Emergency command large model adaptation and reasoning method for multi-disaster coupling scenarios

Abstract

The application discloses an emergency command large model adaptation and reasoning method for a multi-disaster coupling scene, extracts knowledge and instruction in a multi-disaster emergency knowledge graph to fine-tune a general large model, and obtains an emergency command large model with built-in counterfactual reasoning constraints; dynamic spatio-temporal knowledge graphs are generated by fusing multi-modal data on the scene and the weights are updated; the multi-hop coupled disaster chain path in the graph is extracted and input into a physical simulation engine to obtain simulation results, which are input into the large model together with the fusion situation data for collaborative deduction to obtain the multi-disaster coupling evolution trend; the consequences of secondary disasters that may be triggered by the current candidate word are verified using the trend, and if a violation risk is triggered, the constraints are activated to reduce the decoding probability of the candidate word; and a compliant emergency scheduling plan is output through iterative sampling. The application can solve the pain point of lacking safety boundaries in the online self-regression generation of large models, and significantly improve the physical feasibility and compliance safety of emergency scheduling instructions by real-time verification and accurate interception of potential secondary disaster chains.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and emergency command technology, specifically a method for adapting and reasoning a large emergency command model for multi-hazard coupled scenarios. Background Technology

[0002] In recent years, with the increasing complexity of emergencies such as natural disasters and accidents, collaborative scheduling and scientific decision-making have become core requirements of modern emergency command systems. Large-scale models, due to their powerful natural language understanding and text generation capabilities, are being gradually introduced into the field of emergency command to improve the intelligence level of situation assessment and emergency plan generation.

[0003] However, emergency management scenarios have industrial-grade requirements for high reliability and strong professional expertise. When general-purpose models are directly applied to this vertical field, many technical bottlenecks are exposed, as follows: (1) In extreme cases, multiple disasters such as rainstorms, mudslides, road interruptions or hazardous chemical leaks often exhibit chain-like coupled evolution. Existing emergency systems mostly rely on single-disaster mechanism simulation engines, while large models lack a deep understanding of the underlying physical laws, making it difficult to accurately capture and jointly deduce the dynamic spatiotemporal evolution trend of disaster chains under complex multimodal situations.

[0004] (2) Due to the extremely limited historical case samples of extreme disasters, the model faces a severe challenge of small sample or even zero sample adaptation. In the absence of strong vertical knowledge constraints, large models are prone to illusions when decoding and generating emergency decisions and dispatch instructions, outputting erroneous instructions that do not conform to emergency management regulations and operating procedures or are unrealistic. This is unacceptable in emergency command practice where every second counts and the fault tolerance rate is zero.

[0005] The aforementioned deficiencies severely limit the inference accuracy of the emergency command model in multi-hazard coupled scenarios. Summary of the Invention

[0006] To address the technical problems mentioned in the background, the purpose of this application is to provide a system for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios.

[0007] According to a first aspect of this application, a method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios is provided, the method comprising: A multi-hazard emergency knowledge graph is constructed based on multi-source data. Knowledge texts and compliance instructions are extracted from it as fine-tuning data to fine-tune the general model, resulting in an emergency command model with built-in counterfactual reasoning constraints. Multimodal situational data on site is fused to obtain fused situational data, which is then used to update the weights of nodes and edges in the multi-hazard emergency knowledge graph to generate a dynamic spatiotemporal knowledge graph. Multi-hop coupled disaster chain paths are extracted from the dynamic spatiotemporal knowledge graph and input into the physical simulation engine to obtain physical simulation results. The fused situational data, the physical simulation results, and the multi-hop coupled disaster chain paths are input into the emergency command big model for collaborative simulation to obtain the multi-hazard coupling evolution trend. By leveraging the multi-hazard coupling evolution trend, the potential secondary disaster consequences of the current candidate word are inferred. If the secondary disaster consequence triggers a violation risk, counterfactual reasoning constraints are activated to reduce the decoding probability of the candidate word. A safe text sequence is generated by iteratively sampling the remaining candidate words, and an emergency dispatch plan that meets compliance requirements is output.

[0008] According to a second aspect of this application, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface are connected via the communication bus; the memory is used to store a set of executable instructions that cause the processor to execute to implement the method as described in any of the preceding claims.

[0009] According to a third aspect of this application, a computer storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method as described in any of the preceding claims.

[0010] According to a fourth aspect of this application, a computer program product is provided, including computer instructions that, when executed by a processor, implement the method as described in any of the preceding claims.

[0011] The solution of this invention can significantly improve the physical feasibility and compliance security of emergency dispatch instructions, effectively intercept and block potential secondary disaster chains, and is especially suitable for high-reliability emergency command and decision-making in multi-hazard coupling scenarios. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1A flowchart illustrating an emergency command large-scale model adaptation and inference method for multi-hazard coupled scenarios provided in this application embodiment; Figure 2 A flowchart illustrating the specific implementation process of step S3 provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0015] This application discloses a method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios. This embodiment uses a scenario of a coastal chemical industrial park experiencing a multi-hazard coupled extreme event, including a super typhoon, torrential rain and flooding, and the explosion and combustion of hazardous chemical storage tanks, to illustrate the specific implementation process of this invention in detail.

[0016] It should be understood that the solutions of this invention are not limited to the multi-hazard coupling scenarios in coastal chemical industrial parks mentioned above, but can also be applied to other emergency command fields with cascading effects and high safety risks. For example, in urban rail transit systems, a sudden fire in a subway station may trigger multiple coupled disasters such as power grid tripping, smoke exhaust failure, stampedes, and paralysis of ground traffic; in the scenario of a severe nuclear power plant accident, a complete power outage may lead to a chain reaction of cooling loss, reactor core overheating, hydrogen accumulation, and even the spread of radioactive materials; in wildfire control in forest-town boundary areas (WUI), the spread of crown fire may generate flying sparks that cross the firebreak, ignite gas cylinders in residential areas, and trigger a series of explosions. In all of the above scenarios, this invention can significantly improve the physical feasibility and compliance security of emergency command decisions in complex environments by integrating multi-source heterogeneous data, dynamically constructing spatiotemporal knowledge graphs, and using joint physical simulation engines for collaborative inference, and by using counterfactual reasoning constraints to intercept high-risk instructions online.

[0017] Please see Figure 1 The method includes the following steps: S1. Construct a multi-hazard emergency knowledge graph based on multi-source data, extract the knowledge text and compliance instructions from it as fine-tuning data to fine-tune the general model, and obtain an emergency command model with built-in counterfactual reasoning constraints. In this embodiment, during the offline phase, the general-purpose large model needs to possess professional knowledge and safety boundaries for multi-hazard scenarios. To this end, multi-source data is collected, including historical meteorological disaster data of the park and its surrounding area, various material reserves, spatial distribution of hazardous sources, safety production regulations, and emergency response plans.

[0018] As an example, a multi-hazard emergency knowledge graph is constructed based on multi-source data, including: S11, extract disaster entities, material entities, and causal relationships between entities from the multi-source data, and transform the extracted disaster entities, material entities, and causal relationships into triplet structures to construct a multi-hazard emergency knowledge graph.

[0019] For example, disaster and material entities such as "storage tanks," "heavy rainfall," and "benzene overflow" are extracted from the aforementioned multi-source data, and causal relationships between them are identified, such as "heavy rainfall" causing "waterlogging," and "waterlogging" leading to "power grid failure." These entities and causal relationships are transformed into structured triples, such as <heavy rainfall, causing, waterlogging>, thereby constructing a multi-hazard emergency knowledge graph.

[0020] Furthermore, key data is extracted from the constructed multi-hazard emergency knowledge graph for fine-tuning of the general model. This data includes: knowledge texts, including but not limited to unstructured domain knowledge such as disaster evolution patterns, disaster-causing mechanisms, historical case review reports, and material allocation logic; and compliance instructions, including key clauses in safety production regulations, emergency rescue operating procedures, inter-departmental coordination plans, and various mandatory standards. This data is organized into a "situation-decision" question-and-answer format as monitoring and fine-tuning samples to adjust the general model.

[0021] To incorporate counterfactual reasoning constraints into the model, negative sample contrastive learning is introduced into the fine-tuning data. Specifically, for a specific disaster evolution node (e.g., "tank pressure exceeds the safety threshold"), several candidate actions that violate physical or regulatory rules are constructed (e.g., "immediately open all safety valves to release pressure"), and the potential secondary disaster consequences of these actions are labeled (e.g., "instantaneous release of high-pressure medium leading to a chain explosion"). Through this positive and negative example contrastive training of "if X is executed, it will lead to disaster Y," the model gradually develops an inherent ability to inhibit illegal commands during parameter updates.

[0022] After the aforementioned fine-tuning, the general-purpose model is transformed into an emergency command model with built-in counterfactual reasoning constraints. This model not only generates emergency texts that conform to domain specifications, but more importantly, it can assess the risk tendency of candidate words during the autoregressive decoding process.

[0023] S2, fuse the multimodal situational data on site to obtain fused situational data, and use it to update the weights of nodes and edges in the multi-hazard emergency knowledge graph to generate a dynamic spatiotemporal knowledge graph; Once the real-time support phase begins, the situation on-site changes rapidly, requiring real-time data to be injected into the model to reflect the dynamic evolution process.

[0024] As an example, multimodal situational data from the field is fused to obtain fused situational data, including: S21. Collect on-site multimodal situational data including monitoring video, on-site audio and sensor values. Perform spatiotemporal alignment of the on-site multimodal situational data according to a unified time axis and spatial geographic coordinates. Vectorize and stitch together the aligned multimodal features to generate the fused situational data.

[0025] When the typhoon makes landfall, the park's monitoring videos, on-site audio, and various liquid level and pressure sensors report massive amounts of multimodal situational data in real time.

[0026] For example, by unifying the timeline with millisecond-level precision and using the park's geographic information system as the spatial reference, the pressure sensor value of "Tank A" at "13:05:22", the key frame of the monitoring video at that time, and the surrounding audio features are vectorized and stitched together to form a high-density fused situational data record.

[0027] After obtaining the fused situational awareness data, the offline-constructed multi-hazard emergency knowledge graph is dynamically updated using this data. Specifically, for real-time status information reflected in the fused situational awareness data (such as "sudden increase in tank pressure" or "liquid level exceeds warning line"), the corresponding disaster entity nodes in the knowledge graph (such as "tank overpressure") are identified, and the weight value of the node is dynamically adjusted according to the degree of deviation between the current value and the preset threshold; that is, the higher the degree of anomaly, the greater the node weight. At the same time, for causal edges between nodes (such as "cooling failure → tank overpressure"), the weight coefficient of the edge is updated according to the temporal correlation and intensity association of the two events in the real-time data stream, so that it more accurately reflects the coupling strength in the current dynamic environment.

[0028] Through the aforementioned real-time updates, the multi-hazard emergency knowledge graph, originally built based on historical static data, has gradually evolved into a dynamic spatiotemporal knowledge graph that can reflect changes in the on-site situation and dynamically adjust the coupling intensity between different types of hazards in real time.

[0029] S3. Extract multi-hop coupled disaster chain paths from the dynamic spatiotemporal knowledge graph and input them into the physical simulation engine to obtain physical simulation results. Input the fused situation data, the physical simulation results, and the multi-hop coupled disaster chain paths into the emergency command big model for collaborative simulation to obtain the multi-hazard coupling evolution trend. After obtaining the dynamic spatiotemporal knowledge graph, the most dangerous and potentially evolving disaster transmission links are further identified. Since each node (disaster entity) and each directed edge (causal relationship) in the graph is assigned a real-time weight, the node weight reflects the actual severity of the disaster (e.g., the weight of the node "sudden increase in tank pressure" increases as the pressure value rises), and the edge weight reflects the coupling strength between two disasters in the current environment (e.g., the weight of the edge "power outage in cooling system → sudden increase in tank pressure" is affected by real-time temperature and equipment status). Therefore, graph search algorithms (e.g., weight-based depth-first search or Dijkstra's longest path algorithm) can be used to extract multi-hop coupled disaster chain paths.

[0030] Specifically, starting from the initial disaster node (e.g., "typhoon damages the external power grid"), the path is recursively expanded along the directed edge direction, taking into account both node weights and edge weights. At each step, priority is given to transmission relationships where the current node has a high weight, the edge weight between it and its successor node is high, and the successor node itself also has a high weight, thus forming a directed path with the greatest risk accumulation. The path terminates when it reaches the preset maximum number of hops or reaches the final disaster node (e.g., "deflagration damages adjacent storage tanks"). Each hop in this path represents a cause-and-effect disaster transmission relationship, and the path as a whole reflects the most likely or most dangerous evolutionary chain from the initial event to the final secondary disaster.

[0031] In this embodiment, based on a real-time updated dynamic spatiotemporal knowledge graph, the typical high-risk path extracted is: typhoon damages external power grid → cooling system power outage → tank pressure surges → explosion damages adjacent tanks.

[0032] As an example, see Figure 2 As shown, the multi-hop coupled disaster chain path is input into the physics simulation engine to obtain physics simulation results, including: S31, parse the multi-hop coupled disaster chain path to extract the corresponding disaster mechanism control parameters, input the disaster mechanism control parameters into the corresponding disaster simulation module in the physical simulation engine for mathematical model calculation to obtain the numerical matrix of disaster spread trend, as the physical simulation result.

[0033] Specifically, each causal relationship in the aforementioned disaster chain path is analyzed to extract the key physical parameters driving the disaster's evolution. For example, for the stage of "cooling system power outage → sudden increase in tank pressure," the pressure rise rate parameter is extracted based on the tank's volume, medium properties (such as the saturated vapor pressure curve), initial temperature, and the rate of heat accumulation after the power outage. For the stage of "sudden increase in tank pressure → deflagration," parameters such as the tank's ultimate pressure, safety valve relief capacity, and ignition energy threshold are extracted. These parameters are the disaster mechanism control parameters.

[0034] The aforementioned disaster mechanism control parameters are input into the corresponding disaster simulation modules within the physical simulation engine. It should be understood that the physical simulation engine pre-integrates mathematical model modules for different disaster types (such as fire, explosion, flood, toxic gas diffusion, etc.). For example, the fire module is based on a flame spread model and a thermal radiation model, while the explosion module is based on the TNT equivalent method or a computational fluid dynamics model. The extracted disaster mechanism control parameters are then fed into the corresponding modules for numerical calculation, outputting a numerical matrix describing the spatial and temporal spread trend of the disaster. This matrix can represent the fire temperature distribution, thermal radiation flux field, shock wave overpressure distribution, or toxic substance concentration field, etc., as the physical simulation result.

[0035] As an example, see Figure 2 As shown, the fused situational data, the physical simulation results, and the multi-hop coupled disaster chain paths are input into the emergency command big data model for collaborative simulation to obtain the multi-hazard coupled evolution trend, including: S32, extract the bidirectional feedback correlation between different disaster nodes in the multi-hop coupled disaster chain path, and calculate the critical value of energy mutation caused by the coupling of adjacent disasters based on the bidirectional feedback correlation and the physical simulation results; Graph structure analysis was performed on the extracted multi-hop coupled disaster chain paths to identify bidirectional feedback relationships between nodes of different disaster types. Bidirectional feedback relationships refer to mutually reinforcing or inhibiting effects between two disaster nodes, rather than unidirectional causal transmission.

[0036] In this embodiment, taking the fire node (denoted as node A) and the hazardous chemical explosion node (denoted as node B) as examples, their bidirectional feedback relationship can be described as follows: the heat radiation from the fire causes the internal pressure of the storage tank to rise, which in turn triggers an explosion. The shock wave and projectiles generated by the explosion further fuel the spread of the fire. The type of this feedback relationship (positive reinforcement or negative inhibition) and the initial feedback intensity coefficient are read from the edge attributes of the dynamic spatiotemporal knowledge graph.

[0037] To quantify the risk of nonlinear mutation caused by this bidirectional feedback, the critical value for energy mutation is further calculated. It should be understood that this critical value for energy mutation characterizes the physical threshold from stable coupling to runaway cascade. Taking the thermo-mechanical coupling between adjacent storage tanks as an example, based on physical simulation results (such as the thermal radiation flux matrix) and the physical parameters of the storage tanks (such as the yield strength of the wall material and the saturated vapor pressure curve of the medium), the critical value for energy mutation is calculated using the following storage tank thermal failure model: Assume the yield strength of the tank wall material With temperature An increase followed by a linear decrease can be expressed as: in, The yield strength at room temperature For ambient temperature, This refers to the material's melting temperature (or the temperature at which its strength drops to zero). The pressure inside the tank. Circumferential stress generated on the wall surface for: In the formula, Let the radius of the storage tank be 1. The wall thickness is [value]. The failure criterion is [value]. From this, the critical temperature can be solved. .

[0038] Wall temperature The change over time is determined by both thermal radiation absorption and heat dissipation, and its differential equation is: in, The wall surface thermal radiation absorptivity, This is the real-time thermal radiation flux output from the physical simulation. , , These are the wall material density, specific heat capacity, and thickness, respectively. The overall heat dissipation coefficient.

[0039] Energy mutation critical value Defined from the initial time To reach the critical temperature The moment Between these points, the total heat absorbed by the tank wall is: By numerically solving the above differential equation, we can obtain... When the real-time accumulated heat absorption exceeds this critical value, it is determined that the storage tank will experience a coupled deflagration mutation.

[0040] S33, Based on the energy mutation critical value and the fusion situation data, a dynamic spatiotemporal coupling graph is constructed in the hidden layer representation space of the emergency command big model, and a corresponding graph topology mask matrix is ​​generated based on the graph topology structure of the dynamic spatiotemporal coupling graph; wherein, the nodes in the dynamic spatiotemporal coupling graph represent the current evolution intensity of each disaster type, and the edges represent the dynamic feedback coefficients between disaster types. After obtaining the critical value of energy mutation, the coupling relationships of the physical world are further injected into the internal representation of the large model, enabling the model to dynamically perceive the real-time interactions between disaster types during the inference process. To this end, this invention constructs a dynamic spatiotemporal coupling graph within the hidden layer representation space of the emergency command large model. The construction process includes: First, determine the set of nodes and their initial features. Assume the large model is currently processing... Each of the following is a separate disaster entity (e.g., multiple storage tanks, fire points in different locations, leak sources, etc.). For each disaster entity... Encode its current multidimensional state information into a Hidden vectors of dimension The multidimensional state information includes, but is not limited to: the entity's real-time physical field values ​​(such as thermal radiation flux and pressure values), spatial coordinates, timestamps, and semantic features extracted from a dynamic spatiotemporal knowledge graph. These hidden vectors ( This refers to the set of nodes that constitute the dynamic spatiotemporal coupled graph. The evolution strength of each node is determined by the norm of its hidden vector (e.g., ...). It can be characterized by a value of a specific dimension, or by a larger value, indicating that the disaster type is currently more active and contributes more to the overall situation.

[0041] It should be noted that these hidden vectors are intermediate representations naturally generated by the large model through its embedding layer and bottom encoder when processing the input sequence (a serialized representation of the fused situational data, physical simulation results, and disaster chain paths). Therefore, the dynamic spatiotemporal coupled graph is actually a physical semantic annotation and topological reorganization of the existing hidden states of the large model, rather than a reconstruction of an independent graph structure.

[0042] Furthermore, the directed edges between nodes and their dynamic feedback coefficients are calculated. For any two nodes... and To determine whether there exists a path from a multi-hop coupled disaster chain and a dynamic spatiotemporal knowledge graph. arrive The direct causal or feedback relationship. If it exists, the dynamic feedback coefficient is calculated according to the following formula based on the energy mutation threshold and the current fusion situation data. : in: It is the nominal coupling strength coefficient obtained from the knowledge graph, reflecting the inherent degree of correlation between the two disasters (the value ranges from 0 to 1).

[0043] It is a node The corresponding real-time dominant physical field intensity (e.g., if) If it is a fire node, then The thermal radiation flux can be used; if it is a pressure vessel node, the pressure value is used. This value is read in real time from the fused situational data.

[0044] It is a node The energy mutation threshold is a fixed physical property for a given disaster entity, but it varies between different entities.

[0045] ratio This reflects how close the current state is to the mutation threshold. When the fire intensifies or the pressure increases... As the ratio increases, the dynamic feedback coefficient also increases, indicating a sharp increase in the coupling effect when the system is on the verge of losing control.

[0046] It is a disaster entity and The real-time spatial distance between them is provided by GIS data fusion with on-site changes (such as the relative change in the position of the storage tank caused by the scattering of explosion debris), and this value may also change over time; This is a preset reference distance. (Denominator term) The weights decrease as distance increases, and the feedback coefficients are adjusted accordingly when the distance changes due to the development of the disaster.

[0047] If there is no direct coupling relationship, then For bidirectional feedback correlations (e.g., mutual reinforcement between fire and explosion), simultaneous calculation and The two are usually not equal and are updated independently and dynamically, because the physical driving mechanism from fire to explosion and the mechanism from explosion to fire may have different intensity coefficients and thresholds.

[0048] The calculated graph topology is mapped to a mask matrix and a self-attention mechanism is injected. All Organized into a weight matrix ( This allows us to obtain a directed graph within the hidden layer space. Based on the graph's topology (i.e., the connections between non-zero edges), a binary graph topology mask matrix is ​​generated. Its dimensions are also The element is defined as: It should be noted that this mask matrix The hidden vector is not directly modified. It is not itself, but rather serves as an attention routing table to correct the information propagation path in the self-attention mechanism of large models. Specifically, in the subsequent step S34, the mask matrix... It will perform element-wise operations with the original attention weight matrix, so that only Information is only allowed to be passed between node pairs. With this setting, the attention network, which was originally fully connected in the hidden layer, is forced to propagate features only along the directed edges of the dynamic spatiotemporal coupling graph, thereby embedding the coupling constraints of the physical world into the reasoning process of the large model.

[0049] S34. The self-attention mechanism of the emergency command big model is corrected by the graph topology mask matrix. The state features are propagated bidirectionally across disaster types along the directed edge direction of the dynamic spatiotemporal coupling graph to capture the nonlinear transition of the intensity of the secondary disaster chain caused by feedback deadlock and energy mutation, and output the multi-disaster coupling evolution trend.

[0050] In obtaining the graph topology mask matrix Subsequently, it was applied to the multi-head self-attention mechanism of the emergency command large-scale model to achieve information propagation under physical constraints. The standard Transformer's self-attention mechanism in the first... Layer-by-layer calculation of attention weight matrix In this case, a fully connected form is usually used: in, , These are the query and key matrices, respectively. is the scaling factor. This fully connected structure allows each node to focus on all other nodes, which is reasonable in general language modeling, but introduces spurious associations that are physically impossible in multi-hazard coupling derivations. Therefore, a mask matrix is ​​used. The original attention weights are modified as follows: in, The meaning is: for The position where 0 is added This makes the corresponding weights zero after softmax; for At positions where 1 is found, add 0, keeping the original value. This results in the corrected attention matrix. Only those that meet the requirements Information is only allowed to be passed between node pairs. In other words, the information flow path of the model is strictly restricted to the set of directed edges in the dynamic spatiotemporal coupled graph. superior.

[0051] After obtaining physically constrained attention weights, a cross-hazard bidirectional iterative propagation mechanism is further initiated. Traditional self-attention performs only one message pass within a single layer, and the direction is bidirectional (although bidirectional propagation is still allowed after masking), but it cannot simulate the dynamic feedback enhancement process. Therefore, graph-guided attention updates are repeatedly executed in multiple time steps, forming an iterative propagation process.

[0052] Specifically, let the first Node at the next iteration The hidden state is (initial (Corresponding to the state when entering this layer). The update rule for each iteration is: in, To correct the attention matrix from nodes To the node The weight, This is the value transformation matrix. This expression represents the transformation matrix for each node. Aggregate all predecessor nodes that can reach it (Right now The information in the dynamic spatiotemporal coupling graph is such that it contains both forward causal edges (e.g., fire → explosion) and backward feedback edges (e.g., explosion → fire). Therefore, the propagation naturally achieves bidirectional feedback.

[0053] To simulate the nonlinear energy transitions caused by positive feedback between disasters, each node is checked after each iteration. Corresponding physical field strength With energy mutation critical value The relationship. If If the threshold is exceeded, a nonlinear transition operator is applied to the hidden state of the node: in, (For example, take 2.0 or calculate dynamically based on the energy overshoot ratio). This operator amplifies the representation intensity of the node, and then spreads this abrupt change effect along the directed edge to its adjacent nodes through attention propagation in the next iteration. After multiple iterations (e.g., 5-10 iterations), it can capture the nonlinear intensity transition phenomenon caused by feedback deadlock (such as repeated mutual reinforcement of fire and explosion), that is, the disaster chain suddenly changes from slow development to exponential deterioration.

[0054] After the iterative propagation is complete, the large model exits the final hidden state. The decoder outputs predictions for multiple future time windows. Specifically, a linear output layer with an activation function is used to map the hidden vector of each node to the corresponding disaster type in subsequent time windows. Evolution intensity predictions within 15 minutes, 30 minutes, and 1 hour: in, Represents a node In the The predicted intensity within a time window (such as the thermal radiation flux of a fire, the probability of explosion, and the concentration of toxic gases). The activation function is (e.g., Sigmoid or ReLU). The prediction results of all nodes are combined into a multi-dimensional time series, i.e., the multi-hazard coupled evolution trend.

[0055] This evolutionary trend is output in a structured form, for example A numerical matrix, or a JSON object with time and space labels.

[0056] Through the above processing, the emergency command big model of the present invention can overcome the defects of traditional big models such as unidirectional autoregression and lack of physical constraints, and thus generate high-fidelity and physically feasible evolution predictions in multi-hazard coupled scenarios.

[0057] S4 utilizes the multi-hazard coupling evolution trend to infer the secondary disaster consequences that the current candidate word may cause. If the secondary disaster consequences trigger the risk of violation, counterfactual reasoning constraints are activated to reduce the decoding probability of the candidate word. A safe text sequence is generated by iterative sampling of the remaining candidate words, and an emergency dispatch plan that meets compliance requirements is output.

[0058] This step aims to address the technical problem in open autoregressive decoding where large models generate emergency response instructions that lead to the output of non-compliant or unsafe dispatch plans due to a lack of secondary disaster inference regarding the consequences of their own speech. During the autoregressive decoding process of generating emergency dispatch plans, the large model activates a safety filtering mechanism whenever a candidate word for the current output is predicted (e.g., when the model generates the sequence "dispatch the fire brigade to pump water from area A," the current candidate word is "area A").

[0059] As an example, the potential secondary disaster consequences of current candidate words can be inferred by utilizing the multi-hazard coupling evolution trend, including: S41, the current candidate word is converted into a text feature vector and concatenated with the multi-hazard coupling evolution trend to obtain the consequence prediction feature; Let the candidate words output by the model in the current decoding step be... The candidate word is then transformed into a dense vector through the embedding layer of the large model. Simultaneously, the time window and disaster node most relevant to the current decision-making process in the multi-hazard coupled evolution trend are selected and flattened into a state vector. Then, the two are concatenated as vectors to obtain the consequence prediction features. .

[0060] S42, the consequence prediction features are input into a preset secondary disaster assessment network for forward propagation calculation, and the corresponding potential disaster type and damage intensity index are output as the secondary disaster consequences. The secondary disaster assessment network is a feature prediction branch of the emergency command big data model, which is constructed by training on historical secondary evolution samples of various disaster types in the multi-source data.

[0061] In this embodiment, the secondary disaster assessment network can be a multilayer perceptron (MLP) structure, containing one or more hidden layers and an output layer. Its forward propagation computation process can be represented as follows: in, It is a probability distribution vector, with each dimension corresponding to a potential disaster type (such as "personal poisoning", "secondary explosion", "building collapse" etc.), and the type with the highest probability is the predicted disaster type; These are continuous numerical values ​​representing the intensity of damage (e.g., values ​​ranging from 0 to 1, with values ​​above 0.8 indicating extremely high intensity). The network has been trained offline using historical samples. The input to the training samples are "historical decision words" and "the evolutionary trend at that time," and the output is "the actual type and intensity of secondary disasters that occurred."

[0062] As an example, activating counterfactual reasoning constraints to reduce the decoding probability of the candidate word includes: S43, input the candidate word into the preset counterfactual rule base for matching and judgment. If the judgment result is that the candidate word triggers a violation risk, then reduce the log probability value of the candidate word in the current prediction probability distribution to reduce the decoding probability of the candidate word.

[0063] If the aforementioned secondary disaster consequences trigger a violation risk, then the risk (disaster type and intensity index) is sent to the safety assessment logic. Specifically, a predefined counterfactual rule base is maintained, containing rules such as "if the destructive intensity > 0.7 and the disaster type includes 'poisoning' or 'explosion,' then a violation is triggered." When a violation risk is matched, the counterfactual reasoning constraint is activated, penalizing the logit value of the current candidate word. in, The log-odds ratio of the original output of the large model. It is a positive penalty coefficient (which can be preset, for example, 5.0). This is an indicator function, equal to 1 if a violation occurs, and 0 otherwise. After the penalty, the decoding probability of the candidate word after softmax renormalization will be significantly reduced, and it may even be excluded from the sampling candidate set.

[0064] Furthermore, the remaining unpunished candidate words are iteratively sampled according to the updated probability distribution to generate subsequent text tokens one by one until a complete emergency dispatch plan text sequence that meets compliance requirements is formed. Since each candidate word has undergone the above verification and interception during generation, the final output plan is physically feasible, legally compliant, and logically secure.

[0065] For example, in the context of this embodiment, if the candidate word "Area A" triggers a violation judgment of "No Entry to Dangerous Areas," its logarithmic probability is reduced, and the model samples other safe candidate words (such as "Area B" or "Unmanned Decontamination Robot"). The final output sequence generated is "Prioritize dispatching unmanned decontamination robots to Area A for containment, and gather rescue personnel in the safe upwind direction of Area B." This achieves highly safe emergency decision generation.

[0066] See Figure 3 As shown in the illustration, this application also discloses an electronic device 100, including: a processor 101, a memory 102, a communication interface 103, and a communication bus 104. The processor 101, the memory 102, and the communication interface 103 are connected via the communication bus 104 to complete communication between them. The memory 102 is used to store a set of executable instructions, including computer programs, operating systems, drivers, etc. When the executable instructions are executed by the processor 101, the processor 101 is configured to execute the emergency command large model adaptation and inference method for multi-hazard coupled scenarios described in any of the preceding embodiments.

[0067] Specifically, the processor, by executing instructions stored in the memory, can achieve offline multi-hazard emergency knowledge graph construction and large model fine-tuning, real-time on-site multimodal data fusion and dynamic spatiotemporal knowledge graph updating, multi-hop coupled disaster chain path extraction and physical simulation collaborative inference, and secondary disaster consequence verification and counterfactual constraint interception during online autoregressive decoding, ultimately outputting an emergency dispatch plan that meets physical feasibility and compliance security requirements. The communication interface 103 is used to exchange data with external devices (such as on-site sensors, surveillance cameras, drones, physical simulation servers, command terminals, etc.) to receive multimodal situational data and send generated dispatch instructions. The electronic device can be a single server, a cluster of multiple servers, an edge computing node, or a cloud server; its specific form does not constitute a limitation of the present invention.

[0068] This application also discloses a computer storage medium storing a computer program. The computer storage medium can be a non-volatile memory, such as a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), flash memory, a solid-state drive (SSD), a hard disk drive (HDD), an optical disc (CD-ROM, DVD), etc. When the computer program is executed by a processor, it implements the method described in any of the foregoing embodiments. The computer storage medium can be sold separately or provided to users as part of a server or terminal device.

[0069] This application also discloses a computer program product, including computer instructions. These computer instructions can be stored in a computer storage medium as described above, or transmitted and distributed via a network (such as the Internet or a local area network). When the computer instructions are executed by a processor, they implement the method described in any of the foregoing embodiments.

[0070] The computer program product can exist in the form of executable files (such as .exe), dynamic link libraries (.dll), Java bytecode, Python scripts, or container images. Users can obtain the computer program product through online download, CD installation, or network push, and deploy it on the server of the emergency command center, the on-site command terminal, or the cloud computing platform. After deployment and operation, the computer program product executes the emergency command large model adaptation and inference method of the present invention, assisting commanders in quickly generating safe, compliant, and physically feasible dispatching plans in multi-hazard coupled scenarios.

[0071] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios, characterized in that, include: A multi-hazard emergency knowledge graph is constructed based on multi-source data. Knowledge texts and compliance instructions are extracted from it as fine-tuning data to fine-tune the general model, resulting in an emergency command model with built-in counterfactual reasoning constraints. Multimodal situational data on site is fused to obtain fused situational data, which is then used to update the weights of nodes and edges in the multi-hazard emergency knowledge graph to generate a dynamic spatiotemporal knowledge graph. Multi-hop coupled disaster chain paths are extracted from the dynamic spatiotemporal knowledge graph and input into the physical simulation engine to obtain physical simulation results. The fused situational data, the physical simulation results, and the multi-hop coupled disaster chain paths are input into the emergency command big model for collaborative simulation to obtain the multi-hazard coupling evolution trend. By leveraging the multi-hazard coupling evolution trend, the potential secondary disaster consequences of the current candidate word are inferred. If the secondary disaster consequence triggers a violation risk, counterfactual reasoning constraints are activated to reduce the decoding probability of the candidate word. A safe text sequence is generated by iteratively sampling the remaining candidate words, and an emergency dispatch plan that meets compliance requirements is output.

2. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 1, characterized in that, A multi-hazard emergency knowledge graph is constructed based on multi-source data, including: Disaster entities, material entities, and causal relationships between entities are extracted from the multi-source data. The extracted disaster entities, material entities, and causal relationships are transformed into triplet structures to construct a multi-hazard emergency knowledge graph.

3. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 1, characterized in that, Multimodal situational awareness data from the field is fused to obtain fused situational awareness data, including: Collect on-site multimodal situational data including surveillance video, on-site audio, and sensor values. Perform spatiotemporal alignment of the on-site multimodal situational data according to a unified time axis and spatial geographic coordinates. Vectorize and stitch together the aligned multimodal features to generate the fused situational data.

4. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 1, characterized in that, Multi-hop coupled disaster chain paths are extracted from a dynamic spatiotemporal knowledge graph and input into a physics simulation engine to obtain physics simulation results, including: The multi-hop coupled disaster chain path is analyzed to extract the corresponding disaster mechanism control parameters. The disaster mechanism control parameters are then input into the corresponding disaster simulation module in the physical simulation engine to perform mathematical model calculations to obtain a numerical matrix of the disaster spread trend, which serves as the physical simulation result.

5. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 3, characterized in that, The fused situational data, the physical simulation results, and the multi-hop coupled disaster chain paths are input into the emergency command big data model for collaborative simulation to obtain the multi-hazard coupled evolution trend, including: Extract the bidirectional feedback correlation between nodes of different disaster types in the multi-hop coupled disaster chain path, and calculate the critical value of energy mutation caused by the coupling of adjacent disaster types based on the bidirectional feedback correlation and the physical simulation results. Based on the energy mutation critical value and the fusion situation data, a dynamic spatiotemporal coupling graph is constructed in the hidden layer representation space of the emergency command big model, and a corresponding graph topology mask matrix is ​​generated based on the graph topology structure of the dynamic spatiotemporal coupling graph; wherein, the nodes in the dynamic spatiotemporal coupling graph represent the current evolution intensity of each disaster type, and the edges represent the dynamic feedback coefficients between disaster types. The self-attention mechanism of the emergency command big model is corrected by using the graph topology mask matrix. The state features are propagated bidirectionally across disaster types along the directed edge direction of the dynamic spatiotemporal coupling graph to capture the nonlinear transition of the intensity of the secondary disaster chain caused by feedback deadlock and energy mutation, and output the multi-disaster coupling evolution trend.

6. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 1, characterized in that, The potential secondary disaster consequences of current candidate words are inferred by utilizing the multi-hazard coupling evolution trend, including: The current candidate words are converted into text feature vectors and concatenated with the multi-hazard coupling evolution trend to obtain consequence prediction features; The predicted consequences are input into a preset secondary disaster assessment network for forward propagation calculation, and the corresponding potential disaster type and damage intensity index are output as the secondary disaster consequences. The secondary disaster assessment network is a feature prediction branch of the emergency command big data model, which is constructed by training on historical secondary evolution samples of various disaster types in the multi-source data.

7. The method for adapting and reasoning a large-scale emergency command model for multi-hazard coupled scenarios as described in claim 6, characterized in that, Activating counterfactual reasoning constraints to reduce the decoding probability of the candidate word includes: The candidate words are input into a preset counterfactual rule base for matching and judgment. If the judgment result is that the candidate word triggers a violation risk, the log probability value of the candidate word in the current prediction probability distribution is reduced to decrease the decoding probability of the candidate word.

8. An electronic device, comprising: A processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface are connected via the communication bus; characterized in that the memory is used to store a set of executable instructions, which cause the processor to execute to implement the method of any one of claims 1-7.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-7.

10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the method of any one of claims 1-7.

Images