Emergency disaster dynamic decision-making method and system based on multimodal AI large model

Abstract

The present invention provides a method and system for dynamic decision-making in emergency disasters based on a multimodal AI large model. First, a real-time disaster monitoring data set of the target disaster area is obtained, covering disaster image acquisition, environmental sensing, historical response records and geographical distribution topology data. Then, feature extraction is performed on the real-time disaster monitoring data set to obtain a multimodal disaster feature set. Then, a pre-trained multimodal dynamic decision-making model is called to analyze the disaster scene matching degree of the multimodal disaster feature set, generate a matching score and a disaster response optimization strategy set, and then screen the target optimization strategy based on the matching score. An emergency decision instruction set is generated and sent to a disaster response terminal. Finally, decision execution feedback data of the disaster response terminal is obtained, and incremental parameter optimization processing is performed on the multimodal dynamic decision-making model to achieve more accurate and effective dynamic decision-making in emergency disasters.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for dynamic decision-making in emergency disaster response based on a multimodal AI large model. Background Art

[0002] In the field of emergency disaster management, traditional disaster decision-making methods rely primarily on manual experience or simple rule-based matching. These methods often fall short when faced with complex and ever-changing disaster scenarios. Specifically, existing technologies lack the effective integration and utilization of multi-source heterogeneous data, resulting in incomplete and inaccurate descriptions of disaster scenarios. Furthermore, traditional disaster feature extraction methods often only capture the static characteristics of disasters, failing to capture their dynamic evolution and their relationship to environmental factors, thus compromising the scientific nature and timeliness of decision-making.

[0003] Existing technologies for decision-making often rely on rule-based or template-based matching methods. While these approaches can achieve rapid responses to a certain extent, they lack flexibility and adaptability, making them difficult to cope with the complexity and uncertainty of disaster scenarios. Furthermore, due to the lack of effective feedback mechanisms, existing decision-making methods are unable to dynamically adjust and optimize decision-making strategies based on the actual effectiveness of disaster responses, resulting in often unsatisfactory decision-making results. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for dynamic decision-making in emergency disaster response based on a multimodal AI large model, the method comprising:

[0005] Acquire a real-time disaster monitoring data set for the target disaster area, wherein the real-time disaster monitoring data set includes disaster image acquisition data, disaster environment sensor data, disaster historical response record data, and disaster geographical distribution topology data;

[0006] Extracting features from the real-time disaster monitoring data set to obtain a multimodal disaster feature set of the disaster scene, wherein the multimodal disaster feature set includes dynamic features of disaster images, disaster environment association features, disaster response time series features, and geographic topological distribution features;

[0007] Calling a pre-trained multimodal dynamic decision model to perform disaster scenario matching analysis on the multimodal disaster feature set, and generating a disaster scenario matching score and a disaster response optimization strategy set for the target disaster area;

[0008] Selecting a target optimization strategy from the disaster response optimization strategy set based on the disaster scenario matching score, generating a disaster emergency decision instruction set, and sending the disaster emergency decision instruction set to a disaster response terminal;

[0009] Decision execution feedback data returned by the disaster response terminal is obtained, and incremental parameter optimization processing is performed on the multimodal dynamic decision model based on the decision execution feedback data.

[0010] On the other hand, an embodiment of the present invention also provides an emergency disaster dynamic decision-making system based on a multimodal AI large model, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiments of the present invention significantly improve the intelligence level and decision-making efficiency of emergency disaster response. Specifically, by integrating multimodal information such as disaster image acquisition data, environmental sensor data, historical response records and geographical distribution topology data, a feature set that comprehensively reflects the dynamic characteristics of disaster scenes is constructed. Furthermore, the pre-trained multimodal dynamic decision-making model deeply explores the inherent correlation and complementarity between multimodal features to achieve accurate matching evaluation of disaster scenes and intelligent generation of response strategies. It not only improves the scientificity and pertinence of decision-making strategies, but also ensures the optimization of strategy selection through the matching scoring mechanism, effectively avoiding the subjectivity and experience dependence of manual decision-making. In addition, by constructing a decision execution feedback mechanism, the dynamic adjustment and optimization of model parameters are realized, which significantly enhances the adaptability and robustness of the multimodal dynamic decision-making model to complex and changeable disaster scenes. As a result, a fundamental transformation of emergency disaster decision-making from passive response to active prevention, and from experience-driven to data-driven is achieved, which helps to improve disaster response efficiency and reduce disaster losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the emergency disaster dynamic decision-making method based on the multimodal AI large model provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of exemplary hardware and software components of the emergency disaster dynamic decision-making system based on the multimodal AI big model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a dynamic decision-making method for emergency disasters based on a multimodal AI large model provided by an embodiment of the present invention. The dynamic decision-making method for emergency disasters based on a multimodal AI large model is introduced in detail below.

[0015] Step S110: Acquire a real-time disaster monitoring data set of the target disaster area, wherein the real-time disaster monitoring data set includes disaster image acquisition data, disaster environment sensing data, disaster historical response record data, and disaster geographical distribution topology data.

[0016] In this embodiment, in order to make accurate emergency decisions for the target disaster area, it is necessary to fully obtain real-time disaster monitoring data for the target disaster area. The specific acquisition process is as follows:

[0017] Step S111: calling satellite remote sensing equipment to collect disaster image acquisition data of the target disaster area, wherein the disaster image acquisition data includes a pixel matrix of landform changes, thermal distribution gradient information and water coverage change trajectory of the disaster area.

[0018] For example, for a target area affected by an earthquake, satellite remote sensing equipment is used to collect images of the area. The satellite remote sensing equipment captures images of the area at different times. By analyzing these images, a pixel matrix of landform changes can be obtained. Suppose the pixel value matrix of a region in the image during the first acquisition is [100, 110, 120; 130, 140, 150; 160, 170, 180], and during the second acquisition, the pixel value matrix of the region changes to [105, 112, 125; 132, 145, 158; 163, 175, 188]. By comparing these two pixel value matrices, the landform changes can be detected. Thermal gradient information can be obtained by analyzing the temperature of different regions in the image. For example, areas with higher temperatures and larger temperature gradients may be areas of fire caused by the earthquake. Changes in water cover can be determined by identifying and tracking water areas in images at different times. For example, a river may change its course or its water level may change after an earthquake.

[0019] Step S112: Obtain disaster environment sensing data uploaded in real time by multiple environmental sensing nodes deployed in the target disaster area, wherein the disaster environment sensing data includes wind speed fluctuation sequence, temperature change gradient, humidity distribution curve and geological vibration frequency spectrum.

[0020] Multiple environmental sensor nodes have been pre-deployed within the target disaster area, uploading real-time environmental data. For example, in the case of earthquakes, wind speed fluctuations can be recorded by wind speed sensors. Suppose, over a certain period of time, the wind speed values ​​recorded by the wind speed sensors are 3 m / s, 4 m / s, 5 m / s, 3 m / s, 2 m / s, and so on, forming a wind speed fluctuation sequence. Temperature gradients can be calculated by measuring temperatures at different locations and over time using temperature sensors. For example, if the temperature at a certain location rises from 20°C to 22°C within an hour, the temperature gradient is 2°C / hour. Humidity distribution curves can be obtained by measuring humidity at different locations using humidity sensors. These humidity values ​​can be plotted in a coordinate system to create a humidity distribution curve. The geological vibration frequency spectrum, recorded by devices such as seismographs, records the vibration frequencies during an earthquake. For example, the recorded vibration frequencies may range from 1 Hz, 2 Hz, to 5 Hz.

[0021] Step S113: extracting historical disaster response record data of the target disaster area from the disaster emergency database, wherein the historical disaster response record data includes historical disaster type labels, historical emergency response timestamp sequences, historical resource dispatch paths, and historical loss assessment reports.

[0022] Historical response record data for the target disaster area is extracted from a database storing disaster emergency response data. For earthquake-affected areas, historical disaster type labels may indicate that earthquakes, landslides, and other disasters have occurred in the area in the past. The historical emergency response timestamp sequence records the response time for each past disaster. For example, the first earthquake occurred at 10:00 a.m., and the emergency response began at 10:30 a.m.; the second earthquake occurred at 2:00 p.m., and the emergency response began at 2:15 p.m., etc. The historical resource dispatch path records how resources were dispatched during past disasters, such as which road was used to transport supplies from a certain material storage point to the affected area. The historical loss assessment report records in detail the casualties and property losses caused by each past disaster. For example, the first earthquake caused 10 deaths, 50 injuries, and economic losses of 5 million yuan.

[0023] Step S114: calling a geographic information system to analyze the geographic coordinate boundaries of the target disaster area and generate disaster geographic distribution topology data, wherein the disaster geographic distribution topology data includes an elevation gradient distribution map, a road connectivity network, and a coordinate set of residential areas.

[0024] A geographic information system (GIS) is used to analyze the geographic coordinate boundaries of the target disaster area. For example, in an earthquake disaster zone, a GIS can generate an elevation gradient distribution map of the area, representing areas of varying altitude with different colors or lines, such as blue for lower elevations and yellow for higher elevations. A road connectivity network can display the road conditions within the area, including which roads are connected and which may be disrupted due to disasters such as earthquakes. A residential area coordinate set records the geographic coordinates of each residential area within the area, such as (110, 200) for one residential complex and (120, 210) for another.

[0025] Step S115: performing time stamp alignment processing on the disaster image acquisition data, the disaster environment sensing data, the disaster historical response record data and the disaster geographic distribution topology data to obtain the real-time disaster monitoring data set.

[0026] After acquiring the aforementioned data, timestamp alignment is necessary because the data may have been collected at different times. For example, disaster image data may be collected at 10:00 AM, a sensor in the disaster environment sensor data may have collected data at 10:10 AM, an event in the historical disaster response record data may have occurred at 9:00 AM, and the geographic distribution topology data was last updated at 9:30 AM. Through timestamp alignment, these data are organized according to a unified time standard, ensuring temporal consistency and ultimately resulting in a complete set of real-time disaster monitoring data.

[0027] Step S120: extracting features from the real-time disaster monitoring data set to obtain a multimodal disaster feature set of the disaster scene, wherein the multimodal disaster feature set includes dynamic features of disaster images, disaster environment association features, disaster response time series features, and geographic topological distribution features.

[0028] In this embodiment, after obtaining a real-time disaster monitoring data set, it is necessary to extract features from it to obtain a multimodal disaster feature set that can reflect the disaster scenario. The specific steps are as follows:

[0029] Step S121: calling a pre-trained disaster image encoder to perform convolution feature extraction processing on the disaster image acquisition data to obtain dynamic features of the disaster image, wherein the dynamic features of the disaster image include landform deformation gradient features, thermal diffusion direction vectors, and water coverage area change rates.

[0030] Taking earthquake disasters as an example, the pre-trained disaster image encoder is trained on a large amount of disaster image data. It contains multiple convolutional layers, pooling layers, and other structures. The acquired disaster image data is first input into the encoder's first convolutional layer. Assuming the input disaster image is a 256×256 pixel color image with three channels (red, green, and blue), the input data dimensions are 256×256×3.

[0031] In the first convolutional layer, convolution operations are performed using multiple different convolution kernels. For example, 16 3×3 convolution kernels are used. Each convolution kernel slides over the input image, performing element-wise multiplication and summing operations to generate a feature map. For example, as a convolution kernel slides over the image, it multiplies and accumulates the pixel values ​​at the corresponding position in the image, generating a new pixel value and ultimately a feature map. After 16 convolution kernels are applied, 16 feature maps are generated, and the data dimensions become 254×254×16 (because 256-3+1=254).

[0032] Next, these feature maps are pooled, using, for example, 2×2 max pooling. Max pooling selects the maximum value within a 2×2 region as the output for that region, reducing the data dimension while retaining important feature information. After max pooling, the data dimensions become 127×127×16.

[0033] As the data is passed layer by layer in the encoder, the convolution and pooling operations are repeated, the dimension of the data gradually becomes smaller and the features gradually become more abstract.

[0034] To extract terrain deformation gradient features, the encoder focuses on changes in the terrain within the image. For example, by comparing images collected at different times and analyzing changes in pixel values, the encoder determines terrain deformation. Assuming that in a certain area, the change in pixel values ​​between two images is significant, the encoder quantifies this deformation by calculating the gradient of the pixel values. Horizontally, the difference between adjacent pixel values ​​is calculated, and the same calculation is performed vertically to obtain horizontal and vertical gradient values. These gradient values ​​are combined to form the terrain deformation gradient feature. For example, within a 10×10 area, the horizontal gradient values ​​are [0.1, 0.2, 0.15, ...], and the vertical gradient values ​​are [0.05, 0.12, 0.08, ...]. These values ​​are concatenated to form a 20-dimensional terrain deformation gradient feature vector.

[0035] Extracting the heat diffusion direction vector relies on the thermal distribution gradient information in the image. The encoder identifies areas of higher temperature in the image and analyzes the temperature trends in these areas. The direction of heat diffusion is determined by calculating the temperature gradients at different locations. For example, within a local area, by comparing the temperature change rates in different directions, it is found that the temperature rises fastest in a certain direction. This direction is then the primary direction of heat diffusion. This direction is represented by a vector. For example, on a two-dimensional plane, the heat diffusion direction vector can be expressed as (0.3, 0.4), where 0.3 and 0.4 represent the components in the x and y directions, respectively.

[0036] The rate of change in water coverage is calculated by comparing the areas of water bodies in images collected at different times. First, the image needs to be segmented to separate the water areas from other areas. A threshold-based segmentation method can be used. Based on the color characteristics of water bodies in the image, an appropriate threshold is set. Areas with pixel values ​​greater than this threshold are identified as water bodies. Next, the number of pixels in the water area in the images collected at different times is calculated to determine the water coverage area. Assuming that the number of pixels corresponding to the water coverage area in the first image is 1000 and that in the second image is 1100, the rate of change in water coverage is (1100 - 1000) / 1000 = 0.1.

[0037] Step S122: performing time series analysis on the disaster environment sensor data to extract disaster environment correlation features, which include the wind speed and temperature covariance matrix, the humidity and geological vibration frequency domain correlation, and the time lag correlation coefficient of multiple sensor data.

[0038] For disaster environment sensing data, wind speed fluctuation series, temperature change gradient, humidity distribution curve and geological vibration frequency spectrum are all data sequences that change with time.

[0039] When calculating the wind speed and temperature covariance matrix, first ensure that the timestamps of the wind speed fluctuation sequence and the temperature gradient sequence are aligned. Assume that the wind speed fluctuation sequence is [3, 4, 5, 3, 2] m / s, corresponding to time points t1, t2, t3, t4, and t5, and the temperature gradient sequence is [0.5, 0.6, 0.7, 0.5, 0.4] °C / hour, also corresponding to time points t1, t2, t3, t4, and t5.

[0040] First, calculate the mean of the wind speed sequence by adding up all the values ​​in the sequence and dividing by the length of the sequence: (3 + 4 + 5 + 3 + 2) / 5 = 3.4 m / s. Similarly, calculate the mean of the temperature gradient sequence: (0.5 + 0.6 + 0.7 + 0.5 + 0.4) / 5 = 0.54°C / hour.

[0041] Then, for each time point, calculate the difference between the wind speed value and the mean wind speed, as well as the difference between the temperature gradient value and the mean temperature, and multiply these two differences together. For example, at time t1, the wind speed difference is 3 - 3.4 = -0.4 m / s, and the temperature difference is 0.5 - 0.54 = -0.04°C / hour. The product of these is (-0.4) × (-0.04) = 0.016. Repeat this calculation for all time points to obtain a set of product values.

[0042] Finally, add these product values ​​and divide by the length of the sequence minus 1 to obtain the covariance value. Here, the covariance value is ((-0.4) × (-0.04) + (4-3.4) × (0.6-0.54) + (5-3.4) × (0.7-0.54) + (3-3.4) × (0.5-0.54) + (2-3.4) × (0.4-0.54)) / (5-1) = 0.07. Since we are calculating the covariance matrix, and this is a two-dimensional case, the covariance matrix is ​​[[variance of wind speed, covariance], [covariance, variance of temperature]], where the variance of wind speed and temperature are calculated similarly. The variance of wind speed is ((3-3.4)² + (4-3.4)² + (5-3.4)² + (3-3.4)² + (2-3.4)²) / (5-1) = 1.3, and the variance of temperature is ((0.5-0.54)² + (0.6-0.54)² + (0.7-0.54)² + (0.5-0.54)² + (0.4-0.54)²) / (5-1) = 0.013, so the covariance matrix is ​​[[1.3, 0.07], [0.07, 0.013]].

[0043] Calculating the frequency-domain correlation between humidity and geological vibration requires first performing frequency-domain analysis on the humidity distribution curve and the geological vibration frequency spectrum. The humidity distribution curve is converted from the time domain to the frequency domain using a Fourier transform, obtaining the humidity frequency components and their corresponding amplitudes. Similarly, the geological vibration frequency spectrum is also Fourier transformed. The correlation between these two frequency-domain signals at the same frequency components is then analyzed. For example, at a specific frequency f, the amplitude of the humidity frequency-domain signal is A1, and the amplitude of the geological vibration frequency-domain signal is A2. By calculating the amplitude relationship at multiple frequency points, the frequency-domain correlation between humidity and geological vibration is determined. This correlation can be quantified using methods such as the correlation coefficient. For example, if the average correlation coefficient calculated across multiple frequency points is 0.6, then the frequency-domain correlation between humidity and geological vibration can be expressed as 0.6.

[0044] The calculation of the time-lagged correlation coefficient of multi-sensor data is intended to analyze the time delay relationship between different sensor data. Taking wind speed and temperature data as an example, the wind speed series is first lagged by different time steps. For example, the wind speed series is lagged by one time step to obtain a new series. The correlation coefficient between this lagged series and the temperature series is then calculated. The correlation coefficient can be calculated using the Pearson correlation coefficient formula: first calculate the covariance of the two series and then divide it by the product of their standard deviations. By varying the lag time step, the correlation coefficient is calculated for different lag conditions to find the lag time step that maximizes the correlation coefficient. Assuming that the correlation coefficient with the temperature series is maximized when the wind speed series is lagged by two time steps, the time-lagged correlation coefficient for the multi-sensor data in this case is 0.8, with a lag time step of 2.

[0045] Step S123: Input the historical disaster response record data into the time series feature extraction network to generate disaster response time series features, which include a response delay time distribution histogram, a resource scheduling path optimization potential coefficient, and a regression fitting parameter of historical losses and response time.

[0046] Step S1231: performing time slicing processing on the historical disaster response record data to generate a response action sequence within an equally spaced time window.

[0047] Assume that historical disaster response data records the response within 24 hours of the disaster. Divide these 24 hours into 2-hour intervals, resulting in 12 equally spaced time windows. For each time window, carefully analyze the response actions recorded. For example, in the first 2-hour window, possible response actions might include activating the emergency command center and issuing a warning; in the second 2-hour window, response actions might include deploying rescue teams and preparing relief supplies. Organize the response actions within each time window into a sequence. Assuming there are five types of response actions, represented by numbers 1-5, the response action sequence for the first time window might be [1, 2], and for the second window, [3, 4].

[0048] Step S1232: extracting the response action type distribution vector, resource scheduling quantity change curve and loss accumulation rate in each time window.

[0049] In each time window, for the response action type distribution vector, count the number of times each response action type appears. Suppose that in a certain time window, the number of times each of the five response action types appears is 2, 1, 3, 0, and 1, respectively. Then the response action type distribution vector is [2, 1, 3, 0, 1].

[0050] To extract a resource dispatch quantity change curve, we need to focus on the dispatch status of different resources within each time window. For example, let's assume that within a 2-hour window, 2 rescue vehicles are dispatched at minute 0, 3 at minute 30, 1 at minute 60, 2 at minute 90, and 1 at minute 120. By connecting these data points, with time as the horizontal axis and the dispatch quantity as the vertical axis, we can form a resource dispatch quantity change curve.

[0051] The cumulative loss rate is calculated by comparing the losses at the beginning and end of a time window. For example, if the estimated losses from a disaster are 1 million yuan at the beginning of a time window and 1.2 million yuan at the end, the cumulative loss rate is (120 - 100) / 2 = 100,000 yuan per hour.

[0052] Step S1233: calling a long short-term memory network to perform temporal dependency modeling on the response action sequence to generate a hidden state vector.

[0053] A long short-term memory (LSTM) network consists of an input gate, a forget gate, an output gate, and a cell state. The response action sequence generated earlier is sequentially input into the LSTM network. Assume that the response action sequence has 12 elements (corresponding to 12 time windows), each element being a response action type distribution vector with a dimension of 5.

[0054] At the first time step, the input is the response action type distribution vector for the first time window. The LSTM network first uses a forget gate to determine how much information from the previous time step's cell state to retain. The forget gate uses a sigmoid function to linearly combine the input vector with the hidden state vector from the previous time step, then uses the sigmoid function to map the result to a value between 0 and 1. Assuming the hidden state vector from the previous time step is zero (because it is the first time step), and the input vector is [2, 1, 3, 0, 1], the forget gate's weight matrix and bias are trained. After calculation, the forget gate outputs a vector between 0 and 1, which determines the proportion of the cell state to be retained.

[0055] Next, the input gate determines how much new information should be added to the cell state. The input gate also uses a sigmoid function and a tanh function. The sigmoid function determines which values ​​need to be updated, while the tanh function creates a new candidate vector. These two results are combined to update the cell state.

[0056] Finally, the output gate determines the hidden state vector for the current time step. The output gate also uses the sigmoid function to linearly combine the input vector and the hidden state vector from the previous time step. The sigmoid function then generates a vector between 0 and 1. This vector is then multiplied by the cell state vector processed by the tanh function to obtain the hidden state vector for the current time step.

[0057] As time steps progress, the above process is repeated, ultimately resulting in a sequence of 12 hidden state vectors. Assuming the hidden state vector has 32 dimensions, then after processing it through the LSTM network, a 12×32 matrix is ​​obtained, with each row representing the hidden state vector for a time step.

[0058] Step S1234: Calculate the association weights between hidden state vectors of different time windows through the self-attention mechanism to generate global temporal attention features.

[0059] The self-attention mechanism allows the model to focus on relevant information in other time windows while processing the hidden state vector of each time window. For each hidden state vector, the association weight is determined by calculating its similarity with all other hidden state vectors.

[0060] First, the hidden state vector sequence is multiplied by three different weight matrices to obtain the query vector (Q), key vector (K), and value vector (V). Assume that the dimension of the hidden state vector sequence is 12×32, the dimension of the query weight matrix is ​​32×32, the dimension of the key weight matrix is ​​32×32, and the dimension of the value weight matrix is ​​32×32. After matrix multiplication, the resulting query vector sequence, key vector sequence, and value vector sequence all have a dimension of 12×32.

[0061] Then, for each query vector, its similarity with all key vectors is calculated. This similarity can be calculated using a dot product operation, which multiplies the query vector element-wise with the key vector and sums the results. For example, the dot product of the first query vector with the first key vector is one value, and the dot product with the second key vector is another value. This calculation results in a similarity vector containing 12 values.

[0062] Next, the similarity vector is normalized and converted into a probability distribution using the softmax function to obtain the association weight vector. Each element in the association weight vector represents the degree of attention of the current query vector to the hidden state vector of other time windows.

[0063] Finally, the association weight vector is weighted and summed with the value vector sequence to obtain a new vector, which is the global temporal attention feature corresponding to the current query vector. This operation is repeated for all query vectors, ultimately resulting in a sequence of 12 global temporal attention feature vectors. Assuming that the dimension of the global temporal attention feature vector is also 32, the resulting sequence dimension is 12×32.

[0064] Step S1235: Perform weighted fusion on the global temporal attention feature and the hidden state vector of each time window to obtain the disaster response temporal feature.

[0065] Assign weights to the global temporal attention feature and the hidden state vector of each time window. Assume that the weight of the global temporal attention feature is 0.4 and the weight of the hidden state vector of each time window is 0.6.

[0066] For the hidden state vector of the first time window and the corresponding global temporal attention feature vector, multiply the hidden state vector by 0.6 and the global temporal attention feature vector by 0.4, and then add the two results to obtain the first fused vector. Assuming the hidden state vector is [h11, h12, ..., h132] and the global temporal attention feature vector is [a11, a12, ..., a132], then the fused vector is [0.6×h11+0.4×a11, 0.6×h12+0.4×a12, ..., 0.6×h132+0.4×a132].

[0067] This operation is performed sequentially on the hidden state vector and global temporal attention feature vector for each time window, ultimately resulting in a fused set of vectors. These fused vectors are concatenated in the order of the time windows to form the disaster response temporal features. Assuming each hidden state vector and global temporal attention feature is 32-dimensional, the fused vector is also 32-dimensional. The dimensionality of the concatenated disaster response temporal features depends on the number of time windows. For example, if there are 12 time windows, the dimensionality of the disaster response temporal features is 32 × 12 = 384.

[0068] Step S124: Connectivity analysis is performed on the disaster geographic distribution topology data through a graph neural network to generate geographic topology distribution features. The geographic topology distribution features include a road network survivability score, the shortest path distance between residential areas and disaster centers, and a mapping relationship between elevation gradient and disaster diffusion speed.

[0069] Before conducting connectivity analysis, it's necessary to define the specific form of the disaster geographic distribution topology data. For example, in the case of an earthquake disaster scenario, the elevation gradient distribution map within the geographic distribution topology data can be converted into a two-dimensional matrix, with each element representing the elevation of the area. A road connectivity network can be represented by a graph structure, where nodes represent intersections or endpoints of roads, and edges represent the roads connecting these nodes. Edge attributes can include road length, width, and carrying capacity. A residential area coordinate set is a series of two-dimensional coordinate points representing the locations of individual residential areas.

[0070] In this embodiment, the disaster geographic distribution topology data is converted into a graph structure suitable for graph neural network processing. For the road connectivity network, each road intersection or endpoint is used as a node of the graph, and a certain feature vector is assigned to each node. For example, node features may include the elevation value of the node location (obtained from the elevation gradient distribution map), the surrounding population density (which can be estimated based on the coordinate set of the residential area), etc. Assuming that the elevation value ranges from 0 to 1000 meters and the population density ranges from 0 to 1000 people per square kilometer, after normalizing the elevation value and population density, the feature vector of each node can be represented as a two-dimensional vector, such as [0.2, 0.5], where 0.2 is the normalized elevation value and 0.5 is the normalized population density.

[0071] For edges connecting nodes, edge features can include road length and road type (e.g., expressway, highway, etc.). Road length can be calculated based on geographic coordinates. Assuming road lengths range from 0 to 100 kilometers, this is normalized and used as a dimension of the edge feature. Road type can be one-hot encoded. Assuming there are three road types: expressway, highway, and country road, the edge feature can be represented as a four-dimensional vector, such as [0.3, 1, 0, 0], where 0.3 is the normalized road length, 1 indicates that the road is a expressway, and the following two zeros indicate that it is not a highway or country road.

[0072] Graph neural networks update node features through a message passing mechanism. In each layer of the graph neural network, a node receives messages from its neighboring nodes and updates its own features based on these messages.

[0073] First, for each node, calculate the messages sent by its neighboring nodes. Assume the current node is v and its set of neighboring nodes is N(v). The message sent by neighbor node u to node v can be calculated as follows: concatenate the eigenvector of neighbor node u with the eigenvector of the edge (u, v), then perform a linear transformation (which can be represented as a matrix multiplication and the addition of a bias term) to obtain the message vector. For example, if the eigenvector of neighbor node u is [0.2, 0.5] and the eigenvector of the edge (u, v) is [0.3, 1, 0, 0], the concatenation yields [0.2, 0.5, 0.3, 1, 0, 0]. Assume the linear transformation matrix is ​​a 6×5 matrix W and a 5-dimensional bias vector b. After matrix multiplication and addition, a 5-dimensional message vector is obtained.

[0074] Node v then aggregates the messages sent by all its neighboring nodes. This aggregation can be done by summing or averaging, but we'll use summation as an example. The message vectors sent by all neighboring nodes are added together to obtain the aggregated message vector for node v.

[0075] Finally, node v updates its own feature vector based on the aggregated message vector. The aggregated message vector and the original feature vector of node v can be combined through another linear transformation. For example, the aggregated message vector and the original feature vector of node v are concatenated and then subjected to a linear transformation to obtain the updated feature vector.

[0076] The road network survivability score reflects the stability and reliability of the road network in the event of a disaster. It can be calculated based on the updated node and edge features of the graph neural network.

[0077] For each road (edge), consider factors such as its carrying capacity, length, and the elevation of the area in which it is located. Carrying capacity can be estimated based on the road type and width. Assume that the carrying capacity of a highway is 1,000 vehicles / hour, that of a regular road is 500 vehicles / hour, and that of a rural road is 100 vehicles / hour, and these values ​​are normalized.

[0078] Define a survivability evaluation function. For example, for an edge (u, v), its survivability score can be expressed as: survivability score = carrying capacity score × length score × elevation impact score. The carrying capacity score is the normalized carrying capacity value, the length score is calculated as 1 / normalized road length (the shorter the length, the higher the score), and the elevation impact score can be determined based on the elevation values ​​of nodes u and v. For example, if the elevation value is high, it may indicate that the area is more vulnerable to disasters, so the elevation impact score can be set to a lower value, such as 0.8; if the elevation value is low, the elevation impact score can be set to 1.

[0079] For the entire road network, the invulnerability scores of all edges are summed or weighted to obtain the road network invulnerability score. For example, if there are 10 edges in the road network and the invulnerability scores of each edge are [0.8, 0.9, 0.7, 0.6, 0.85, 0.95, 0.75, 0.65, 0.8, 0.9], then the road network invulnerability score is calculated as follows: 0.8 + 0.9 + 0.7 + 0.6 + 0.85 + 0.95 + 0.75 + 0.65 + 0.8 + 0.9 = 7.9.

[0080] The updated graph structure of the graph neural network is used in combination with graph algorithms (such as the Dijkstra algorithm) to calculate the shortest path distance between residential areas and disaster centers.

[0081] First, determine the location of the disaster center. Assume that the disaster center corresponds to a node c in the graph. For each residential area coordinate, find the node closest to the residential area in the graph and use it as the node representing the residential area.

[0082] Taking the Dijkstra algorithm as an example, a distance array is initialized, with the distance to the disaster center node c set to 0 and the distances to all other nodes set to infinity. Then, the node with the smallest distance from the distance array is selected, marked as visited, and the distances to its neighboring nodes are updated. If the distance to the neighboring node through the current node is smaller than the original distance to the neighboring node, the neighboring node distance is updated.

[0083] Repeat the above steps until all nodes have been visited or the shortest path distances to all residential area representative nodes are found. For example, suppose there are 5 residential area representative nodes, and the Dijkstra algorithm calculates their shortest path distances to the disaster center as [5, 8, 3, 6, 7] kilometers respectively.

[0084] Finally, the relationship between the elevation gradient distribution map and the disaster diffusion speed is analyzed. The elevation gradient distribution map can be divided into multiple regions, each with a similar elevation value range.

[0085] For each region, the disaster diffusion rate data within the region is collected. The disaster diffusion rate can be obtained from historical disaster data or simulation data. For example, in earthquake disasters, the disaster diffusion rate can be estimated based on the propagation speed of seismic waves and the impact range.

[0086] Assume that the elevation gradient distribution map is divided into five regions, with elevation values ​​in each region ranging from 0-200 m, 200-400 m, 400-600 m, 600-800 m, and 800-1000 m. The disaster diffusion speed within each region is calculated, yielding the following data: the average disaster diffusion speed in Region 1 (0-200 m) is 10 m / s, in Region 2 (200-400 m) it is 8 m / s, in Region 3 (400-600 m) it is 6 m / s, in Region 4 (600-800 m) it is 4 m / s, and in Region 5 (800-1000 m) it is 2 m / s.

[0087] By fitting these data, we can establish a mapping relationship between elevation gradient and disaster spread rate. This can be done using methods such as linear regression. Assume the resulting linear regression equation is: Disaster spread rate = -0.01 × elevation + 12. Thus, for any given elevation value, this mapping relationship can be used to estimate the corresponding disaster spread rate.

[0088] Step S125: aligning and splicing the dynamic features of the disaster image, the disaster environment association features, the disaster response time series features, and the geographic topology distribution features to generate the multimodal disaster feature set.

[0089] Before performing feature dimension alignment and splicing, it is necessary to clarify the dimensions of each feature. Assume that the dimension of the dynamic features of disaster images is 50, the dimension of the disaster environment association features is 30, the dimension of the disaster response time series features is 384, and the dimension of the geographic topology distribution features is 20.

[0090] Specifically, you can check whether the dimensions of each feature are suitable for concatenation. If the dimensions of some features are too high or too low, you may need to reduce or increase the dimensions.

[0091] For dynamic features of disaster images with high dimensionality, methods such as principal component analysis (PCA) can be used for dimensionality reduction. PCA calculates the covariance matrix of the features, finds the principal component directions, and then projects the features onto these principal component directions, thereby reducing the dimensionality of the features. Suppose PCA is used to reduce the dynamic features of disaster images from 50 dimensions to 40 dimensions.

[0092] For low-dimensional features, such as geographic topological distribution features, dimensionality can be increased by padding zero vectors. For example, if the geographic topological distribution feature is increased from 20 dimensions to 30 dimensions, 10 zero values ​​are padded after the original feature vector.

[0093] After dimensional alignment, the disaster image dynamic features, disaster environment correlation features, disaster response time series features, and geographic topological distribution features are spliced ​​in a certain order. For example, the disaster image dynamic features, disaster environment correlation features, disaster response time series features, and geographic topological distribution features are spliced ​​in this order.

[0094] The dimensionality of the concatenated multimodal disaster feature set is 40 + 30 + 384 + 30 = 484. Each feature vector is concatenated sequentially to form a new 484-dimensional feature vector, which represents the multimodal disaster feature set for the disaster scenario. This integration of different types of features provides more comprehensive information for subsequent disaster scenario matching analysis and decision-making.

[0095] Step S130: calling a pre-trained multimodal dynamic decision model, performing disaster scenario matching analysis on the multimodal disaster feature set, and generating a disaster scenario matching score and a disaster response optimization strategy set for the target disaster area.

[0096] In this embodiment, the pre-trained multimodal dynamic decision-making model is trained based on a large amount of disaster case data. It includes multiple feature matching channels for processing different types of disaster features. After obtaining the multimodal disaster feature set, it is input into the model for disaster scenario matching analysis. The specific steps are as follows:

[0097] Step S131: input the multimodal disaster feature set into the multimodal dynamic decision model, and calculate the image matching coefficient set, environment matching coefficient set, time series matching coefficient set and topology matching coefficient set between the multimodal disaster feature set and pre-stored disaster cases.

[0098] Step S1311: Input the multimodal disaster feature set into the image feature matching channel of the multimodal dynamic decision model, calculate the pixel gradient similarity between the dynamic features of the disaster image and each case in the image feature library of pre-stored disaster cases, and generate an image matching coefficient set.

[0099] Assume that the dynamic feature of a disaster image is a 40-dimensional vector, there are 100 cases in the image feature library of pre-stored disaster cases, and the image feature of each case is also a 40-dimensional vector.

[0100] In the image feature matching channel, for the dynamic feature vector of the disaster image and each case vector in the image feature library, their pixel gradient similarity is calculated. The specific calculation process is as follows:

[0101] First, subtract the elements at corresponding positions in the two vectors to obtain a difference vector. For example, if the dynamic feature vector of a disaster image is [a1, a2, …, a40] and the vector of a case in the image feature library is [b1, b2, …, b40], the difference vector is [|a1-b1|, |a2-b2|, …, |a40-b40|].

[0102] Then, each element in the difference vector is squared to obtain the square difference vector [(a1-b1)^2, (a2-b2)^2, …, (a40-b40)^2].

[0103] Next, all elements in the squared difference vector are added together to obtain a sum value. Let the sum value be S.

[0104] Finally, to get the similarity, we use the idea of ​​an inverse proportional function and divide 1 by (1+S) to get the pixel gradient similarity. For example, if S=5, then the pixel gradient similarity is 1 / (1+5)=1 / 6≈0.17.

[0105] The above calculation is performed on 100 cases in the image feature library in turn, and an image matching coefficient set containing 100 similarity values ​​is obtained, such as [0.17, 0.2, 0.15, ...].

[0106] Step S1312: Input the disaster environment-related features into the environmental feature matching channel of the multimodal dynamic decision-making model, extract the environmental parameter covariance matrix of each case in the environmental feature library of pre-stored disaster cases, calculate the spectral norm ratio of the covariance matrix, and generate a set of environmental matching coefficients.

[0107] Assume that the disaster environment-related characteristics are represented by a 30-dimensional vector, and there are 100 cases in the environmental feature library of pre-stored disaster cases, and each case is also represented by a 30-dimensional vector.

[0108] First, for each case vector in the disaster-environment correlation database and the disaster-environment correlation feature vector, we calculate the environmental parameter covariance matrix. Taking the disaster-environment correlation feature vector as an example, assume there is a series of observations, which are arranged in a certain time sequence into a matrix X (assuming the number of observations is n, then X is an n×30 matrix).

[0109] The steps to calculate the covariance matrix are as follows:

[0110] First, calculate the mean vector of each feature. For each column of X, add all the elements of the column and divide it by the number of observations n to get a 30-dimensional mean vector μ.

[0111] Then, the mean vector μ is subtracted from each row in the matrix X to obtain a new matrix X'.

[0112] Next, calculate the covariance matrix C = (X'^T*X') / (n-1), where X'^T represents the transposed matrix of X'.

[0113] Similarly, the covariance matrix is ​​calculated for each case vector in the environmental feature library.

[0114] Next, calculate the spectral norm of the covariance matrix. The spectral norm is the largest singular value of a matrix and can be calculated using singular value decomposition (SVD). For a matrix A, perform the singular value decomposition A = UΣV^T, where Σ is a diagonal matrix whose diagonal elements are the singular values ​​of matrix A. The spectral norm is the largest diagonal element in Σ.

[0115] Calculate the spectral norm ratio of the covariance matrix C1 of the disaster environment-related features and the covariance matrix C2 of a case in the environmental feature database. Assume that the spectral norm of C1 is σ1 and the spectral norm of C2 is σ2, and the spectral norm ratio is σ1 / σ2.

[0116] The above calculations are performed on 100 cases in the environmental feature library in turn to obtain an environmental matching coefficient set containing 100 spectral norm ratios, such as [0.8, 0.9, 0.75, ...].

[0117] Step S1313: Input the disaster response time series characteristics into the time series feature matching channel of the multimodal dynamic decision model, perform dynamic time regularization alignment with the response delay time distribution of each case in the time series feature library of pre-stored disaster cases, and generate a set of time series matching coefficients.

[0118] Assume that the disaster response time series feature is a 384-dimensional vector, there are 100 cases in the pre-stored disaster case time series feature library, and the response delay time distribution of each case is represented by a time series. Assume that the length of the time series is m.

[0119] The Dynamic Time Warping (DTW) algorithm is used to calculate the similarity between two time series. In this step, the response delay-related portion of the disaster response time series features is extracted to form a time series T1. The response delay distribution of a case in the pre-stored disaster case time series feature database is formed into a time series T2.

[0120] The specific steps of the DTW algorithm are as follows:

[0121] Create an (m+1)×(m+1) matrix D to store intermediate calculation results. Initialize D[0,0]=0. For i>0, D[i,0]=infinity, and for j>0, D[0,j]=infinity.

[0122] Then, for i from 1 to m, j from 1 to m, calculate the value of D[i, j]. The calculation formula of D[i, j] is: D[i, j] = |T1[i-1]-T2[j-1]| + min(D[i-1, j], D[i, j-1], D[i-1, j-1]).

[0123] Finally, the DTW distance is D[m,m]. To get the similarity, divide 1 by (1 + DTW distance) to get a similarity value.

[0124] DTW calculations are performed on 100 cases in the time series feature library in turn to obtain a time series matching coefficient set containing 100 similarity values, such as [0.7, 0.65, 0.8, ...].

[0125] Step S1314: Input the geographic topological distribution features into the topological feature matching channel of the multimodal dynamic decision model, compare the road network survivability score and the shortest path distance to the residential area of ​​each case in the topological feature library of pre-stored disaster cases, and generate a set of topological matching coefficients.

[0126] Assume that the geographic topological distribution characteristics are represented by a 30-dimensional vector, which contains information such as the road network survivability score and the shortest path distance to residential areas. There are 100 cases in the topological feature library of pre-stored disaster cases, and each case also contains the road network survivability score and the shortest path distance to residential areas.

[0127] For the road network invulnerability score R1 in the geographic topology distribution feature and the road network invulnerability score R2 of a case in the topology feature library, calculate the absolute value of their difference |R1-R2|. For the shortest path distance D1 in the geographic topology distribution feature and the shortest path distance D2 of a case in the topology feature library, similarly calculate the absolute value of their difference |D1-D2|.

[0128] Then, weights are assigned to the road network invulnerability score and the shortest path distance to residential areas respectively. Assume that the weight of the road network invulnerability score is 0.6 and the weight of the shortest path distance to residential areas is 0.4.

[0129] Comprehensive score = 0.6×|R1-R2|+0.4×|D1-D2|.

[0130] To get the similarity, divide 1 by (1 + overall score) to get a similarity value.

[0131] The above calculations are performed on the 100 cases in the topological feature library in turn to obtain a set of topological matching coefficients containing 100 similarity values, such as [0.85, 0.9, 0.8, ...].

[0132] Step S132: performing normalized weighted summation on the image matching coefficient set, the environment matching coefficient set, the time series matching coefficient set, and the topology matching coefficient set to generate a global matching score for each pre-stored disaster case.

[0133] First, normalize the image matching coefficient set, the environment matching coefficient set, the time series matching coefficient set, and the topology matching coefficient set. Taking the image matching coefficient set as an example, find the maximum value (max1) and the minimum value (min1) in the set. For each element x in the set, the normalized element x' = (x - min1) / (max1 - min1). Similarly, normalize the environment matching coefficient set, the time series matching coefficient set, and the topology matching coefficient set.

[0134] Assume that the weights of the image matching coefficient set, the environment matching coefficient set, the temporal matching coefficient set, and the topology matching coefficient set are 0.3, 0.2, 0.2, and 0.3, respectively.

[0135] For the first case in the pre-stored disaster cases, the normalized image matching coefficient is 0.8, the normalized environment matching coefficient is 0.7, the normalized time series matching coefficient is 0.6, and the normalized topology matching coefficient is 0.8.

[0136] Global matching score = 0.3 × 0.8 + 0.2 × 0.7 + 0.2 × 0.6 + 0.3 × 0.8

[0137] =0.24+0.14+0.12+0.24

[0138] =0.74.

[0139] The 100 cases in the pre-stored disaster cases are calculated in turn to obtain the global matching score of each case, forming a set of 100 scoring values, such as [0.74, 0.78, 0.72, ...].

[0140] Step S133: Arrange the pre-stored disaster cases in descending order according to the global matching scores, select cases with global matching scores higher than a set scoring threshold as a similar case set, and extract the historical response strategy and loss reduction rate corresponding to each case in the similar case set.

[0141] Assume that the score threshold is set to 0.7. Arrange the global matching score set from largest to smallest, for example, the arranged set is [0.85, 0.8, 0.78, 0.76, 0.74, 0.72, ...].

[0142] Select cases with scores higher than 0.7. Assume that 20 cases are obtained after screening, and these cases constitute a set of similar cases.

[0143] For each case in the similar case set, the corresponding historical response strategy and loss reduction rate are extracted from a pre-existing database. Historical response strategies may include resource allocation plans, evacuation plans, etc. The loss reduction rate is the percentage reduction in disaster losses after adopting the historical response strategy compared to the situation without the strategy. For example, the historical response strategy for the first case was to prioritize evacuating residents to a safe area, resulting in a loss reduction rate of 30%; the historical response strategy for the second case was to quickly deploy relief supplies, resulting in a loss reduction rate of 25%, and so on.

[0144] Step S134: Adaptively adjust the strategy parameters of the historical response strategy and the current disaster environment parameters, including adjusting the starting coordinates of the resource scheduling path, the evacuation route density coefficient and the equipment deployment time window, to generate a candidate optimization strategy set.

[0145] For each historical response strategy in a set of similar cases, parameter adjustments are made.

[0146] Taking the adjustment of the starting coordinates of the resource scheduling path as an example, assuming that the starting coordinates of the resource scheduling path of a certain historical response strategy are (x1, y1), in the current disaster environment, based on the location of available resources and road conditions, it is found that the coordinates closer to the starting point and more suitable as the starting point are (x2, y2), then the starting coordinates of the resource scheduling path are adjusted to (x2, y2).

[0147] Adjustments to the evacuation route density coefficient are made based on factors such as the current disaster area's population density and road carrying capacity. Assuming the evacuation route density coefficient in the historical response strategy was 0.8, given that the current disaster area's population density is higher than in historical cases and road carrying capacity is limited, the evacuation route density coefficient was adjusted to 0.7 after evaluation.

[0148] When adjusting the device deployment window, consider the current disaster stage and device availability. For example, if the historical response strategy planned to deploy a device between 10:00 AM and 11:00 AM, but the current disaster is developing rapidly and the device is available between 9:00 AM and 10:00 AM, adjust the device deployment window to 9:00 AM and 10:00 AM.

[0149] The above parameter adjustments are performed on the 20 historical response strategies in the similar case set to obtain 20 adjusted strategies, which constitute the candidate optimization strategy set.

[0150] Step S135: Calculate the execution priority score of each candidate optimization strategy based on the loss reduction rate and the adjusted strategy parameters, use the candidate optimization strategy set sorted by the execution priority score as the disaster response optimization strategy set, and use the global matching score as the disaster scenario matching score.

[0151] Assign weights to the loss reduction rate and the adjusted strategy parameters respectively. Assume that the weight of the loss reduction rate is 0.6 and the weight of the comprehensive evaluation score of the adjusted strategy parameters is 0.4.

[0152] For each strategy in the candidate optimization strategy set, the adjusted comprehensive evaluation score of the strategy parameters can be calculated by evaluating factors such as the rationality of the resource scheduling path, the feasibility of the evacuation route, and the timeliness of equipment deployment. Assuming the loss reduction rate of the first candidate optimization strategy is 30%, the adjusted comprehensive evaluation score of the strategy parameters is 0.8.

[0153] Execution priority score = 0.6 × 30% + 0.4 × 0.8

[0154] =0.18+0.32

[0155] =0.5.

[0156] The 20 strategies in the candidate optimization strategy set are calculated sequentially to obtain an execution priority score for each strategy. The candidate optimization strategy set is sorted from high to low based on execution priority score. The sorted candidate optimization strategy set becomes the disaster response optimization strategy set. At the same time, the previously calculated global matching score is used as the disaster scenario matching score. For example, for the first candidate optimization strategy, its corresponding disaster scenario matching score is the previously calculated global matching score of 0.85 for that case.

[0157] Step S140: Filtering a target optimization strategy from the disaster response optimization strategy set based on the disaster scenario matching score, generating a disaster emergency decision instruction set, and sending the disaster emergency decision instruction set to a disaster response terminal.

[0158] In this embodiment, after obtaining the disaster scenario matching score and the disaster response optimization strategy set, it is necessary to screen the target optimization strategy based on the disaster scenario matching score, and generate a corresponding disaster emergency decision-making instruction set and send it to the disaster response terminal. The specific steps are as follows:

[0159] Step S141: Determine the weight distribution ratio of each candidate optimization strategy in the disaster response optimization strategy set according to the difference between the disaster scenario matching score and the set score threshold.

[0160] Assuming a scoring threshold of 0.7 and different disaster scenario matching scores, each candidate optimization strategy in the set of disaster response optimization strategies is weighted based on the difference between the disaster scenario matching score and the set scoring threshold. For example, when the disaster scenario matching score is 0.8, the difference is 0.8-0.7=0.1. For the first candidate optimization strategy, based on the difference and the preset weighting rule, its weight is determined to be 0.3; the second candidate optimization strategy has a weight of 0.25; the third candidate optimization strategy has a weight of 0.2; the fourth candidate optimization strategy has a weight of 0.15; and the fifth candidate optimization strategy has a weight of 0.1. The weighting rule here can be a linear distribution based on the size of the difference or a more complex nonlinear function, depending on the actual application requirements and model training results.

[0161] Step S142: Extract the resource scheduling path starting point coordinates, evacuation route density coefficient and equipment deployment time window of each strategy in the candidate optimization strategy set, and perform conflict detection with the available resource coordinate set, real-time traffic density data and equipment status data of the current disaster response terminal.

[0162] From the candidate optimization strategy set, key parameters for each strategy are extracted, such as the starting coordinates of the resource scheduling path, the evacuation route density coefficient, and the equipment deployment time window. Taking the starting coordinates of the resource scheduling path as an example, assume that the starting coordinates of the resource scheduling path for the first candidate optimization strategy are (110, 210), the starting coordinates for the second candidate optimization strategy are (120, 220), and so on. Simultaneously, the set of available resource coordinates for the current disaster response terminal is obtained. Assume that there are three available resource points in this set, with coordinates at (105, 205), (115, 215), and (125, 225). Each candidate optimization strategy has a corresponding value for the evacuation route density coefficient, such as 0.9 for the first candidate and 0.85 for the second candidate. Real-time traffic density data can be obtained from traffic monitoring equipment. For example, the real-time traffic density of a road is 50 vehicles per kilometer. Equipment status data includes information such as whether the equipment is available and the time of availability. For example, a certain rescue equipment is available from 11:00 AM to 2:00 PM. These parameters are compared to detect any conflicts. For example, if the starting coordinates of the resource scheduling path of a candidate optimization strategy are far apart from the coordinates of all available resources, it may mean that resource allocation is difficult and there is a conflict; if the real-time traffic density of the evacuation route exceeds the range allowed by the evacuation route density coefficient of the candidate optimization strategy, there is also a conflict; if the equipment deployment time window does not match the available time in the equipment status data, there is also a conflict.

[0163] Step S143: For candidate optimization strategies with resource coordinate conflicts, replan the path starting point coordinates to the nearest available resource coordinates, and update the path length and estimated arrival time.

[0164] When a resource coordinate conflict is detected for a candidate optimization strategy, the path's starting coordinates need to be replanned. Suppose the original resource scheduling path's starting coordinates for the first candidate optimization strategy are (110, 210). After comparing the available resource coordinates to the set of available resource coordinates, the closest available resource coordinates are (115, 215). The path's starting coordinates for this candidate optimization strategy are then updated to (115, 215). Next, the path length and estimated time of arrival (ETA) are updated. The path length can be calculated using a geographic information system (GIS) based on the new starting coordinates and the coordinates of the target disaster area. Suppose the original path length was 20 kilometers. After replanning, the GIS calculates the new path length to be 18 kilometers. The estimated time of arrival (ETA) can be calculated based on the path length and the average speed of the transportation vehicle. Assuming the average speed is 60 kilometers per hour, the original ETA is 20 ÷ 60 × 60 = 20 minutes, while the updated ETA is 18 ÷ 60 × 60 = 18 minutes.

[0165] Step S144: For an evacuation route with excessive traffic density, the route is divided into multiple sub-segments according to the density coefficient threshold, and an alternative route identifier and a turn instruction are allocated to each sub-segment.

[0166] For evacuation routes with excessive traffic density, they are processed according to the evacuation route density coefficient threshold. Assume that the evacuation route density coefficient threshold is 40 vehicles per kilometer, and the real-time traffic density of a certain evacuation route is 50 vehicles per kilometer, which exceeds the threshold. Divide this evacuation route into multiple sub-segments according to certain rules, such as according to milestones or key intersections of the road. Suppose it is divided into three sub-segments. Assign an alternative route identifier and turning instructions to each sub-segment. The alternative route identifier can be a unique number. For example, the alternative route identifier of sub-segment 1 is R001, that of sub-segment 2 is R002, and that of sub-segment 3 is R003. Turning instructions can be determined based on geographic information systems and traffic monitoring data. For example, at a certain intersection in sub-segment 1, you need to turn left to enter the alternative route R001.

[0167] Step S145: For the policy in which the device deployment time window does not match the device status, the deployment sequence is reallocated according to the device available time, and a device deployment instruction with aligned timestamps is generated.

[0168] When the device deployment window for a candidate optimization policy doesn't match the device status, the deployment sequence needs to be reassigned. For example, suppose a candidate optimization policy plans to deploy a device at 10:00 AM, but the device's availability is from 11:00 AM to 2:00 PM. Based on the device's availability, adjust the device's deployment time to 11:00 AM and reschedule the deployment sequence for other devices to ensure the rationality of the entire deployment process. Generate device deployment instructions with aligned timestamps, for example, specifying that device 1 starts deployment at 11:00 AM and device 2 starts deployment at 11:30 AM.

[0169] Step S146: Sort the updated candidate optimization strategies by execution priority score, select the top N strategies to generate the disaster emergency decision instruction set, wherein the disaster emergency decision instruction includes the adjusted resource scheduling path coordinate sequence, evacuation sub-section steering instructions and equipment deployment timestamp.

[0170] After the conflict resolution and adjustments described above, the candidate optimization strategies are sorted according to their execution priority scores. Assume that the execution priority scores, from highest to lowest, are 0.5 for the first candidate optimization strategy, 0.45 for the second, 0.4 for the third, and so on. Select the top N strategies (assuming N is 3 here), i.e., the three strategies with the highest execution priority scores. Extract the adjusted resource scheduling path coordinate sequence, evacuation sub-section turn instructions, and device deployment timestamps from these three strategies to generate a set of disaster emergency decision instructions. For example, the resource scheduling path coordinate sequence for the first strategy is (115, 215), (120, 220), (125, 225); the evacuation sub-section turn instruction is to turn left at intersection A; and the device deployment timestamp is 11:00 AM. This information is integrated into the set of disaster emergency decision instructions.

[0171] Step S147: Encapsulate the disaster emergency decision instruction set into a disaster response terminal protocol format, add the execution time identifier and retry limit parameters of each disaster emergency decision instruction, and send it to the corresponding disaster response terminal.

[0172] The generated disaster emergency decision-making instruction set is encapsulated according to the disaster response terminal protocol format. The disaster response terminal protocol format specifies the transmission format and data structure of the instructions to ensure that the disaster response terminal can correctly parse and execute these instructions. An execution time identifier is added to each disaster emergency decision-making instruction, for example, stipulating that a certain instruction must be executed before 3 pm. At the same time, a retry limit parameter is added. Assuming that the retry limit for a certain instruction is 3 times, if the instruction fails during execution, it can be retried up to 3 times. Finally, the encapsulated disaster emergency decision-making instruction set is sent to the corresponding disaster response terminal. These terminals can be rescue vehicles, rescue personnel's handheld devices, etc.

[0173] Step S150: Acquire the decision execution feedback data returned by the disaster response terminal, and perform incremental parameter optimization processing on the multimodal dynamic decision model based on the decision execution feedback data.

[0174] In this embodiment, after the disaster emergency decision instruction set is sent to the disaster response terminal, it is necessary to obtain decision execution feedback data and perform incremental parameter optimization on the multimodal dynamic decision model. The specific steps are as follows:

[0175] Step S151: receiving the instruction execution status parameters uploaded by the disaster response terminal, wherein the instruction execution status parameters include the actual arrival time deviation value of the resource scheduling path, the real-time congestion coefficient of the evacuation sub-section, and the actual startup delay time of the equipment deployment.

[0176] The disaster response terminal will upload the instruction execution status parameters. Taking the actual arrival time deviation value of the resource scheduling path as an example, assuming that the estimated arrival time of a resource scheduling instruction is 11:00 a.m. and the actual arrival time is 11:10 a.m., then the actual arrival time deviation value is 10 minutes. The real-time congestion coefficient of the evacuation sub-section can be obtained through sensors installed on the road or real-time feedback from vehicles. For example, the real-time congestion coefficient of an evacuation sub-section is 0.8, indicating that the congestion level of the section is high. The actual startup delay time of the equipment deployment refers to the difference between the actual startup time of the equipment and the startup time specified in the instruction. Assuming that the instruction stipulates that the equipment starts at 11:00 a.m. and it actually starts at 11:15 a.m., then the actual startup delay time is 15 minutes.

[0177] Step S152: Calculate the absolute error between the actual arrival time deviation value and the instruction estimated time, the ratio of the real-time congestion coefficient to the preset threshold, and the difference between the actual startup delay time and the deployment timestamp to generate a policy execution deviation indicator set.

[0178] For the absolute error between the actual arrival time and the commanded estimated time, using the example above, the estimated arrival time is 11:00 AM, or 660 minutes (calculated from midnight), and the actual arrival time is 11:10 AM, or 670 minutes. The absolute error is 670 - 660 = 10 minutes. For the ratio of the real-time congestion coefficient to the preset threshold, assuming the preset threshold is 0.6 and the real-time congestion coefficient of a particular evacuation sub-segment is 0.8, the ratio is 0.8 ÷ 0.6 = 1.33. For the difference between the actual start delay and the deployment timestamp, assuming the deployment timestamp is 11:00 AM, or 660 minutes, and the actual start time is 11:15 AM, or 675 minutes, the difference is 675 - 660 = 15 minutes. These calculation results are combined to generate a set of policy execution deviation indicators, for example, [10, 1.33, 15].

[0179] Step S153: Collect real-time environmental data of the current disaster area, including the latest sampling value of the wind speed fluctuation sequence, the current slope of the temperature change gradient and the peak displacement of the geological vibration frequency spectrum, and generate environmental status update parameters.

[0180] Collect real-time environmental data for the current disaster area. The latest sampled value of the wind speed fluctuation sequence can be obtained from a wind speed sensor. Assume the latest sampled value is 5 m / s. The current slope of the temperature gradient can be determined by analyzing temperature change data over a period of time. For example, if the temperature rose from 20°C to 22°C over the past hour, the current slope of the temperature gradient is (22 - 20) ÷ 1 = 2°C / hour. The peak displacement of the geological vibration frequency spectrum can be measured using equipment such as a seismograph. Assume the peak displacement is 0.5 cm. Combine this data to generate environmental state update parameters, such as [5, 2, 0.5].

[0181] Step S154: Input the strategy execution deviation index set and the environmental state update parameter into the feedback learning channel of the multimodal dynamic decision model, and calculate the historical matching score correction amount of each candidate optimization strategy, where the historical matching score correction amount is a linear combination of the deviation index and the environmental parameter.

[0182] The policy execution deviation indicator set and the environment state update parameters are input into the feedback learning channel of the multimodal dynamic decision-making model. Assume that the policy execution deviation indicator set is [10, 1.33, 15] and the environment state update parameters are [5, 2, 0.5]. Assign corresponding weights to each parameter. For example, the weights for the policy execution deviation indicator set are 0.4, 0.3, and 0.3, respectively, and the weights for the environment state update parameters are 0.2, 0.3, and 0.5, respectively. For each candidate optimization policy, calculate its historical matching score correction using a linear combination method: 10 × 0.4 + 1.33 × 0.3 + 15 × 0.3 + 5 × 0.2 + 2 × 0.3 + 0.5 × 0.5 = 4 + 0.399 + 4.5 + 1 + 0.6 + 0.25 = 10.749. Calculate the historical matching score correction for each candidate optimization policy in turn.

[0183] Step S155: Adjust the pixel gradient similarity weight of the image feature matching channel, the covariance matrix spectral norm weight of the environmental feature matching channel, the dynamic time warping alignment weight of the temporal feature matching channel, and the anti-destruction score weight of the topological feature matching channel in the multimodal dynamic decision model according to the historical matching score correction amount.

[0184] The weights of each channel in the multimodal dynamic decision model are adjusted based on the calculated historical matching score correction. Assuming the historical matching score correction is 10.749, the pixel gradient similarity weight for the image feature matching channel, originally 0.3, is adjusted to 0.3 + 10.749 × 0.01 = 0.40749 based on a specific adjustment rule. The covariance matrix spectral norm weight for the environmental feature matching channel, originally 0.2, is adjusted to 0.2 + 10.749 × 0.005 = 0.253745. The dynamic time warping alignment weight for the temporal feature matching channel, originally 0.2, is adjusted to 0.2 + 10.749 × 0.008 = 0.285992. The invulnerability score weight for the topological feature matching channel, originally 0.3, is adjusted to 0.3 + 10.749 × 0.012 = 0.428988. The adjustment rules here can be set according to the model's training and actual application to ensure that the model can better adapt to new disaster scenarios and decision-making feedback.

[0185] Step S156: Use a sliding window mechanism to update the image feature library, environmental feature library, time series feature library and topology feature library of the pre-stored disaster cases, and store the current disaster feature set and the adjusted strategy parameters as new cases in the corresponding feature libraries.

[0186] Step S1561: setting the maximum number of storage cases in the feature library of the pre-stored disaster cases, and starting the sliding window elimination mechanism when the total number of newly added cases exceeds the limit.

[0187] Set the maximum number of cases that can be stored in the image feature library, environmental feature library, temporal feature library, and topological feature library for pre-stored disaster cases. Assume that the maximum number of cases that can be stored in each feature library is 100. When the current disaster feature set and adjusted policy parameters are added to the feature library as new cases, if the total number of cases exceeds 100, the sliding window elimination mechanism is activated.

[0188] Step S1562: Calculate the product of the most recent access timestamp of each historical case and the global matching score as the case activity indicator.

[0189] For each historical case in the pre-stored disaster case database, calculate the product of its most recent access timestamp and the global matching score as the case activity index. Assuming the most recent access timestamp of a historical case is 1000 (calculated with a start time of 0) and the global matching score is 0.8, the case activity index for this case is 1000 × 0.8 = 800. Calculate the case activity index for each historical case in turn.

[0190] Step S1563: Arrange historical cases in ascending order according to the case activity index, and eliminate the M cases with the lowest case activity index to free up storage space.

[0191] Sort all historical cases in ascending order by their activity index. Assuming M is 5, the five cases with the lowest activity indexes are eliminated. For example, after sorting, the five cases with the lowest activity indexes are 100, 200, 300, 400, and 500, respectively. These five cases are deleted from the corresponding feature database to free up storage space.

[0192] Step S1564: Encode the disaster image dynamic features, disaster environment association features, disaster response time series features, and geographic topology distribution features of the current disaster feature set into standardized feature vectors respectively.

[0193] Each feature in the current disaster feature set is encoded into a standardized feature vector. Taking the dynamic features of disaster images as an example, assuming they include landform deformation gradient features, thermal diffusion direction vectors, and water coverage area change rates, these features are converted into standardized feature vectors according to certain encoding rules. For example, the landform deformation gradient feature is 0.1 horizontally and 0.2 vertically, the thermal diffusion direction vector is (0.3, 0.4), and the water coverage area change rate is 0.1. These values ​​are combined and normalized to obtain a standardized feature vector [0.1, 0.2, 0.3, 0.4, 0.1]. Similarly, the disaster environment association features, disaster response temporal features, and geographic topological distribution features are encoded to obtain corresponding standardized feature vectors.

[0194] Step S1565: Associating the standardized feature vector with the corresponding adjusted strategy parameters and execution deviation indicators to generate a new case data block.

[0195] The encoded standardized feature vector is then associated with the corresponding adjusted policy parameters and execution deviation indicators. For example, for the standardized feature vector of the disaster image's dynamic features, [0.1, 0.2, 0.3, 0.4, 0.1], the corresponding adjusted policy parameters are the starting coordinates of the resource scheduling path (115, 215), and the execution deviation indicator is the actual arrival time deviation of 10 minutes. This information is associated to generate a new case data block. Similarly, the standardized feature vectors of other features are processed similarly.

[0196] Step S1566: Append the new case data block to the end of the image feature library, environmental feature library, temporal feature library and topological feature library of the pre-stored disaster cases according to feature type, and update the index identifier of each corresponding feature library.

[0197] The generated new case data blocks are appended to the end of the corresponding feature library according to the feature type. For example, the new case data blocks of the dynamic features of disaster images are appended to the end of the image feature library, the new case data blocks of the disaster environment-related features are appended to the end of the environmental feature library, the new case data blocks of the disaster response time series features are appended to the end of the time series feature library, and the new case data blocks of the geographic topological distribution features are appended to the end of the topological feature library. After the appending is completed, the index identifiers of the corresponding feature libraries are updated to ensure that these new case data can be accurately located and retrieved. For example, the index identifiers of the original image feature library are from 1 to 95. After appending the new case data blocks, the index identifiers are updated to from 1 to 96. In this way, the feature library of the pre-stored disaster cases is updated, allowing the model to use the latest disaster information for subsequent decision analysis.

[0198] Throughout the dynamic disaster emergency decision-making process, various disaster scenarios can be addressed more accurately, improving the efficiency and accuracy of disaster emergency decision-making and reducing disaster losses. For example, when acquiring real-time disaster monitoring data, the system ensures data integrity and accuracy through the integrated collection of multiple data sources and timestamp alignment. During feature extraction, different extraction methods are applied to different types of data, resulting in a multimodal disaster feature set that reflects the disaster scenario. When analyzing disaster scenario matching, multi-channel matching calculations and weighted summation accurately identify similar cases and generate optimization strategies. During the decision execution and feedback optimization phases, the model is adjusted based on actual execution, enabling it to continuously adapt to new disaster scenarios and improving decision reliability.

[0199] Figure 2 The following diagram illustrates exemplary hardware and software components of a multimodal AI large model-based emergency disaster dynamic decision-making system 100, which can implement the concepts of the present application, as provided in some embodiments of the present application. For example, the processor 120 can be used in the multimodal AI large model-based emergency disaster dynamic decision-making system 100 to perform the functions described in the present application.

[0200] The emergency disaster dynamic decision-making system 100 based on a multimodal AI large model can be a general-purpose server or a special-purpose server, both of which can be used to implement the emergency disaster dynamic decision-making method based on a multimodal AI large model of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0201] For example, the emergency disaster dynamic decision-making system 100 based on the multimodal AI large model may include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the emergency disaster dynamic decision-making system 100 based on the multimodal AI large model may also include program instructions stored in ROM, RAM, or other types of non-temporary storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The emergency disaster dynamic decision-making system 100 based on the multimodal AI large model also includes an I / O interface 150 between the computer and other input and output devices.

[0202] For ease of explanation, only one processor is described in the emergency disaster dynamic decision-making system 100 based on a multimodal AI large model. However, it should be noted that the emergency disaster dynamic decision-making system 100 based on a multimodal AI large model in this application can also include multiple processors, so the steps performed by one processor described in this application can also be performed jointly or individually by multiple processors. For example, if the processor of the emergency disaster dynamic decision-making system 100 based on a multimodal AI large model executes step A and step B, it should be understood that step A and step B can also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0203] In addition, an embodiment of the present invention also provides a readable storage medium, which has computer-executable instructions preset in the readable storage medium. When the processor executes the computer-executable instructions, the above-mentioned emergency disaster dynamic decision-making method based on the multimodal AI large model is implemented.

[0204] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A dynamic decision-making method for emergency disasters based on a multimodal AI large model, characterized by: The method comprises: Acquire a real-time disaster monitoring data set for the target disaster area, wherein the real-time disaster monitoring data set includes disaster image acquisition data, disaster environment sensor data, disaster historical response record data, and disaster geographical distribution topology data; Extracting features from the real-time disaster monitoring data set to obtain a multimodal disaster feature set of the disaster scene, wherein the multimodal disaster feature set includes dynamic features of disaster images, disaster environment association features, disaster response time series features, and geographic topological distribution features; Calling a pre-trained multimodal dynamic decision model to perform disaster scenario matching analysis on the multimodal disaster feature set, and generating a disaster scenario matching score and a disaster response optimization strategy set for the target disaster area; Selecting a target optimization strategy from the disaster response optimization strategy set based on the disaster scenario matching score, generating a disaster emergency decision instruction set, and sending the disaster emergency decision instruction set to a disaster response terminal; Obtaining decision execution feedback data returned by the disaster response terminal, and performing incremental parameter optimization processing on the multimodal dynamic decision model based on the decision execution feedback data; The calling of the pre-trained multimodal dynamic decision model, performing disaster scenario matching analysis on the multimodal disaster feature set, and generating a disaster scenario matching score and a disaster response optimization strategy set for the target disaster area include: Inputting the multimodal disaster feature set into the multimodal dynamic decision model, and calculating an image matching coefficient set, an environment matching coefficient set, a time series matching coefficient set, and a topology matching coefficient set between the multimodal disaster feature set and pre-stored disaster cases; Performing normalized weighted summation on the image matching coefficient set, the environment matching coefficient set, the time series matching coefficient set, and the topology matching coefficient set to generate a global matching score for each pre-stored disaster case; Arrange the pre-stored disaster cases in descending order according to the global matching score, select cases with a global matching score higher than a set scoring threshold as a similar case set, and extract the historical response strategy and loss reduction rate corresponding to each case in the similar case set; Adaptively adjust the strategy parameters of the historical response strategy and the current disaster environment parameters, including adjusting the starting coordinates of the resource scheduling path, the evacuation route density coefficient, and the equipment deployment time window, to generate a set of candidate optimization strategies; Based on the loss reduction rate and the adjusted strategy parameters, the execution priority score of each candidate optimization strategy is calculated, the set of candidate optimization strategies sorted by the execution priority scores is used as the disaster response optimization strategy set, and the global matching score is used as the disaster scenario matching score.

2. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 1 is characterized in that: The step of obtaining a real-time disaster monitoring data set for a target disaster area includes: Invoking satellite remote sensing equipment to collect disaster image data of the target disaster area, wherein the disaster image data includes a pixel matrix of landform changes, thermal distribution gradient information, and water coverage change trajectory of the disaster area; Acquire disaster environment sensing data uploaded in real time by multiple environmental sensing nodes deployed in the target disaster area, wherein the disaster environment sensing data includes a wind speed fluctuation sequence, a temperature change gradient, a humidity distribution curve, and a geological vibration frequency spectrum; Extracting historical disaster response record data of the target disaster area from a disaster emergency database, wherein the historical disaster response record data includes historical disaster type labels, historical emergency response timestamp sequences, historical resource dispatch paths, and historical loss assessment reports; Calling a geographic information system to analyze the geographic coordinate boundaries of the target disaster area and generate disaster geographic distribution topology data, wherein the disaster geographic distribution topology data includes an elevation gradient distribution map, a road connectivity network, and a coordinate set of residential areas; The disaster image acquisition data, the disaster environment sensing data, the disaster history response record data and the disaster geographic distribution topology data are timestamp aligned to obtain the real-time disaster monitoring data set.

3. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 1 is characterized in that: The feature extraction of the real-time disaster monitoring data set to obtain a multimodal disaster feature set of the disaster scene includes: Calling a pre-trained disaster image encoder to perform convolution feature extraction on the disaster image acquisition data to obtain dynamic features of the disaster image, wherein the dynamic features of the disaster image include landform deformation gradient features, thermal diffusion direction vectors, and water coverage area change rates; Performing time series analysis on the disaster environment sensor data to extract disaster environment correlation features, wherein the disaster environment correlation features include a wind speed and temperature covariance matrix, a humidity and geological vibration frequency domain correlation, and a time lag correlation coefficient of multiple sensor data; Inputting the historical disaster response record data into a time series feature extraction network to generate disaster response time series features, wherein the disaster response time series features include a response delay time distribution histogram, a resource scheduling path optimization potential coefficient, and a regression fitting parameter of historical losses and response time; Performing connectivity analysis on the disaster geographic distribution topology data using a graph neural network to generate geographic topology distribution features, including a road network resilience score, the shortest path distance between residential areas and disaster centers, and a mapping relationship between elevation gradient and disaster diffusion speed. The dynamic features of the disaster image, the disaster environment association features, the disaster response time series features and the geographic topological distribution features are aligned and spliced ​​to generate the multimodal disaster feature set.

4. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 1 is characterized in that: Inputting the multimodal disaster feature set into the multimodal dynamic decision model, and calculating an image matching coefficient set, an environment matching coefficient set, a time series matching coefficient set, and a topology matching coefficient set between the multimodal disaster feature set and pre-stored disaster cases, includes: Inputting the multimodal disaster feature set into the image feature matching channel of the multimodal dynamic decision model, calculating the pixel gradient similarity between the dynamic features of the disaster image and each case in the image feature library of pre-stored disaster cases, and generating a set of image matching coefficients; Inputting the disaster environment-related features into the environmental feature matching channel of the multimodal dynamic decision-making model, extracting the environmental parameter covariance matrix of each case in the environmental feature library of pre-stored disaster cases, calculating the spectral norm ratio of the covariance matrix, and generating a set of environmental matching coefficients; Inputting the disaster response time series features into the time series feature matching channel of the multimodal dynamic decision model, performing dynamic time warping alignment with the response delay time distribution of each case in the time series feature library of pre-stored disaster cases, and generating a set of time series matching coefficients; The geographic topological distribution features are input into the topological feature matching channel of the multimodal dynamic decision model, and the road network survivability score and the shortest path distance to the residential area of ​​each case in the topological feature library of pre-stored disaster cases are compared to generate a set of topological matching coefficients.

5. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 1 is characterized in that: The method of selecting a target optimization strategy from the disaster response optimization strategy set based on the disaster scenario matching score, generating a disaster emergency decision instruction set, and sending the disaster emergency decision instruction set to a disaster response terminal includes: Determining a weight distribution ratio for each candidate optimization strategy in the disaster response optimization strategy set according to a difference between the disaster scenario matching score and a set score threshold; Extract the resource scheduling path starting point coordinates, evacuation route density coefficient, and equipment deployment time window of each strategy in the candidate optimization strategy set, and perform conflict detection with the available resource coordinate set, real-time traffic density data, and equipment status data of the current disaster response terminal; For candidate optimization strategies with resource coordinate conflicts, replan the path starting point coordinates to the nearest available resource coordinates and update the path length and estimated arrival time. For evacuation routes with excessive traffic density, the routes are divided into multiple sub-segments according to the density coefficient threshold, and each sub-segment is assigned an alternative route identifier and a turn instruction; For policies where the device deployment time window does not match the device status, the deployment sequence is reallocated according to the device's available time, generating device deployment instructions with aligned timestamps. The updated candidate optimization strategies are sorted by execution priority scores, and the top N strategies are selected to generate the disaster emergency decision-making instruction set, wherein the disaster emergency decision-making instruction includes the adjusted resource scheduling path coordinate sequence, the evacuation sub-section turning instruction, and the equipment deployment timestamp; The disaster emergency decision instruction set is encapsulated into a disaster response terminal protocol format, and the execution time identifier and retry limit parameters of each disaster emergency decision instruction are added, and sent to the corresponding disaster response terminal.

6. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 1, characterized in that: The obtaining of the decision execution feedback data returned by the disaster response terminal and performing incremental parameter optimization processing on the multimodal dynamic decision model based on the decision execution feedback data includes: Receiving instruction execution status parameters uploaded by the disaster response terminal, the instruction execution status parameters including an actual arrival time deviation value of a resource scheduling path, a real-time congestion coefficient of an evacuation sub-section, and an actual startup delay time of equipment deployment; Calculating the absolute error between the actual arrival time deviation value and the instruction estimated time, the ratio of the real-time congestion coefficient to a preset threshold, and the difference between the actual start delay time and the deployment timestamp to generate a policy execution deviation indicator set; Collect real-time environmental data of the current disaster area, including the latest sampling value of the wind speed fluctuation series, the current slope of the temperature change gradient, and the peak displacement of the geological vibration frequency spectrum, to generate environmental status update parameters; Inputting the strategy execution deviation index set and the environmental state update parameter into the feedback learning channel of the multimodal dynamic decision model, and calculating a historical matching score correction value for each candidate optimization strategy, wherein the historical matching score correction value is a linear combination of the deviation index and the environmental parameter; Adjusting the pixel gradient similarity weight of the image feature matching channel, the covariance matrix spectral norm weight of the environment feature matching channel, the dynamic time warping alignment weight of the temporal feature matching channel, and the invulnerability score weight of the topological feature matching channel in the multimodal dynamic decision model according to the historical matching score correction amount; A sliding window mechanism is used to update the image feature library, environmental feature library, temporal feature library and topological feature library of the pre-stored disaster cases, and the current disaster feature set and the adjusted strategy parameters are stored as new cases in the corresponding feature libraries.

7. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 6 is characterized in that: The sliding window mechanism is used to update the image feature library, environmental feature library, temporal feature library and topological feature library of the pre-stored disaster case, and the current disaster feature set and the adjusted strategy parameters are stored as new cases in the corresponding feature libraries, including: Setting the maximum number of storage cases in the feature library of pre-stored disaster cases, and activating a sliding window elimination mechanism when new cases cause the total number to exceed the limit; Calculate the product of the most recent access timestamp of each historical case and the global matching score as the case activity indicator; Arrange historical cases in ascending order by case activity index, and eliminate the M cases with the lowest case activity index to free up storage space; The disaster image dynamic features, disaster environment correlation features, disaster response time series features and geographical topological distribution features of the current disaster feature set are encoded into standardized feature vectors respectively; Associating the standardized feature vector with the corresponding adjusted strategy parameters and execution deviation indicators to generate a new case data block; The new case data block is appended to the end of the image feature library, environmental feature library, temporal feature library and topological feature library of the pre-stored disaster cases according to feature type, and the index identifier of each corresponding feature library is updated.

8. The method for dynamic decision-making in emergency disaster response based on a multimodal AI large model according to claim 3 is characterized in that: The step of inputting the historical disaster response record data into a time series feature extraction network to generate disaster response time series features includes: Performing time slicing processing on the historical disaster response record data to generate a response action sequence within an equally spaced time window; Extract the response action type distribution vector, resource scheduling quantity change curve and loss accumulation rate in each time window; Calling a long short-term memory network to model the temporal dependency of the response action sequence and generate a hidden state vector; The association weights between hidden state vectors of different time windows are calculated through the self-attention mechanism to generate global temporal attention features; The global temporal attention feature is weightedly fused with the hidden state vector of each time window to obtain the disaster response temporal feature.

9. A dynamic decision-making system for emergency disasters based on a multimodal AI large model, characterized by: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the emergency disaster dynamic decision-making method based on the multimodal AI large model as described in any one of claims 1 to 8.

Images