An event-driven emergency command and dispatch method

Abstract

The present application discloses an event-driven emergency command and dispatch method, which relates to the field of emergency command and dispatch, and includes: using feature engineering to extract features from historical emergency event data; wherein, the feature data set includes an event type feature vector, an event frequency feature vector, a distribution density feature vector, and a text feature vector; training an emergency event classification model based on a CART decision tree using the feature data set; obtaining real-time emergency event data, and performing feature extraction on the real-time emergency event data to obtain a real-time event feature vector; using the obtained real-time event feature vector as input, and identifying emergency events using the emergency event classification model; according to the identified emergency events, adopting a rule-based case reasoning method to retrieve and generate an emergency command and dispatch plan from a preset emergency plan knowledge base. Aiming at the problem of low personalization degree of emergency command and dispatch plans in the prior art, the present application retrieves and generates a personalized emergency command and dispatch plan.

Description

Technical Field

[0001] This application relates to the field of emergency command and dispatch, and particularly to an event-driven emergency command and dispatch method. Background Art

[0002] With the rapid development of social economy and the increasing frequency of human activities, various emergencies and natural disasters present complex characteristics such as multiple occurrences, frequent occurrences, and coupling. The effective response and handling of emergency events are facing greater and greater challenges. Emergency command and dispatch is the core link of emergency management, and the scientificity and effectiveness of its plan directly affect the emergency handling effect. However, due to the complex and changeable nature of emergency events, traditional emergency command and dispatch methods are difficult to meet the growing personalized needs. Therefore, researching new intelligent and precise emergency command and dispatch methods to improve the personalized level and real-time response ability of emergency plans is of great significance for enhancing the emergency management level of sudden events.

[0003] In recent years, the development of emerging information technologies such as artificial intelligence and big data has provided new ways for the intelligent transformation of emergency command. Introducing artificial intelligence technologies such as machine learning and knowledge engineering into the field of emergency command and constructing an intelligent decision-making model driven by the integration of data and knowledge are research hotspots and development trends in the field of emergency management. However, current related research mainly focuses on aspects such as the optimization of emergency resource scheduling and the planning of rescue paths, and the research on the real-time generation and personalized construction of emergency plans is still relatively weak.

[0004] Existing emergency command and dispatch methods mainly include a matching method based on emergency plans and an artificial formulation method based on expert experience. The method based on plan matching retrieves the reference dispatch plan from the emergency plan library according to the type and characteristics of the emergency event by establishing an emergency plan library. However, most of the plans in the plan library are general plans designed for typical scenarios, which are difficult to accurately match complex and changeable actual emergency events, and the pertinence and personalization degree of the plans are insufficient. The artificial formulation method based on expert experience depends on the experience and ability of the command personnel, has problems such as strong subjectivity, low formulation efficiency, and difficulty in dealing with new situations, and it is difficult to guarantee the timeliness and reliability of the plan. Summary of the Invention

[0005] Aiming at the problem of low personalization degree of the existing emergency command and dispatch plan, this application provides an event-driven emergency command and dispatch method, which retrieves and generates a personalized emergency command and dispatch plan through ontology-based and case-based reasoning, etc.

[0006] The purpose of this application is achieved through the following technical solutions.

[0007] The present application provides an event-driven emergency command and dispatch method, including: obtaining historical emergency event data, which includes event type, event time, event location, and disposal plan; using feature engineering to extract features from the historical emergency event data to obtain a feature dataset; wherein, the feature dataset includes an event type feature vector, an event frequency feature vector, a distribution density feature vector, and a text feature vector; training an emergency event classification model based on a CART decision tree using the feature dataset; obtaining real-time emergency event data, and extracting features from the real-time emergency event data to obtain a real-time event feature vector; using the obtained real-time event feature vector as an input, and identifying the emergency event using the emergency event classification model; according to the identified emergency event, adopting a rule-based case-based reasoning method to retrieve and generate an emergency command and dispatch plan from a preset emergency plan knowledge base.

[0008] Wherein, CART is an abbreviation of "Classification and Regression Tree", referring to Classification and Regression Tree. In the present application, a CART decision tree model is trained using the feature dataset to obtain an emergency event classification model. This classifier can be used to determine which type a new emergency event belongs to. Case-based Reasoning (CBR) is an analogical reasoning method. By retrieving existing cases similar to the new problem, the solutions of historical cases are obtained, and then adjusted or modified to adapt to the new problem. In the present application, according to the event classification result, the relevant emergency plan is found in the case library (emergency plan knowledge base) using the rule-based CBR method, and a specific emergency dispatch plan is generated accordingly. In the present application, the historical emergency event data: is used for model training, that is, to generate an emergency event classifier. The historical data includes event type, occurrence time, location, adopted disposal plan, etc. The real-time emergency event data: refers to the information of sudden emergency events and is the input of the model. By extracting features from the real-time data and then using the trained classifier for identification. The identification result will be used for subsequent plan retrieval.

[0009] Furthermore, feature engineering is used to extract features from historical emergency event data, including: using One-Hot encoding with the same number of binary digits as the number of event types to encode the event types and obtain event type feature vectors; using the sliding time window method to perform time window statistics on the event time to obtain event frequency feature vectors at different time scales; dividing the event locations into geographical grids and using the kernel density estimation method to estimate the number of events in each grid to obtain distribution density feature vectors of the events in different regions; using the TF-IDF algorithm to extract features from the text descriptions of the disposal plans to obtain text feature vectors of the disposal plans; and concatenating the event type feature vectors, event frequency feature vectors, distribution density feature vectors, and text feature vectors to obtain a feature dataset.

[0010] Among them, the event type refers to the types of emergency events or sudden events. Different emergency events have different natures, scales, impacts, and disposal methods, so they need to be classified. The number of event types refers to the total number of emergency event types. This requires a comprehensive sorting and induction of various events according to the actual needs of emergency management. Usually, events can be classified from perspectives such as time (e.g., forest fires mostly occur in summer), space (e.g., floods mostly occur in river basins), and disaster types (e.g., geological disasters, accident disasters). The disposal plan refers to a series of measures and actions taken to respond to emergency events to control the situation, eliminate hidden dangers, and reduce losses. In the emergency plan, each emergency event will have a corresponding disposal plan. The disposal plan is usually described in text form, including specific arrangements for pre-event prevention, on-site response, in-event disposal, and post-event recovery. The content involves organizational command, team assembly, information aggregation, on-site investigation, force dispatching, personnel evacuation, material support, post-disaster reconstruction, summary evaluation, etc.

[0011] Furthermore, use the feature dataset to train an emergency event classification model based on the CART decision tree, including: using the feature dataset as the training set, selecting the optimal splitting feature using the Gini index, and generating a CART decision tree through recursion; among them, the feature dataset is used as the input of the CART decision tree, and the event type is used as the output of the CART decision tree; using the K-fold cross-validation method to divide the feature dataset into K mutually exclusive subsets, selecting one subset as the validation set each time, and the remaining K - 1 subsets as the training set; using the training set to train the CART decision tree and evaluating the classification performance of the CART decision tree on the validation set; repeating the K-fold cross-validation process K times to obtain an emergency event classification model.

[0012] Preferably, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is used to oversample the event types with a small number of samples to balance the quantity distribution of different types of events; the balanced feature dataset is randomly divided into a training set, a validation set, and a test set; the root node is initialized with all the samples in the training set and regarded as the node to be divided; for each node to be divided, the feature vectors of the samples it contains are extracted to obtain each feature dimension and its possible values; for each possible value of each feature dimension, it is used as the division threshold to attempt to divide the samples of the current node into two subsets above and below the threshold, and the Gini index after division is calculated; the feature dimension with the smallest Gini index and its corresponding threshold are selected as the optimal division basis, and the sample set of the current node is divided into two child nodes accordingly; the newly generated child nodes are added to the list of nodes to be divided; the above steps are repeatedly executed until all nodes meet the preset stop conditions; finally, the first-layer CART decision tree is obtained; the first-layer CART decision tree is used to predict each sample in the training set, and the residual between the predicted value and the true event type label is calculated; taking the feature vectors of the training set samples as the input and the residual as the new target value, a training sample set for residual learning is constructed; using the training sample set for residual learning, the above steps are repeatedly executed to obtain the second-layer CART decision tree; the above steps are repeatedly executed to generate one layer of CART decision tree in each iteration until the preset maximum number of layers is reached to obtain a multi-layer CART decision tree; for each sample in the validation set, all layers of CART decision trees are used for prediction to obtain the prediction probability vector of each layer of decision tree; the prediction probability vectors of each sample on each layer of decision tree are concatenated to generate a new Stacking feature vector; taking the Stacking feature vector as the input and the event type as the target value, a Logistic regression model is trained as the Stacking meta-model; using the trained multi-layer CART decision tree and the Stacking meta-model, the new emergency event data is classified and predicted: the feature vectors of the new event data are extracted; the feature vectors are input into each layer of CART decision tree to obtain the prediction probability vector of each layer; the prediction probability vectors of all layers are concatenated into a Stacking feature vector and input into the Stacking meta-model to obtain the final event type prediction label.

[0013] Further, an emergency command and dispatch plan is generated, including: adopting a knowledge representation method based on ontology to structurally represent and store the emergency plan in the form of ontology, and constructing an emergency plan knowledge base; according to the identified emergency event, retrieving the emergency plan knowledge base through the ontology reasoning mechanism to obtain multiple emergency plans corresponding to the emergency event; calculating the semantic similarity between the identified emergency event and the obtained multiple emergency plans, and selecting one or more emergency plans with a semantic similarity greater than a preset threshold as the matching result; extracting the differential features between the identified emergency event and the matching result, and adopting a heuristic search algorithm to generate a personalized emergency sub-plan for the identified emergency event; fusing the matched one or more emergency plans with the generated personalized emergency sub-plan to obtain the emergency command and dispatch plan.

[0014] Among them, the emergency plan knowledge base is a knowledge base system for storing emergency plans. Different from traditional relational databases, it uses methods of knowledge representation and knowledge organization, based on semantic web technology, to structurally represent and store concepts, relationships, rules, etc. in the emergency plan in the form of ontology. Ontology reasoning refers to the process of using the concept hierarchy, relationship network, constraint rules, etc. in the ontology knowledge base to deduce implicit knowledge, answer questions or solve problems. By describing the semantic associations of domain concepts, ontology endows knowledge with clear semantics, enabling the computer to perform logical reasoning based on semantics. In this application, when a new emergency event is identified, relevant plans need to be retrieved from the emergency plan knowledge base. This retrieval is not a simple keyword match but a semantic retrieval. Through ontology reasoning, those plans that are conceptually and semantically related to the input event can be found. For example, wing failures of winged aircraft mostly occur in the wings, etc. Common ontology reasoning mechanisms include: rule-based reasoning (using if-then form rules), description logic-based reasoning (using formal definitions of concepts and relationships), etc. The reasoning process may involve traversing the ontology hierarchy, searching the relationship network, matching the rule base, etc. The reasoning results can be in the form of identified concept instances, triple relationships, logical predicates, etc.

[0015] In this application, the differential feature refers to the differential features between the identified emergency event and the retrieved standard emergency plan. What is stored in the emergency plan knowledge base are some standardized and templated plans, but the actual emergency events that occur often have their particularities. For example, the time, location, magnitude, etc. of an earthquake, and the climate, terrain, wind direction, etc. of a wildfire will all be different from the standard plan.

[0016] Further, obtaining multiple emergency response plans corresponding to an emergency event includes: using an ontology construction method to represent the identified emergency event ontologically to obtain an event ontology; extracting key concepts in the event ontology to form an event query ontology; where the key concepts include event type, event time, and event location; according to the event query ontology, using an ontology reasoning mechanism based on description logic, retrieving from the constructed emergency response plan knowledge base the emergency response plan ontologies that contain all the key concepts in the event query ontology as a candidate emergency response plan ontology set; where the emergency response plan ontology includes plan type and disposal plan; for each emergency response plan ontology in the candidate emergency response plan ontology set, extracting the corresponding plan type and disposal plan to form a key concept set of the emergency response plan ontology; calculating the semantic similarity between the key concepts in the event query ontology and the key concept set of the emergency response plan ontology to obtain the similarity of each candidate emergency response plan ontology; taking the candidate emergency response plan ontologies with similarity greater than a preset threshold as the matching emergency response plan ontologies; obtaining the emergency response plans corresponding to the matching emergency response plan ontologies as the emergency response plans corresponding to the identified emergency event.

[0017] Further, obtaining the similarity of each candidate emergency response plan ontology includes: obtaining the key concepts in the event query ontology ; obtaining each key concept in the key concept set of the emergency response plan ontology ; obtaining the synset of the key concept from the WordNet knowledge base ; obtaining the synset of the key concept from the WordNet knowledge base ; calculating and 's semantic similarity : , where represents and 's synset of the closest common ancestor concept in WordNet; represents the shortest path length from the synset to the WordNet root node; calculating the maximum value of the semantic similarity between the key concept and all the key concepts in the key concept set of the emergency response plan ontology as the similarity component of the key concept , and the calculation formula is: , where O represents the key concept set of the emergency response plan ontology; calculating the arithmetic mean of the similarity components of all the key concepts in the event query ontology as the similarity between the event query ontology and the current candidate emergency response plan ontology: , where Represents the event query ontology; Represents the candidate emergency plan ontology; n is the number of key concepts in the event query ontology.

[0018] Furthermore, select one or more emergency plans with semantic similarity greater than a preset threshold as the matching results, including: obtaining the event type of the identified emergency event, converting the event type into an event type vector through One-hot encoding; performing word segmentation and part-of-speech tagging on the obtained multiple emergency plans, extracting nouns and verbs as keywords, and extracting the titles of the emergency plans; concatenating the extracted keywords and titles as the emergency plan vector; mapping each element in the event type vector to the corresponding word embedding vector of the event type, and obtaining the event type embedding vector by weighted averaging the word embedding vectors corresponding to non-zero elements; where the weight is the frequency of each event type appearing in the historical data; mapping each word in the emergency plan vector to the corresponding word embedding vector, and obtaining the emergency plan embedding vector by weighted averaging all the word embedding vectors; where the weight is the TF-IDF value of each word; calculating the similarity between the event type embedding vector and the emergency plan embedding vector using cosine similarity to obtain a similarity list; obtaining one or more emergency plans from the similarity list according to the preset threshold as the matching results.

[0019] Furthermore, adopt a heuristic search algorithm to generate a personalized emergency sub-plan for the identified emergency event, including: decomposing the attributes of the identified emergency event, extracting the event type, event time, and location of the event to form an event attribute vector; decomposing the attributes of the one or more matched emergency plans, extracting the event type, event time, and location of the plan to form a plan attribute vector; calculating the weighted Euclidean distance between the event attribute vector and each plan attribute vector as the difference degree to generate a difference attribute set; retrieving plan fragments from the emergency plan knowledge base with semantic similarity greater than the threshold to the difference attributes according to the difference attribute set to form a plan fragment library; where a plan fragment represents a step or measure in the emergency plan; using the event attribute vector in the difference attribute set as the initial state of the search and using the corresponding plan attribute vector in the difference attribute set as the target state of the search ; in the plan fragment library, search for the optimal path from the initial state to the target state as the personalized emergency sub-plan. Among them, a plan fragment refers to an independent operation step or disposal measure in the emergency plan and is the basic unit that constitutes a complete plan. An emergency plan usually consists of multiple plan fragments combined, and there is a certain logical relationship and execution order between the fragments.

[0020] Furthermore, search from the initial state to the target state The optimal path, as a personalized emergency sub - solution, includes: for each scenario segment in the scenario segment library , calculate the cost from the initial state to : , where is the actual cost from the initial state to ; is the estimated cost from to the target state ; Select the scenario segment with the minimum value, mark it as visited, and update to the corresponding scenario segment attribute vector , denoted as ; Use as the new initial state, repeat the above marking steps until reaching the target state or the preset termination condition; If reaching the target state , then take the sequence of scenario segments visited during the search process as the optimal path, extract the content of the scenario segments corresponding to the path, and combine them in order to form a personalized emergency sub - solution; If the preset termination condition is reached, the personalized emergency sub - solution is empty.

[0021] Furthermore, the actual cost is calculated through the following formula: , where and respectively represent the k - th component of the initial state event attribute vector and the attribute vector of the scenario segment ; is the weight of the k - th component, reflecting the importance of this component; is the value range of the k - th component, used for normalization processing; The estimated cost is calculated through the following formula: , where and respectively represent the k - th component of the scenario segment attribute vector and the target state scenario attribute vector ; is the weight of the k - th component.

[0022] Compared with the prior art, the advantages of this application are:

[0023] The ontology knowledge representation method is used to structurally represent the emergency plan, and an emergency plan knowledge base with rich semantics and standardized structure is constructed. By calculating the semantic similarity between the real-time emergency event and the emergency plans in the plan knowledge base, the most relevant emergency plan is automatically retrieved and matched, and a reference plan for event handling is given personalized.

[0024] By analyzing the differential features of the real-time event and the matched plan, combined with the heuristic search algorithm, targeted personalized plan fragments are searched from the knowledge base, and an event-driven personalized emergency sub-plan is automatically generated. This method makes full use of the plan fragments in the knowledge base, and generates a plan through the assembly of fragments driven by differential features, so that the generated scheduling plan can accurately match the personalized needs of real-time events.

[0025] When generating the personalized emergency sub-plan, by defining the cost function in the heuristic search algorithm, mechanisms such as the relevance weight of event attributes and the normalization of the attribute values of plan fragments are introduced, so that the search process can comprehensively weigh the matching degree between event features and plan features, search for the optimal combination of personalized plans, and improve the applicability and effectiveness of the emergency scheduling plan.

[0026] The emergency plan retrieved and matched in the knowledge base is integrated with the personalized sub-plan generated by reasoning to form a complete event-driven emergency command and dispatch plan. The integrated plan takes into account both the general plan in the plan knowledge base and the personalized plan generated specifically, and can meet both the general needs and specific needs of emergency event handling. Brief Description of the Drawings

[0027] This application will be further described in the form of exemplary embodiments, and these exemplary embodiments will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where:

[0028] Figure 1 is an exemplary flowchart of an event-driven emergency command and dispatch method according to some embodiments of this application;

[0029] Figure 2 is an exemplary flowchart of generating a feature data set according to some embodiments of this application;

[0030] Figure 3 is an exemplary flowchart of generating an emergency command and dispatch plan according to some embodiments of this application. Detailed Embodiments

[0031] The methods and systems provided in the embodiments of this application will be described in detail below with reference to the drawings.

[0032] As Figure 1As shown, obtain historical emergency event data, which includes event type, event time, event location, and disposal plan; use feature engineering to extract features from the historical emergency event data to obtain a feature dataset; among them, the feature dataset includes an event type feature vector, an event frequency feature vector, a distribution density feature vector, and a text feature vector; use the feature dataset to train an emergency event classification model based on the CART decision tree; obtain real-time emergency event data, and extract features from the real-time emergency event data to obtain a real-time event feature vector; use the obtained real-time event feature vector as input, and use the emergency event classification model to identify the emergency event; according to the identified emergency event, adopt a rule-based case reasoning method to retrieve and generate an emergency command and dispatch plan from a preset emergency plan knowledge base.

[0033] Specifically, obtain historical emergency event data from the publicly available data in the public database. The historical emergency event data includes event type, event time, event location, and disposal plan; for the selected data source, data collection tools such as crawlers can be used to automatically extract the required structured event information from unstructured or semi-structured data such as web pages or files. For example, for the accident notice news page published on the official website of the emergency management department, write targeted web parsing rules to extract keyword fields such as the time, location, accident type, process description, and disposal situation of the accident. For the existing structured accident database, historical event records can be obtained in batches directly through database query statements. Preprocess the obtained data, and import the preprocessed historical emergency event data into a relational database such as MySQL or PostgreSQL for centralized storage and management, and establish an emergency event historical database. The database table structure design should cover but is not limited to the following fields: Event ID: A serial number that uniquely identifies an emergency event. Event type: Such as fire, traffic accident, geological disaster, etc., which can be represented by numerical codes. Event time: The time of the incident, in the unified format of year-month-day hour:minute:second. Event location: A detailed description of the incident address, such as No. XX Road, XX District, XX City, XX Province. Disposal plan: A text field that records the specific disposal measures during the disposal process of the event.

[0034] Such as Figure 2As shown in the figure, feature engineering is used to extract features from historical emergency event data to obtain a feature dataset. Among them, the feature dataset includes an event type feature vector, an event frequency feature vector, a distribution density feature vector, and a text feature vector. Specifically, for event type feature extraction, for the event type attribute in the historical emergency event data, the One-Hot encoding method is used for feature extraction. First, enumerate the event types to obtain an event type set, such as {fire, traffic accident, geological disaster}. Then, determine the number of binary digits of the One-Hot encoding according to the size of the event type set, and the number of digits is the same as the number of types. For each emergency event, set the binary position corresponding to its event type to 1 and the rest to 0 to obtain the feature vector representation of the event type. For example, assuming the event type set is {fire, traffic accident, geological disaster}, the feature vector of the "fire" event is [1, 0, 0], and the feature vector of the "traffic accident" event is [0, 1, 0]. For event frequency feature extraction, for the event time attribute in the historical emergency event data, the sliding time window method is used for frequency feature extraction. First, set different time window scales according to application requirements, such as hours, days, weeks, months, etc. Then, use the sliding window method to count the number of events occurring in each time window to obtain the event frequencies at different time scales. For example, set a 24-hour sliding window to count the number of events within every 24 hours to obtain the event frequency vector at the hour scale, and then set a 7-day sliding window to count the number of events within every 7 days to obtain the event frequency vector at the day scale. By concatenating the frequency vectors at multiple time scales, the event frequency feature vector is obtained.

[0035] Specifically, for event distribution density feature extraction, input: historical emergency event dataset D, and each event data includes attributes such as event ID and event location (longitude and latitude coordinates). According to the given geographical range R of the target area and the grid granularity parameter g, divide the area R into m×n grid cells of equal size. Each grid cell is represented as G(i, j), where i is the row number and j is the column number, i ∈ [1, m], j ∈ [1, n]. Let the set of grid cells be G. For each historical event e in the dataset D, calculate the grid cell number (i, j) where the event is located according to its location coordinates (longitude lng, latitude lat). Add the event ID to the event list L(i, j) of the corresponding grid. After traversing all events, obtain the event list corresponding to each grid cell. For each grid cell G(i, j) in the grid set G, with its center coordinates as the core, use the kernel density estimation method to estimate the event density of this grid. Set the kernel function as K, and common kernel functions include Gaussian kernel, Epanechnikov kernel, etc. Here, the Gaussian kernel is taken as an example. The formula is: ; Set the bandwidth parameter of the kernel function as h. According to the empirical rule, it can be set , where σ is the standard deviation of event coordinates and n is the total number of events. Then, for each event e' in the grid , calculate its distance from the grid center : ; According to the distance value d and bandwidth h, calculate the contribution coefficient of event to the density estimation of the grid : ; Finally, sum the contribution coefficients of all events in the grid to obtain the event density estimation value of this grid: . Traverse all grid cells in the grid set G, sequentially obtain the event density estimation values of each grid, and construct the event distribution density feature vector : , where has a length of m×n, representing the total number of grid cells. depicts the spatial distribution pattern of emergency events in the entire target area, which helps to mine the regional characteristics and hot spots of events.

[0036] Specifically, extract the text feature vector of the disposal plan. Extract the text descriptions of the disposal plans of all events in the data set D to form the original corpus C. Perform text preprocessing on the corpus C; segment each text description according to semantic content words to obtain a word segmentation sequence. Chinese word segmentation tools such as Jieba and THULAC can be used. After preprocessing, the standardized corpus is obtained. According to the corpus , count the occurrence frequency of each word w in the corpus to obtain the term frequency TF(w): TF(w) = (the number of occurrences of w in the corpus) / (the total number of words in the corpus); count the number of texts containing the word w to obtain the inverse document frequency IDF(w): IDF(w) = log((the total number of corpus texts + α) / (the number of texts containing w + α)), where α is a smoothing parameter, usually taken as 1. Calculate the TF-IDF weight of each word w: ; Select the N words with the highest TF-IDF weights to form the feature dictionary W of the disposal plan text. For each event e in the data set D, extract its text description t(e) of the disposal plan, perform preprocessing such as word segmentation and stop word removal to obtain a standardized word sequence. Then, according to the feature dictionary W, count the TF-IDF weights of each feature word in the word sequence to construct the text feature vector : ; where is the i-th feature word in the dictionary W, n is the dictionary size N, is The TF-IDF weights in text t(e). Traverse all events in dataset D to obtain the text feature vectors of the disposal solutions for each event. The vector of the text features of the historical emergency event disposal solutions Characterizes the semantic information and key measures of the event disposal solutions, which helps to mine the similarities and differences in the disposal measures of different events.

[0037] Use the feature dataset to train an emergency event classification model based on the CART decision tree; specifically, divide the feature dataset D into a feature matrix X and a label vector y. Among them, each row of X represents the feature vector of an event, including event type features, frequency features, density features, and text features, etc.; each element of y represents the type label of the corresponding event, such as "fire", "traffic accident", etc. Use the CART (Classification and Regression Tree) algorithm to train a decision tree according to the feature matrix X and the label vector y. Calculate the Gini index Gini(D) of the current node, and select the optimal splitting feature a and the optimal splitting point v according to the feature matrix X to minimize the sum of the Gini indices of the sub-nodes after splitting. The calculation formula of the Gini index is: , ; where D is the sample set of the current node, is the sample subset of class k, and K is the total number of classes. According to the optimal splitting feature a and the splitting point v, divide the sample set D of the current node into the sample sets of the left and right sub-nodes and . Recursively divide the sub-nodes until one of the following stopping conditions is met: the sample set of the current node belongs to the same class; the number of samples in the current node is less than the preset threshold ; the depth of the current node reaches the preset maximum depth ; no better splitting feature can be found. For each leaf node, mark the node as the corresponding class according to the majority class label in its sample set. Use the K-fold cross-validation method to evaluate the generalization performance of the decision tree model. Divide the feature dataset D into K mutually exclusive subsets . Perform K-fold cross-validation, and each time select a subset as the validation set, and the remaining K - 1 subsets form the training set . Use the training set to train a decision tree model , and evaluate the classification performance of the model on the validation set . Commonly used evaluation metrics include accuracy, precision, recall, and F1 value, etc. Repeat K times to obtain K decision tree models And the corresponding performance evaluation results, take the average of the K performance evaluation results to obtain the comprehensive performance evaluation result of the decision tree model. According to the evaluation results of cross-validation, select the decision tree model with the optimal performance As the final emergency event classification model.

[0038] Obtain the latest emergency event report data. The data content includes the time, location, preliminary description, etc. of the event. Parse the obtained real-time event data and extract the key field information. According to the data formats of different data sources, use methods such as regular expression, XML parsing, JSON parsing, etc. to extract structured event attributes, such as event type, time, location, etc. from unstructured or semi-structured data. According to the keywords in the event description, use the pre-trained event type classification model to identify the possible types of the event and generate event type features. According to the event time, update the event frequency statistical features within the time window. According to the event location coordinates, locate the geographical grid where the event is located, update the event density estimation value of the grid, and generate distribution density features. Perform preprocessing such as word segmentation and stop word removal on the text description of the event, and use the pre-trained TF-IDF model to extract the text feature vector. Concatenate the feature vectors such as event type, frequency, density, and text to obtain the comprehensive feature vector of the real-time event, which is used as the input for subsequent event classification. Adopt a stream processing framework such as Apache SparkStreaming, Flink, etc. to organize steps such as accessing, parsing, and feature extraction of real-time event data into a real-time processing pipeline. S5, use the obtained real-time event feature vector as the input, and use the emergency event classification model to identify the emergency event.

[0039] Such as Figure 3As shown, according to the identified emergency events, a rule-based case reasoning method is adopted to retrieve and generate an emergency command and dispatch plan from a preset emergency plan knowledge base; the construction of the emergency plan ontology, collecting emergency plan texts from channels such as emergency management departments and public databases, the content including the type of the plan, applicable conditions, organizational structure, disposal process, etc. Perform natural language processing on the emergency plan text, adopt techniques such as named entity recognition and keyword extraction to identify the key concepts in the plan, such as "earthquake", "rescue", "evacuation", etc. Analyze the semantic relationships between the concepts, define the attributes and relationships of the ontology, such as "is-a", "part-of", etc. Construct a concept classification hierarchy and relationship network. Adopt an ontology description language such as OWL (Web Ontology Language) to represent concepts, attributes, relationships, etc. in the form of formal triples (subject-predicate-object). For example: (Earthquake is-a Natural Disaster), (Rescue part-of Emergency Disposal). Store the constructed emergency plan ontology in a graph database such as Neo4j in formats such as RDF (Resource Description Framework) to form an emergency plan knowledge base.

[0040] Emergency plan retrieval based on ontology, adopting the ontology construction method, formally representing the identified emergency events as event ontology E. Extract key attributes such as event type, occurrence time, occurrence location, etc. from the event description. According to an ontology language such as OWL, represent the event attributes as ontology concepts (Class) and relationships (Property). For example, represent "An earthquake of magnitude 5.8 occurred in a certain place on May 12, 2023" as: (E rdf:type Earthquake), (E has Occur Time "2023-05-12"), (E has Occur Location a certain place), (E has Magnitude “5.8”). Generate an event query ontology, extract key concepts such as event type (event Type), event time (event Time), event location (event Loc) from the event ontology E to form a query ontology Q. The query ontology Q is represented as: (Q rdf:type [event Type]), (Q hasOccur Time [event Time]), (Q has Occur Location [event Loc]), where [eventType], [event Time], [event Loc] are specific event attribute values.

[0041] Retrieve the candidate plan ontology. Using the ontology reasoning mechanism based on Description Logic (DL), retrieve the plan ontology that satisfies the constraints of the query ontology Q from the emergency plan knowledge base. DL ontology reasoning defines the semantic relationships between ontologies through preset axioms. Common axioms include: Sub Class Of (C1, C2): Concept C1 is a subclass of concept C2. EquivalentClasses (C1, C2): Concepts C1 and C2 are equivalent. Dis joint Classes (C1, C2): Concepts C1 and C2 are disjoint. Define the axiom: Sub Class Of (Earthquake Plan, Emergency Plan), that is, "earthquake plan" is a subclass of "emergency plan". Query the plan ontology P that satisfies Q: (P rdf:type [event Type]), (P has Refer Loc [eventLoc]). Find all eligible Ps through semantic relationships such as subclasses and equivalent classes as the candidate plan ontology set {P1, P2,...}. Extract the key concepts of the candidate plans. For each plan ontology Pi in the candidate plan set, extract the key concepts such as the plan type (plan Type) and the disposal plan (plan Scheme).

[0042] Example of the plan type concept: (Pi rdf:type Building Collapse), indicating that the plan Pi is applicable to building collapse events. Example of the disposal plan concept: (Pi has Step Evacuation), (Pi has Step RescueTrapped), indicating that the plan Pi includes the steps of "personnel evacuation" and "rescuing trapped personnel". The extracted key concepts form a concept set {Ci1, Ci2,...}. Calculate the event-plan similarity. Calculate the semantic similarity between each key concept in the query ontology Q and the concept set of the candidate plan Pi. For each concept q in Q: Obtain the synonym set synset(q); for each concept c in Pi: Obtain the synonym set synset(c); calculate the similarity between synset(q) and synset(c): , where is the most recent common ancestor concept of q and c on WordNet; is the distance from q to the WordNet root node; Calculate the maximum similarity between q and all concepts ; Calculate the average value of the similarities of all concepts in Q as the similarity between Q and : . Match the emergency plan. Set the similarity threshold , and select from the candidate plan set those that satisfy pre - plan , extract The corresponding pre - plan content (i.e., non - ontology form) is used as the final retrieval result. The events and pre - plans are uniformly represented as ontologies, and DL reasoning is used to quickly narrow the search scope. The semantic similarity between the query concept and the pre - plan concept is calculated on the candidate pre - plan set, breaking through the limitation of keyword matching and achieving semantic - level similarity judgment.

[0043] Calculate the semantic similarity between the identified emergency events and the obtained multiple emergency pre - plans, and select one or more emergency pre - plans with semantic similarity greater than the preset threshold as the matching results, including: Represent the event type in a vectorized form, define the set of emergency event types, such as {fire, earthquake, rainstorm, explosion,...}, a total of N types; Use One - hot encoding to convert the event type into an N - dimensional binary vector. For example, if the identified event type is "earthquake", the corresponding vector is [0, 1, 0, 0,...]. Tokenize and perform part - of - speech tagging on each candidate pre - plan text. Extract the words with noun and verb parts of speech as keywords. Extract the title of the pre - plan (if any). Concatenate the keywords and the title into a string to represent the pre - plan text.

[0044] Generate the event - type embedding vector, load the pre - trained word - embedding model, such as Word2Vec, GloVe, etc. Count the occurrence frequency of each event type in the historical emergency event data . For each non - zero element of the event - type vector: Obtain the word vector of this event type through the word - embedding model of the word vector . Calculate the weighted embedding vector of: ; Take the average of the weighted embedding vectors of all non - zero elements to obtain the event - type embedding vector: . Generate the emergency - pre - plan embedding vector for the keyword list of each pre - plan text : Obtain the word vector of each keyword through the word - embedding model of the word vector . Calculate the TF - IDF weight of each keyword , as the weighting coefficient. Take the weighted average of all word vectors to obtain the embedding vector of the pre - plan text: .

[0045] Obtain the final matching result. For each pre - plan embedding vector : Calculate its cosine similarity with the event - type embedding vector : ; Save the similarity result to the similarity list. Output the similarity list . Set the similarity threshold Select all the pre - plans that meet from the similarity list as the final matching result. If the similarity of no pre - plan exceeds the threshold, the threshold can be appropriately reduced, or the top - K pre - plans with the highest similarity can be selected. By embedding both the event type and the pre - plan text into the same semantic space, the similarity degree between them in this space can be calculated, overcoming the limitations of keyword matching. When embedding events, the prior distribution of different event types is considered, and higher weights are assigned to high - frequency event types to adapt to the long - tail distribution of the data. When embedding pre - plans, TF - IDF weighting is adopted to highlight the key information in the pre - plan text.

[0046] Extract the differential features between the identified emergency event and the matching result, and use a heuristic search algorithm to generate a personalized emergency sub - plan for the identified emergency event. First, extract the event attributes, decompose the attributes of the input emergency event, and extract the following key attributes: Event type: such as earthquake, fire, etc.; Event time: the time when the event occurs; Event location: the geographical location where the event occurs; Map the extracted attribute values into numerical vectors to form an event attribute vector Extract the pre - plan attributes, decompose the attributes of each input matching pre - plan text, and extract the following attributes: Applicable event type of the pre - plan: such as earthquake pre - plan, fire pre - plan, etc.; Applicable time of the pre - plan: such as weekdays, nights, etc.; Applicable location of the pre - plan: such as schools, communities, etc.; Map the extracted attribute values into numerical vectors to form a pre - plan attribute vector .

[0047] Calculate the attribute difference. For each matching pre - plan, calculate the weighted Euclidean distance between its pre - plan attribute vector and the event attribute vector : , where is the weight of the k - th attribute, and is the value range of the k - th attribute. Save the pre - plan attributes with significant differences (such as diff greater than a certain threshold) and their difference values as differential attributes into the differential attribute set.

[0048] Retrieval plan fragment, calculated based on the differences between event attributes and plan attributes, representing the mismatches between the current event and existing plans, such as {"event type": "mountain flood", "event time": "early morning", "event location": "village"}; Emergency plan knowledge base: contains a large number of plan texts, segmented into several plan fragments according to granularity such as chapters and paragraphs. Therefore, expand the differential attributes into a richer semantic representation to improve the recall rate of retrieval. Use pre-trained word embedding models, such as Word2Vec, GloVe, etc.; for each attribute value in the differential attribute set, find semantically similar words in the word embedding space; select the top K words with the highest similarity as the extended words for this attribute value; combine the original attribute value and the extended words to form the extended differential attributes. For example, the original differential attribute: "event type": "mountain flood"; the extended differential attribute: "event type": ["mountain flood", "flood", "mudslide", "flood disaster"]. Use the bag-of-words model or document vectorization methods, such as TF-IDF, Doc2Vec, etc., to perform preprocessing such as word segmentation and stop word removal on each plan fragment in the plan knowledge base; based on the bag-of-words model, count the word frequencies in each fragment to generate a word frequency vector; or directly map the fragment to a dense vector using the document vectorization method; normalize the fragment vector, such as L2 normalization. For example, the plan fragment: "When a mountain flood occurs, the warning system should be activated immediately, and residents should be notified to evacuate to a safe area." Fragment vector: [0.2, 0.0, 0.1,......, 0.0, 0.3].

[0049] Regard the extended differential attributes as a "virtual document", and adopt the same vectorization method as the plan fragment. For each differential attribute, concatenate its attribute name and attribute value (including extended words) into a text string, perform preprocessing such as word segmentation and stop word removal on the text string, and map the text string to a vector based on the bag-of-words model or document vectorization method, and normalize the attribute vector. For example, the differential attribute: "event type": ["mountain flood", "flood", "mudslide", "flood disaster",...]; the attribute text string: "event type mountain flood flood mudslide flood disaster"; the attribute vector: [0.1, 0.0, 0.2,......., 0.0, 0.4].

[0050] Calculate the similarity between each plan fragment and the differential attributes, and select the fragments with high similarity. Use similarity measurement methods such as cosine similarity. For each plan fragment vector and the differential attribute vector , calculate their cosine similarity: ; traverse all plan fragments, select the fragments whose similarity to any differential attribute is greater than the preset threshold, and add them to the alternative set; sort the alternative fragments in descending order according to the similarity scores. For example, the plan fragment vector : [0.2, 0.0, 0.1,......, 0.0, 0.3]; Differential attribute vector : [0.1, 0.0, 0.2,......, 0.0, 0.4]; Cosine similarity: ; If the set threshold is 0.6, then is selected into the alternative set to form the alternative plan fragment library.

[0051] Generate personalized emergency sub - plans based on the search algorithm. Mark the event attribute vector as the initial state of the search, and the plan attribute vector as the target state; Define the open list (Open List) and closed list (Closed List) for the search. The open list stores the plan fragments to be visited, which is empty initially. The closed list stores the plan fragments that have been visited, which initially only contains . For each plan fragment , record its parent fragment , that is, the predecessor fragment in the search process, : The actual cost from the initial state to , initially : The estimated cost from to the target state , initially , , that is, the estimated total cost from via to .

[0052] Calculate the actual cost . For the plan fragment and the attribute vectors of its parent fragment and , calculate the difference degree between them, , , is the weight of the k - th attribute component, reflecting the importance of this component, is the value range of the k - th component for normalization processing. Calculate the estimated cost . For the attribute vector of the plan fragment and the target state , calculate the similarity between them, , is the weight of the k - th attribute component.

[0053] Search process: Select the scenario fragment p with the smallest f(p) from the open list, mark it as the current fragment, remove the current fragment p from the open list, and add it to the closed list. If the attribute vector of the current fragment p has a difference degree less than the preset threshold with the target state , the search is successful; otherwise, continue the search. Generate the successor fragments of the current fragment p, that is, select several fragments semantically related to p in the scenario fragment library . The successor fragments can be selected by methods such as keyword matching and semantic similarity; for each successor fragment : If is already in the closed list, it means it has been visited before, skip it; if is not in the open list, add it to the open list, set its parent fragment as p, and calculate , and ; if is already in the open list, compare its original g value with the newly calculated g value. If the new g value is smaller, update 's parent fragment to p, and recalculate and ; if the open list is empty, the search fails.

[0054] Generate a personalized emergency sub - plan. If the search is successful: Start from the target fragment , trace back the search path along the parent pointer until the initial state ; Extract the content of the scenario fragments on the traced - back path in reverse order and combine them into a complete personalized emergency sub - plan. Output the emergency sub - plan. If the search fails: The personalized emergency sub - plan is empty, and an error message indicating the search failure is output. In this application, by defining the actual cost and estimated cost between scenario fragments and using the f(p) value to guide the search direction, the optimal path from the event attributes to the target scenario attributes is found in the scenario fragment library, and the fragment content on the path is extracted to form a personalized plan.

[0055] Fuse one or more matched emergency plans with the generated personalized emergency sub - plan to obtain an emergency command and dispatch plan; specifically, for each matched emergency plan, extract its key information and convert it into a structured plan object, including: metadata such as plan name, version, applicable conditions, etc.; the organizational structure of the plan, including hierarchical information such as chapters and paragraphs; the key attributes of each chapter or paragraph, such as risk sources, disposal measures, resource requirements, etc.; the key entities and their relationships in the plan, such as events, departments, personnel, places, etc.; for the personalized emergency sub - plan, also extract its key information, represented as one or more plan fragments; each plan fragment includes information such as its description content, key attributes, and front - back relationships.

[0056] Traverse each chapter and paragraph of the emergency plan and compare them with the fragments of the personalized sub-plan. The basis for comparison includes the similarity of key attributes, the co-occurrence of entities, etc. A similarity threshold can be set to establish connections between the plan paragraphs and the plan fragments that exceed the threshold. Through comparison, find the corresponding relationship between the personalized sub-plan and the plan content, identify the personalized disposal measures already included in the plan, and identify the supplements or amendments to the standard plan in the personalized sub-plan.

[0057] Based on a certain matching plan, generate an initial command and dispatch plan, and traverse each fragment in the personalized emergency sub-plan: if the fragment highly matches the attributes of a certain paragraph in the plan, replace the original paragraph with the personalized fragment; if the fragment is a supplement to the plan, insert it into the appropriate position in the plan according to its context; if the fragment conflicts with the plan (such as incompatible disposal measures), issue a warning and let the emergency experts decide how to handle it, and generate an emergency command and dispatch plan that includes all the new and modified fragments. This incremental integration method can maximize the retention of the complete structure of the standard plan, while supplementing and optimizing the personalized disposal measures in a targeted manner, and generating a comprehensive, dynamic, and systematic command and dispatch plan.

[0058] Preferably, conduct a global consistency and feasibility check on the generated command and dispatch plan, mainly including: resource requirement conflicts, such as the demand for the same type of emergency resource by different disposal measures exceeding the existing supply capacity; timing logic conflicts, such as inconsistent or circular dependencies in the required action sequences of different fragments; spatial location conflicts, such as position conflicts in the emergency sites, evacuation routes, etc. involved in different fragments. Automatically or semi-automatically identify potential conflicts, prompt the emergency commanders to make decisions, and further modify and optimize the plan through human-computer interaction until all serious conflicts are eliminated. For inevitable situations such as resource contention, provide multiple decision alternatives and give risk assessments.

[0059] Extract the key information of the finally generated emergency command and dispatch plan, convert it into a friendly visual interface, and use graphical methods such as Gantt charts and flowcharts to display the time progress and task dependencies of the plan. Overlay spatial information such as emergency resources, risk hazards, and disposal actions on the map, provide a text view of the plan content, allow viewing and exporting the complete dispatch plan document, and help the emergency commanders quickly understand and execute the plan and monitor the execution progress of the plan in real time through visualization means.

Claims

1. An event-driven emergency command and dispatch method, characterized in that: include: Obtain historical emergency event data, which includes event type, event time, event location and disposal plan; Feature engineering is used to extract features from historical emergency event data to obtain a feature data set; wherein the feature data set includes event type feature vectors, event frequency feature vectors, distribution density feature vectors, and text feature vectors; Use the feature data set to train the emergency event classification model based on the CART decision tree; Acquire real-time emergency event data, and perform feature extraction on the real-time emergency event data to obtain a real-time event feature vector; The obtained real-time event feature vector is used as input to identify emergency events using the emergency event classification model; According to the identified emergency events, the rule-based case reasoning method is used to retrieve and generate emergency command and dispatch plans from the preset emergency plan knowledge base; Generate emergency command and dispatch plan, including: adopting ontology-based knowledge representation method, structured representation and storage of emergency plan in the form of ontology, and building emergency plan knowledge base; According to the identified emergency events, the emergency plan knowledge base is retrieved through the ontology reasoning mechanism to obtain multiple emergency plans corresponding to the emergency events, including: calculating the semantic similarity between the key concepts in the event query ontology and the key concept set of the emergency plan ontology, and obtaining the similarity of each candidate emergency plan ontology; Calculate the semantic similarity between the identified emergency event and the acquired multiple emergency plans, and select one or more emergency plans with semantic similarity greater than a preset threshold as the matching result; Extract the difference features between the identified emergency events and the matching results, and use a heuristic search algorithm to generate a personalized emergency sub-plan for the identified emergency events, including: decomposing the attributes of the identified emergency events, extracting the event type, event time and location of the event, and forming an event attribute vector; decomposing the attributes of one or more matched emergency plans, extracting the event type, event time and location of the plan, and forming a plan attribute vector; calculating the weighted Euclidean distance between the event attribute vector and each plan attribute vector as the difference degree, and generating a difference attribute set; according to the difference attribute set, retrieving plan fragments whose semantic similarity with the difference attribute is greater than a threshold from the emergency plan knowledge base to form a plan fragment library; wherein the plan fragment represents a step or measure in the emergency plan; taking the event attribute vector in the difference attribute set as the initial state of the search , take the corresponding plan attribute vector in the difference attribute set as the target state of the search ; In the plan fragment library, search from the initial state To the target state The optimal path is used as a personalized emergency sub-plan; The matched one or more emergency plans are integrated with the generated personalized emergency sub-plans to obtain an emergency command and dispatch plan.

2. The event-driven emergency command and dispatch method according to claim 1 is characterized in that: Feature engineering is used to extract features from historical emergency event data, including: One-hot encoding with the same number of binary bits as the event type is used to encode the event type and obtain the event type feature vector; The sliding time window method is used to perform time window statistics on event time to obtain event frequency feature vectors at different time scales; The event locations are divided into geographic grids, and the number of events in each grid is estimated using the kernel density estimation method to obtain the distribution density feature vectors of events in different regions. The TF-IDF algorithm is used to extract features from the text description of the treatment plan to obtain the text feature vector of the treatment plan; The event type feature vector, event frequency feature vector, distribution density feature vector and text feature vector are concatenated through vectors to obtain a feature data set.

3. The event-driven emergency command and dispatch method according to claim 2 is characterized in that: The feature data set is used to train the emergency event classification model based on the CART decision tree, including: The feature data set is used as the training set, the Gini index is used to select the optimal partitioning features, and a CART decision tree is generated recursively; the feature data set is used as the input of the CART decision tree, and the event type is used as the output of the CART decision tree; The K-fold cross-validation method is used to divide the feature data set into K mutually exclusive subsets. One subset is selected as the validation set each time, and the remaining K-1 subsets are used as training sets. The CART decision tree is trained using the training set, and the classification performance of the CART decision tree is evaluated on the validation set. The cross-validation process is repeated K times to obtain the emergency event classification model.

4. The event-driven emergency command and dispatch method according to claim 1 is characterized in that: Obtain multiple emergency response plans corresponding to emergency events, including: The ontology construction method is used to represent the identified emergency events and obtain the event ontology; Extract key concepts from the event ontology to form an event query ontology; the key concepts include event type, event time and event location; According to the event query ontology, the ontology reasoning mechanism based on description logic is adopted to retrieve the emergency plan ontology containing all the key concepts in the event query ontology from the constructed emergency plan knowledge base as the candidate emergency plan ontology set; wherein, the emergency plan ontology contains the plan type and disposal plan; For each emergency plan ontology in the candidate emergency plan ontology set, the corresponding plan type and disposal plan are extracted to form a key concept set of the emergency plan ontology; The candidate emergency plan ontology with a similarity greater than a preset threshold is taken as the matched emergency plan ontology; The emergency plan corresponding to the matched emergency plan ontology is obtained as the emergency plan corresponding to the identified emergency event.

5. The event-driven emergency command and dispatch method according to claim 4 is characterized in that: The similarity of each candidate emergency plan ontology is obtained, including: Get the key concepts in the event query ontology ; Get each key concept in the key concept set of the emergency plan ontology ; Get key concepts from the WordNet knowledge base Synonyms of ; Get key concepts from the WordNet knowledge base Synonyms of ; calculate and The semantic similarity of : ； in, express and The synset of the lowest common ancestor concept in WordNet; Indicates the shortest path length from the synonym set synset to the WordNet root node; Computing Key Concepts All key concepts in the key concept set of the emergency plan ontology The maximum semantic similarity of , as a key concept The similarity component is calculated as: ； Among them, O represents the key concept set of the emergency plan ontology; Compute all key concepts in the event query ontology The arithmetic mean of the similarity components is taken as the similarity between the event query ontology and the current candidate emergency plan ontology: ； in, Represents the event query ontology; represents the candidate emergency plan ontology; n is the number of key concepts in the event query ontology.

6. The event-driven emergency command and dispatch method according to claim 1 is characterized in that: One or more emergency plans with semantic similarity greater than a preset threshold are selected as matching results, including: Obtain the event type of the identified emergency event and convert the event type into an event type vector through One-hot encoding; Perform word segmentation and part-of-speech tagging on the multiple emergency plans obtained, extract nouns and verbs as keywords, and extract the titles of the emergency plans; The extracted keywords and titles are concatenated as emergency plan vectors; Map each element in the event type vector to the word embedding vector of the corresponding event type, and obtain the event type embedding vector by taking the weighted average of the word embedding vectors corresponding to the non-zero elements; the weight is the frequency of each event type in the historical data; Map each word in the emergency plan vector to a corresponding word embedding vector, and obtain the emergency plan embedding vector by weighted averaging all word embedding vectors; where the weight is the TF-IDF value of each word; The cosine similarity is used to calculate the similarity between the event type embedding vector and the emergency plan embedding vector to obtain a similarity list; According to a preset threshold, one or more emergency plans are obtained from the similarity list as matching results.

7. The event-driven emergency command and dispatch method according to claim 1 is characterized in that: Search from the initial state To the target state The optimal path, as a personalized emergency sub-plan, includes: For each plan segment in the plan segment library , calculate from the initial state arrive The cost : ； in, From the initial state arrive the actual cost of For To the target state the estimated cost of Select the one with the minimum Value Scenario Snippet , mark it as visited, and Updated to The corresponding plan segment attribute vector , denoted as ; by is the new initial state, repeat the above marking steps until the target state is reached or preset termination conditions; If the target state is reached , then the sequence of plan fragments accessed during the search process As the optimal path, extract the plan fragment content corresponding to the corresponding path, and combine them in order to form a personalized emergency sub-plan; If the preset termination condition is reached, the personalized emergency sub-plan is empty.

8. The event-driven emergency command and dispatch method according to claim 7 is characterized in that: Actual cost , calculated by the following formula: ； in, and Represents the initial state event attribute vector And the plan fragment The attribute vector The kth component of ; is the weight of the kth component, reflecting the importance of the component; is the value range of the kth component, used for normalization processing; Estimated cost , calculated by the following formula: ； in, and Represent the attribute vectors of the plan fragments and the target state plan attribute vector The kth component of ; is the weight of the kth component.

Images