Intelligent processing method, device and equipment for emergencies and storage medium

Abstract

The invention relates to an intelligent processing method, device and equipment for emergencies and a storage medium, and relates to the technical field of emergency processing. The method comprises the following steps: receiving a multi-modal data stream from an emergency scene; and performing semantic fusion and understanding processing on the multi-modal data to generate a structured real-time situation map. And through a preset intelligent deduction algorithm, carrying out strategy deduction processing on the real-time situation map, and generating a disposal strategy containing a disposal action sequence. Determining a target execution device in a preset heterogeneous execution device set according to the processing strategy; the processing strategy is analyzed and packaged, an executable instruction matched with the target execution equipment is generated, and the executable instruction is issued to the target execution equipment, so that the target execution equipment processes the emergency according to the received executable instruction. The problems existing in the prior art are solved, the system expansion cost of the equipment resource pool can be reduced, and the flexibility of the dynamic control equipment can be ensured under the scene of processing emergencies.

Description

Technical Field

[0001] This invention relates to the field of emergency response technology, and in particular to an intelligent emergency response method, apparatus, equipment, and storage medium. Background Technology

[0002] Currently, in the field of emergency command for sudden incidents, in order to quickly develop response plans for sudden incidents, it is necessary to understand the actual situation of the sudden incident in real time and make correct decisions.

[0003] In existing technologies, the system internally includes a rule base or knowledge base based on an "event type-response action" mapping relationship and integrates a geographic information system. When an emergency occurs, command personnel manually determine the event type based on the aggregated information and input it into the system to generate decision instructions. Then, command personnel manually draw polygons on a map or select preset administrative divisions to select the target area that needs to be covered by the response information. Based on the selected geographical area, the system automatically retrieves a list of all registered terminal devices with matching types within that area. Subsequently, the system sends the decision instructions generated in the previous step to the target devices for execution through dedicated interfaces pre-configured for various types of devices, such as API calls or socket communication with specific data formats.

[0004] However, in existing technologies, command issuance relies on dedicated interfaces pre-configured for different manufacturers' devices. This "siloed" integration architecture leads to significant technical barriers. Whenever a new model or manufacturer's device needs to be integrated, independent development and interfacing are required, resulting in high system expansion costs, long development cycles, and difficulty in forming a flexible and unified device resource pool. More importantly, this point-to-point command translation logic is hard-coded, tightly coupling upper-layer strategies with lower-layer execution. This makes it difficult to dynamically and intelligently orchestrate complex linkage strategies for multiple devices to work together based on field conditions, and hinders flexible dynamic control of devices to handle unexpected events. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides an intelligent method, apparatus, device, and storage medium for handling emergencies, thereby resolving the problems existing in the prior art, reducing the system expansion cost of the device resource pool, and ensuring the flexibility of dynamic control of the device in handling emergency scenarios.

[0006] Firstly, an intelligent method for handling emergencies is provided, which may include: Receive multimodal data streams from the scene of an emergency; perform semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map; The real-time situation map is processed by a preset intelligent inference algorithm to generate a response strategy that includes a sequence of response actions. According to the handling strategy, a target execution device is determined from a preset set of heterogeneous execution devices; the handling strategy is parsed and encapsulated to generate an executable instruction that matches the target execution device, and the executable instruction is sent to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

[0007] Secondly, an intelligent emergency response device is provided, which may include: The fusion module is used to receive multimodal data streams from the scene of an emergency; perform semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map; The deduction module is used to perform strategy deduction processing on the real-time situation map through a preset intelligent deduction algorithm to generate a disposal strategy containing a sequence of disposal actions. The delivery module is used to determine the target execution device from a preset set of heterogeneous execution devices according to the handling strategy; to parse and encapsulate the handling strategy to generate an executable instruction that matches the target execution device; and to deliver the executable instruction to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

[0008] Thirdly, an electronic device is provided, which includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described in the first aspect above.

[0009] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.

[0010] This application provides an intelligent handling method, apparatus, device, and storage medium for emergencies. It receives multimodal data streams from the emergency scene; performs semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map. Through a preset intelligent inference algorithm, it performs strategy inference processing on the real-time situation map to generate a handling strategy containing a sequence of handling actions. Based on the handling strategy, a target execution device is determined from a preset set of heterogeneous execution devices; the handling strategy is parsed and encapsulated to generate executable instructions matching the target execution device, and these executable instructions are sent to the target execution device so that the target execution device can handle the emergency according to the received executable instructions. The beneficial effects of this invention are that it breaks down the technical barriers between different types of physical devices in a heterogeneous execution device set, constructs a task-oriented autonomous collaborative capability between heterogeneous devices, significantly improves the utilization efficiency of on-site resources and the smoothness of task execution, realizes intelligent and refined emergency response, solves the aforementioned problems existing in the prior art, reduces the system expansion cost of the device resource pool, and ensures the flexibility of dynamically controlled devices in emergency response scenarios. Attached Figure Description

[0011] Figure 1 A flowchart illustrating an intelligent method for handling emergencies provided by the present invention; Figure 2 A flowchart illustrating an intelligent method for handling emergencies provided by the present invention; Figure 3 A flowchart illustrating an intelligent emergency response device provided by the present invention; Figure 4 An electronic device provided by the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Currently, in the field of emergency command for sudden incidents, in order to quickly develop response plans for sudden incidents, it is necessary to understand the actual situation of the sudden incident in real time and make correct decisions.

[0014] In one example, the system has a pre-built rule base or knowledge base based on the mapping relationship between "event type and response action" and integrates a geographic information system. When an emergency occurs, the command personnel manually determine the type of the event based on the aggregated information and input it into the system to generate decision instructions. Then, the command personnel select the target area that needs to be covered by the response information on the map by manually drawing polygons or selecting preset administrative divisions. Based on the selected geographical area, the system automatically retrieves a list of all registered terminal devices with matching types within that area. Subsequently, the system sends the decision instructions generated in the previous step to the target devices for execution through dedicated interfaces pre-configured for various types of devices, such as API calls or socket communication with specific data formats.

[0015] However, in existing technologies, command issuance relies on dedicated interfaces pre-configured for different manufacturers' devices. This "siloed" integration architecture leads to significant technical barriers. Whenever a new model or manufacturer's device needs to be integrated, independent development and interfacing are required, resulting in high system expansion costs, long development cycles, and difficulty in forming a flexible and unified device resource pool. More importantly, this point-to-point command translation logic is hard-coded, tightly coupling upper-layer strategies with lower-layer execution. This makes it difficult to dynamically and intelligently orchestrate complex linkage strategies for multiple devices to work together based on field conditions, and hinders flexible dynamic control of devices to handle unexpected events.

[0016] In one example, regarding the aggregation and presentation of multi-source information, existing technical solutions can aggregate data from different business systems through standard interfaces during the event perception and recognition phase. For instance, alarm text records from the public hotline center can be accessed via a telephone gateway, video streams from designated cameras can be accessed via video platform protocols, and preset urban component layers, such as the locations of fire hydrants and shelters, can be loaded via a geographic information system. This information is presented on the command center's interface in the form of raw, independent data streams for commanders to manually review and make comprehensive judgments. However, while existing technical solutions can aggregate multi-source data, their processing method remains at the level of "data presentation" rather than "information understanding." For example, the anxious tone in an alarm call, the trend of smoke spreading in on-site images, and the rapidly changing data from environmental sensors are all isolated data items in the system. The system itself cannot automatically correlate them and infer that "this is a rapidly deteriorating fire, and there may be panic among people at the scene." This deep semantic correlation and judgment relies entirely on the commander's brain, which, under information overload and high-pressure environments, can easily lead to delays in judgment or biased judgments due to the omission of a key piece of information.

[0017] In one example, for manual response scheduling based on a preset template, the system has a pre-built rule base or knowledge base based on the mapping relationship between "event type - response action" during the decision-making and instruction generation phase. When the command personnel manually determine the type of event based on the aggregated information, such as "Level 1 fire" or "Level 2 flooding," they retrieve the corresponding response plan template from the system. This template typically contains a series of fixed response procedures and parameter fields to be filled in. The command personnel need to manually fill in the key information of the event, such as the coordinates of the incident location, the time of occurrence, and the scope of impact, into the corresponding fields of the template to generate specific response content, such as broadcast announcements or on-screen text. However, the system heavily relies on the preset template library. The fundamental flaw of this model lies in its rigidity. For complex events that exceed the scope of the plan and involve multiple concurrent disasters (e.g., a gas leak occurring on a flooded main road), the system often lacks a directly applicable template. At this point, the generation of response strategies degenerates entirely into the on-the-spot creation of commanders, and their quality is highly dependent on personal knowledge and experience. The lack of systematic support and constraints may lead to oversights or inappropriate response measures.

[0018] In one example, for execution receipts based on device status, the system has the capability to receive terminal device status receipts after deployment. After receiving an instruction or completing a playback / display action, the terminal device sends a basic status code or signaling back to the central system. This receipt information typically only includes the device's own operational status, such as "device online," "instruction received," or "task executed," to prove that the instruction link from the center to the terminal is smooth and the device is available. However, the system can only obtain basic receipts at the "instruction delivered" level. This technical reality leads to a critical missing loop: the command center cannot know whether the response information has truly and effectively impacted the target. Due to the lack of perception and quantitative evaluation methods for the actual effectiveness of the response actions, the entire emergency response process remains in an open-loop state of "deployment ends immediately," unable to make timely strategy adjustments based on on-site feedback, and unable to accumulate effective data for post-event review and process optimization.

[0019] Example 1: Reference Figure 1 , Figure 1 This is a flowchart illustrating an intelligent method for handling emergencies provided in an embodiment of this application. Figure 1 As shown, the method may include: Step S101: Receive multimodal data streams from the scene of the emergency; perform semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map.

[0020] For example, the executing entity of this embodiment can be an electronic device, a terminal device, an intelligent processing device or device for emergencies, or other device or device capable of executing this embodiment, and there is no limitation thereto. In this embodiment, the executing entity is described as an electronic device.

[0021] In this step, multiple multimodal data streams from the scene of the emergency are continuously received, such as real-time video streams from surveillance cameras, text information streams from social media platforms and emergency call centers, and time-series data streams from IoT environmental sensors, such as smoke detectors and crowd density meters. To overcome the "semantic gap" between different modalities—that is, the inherent heterogeneity in feature representation among various data types—a simple information splicing approach was not adopted. Instead, a hierarchical cross-modal attention fusion network was designed and implemented. Its purpose is to project feature vectors from different sources into a unified, high-dimensional joint semantic space, achieving deep association and alignment within this space. Specifically, based on this network, semantic fusion and understanding processing of multimodal data is performed to generate a structured real-time situation map. This structured, machine-readable real-time situation map fully encapsulates a deep semantic understanding of the event's "who, what, when, where, and how," and is then used as core input, passed in real-time to subsequent dynamic strategy generation steps, laying a solid cognitive foundation for scientific and accurate emergency strategy simulation.

[0022] Step S102: Using a preset intelligent deduction algorithm, perform strategy deduction processing on the real-time situation map to generate a disposal strategy containing a sequence of disposal actions.

[0023] For example, a real-time updated situational map, in the form of a graph data structure, accurately depicts the entities, events, and their dynamic relationships at the scene of an emergency. Faced with rapidly changing and complex situations, simple rule-based or template-based responses are no longer sufficient to meet the demands for efficient and precise handling. Therefore, within a vast decision-making space, a pre-set intelligent inference algorithm is used to perform strategy inference processing on the real-time situational map, quickly searching for and inferring the optimal or near-optimal sequence of actions to achieve predetermined emergency objectives (such as personnel rescue, fire control, and secondary disaster containment) in the shortest possible time and at the lowest cost, ultimately generating a response strategy containing the sequence of actions.

[0024] Step S103: According to the handling strategy, determine the target execution device in the preset heterogeneous execution device set; parse and encapsulate the handling strategy to generate an executable instruction that matches the target execution device, and send the executable instruction to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

[0025] For example, the physical devices in the heterogeneous execution device set correspond one-to-one with the virtual device entities in the virtual execution device set. The virtual devices are all pre-converted into standardized interfaces that are independent of electronic devices and can be called over the network. Thus, the electronic devices no longer need to care about how to call the underlying SDK of a specific brand of drone, but only need to publish a standardized service request to the virtual device entity. The physical device corresponding to the virtual device entity can then accept and execute the executable instructions.

[0026] In this step, based on the task requirements of the multi-objective optimization model and the handling strategy, a capability service combination scheme that meets the preset constraints is searched and output from the virtual execution device set corresponding to the preset heterogeneous execution device set. Then, the handling strategy is parsed into standard instructions that match the capability service combination scheme, the standard instructions are converted into an instruction format executable by the target execution device, and the converted standard instructions are encapsulated and sent to the target execution device so that the target execution device can handle the sudden event according to the received executable instructions.

[0027] The method provided in this application receives a multimodal data stream from the scene of an emergency; performs semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map. Through a preset intelligent inference algorithm, the real-time situation map is processed to generate a response strategy containing a sequence of actions. Based on the response strategy, a target execution device is determined from a preset set of heterogeneous execution devices; the response strategy is parsed and encapsulated to generate executable instructions matching the target execution device, and these executable instructions are sent to the target execution device so that it can handle the emergency according to the received instructions. The beneficial effects of this invention are that it breaks down technical barriers between different types of physical devices in a heterogeneous execution device set, constructs a task-oriented autonomous collaborative capability between heterogeneous devices, significantly improves the utilization efficiency of on-site resources and the smoothness of task execution, realizes intelligent and refined emergency response, solves the aforementioned problems existing in the prior art, reduces the system expansion cost of the device resource pool, and ensures the flexibility of dynamically controlled devices in emergency response scenarios.

[0028] Example 2: Figure 2 A flowchart illustrating an intelligent emergency response method provided in this application is shown below. Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the method is described in detail below, and the method includes: Step S201: Receive multimodal data streams from the scene of the emergency; perform semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map.

[0029] In one example, S201 includes: extracting feature vectors of each modality of data through a preset deep neural network; calculating the correlation weights between feature vectors of different modalities based on a preset cross-modal attention mechanism; fusing the feature vectors based on the correlation weights through a preset gating fusion mechanism to generate fused features; and generating a structured real-time situation map based on the fused features; wherein the real-time situation map includes nodes used to describe entities and edges used to describe the relationships between entities.

[0030] For example, multiple multimodal data streams from the event site are continuously received. Deep features are extracted from each individual modality of data using a pre-defined deep neural network. For instance, for video data, spatial and temporal visual features are extracted using convolutional neural networks and temporal networks to obtain a sequence of visual feature vectors. ,in, Representative from video number Extracted from keyframes or segments The feature vector is dimensional, where n is the sequence length. For text data, a pre-trained language model, such as BERT, is used for encoding to obtain a set of text feature vector sequences. ,in, The first in the representative text Each word or sentence The feature vector is dimensional, where m is the sequence length. A key step in the fusion process is the cross-modal attention mechanism. Taking text feature enhancement of visual features as an example, the network uses the text feature sequence T as the "query" and the visual feature sequence V as the "key" and "value," calculating a text-guided visual context representation. This process can be defined as: , In this formula, , and It is a learnable linear projection matrix that maps the original text features and visual features to the query space, key space, and value space, respectively. This is the dimension of the key vector, used to scale the dot product result to prevent the gradient from becoming too small. The essence of this attention mechanism is to calculate the association weight between each semantic unit in the text and each visual segment (key) in the video, and then perform a weighted summation of the visual segments (values) based on this weight, thereby obtaining a visual context representation that is relevant to the text content. Conversely, a visually guided textual context representation can also be computed. .

[0031] To achieve more refined fusion and prevent interference from irrelevant information, a gating fusion mechanism was further introduced. After obtaining cross-modal context representations (e.g., Instead of directly adding it to the original feature, it uses a dynamically calculated gating coefficient. To perform adaptive weighting and update visual features. For example, the new features after fusion The calculation is as follows: in, Is with Aligned text context representation, It is after linear transformation By matching dimension; This indicates concatenating two vectors; and These are the weights and biases of the gated network; It is the Sigmoid activation function, and its output range is... This represents a "door" that is open or closed; This represents element-wise multiplication. This gating mechanism allows the model to dynamically determine, based on current data, the extent to which original visual information is retained and the extent to which contextual supplementation from textual information is incorporated, thereby achieving intelligent information filtering and enhancement. Through pairwise and higher-order hierarchical fusion of all modal features, a set of fused feature vectors aligned and mutually reinforcing in a unified semantic space is ultimately obtained. Based on this set of fused features, entity recognition, event detection, and relation extraction are further performed, automatically constructing and updating in real-time a dynamic knowledge graph describing the global situation of the event. This structured, machine-readable situation graph fully encapsulates a deep semantic understanding of the event's "who, what, when, where, and how," and is subsequently used as core input, passed in real-time to subsequent dynamic strategy generation steps, laying a solid cognitive foundation for scientific and accurate emergency strategy simulation.

[0032] Step S202: Using a preset intelligent deduction algorithm, perform strategy deduction processing on the real-time situation map to generate a disposal strategy containing a sequence of disposal actions.

[0033] In one example, S202 includes: encoding the real-time situation map using a preset graph attention network to obtain graph features of nodes and relationships, as well as action prior probabilities; using the encoded graph features as input states, performing policy deduction processing using a Monte Carlo tree search algorithm and action prior probabilities; and outputting at least one disposal policy whose estimated utility satisfies the preset estimated utility conditions through iterative search.

[0034] For example, in the first approach to generating the disposition strategy, this step provides a Monte Carlo Tree Search Inference Algorithm (GI-MCTS algorithm) based on graph attention networks. This algorithm organically combines policy evaluation and planning search in reinforcement learning, and is particularly suitable for handling complex decision problems where the state is represented by a graph structure. Specifically, at any decision time t, the current state can be precisely defined as a state. That is, the situation map generated by the previous module. ,in, It is a set of nodes in the graph (representing entities such as people, vehicles, facilities, and fire points). It is a set of edges (representing spatial, logical, or causal relationships between entities). Action space. It consists of a series of parameterized macro commands, such as "dispatch (unit type, quantity, target location)", "evacuate (area, route)" or "establish isolation zone (center point, radius)".

[0035] The GI-MCTS algorithm iteratively executes four steps: selection, expansion, simulation, and backtracking. Within a finite computational time, it constructs an asymmetric search tree to explore the potential situational evolution caused by different future action sequences. In the selection phase, the algorithm starts from the root node (representing the current state). Starting with a confidence upper bound strategy, child nodes are recursively selected until a leaf node that has not been fully expanded is reached. This strategy aims to balance the "exploitation" of known high-reward paths with the "exploration" of unknown paths. Next action is selected. The formula is as follows: , in, It is in the state The estimate of the long-term value of the following action a represents the "utilization" portion. It is a state-action pair The number of visits. c is an exploration constant used to adjust the weight of exploration. It is a priori probability, calculated by a policy network based on the current situation map. As input, it directly outputs the recommended probability of each candidate action under the current situation; the recommended probability is the prior probability. This allows the search to favor more promising directions from the beginning based on the prior probability, greatly improving search efficiency.

[0036] To enable the policy network to deeply understand the complex situations inherent in the graph, a graph attention network (GAT) is used as its core encoder. The GAT dynamically assigns different attention weights to the neighbors of each node in the graph, thereby capturing the differences in importance between different entities. For any node i in the graph, its feature vector h... i The update process is as follows: First, calculate the relationship between node i and its neighboring nodes. Attention coefficient between : , Where W is a learnable linear transformation weight matrix used to map input features to a higher-dimensional space; This represents a vector concatenation operation; 'a' is a learnable weight vector of a single-layer feedforward neural network used to map the concatenated node features to a scalar; LeakyReLU is a modified linear unit activation function with leakage. This represents the transpose of 'a'; This represents the input feature vector of node j; This represents the input feature vector of node i; This represents the input feature vector of node k; The whole represents the set of neighbors of node i. This attention coefficient... The importance of information from node j to node i was measured. Subsequently, new features of node i were discussed. It is obtained by weighted summation of the features of its neighbors: , in, It is a non-linear activation function. By stacking multiple layers of GAT, the network can aggregate increasingly wider ranges of neighborhood information, thereby gaining global context awareness of the entire graph. Finally, by pooling the features of all nodes, the representation vector of the entire graph is obtained. This vector is then fed into the fully connected layer of the policy network to output the probability distribution. .

[0037] During the expansion phase, the algorithm creates one or more new child nodes for selected leaf nodes, corresponding to actions executable from that state. During the simulation phase, the algorithm starts from the newly expanded nodes and performs a rapid "deduction" without storing intermediate nodes until a terminating state is reached, such as event resolution or the maximum deduction depth is reached. Traditional MCTS deductions are stochastic, while the innovation of this scheme lies in the fact that the deduction result does not come from a random policy, but is directly predicted by a value network that shares the GAT backbone network with the policy network. The value network takes a state as input and outputs a scalar value. This represents the degree of superiority or inferiority of the state. This avoids time-consuming and unstable stochastic simulations, making the evaluation of the value of leaf nodes more accurate and efficient.

[0038] Finally, during the backtracking phase, the simulated reward value or the value predicted by the value network is used to update the values ​​of all nodes on the path from the leaf node to the root node from the bottom up. and After numerous iterations, the statistics of each child node under the root node—that is, the currently executable initial actions—will tend to stabilize. At this point, action 'a' with the highest number of visits is considered the current optimal decision. Finally, the output is not a single instruction, but a sorted list of contingency plans containing multiple handling strategies. Each strategy includes a detailed action sequence and an estimated utility value. This structured, quantifiable contingency plan will serve as the final decision instruction, seamlessly transmitted downstream for feasibility analysis, resource matching, and task assignment, thus forming a complete closed loop from perception, cognition, decision-making to execution.

[0039] Therefore, it balances the "speed" and "quality" of decision-making. The rapid initiation of case-based reasoning significantly shortens the initial decision-making time for emergency response; while the refined deduction through Monte Carlo tree search ensures the adaptability and optimality of the final solution to the current situation. This combination effectively avoids the shortcomings of a single algorithm, making the generated contingency plan reliable, efficient, and flexible and creative in dealing with unknown and unexpected situations.

[0040] Optionally, in the second approach to generating response strategies, an end-to-end strategy generation scheme based on reinforcement learning can also be used. Specifically, the original scheme employs a two-stage model of "case-based reasoning + Monte Carlo tree search." This alternative model models the entire emergency command problem as a Markov Decision Process (MDP) and directly trains a policy network using deep reinforcement learning algorithms (such as PPO, DQN, etc.). This network takes real-time situational information as input and directly outputs the optimal response action (or action sequence). The training process can be conducted in a highly realistic environment. Through extensive trial-and-error learning, the agent can autonomously discover the optimal command strategy without relying on a pre-set case library or explicit inference search. This approach also achieves the technical effect of dynamically generating scientific decisions based on the situation, and through end-to-end learning, it may discover innovative tactics that surpass human experience, achieving a higher level of intelligent decision-making.

[0041] Optionally, in the third approach to generating response strategies, a scheme based on Planning Domain Definition Language (PDDL) and an automated planner can also be adopted. Specifically, in this alternative, the emergency command domain is formally described as a PDDL problem, including a series of "objects" (such as firefighters, fire trucks, and fire points), "predicates" (describing the world state, such as at(fireman, locationA)), and "actions" (defining the preconditions and effects of the actions, such as move(fireman, from, to)). When a task objective is received, an automated planner (such as FF Planner or OPTIC) automatically searches for a sequence of actions from the initial state to the target state; this sequence constitutes the response plan. This approach also achieves the technical effect of automatically generating response plans, with the advantages of logical rigor and interpretability of the results. Each step generated has a clear causal relationship, ensuring the logical completeness of the plan.

[0042] Step S203: Based on the handling strategy, determine the target execution device from the preset heterogeneous execution device set.

[0043] In one example, S203 includes: determining the task requirements for the disposal strategy; based on the task requirements, and according to a multi-objective optimization model, searching and outputting a capability service combination scheme that satisfies preset constraints in the virtual execution device set corresponding to the preset heterogeneous execution device set; wherein, the physical devices in the heterogeneous execution device set correspond one-to-one with the virtual device entities in the virtual execution device set; and determining the target execution device according to the virtual execution device corresponding to the capability service combination scheme.

[0044] For example, this step parses the dynamically generated response strategy into a series of specific, logical task instructions, such as "conduct high-altitude fire reconnaissance in area A" or "deliver 50 kg of relief supplies to point B." However, the physical equipment at the emergency site, such as drones, robots, sensors, and communication base stations, comes from diverse sources, has different models, and uses vastly different communication protocols and data formats. This inherent heterogeneity constitutes a significant obstacle to the effective execution of command instructions. The core mission of this step and step S204 is precisely to eliminate this heterogeneity gap by abstracting and encapsulating the complex capabilities of physical equipment into unified, standardized, and programmable software services, thereby providing a transparent and consistent resource view for upper-layer applications.

[0045] Specifically, in the first encapsulation scheme, the workflow of this step begins with the dynamic access and registration of the physical device. When any new physical device, such as a multi-rotor drone or a ground environmental monitoring station, enters the emergency command network, the electronic device first authenticates and declares its capabilities through a pre-defined device gateway, and then creates a corresponding virtual device entity for the new physical device. This virtual device entity is a structured digital object that not only mirrors the real-time status of the physical device, such as its geographical location, battery level, network connection status, and payload, but more importantly, it formally defines all the functions that the device can perform through a standardized capability description language. For example, the capability set of a drone might be described as including "flying to a designated coordinate," "cruising along a pre-defined route," "collecting visible light / infrared video streams," and "dropping payloads." Each capability is encapsulated as an independent capability service.

[0046] The essence of this service-oriented encapsulation is to transform operation instructions for specific hardware into standardized interfaces that are independent of electronic devices and can be invoked over the network. Electronic devices no longer need to worry about how to call the underlying SDK of a specific brand of drone; instead, they only need to publish a standardized service request to the virtual device entity corresponding to the drone, such as `Invoke("FlyTo",{lat:39.9,lon:116.3,alt:100,speed:15})`. The protocol translation engine inside the electronic device is then responsible for converting this standardized service request into proprietary instructions that the drone can understand and sending them through the appropriate communication link.

[0047] For complex tasks assigned from upstream, the capabilities of a single device often cannot be completed independently. For example, the task of "creating a panoramic model of a fire scene" may require multiple drones to collaboratively collect data from different altitudes and angles. In this case, how to intelligently select and combine capability services from numerous available virtual device entities to complete the complex task in the optimal way becomes a core combinatorial optimization problem. To address this, this embodiment proposes a task-oriented multi-objective capability combinatorial optimization algorithm.

[0048] The algorithm aims to solve a given task requirement R using all currently available virtual devices. In this process, find an optimal combination of capability services. The task requirement R itself is defined as a set that includes functional requirements, such as "acquiring video streams" and "transporting goods", and non-functional requirements, such as "video resolution not less than 4K", "transport load not less than 20kg", and "task completion time not exceeding 10 minutes".

[0049] The core of the algorithm lies in constructing and solving a multi-objective optimization model. First, a decision variable is defined. It is a binary variable if and only if the device Capability Services (in It is equipment When a set of capability services is selected to constitute a solution. The value is 1 if it is not 0 otherwise.

[0050] The merits of a solution typically need to be evaluated from multiple dimensions, such as time cost, energy cost, and execution risk. Therefore, the objective function is designed as a comprehensive utility function $U(S)$, aiming to minimize a weighted combined cost. This function is expressed as: , Here, S represents a candidate solution consisting of all the selected capability services. These are weighted coefficients for time, energy, and risk, which are dynamically set by the commander based on the urgency and nature of the current mission (e.g., during the golden time for life-saving operations). The value will be very high. These represent normalized estimates of the time required to perform capability service c, the energy consumed, and the associated risks, respectively. These estimates are derived from static capability parameters and dynamic state information stored in the virtual device entity. For example, the time it takes for a drone to perform a flight mission. It is not only related to distance, but also to dynamic factors such as current battery level and wind speed.

[0051] This optimization process must be carried out under a series of preset constraints: Functional completeness constraint: Ensure that all selected capability services combined can fully cover all functional requirements of task R.

[0052] Quality compliance constraints: Ensure that for each functional requirement that is met, its related non-functional indicators (such as accuracy and resolution) meet or exceed the thresholds specified in task R.

[0053] Device mutual exclusion and capacity constraints: Ensure that each device executes only one task at a time, and that the resource requirements of the task, such as load and computing power, do not exceed its maximum carrying capacity.

[0054] By solving this integer programming problem, the algorithm can output an optimal capacity service combination scheme. The plan clarifies which specific capabilities of which equipment need to be utilized, and finally, based on the virtual execution devices corresponding to the capability service combination plan, determines the target execution devices.

[0055] Optionally, in the second encapsulation scheme, a capability description and discovery scheme based on semantic Web technology can also be used. Specifically, the original scheme uses JSON / XML format to define device capabilities, which is a syntactic-level standardization. This alternative scheme uses a Web ontology language and resource description framework to build an ontology knowledge base for the emergency response domain. In this scheme, each device's capability, such as "thermal imaging detection" or "water cannon spraying," is no longer an isolated string, but a "class" or "instance" defined in the ontology library, possessing rich attributes such as detection range, spray flow rate, applicable media, and relationships; for example, "water cannon spraying" is a type of "fire extinguishing capability." When the upper-level task requires "suppressing Class A fires," the system can use semantic query languages ​​such as SPARQL to infer and discover all devices with "Class A fire extinguishing capabilities," rather than simply matching keywords. This scheme also achieves the technical effect of shielding the heterogeneity of underlying devices and supporting task-oriented autonomous collaboration, and through semantic reasoning, it can achieve more intelligent and accurate capability matching and resource discovery.

[0056] Optionally, in the third encapsulation scheme, an interface definition and service governance scheme based on gRPC / Protobuf can also be used. This scheme can replace the service-oriented model based on RESTful APIs and a registry center in the original scheme. In this alternative, each device capability is defined as a gRPC service, and its interface is strictly defined through a protocol buffer. Protobuf is a language-independent, platform-independent binary serialization format that is more efficient and type-safer than JSON / XML. The command center and devices communicate via gRPC for efficient remote procedure calls. Service discovery can be implemented using mature service governance frameworks such as Consul or Etcd. This scheme can also encapsulate device capabilities into standardized, remotely invokeable services, achieving the technical effects of hardware decoupling and unified scheduling, and has advantages in communication efficiency, interface robustness, and service governance maturity.

[0057] Step S204: The handling strategy is parsed and encapsulated to generate an executable instruction that matches the target execution device, and the executable instruction is sent to the target execution device so that the target execution device can handle the emergency according to the received executable instruction.

[0058] In one example, S204 includes: parsing the disposal strategy into standard instructions that match the capability service combination scheme; wherein the capability service combination scheme is determined based on the task requirements and multi-objective optimization model of the disposal strategy; converting the standard instructions into an instruction format executable by the target execution device, encapsulating the standard instructions that have been converted into the instruction format, and issuing the encapsulated instructions to the target execution device.

[0059] For example, the execution engine of an electronic device transforms a capability service composition scheme into a series of concurrent or serial service call sequences targeting specific virtual device entities. Specifically, it parses the handling strategy into standard instructions that match the capability service composition scheme, converts the standard instructions into an instruction format executable by the target execution device, encapsulates the standard instructions that have been converted into the instruction format, and issues the encapsulated instructions to the target execution device.

[0060] Optionally, during instruction execution, the real-time status of each physical device is continuously monitored, and this status is fed back to the virtual device entity. In the event of an anomaly such as device failure, battery depletion, or communication interruption, the virtual device entity status of that device is immediately updated, such as setting availability to "offline." This status change triggers a replanning, and the multi-objective capability combination optimization algorithm will be based on the updated set of available devices. The system is restarted, quickly generating alternative solutions, thus ensuring the robustness and adaptability of the entire emergency response system. Ultimately, this series of meticulously orchestrated and dynamically adjusted underlying operations collectively realize the macro-strategic intentions communicated from the upper levels, forming a seamless closed loop from abstract decision-making to concrete physical action. Step S205: Collect execution feedback information corresponding to executable instructions.

[0061] For example, execution feedback information corresponding to executable instructions is collected and fused in real time, i.e., multi-source heterogeneous information. This multi-source heterogeneous information includes, but is not limited to: field data streams transmitted from front-end devices such as drones, ground robots, and various sensors, such as high-definition video, infrared thermal imaging, gas concentration, and vital signs; task execution status reports, such as task success, failure, interruption, and resource consumption; and manual observation information that may be input by on-site command personnel. This raw, fragmented data is first transformed into a structured information stream under a unified spatiotemporal benchmark through an internal data preprocessing and alignment engine, laying the foundation for subsequent in-depth performance analysis.

[0062] Step S206: Evaluate the effectiveness of the handling strategy based on the execution feedback information.

[0063] In one example, S206 includes: constructing a dynamic Bayesian network model that describes the state evolution and impact of response actions at the scene of an emergency; based on the execution feedback information and the dynamic Bayesian network model, performing probabilistic inference on the preset core emergency states to obtain the inference results; and based on the inference results, evaluating the execution effectiveness of the response strategy to generate a comprehensive effectiveness score.

[0064] For example, in the first scheme of the effectiveness prediction and evaluation scheme, a dynamic Bayesian network-based algorithm for inferring the effectiveness of response measures is used to probabilistically infer the macroscopic effectiveness of response strategies in complex environments full of uncertainty. The core idea of ​​this algorithm is to construct a time-series probabilistic graphical model that reflects the causal relationship of the emergency event's evolution, and to continuously update the understanding of the on-site situation and response effectiveness using real-time observation data. Let... This represents the set of system state variables at time slice t. This is a set of latent variables that cannot be directly observed, such as the "actual fire spread area" or the "criticality level of trapped personnel" in a fire scenario. Representative in time slice The set of actions performed, such as "dropping fire extinguishing bombs to area C" and "dispatching a drone to coordinates (x,y) for reconnaissance," are known and issued by the upstream module. This represents the set of observational evidence obtained at time slice t, namely the aforementioned fused multi-source information, such as "temperature readings of infrared sensors in region C" and "the number of suspected trapped personnel identified by video analysis".

[0065] The goal of the algorithm is to compute the posterior probability distribution of the current system state, given all historical actions and observational evidence. Based on the Markov assumption of dynamic Bayesian networks (i.e., the current state depends only on the state and action of the previous time step, and the current observation depends only on the current state), this computation process can be decomposed into two alternating steps: Prediction Steps: Based on the state and actions performed in the previous time step, predict the state distribution at the current time step. The mathematical expression for the state distribution at the current time step is: , In this formula, It is a state transition model that describes the inherent evolution of emergency events and the impact of response actions on the event state. For example, it quantifies the state of a fire at a given moment in time. Firefighting actions Under these circumstances, the current fire situation The probability distribution. This model can be pre-built based on domain expert knowledge, historical data, or physical simulation. This is the posterior probability calculated in the previous round, reflecting the iterative update of knowledge.

[0066] Update steps: Utilize newly acquired observational evidence at the current moment This corrects the state distribution obtained in the prediction stage, yielding a more accurate posterior probability. According to Bayes' theorem, its mathematical expression is: , In this formula, It is an observation model that describes the state of a real system. Below, specific observation data is generated. The probability. For example, it quantifies the likelihood that an infrared sensor reading of 790 degrees Celsius is possible when the actual fire temperature is 800 degrees Celsius. The scaling factor, on the other hand, calculates the probability of all possible values. Normalization is performed to determine this.

[0067] By repeatedly executing prediction and update steps, the true state of an emergency event can be observed in real time and probabilistically. This is based on the inferred posterior distribution of the state. This allows for the calculation of a series of quantitative performance indicators. For example, a comprehensive performance score can be defined. The calculation method is as follows: ,in, It is a utility function that maps the system state to one or more performance metrics (such as the controlled disaster area, the number of people rescued, etc.). This represents the mathematical expectation under the posterior probability distribution, while It is a function that combines multiple performance metrics into a single composite score.

[0068] This overall performance score The current performance score and its changing trend over time form the core basis for closed-loop adjustments. This step compares the calculated current performance score with the expected performance curve predicted when the plan was formulated, thereby obtaining the performance deviation. When this deviation exceeds a preset threshold, or the rate of improvement in the performance score is far lower than expected, the system will determine that the current handling strategy is ineffective or inefficient.

[0069] Therefore, the implicit state assessment and attribution technology for emergency response effectiveness based on dynamic Bayesian networks can penetrate complex surface data and grasp the deep core state and dynamic evolution trend of an emergency event in real time and quantitatively. It transforms the vague feeling of "the situation is improving" into the precise "the probability of the fire being under control has increased from 30% to 70% in the past 5 minutes," providing objective and reliable data for decision-making adjustments. Simultaneously, by analyzing the impact of response actions on the probability of state transition, it can also achieve effectiveness attribution, that is, determine which actions are effective and which are ineffective, thereby guiding the next optimization direction.

[0070] Optionally, in the second approach to performance prediction and evaluation, performance prediction and evaluation can also be based on recurrent neural networks such as Long Short-Term Memory (LSTM) networks. Specifically, this approach provides a data-driven method based on deep learning. An LSTM or GRU network can be constructed, with historical time-series data as input, including sequences of our actions. Emergency event on-site situation observation sequence The network output can directly be the performance score for the next time step. The predicted values, or estimates of key state variables, are used. By training on a large amount of historical or simulation data, the network can learn complex, nonlinear dynamic evolution patterns. This approach also achieves the technical effect of quantitatively assessing and predicting response effectiveness. Its advantage lies in the fact that it does not require building explicit physical or causal models, and can automatically learn the dynamic characteristics of the system from the data, making it more adaptable.

[0071] Optionally, in the third scheme of the effectiveness prediction and evaluation scheme, state estimation and effectiveness evaluation can also be performed based on Kalman filtering and its variants. Specifically, the original scheme uses a dynamic Bayesian network for latent variable inference. This alternative scheme uses Kalman filtering, extended Kalman filtering, or unscented Kalman filtering to estimate the core state of the emergency event site in real time, such as the fire spread rate and the vital signs of trapped personnel. These filtering algorithms can effectively fuse noisy sensor observation data through a "prediction-update" loop to provide the optimal estimate of the dynamic state of the system. Based on these estimated state variables, an effectiveness evaluation function can also be constructed to calculate the quantified response effectiveness. This scheme can also achieve the technical effect of inferring the core state of the emergency event site that cannot be directly observed and performing quantified effectiveness evaluation, especially in scenarios where the system model is close to linear and the noise follows a Gaussian distribution, its computational efficiency is even higher.

[0072] Step S207: Based on the assessment results, adaptively adjust the follow-up response strategy for the emergency.

[0073] For example, based on the assessment results, the follow-up response strategy for the emergency is adaptively adjusted, and a structured adaptive adjustment recommendation report is generated. This report includes: the most probable on-site state of the emergency event as currently inferred. Performance deviation The report provides specific numerical values ​​and trends, as well as root cause inferences from attribution analysis of Bayesian networks, such as "the fire spread much faster than the model predicted in the northeast direction, possibly due to unexpected strong winds." This report is immediately fed back to step S202 as insightful input for initiating a new round of simulations and decision-making. Based on this precise and quantitative feedback, step S202 adjusts its assumptions about the emergency response environment, corrects the simulation model, and generates a more realistic and targeted new response plan.

[0074] The method provided in this application embodiment receives a multimodal data stream from the scene of an emergency; performs semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map. Using a preset intelligent inference algorithm, the real-time situation map is processed to generate a response strategy containing a sequence of actions. Based on the response strategy, a target execution device is determined from a preset set of heterogeneous execution devices. The response strategy is parsed and encapsulated to generate executable instructions matching the target execution device, and these instructions are sent to the target execution device so that it can handle the emergency according to the received instructions. Execution feedback information corresponding to the executable instructions is collected. Based on the execution feedback information, the execution effectiveness of the response strategy is evaluated. Based on the evaluation results, subsequent response strategies for the emergency are adaptively adjusted. The beneficial effects of this invention are reflected in the fact that it breaks down the technical barriers of different types of physical devices in a heterogeneous execution device cluster, builds a task-oriented autonomous collaborative capability of heterogeneous devices, greatly improves the utilization efficiency of on-site resources and the smoothness of task execution, realizes intelligent and refined emergency response, solves the above-mentioned problems existing in the prior art, can reduce the system expansion cost of the device resource pool, and can ensure the flexibility of dynamic control of devices in handling emergency scenarios.

[0075] In one example, this application provides an intelligent emergency response system, which includes: a multimodal semantic fusion and situational awareness module, a dynamic strategy generation and intelligent deduction module, a device capability virtualization and service encapsulation module, and a response efficiency evaluation and closed-loop adaptive adjustment module. The multimodal semantic fusion and situational awareness module uses a deep learning model to perform cross-modal correlation and semantic understanding of information from heterogeneous data sources such as video, text, and IoT sensors. It automatically identifies and constructs a structured, high-dimensional situational map of events, replacing simple data listing and manual interpretation. Based on this situational map, the dynamic strategy generation and intelligent deduction module uses reinforcement learning or programming algorithms to generate the optimal combination of handling strategies that highly matches the current situation in real time, rather than calling rigid pre-plan templates. This strategy instruction is transmitted to the device capability virtualization and service encapsulation module, which abstracts all physical terminals with different protocols into a unified and standardized "capability service interface," such as "area broadcast capability" and "path guidance display capability," thereby decoupling the upper-layer strategy from the lower-layer device hardware. Finally, the handling efficiency evaluation and closed-loop adaptive adjustment module analyzes efficiency data such as changes in on-site pedestrian flow heatmaps and social media public opinion feedback to quantitatively evaluate the actual execution effect of the current strategy and feeds the evaluation results back to the dynamic strategy generation module to drive the real-time optimization or reconstruction of the strategy, forming a complete intelligent closed loop.

[0076] Therefore, this application constructs a complete closed-loop intelligent command system, from underlying capability abstraction and mid-level intelligent decision-making to top-level performance evaluation. Compared with existing technologies, it achieves significant progress in emergency command efficiency, response speed, resource utilization, and system intelligence level. Specific technical effects are described below: 1. Significantly improved the efficiency and scientific nature of emergency command decision-making, realizing a fundamental shift from "experience-driven" to "data and model-driven".

[0077] Existing emergency command technologies heavily rely on commanders' personal experience and on-the-spot intuition. Faced with information overload and rapidly changing complex disaster scenes, human cognitive bandwidth is limited, easily leading to decision-making delays, inadequate consideration, or misjudgments. The dynamic strategy generation and intelligent simulation module in this invention can generate multiple response plans in parallel within a very short time based on the real-time situation at the emergency site, and predict the possible outcomes of each plan through a built-in simulation model. This is equivalent to equipping commanders with an omniscient and omnipotent "intelligent staff." Commanders no longer need to conceive plans from scratch, but instead select, optimize, and confirm from a system-recommended set of high-quality plans that have undergone preliminary quantitative evaluation. This process elevates the starting point of decision-making from "nothing" to "excellence," significantly shortening the decision-making cycle. Furthermore, because the decision is based on data analysis and model simulation, rather than purely personal experience, it effectively avoids cognitive biases, making the final decision more scientific, rigorous, and reliable.

[0078] 2. It has enabled a deep and penetrating understanding of the situation at the scene of an emergency, elevating situational awareness from "surface observation" to a higher dimension of "internal insight".

[0079] Traditional command systems typically present commanders directly with video and data collected by front-end sensors, creating a superficial "what you see is what you get" situational map. This approach results in fragmented and unconnected information, requiring commanders to fill in the gaps and make connections to assess deeper situations such as the spread of fire and the true severity of the urgency of trapped personnel. The response effectiveness assessment and closed-loop adaptive adjustment module in this invention innovatively employs a response effectiveness inference algorithm based on dynamic Bayesian networks. This algorithm can fuse multi-source heterogeneous observation data ($E_t$) and, based on a state transition model ($P(X_t |X_{t-1}, A_{t-1})$) and an observation model ($P(E_t | X_t)$), perform probabilistic inference on the latent variables ($X_t$) of the system's core state that cannot be directly observed. This means the system no longer simply displays "sensor A reading is Y," but can provide insightful judgments such as "the probability that the current temperature in the core area of ​​the fire is Z is 95%" and "the risk of the trapped personnel's vital signs deteriorating within the next 5 minutes is 80%." This ability to infer from observed data to the underlying state allows the command system to penetrate the fog of information and reach the essence of the event, providing an unprecedentedly solid foundation for accurate decision-making.

[0080] 3. It has built a task-oriented autonomous collaboration capability for heterogeneous equipment, which has greatly improved the utilization efficiency of on-site resources and the smoothness of task execution.

[0081] In existing technologies, coordinating the command of equipment from different manufacturers and models, such as drones and robots, typically requires multiple operators to perform tedious manual coordination, resulting in high communication costs, low coordination efficiency, and difficulty in achieving complex tactical cooperation. The device capability virtualization and service encapsulation module proposed in this invention abstracts and encapsulates the specific functions of underlying equipment (such as flight, water spraying, and detection) into standardized "capability services," such as "area reconnaissance service" and "targeted fire suppression service." The upper-level decision-making system only needs to issue task-oriented instructions, such as "conduct fire reconnaissance in area A," and the module can automatically analyze the task and intelligently select the most suitable equipment or equipment combination, planning specific execution paths and action sequences to achieve autonomous coordination. This "task-centralized" command model shields the complexity and heterogeneity of the underlying hardware, freeing commanders from tedious micro-operations and allowing them to focus on macro-tactical planning. Simultaneously, resource scheduling is optimized by the system based on the capability model, avoiding delays and conflicts caused by manual coordination, ensuring that resources are mobilized in the shortest time and in the optimal way to respond to tasks, greatly improving overall execution efficiency.

[0082] 4. A high-frequency, quantitative closed-loop feedback and adaptive adjustment mechanism for handling effectiveness has been established, giving the command system the ability to continuously learn and self-evolve.

[0083] Traditional command processes often rely on delayed and qualitative feedback mechanisms, typically involving manual debriefing after an operation concludes. This approach fails to provide real-time, precise performance correction for ongoing tasks. This invention revolutionizes this by employing a performance evaluation and closed-loop adaptive adjustment module. This module continuously calculates a comprehensive performance score ($\eta_t$) and compares it in real-time with the expected performance, resulting in a quantified performance deviation ($\Delta \eta_t$). If the deviation exceeds a threshold, the system not only issues a warning but, more importantly, generates an adaptive adjustment recommendation report containing inferences about the root causes of the deviation. This report is directly fed back to the dynamic strategy generation and intelligent deduction module, triggering a new round of strategy optimization. This "evaluation-feedback-re-decision" closed loop operates at a frequency far exceeding human reaction speed, enabling the command system to keenly detect any subtle deviations in strategy execution and make rapid, precise corrections. Essentially, this is an online learning process; the system continuously calibrates its internal model through interaction with the real environment, making subsequent decisions increasingly realistic. This achieves dynamic self-optimization and continuous evolution of the entire command system during task execution.

[0084] Therefore, the closed-loop adaptive control of "capability abstraction - intelligent decision-making - effectiveness evaluation" for emergency command in this embodiment fundamentally changes the traditional emergency command paradigm, realizing a shift from "human-centered and experience-driven" to "data-centered and intelligence-driven". It can transform massive, heterogeneous emergency resources into unified, computable "capability units", achieving seamless collaboration and optimal allocation of cross-domain and cross-departmental resources, significantly improving emergency response speed and resource utilization efficiency. More importantly, its closed-loop feedback mechanism endows the command system with the ability to "self-evolve", continuously optimizing tactics based on actual combat results and exhibiting strong adaptability to sudden changes in the situation.

[0085] Corresponding to the above method, this application also provides a flowchart of an intelligent handling device for emergencies, as shown in the embodiment. Figure 3 As shown, the device includes: The fusion module 41 is used to receive multimodal data streams from the scene of the emergency; perform semantic fusion and understanding processing on the multimodal data to generate a structured real-time situation map; The deduction module 42 is used to perform strategy deduction processing on the real-time situation map through a preset intelligent deduction algorithm to generate a disposal strategy containing a sequence of disposal actions. The issuing module 43 is used to determine the target execution device from a preset set of heterogeneous execution devices according to the handling strategy; to parse and encapsulate the handling strategy to generate an executable instruction that matches the target execution device; and to issue the executable instruction to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

[0086] The functions of each functional unit of the urban governance event control device provided in the above embodiments of this application can be implemented through the above methods and steps. Therefore, the specific working process and beneficial effects of each unit in the urban governance event control device provided in the embodiments of this application will not be repeated here.

[0087] This application also provides an electronic device, such as... Figure 4 As shown, it includes a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540.

[0088] Memory 530 is used to store computer programs; The processor 510 performs the above steps when executing the program stored in the memory 530.

[0089] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0090] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0091] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0092] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0093] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments for solving the problem can be found in [reference needed]. Figure 1 The steps in the illustrated embodiments are used to implement the electronic device. Therefore, the specific working process and beneficial effects of the electronic device provided in this application will not be repeated here.

[0094] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform any of the intelligent event handling methods described in the above embodiments.

[0095] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the intelligent handling methods for sudden events described in the above embodiments.

[0096] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent method for handling emergencies, characterized in that, include: Receive multimodal data streams from the scene of an emergency; The multimodal data is semantically fused and understood to generate a structured real-time situation map; The real-time situation map is processed by a preset intelligent inference algorithm to generate a response strategy that includes a sequence of response actions. According to the aforementioned processing strategy, the target execution device is determined from a preset set of heterogeneous execution devices; The handling strategy is parsed and encapsulated to generate an executable instruction that matches the target execution device, and the executable instruction is sent to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

2. The method according to claim 1, characterized in that, The multimodal data undergoes semantic fusion and understanding processing to generate a structured real-time situation map, including: The feature vectors of each modality data are extracted using a pre-defined deep neural network. Based on a pre-defined cross-modal attention mechanism, the correlation weights between feature vectors of different modal data are calculated. Based on the association weights, the feature vectors are fused using a preset gating fusion mechanism to generate fused features; Based on the fusion features, a structured real-time situation map is generated; wherein, the nodes in the real-time situation map are used to describe entities, and the edges are used to describe the relationships between the entities.

3. The method according to claim 1, characterized in that, Using a preset intelligent deduction algorithm, the real-time situation map is processed to generate a response strategy containing a sequence of response actions, including: The real-time situation map is encoded using a preset graph attention network to obtain the graph features of nodes and relationships, as well as the prior probabilities of actions. Using the encoded graph features as input, the Monte Carlo tree search algorithm and the prior probability of the action are used to perform policy deduction; and through iterative search, at least one disposal policy whose estimated utility satisfies the preset estimated utility condition is output.

4. The method according to claim 1, characterized in that, According to the aforementioned processing strategy, target execution devices are identified from a pre-defined set of heterogeneous execution devices, including: Determine the task requirements for the proposed handling strategy; Based on the task requirements, and according to the multi-objective optimization model, a capability service combination scheme that satisfies the preset constraints is searched and output from the virtual execution device set corresponding to the preset heterogeneous execution device set; wherein, the physical devices in the heterogeneous execution device set correspond one-to-one with the virtual device entities in the virtual execution device set. Determine the target execution device based on the virtual execution device corresponding to the capability service combination scheme.

5. The method according to claim 1, characterized in that, The processing strategy is parsed and encapsulated to generate executable instructions that match the target execution device, and the executable instructions are sent to the target execution device, including: The disposal strategy is parsed into standard instructions that match the capability service combination scheme; wherein the capability service combination scheme is determined based on the task requirements and multi-objective optimization model of the disposal strategy. The standard instructions are converted into an instruction format executable by the target execution device, and the converted standard instructions are encapsulated and sent to the target execution device.

6. The method according to any one of claims 1-5, characterized in that, After parsing and encapsulating the processing strategy to generate executable instructions matching the target execution device, and issuing the executable instructions to the target execution device, the method further includes: Collect execution feedback information corresponding to the executable instructions; The effectiveness of the handling strategy is evaluated based on the execution feedback information. Based on the assessment results, the follow-up response strategy for the aforementioned emergency will be adaptively adjusted.

7. The method according to claim 6, characterized in that, Based on the execution feedback information, the execution effectiveness of the handling strategy is evaluated, and the method further includes: Construct a dynamic Bayesian network model to describe the state evolution at the scene of the emergency and the impact of response actions; Based on the execution feedback information and the dynamic Bayesian network model, a probabilistic inference is made on the preset core emergency states to obtain the inference results. Based on the inference results, the execution effectiveness of the disposal strategy is evaluated to generate a comprehensive effectiveness score.

8. An intelligent emergency response device, characterized in that, include: The fusion module is used to receive multimodal data streams from the scene of an emergency. The multimodal data is semantically fused and understood to generate a structured real-time situation map; The deduction module is used to perform strategy deduction processing on the real-time situation map through a preset intelligent deduction algorithm to generate a disposal strategy containing a sequence of disposal actions. The distribution module is used to determine the target execution device from a preset set of heterogeneous execution devices according to the processing strategy. The handling strategy is parsed and encapsulated to generate an executable instruction that matches the target execution device, and the executable instruction is sent to the target execution device so that the target execution device can handle the sudden event according to the received executable instruction.

9. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images