A method and system for generating an emergency plan for a sudden environmental event

Abstract

The application relates to the technical field of preplan generation, in particular to a method and system for generating an emergency preplan for an environmental emergency, which comprises the following steps: obtaining an event description text and real-time monitoring data stream of an environmental emergency, inputting the event description text into a pre-trained language model for semantic coding to obtain an event semantic feature vector; performing similarity retrieval in a preset case library based on the event semantic feature vector to obtain a historical emergency preplan set matched with the environmental emergency, wherein the preset case library comprises historical emergency preplans and corresponding feature vectors; extracting standard disposal measures of each preplan in the historical emergency preplan set, and inputting the real-time monitoring data stream and the standard disposal measures into a feature fusion model. The application can quickly generate an emergency preplan by performing semantic coding and feature extraction on the event description text and the real-time monitoring data stream, thereby improving the automation level and real-time performance of emergency response.

Description

Technical Field

[0001] This invention relates to the field of emergency response plan generation technology, specifically to a method and system for generating emergency response plans for sudden environmental events. Background Technology

[0002] Currently, traditional methods typically rely on manual input of event information, and contingency plans are often generated based on past experience or templates, resulting in long response times, a lack of real-time dynamic adjustment capabilities, and an inability to quickly respond to emergencies. Furthermore, traditional methods often lack the support of real-time monitoring data and the ability to make dynamic adjustments based on real-time changes. Event handling is usually static and lacks a flexible emergency response mechanism, which makes it impossible for traditional methods to effectively cope with complex emergencies in a rapidly changing environment.

[0003] Furthermore, in terms of resource allocation, traditional methods often rely on manual judgment or prior planning, which may lead to situations of resource surplus or shortage. They also fail to consider the optimization and adjustment of time nodes and resource conflicts, resulting in unreasonable resource allocation and affecting the overall emergency response effect. Moreover, traditional methods usually rely on the experience of specific departments or personnel, lack effective historical data support, and cannot fully draw on past successful experiences in emergency response. Each time a new event is dealt with, a reassessment may be required, increasing the uncertainty and the possibility of errors. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for generating an emergency response plan for sudden environmental incidents, comprising:

[0005] The event description text and real-time monitoring data stream of the sudden environmental event are obtained, and the event description text is input into a pre-trained language model for semantic encoding to obtain the event semantic feature vector.

[0006] Based on the semantic feature vector of the event, a similarity search is performed in the preset case library to obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors.

[0007] Extract the standard response measures for each plan from the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes.

[0008] The initial response strategy is input into the pre-trained strategy adjustment model to generate a logistic matrix. The logistic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures.

[0009] The selection probabilities in the logic matrix are processed to increase the probability values ​​of key response measures, resulting in an adjusted logic matrix. The key response measures are determined based on high-frequency measures in the historical emergency plan set.

[0010] Based on the adjusted logic matrix, specific emergency measures for sudden environmental events are determined, and intermediate emergency plans are generated based on the control parameters in the preliminary response strategy.

[0011] Preferably, after generating an intermediate emergency response plan based on the control parameters in the initial response strategy, the method further includes:

[0012] The emergency resource knowledge graph is invoked to identify resource demand nodes in the intermediate emergency plan, and a list of candidate emergency resources is determined based on the connection relationship of the resource nodes.

[0013] Based on the event level of the sudden environmental incident and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated.

[0014] Insert the resource scheduling plan into the corresponding time node of the intermediate emergency plan, check whether there is a resource or time conflict, and if there is a conflict, adjust the time node based on the preset rules to generate a conflict-free target emergency plan.

[0015] Output the target emergency response plan and add it as new historical data to the preset case library to update the parameters of the pre-trained strategy to adjust the model.

[0016] Preferably, a similarity search is performed in a pre-defined case library based on the event semantic feature vector to obtain a set of historical emergency response plans that match the sudden environmental event, including:

[0017] Calculate the cosine similarity between the event semantic feature vector and the feature vector of each historical emergency plan in the preset case library to obtain the vector similarity score;

[0018] Extract the textual structural features of the event description text and compare them with the textual structural features of historical emergency plans to obtain a structural similarity score;

[0019] Identify the event type labels of sudden environmental events and calculate their overlap with the event type labels of historical emergency plans to obtain a label matching score;

[0020] The vector similarity score, structural similarity score, and label matching score are weighted and summed to obtain the total similarity score for each historical emergency plan.

[0021] Historical emergency response plans with a total similarity score greater than a preset threshold are selected and compiled into a set of historical emergency response plans.

[0022] Preferably, the real-time monitoring data stream and standard handling measures are input into a feature fusion model to obtain a dynamically adjusted preliminary handling strategy, including:

[0023] The pollutant concentration data and meteorological data in the real-time monitoring data stream are normalized to obtain the environmental state vector;

[0024] The environmental state vector is concatenated with the standard response measures in the historical emergency plan set and input into the neural network layer of the feature fusion model;

[0025] The correction coefficients for standard response measures are calculated using a neural network layer to determine the environmental state vector. These correction coefficients are then used to adjust the priority of evacuation routes.

[0026] Based on the correction coefficient, the evacuation routes and control parameters in the standard response measures are weighted and adjusted to generate a preliminary response strategy.

[0027] Preferably, the selection probabilities in the logic matrix are processed to increase the probability values ​​of key actions, resulting in an adjusted logic matrix, including:

[0028] By using a pre-training strategy to adjust the fully connected layers in the model and perform a linear transformation on the logistic matrix, the probability distribution to be evaluated is obtained.

[0029] The basic selection probability of candidate emergency measures is obtained by processing the probability distribution to be evaluated using a normalization function;

[0030] Identify the measures corresponding to the highest probability value in the basic selection probabilities, and determine the target-specified candidate emergency measures for non-highest probability values;

[0031] The base selection probability of the target-specified candidate emergency measures is adjusted to a preset multiple of the maximum probability value, or random sampling is performed within a preset multiple range of the maximum probability value to obtain the adjusted selection probability;

[0032] The adjusted logic matrix is ​​composed of adjusted selection probabilities.

[0033] Preferably, based on the adjusted logic matrix, specific emergency measures for sudden environmental events are determined, and combined with the control parameters in the preliminary response strategy, an intermediate emergency plan is generated, including:

[0034] The final selection probability of each candidate emergency measure is determined based on the adjusted logic matrix;

[0035] A probabilistic sampling strategy is used to select specific emergency measures from candidate emergency measures based on the final selection probability, thereby generating a measure sequence;

[0036] By combining the sequence of measures with the pollutant control parameters and evacuation routes in the initial response strategy, an intermediate emergency plan text is generated.

[0037] Preferably, an emergency resource knowledge graph is invoked to identify resource demand nodes in the intermediate emergency plan, and a candidate emergency resource list is determined based on the connection relationships of the resource nodes, including:

[0038] Entity extraction was performed on the intermediate emergency response plan to identify key handling points involving pollutant interception and medical treatment;

[0039] Search the emergency resource knowledge graph for resource entity nodes that have direct edge connections with key response nodes;

[0040] Extract the attribute information of the resource entity nodes, including resource type, storage location, and available quantity;

[0041] All resource entity nodes that meet the requirements of critical response nodes are aggregated to generate a candidate emergency resource list.

[0042] Preferably, based on the event level of the sudden environmental incident and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated, including:

[0043] Obtain the event level of the sudden environmental incident and determine the urgency coefficient of the resource requirement based on the event level;

[0044] Obtain the impact range data from the real-time monitoring data stream and calculate the resource demand within the impact range;

[0045] Input the urgency factor and resource demand into the filtering model to filter out the target emergency resources that meet the demand and have the shortest transportation time from the candidate emergency resource list.

[0046] Based on the geographical location of key response nodes and the storage location of target emergency resources, plan multi-target transportation routes;

[0047] By combining target emergency resources, transportation routes, and time windows, a resource scheduling plan is generated.

[0048] Preferably, the resource scheduling plan is inserted into the corresponding time node of the intermediate emergency plan, and the existence of resource or time conflicts is detected. If a conflict exists, the time node is adjusted based on preset rules to generate a conflict-free target emergency plan, including:

[0049] Align the resource arrival time in the resource scheduling plan with the response start time in the intermediate emergency plan;

[0050] To detect whether there are multiple resources competing for the same transportation channel at the same time point, or whether the arrival time of resources is later than the time required for disposal;

[0051] If a conflict is detected, the resource transportation order will be adjusted or backup resources will be enabled based on the preset priority rules.

[0052] Repeat the testing until there are no conflicts in all time points and resource allocations, and use the final text as the target contingency plan.

[0053] A system for generating emergency response plans for sudden environmental incidents, applicable to the aforementioned method for generating such plans, includes:

[0054] The semantic encoding unit is used to acquire the event description text and real-time monitoring data stream of sudden environmental events, and input the event description text into a pre-trained language model for semantic encoding to obtain the event semantic feature vector;

[0055] The case retrieval unit is used to perform similarity retrieval in a preset case library based on the semantic feature vector of the event, and obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors.

[0056] The feature fusion unit is used to extract the standard response measures of each plan in the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes.

[0057] The strategy adjustment unit is used to input the initial response strategy into the pre-trained strategy adjustment model and generate a logic matrix. The logic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures.

[0058] The probability processing unit is used to process the selection probabilities in the logic matrix, increase the probability value of key response measures, and obtain the adjusted logic matrix. The key response measures are determined based on high-frequency measures in the historical emergency plan set.

[0059] The contingency plan generation unit is used to determine specific emergency measures for sudden environmental events based on the adjusted logic matrix, and to generate intermediate emergency plans based on the control parameters in the preliminary response strategy.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] (1) By semantically encoding and feature extraction of event description text and real-time monitoring data stream, this invention can quickly generate emergency plans, improve the automation level and real-time performance of emergency response, and effectively learn from past successful experiences by using historical emergency plans in the preset case library through similarity retrieval, reduce uncertainty in handling new events, improve the pertinence and effectiveness of emergency measures, and dynamically adjust the initial handling strategy through feature fusion model, which can flexibly optimize emergency measures according to changes in real-time monitoring data, and ensure that efficient response capabilities are maintained in a rapidly changing environment.

[0062] (2) By calling the emergency resource knowledge graph, this invention can accurately identify the required resources and generate a reasonable resource scheduling plan, ensuring that resources are used optimally in emergency response, reducing the risk of resource waste and allocation errors. Moreover, during the resource scheduling process, the conflict between time nodes and resources is detected and adjusted to ensure the smooth implementation of various measures, improve the feasibility and scientific nature of the overall emergency plan, and realize the accumulation and updating of knowledge by adding the generated target emergency plan as new historical data to the case library, so that the emergency plan generation method can be continuously optimized to adapt to different types of sudden environmental events that may occur in the future. Attached Figure Description

[0063] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0065] In the diagram: 1. Semantic encoding unit; 2. Case retrieval unit; 3. Feature fusion unit; 4. Strategy adjustment unit; 5. Probability processing unit; 6. Draft plan generation unit. Detailed Implementation

[0066] 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.

[0067] Example 1, please refer to Figure 1 This invention provides a technical solution: a method for generating an emergency response plan for sudden environmental events, comprising:

[0068] S1. Obtain the event description text and real-time monitoring data stream of the sudden environmental event, and input the event description text into the pre-trained language model for semantic encoding to obtain the event semantic feature vector;

[0069] S2. Based on the event semantic feature vector, perform similarity retrieval in the preset case library to obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors.

[0070] S3. Extract the standard response measures for each plan in the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes.

[0071] S4. Input the initial response strategy into the pre-trained strategy adjustment model to generate a logistic matrix. The logistic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures.

[0072] S5. Process the selection probabilities in the logic matrix to increase the probability value of key response measures, and obtain the adjusted logic matrix. The key response measures are determined based on high-frequency measures in the historical emergency plan set.

[0073] S6. Based on the adjusted logic matrix, determine the specific emergency measures for sudden environmental events, and generate intermediate emergency plans based on the control parameters in the preliminary handling strategy.

[0074] It should be noted that by acquiring descriptions of sudden environmental events (such as text descriptions or reports) and real-time monitoring data (such as pollution concentrations, weather data, etc.), these data provide a basis for subsequent decision-making. For example, if a chemical spill occurs in a certain area, the event description may include "Leakage type: chemical substance X, location: city center, leakage volume: 1000 liters"; at the same time, the monitoring data will show pollutant concentrations, wind direction, temperature, etc.

[0075] The event description text is input into a pre-trained language model (such as BERT) for semantic encoding, transforming it into a "feature vector." This vector reflects the semantic information of the event. Next, a pre-defined case library is used to search for historical emergency plans similar to the current event, and a matching emergency plan is obtained. For example, if there have been similar chemical spills in the past, the system will compare the text description with historical emergency plans to find emergency plans similar to the current event. For instance, historical plans may include standard operating procedures (SOPs) for handling chemical spills.

[0076] Extract standard emergency response measures (e.g., evacuation routes, cleanup measures, etc.) from historical emergency plans; then, integrate real-time monitoring data (e.g., pollutant concentrations, meteorological conditions) with these standard emergency measures to generate a dynamically adjusted preliminary response strategy; for example, if a historical plan recommends using a specific evacuation route to avoid the contaminated area, but real-time data indicates high wind speeds, it may be necessary to adjust the evacuation route to prevent the spread of pollutants.

[0077] Preliminary emergency strategies (e.g., pollutant control measures and evacuation routes) are input into a strategy adjustment model. By optimizing the selection probabilities, a "logic matrix" is generated. This matrix represents the selection probability of various emergency measures, so as to select the most appropriate measure according to different situations. For example, if the calculation result of the logic matrix shows that the selection probability of evacuation route A is higher than that of evacuation route B, then in most cases, the system will prioritize recommending route A. However, if real-time data changes (such as changes in wind speed), the system will dynamically adjust the recommended plan.

[0078] Based on the adjusted logic matrix, the specific emergency measures are finally determined. At the same time, the system will generate an intermediate emergency plan based on the control parameters in the preliminary disposal strategy (such as the concentration requirements for pollutant control). For example, if the preliminary strategy indicates that the pollutant concentration needs to be reduced to a certain safety standard and people in a certain area need to be evacuated, then the intermediate emergency plan may include specific measures for mobilizing emergency equipment, evacuating personnel, and implementing pollutant treatment.

[0079] The process involves updating the parameters of the initial policy adjustment model using the loss value to obtain a pre-trained policy adjustment model, including:

[0080] The first probability distribution is formed using the basic selection probabilities, and the second probability distribution is formed using the adjusted selection probabilities.

[0081] Calculate the relative entropy between the first probability distribution and the second probability distribution, and use the relative entropy as the loss value;

[0082] The loss value is used to update the gradient of the fully connected layer parameters of the model by adjusting the initial policy through the backpropagation algorithm;

[0083] Repeat the above steps until the loss value converges to obtain the pre-trained policy adjustment model.

[0084] In an optional embodiment, after generating an intermediate emergency plan based on the control parameters in the initial response strategy, the method further includes:

[0085] The emergency resource knowledge graph is invoked to identify resource demand nodes in the intermediate emergency plan, and a list of candidate emergency resources is determined based on the connection relationship of the resource nodes.

[0086] Based on the event level of the sudden environmental incident and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated.

[0087] Insert the resource scheduling plan into the corresponding time node of the intermediate emergency plan, check whether there is a resource or time conflict, and if there is a conflict, adjust the time node based on the preset rules to generate a conflict-free target emergency plan.

[0088] Output the target emergency response plan and add it as new historical data to the preset case library to update the parameters of the pre-trained strategy to adjust the model.

[0089] It should be noted that after generating the intermediate emergency plan, the system will invoke a knowledge graph containing information related to emergency resources, such as equipment, personnel, and supplies. The system identifies the required resource demand nodes by analyzing the control parameters in the intermediate emergency plan. For example, suppose the intermediate emergency plan mentions the need for specific cleaning equipment (such as vacuum cleaners, chemical neutralizers, etc.) and professionals (such as chemical engineers and firefighters); the knowledge graph will help identify these resource requirements.

[0090] Based on the identified resource demand nodes, the system will search for and determine a list of candidate emergency resources related to the connection relationships of these nodes; these resources may be equipment, personnel and other support available in the region; for example, if cleaning equipment and professionals are needed, the system will filter out local emergency response teams, available cleaning equipment and available expert resources, and generate a list of candidate resources.

[0091] Based on the severity of the event (e.g., minor, moderate, severe) and the scope of impact in the real-time monitoring data stream, the most suitable target emergency resources are further filtered from the candidate resource list. This will form a resource scheduling plan to ensure that the required resources are deployed in the appropriate time. For example, if the event is assessed as severe and the impact covers a large area, the system may dispatch more cleanup equipment and personnel and design a scheduling plan to ensure that these resources can arrive at the accident site in the shortest possible time.

[0092] Once the resource scheduling plan is determined, the system will insert it into the time nodes of the intermediate emergency plan. At this time, the system needs to detect whether there are resource conflicts (e.g., the same resource is scheduled multiple times) or time conflicts (e.g., the time of a certain step overlaps). For example, suppose the scheduling plan requires the use of a specific cleaning device at a certain time, but at the same time, this device has been assigned to another event; the system will detect this conflict.

[0093] If a conflict is detected, the system will make adjustments based on preset rules; for example, it may postpone the time point of a certain step or rearrange the order of resource usage to ensure that all resources can be effectively utilized when needed; for example, if the cleaning equipment is unavailable in the initial stage of the incident handling, the system may postpone the time point for using the equipment or select a backup equipment to replace it, thereby ensuring the smooth implementation of the entire emergency plan.

[0094] The adjusted and optimized target emergency plan will be output; this plan not only includes emergency measures, but also details the resource allocation plan and time schedule; for example, the target emergency plan may include: dispatching two cleanup vehicles and ten personnel to the site at a certain time, while dispatching chemical engineers to provide guidance, and stipulating that an effectiveness evaluation be carried out after the cleanup is completed;

[0095] The generated target emergency plan is added to the pre-set case library as new historical data. This step is to continuously enrich and update the knowledge base, making future emergency responses more efficient. It also helps to optimize the pre-training strategy and adjust the model parameters to improve the accuracy of future decisions. For example, this new emergency plan will become part of the historical cases for reference when encountering similar events in the future, helping the system to generate emergency plans more accurately.

[0096] In an optional embodiment, a similarity search is performed in a preset case library based on the event semantic feature vector to obtain a set of historical emergency plans matching the sudden environmental event, including:

[0097] Calculate the cosine similarity between the event semantic feature vector and the feature vector of each historical emergency plan in the preset case library to obtain the vector similarity score;

[0098] Extract the textual structural features of the event description text and compare them with the textual structural features of historical emergency plans to obtain a structural similarity score;

[0099] Identify the event type labels of sudden environmental events and calculate their overlap with the event type labels of historical emergency plans to obtain a label matching score;

[0100] The vector similarity score, structural similarity score, and label matching score are weighted and summed to obtain the total similarity score for each historical emergency plan.

[0101] Historical emergency response plans with a total similarity score greater than a preset threshold are selected and compiled into a set of historical emergency response plans.

[0102] It should be noted that the description of a sudden environmental event is converted into a semantic feature vector (i.e., a digital vector representation). Then, by calculating the cosine similarity between this event vector and the feature vectors of each historical emergency plan, the system can obtain a "similarity score." This score represents the degree of similarity between the event and the historical plan. For example, assuming a sudden event is a "chemical leak," the system will first convert this event description into a vector. If there is a historical emergency plan for a "chemical leak accident," then the feature vector of that plan will have a high similarity score with the vector of the sudden event, indicating that the two are very well matched.

[0103] In addition to semantic features, the system also analyzes the "textual structure features" of the event description, such as the sentence structure, the order of keywords, and the level of description. These structural features are compared with the textual structure features of historical emergency plans to calculate a structural similarity score. Structural similarity helps the system determine whether the structure of the event description is similar to that of the historical plan. For example, if the description structure of the emergency is: "A chemical leak has occurred, causing environmental pollution, and nearby personnel need to be evacuated immediately," and the structure in the historical emergency plan is similar, with the same order of description (e.g., "Chemical leak, evacuation of personnel, pollution control"), then the structural similarity will be high.

[0104] The system identifies "event type tags" for emergencies, such as whether they are "chemical leaks," "fires," or "water pollution." Then, it compares the emergency's tag with tags in historical emergency plans, calculating a tag matching score. This score reflects the degree of matching between event types. For example, if the emergency is a "fire," the system automatically tags it with "fire" and matches it against historical emergency plans that have the "fire" tag. If a historical plan also has the "fire" tag, the tag matching score is considered high.

[0105] After calculating the three types of similarity mentioned above, the system will sum the scores of each historical emergency plan according to certain weights to obtain the "total similarity score". Different factors (semantics, structure, and tags) may have different importance, and the system will assign different weights to them according to preset rules. For example, if semantic similarity accounts for 60%, text structure similarity accounts for 30%, and tag matching accounts for 10%, the final total similarity score will combine the weighted results of these three factors. For example, if a historical plan has very similar semantics and structure, but low tag matching, it may still have a high total similarity score.

[0106] Historical emergency response plans with a total similarity score greater than a preset threshold are selected. Only those plans with a high degree of matching will be included in the "historical emergency response plan set" for actual emergency response. For example, if the threshold is set to 70%, a historical plan with a total similarity score of 75% will be selected, while another historical plan with a total similarity score of only 65% ​​will be excluded.

[0107] In an optional embodiment, the real-time monitoring data stream and standard handling measures are input into a feature fusion model to obtain a dynamically adjusted preliminary handling strategy, including:

[0108] The pollutant concentration data and meteorological data in the real-time monitoring data stream are normalized to obtain the environmental state vector;

[0109] The environmental state vector is concatenated with the standard response measures in the historical emergency plan set and input into the neural network layer of the feature fusion model;

[0110] The correction coefficients for standard response measures are calculated using a neural network layer to determine the environmental state vector. These correction coefficients are then used to adjust the priority of evacuation routes.

[0111] Based on the correction coefficient, the evacuation routes and control parameters in the standard response measures are weighted and adjusted to generate a preliminary response strategy.

[0112] It should be noted that the real-time monitoring data stream includes environmental pollutant concentrations and meteorological information (such as wind direction, wind speed, temperature, humidity, etc.). In order for the model to process this data, it needs to be normalized first, that is, the data is converted into a uniform dimension range to form an "environmental state vector". For example, suppose a chemical leak occurs near a factory, and the monitoring station measures an ammonia concentration of 50 ppm, a wind speed of 12 m / s, and a wind direction of east. Through normalization, these data are converted into an environmental state vector with values ​​between 0 and 1, such as [0.5, 0.6, 0.8], which represent the state of pollution concentration, wind speed, and wind direction, respectively.

[0113] In the collection of historical emergency response plans, each plan has standard handling measures, including evacuation routes and control equipment operation. These standard handling measures (which can also be converted into vector representations) are concatenated with the current environmental state vector and fed into a feature fusion model (usually a neural network). For example: Standard handling measure vector: evacuation route priority [0.7, 0.3, 0.0] (representing the initial priority of the main exit, side exit, and backup exit); Current environmental state vector: [0.5, 0.6, 0.8] (pollution concentration, wind speed, wind direction); The concatenated input to the model is [0.5, 0.6, 0.8, 0.7, 0.3, 0.0].

[0114] The neural network layer calculates "correction coefficients" based on the input. These coefficients represent the degree to which standard response measures need to be adjusted under the current environmental conditions. The correction coefficients can be understood as dynamic adjustment factors for the priority of evacuation routes. For example, the model calculates correction coefficients [0.9, 1.1, 1.0]. Multiplying the priority of the first exit by 0.9 indicates that the risk is slightly higher and the evacuation priority needs to be reduced. Multiplying the priority of the second exit by 1.1 indicates that the route is relatively safe and the priority can be increased. The third exit remains unchanged.

[0115] By using correction coefficients to weight and adjust standard response measures, a dynamically optimized preliminary response strategy can be generated, which includes new evacuation route priorities and control parameters (such as access control, alarm activation sequence, etc.). For example: the original evacuation route priority is [0.7, 0.3, 0.0]; after adjustment, it becomes [0.7*0.9, 0.3*1.1, 0.0*1.0] = [0.63, 0.33, 0.0]. This result tells the emergency command center: prioritize evacuation from the second exit (relatively upgraded); the main exit is still used, but its priority is slightly reduced; the backup exit is not activated for the time being.

[0116] In an optional embodiment, the selection probabilities in the logic matrix are processed to increase the probability values ​​of key actions, resulting in an adjusted logic matrix, including:

[0117] By using a pre-training strategy to adjust the fully connected layers in the model and perform a linear transformation on the logistic matrix, the probability distribution to be evaluated is obtained.

[0118] The basic selection probability of candidate emergency measures is obtained by processing the probability distribution to be evaluated using a normalization function;

[0119] Identify the measures corresponding to the highest probability value in the basic selection probabilities, and determine the target-specified candidate emergency measures for non-highest probability values;

[0120] The base selection probability of the target-specified candidate emergency measures is adjusted to a preset multiple of the maximum probability value, or random sampling is performed within a preset multiple range of the maximum probability value to obtain the adjusted selection probability;

[0121] The adjusted logic matrix is ​​composed of adjusted selection probabilities.

[0122] It should be noted that the logistic matrix is ​​a table that records the logical relationships or selection weights between various candidate emergency measures. Each row represents the probability of selecting each measure under a scenario. By adjusting the model through pre-training strategies, the logistic matrix undergoes a linear transformation through the fully connected layer of the neural network to generate the "probability distribution to be evaluated" for each measure. For example, suppose there are 3 evacuation measures: main exit, side exit, and backup exit. The probability distribution corresponding to the original logistic matrix (unadjusted) is [0.4, 0.35, 0.25]. After the linear transformation by the fully connected layer, the "probability distribution to be evaluated" is obtained as [0.45, 0.33, 0.22].

[0123] To ensure that the sum of probabilities is 1 or within a reasonable range, the probability distribution to be evaluated needs to be normalized to obtain the basic selection probability, which is the basic probability of each measure being selected. For example, after normalization, the basic selection probabilities are [0.45, 0.33, 0.22]. The main exit has the highest probability of being selected, followed by the side exit, and then the backup exit.

[0124] The measure corresponding to the maximum probability value is the current optimal choice (critical measure); non-maximum probability values ​​are candidate measures that need to be adjusted; for example, the maximum probability value is 0.45, corresponding to the main exit; the target-specified candidate measures are the side exit and the backup exit (non-maximum value measures).

[0125] The basic selection probability of the target candidate measure is increased to a preset multiple of the maximum probability value, or random sampling is performed within this multiple range. This increases the selection probability of the critical measure while ensuring that other measures also have a certain chance of being selected. For example, if the preset multiple is 1.2, the side exit probability is originally 0.33 → adjusted to 0.45 × 1.2 = 0.54 (or randomly selected between 0.5 and 0.55); the backup exit probability is originally 0.22 → adjusted to 0.45 × 1.2 = 0.54 (similar to random sampling). In this way, the probability of the critical measure is appropriately increased.

[0126] In an optional embodiment, specific emergency measures for sudden environmental events are determined based on the adjusted logic matrix, and an intermediate emergency plan is generated by combining the control parameters in the initial response strategy, including:

[0127] The final selection probability of each candidate emergency measure is determined based on the adjusted logic matrix;

[0128] A probabilistic sampling strategy is used to select specific emergency measures from candidate emergency measures based on the final selection probability, thereby generating a measure sequence;

[0129] By combining the sequence of measures with the pollutant control parameters and evacuation routes in the initial response strategy, an intermediate emergency plan text is generated.

[0130] It should be noted that the final selection probability of each candidate emergency measure is obtained based on the adjusted logical matrix. These probabilities, after adjustment, reflect the priority and likelihood of each emergency measure being selected. For example, assuming a pollution emergency scenario, the adjusted logical matrix yields the following final selection probabilities: Measure A: controlling the source, final selection probability 0.6; Measure B: air purification, final selection probability 0.3; Measure C: evacuating people, final selection probability 0.1. These final selection probabilities reflect the weight of each emergency measure, and Measure A, which controls the source, is clearly the highest priority.

[0131] Using a probabilistic sampling strategy, actual emergency measures are selected from candidate emergency measures based on the final selection probability of each measure. Probabilistic sampling means that the chance of each emergency measure being selected is proportional to its final selection probability. For example, suppose the system needs to select three measures from these candidate measures as part of an emergency plan. Using probabilistic sampling, measures might be selected as follows: Measure A (source control) has a 60% probability of being selected; Measure B (air purification) has a 30% probability of being selected; Measure C (crowd evacuation) has a 10% probability of being selected. Suppose that after sampling, the resulting emergency measure sequence is: Measure A (source control); Measure A (source control); Measure B (air purification). Through sampling, Measure A is selected twice due to its higher probability, and Measure B is selected once.

[0132] Once specific emergency measures are selected, these measures need to be combined with the control parameters (such as pollutant control and evacuation routes) in the initial response strategy to form a complete emergency plan. For example, suppose the control parameters in the initial response strategy include: pollutant concentration control: requiring the pollution source to be sealed off and ensuring that air purification equipment is fully activated; evacuation routes: the evacuation routes should be set up to avoid the polluted area as much as possible to ensure smooth evacuation. Combining these control parameters with the selected sequence of emergency measures (measure A, measure B), the following intermediate emergency plan text can be generated:

[0133] Emergency measures: Close the pollution source to ensure effective control of pollutant emissions; strengthen monitoring to ensure no further pollutant leakage after the pollution source is closed; activate the air purification system to ensure that the pollutant concentration is reduced to a safe range; control parameters: the pollution source must be completely sealed off, and the concentration of pollutants in the air must be strictly monitored; avoid passing through the area where the pollution source is located during evacuation, and evacuation routes must be guaranteed to be free of pollutant leakage to ensure the safety of the population.

[0134] In an optional embodiment, an emergency resource knowledge graph is invoked to identify resource demand nodes in intermediate emergency plans, and a candidate emergency resource list is determined based on the connection relationships of the resource nodes, including:

[0135] Entity extraction was performed on the intermediate emergency response plan to identify key handling points involving pollutant interception and medical treatment;

[0136] Search the emergency resource knowledge graph for resource entity nodes that have direct edge connections with key response nodes;

[0137] Extract the attribute information of the resource entity nodes, including resource type, storage location, and available quantity;

[0138] All resource entity nodes that meet the requirements of critical response nodes are aggregated to generate a candidate emergency resource list.

[0139] It should be noted that the first step is to identify the key "nodes" or "events" in the existing emergency response plan texts; special attention should be paid to two types of response nodes: pollutant interception nodes: for example, actions to intercept and isolate pollutants when there is a chemical leak in a certain area; and medical treatment nodes: involving the treatment of the injured, the deployment of emergency stations, and the demand for medical equipment. Identifying these key nodes is equivalent to determining the "emergency demand points".

[0140] An emergency resource knowledge graph is a network graph that contains various resources and their interrelationships.

[0141] For each critical response node (contaminant interception or medical treatment), find the resource nodes that are directly connected to it in the knowledge graph; these resource nodes are the tools, equipment or materials that can directly support the response node; for example, for the contaminant interception node, the resources that may be connected include adsorbents, protective clothing, barrier materials, etc.; for the medical treatment node, the resources that may be connected include stretchers, first aid kits, medicines, etc.

[0142] Each resource node is more than just a name; it also has detailed attributes that help assess availability and deployment efficiency. Key attributes include: Resource type: such as medicine, equipment, personnel, or protective materials; Storage location: which warehouse or geographical location the resource is located in; Available quantity: the current amount of available inventory or the number of available personnel.

[0143] Organize all resource information directly related to key nodes to form a candidate emergency resource list;

[0144] Generating a candidate list of aggregated resources can help decision-makers quickly see which resources can be used for the current contingency plan's needs; this facilitates the next steps of resource scheduling, optimization, or emergency response deployment.

[0145] In an optional embodiment, based on the event level of the sudden environmental event and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated, including:

[0146] Obtain the event level of the sudden environmental incident and determine the urgency coefficient of the resource requirement based on the event level;

[0147] Obtain the impact range data from the real-time monitoring data stream and calculate the resource demand within the impact range;

[0148] Input the urgency factor and resource demand into the filtering model to filter out the target emergency resources that meet the demand and have the shortest transportation time from the candidate emergency resource list.

[0149] Based on the geographical location of key response nodes and the storage location of target emergency resources, plan multi-target transportation routes;

[0150] By combining target emergency resources, transportation routes, and time windows, a resource scheduling plan is generated.

[0151] It should be noted that assessing a sudden environmental incident and determining its "incident level" is usually based on factors such as the severity of the incident and its potential impact.

[0152] Purpose: The event level will be used to determine the urgency of resource requests, and this urgency level will affect the priority of subsequent resource scheduling;

[0153] Based on the event level just obtained, determine an "urgency factor"; the higher the event level, the greater the urgency factor, which means that corresponding resources need to be allocated more quickly and with higher priority.

[0154] Data related to the event is collected through a real-time monitoring system. This data includes the geographical extent of the event's impact and the affected areas. Understanding the scope of impact is crucial because it will help determine the scale of resource input required.

[0155] After determining the geographical extent of the event's impact, assess the amount of resources that may be needed in that area; if a region is contaminated, how many tons of absorbent, how much medical equipment, etc., may be required, all of these must be taken into account.

[0156] Using urgency level and resource demand as input, a screening model is used to identify suitable emergency resources from a list of candidate emergency resources. During this screening process, the shortest transportation time for each resource must also be considered, as rapid response is crucial for emergency management. Ultimately, a list of "target emergency resources" is generated, which can meet the demand and arrive at the scene quickly.

[0157] Design an optimal transportation route based on key disposal points (such as the location of pollutant interception) and the storage location of target emergency resources; consider multiple factors, such as transportation efficiency, traffic conditions, and distance, to ensure that resources can be delivered as quickly as possible;

[0158] By combining target emergency resources, transportation routes, and time windows, a complete resource scheduling plan is formed; the generated resource scheduling plan details which resources need to be delivered to the designated location at which time and via which route, ensuring the timeliness and effectiveness of emergency response.

[0159] In an optional embodiment, the resource scheduling scheme is inserted into the corresponding time node of the intermediate emergency plan, and the existence of resource or time conflicts is detected. If a conflict exists, the time node is adjusted based on preset rules to generate a conflict-free target emergency plan, including:

[0160] Align the resource arrival time in the resource scheduling plan with the response start time in the intermediate emergency plan;

[0161] To detect whether there are multiple resources competing for the same transportation channel at the same time point, or whether the arrival time of resources is later than the time required for disposal;

[0162] If a conflict is detected, the resource transportation order will be adjusted or backup resources will be enabled based on the preset priority rules.

[0163] Repeat the testing until there are no conflicts in all time points and resource allocations, and use the final text as the target contingency plan.

[0164] It should be noted that the estimated arrival time of each resource in the resource scheduling plan should first be matched with the planned start time of the emergency response in the intermediate emergency plan; this ensures that the resources arrive on time, neither too early nor too late, so as to support the execution of each response node.

[0165] Check whether the following situations occur within the same time period: different resources need to use the same transportation channel or means of transport, which may cause congestion or delays; the arrival time of some resources is later than the time when the node needs to start processing, which affects the processing; identify problems in time to prevent bottlenecks or delays in actual execution;

[0166] If a conflict is found, adjustments need to be made; based on pre-set priority rules, adjust which resources should be transported first and which can be transported later; if the main resources cannot arrive on time, backup resources can be activated to ensure that the processing can proceed normally; through these measures, delays and resource conflicts should be avoided as much as possible.

[0167] After adjustments, perform another check to confirm if conflicts still exist; if conflicts still exist, continue adjusting until all time points and resource allocations are problem-free.

[0168] Once all conflicts are resolved, these adjusted timelines and resource allocations will be integrated into a final conflict-free emergency plan. The final emergency plan can be directly used for actual emergency operations, with corresponding resources for each response node and reasonable time arrangements to prevent conflicts or delays.

[0169] Example 2, please refer to Figure 2 This invention provides a technical solution: a system for generating emergency response plans for sudden environmental incidents, applicable to the aforementioned method for generating emergency response plans for sudden environmental incidents, comprising:

[0170] Semantic encoding unit 1 is used to acquire event description text and real-time monitoring data stream of sudden environmental events, and input the event description text into a pre-trained language model for semantic encoding to obtain event semantic feature vector;

[0171] Case retrieval unit 2 is used to perform similarity retrieval in a preset case library based on the event semantic feature vector to obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors.

[0172] Feature fusion unit 3 is used to extract the standard response measures of each plan in the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes.

[0173] The strategy adjustment unit 4 is used to input the initial response strategy into the pre-trained strategy adjustment model and generate a logic matrix. The logic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures.

[0174] The probability processing unit 5 is used to process the selection probability in the logic matrix, increase the probability value of the key response measures, and obtain the adjusted logic matrix. The key response measures are determined based on the high-frequency measures in the historical emergency plan set.

[0175] The contingency plan generation unit 6 is used to determine specific emergency measures for sudden environmental events based on the adjusted logic matrix, and to generate intermediate emergency plans based on the control parameters in the preliminary handling strategy.

[0176] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for generating an emergency response plan for a sudden environmental incident, characterized in that, include: The event description text and real-time monitoring data stream of the sudden environmental event are obtained, and the event description text is input into a pre-trained language model for semantic encoding to obtain the event semantic feature vector. Based on the semantic feature vector of the event, a similarity search is performed in the preset case library to obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors. Extract the standard response measures for each plan from the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes. The initial response strategy is input into the pre-trained strategy adjustment model to generate a logistic matrix. The logistic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures. The selection probabilities in the logic matrix are processed to obtain the adjusted logic matrix; Based on the adjusted logic matrix, specific emergency measures for sudden environmental events are determined, and intermediate emergency plans are generated based on the control parameters in the preliminary response strategy. After generating an intermediate emergency response plan based on the control parameters in the initial response strategy, the method further includes: The emergency resource knowledge graph is invoked to identify resource demand nodes in the intermediate emergency plan, and a list of candidate emergency resources is determined based on the connection relationship of the resource nodes. Based on the event level of the sudden environmental incident and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated. Insert the resource scheduling plan into the corresponding time node of the intermediate emergency plan, check whether there is a resource or time conflict, and if there is a conflict, adjust the time node based on the preset rules to generate a conflict-free target emergency plan. Output the target emergency response plan and add it as new historical data to the preset case library to update the parameters of the model by adjusting the pre-trained strategy; The method of inputting real-time monitoring data streams and standard handling measures into a feature fusion model to obtain a dynamically adjusted preliminary handling strategy includes: The pollutant concentration data and meteorological data in the real-time monitoring data stream are normalized to obtain the environmental state vector; The environmental state vector is concatenated with the standard response measures in the historical emergency plan set and input into the neural network layer of the feature fusion model; The correction coefficients for standard response measures are calculated using a neural network layer to determine the environmental state vector. These correction coefficients are then used to adjust the priority of evacuation routes. Based on the correction coefficient, the evacuation routes and control parameters in the standard response measures are weighted and adjusted to obtain the preliminary response strategy after dynamic adjustment. The process of processing the selection probabilities in the logic matrix to obtain the adjusted logic matrix includes: By using a pre-training strategy to adjust the fully connected layers in the model and perform a linear transformation on the logistic matrix, the probability distribution to be evaluated is obtained. The basic selection probability of candidate emergency measures is obtained by processing the probability distribution to be evaluated using a normalization function; Identify the measures corresponding to the highest probability value in the basic selection probabilities, and determine the target-specified candidate emergency measures for non-highest probability values; The base selection probability of the target-specified candidate emergency measures is adjusted to a preset multiple of the maximum probability value, or random sampling is performed within a preset multiple range of the maximum probability value to obtain the adjusted selection probability; The adjusted logic matrix is ​​composed of adjusted selection probabilities.

2. The method for generating an emergency response plan for a sudden environmental incident according to claim 1, characterized in that, Based on the semantic feature vector of the event, a similarity search is performed in a pre-set case database to obtain a set of historical emergency response plans that match the sudden environmental event, including: Calculate the cosine similarity between the event semantic feature vector and the feature vector of each historical emergency plan in the preset case library to obtain the vector similarity score; Extract the textual structural features of the event description text and compare them with the textual structural features of historical emergency plans to obtain a structural similarity score; Identify the event type labels of sudden environmental events and calculate their overlap with the event type labels of historical emergency plans to obtain a label matching score; The vector similarity score, structural similarity score, and label matching score are weighted and summed to obtain the total similarity score for each historical emergency plan. Historical emergency response plans with a total similarity score greater than a preset threshold are selected and compiled into a set of historical emergency response plans.

3. The method for generating an emergency response plan for a sudden environmental incident according to claim 2, characterized in that, Based on the adjusted logic matrix, specific emergency measures for sudden environmental events are determined, and combined with the control parameters in the preliminary response strategy, an intermediate emergency plan is generated, including: The final selection probability of each candidate emergency measure is determined based on the adjusted logic matrix; A probabilistic sampling strategy is used to select specific emergency measures from candidate emergency measures based on the final selection probability, thereby generating a measure sequence; By combining the sequence of measures with the pollutant control parameters and evacuation routes in the initial response strategy, an intermediate emergency plan text is generated.

4. The method for generating an emergency response plan for a sudden environmental incident according to claim 3, characterized in that, The emergency resource knowledge graph is invoked to identify resource requirement nodes in intermediate emergency plans, and a candidate emergency resource list is determined based on the connection relationships between these resource nodes, including: Entity extraction was performed on the intermediate emergency response plan to identify key handling points involving pollutant interception and medical treatment; Search the emergency resource knowledge graph for resource entity nodes that have direct edge connections with key response nodes; Extract the attribute information of the resource entity nodes, including resource type, storage location, and available quantity; All resource entity nodes that meet the requirements of critical response nodes are aggregated to generate a candidate emergency resource list.

5. The method for generating an emergency response plan for a sudden environmental incident according to claim 4, characterized in that, Based on the event level of the sudden environmental incident and the impact range data in the real-time monitoring data stream, target emergency resources are selected from the candidate emergency resource list, and a resource scheduling plan is generated, including: Obtain the event level of the sudden environmental incident and determine the urgency coefficient of the resource requirement based on the event level; Obtain the impact range data from the real-time monitoring data stream and calculate the resource demand within the impact range; Input the urgency factor and resource demand into the filtering model to filter out the target emergency resources that meet the demand and have the shortest transportation time from the candidate emergency resource list. Based on the geographical location of key response nodes and the storage location of target emergency resources, plan multi-target transportation routes; By combining target emergency resources, transportation routes, and time windows, a resource scheduling plan is generated.

6. The method for generating an emergency response plan for a sudden environmental incident according to claim 5, characterized in that, The resource scheduling plan is inserted into the corresponding time node of the intermediate emergency plan to check for resource or time conflicts. If a conflict exists, the time node is adjusted based on preset rules to generate a conflict-free target emergency plan, including: Align the resource arrival time in the resource scheduling plan with the response start time in the intermediate emergency plan; To detect whether there are multiple resources competing for the same transportation channel at the same time point, or whether the arrival time of resources is later than the time required for disposal; If a conflict is detected, the resource transportation order will be adjusted or backup resources will be enabled based on the preset priority rules. Repeat the testing until there are no conflicts in all time points and resource allocations, and use the final text as the target contingency plan.

7. A system for generating emergency response plans for sudden environmental incidents, applicable to the method for generating emergency response plans for sudden environmental incidents as described in any one of claims 1-6, characterized in that, include: The semantic encoding unit is used to acquire the event description text and real-time monitoring data stream of sudden environmental events, and input the event description text into a pre-trained language model for semantic encoding to obtain the event semantic feature vector; The case retrieval unit is used to perform similarity retrieval in a preset case library based on the semantic feature vector of the event, and obtain a set of historical emergency plans that match the sudden environmental event. The preset case library includes historical emergency plans and their corresponding feature vectors. The feature fusion unit is used to extract the standard response measures of each plan in the historical emergency response plan set, and input the real-time monitoring data stream and the standard response measures into the feature fusion model to obtain the dynamically adjusted preliminary response strategy. The preliminary response strategy includes pollutant control parameters and evacuation routes. The strategy adjustment unit is used to input the initial response strategy into the pre-trained strategy adjustment model and generate a logic matrix. The logic matrix is ​​used to represent the selection probability of candidate emergency measures. The pre-trained strategy adjustment model is used to optimize the selection probability distribution of measures. The probability processing unit is used to process the selection probabilities in the logic matrix to obtain the adjusted logic matrix. The contingency plan generation unit is used to determine specific emergency measures for sudden environmental events based on the adjusted logic matrix, and to generate intermediate emergency plans based on the control parameters in the preliminary response strategy.

Images