added at the output end of the STGCN, so that the spatio-temporal graph convolution network coupling model perceives the disaster evolution stage, and the calculation model of the time position coding is:

[0044] wherein, is a time position encoding; is a time step; is a dimension index; is a total number of feature dimensions.

[0045] The time step and the dimension index are mapped into the position encoding by a sine function, so that the spatio-temporal graph convolution network coupling model can capture the sequential relationship and periodic characteristics in the time series, retain the continuity of the time information, and also enhance the modeling ability of the spatio-temporal graph convolution network coupling model for long-term dependencies, providing a powerful mathematical tool for time dimension modeling, thereby improving the identification accuracy of the spatio-temporal graph convolution network coupling model for key time nodes in the disaster evolution process.

[0046] In some embodiments, as shown in Figure 4 the decision response step S4 specifically comprises: a reward function definition step S41: in the Markov decision process model, a reward function is defined, and a total reward value is calculated according to the reward function; In some embodiments, the calculation model of the state vector is:

[0047] wherein, is a state vector; is a vector of situation characterization results; is a real-time available resource state vector; is an environmental constraint vector.

[0048] Specifically, the vector of situation characterization results includes risk point coordinates and grades, an influence range vector, and a prediction trend vector; The real-time available resource state vector includes a rescue team position vector, a rescue team state vector, a rescue team capability vector, a material inventory vector, a vehicle position vector, and a vehicle state vector; The environmental constraint vector includes a real-time traffic condition matrix, a weather condition code, a time window constraint, and a priority rule code; By organically combining the vector of situation characterization results, the real-time available resource state vector, and the environmental constraint vector, the state characteristics of the emergency scene are comprehensively depicted, and the multi-objective optimization problem can be solved in real time and adaptively under the conditions of incomplete information and highly dynamic environment, to generate a resource scheduling and action plan that approximates the optimal solution, far exceeding static optimization methods; The perception ability of the Markov decision process model for complex environments is enhanced, and rich information input is provided for the policy network, thereby improving the accuracy and adaptability of the decision, and laying a solid foundation for efficient execution of emergency response.

[0049] In particular, the action vector is defined as ; where the action space comprises discrete and / or continuous decision instruction combinations; For example, a discrete action can be [Dispatch, Team_A, Location_X, Rescue], indicating dispatching Team_A to Location_X for a rescue mission; A continuous or discrete action can be [Allocate, Medical_Supplies, Warehouse_Y, Location_Z, 50], indicating allocating 50 units of medical supplies from Warehouse_Y to Location_Z.

[0050] In particular, the reward function is defined as ; In some embodiments, the calculation model of the total reward value is:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] wherein, is the total reward value; is the success reward value; is the time penalty value; is the resource consumption penalty value; is the constraint violation penalty value; is the risk penalty value.

[0057] is the success reward weight; is the number of people successfully rescued; is the number of risk points eliminated; is the time penalty weight; is the average response delay; is the resource consumption penalty weight; is the resource consumption cost; is the constraint violation penalty weight; is the constraint violation penalty term, such as overtime, entering a forbidden zone; is the risk penalty weight; is the occurrence of secondary disasters; Penalties will be imposed for raising the risk level.

[0058] By using a total reward value calculation model, success rewards, time penalties, resource consumption penalties, constraint violation penalties, and risk penalties are comprehensively considered, providing a comprehensive optimization objective for the training of the policy network. This balances various key factors in emergency response and avoids unreasonable decisions through a penalty mechanism, thereby ensuring the accuracy of the action vectors generated by the policy network.

[0059] Decision evaluation step S42: Train the policy network and evaluation network using the proximal policy optimization algorithm. The policy network generates action decisions by combining the reward function, and the evaluation network evaluates the action decisions by combining the reward function.

[0060] Specifically, the policy network and the evaluation network are defined as follows: It employs a multilayer perceptron (MLP) structure; in the policy network, based on the state vector... Generate action decisions Action planning includes resource allocation and action decision-making instructions; in the evaluation network, action planning is... Conduct an assessment; The online adjustment of the policy network is achieved by simultaneously updating the parameters of the policy network and the evaluation network through an optimizer; the optimizer can be the Adam optimizer. By defining a balancing strategy for learning rate and discount factor, the stability and convergence speed of network updates can be improved. By constructing a Markov decision process model, defining the state vector based on the situation representation results, and introducing action vectors and reward functions, a scientific framework is provided for training the policy network. The near-end policy optimization algorithm generates action vectors by optimizing the policy network, realizing dynamic response to emergency events. It can adapt to the dynamic changes at the disaster site and ensures the rationality and efficiency of decision-making through the guidance of the reward function, significantly improving the intelligence level of emergency response.

[0061] In some embodiments, the decision response step S4 further includes: Model training steps: Construct an emergency response simulation environment based on historical cases and preset physical rules. In the emergency response simulation environment, train the Markov decision process model through the near-end policy optimization algorithm and the experience back-visit pool.

[0062] Optionally, after the Markov decision process model is deployed, online policy fine-tuning can be performed using new data under security constraints.

[0063] Specifically, the preset physical rules may include disaster spread models, resource movement models, and time-progression models; By constructing an emergency response simulation environment and using an experience replay pool to train the Markov decision process model, efficient learning and improved generalization ability of the Markov decision process model are achieved. The simulation environment constructed based on historical cases and preset physical rules provides rich and realistic training scenarios; the experience replay pool mechanism improves data utilization efficiency and enhances the stability of training, accelerates the convergence of the policy network, and enhances the generalization ability of the Markov decision process model through simulation of diverse scenarios, enabling it to adapt to various emergency events, overcoming the problems of insufficient real-world training data and high trial-and-error costs.

[0064] In some embodiments, the emergency event response method further comprises: A decision assistance step: adding a real-time fusion data layer, a situation assessment layer, and a decision instruction layer on the digital twin map platform to display the global emergency situation tensor, the situation characterization result, and the action vector, respectively; A risk point marking layer and a prediction trend animation layer are also added to the digital twin map platform, and visualization tools are displayed, for example, using a WebGL-based GIS engine; Specifically, the digital twin map platform is an interactive geographic information system based on digital twin technology, which reconstructs the emergency scene of the physical world 1:1 in a virtual space through real-time data fusion, dynamic modeling, and visual simulation, supporting disaster monitoring, decision-making, and human-machine collaboration.

[0065] The cross-modal attention weight is fused and explained to show which data or features contribute most to the current situation; The key time nodes are predicted and explained, providing attention-based or feature-based attribution (such as calculating SHAP values using the Captum library), for example, predicting that the area is at high risk, mainly based on upstream water level exceeding the warning, geological vulnerability map, and historical dam collapse records; The action decision is explained, i.e., explaining the reason why the policy network chooses a certain action, for example, sending Team A instead of Team B because Team A is 30% closer and has a higher matching degree of equipment.

[0066] Optionally, on the digital twin map platform, the commander can view the explanation, adjust or veto the decision in the action vector, set or modify the optimization target weight (such as temporarily increasing the priority of a certain area), inject expert rules, and the adjusted decision in the action vector can be fed back to the digital twin map platform for updating the Markov decision process model.

[0067] By displaying the global emergency situation tensor, the situation characterization result, and the action vector in real time on the digital twin map platform, an intuitive visualization tool is provided for the commander; The cross-modal attention weight, key time node and action vector are explained, the transparency and interpretability are enhanced, the scientificity and credibility of the decision are improved, and the optimization efficiency of the emergency response is facilitated.

[0068] Based on the above-mentioned emergency event response method, in the scene of urban heavy rain causing waterlogging and mountain flood geological disaster risk, input real-time radar rainfall map, river water level sensor, low-lying area waterlogging monitoring, AI identification of water depth or vehicle anchoring in traffic camera, NLP extraction of location or danger in social media help information, geological hazard point GIS data, rescue team location or boat number, sandbag number or water pump location or inventory in material warehouse, real-time traffic conditions.

[0069] The cross-modal fusion model fuses rainfall intensity, water level rising trend, water accumulation point image features, and text help information (such as water depth at XX road and waist), and identifies the global emergency situation tensor including A overpass severe water accumulation and B mountain slope landslide high risk.

[0070] The spatio-temporal graph convolution network coupling model predicts that the water accumulation at A bridge will block the main road in the next 2 hours, the landslide probability at B mountain slope reaches 70%, and it may block the underlying river to cause secondary flood, and generates the situation representation result of high-risk area heat map.

[0071] After the strategy network learns in the simulation environment, the real-time action decision is output as: dispatching boat team 1 to A bridge to transfer stranded vehicle personnel, while dispatching engineering team 2 to carry water pumps; dispatching monitoring team 3 to B slope to set up monitoring instruments and prepare to evacuate downstream residents; instructing warehouse 4 to transport sandbags to river weak point C; and issuing detour information through traffic induction screen.

[0072] The commander sees the decision explanation of the action decision on the digital twin map platform, such as dispatching boat 1 because it is closest and has large capacity, combined with on-site feedback, such as A bridge has a patient needing emergency treatment, manually adjusting to increase the medical team to depart with boat 1, which is recorded and used for subsequent Markov decision process model update.

[0073] Through the above-mentioned emergency event response method, the response time is shortened by 35%, the resource scheduling error is improved by 71.8%, the secondary disaster prediction AUC is improved by 26.1%, the decision adoption rate is improved by 46.6%, 15% more people are rescued, the rescue delay caused by traffic paralysis is reduced by 50%, and the secondary disaster loss is avoided by early warning.

[0074] It should be noted that the above is a reference way of an emergency event response method, and the present application is not limited thereto.

[0075] The embodiment of the present application realizes high-precision and interpretable evaluation and prediction of complex emergency situation, provides reliable basis for decision-making, improves real-time and accuracy of emergency response, and solves the technical problems that the prior art cannot realize semantic-level fusion, leads to situation awareness fragmentation, a single model is difficult to capture the spatio-temporal evolution law of disaster, the prediction result is poor in real-time and uninterpretable, a static optimization model cannot dynamically adjust the strategy, leading to lag or mismatch of resource scheduling, the black-box decision of AI lacks transparency and interpretability, commanders cannot understand the logic or intervene, and the emergency efficiency is reduced.

[0076] It should be noted finally that: each of the embodiments in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. The above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application; although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent; without departing from the spirit of the technical scheme of the present application, they should be covered in the technical scheme range of the present application claimed by the present application.

Claims

1. An emergency event response method, characterized by, The method comprises the following steps: a data acquisition step: acquiring multi-modal emergency data of a collection end through a preset interface; a feature fusion step: inputting the multi-modal emergency data into a preset cross-modal fusion model to perform feature extraction, attention weight calculation and data fusion, and obtaining a global emergency situation tensor; a situation representation result output step: inputting the global emergency situation tensor into a preset spatio-temporal graph convolution network coupling model, extracting spatial features based on a spatial dimension to obtain a spatial feature vector, extracting temporal features based on a time dimension to obtain a temporal feature vector, and obtaining a situation representation result according to the spatial feature vector and the temporal feature vector; a decision response step: constructing a Markov decision process model, defining an action vector, defining a state vector according to the situation representation result, generating an action decision through a preselected policy network according to the state vector and the action vector, and responding to an emergency event according to the action decision through the Markov decision process model to obtain a response result.

2. The emergency event response method of claim 1, wherein, The feature fusion step specifically comprises: a feature extraction step: inputting the multi-modal emergency data into a modal-specific encoder in the cross-modal fusion model to perform modal encoding and feature extraction, and obtaining multi-modal features; an attention weight calculation step: performing cross-modal attention weight calculation on the multi-modal features through a cross-modal attention layer in the cross-modal fusion model to obtain a weight matrix, and performing weighted aggregation on the weight matrix and the multi-modal features to obtain cross-modal interaction features; a data fusion step: performing feature splicing or weighted fusion on the cross-modal interaction features and the multi-modal features to obtain the global emergency situation tensor.

3. The emergency event response method of claim 2, wherein, The situation representation result output step specifically comprises: a spatial feature extraction step: in an STGCN layer of the spatio-temporal graph convolution network coupling model, generating a graph structure according to a spatial topology structure, performing graph convolution operation on the global emergency situation tensor, and extracting spatial features based on a spatial dimension to obtain a spatial feature vector; a temporal feature extraction step: adding a time position encoding at an output end of the STGCN layer, performing spatio-temporal encoding on the spatial feature vector according to the time position encoding, calculating correlation weights of time steps through a self-attention mechanism in a Transformer layer of the spatio-temporal graph convolution network coupling model to dynamically analyze key time nodes, extracting temporal features based on a time dimension through the Transformer layer to obtain a temporal feature vector.

4. The emergency event response method of claim 3, wherein, The spatial feature extraction step specifically comprises: generating a graph structure according to a spatial topology structure, and the graph structure is represented by an adjacency matrix as follows: wherein, is a node is a node is a weight of adjacency between nodes is a node is a node is a geographical distance between nodes, the nodes being geographical entities related to the emergency event; is a scale parameter controlling the decay of the weight.

5. The emergency event response method of claim 3, wherein, The calculation model of the time position encoding is as follows: wherein, is a time position encoding; is a time step; is a dimension index; is a total number of feature dimensions.

6. The emergency event response method of claim 3, wherein, The method further comprises: a decision assistance step: adding a real-time fusion data layer, a situation assessment layer and a decision instruction layer on a digital twin map platform, and displaying the global emergency situation tensor, the situation representation result and the action vector respectively.

7. The emergency event response method of any one of claims 1-6, wherein, The decision response step specifically comprises: a reward function definition step: defining a reward function in the Markov decision process model, and calculating a total reward value according to the reward function; The decision evaluation step comprises: training a policy network and an evaluation network by a proximal policy optimization algorithm, the policy network generating an action decision in combination with the reward function, and the evaluation network evaluating the action decision in combination with the reward function.

8. The emergency event response method of claim 7, wherein, A calculation model of the state vector is: wherein, is a state vector; is a vector of situation representation results; is a real-time available resource state vector; is an environmental constraint vector.

9. The emergency event response method of claim 7, wherein, A calculation model of the total reward value is: wherein, is a total reward value; is a success reward value; is a time penalty value; is a resource consumption penalty value; is a constraint violation penalty value; is a risk penalty value.

10. The emergency event response method of claim 7, wherein, The decision response step specifically further comprises: A model training step: constructing an emergency response simulation environment according to historical cases and preset physical rules, and training the Markov decision process model in the emergency response simulation environment by the proximal policy optimization algorithm and an experience replay pool.

Images