Intelligent analysis and decision support method for emergency scenarios based on AI big model

Through intelligent emergency scenario analysis and decision-making support methods based on AI large model, the dependence on high-quality data and real-time data in emergency management is solved, the adaptability and generalization capabilities of the model are improved, and efficient emergency response decision-making is achieved Support has improved the scientificity and accuracy of emergency management.

CN118378912BActive Publication Date: 2024-11-15SHENZHEN GENEW INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410529077.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-29
Publication Date
2024-11-15
Estimated Expiration
2044-04-29

AI Technical Summary

Technical Problem

When handling complex and dynamically changing emergency scenarios, existing emergency management technologies have problems such as dependence on high-quality data, complexity of scenario-aware framework construction, insufficient generalization capabilities of model, real-time data and processing delays, resulting in inflexible decision-making responses. and accurate.

Method used

Adopting intelligent analysis and decision-making support methods based on AI large model, combining with Bayer , Deep Belief Network, Recurrent Neural Network, Stream Computing Architecture and Reinforcement Learning Model, reduce dependence on high-quality data, and improve model adaptability and real-timeness.

Benefits of technology

It significantly enhances the scientificity and accuracy of emergency management, improves the adaptability and generalization ability to complex scenarios, ensures real-time and accuracy of decisions, reduces processing delays, and improves the flexibility and efficiency of emergency response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118378912B_ABST
    Figure CN118378912B_ABST
Patent Text Reader

Abstract

The present invention proposes an intelligent analysis and decision support method for emergency scenarios based on an AI big model, which includes six parts: multi-source data integration and preprocessing, dynamic scenario modeling, decision perspective and behavior analysis, real-time data processing and feedback, adaptive decision support, and knowledge-driven learning and optimization. This method integrates the Bayesian network and the deep belief network to preprocess multi-source data to improve the credibility and consistency of information; through dynamic scenario modeling, combined with adaptive deep learning networks and graph neural networks, the adaptability and generalization ability of the model are improved; the real-time data processing architecture ensures low latency in data processing, so that the system can respond quickly to emergencies; the adaptive decision support mechanism is combined with reinforcement learning to improve the flexibility and accuracy of decision-making. The present invention provides scientific, rigorous and efficient support for emergency management, significantly enhances the ability to respond to emergencies, and ensures the real-time and accuracy of decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data recognition and analysis, and more specifically, to an emergency scenario intelligent analysis and decision support method based on an AI big model. Background Art

[0002] In the current global environment, the need for emergency management is becoming more and more prominent. Emergency scenarios, such as fires, earthquakes, and floods, are usually characterized by high suddenness, unpredictability, urgency, and complexity. These characteristics require that relevant decision support systems must be able to quickly and accurately identify and evaluate events and formulate effective response strategies. Traditional emergency decision-making methods mainly rely on rule-driven models and manual experience judgment, which often show limitations when dealing with complex and dynamically changing emergency scenarios.

[0003] With the advancement of technology, especially the development of information technology and artificial intelligence, the field of emergency management has begun to introduce more advanced technologies to improve response efficiency and decision-making quality. In traditional emergency management methods, scenarios such as scenario construction, scenario matching, and scenario response require a large amount of data input and real-time processing capabilities. In this context, artificial intelligence technology, especially AI-based big models, has become a key tool for improving emergency management capabilities due to its excellent data processing and pattern recognition capabilities. Big models can not only process and analyze large data sets, but also improve the accuracy of predictions by learning historical data, thereby providing more accurate scenario analysis in emergency scenarios. AI big models show unique advantages in simulating complex emergency situations. These models can achieve dynamic learning and adaptation in a changing environment, improving the real-time and sensitivity of decision-making. For example, in fire emergency management, AI models can analyze the possible paths and speeds of fire spread in real time, helping decision makers to make quick decisions on evacuation or fire extinguishing. This intelligent analysis based on big data and algorithms not only enhances the speed and effectiveness of emergency response, but also significantly reduces casualties and property losses.

[0004] Therefore, applying AI big models to emergency scenario analysis and decision support is not only an inevitable trend of technological progress, but also a key step to improve the efficiency and effectiveness of emergency management. By utilizing these advanced technologies, we can better build a situational awareness framework and optimize the co-coupling of situational elements, so as to make more scientific, rapid and accurate responses to the increasing number of emergency situations in the future.

[0005] Although the existing technology also applies learning models to emergency scenario analysis or decision-making, such as the Chinese invention patent with application number 202111053090.7, which proposes a fire scene data acquisition method, medium and electronic equipment, this technical solution provides a method that integrates machine learning models and sensor networks for fire scene data acquisition, with the aim of acquiring and processing environmental information near the fire source of underground buildings in real time and accurately, so as to obtain key fire parameters and scene data. Although this solution has significant advantages, such as the ability to provide fast, real-time data processing and prediction, it also has some technical defects or limitations, such as reliance on high-quality data. The system is highly dependent on the quality and accuracy of environmental information obtained from the front-end perception module. Any sensor failure or data quality issues, such as noise, data loss or corruption, may directly affect the accuracy and reliability of the machine learning model; the generalization ability of the machine learning model. Although machine learning models can provide accurate predictions, their performance is usually limited by the scope and quality of the training data. If they encounter new or extreme situations that are significantly different from the training data, the model's predictions may be inaccurate; real-time and processing delays. Although the goal is to achieve real-time data updates and processing, in actual applications, each step from data collection, transmission to processing may cause delays. This delay may affect the real-time response of the system to fire, especially in rapidly developing fire scenarios.

[0006] For example, the Chinese invention patent with application number 202310998825.6 proposes a fire simulation evacuation drill method based on a three-dimensional model. This technical solution provides a more advanced simulation method. By integrating scanning data, escape routes, fire-fighting facilities, network topology diagrams and other multi-dimensional data, a three-dimensional model is constructed and the fire situation and personnel parameters are simulated on this basis. Although this method is innovative, it also has some potential technical defects and challenges: data integration and synchronization issues. When integrating multiple data such as building scanning data, escape routes, firefighting facility data, and network topology diagrams into a three-dimensional model, ensuring the real-time and accuracy of the data is a challenge. Data out of sync or outdated data may lead to inaccurate simulation results and affect the effectiveness of the evacuation strategy; model generalization and adaptability. Although this method provides the ability to flexibly configure fire parameters and personnel parameters, the model's generalization ability and adaptability to new and unseen scenarios may not be sufficient, which may lead to a large deviation between the simulation effect and the actual situation when encountering atypical or extreme fire situations; flexibility and emergency response. Although the model can dynamically adjust the escape route and fire extinguishing strategy, in actual applications, how to respond quickly and update the model in real time to adapt to sudden changes is a technical challenge. In real emergency situations, the response speed and accuracy of the system are key.

[0007] Another example is the Chinese invention patent with application number 202311124872.4, which proposes a quantitative risk assessment method and system based on real-time monitoring of multiple disasters in the park. The patented technical means include detection modules and storage modules, which monitor temperature, gas concentration and fire information in real time, and use FMEA methods, bow tie diagrams and Bayesian networks to model accident scenarios, update risks and calculate consequence analysis. The beneficial effect of the patent is that it solves the problem of inadequate and inefficient manual inspections, and improves the accuracy and efficiency of safety inspections in chemical parks. However, although the patent adopts multiple models such as FMEA method, bowtie diagram and Bayesian network, the system is highly dependent on the accuracy and completeness of real-time monitoring data. If there are errors or data loss in the monitoring equipment, it will directly affect the reliability of risk assessment; model generalization and adaptability. The model mentioned in the patent may require parameter adjustment and optimization for specific parks, and may lack sufficient adaptability and predictive capabilities when facing unknown or unconventional disaster modes; real-time processing and computing efficiency. With the increase in monitoring points and data volume, the real-time processing capability and computing cost of the algorithm may become bottlenecks, especially in the case of high load or data peaks. There are challenges in maintaining system response speed and processing efficiency.

[0008] In summary, although modern information technology and artificial intelligence provide new tools for the analysis and decision-making of emergency scenarios, existing methods are still insufficient. These deficiencies include dependence on high-quality data, the complexity of building a scenario-aware framework, insufficient model generalization capabilities, and challenges in data real-time and processing delays. In addition, flexibility and emergency decision-making response capabilities in emergency situations also need to be improved. In view of this, it is particularly urgent to develop emergency scenario intelligent analysis technology based on AI large models to improve data processing capabilities and decision-making quality, optimize the co-coupling of scenario elements, and enhance the adaptability and real-time performance of the model, thereby effectively improving the scientificity and accuracy of emergency management. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide an emergency scenario intelligent analysis and decision support method based on an AI big model in response to the above-mentioned defects of the prior art.

[0010] The technical solution adopted by the present invention to solve its technical problem is:

[0011] The emergency scenario intelligent analysis and decision support method based on AI big model includes the following steps:

[0012] Multi-source data integration and preprocessing: Receive input from multiple data sources through a data integration module, the data sources include several sensor collection data, historical databases and real-time communication networks; use Bayesian network combined with deep belief network to perform probability fusion on the input data sources and evaluate the credibility of data points. The probability fusion is performed by weighting the trust scores of different data sources. The credibility of each data point is determined by the error rate and timestamp of its source; the data source data is denoised, formatted and standardized through the preprocessing unit, and data cleaning and denoising are performed to reduce the dependence on the quality of the original data. The formula used for preprocessing the data set of the data source is:

[0013] D′=f(D)where f(·)denotes the standardization function

[0014] Wherein, D is the original data set of the data source; D′ is the data set after standardization; f(·) is the standardization function, and the formula is used to convert the original data into a standard or standardized format;

[0015] Dynamic scenario modeling: Using a deep learning network with adaptive structural adjustment capabilities, the emergency scenario model is dynamically constructed and optimized based on the real-time and historical data of the data source after standardization. The deep learning network adopts a recursive neural network RNN ​​architecture composed of three layers of long short-term memory LSTM networks. Each layer of the long short-term memory LSTM automatically adjusts its network weights to optimize the prediction accuracy of emergency scenarios by continuously learning the time dependency in the data source. The adjustment of the network weights is implemented using the back propagation algorithm to ensure that the model reflects the real-time data changes more accurately after multiple iterations. The recursive neural network RNN ​​architecture dynamically adjusts the network weights through the gradient descent method. The update formula of the network weights is:

[0016]

[0017] Among them, θ t represents the network weight, α is the learning rate, and L is the loss function, which is used to evaluate the deviation between the model prediction results and the actual situation;

[0018] Decision-making perspective and behavior analysis: The decision analysis module evaluates and simulates the behavior of key decision makers and their impact on emergency response, and uses support vector machines (SVMs) to classify and predict key actors to provide a basis for real-time adjustment of decision support strategies. The behavior patterns of key actors are classified into two categories using support vector machines to distinguish different emergency response results that may be caused by their behaviors. The classification of key actors is solved by optimizing the problem using the following formula:

[0019]

[0020] Wherein, w is a weight vector, which determines the direction of the best separation hyperplane in the support vector machine SVM;

[0021] b is a bias term, which together with the weight vector w determines the position of the hyperplane;

[0022] ||w|| 2 is the square norm of the weight vector, the inverse of which is a measure of the size of the interval of the hyperplane. Minimizing this norm is to maximize the interval;

[0023] are coefficients used for mathematical processing, especially for ||w|| 2 Derivation;

[0024] C is the regularization parameter, which determines the degree of penalty for data points that violate the interval rule. The larger the C value, the greater the penalty for data points that violate the interval.

[0025] ξ i is a slack variable used to handle incompletely separable situations and allow data points to violate the interval rule;

[0026] is the sum of all slack variables;

[0027] Real-time data processing and feedback: Implement a data processing architecture based on stream computing, use a stream processing framework to collect, process and distribute data, ensure the real-time nature of data and reduce processing delays, generate emergency scenario analysis results and provide feedback to users. The emergency scenario analysis results include key decision points and potential risk assessments, and are combined with real-time analysis results of emergency scenarios to feed back to the decision model for dynamic updates;

[0028] Adaptive decision support: Configure an adaptive decision support system, which integrates decision optimization technology based on the decision model and reinforcement learning model, dynamically uses model pruning, quantization and knowledge distillation model compression technology to adjust the decision strategy according to real-time feedback and historical data, and finally outputs specific emergency response decision suggestions through the decision support system. The emergency response decision suggestions include emergency measures, resource allocation and response priority to guide actual emergency operations. The reinforcement learning model is learned and optimized through the following reward function:

[0029]

[0030] Among them, R(s, a) is the reward obtained by taking action a in state s;

[0031] r 0 The immediate reward is the immediate reward obtained by taking action a in state s;

[0032] γ is the discount factor, which ranges from 0 to 1. It determines the importance of future returns relative to immediate returns. The closer γ is to 1, the greater the importance of future returns.

[0033] Q(s′, a′) is the Q function, which represents the expected reward of taking action a′ in the new state s′;

[0034] It means that for all possible actions a′, the action that maximizes the Q value is selected, that is, the best action in the new state s′;

[0035] s and s′ are the current and next states, respectively, a and a′ are the current and subsequent actions, respectively;

[0036] Knowledge-driven learning and optimization: Operational knowledge is extracted from decision implementation through a continuous learning mechanism, and model learning and data simulation of historical events are realized by using a generative adversarial network (GAN) to promote the effective integration of scenario elements and enhance model adaptability. The generator G and discriminator D in the training process of the generative adversarial network GAN are optimized through the following adversarial losses:

[0037]

[0038] in, It is a confrontation game between the generator G and the discriminator D during the training process, where the goal of the generator G is to minimize the function V, while the goal of the discriminator D is to maximize V;

[0039] is the expected value expression, which means that for the real data distribution p data The logarithm of the probability that the discriminator D classifies the sample x as true, and the discriminator D attempts to correctly identify the actual data sample;

[0040] The expected value expression represents the expected value of the noise distribution p. z The logarithm of the probability that the false sample G(z) generated by the generator G is classified as false by the discriminator D, and the generator attempts to generate samples that can deceive the discriminator so that the discriminator D misclassifies the false sample as a real sample;

[0041] D(x) is the discriminator network, which outputs a scalar indicating the probability that the sample x is a true sample;

[0042] G(z) is a generator network that accepts the noise sample z as input and generates a data sample;

[0043] p datais the probability distribution of the real data;

[0044] p z Probability distribution of the noise at the generator input.

[0045] Furthermore, the multi-source data includes:

[0046] Headcount, the unit is person, expressed as an integer;

[0047] Spatial data, in square meters, with areas represented as continuous floating-point values;

[0048] Temperature, in degrees Celsius, expressed as a continuous floating point value;

[0049] Humidity, in percentage %, expressed as a continuous floating point value;

[0050] Wind speed, in meters per second, expressed as a continuous floating point value;

[0051] Infrastructure status, in terms of service availability or damage, expressed as a percentage;

[0052] Traffic volume and mode, in person-times / hour, expressed as an integer;

[0053] Communication network data, the unit is data flow bits / second, expressed as an integer;

[0054] Resource availability, including emergency service personnel and rescue equipment, in units of the quantity of the resource, expressed as an integer.

[0055] Furthermore, in the multi-source data integration and preprocessing step, the step of probabilistically fusing the input data sources using a Bayesian network combined with a deep belief network specifically includes the following steps:

[0056] S1: including using a Bayesian network to perform a preliminary probability analysis on the input data from several of the data sources, wherein the Bayesian network calculates the probability P(C k |x), where C k |x is the category, x is the input data point, using the Bayesian formula:

[0057]

[0058] Among them, P(C k ) is the prior probability, P(C k |x) is the likelihood of a data point given its category, and P(x) is the marginal probability of the data point;

[0059] S2: Fusion deep belief network DBN, the deep belief network DBN receives the intermediate probability distribution output by the Bayesian network as input, and performs in-depth learning processing through its multi-layer structure to improve the accuracy of probability analysis; the deep belief network DBN learns data representation layer by layer through the pre-trained restricted Boltzmann machine RBM layer, each of the restricted Boltzmann machine RBM layer adjusts its weight to maximize the reconstruction probability of the data, and finally outputs a probability distribution P(C k |x);

[0060] S3: The fusion output stage of the Bayesian network and the deep belief network DBN, the output P′(C k |x) and the preliminary output P(C k |x) to fuse the probability evaluations of the two networks using a weighted average statistical method, and finally form a comprehensive credibility evaluation for each of the data points. The formula used for the fusion is:

[0061] P″(C k |x)=αP(C k |x)+(1-α)P′(C k ||x)

[0062] Among them, α is an adjustment parameter used to balance the weights of the two probability contributions; P″(C k |x) is the probability result after fusion.

[0063] Furthermore, the multi-source data integration and preprocessing step further includes a data preprocessing step of introducing an autoencoder or a variational autoencoder:

[0064] The preprocessing unit further includes at least one autoencoder or variational autoencoder for denoising and cleaning data; the autoencoder or variational autoencoder is responsible for receiving raw data from a variety of data sources, and automatically learning important features of the data through its network structure, while excluding noise and irrelevant information; the autoencoder converts the input data x into a lower-dimensional potential representation z through an encoder network, which is expressed as:

[0065] z=f θ (x)

[0066] Among them, f θ represents the encoder function, θ is the network parameter; then, a decoder network tries to reconstruct the input data x′ from the potential representation z to train the model to minimize the reconstruction error ||xx′||, helping the model learn to suppress noise while retaining the key features of the data;

[0067] The training process of the autoencoder or the variational autoencoder also includes using regression techniques to optimize network parameters to ensure that the data is converted from the original format to a cleaner and more standardized format.

[0068] Furthermore, the dynamic scenario modeling step also includes a scenario model construction technology using a graph neural network GNN, which specifically includes the following steps:

[0069] Node construction step: construct the nodes of the graph neural network according to the characteristics of the scenario elements. The characteristics of the scenario elements include position, type, and state. Each node represents a unique scenario element. The feature vector x of the node i It is obtained by encoding its related attributes and can be expressed as:

[0070] x i =Encode(feature 1 ,feature 2 ,...,feature n )

[0071] The Encode function is used to convert the characteristics of the scenario elements into a numerical feature vector;

[0072] Edge construction step: construct the edge of the graph neural network according to the relationship between the scenario elements. The edge represents the interaction or influence between node i and node j. The weight of the edge can be calculated based on the physical distance and functional connection index between the two nodes. The formula can be:

[0073] w ij =Weight(relation ij )

[0074] Among them, w ij Represents the weight of the edge connecting the node i and the node j. The Weight function is used to calculate the weight of the edge in the graph neural network. The weight represents the strength or importance of the connection between two nodes in the graph.

[0075] Dynamic adjustment step: monitor the changes of scenario elements in real time, and dynamically adjust the attributes of nodes and edges according to the changes. The attributes include updating the feature vectors of nodes and the weights of edges to reflect the latest emergency scenario status. The dynamic adjustment can use a time-based algorithm to update the graph structure through an incremental learning method, where the attribute updates of nodes and edges can be expressed as:

[0076]

[0077]

[0078] Where Δx i and Δw ijThey are the changes in node and edge attributes calculated based on the latest data, t represents the current time step, and t+1 represents the next time step, that is, the moment immediately after t.

[0079] Furthermore, in the adaptive decision support step, a deep Q network DQN is used as a reinforcement learning model. The reinforcement learning model optimizes the decision process by combining a Q learning algorithm and a deep neural network. The deep Q network DQN can dynamically adjust the strategy to achieve the optimal decision based on the reward feedback of the state-action pair learned from real-time and historical data. The decision optimization also includes the following sub-steps:

[0080] State representation: Represent the results of real-time scenario analysis as the state of the decision-making system;

[0081] Strategy learning: Use reinforcement learning algorithms to learn the optimal strategy to take in different states;

[0082] Action execution: According to the learned strategy, corresponding decision actions are executed, and feedback results are collected for further optimization of the strategy.

[0083] Furthermore, in the knowledge-driven learning and optimization step, the generative adversarial network GAN is used to simulate historical emergency event data, and the generator learns to create data samples with high sense of reality, while the discriminator learns to distinguish between real and generated data, thereby improving the accuracy and adaptability of the model in predicting future emergency scenarios. Specifically, the following steps are also included:

[0084] Context-aware step: Identify the multimodal data and contextual information most relevant to the current emergency scenario through the attention mechanism;

[0085] Data integration step: Integrate the identified data and information to form a knowledge base for decision support;

[0086] Model training steps: Use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, and train the model to improve prediction and decision-making capabilities.

[0087] Furthermore, the decision perspective and behavior analysis step also includes applying a decision tree algorithm to construct a decision model for key actors. The decision tree algorithm generates a series of decision nodes and branches, and performs probability evaluation on each decision node, thereby analyzing possible behavior paths and predicting decision results, which is used to optimize strategy adjustments in decision support systems.

[0088] Furthermore, the model compression technology in the adaptive decision support step specifically includes the following steps:

[0089] Model pruning: Execute an iterative pruning algorithm that first identifies and removes neuronal connections in the deep learning model whose weight contribution is below a predetermined threshold, then retrains the adjusted model to restore performance. The process can be iterated until a balance between performance and model complexity is achieved.

[0090] Model quantization process: Implement a quantization strategy to reduce the required computational resources and storage requirements by reducing the bit width of the data representation used in the model. This includes converting floating-point weights to fixed-point or lower-precision formats and applying quantization-aware training methods to optimize model parameters to maintain prediction accuracy.

[0091] Knowledge distillation implementation: Knowledge distillation technology is used to transfer the knowledge of a large pre-trained deep learning model to a smaller model. This is achieved by training the small model to mimic the output distribution of the large model, thereby improving the performance of the small model.

[0092] Furthermore, in the real-time data processing and feedback step, the stream processing framework is based on the distributed data stream processing architecture of Apache Kafka, and utilizes the high throughput and low latency characteristics of Kafka to collect data from multiple data sources, and performs real-time filtering, aggregation and windowing of the data through the Kafka Streams API, and distributes the processed data and emergency scenario analysis results to the downstream system in real time. The specific steps include:

[0093] Data collection sub-step: Collect data from multiple data sources, including sensor data, social media data, and network traffic data, through Apache Kafka's producer client, and use Kafka's high throughput characteristics to process large-scale data streams;

[0094] Data processing sub-step: Use the Kafka Streams API to implement real-time data processing, including data filtering, aggregation, and windowing operations, and apply complex event processing logic to detect and respond to specific data patterns and event sequences;

[0095] Data distribution sub-step: Use Kafka's consumer client to distribute the processed data and emergency scenario analysis results to downstream systems in real time, ensuring the real-time and high reliability of the data;

[0096] Feedback generation sub-step: Based on the results of processing and analysis, feedback on key decision points and potential risk assessments is provided to users in real time through visual dashboards and mobile device applications to support users' rapid decision-making and emergency response.

[0097] The beneficial effects of the present invention are as follows: the present invention provides an intelligent analysis and decision support method for emergency scenarios based on an AI big model, which integrates multi-source data and applies data set preprocessing formula technology to reduce dependence on high-quality data, while using Bayesian networks and data standardization technology to improve the credibility and consistency of information; through dynamic scenario modeling, combined with adaptive deep learning networks and graph neural networks, the model building process is simplified, and the adaptability and generalization ability to complex scenarios are improved; the real-time data processing architecture based on stream computing is used to ensure low-latency data processing, so that the decision support system can respond to emergencies quickly and effectively; combined with the adaptive decision support mechanism of reinforcement learning, the operational flexibility and decision accuracy are improved; through the integration of big model technology, scientific, rigorous and efficient analysis support is provided for emergency scenario management, which significantly enhances the ability to respond to emergencies and ensures the real-time and accuracy of decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work:

[0099] Figure 1 It is a schematic diagram of a framework of an emergency scenario intelligent analysis and decision support method based on an AI big model according to an embodiment of the present invention;

[0100] Figure 2 is a schematic diagram of a multi-source data integration and preprocessing process according to an embodiment of the present invention;

[0101] Figure 3 It is a schematic diagram of a dynamic scenario modeling and decision support system according to an embodiment of the present invention.

[0102] Figure 4 It is a schematic diagram of the real-time data processing and feedback mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the following will be described clearly and completely in combination with the technical solution in the embodiments of the present invention. Obviously, the described embodiments are partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0104] like Figures 1 to 4 As shown in the figure, the emergency scenario intelligent analysis and decision support method based on the AI ​​big model includes the following steps:

[0105] See also Figure 2 , Multi-source data integration and preprocessing: receiving input from multiple data sources through a data integration module, the data sources include several sensor collection data, historical databases and real-time communication networks; the multi-source data includes:

[0106] Headcount, the unit is person, expressed as an integer;

[0107] Spatial data, in square meters, with areas represented as continuous floating-point values;

[0108] Temperature, in degrees Celsius, expressed as a continuous floating point value;

[0109] Humidity, in percentage %, expressed as a continuous floating point value;

[0110] Wind speed, in meters per second, expressed as a continuous floating point value;

[0111] Infrastructure status, in terms of service availability or damage, expressed as a percentage;

[0112] Traffic volume and mode, in person-times / hour, expressed as an integer;

[0113] Communication network data, the unit is data flow bits / second, expressed as an integer;

[0114] Resource availability, including emergency service personnel and rescue equipment, in units of the quantity of the resource, expressed as an integer.

[0115] The sensor comprises:

[0116] Infrared sensors or video analytics sensors: used to monitor and count the number of people in a specific area;

[0117] Laser rangefinder or ultrasonic sensor: used to measure and monitor space size or area;

[0118] Thermocouple or thermistor: used to measure ambient temperature;

[0119] Capacitive humidity sensor: used to monitor the relative humidity in the environment;

[0120] Anemometer: used to measure wind speed, often used for space wind observation;

[0121] Infrastructure Status Sensors:

[0122] Vibration sensors or acoustic emission sensors: used to monitor the structural integrity and operating status of buildings or facilities;

[0123] People flow and pattern monitoring sensors:

[0124] Infrared sensors or video surveillance systems: used to record and analyze the number and movement patterns of people;

[0125] Network traffic monitoring tool: used to monitor data traffic and network status in real time;

[0126] Resource availability monitoring sensors:

[0127] RFID or barcode scanners: Used to track the location and number of emergency service personnel and rescue equipment.

[0128] The historical database is a system for storing and managing past scene activity data records, which may include various types of historical information for analysis, reporting, decision support or reference.

[0129] In emergency scenarios, the real-time communication network plays a vital role, ensuring the timely transmission and receipt of key information, thereby enabling decision makers to respond quickly to emergencies.

[0130] The Bayesian network combined with the deep belief network is used to perform probability fusion on the input data source to evaluate the credibility of the data point. The probability fusion is performed by weighting the trust scores of different data sources. The credibility of each data point is determined by the error rate and timestamp of its source. The data source data is denoised, formatted and standardized by the preprocessing unit, and data cleaning and denoising are performed to reduce the dependence on the quality of the original data. The formula used for preprocessing the data set of the data source is:

[0131] D'=f(D)where f(·)denotes the standardization function

[0132] Wherein, D is the original data set of the data source; f(·) is a normalization function, which is used to convert the original data into a standard or standardized format; D′ is the data set after normalization;

[0133] The purpose of using the Bayesian network combined with the deep belief network to perform probability fusion on the input data source is to jointly evaluate the credibility of the data points. The steps specifically include the following steps:

[0134] S1: including using a Bayesian network to perform a preliminary probability analysis on the input data from several of the data sources, wherein the Bayesian network calculates the probability P(C k |x), where C k |x is the category, x is the input data point, using the Bayesian formula:

[0135]

[0136] Among them, P(C k ) is the prior probability, P(C k |x) is the likelihood of a data point given its category, and P(x) is the marginal probability of the data point;

[0137] S2: Fusion deep belief network DBN, the deep belief network DBN receives the intermediate probability distribution output by the Bayesian network as input, and performs in-depth learning processing through its multi-layer structure to improve the accuracy of probability analysis; the deep belief network DBN learns data representation layer by layer through the pre-trained restricted Boltzmann machine RBM layer, each of the restricted Boltzmann machine RBM layer adjusts its weight to maximize the reconstruction probability of the data, and finally outputs a probability distribution P(C k |x);

[0138] S3: The fusion output stage of the Bayesian network and the deep belief network DBN, the output P′(C k |x) and the preliminary output P(C k |x) to fuse the probability evaluations of the two networks using a weighted average statistical method, and finally form a comprehensive credibility evaluation for each of the data points. The formula used for the fusion is:

[0139] P"(C k |x)=αP(C k |x)+(1-α)P′(C k |x)

[0140] Among them, α is an adjustment parameter used to balance the weights of the two probability contributions; P″(C k |x) is the probability result after fusion.

[0141] The multi-source data integration and preprocessing step also includes a data preprocessing step of introducing an autoencoder or a variational autoencoder:

[0142] The preprocessing unit further includes at least one autoencoder or variational autoencoder for denoising and cleaning data; the autoencoder or variational autoencoder is responsible for receiving raw data from a variety of data sources, and automatically learning important features of the data through its network structure, while excluding noise and irrelevant information; the autoencoder converts the input data x into a lower-dimensional potential representation z through an encoder network, which is expressed as:

[0143] z=f θ (x)

[0144] Among them, f θrepresents the encoder function, θ is the network parameter; then, a decoder network tries to reconstruct the input data x′ from the potential representation z to train the model to minimize the reconstruction error ||xx′||, helping the model learn to suppress noise while retaining the key features of the data;

[0145] For the variational autoencoder, in addition to the above process, this technique generates statistical distribution parameters (mean μ and variance σ) representing the latent space through the encoder network. 2 ), and sample the latent representation z from this distribution:

[0146]

[0147] This allows the variational autoencoder to learn data representation while generating new data points that reflect the statistical properties of the original data, which can be used for further processing such as data augmentation and model training.

[0148] The training process of the autoencoder or the variational autoencoder also includes using regression techniques to optimize network parameters to ensure that the data is converted from the original format to a cleaner, more standardized format, thereby reducing dependence on the quality of the original data and improving the efficiency and accuracy of subsequent data processing and analysis.

[0149] like Figure 3 As shown, dynamic scenario modeling: using a deep learning network with adaptive structural adjustment capabilities, dynamically construct and optimize the emergency scenario model according to the real-time and historical data of the data source after standardization. The deep learning network adopts a recursive neural network RNN ​​architecture composed of three layers of long short-term memory LSTM networks. Each layer of the long short-term memory LSTM automatically adjusts its network weights to optimize the prediction accuracy of emergency scenarios by continuously learning the time dependency in the data source. The adjustment of the network weights is implemented by the back propagation algorithm to ensure that the model reflects the real-time data changes more accurately after multiple iterations. The recursive neural network RNN ​​architecture dynamically adjusts the network weights by the gradient descent method. The update formula of the network weights is:

[0150] Among them, θ t represents the network weight, α is the learning rate, and L is the loss function, which is used to evaluate the deviation between the model prediction results and the actual situation;

[0151] The dynamic scenario modeling step also includes a scenario model construction technology using a graph neural network GNN, which specifically includes the following steps:

[0152] Node construction step: construct the nodes of the graph neural network according to the characteristics of the scenario elements. The characteristics of the scenario elements include position, type, and state. Each node represents a unique scenario element. The feature vector x of the nodei It is obtained by encoding its related attributes and can be expressed as:

[0153] x i =Encode(feature 1 ,feature 2 ,...,feature n )

[0154] The Encode function is used to convert the characteristics of the scenario elements into a numerical feature vector;

[0155] Edge construction step: construct the edge of the graph neural network according to the relationship between the scenario elements. The edge represents the interaction or influence between node i and node j. The weight of the edge can be calculated based on the physical distance and functional connection index between the two nodes. The formula can be:

[0156] w ij =Weight(relation ij )

[0157] Among them, w ij Represents the weight of the edge connecting the node i and the node j. The Weight function is used to calculate the weight of the edge in the graph neural network. The weight represents the strength or importance of the connection between two nodes in the graph.

[0158] Decision-making perspective and behavior analysis: The decision analysis module evaluates and simulates the behavior of key decision makers and their impact on emergency response, and uses support vector machines (SVMs) to classify and predict key actors to provide a basis for real-time adjustment of decision support strategies. The behavior patterns of key actors are classified into two categories using support vector machines to distinguish different emergency response results that may be caused by their behaviors. The classification of key actors is solved by optimizing the problem using the following formula:

[0159]

[0160] Wherein, w is a weight vector, which determines the direction of the best separation hyperplane in the support vector machine SVM;

[0161] b is a bias term, which together with the weight vector w determines the position of the hyperplane;

[0162] ||w|| 2 is the square norm of the weight vector, the inverse of which is a measure of the size of the interval of the hyperplane. Minimizing this norm is to maximize the interval;

[0163] are coefficients used for mathematical processing, especially for ||w|| 2 Derivation;

[0164] C is the regularization parameter, which determines the degree of penalty for data points that violate the interval rule. The larger the C value, the greater the penalty for data points that violate the interval.

[0165] ξ i is a slack variable used to handle incompletely separable situations and allow data points to violate the interval rule;

[0166] is the sum of all slack variables, and the support vector machine SVM tries to minimize this value, thereby allowing as few data points as possible to violate the interval rule;

[0167] The decision perspective and behavior analysis step also includes applying a decision tree algorithm to construct a decision model for key actors. The decision tree algorithm generates a series of decision nodes and branches, and performs a probability evaluation on each decision node, thereby analyzing possible behavior paths and predicting decision results, which is used to optimize policy adjustments in decision support systems.

[0168] Dynamic adjustment step: monitor the changes of scenario elements in real time, and dynamically adjust the attributes of nodes and edges according to the changes. The attributes include updating the feature vectors of nodes and the weights of edges to reflect the latest emergency scenario status. The dynamic adjustment can use a time-based algorithm to update the graph structure through an incremental learning method, where the attribute updates of nodes and edges can be expressed as:

[0169]

[0170]

[0171] Where Δx i and Δw ij They are the changes in node and edge attributes calculated based on the latest data, t represents the current time step, and t+1 represents the next time step, that is, the moment immediately after t.

[0172] like Figure 4 As shown, real-time data processing and feedback: implement a data processing architecture based on stream computing, use a stream processing framework to collect, process and distribute data, ensure the real-time nature of data and reduce processing delays, generate emergency scenario analysis results and provide feedback to users. The emergency scenario analysis results include key decision points and potential risk assessments, and are combined with real-time analysis results of emergency scenarios to feed back to the decision model for dynamic update;

[0173] In the real-time data processing and feedback step, the stream processing framework is based on the distributed data stream processing architecture of Apache Kafka, and uses the high throughput and low latency characteristics of Kafka to collect data from multiple data sources, and performs real-time filtering, aggregation and windowing of the data through the Kafka Streams API, and distributes the processed data and emergency scenario analysis results to the downstream system in real time. The specific steps include:

[0174] Data collection sub-step: Collect data from multiple data sources, including sensor data, social media data, and network traffic data, through Apache Kafka's producer client, and use Kafka's high throughput characteristics to process large-scale data streams;

[0175] Data processing sub-step: Use the Kafka Streams API to implement real-time data processing, including data filtering, aggregation, and windowing operations, and apply complex event processing logic to detect and respond to specific data patterns and event sequences;

[0176] Data distribution sub-step: Use Kafka's consumer client to distribute the processed data and emergency scenario analysis results to downstream systems in real time, ensuring the real-time and high reliability of the data;

[0177] Feedback generation sub-step: Based on the results of processing and analysis, feedback on key decision points and potential risk assessments is provided to users in real time through visual dashboards and mobile device applications to support users' rapid decision-making and emergency response.

[0178] Adaptive decision support: configure an adaptive decision support system, which integrates decision optimization technology based on the decision model and reinforcement learning model, and dynamically uses model pruning, quantization and knowledge distillation model compression technology to adjust the decision strategy according to real-time feedback and historical data. The model compression technology in the adaptive decision support step specifically includes the following steps:

[0179] Model pruning: Execute an iterative pruning algorithm that first identifies and removes neuronal connections in the deep learning model whose weight contribution is below a predetermined threshold, then retrains the adjusted model to restore performance. The process can be iterated until a balance between performance and model complexity is achieved.

[0180] Model quantization process: Implement a quantization strategy to reduce the required computational resources and storage requirements by reducing the bit width of the data representation used in the model. This includes converting floating-point weights to fixed-point or lower-precision formats and applying quantization-aware training methods to optimize model parameters to maintain prediction accuracy.

[0181] Knowledge distillation implementation: Knowledge distillation technology is used to transfer the knowledge of a large pre-trained deep learning model to a smaller model. This is achieved by training the small model to mimic the output distribution of the large model, thereby improving the performance of the small model.

[0182] Finally, the decision support system outputs specific emergency response decision suggestions, which include emergency measures, resource allocation, and response priorities to guide actual emergency operations. The reinforcement learning model is learned and optimized through the following reward function:

[0183]

[0184] Among them, R(s, a) is the reward obtained by taking action a in state s;

[0185] r 0 The immediate reward is the immediate reward obtained by taking action a in state s;

[0186] γ is the discount factor, which ranges from 0 to 1. It determines the importance of future returns relative to immediate returns. The closer γ is to 1, the greater the importance of future returns.

[0187] Q(s′, a′) is the Q function, which represents the expected reward of taking action a′ in the new state s′. The Q function is usually approximated by a Q table or a neural network.

[0188] It means that for all possible actions a′, the action that maximizes the Q value is selected, that is, the best action in the new state s′;

[0189] s and s′ are the current and next states, respectively, a and a′ are the current and subsequent actions, respectively;

[0190] A deep Q network DQN is used as a reinforcement learning model. The reinforcement learning model optimizes the decision-making process by combining a Q learning algorithm and a deep neural network. The deep Q network DQN can dynamically adjust the strategy to achieve the optimal decision based on the reward feedback of the state-action pair learned from real-time and historical data. The decision optimization also includes the following sub-steps: State representation: representing the result of real-time scenario analysis as the state of the decision-making system;

[0191] Strategy learning: Use reinforcement learning algorithms to learn the optimal strategy to take in different states;

[0192] Action execution: According to the learned strategy, corresponding decision actions are executed, and feedback results are collected for further optimization of the strategy.

[0193] Knowledge-driven learning and optimization: Operational knowledge is extracted from decision implementation through a continuous learning mechanism, and model learning and data simulation of historical events are realized by using a generative adversarial network (GAN) to promote the effective integration of scenario elements and enhance model adaptability. The generator G and discriminator D in the training process of the generative adversarial network GAN are optimized through the following adversarial losses:

[0194]

[0195] in, It is a confrontation game between the generator G and the discriminator D during the training process, where the goal of the generator G is to minimize the function V, while the goal of the discriminator D is to maximize V;

[0196] is the expected value expression, which means that for the real data distribution p data The logarithm of the probability that the discriminator D classifies the sample x as true, and the discriminator D attempts to correctly identify the actual data sample;

[0197] The expected value expression represents the expected value of the noise distribution p. z The logarithm of the probability that the false sample G(z) generated by the generator G is classified as false by the discriminator D, and the generator attempts to generate samples that can deceive the discriminator so that the discriminator D misclassifies the false sample as a real sample;

[0198] D(x) is the discriminator network, which outputs a scalar indicating the probability that the sample x is a true sample;

[0199] G(z) is a generator network that accepts the noise sample z as input and generates a data sample;

[0200] p data is the probability distribution of the real data;

[0201] p z The probability distribution of the noise at the generator input;

[0202] The goal of the entire function is to optimize the generator G and the discriminator D so that the discriminator D can distinguish between real and generated samples, while enabling the generator G to generate samples that are as close to the real data distribution as possible.

[0203] The knowledge-driven learning and optimization step uses the generative adversarial network GAN to simulate historical emergency event data, and the generator learns to create data samples with high sense of reality, while the discriminator learns to distinguish between real and generated data, thereby improving the accuracy and adaptability of the model in predicting future emergency scenarios, and specifically includes the following steps:

[0204] Context-aware step: Identify the multimodal data and context information most relevant to the current emergency scenario through the attention mechanism; the specific steps are:

[0205] Define input data: First, define and collect multimodal data related to emergency management, including videos, images, text reports, sensor data, etc. These data sources will serve as model inputs for subsequent processing and analysis.

[0206] Apply attention mechanisms: Use attention mechanisms to automatically identify and emphasize the information most relevant to the current emergency scenario. This can be achieved using one or more deep learning models, such as Transformer-based models. For example, an attention model can be designed that learns how to extract key information from complex input data, such as video surveillance and social media updates:

[0207]

[0208] Where X is the input multimodal dataset, Q, K, and V are query, key, and value functions, respectively, which can be learned from the data through neural networks. k is the dimension of the key, used to scale the dot product.

[0209] Information filtering and focus identification: Based on the attention score, the most critical information fragments are filtered out, such as specific video frames, text messages, or sensor readings, which will be marked as the focus of processing and response.

[0210] Data integration step: Integrate the identified data and information to form a knowledge base for decision support; the specific steps are:

[0211] Data preprocessing: standardize and format the collected data to ensure data quality and meet the needs of subsequent analysis;

[0212] Build a knowledge base: Integrate the data filtered by the attention mechanism into a unified platform or database. This knowledge base will support various query operations and can be used to quickly retrieve information and support decision making. Use a database management system (such as SQL or NoSQL database) to organize data and ensure data accessibility and consistency;

[0213] Contextual association analysis: Implement contextual association analysis function in the knowledge base, and identify the correlation and association patterns between different data sources and information through data mining and pattern recognition technology.

[0214] Model training steps: Use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, and use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, and train models to improve prediction and decision-making capabilities. The data generated by the generator can be used to train one or more prediction models, which can be based on traditional machine learning or deep learning. These models will learn how to identify patterns and potential emergency situations from the data. And regularly evaluate model performance, adjust and optimize based on feedback from actual applications. Use cross-validation and other statistical methods to verify the prediction accuracy and reliability of the model.

[0215] The technical solution of this embodiment improves and solves some key deficiencies in the prior art. First, for the problem of reliance on high-quality data, this solution effectively reduces the reliance on a single data source and improves the overall credibility and quality of the data by implementing a multi-source data integration and preprocessing module, combining Bayesian networks and deep belief networks for probabilistic fusion and data cleaning. This helps maintain system performance even when the quality of data sources is uneven.

[0216] Regarding the complexity of constructing a situational awareness framework, the technical solution provides powerful tools for dynamically constructing complex situational awareness frameworks through dynamic situation modeling, especially the use of recursive neural networks (RNN) and long short-term memory networks (LSTM). These tools can effectively adapt to the changing environment and conditions. In addition, the introduction of an adaptive decision support system, combined with decision models and reinforcement learning-based strategy adjustments, significantly enhances the model's generalization ability in the face of diverse situations and disaster patterns, while improving flexibility and responsiveness in emergency situations, which is also impossible with traditional static technologies.

[0217] The technical solution of this embodiment also pays special attention to the problems of data real-time performance and processing delay. By adopting a data processing architecture based on stream computing, especially integrating a stream processing framework such as Apache Kafka, it optimizes the real-time processing capability of data, reduces processing delays during data peaks, and ensures that the system maintains efficient response speed and processing efficiency under high load or data peak conditions.

[0218] However, despite the many improvements provided by the technical solution, as the number of monitoring points and data increases, computing efficiency and real-time processing capabilities may still face challenges. However, this technical solution introduces a reinforcement learning model for adaptive decision support to learn and optimize through reward functions and knowledge-driven learning and optimization, focusing on algorithm optimization and decision support management to maintain the system's performance and learning capabilities under the ever-expanding data and monitoring needs. This is not covered by the technology in this field, and has outstanding substantive characteristics and significant progress.

[0219] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. The intelligent analysis and decision support method for emergency scenarios based on AI big model is characterized by: The following steps are involved: Multi-source data integration and preprocessing: Receive input from multiple data sources through a data integration module, the data sources include several sensor collection data, historical databases and real-time communication networks; use Bayesian network combined with deep belief network to perform probability fusion on the input data sources and evaluate the credibility of data points. The probability fusion is performed by weighting the trust scores of different data sources. The credibility of each data point is determined by the error rate and timestamp of its source; the data source data is denoised, formatted and standardized through the preprocessing unit, and data cleaning and denoising are performed to reduce the dependence on the quality of the original data. The formula used for preprocessing the data set of the data source is: D′=f(D) Wherein, D is the original data set of the data source; D′ is the data set after standardization; f(·) is the standardization function, and the formula is used to convert the original data into a standard or standardized format; Dynamic scenario modeling: Using a deep learning network with adaptive structural adjustment capabilities, the emergency scenario model is dynamically constructed and optimized based on the real-time and historical data of the data source after standardization. The deep learning network adopts a recursive neural network RNN ​​architecture composed of three layers of long short-term memory LSTM networks. Each layer of the long short-term memory LSTM automatically adjusts its network weights to optimize the prediction accuracy of emergency scenarios by continuously learning the time dependency in the data source. The adjustment of the network weights is implemented using the back propagation algorithm to ensure that the model reflects the real-time data changes more accurately after multiple iterations. The recursive neural network RNN ​​architecture dynamically adjusts the network weights through the gradient descent method. The update formula of the network weights is: Among them, θ t represents the network weight, α is the learning rate, and L is the loss function, which is used to evaluate the deviation between the model prediction results and the actual situation; Decision-making perspective and behavior analysis: The decision analysis module evaluates and simulates the behavior of key decision makers and their impact on emergency response, and uses support vector machines (SVMs) to classify and predict key actors to provide a basis for real-time adjustment of decision support strategies. The behavior patterns of key actors are classified into two categories using support vector machines to distinguish different emergency response results that may be caused by their behaviors. The classification of key actors is solved by optimizing the problem using the following formula: Wherein, w is a weight vector, which determines the direction of the best separation hyperplane in the support vector machine SVM; b is a bias term, which together with the weight vector w determines the position of the hyperplane; ||w|| 2 is the square norm of the weight vector, the inverse of which is a measure of the size of the interval of the hyperplane. Minimizing this norm is to maximize the interval; are coefficients used for mathematical processing, especially for ||w|| 2 Derivation; C is the regularization parameter, which determines the degree of penalty for data points that violate the interval rule. The larger the C value, the greater the penalty for data points that violate the interval. ξ i is a slack variable used to handle incompletely separable situations and allow data points to violate the interval rule; is the sum of all slack variables; Real-time data processing and feedback: Implement a data processing architecture based on stream computing, use a stream processing framework to collect, process and distribute data, ensure the real-time nature of data and reduce processing delays, generate emergency scenario analysis results and provide feedback to users. The emergency scenario analysis results include key decision points and potential risk assessments, and are combined with real-time analysis results of emergency scenarios to feed back to the decision model for dynamic updates; Adaptive decision support: Configure an adaptive decision support system, which integrates decision optimization technology based on the decision model and reinforcement learning model, dynamically uses model pruning, quantization and knowledge distillation model compression technology to adjust the decision strategy according to real-time feedback and historical data, and finally outputs specific emergency response decision suggestions through the decision support system. The emergency response decision suggestions include emergency measures, resource allocation and response priority to guide actual emergency operations. The reinforcement learning model is learned and optimized through the following reward function: Among them, R(s, a) is the reward obtained by taking action a in state s; r0 is the immediate reward, which is the immediate reward obtained by taking action a in state s; γ is the discount factor, which ranges from 0 to 1. It determines the importance of future returns relative to immediate returns. The closer γ is to 1, the greater the importance of future returns. Q(s′, a′) is the Q function, which represents the expected reward of taking action a′ in the new state s′; It means that for all possible actions a′, the action that maximizes the Q value is selected, that is, the best action in the new state s′; s and s′ are the current and next states, respectively, a and a′ are the current and subsequent actions, respectively; Knowledge-driven learning and optimization: Operational knowledge is extracted from decision implementation through a continuous learning mechanism, and model learning and data simulation of historical events are realized by using a generative adversarial network (GAN) to promote the effective integration of scenario elements and enhance model adaptability. The generator G and discriminator D in the training process of the generative adversarial network GAN are optimized through the following adversarial losses: in, It is a confrontation game between the generator G and the discriminator D during the training process, where the goal of the generator G is to minimize the function V, while the goal of the discriminator D is to maximize V; is the expected value expression, which means that for the real data distribution p data The logarithm of the probability that the discriminator D classifies the sample x as true, and the discriminator D attempts to correctly identify the actual data sample; The expected value expression represents the expected value of the noise distribution p. z The logarithm of the probability that the false sample G(z) generated by the generator G is classified as false by the discriminator D, and the generator attempts to generate samples that can deceive the discriminator so that the discriminator D misclassifies the false sample as a real sample; D(x) is the discriminator network, which outputs a scalar indicating the probability that the sample x is a true sample; G(z) is a generator network that accepts the noise sample z as input and generates a data sample; p data is the probability distribution of the real data; p z Probability distribution of the noise at the generator input.

2. The method according to claim 1, characterized in that The multi-source data includes: Headcount, the unit is person, expressed as an integer; Spatial data, in square meters, with areas represented as continuous floating-point values; Temperature, in degrees Celsius, expressed as a continuous floating point value; Humidity, in percentage %, expressed as a continuous floating point value; Wind speed, in meters per second, expressed as a continuous floating point value; Infrastructure status, in terms of service availability or damage, expressed as a percentage; Traffic volume and mode, in person-times / hour, expressed as an integer; Communication network data, the unit is data flow bits / second, expressed as an integer; Resource availability, including emergency service personnel and rescue equipment, in units of the quantity of the resource, expressed as an integer.

3. The method according to claim 1, characterized in that In the multi-source data integration and preprocessing step, the step of probabilistically fusing the input data sources using a Bayesian network combined with a deep belief network specifically includes the following steps: S1: includes using a Bayesian network to perform a preliminary probability analysis on the input data from several of the data sources, wherein the Bayesian network calculates the probability P(C k |x), where C k |x is the category, x is the input data point, using the Bayesian formula: Among them, P(C k ) is the prior probability, P(C k |x) is the likelihood of a data point given its category, and P(x) is the marginal probability of the data point; S2: Fusion deep belief network DBN, the deep belief network DBN receives the intermediate probability distribution output by the Bayesian network as input, and performs in-depth learning processing through its multi-layer structure to improve the accuracy of probability analysis; the deep belief network DBN learns data representation layer by layer through the pre-trained restricted Boltzmann machine RBM layer, each of the restricted Boltzmann machine RBM layer adjusts its weight to maximize the reconstruction probability of the data, and finally outputs a probability distribution P(C k |x); S3: The fusion output stage of the Bayesian network and the deep belief network DBN, the output P′(C k |x) and the preliminary output P(C k |x) to fuse the probability evaluations of the two networks using a weighted average statistical method, and finally form a comprehensive credibility evaluation for each of the data points. The formula used for the fusion is: P"(C k |x)=αP(C k |x)+(1-α)P′(C k |x) Among them, α is an adjustment parameter used to balance the weights of the two probability contributions; P″(C k |x) is the probability result after fusion.

4. The method according to claim 1, characterized in that: The multi-source data integration and preprocessing step also includes a data preprocessing step of introducing an autoencoder or a variational autoencoder: The preprocessing unit further includes at least one autoencoder or variational autoencoder for denoising and cleaning data; the autoencoder or variational autoencoder is responsible for receiving raw data from a variety of data sources, and automatically learning important features of the data through its network structure, while excluding noise and irrelevant information; the autoencoder converts the input data x into a lower-dimensional potential representation z through an encoder network, which is expressed as: z=f θ (x) Among them, f θ represents the encoder function, θ is the network parameter; then, a decoder network tries to reconstruct the input data x′ from the potential representation z to train the model to minimize the reconstruction error ||xx′||, helping the model learn to suppress noise while retaining the key features of the data; The training process of the autoencoder or the variational autoencoder also includes using regression techniques to optimize network parameters to ensure that the data is converted from the original format to a cleaner and more standardized format.

5. The method according to claim 1, characterized in that The dynamic scenario modeling step also includes a scenario model construction technology using a graph neural network GNN, which specifically includes the following steps: Node construction step: construct the nodes of the graph neural network according to the characteristics of the scenario elements. The characteristics of the scenario elements include position, type, and state. Each node represents a unique scenario element. The feature vector x of the node i It is obtained by encoding its related attributes and can be expressed as: x i =Encode(feature1,feature2,...,feature n ) The Encode function is used to convert the characteristics of the scenario elements into a numerical feature vector; Edge construction step: construct the edge of the graph neural network according to the relationship between the scenario elements. The edge represents the interaction or influence between node i and node j. The weight of the edge can be calculated based on the physical distance and functional connection index between the two nodes. The formula can be: w ij =Weight(relation ij ) Among them, w ij Represents the weight of the edge connecting the node i and the node j. The Weight function is used to calculate the weight of the edge in the graph neural network. The weight represents the strength or importance of the connection between two nodes in the graph. Dynamic adjustment step: monitor the changes of scenario elements in real time, and dynamically adjust the attributes of nodes and edges according to the changes. The attributes include updating the feature vectors of nodes and the weights of edges to reflect the latest emergency scenario status. The dynamic adjustment can use a time-based algorithm to update the graph structure through an incremental learning method, where the attribute updates of nodes and edges can be expressed as: Where Δx i and Δw ij They are the changes in node and edge attributes calculated based on the latest data, t represents the current time step, and t+1 represents the next time step, that is, the moment immediately after t.

6. The method according to claim 1, characterized in that In the adaptive decision support step, a deep Q network DQN is used as a reinforcement learning model. The reinforcement learning model optimizes the decision process by combining a Q learning algorithm and a deep neural network. The deep Q network DQN can dynamically adjust the strategy to achieve the optimal decision based on the reward feedback of the state-action pair learned from real-time and historical data. The decision optimization also includes the following sub-steps: State representation: Represent the results of real-time scenario analysis as the state of the decision-making system; Strategy learning: Use reinforcement learning algorithms to learn the optimal strategy to take in different states; Action execution: According to the learned strategy, corresponding decision actions are executed, and feedback results are collected for further optimization of the strategy.

7. The method according to claim 1, characterized in that In the knowledge-driven learning and optimization step, the generative adversarial network GAN is used to simulate historical emergency event data, and the generator learns to create data samples with high sense of reality, while the discriminator learns to distinguish between real and generated data, thereby improving the accuracy and adaptability of the model in predicting future emergency scenarios. Specifically, the following steps are also included: Context-aware step: Identify the multimodal data and contextual information most relevant to the current emergency scenario through the attention mechanism; Data integration step: Integrate the identified data and information to form a knowledge base for decision support; Model training steps: Use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, use generative adversarial networks to learn historical emergency event data, generate data for simulated emergency events, and train the model to improve prediction and decision-making capabilities.

8. The method according to claim 1, characterized in that The decision perspective and behavior analysis step also includes applying a decision tree algorithm to construct a decision model for key actors. The decision tree algorithm generates a series of decision nodes and branches, and performs a probability evaluation on each decision node, thereby analyzing possible behavior paths and predicting decision results, which is used to optimize policy adjustments in decision support systems.

9. The method according to claim 6, characterized in that The model compression technology in the adaptive decision support step specifically includes the following steps: Model pruning: Execute an iterative pruning algorithm that first identifies and removes neuronal connections in the deep learning model whose weight contribution is below a predetermined threshold, then retrains the adjusted model to restore performance. The process can be iterated until a balance between performance and model complexity is achieved. Model quantization process: Implement a quantization strategy to reduce the required computational resources and storage requirements by reducing the bit width of the data representation used in the model. This includes converting floating-point weights to fixed-point or lower-precision formats and applying quantization-aware training methods to optimize model parameters to maintain prediction accuracy. Knowledge distillation implementation: Knowledge distillation technology is used to transfer the knowledge of a large pre-trained deep learning model to a smaller model. This is achieved by training the small model to mimic the output distribution of the large model, thereby improving the performance of the small model.

10. The method according to claim 1, characterized in that In the real-time data processing and feedback step, the stream processing framework is based on the distributed data stream processing architecture of Apache Kafka, and uses the high throughput and low latency characteristics of Kafka to collect data from multiple data sources, and performs real-time filtering, aggregation and windowing of the data through the Kafka Streams API, and distributes the processed data and emergency scenario analysis results to the downstream system in real time. The specific steps include: Data collection sub-step: Collect data from multiple data sources, including sensor data, social media data, and network traffic data, through Apache Kafka's producer client, and use Kafka's high throughput characteristics to process large-scale data streams; Data processing sub-step: Use the Kafka Streams API to implement real-time data processing, including data filtering, aggregation, and windowing operations, and apply complex event processing logic to detect and respond to specific data patterns and event sequences; Data distribution sub-step: Use Kafka's consumer client to distribute the processed data and emergency scenario analysis results to downstream systems in real time, ensuring the real-time and high reliability of the data; Feedback generation sub-step: Based on the results of processing and analysis, feedback on key decision points and potential risk assessments is provided to users in real time through visual dashboards and mobile device applications to support users' rapid decision-making and emergency response.