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

Abstract

The invention provides an emergency disaster dynamic decision-making method and system based on a multi-modal AI large model, and the method comprises the steps: firstly obtaining a real-time disaster monitoring data set of a target disaster area, covering disaster image collection, environment sensing, historical response records and geographical distribution topological data, carrying out the feature extraction of the real-time disaster monitoring data set, and carrying out the feature extraction of the real-time disaster monitoring data set; then calling a pre-trained multi-modal dynamic decision-making model, analyzing the disaster scene matching degree of the multi-modal disaster feature set, generating a matching degree score and disaster response optimization strategy set, and screening a target optimization strategy based on the matching degree score, and generating an emergency decision instruction set and sending the emergency decision instruction set to a disaster response terminal, finally obtaining decision execution feedback data of the disaster response terminal, and performing incremental parameter optimization processing on the multi-modal dynamic decision model to realize more accurate and effective emergency disaster dynamic decision.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly, to an emergency disaster dynamic decision-making method and system based on a multi-modal AI large model. Background Art

[0002] In the field of emergency disaster management, traditional disaster decision-making methods mainly rely on manual experience judgment or simple rule matching. These methods often fall short when faced with complex and changing disaster scenarios. Specifically, in the prior art, there is a lack of effective integration and utilization of multi-source heterogeneous data, resulting in an incomplete and inaccurate description of the disaster scenario. In addition, traditional disaster feature extraction methods can often only capture the static features of disasters and are difficult to capture the dynamic evolution process of disasters and their associations with environmental factors, thus affecting the scientificity and timeliness of decision-making.

[0003] In the decision-making process, the prior art mostly adopts rule- or template-based matching methods. Although these methods can achieve rapid response to a certain extent, they lack flexibility and adaptability and are difficult to cope with the complexity and uncertainty of disaster scenarios. At the same time, due to the lack of an effective feedback mechanism, the existing decision-making methods cannot dynamically adjust and optimize the decision-making strategy according to the actual effect of disaster response, resulting in often unsatisfactory decision-making effects. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an emergency disaster dynamic decision-making method based on a multi-modal AI large model, and the method includes: Obtain a real-time disaster monitoring data set of a target disaster area, where 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; Extract features from the real-time disaster monitoring data set to obtain a multi-modal disaster feature set of the disaster scenario, where the multi-modal disaster feature set includes disaster image dynamic features, disaster environment association features, disaster response time series features, and geographical topology distribution features; Call a pre-trained multi-modal dynamic decision-making model to perform disaster scenario matching degree analysis processing on the multi-modal disaster feature set, and generate a disaster scenario matching degree score of the target disaster area and a set of disaster response optimization strategies; Based on the disaster scenario matching degree score, screen target optimization strategies from the set of disaster response optimization strategies, generate a disaster emergency decision instruction set, and send the disaster emergency decision instruction set to a disaster response terminal; Obtain the decision execution feedback data returned by the disaster response terminal, and perform incremental parameter optimization processing on the multi-modal dynamic decision-making model based on the decision execution feedback data.

[0005] In another aspect, an embodiment of the present invention further provides an emergency disaster dynamic decision-making system based on a multi-modal 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.

[0006] Based on the above aspects, the embodiment of the present invention significantly improves the intelligent level and decision-making efficiency of emergency disaster response. Specifically, by integrating multi-modal information such as disaster image acquisition data, environmental sensing data, historical response records, and geographical distribution topology data, a feature set that comprehensively reflects the dynamic characteristics of the disaster scene is constructed. Further, the pre-trained multi-modal dynamic decision-making model realizes the accurate matching degree evaluation of the disaster scene and the intelligent generation of response strategies by deeply mining the internal correlation and complementarity between multi-modal features, which not only improves the scientificity and pertinence of the decision-making strategy, but also ensures the optimization of strategy selection through the matching degree 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, significantly enhancing the adaptability and robustness of the multi-modal dynamic decision-making model to complex and changing disaster scenes. Thus, 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 the efficiency of disaster response and reduce disaster losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a schematic execution flow diagram of an emergency disaster dynamic decision-making method based on a multi-modal AI large model provided by an embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of exemplary hardware and software components of an emergency disaster dynamic decision-making system based on a multi-modal AI large model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flow diagram of an emergency disaster dynamic decision-making method based on a multi-modal AI large model provided by an embodiment of the present invention. The emergency disaster dynamic decision-making method based on the multi-modal AI large model will be introduced in detail below.

[0010] Step S110: Obtain the real-time disaster monitoring data set of the target disaster area, where 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.

[0011] In this embodiment, in order to make accurate emergency decisions for the target disaster area, it is necessary to comprehensively obtain the real-time disaster monitoring data of the target disaster area. The specific acquisition process is as follows: Step S111: Invoke satellite remote sensing equipment to collect the disaster image acquisition data of the target disaster area, where the disaster image acquisition data includes the landform change pixel matrix, thermal distribution gradient information, and water body coverage change trajectory of the disaster area.

[0012] For example, for a target area suffering from an earthquake disaster, satellite remote sensing equipment is invoked to collect images of the area. The satellite remote sensing equipment will obtain images of the area at different times, and the landform change pixel matrix can be obtained by analyzing these images. Suppose the pixel value matrix of a certain area in the image at the first acquisition is [100, 110, 120; 130, 140, 150; 160, 170, 180], and the pixel value matrix of this area becomes [105, 112, 125; 132, 145, 158; 163, 175, 188] at the second acquisition. The change in the landform can be found by comparing these two pixel value matrices. The thermal distribution gradient information can be obtained by analyzing the temperatures of different areas in the image. For example, it can be seen in the image that some areas have higher temperatures and larger temperature change gradients, which may be the fire areas caused by the earthquake. The water body coverage change trajectory can be obtained by identifying and tracking the water body areas in the images at different times. For example, there was originally a river, and after the earthquake, the river may have changed its course or the water level may have changed.

[0013] Step S112: Obtain the disaster environment sensing data uploaded in real time by multiple environment sensing nodes deployed in the target disaster area, where the disaster environment sensing data includes wind speed fluctuation sequence, temperature change gradient, humidity distribution curve, and geological vibration frequency spectrum.

[0014] Multiple environmental sensing nodes are pre-deployed in the target disaster area, and these environmental sensing nodes will upload environmental data in real time. Taking earthquake disasters as an example, the wind speed fluctuation sequence can be recorded by a wind speed sensor. Suppose that within a certain period of time, the wind speed values recorded by the wind speed sensor are 3 m / s, 4 m / s, 5 m / s, 3 m / s, 2 m / s, etc. in sequence, forming a wind speed fluctuation sequence. The temperature change gradient can be calculated by measuring the temperature at different positions and different times with a temperature sensor. For example, within one hour, the temperature at a certain position rises from 20 °C to 22 °C, then the temperature change gradient is 2 °C / hour. The humidity distribution curve can be obtained by measuring the humidity at different positions with a humidity sensor, and the humidity distribution curve can be obtained by plotting these humidity values in a coordinate system. The geological vibration frequency spectrum records the vibration frequencies during earthquakes through devices such as seismographs. For example, the recorded vibration frequencies include vibrations of different frequencies such as 1 Hz, 2 Hz, 5 Hz, etc.

[0015] Step S113: Extract the disaster historical response record data of the target disaster area from the disaster emergency database. The disaster historical response record data includes historical disaster type labels, historical emergency response timestamp sequences, historical resource scheduling paths, and historical loss assessment reports.

[0016] Extract the historical response record data of the target disaster area from the database storing disaster emergency data. For earthquake disaster areas, the historical disaster type labels may show that earthquakes, landslides and other disasters have occurred in this area in the past. The historical emergency response timestamp sequence records the response times at the time of each past disaster. For example, the first earthquake occurred at 10 am and the emergency response started at 10:30 am; the second earthquake occurred at 2 pm and the emergency response started at 2:15 pm, etc. The historical resource scheduling path records how resources were scheduled during past disasters. For example, through which road the materials were transported from a certain material storage point to the disaster area. The historical loss assessment report details the casualties, property losses, etc. caused by each past disaster. For example, the first earthquake caused 10 deaths, 50 injuries, and the economic loss reached 5 million yuan.

[0017] Step S114: Call the geographic information system to analyze the geographic coordinate boundaries of the target disaster area and generate disaster geographic distribution topology data. The disaster geographic distribution topology data includes elevation gradient distribution maps, road connectivity networks, and sets of coordinates of residential aggregation areas.

[0018] Invoke the Geographic Information System (GIS) to parse the geographical coordinate boundaries of the target disaster area. Taking the earthquake disaster area as an example, through GIS, an elevation gradient distribution map of the area can be generated, with areas at different altitudes represented by different colors or lines. For example, areas with lower altitudes are represented by blue, and areas with higher altitudes are represented by yellow. The road connectivity network can show the road conditions in the area, including which roads are connected and which roads may be interrupted due to disasters such as earthquakes. The coordinate set of residential gathering areas records the geographical location coordinates of each residential gathering area in the area. For example, the coordinates of a residential community are (110, 200), and the coordinates of another residential community are (120, 210), etc.

[0019] Step S115: Perform timestamp alignment processing on the disaster image acquisition data, the disaster environment sensing data, the disaster historical response record data, and the disaster geographical distribution topology data to obtain the real-time disaster monitoring data set.

[0020] After obtaining the above various types of data, since the acquisition times of these data may be different, timestamp alignment processing is required. For example, the disaster image acquisition data was collected at 10 am, a certain sensor in the disaster environment sensing data collected data at 10:10 am, a certain event in the historical disaster response record data occurred at 9 am, and the geographical distribution topology data was last updated at 9:30 am. Through timestamp alignment processing, these data are sorted according to a unified time standard, making them consistent in time, so as to obtain a complete real-time disaster monitoring data set.

[0021] Step S120: Extract features from the real-time disaster monitoring data set to obtain a multi-modal disaster feature set of the disaster scene, where the multi-modal disaster feature set includes disaster image dynamic features, disaster environment association features, disaster response time series features, and geographical topology distribution features.

[0022] In this embodiment, after obtaining the real-time disaster monitoring data set, it is necessary to extract features from it to obtain a multi-modal disaster feature set that can reflect the disaster scene. The specific steps are as follows: Step S121: Invoke a pre-trained disaster image encoder to perform convolutional feature extraction processing on the disaster image acquisition data to obtain disaster image dynamic features, where the disaster image dynamic features include landform deformation gradient features, thermal diffusion direction vectors, and water body coverage area change rates.

[0023] Taking earthquake disasters as an example, the pre-trained disaster image encoder is trained based on a large amount of disaster image data, and its internal structure includes multiple convolutional layers, pooling layers, etc. For the acquired disaster image acquisition data, it is first input into the first convolutional layer of the encoder. Suppose the input disaster image is a color image with 256×256 pixels and has three channels (red, green, blue), then the dimension of the input data is 256×256×3.

[0024] In the first convolutional layer, multiple different convolutional kernels are used for convolution operations. For example, convolutional kernels with a size of 3×3 are used, and the number is 16. Each convolutional kernel slides on the input image, performing element multiplication and summation operations to generate a feature map. Taking one convolutional kernel as an example, when it slides on the image, it multiplies and accumulates with the pixel values at the corresponding positions of the image to generate a new pixel value, and finally obtains a feature map. After the operations of 16 convolutional kernels, 16 feature maps are obtained, and at this time the dimension of the data becomes 254×254×16 (because 256 - 3 + 1 = 254).

[0025] Next, pooling operations are performed on these feature maps, such as using 2×2 max pooling. Max pooling selects the maximum value within a 2×2 area as the output of this area, which can reduce the dimension of the data while retaining important feature information. After max pooling, the dimension of the data becomes 127×127×16.

[0026] As the data is passed layer by layer in the encoder, the convolution and pooling operations are continuously repeated, the dimension of the data gradually becomes smaller, and the features gradually become more abstract.

[0027] For the extraction of landform deformation gradient features, the encoder will focus on the changes in the landform in the image. For example, by comparing the images collected at different times, the changes in pixel values are analyzed to determine the deformation of the landform. Suppose in a certain area, the change in pixel values between two consecutive images is obvious, and the encoder will calculate the gradient of the pixel values to quantify this deformation. In the horizontal direction, the difference between adjacent pixel values is calculated, and the same calculation is also performed in the vertical direction to obtain the gradient values in the horizontal and vertical directions. Combining these gradient values forms the landform deformation gradient feature. For example, in a 10×10 area, the gradient values in the horizontal direction are [0.1, 0.2, 0.15,...], and the gradient values in the vertical direction are [0.05, 0.12, 0.08,...], and they are concatenated to obtain a 20-dimensional landform deformation gradient feature vector.

[0028] The extraction of the thermal diffusion direction vector relies on the thermal distribution gradient information in the image. The encoder will identify the areas with higher temperatures in the image and analyze the temperature change trends in these areas. By calculating the temperature gradients at different positions, the direction of thermal diffusion is determined. 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, then this direction is the main direction of thermal diffusion. Represent this direction as a vector. For example, in a two-dimensional plane, the thermal diffusion direction vector can be represented as (0.3, 0.4), where 0.3 and 0.4 represent the components in the x and y directions respectively.

[0029] The calculation of the change rate of the water body coverage area is done by comparing the areas of the water body regions in images collected at different times. First, the image needs to be segmented to separate the water body region from other regions. A threshold-based segmentation method can be used. According to the color characteristics of the water body in the image, a suitable threshold is set, and the regions with pixel values greater than this threshold are determined as the water body. Then, calculate the number of pixels in the water body region in images at different times to obtain the water body coverage area. Suppose the number of pixels corresponding to the water body coverage area in the first image is 1000, and in the second image is 1100, then the change rate of the water body coverage area is (1100 - 1000) / 1000 = 0.1.

[0030] Step S122: Perform time series analysis processing on the disaster environment sensing data to extract disaster environment correlation features, where the disaster environment correlation features include the covariance matrix of wind speed and temperature, the frequency domain correlation degree of humidity and geological vibration, and the time lag correlation coefficient of multi-sensor data.

[0031] For disaster environment sensing data, the wind speed fluctuation sequence, temperature change gradient, humidity distribution curve, and geological vibration frequency spectrum are all data sequences that change over time.

[0032] When calculating the covariance matrix of wind speed and temperature, first ensure that the timestamps of the wind speed fluctuation sequence and the temperature change gradient sequence are aligned. Suppose the wind speed fluctuation sequence is [3, 4, 5, 3, 2] m / s, and the corresponding time points are t1, t2, t3, t4, t5 respectively, and the temperature change gradient sequence is [0.5, 0.6, 0.7, 0.5, 0.4] °C / hour, and the time points are also t1, t2, t3, t4, t5.

[0033] First, calculate the mean of the wind speed sequence. Add all the values in the wind speed sequence and then divide by the length of the sequence, that is, (3 + 4 + 5 + 3 + 2) / 5 = 3.4 m / s. Similarly, calculate the mean of the temperature change gradient sequence as (0.5 + 0.6 + 0.7 + 0.5 + 0.4) / 5 = 0.54 °C / hour.

[0034] Then, for each time point, calculate the difference between the wind speed value and the mean wind speed, and the difference between the temperature change gradient value and the mean temperature, and multiply these two differences. For example, at time t1, the wind speed difference is 3 - 3.4 = -0.4 m / s, the temperature difference is 0.5 - 0.54 = -0.04 °C / hour, and their product is (-0.4) × (-0.04) = 0.016. Perform such calculations for all time points to obtain a set of product values.

[0035] Finally, sum up 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 variances of wind speed and temperature are also calculated in a similar manner. 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]].

[0036] Calculating the frequency-domain correlation degree between humidity and geological vibration requires first performing frequency-domain analysis on the humidity distribution curve and the geological vibration frequency spectrum. For the humidity distribution curve, use Fourier transform to convert it from the time domain to the frequency domain to obtain the frequency components of humidity and their corresponding amplitudes. Similarly, perform Fourier transform on the geological vibration frequency spectrum. Then, analyze the correlation between these two frequency-domain signals at the same frequency components. 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 relationships at multiple frequency points, the correlation degree between humidity and geological vibration in the frequency domain is obtained. Methods such as correlation coefficient can be used to quantify this correlation degree. Assuming that the average value of the correlation coefficients calculated at multiple frequency points is 0.6, then the frequency-domain correlation degree between humidity and geological vibration can be expressed as 0.6.

[0037] The calculation of the time-lag correlation coefficient of multi-sensor data is to analyze the delay relationship in time between different sensor data. Taking wind speed and temperature data as an example, first perform lag operations on the wind speed sequence at different time steps. For example, lag the wind speed sequence by 1 time step to obtain a new sequence. Then calculate the correlation coefficient between this lagged sequence and the temperature sequence. The calculation of the correlation coefficient can use the Pearson correlation coefficient formula, that is, first calculate the covariance of the two sequences, and then divide it by the product of their standard deviations. By continuously changing the lag time step, calculate the correlation coefficients in different lag cases, and find the lag time step when the correlation coefficient is the largest. Suppose when the wind speed sequence is lagged by 2 time steps, the correlation coefficient with the temperature sequence is the largest, which is 0.8. Then the time-lag correlation coefficient of the multi-sensor data in this case is 0.8, and the lag time step is 2.

[0038] Step S123: Input the disaster historical response record data into the time series feature extraction network to generate disaster response time series features, where 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 loss and response time.

[0039] Step S1231: Perform time slicing processing on the disaster historical response record data to generate a response action sequence within an equally spaced time window.

[0040] Suppose the disaster historical response record data records the response situation within 24 hours from the occurrence of the disaster. Divide these 24 hours into intervals of every 2 hours to obtain 12 equally spaced time windows. For each time window, carefully analyze the recorded response actions. For example, within the first 2-hour time window, the possible response actions are to start the emergency command center and issue early warning information; within the second 2-hour time window, the response actions may be to deploy rescue teams and prepare rescue supplies, etc. Organize the response actions within each time window into a sequence. Suppose there are 5 types of response actions, which are represented by numbers 1 - 5 respectively. Then the response action sequence of the first time window may be [1, 2], and the second time window may be [3, 4].

[0041] Step S1232: Extract the response action type distribution vector, the resource scheduling quantity change curve, and the loss accumulation rate within each time window.

[0042] Within each time window, for the response action type distribution vector, count the number of occurrences of each response action type. Suppose within a certain time window, the number of occurrences of the 5 response action types are 2 times, 1 time, 3 times, 0 times, and 1 time respectively. Then the response action type distribution vector is [2, 1, 3, 0, 1].

[0043] To extract the curve of the change in resource scheduling quantity, it is necessary to pay attention to the scheduling situation of different resources within each time window. Taking rescue vehicles as an example, assume that within a 2-hour time window, 2 vehicles are scheduled at the 0th minute, 3 vehicles at the 30th minute, 1 vehicle at the 60th minute, 2 vehicles at the 90th minute, and 1 vehicle at the 120th minute. Using time as the horizontal axis and the scheduling quantity as the vertical axis, connecting these data points forms the curve of the change in resource scheduling quantity.

[0044] The calculation of the cumulative loss rate is by comparing the loss situations at the start and end of this time window. Assume that at the start of a time window, the loss assessment caused by the disaster is 1 million yuan, and at the end it is 1.2 million yuan. Then the cumulative loss rate is (1.2 - 1) / 2 = 0.1 million yuan per hour.

[0045] Step S1233: Invoke a long short-term memory network to model the temporal dependence relationship of the response action sequence and generate a hidden state vector.

[0046] A long short-term memory network (LSTM) consists of an input gate, a forget gate, an output gate, and a cell state. The previously generated response action sequence is input into the LSTM network in sequence. Assume that the response action sequence has 12 elements (corresponding to 12 time windows), and each element is a response action type distribution vector with a dimension of 5.

[0047] At the first time step, the input is the response action type distribution vector of the first time window. The LSTM network first determines through the forget gate how much information of the cell state from the previous time step needs to be retained. The forget gate uses a sigmoid function to linearly combine the input vector and the hidden state vector of the previous time step, and then maps the result to between 0 and 1 through the sigmoid function. Assume that the hidden state vector of the previous time step is a zero vector (because it is the first time step), the input vector is [2, 1, 3, 0, 1], and the weight matrix and bias of the forget gate are obtained through training. After calculation, the forget gate outputs a vector between 0 and 1, determining the proportion of the cell state to be retained.

[0048] Next, the input gate determines how much new information needs to be added to the cell state. The input gate also uses the sigmoid function and the tanh function. The sigmoid function determines which values need to be updated, and the tanh function creates a new candidate vector. Combining these two results updates the cell state.

[0049] Finally, the output gate determines the hidden state vector at the current time step. The output gate also uses the sigmoid function to linearly combine the input vector and the hidden state vector of the previous time step, and then obtains a vector between 0 and 1 through the sigmoid function. This vector is then multiplied by the cell state vector processed by the tanh function to obtain the hidden state vector at the current time step.

[0050] As the time steps progress, the above process is continuously repeated, and finally a sequence containing 12 hidden state vectors is obtained. Assuming that the dimension of the hidden state vector is 32, after being processed by the LSTM network, a 12×32 matrix will be obtained, where each row represents the hidden state vector of a time step.

[0051] Step S1234: Calculate the correlation weights between the hidden state vectors of different time windows through the self-attention mechanism to generate global temporal attention features.

[0052] The self-attention mechanism allows the model to focus on relevant information in other time windows when processing the hidden state vector of each time window. For each hidden state vector, the correlation weights are determined by calculating its similarity with all other hidden state vectors.

[0053] First, multiply the hidden state vector sequence by three different weight matrices respectively 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, a query vector sequence, a key vector sequence, and a value vector sequence are obtained, all with a dimension of 12×32.

[0054] Then, for each query vector, calculate its similarity with all key vectors. The similarity can be calculated through the dot product operation, that is, the query vector and the key vector are multiplied element by element and summed. For example, the dot product of the first query vector and the first key vector is a value, and the dot product with the second key vector is another value. By calculating in turn, a similarity vector containing 12 values is obtained.

[0055] Next, normalize this similarity vector and use the softmax function to convert it into a probability distribution to obtain the correlation weight vector. Each element in the correlation weight vector represents the degree of attention of the current query vector to the hidden state vectors of other time windows.

[0056] Finally, perform a weighted sum of the associated weight vector and the sequence of value vectors to obtain a new vector, which is the global temporal attention feature corresponding to the current query vector. Perform such an operation on all query vectors, and finally obtain a sequence containing 12 global temporal attention feature vectors. Assuming that the dimension of the global temporal attention feature vector is also 32, the dimension of the obtained sequence is 12×32.

[0057] Step S1235: Weightedly fuse the global temporal attention feature with the hidden state vectors of each time window to obtain the disaster response temporal feature.

[0058] Assign weights to the global temporal attention feature and the hidden state vectors of each time window. Assume that the weight of the global temporal attention feature is 0.4, and the weight of the hidden state vectors of each time window is 0.6.

[0059] 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, multiply the global temporal attention feature vector by 0.4, and then add these two results to obtain the first fused vector. Assume that 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].

[0060] Perform such an operation on the hidden state vector and the global temporal attention feature vector of each time window in turn, and finally obtain a set of fused vectors. Concatenate these fused vectors in the order of time windows to obtain the disaster response temporal feature. Assume that each hidden state vector and global temporal attention feature is 32-dimensional, then the fused vector is also 32-dimensional, and the dimension of the disaster response temporal feature after concatenation will depend on the number of time windows. If there are 12 time windows, then the dimension of the disaster response temporal feature is 32×12 = 384 dimensions.

[0061] Step S124: Perform connectivity analysis processing on the disaster geographical distribution topology data through a graph neural network to generate geographical topology distribution features, where the geographical topology distribution features include the road network anti-destruction score, the shortest path distance between the residential area and the disaster center, and the mapping relationship between the elevation gradient and the disaster diffusion speed.

[0062] Before performing connectivity analysis processing, it is necessary to clarify the specific form of the disaster geographical distribution topology data. Taking the earthquake disaster scenario as an example, the elevation gradient distribution map in the geographical distribution topology data can be transformed into a two-dimensional matrix, where each element in the matrix represents the altitude value of the area; the road connectivity network can be represented by a graph structure, where the nodes in the graph represent the intersections or endpoints of the roads, and the edges represent the roads connecting these nodes. The attributes of the edges can include the length, width, load-bearing capacity, etc. of the roads; the set of coordinates of the residential aggregation areas is a series of two-dimensional coordinate points, representing the locations of each residential area.

[0063] In this embodiment, the disaster geographical distribution topology data is transformed 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, the node features can include the elevation value at the location of the node (obtained from the elevation gradient distribution map), the surrounding population density (which can be estimated according to the set of coordinates of the residential aggregation areas), 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 the 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.

[0064] For the edges connecting the nodes, the edge features can include the length of the road and the type of the road (such as highway, ordinary road, etc.). The road length can be calculated according to the geographical coordinates. Assuming that the road length ranges from 0 to 100 kilometers, it is normalized and used as one dimension of the edge feature. The road type can be one-hot encoded. Assuming there are three types of roads: highway, ordinary road, and rural path, then 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 highway, and the latter two 0s indicate that it is not an ordinary road and a rural path.

[0065] The graph neural network updates the features of the nodes through a message passing mechanism. In each layer of the graph neural network, a node receives messages from its neighbor nodes and updates its own features based on these messages.

[0066] First, for each node, calculate the messages sent by its neighbor nodes. Assume the current node is v, and its set of neighbor nodes is N(v). The message sent by neighbor node u to node v can be calculated as follows: Concatenate the feature vector of neighbor node u and the feature vector of edge (u, v), and then obtain the message vector through a linear transformation (which can be represented as a matrix multiplication and an addition of a bias term). For example, if the feature vector of neighbor node u is [0.2, 0.5] and the feature vector of edge (u, v) is [0.3, 1, 0, 0], after concatenation, we get [0.2, 0.5, 0.3, 1, 0, 0]. Assume the matrix of the linear transformation is a 6×5 matrix W and a 5-dimensional bias vector b. After matrix multiplication and addition operations, a 5-dimensional message vector is obtained.

[0067] Then, node v aggregates the messages sent by all neighbor nodes. Aggregation can be performed using methods such as summation or averaging. Here, summation is used as an example. Add up the message vectors sent by all neighbor nodes to obtain the aggregated message vector of node v.

[0068] 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, after concatenating the aggregated message vector and the original feature vector of node v, a linear transformation is performed to obtain the updated feature vector.

[0069] The resilience score of the road network reflects the stability and reliability of the road network during disasters. It can be calculated based on the updated features of nodes and edges in the graph neural network.

[0070] For each road (edge), comprehensively consider factors such as its carrying capacity, length, and elevation of the area where it is located. The carrying capacity can be estimated based on the type and width of the road. Assume the carrying capacity of a highway is 1000 vehicles per hour, that of an ordinary road is 500 vehicles per hour, and that of a rural path is 100 vehicles per hour, and normalize them.

[0071] Define a resilience evaluation function. For example, for edge (u, v), its resilience score can be expressed as: Resilience score = Carrying capacity score × Length score × Elevation impact score. The carrying capacity score is the normalized carrying capacity value. The length score can be calculated by 1 / the normalized road length (the shorter the length, the higher the score). The elevation impact score can be determined based on the elevation values of nodes u and v. For example, if the elevation value is relatively high, it may indicate that the area is more vulnerable during disasters, and the elevation impact score can be set to a relatively low value, such as 0.8; if the elevation value is relatively low, the elevation impact score can be set to 1.

[0072] For the entire road network, sum or weighted sum the resilience scores of all edges to obtain the resilience score of the road network. Suppose there are 10 edges in the road network, and the resilience 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]. Add these scores to get the resilience score of the road network: 0.8 + 0.9 + 0.7 + 0.6 + 0.85 + 0.95 + 0.75 + 0.65 + 0.8 + 0.9 = 7.9.

[0073] Use the graph structure updated by the graph neural network and combine it with graph algorithms (such as Dijkstra's algorithm) to calculate the shortest path distance between the residential area and the disaster center.

[0074] First, determine the location of the disaster center. Assume that the disaster center corresponds to a node c in the graph. For each coordinate of the residential gathering area, find the node in the graph that is closest to the residential gathering area and use it as the node representing the residential area.

[0075] Taking Dijkstra's algorithm as an example, initialize a distance array, set the distance of the disaster center node c to 0, and set the distances of other nodes to infinity. Then, select the node with the smallest distance from the distance array, mark it as visited, and update the distances of its neighbor nodes. If the distance to a neighbor node through the current node is smaller than the original distance of the neighbor node, update the distance of the neighbor node.

[0076] Repeat the above steps until all nodes are visited or the shortest path distances of all representative nodes of the residential areas are found. For example, assume there are 5 representative nodes of the residential areas, and the shortest path distances between them and the disaster center calculated by Dijkstra's algorithm are [5, 8, 3, 6, 7] kilometers.

[0077] Finally, analyze the relationship between the elevation gradient distribution map and the disaster diffusion speed. The elevation gradient distribution map can be divided into multiple regions, and each region has a similar elevation value range.

[0078] For each region, count the disaster diffusion speed data within the region. The disaster diffusion speed can be obtained from historical disaster data or simulation data. For example, in an earthquake disaster, the disaster diffusion speed can be estimated based on the propagation speed and influence range of seismic waves.

[0079] Suppose the elevation gradient distribution map is divided into 5 regions, and the elevation value ranges of each region are [0 - 200 meters, 200 - 400 meters, 400 - 600 meters, 600 - 800 meters, 800 - 1000 meters] respectively. The disaster diffusion speeds within each region are statistically obtained as follows: the average disaster diffusion speed in Region 1 (0 - 200 meters) is 10 m / s, in Region 2 (200 - 400 meters) is 8 m / s, in Region 3 (400 - 600 meters) is 6 m / s, in Region 4 (600 - 800 meters) is 4 m / s, and in Region 5 (800 - 1000 meters) is 2 m / s. [[ID=!]]

[0080] By fitting these data, a mapping relationship between the elevation gradient and the disaster diffusion speed can be established. Methods such as linear regression can be used for fitting. Suppose the obtained linear regression equation is: Disaster diffusion speed = -0.01 × Elevation value + 12. In this way, for any given elevation value, the corresponding disaster diffusion speed can be estimated according to this mapping relationship.

[0081] Step S125: Align and splice the disaster image dynamic features, disaster environment association features, disaster response time - series features, and geographical topology distribution features in terms of feature dimensions to generate the multi - modal disaster feature set.

[0082] Before performing the feature dimension alignment and splicing process, it is necessary to clarify the dimensions of each feature. Suppose the dimension of the disaster image dynamic feature is 50 - dimensional, the dimension of the disaster environment association feature is 30 - dimensional, the dimension of the disaster response time - series feature is 384 - dimensional, and the dimension of the geographical topology distribution feature is 20 - dimensional.

[0083] Specifically, it can be checked whether the dimensions of each feature are suitable for splicing. If the dimensions of some features are too high or too low, dimensionality reduction or dimensionality increase processing may be required.

[0084] For the disaster image dynamic feature, if its dimension is relatively high, methods such as principal component analysis (PCA) can be used for dimensionality reduction. The PCA method calculates the covariance matrix of the features, finds the directions of its principal components, and then projects the features onto these principal component directions, thereby reducing the dimension of the features. Suppose the disaster image dynamic feature is reduced from 50 dimensions to 40 dimensions through PCA.

[0085] For features with relatively low dimensions, such as the geographical topology distribution feature, dimensionality increase can be performed by filling zero vectors, etc. Suppose the geographical topology distribution feature is increased from 20 dimensions to 30 dimensions, and 10 zero values are filled behind the original feature vector.

[0086] After dimension alignment, the dynamic features of the disaster image, the associated features of the disaster environment, the temporal features of disaster response, and the geographical topological distribution features are concatenated in a certain order. For example, they are concatenated in the order of dynamic features of the disaster image, associated features of the disaster environment, temporal features of disaster response, and geographical topological distribution features.

[0087] The dimension of the concatenated multi-modal disaster feature set is 40 + 30 + 384 + 30 = 484 dimensions. The individual feature vectors are sequentially connected to form a new 484-dimensional feature vector, which represents the multi-modal disaster feature set of the disaster scenario. In this way, different types of features are integrated together to provide more comprehensive information for subsequent disaster scenario matching degree analysis and decision-making.

[0088] Step S130: Invoke the pre-trained multi-modal dynamic decision-making model to perform disaster scenario matching degree analysis on the multi-modal disaster feature set, and generate the disaster scenario matching degree score of the target disaster area and the set of disaster response optimization strategies.

[0089] In this embodiment, the pre-trained multi-modal dynamic decision-making model is trained based on a large amount of disaster case data. It contains multiple feature matching channels for processing different types of disaster features respectively. After obtaining the multi-modal disaster feature set, it is input into this model for disaster scenario matching degree analysis. The specific steps are as follows: Step S131: Input the multi-modal disaster feature set into the multi-modal dynamic decision-making model, and calculate the set of image matching coefficients, the set of environment matching coefficients, the set of temporal matching coefficients, and the set of topological matching coefficients between the multi-modal disaster feature set and the pre-stored disaster cases.

[0090] Step S1311: Input the multi-modal disaster feature set into the image feature matching channel of the multi-modal dynamic decision-making model, calculate the pixel gradient similarity between the dynamic features of the disaster image and each case in the image feature library of the pre-stored disaster cases, and generate the set of image matching coefficients.

[0091] Suppose the dynamic features of the disaster image are a 40-dimensional vector, and there are 100 cases in the image feature library of the pre-stored disaster cases, and the image features of each case are also 40-dimensional vectors.

[0092] In the image feature matching channel, for the dynamic feature vector of the disaster image and each case vector in the image feature library, calculate their pixel gradient similarity. The specific calculation process is as follows: First, subtract the elements at the corresponding positions of the two vectors to obtain a difference vector. For example, if the dynamic feature vector of the disaster image is [a1, a2, …, a40], and a certain case vector in the image feature library is [b1, b2, …, b40], the difference vector is [|a1 - b1|, |a2 - b2|, …, |a40 - b40|].

[0093] Then, square each element in the difference vector to obtain a squared difference vector [(a1 - b1)^2, (a2 - b2)^2, …, (a40 - b40)^2].

[0094] Next, add up all the elements in the squared difference vector to obtain a total value. Assume the total value is S.

[0095] Finally, to obtain the similarity, using the idea of an inverse proportional function, divide 1 by (1 + S) to obtain the pixel gradient similarity. For example, if S = 5, then the pixel gradient similarity is 1 / (1 + 5) = 1 / 6 ≈ 0.17.

[0096] Perform the above calculations on 100 cases in the image feature library in sequence to obtain an image matching coefficient set containing 100 similarity values, such as [0.17, 0.2, 0.15, …].

[0097] Step S1312: Input the disaster environment associated features into the environment feature matching channel of the multi-modal dynamic decision model, extract the environmental parameter covariance matrix of each case in the environmental feature library of the pre-stored disaster cases, calculate the spectral norm ratio of the covariance matrix, and generate an environmental matching coefficient set.

[0098] Assume that the disaster environment associated features are represented by a 30-dimensional vector, and there are 100 cases in the environmental feature library of the pre-stored disaster cases, and each case is also represented by a 30-dimensional vector.

[0099] First, for the disaster environment associated feature vector and each case vector in the environmental feature library, calculate their environmental parameter covariance matrices respectively. Taking the disaster environment associated feature vector as an example, assume there is a series of observation values, and arrange these observation values into a matrix X in a certain time order (assuming the number of observations is n, then X is an n×30 matrix).

[0100] The steps to calculate the covariance matrix are as follows: First, calculate the mean vector of each feature. For each column of X, add up all the elements in this column and then divide by the number of observations n to obtain a 30-dimensional mean vector μ.

[0101] Then, subtract the mean vector μ from each row in the matrix X to obtain a new matrix X'.

[0102] Next, calculate the covariance matrix \(C = (X'^T*X') / (n - 1)\), where \(X'^T\) represents the transpose matrix of \(X'\).

[0103] Similarly, calculate the covariance matrix for each case vector in the environmental feature library.

[0104] Next, calculate the spectral norm of the covariance matrix. The spectral norm is the largest singular value of the matrix and can be calculated by singular value decomposition (SVD). For a matrix \(A\), perform singular value decomposition \(A = U\Sigma V^T\), where \(\Sigma\) is a diagonal matrix and the elements on its diagonal are the singular values of matrix \(A\), and the spectral norm is the largest diagonal element in \(\Sigma\).

[0105] For the covariance matrix \(C1\) of the disaster environment correlation features and the covariance matrix \(C2\) of a certain case in the environmental feature library, calculate the ratio of their spectral norms. Assume the spectral norm of \(C1\) is \(\sigma1\) and the spectral norm of \(C2\) is \(\sigma2\), and the spectral norm ratio is \(\sigma1 / \sigma2\).

[0106] Perform the above calculations for 100 cases in the environmental feature library in sequence to obtain an environmental matching coefficient set containing 100 spectral norm ratios, such as \([0.8, 0.9, 0.75, \ldots]\).

[0107] Step S1313: Input the disaster response time series features into the time series feature matching channel of the multi-modal dynamic decision-making model, and perform dynamic time warping alignment with the response delay time distribution of each case in the time series feature library of pre-stored disaster cases to generate a time series matching coefficient set.

[0108] Assume that the disaster response time series features are a 384-dimensional vector, and there are 100 cases in the time series feature library of pre-stored disaster cases. The response delay time distribution of each case is represented by a time series, and assume the length of the time series is \(m\).

[0109] The dynamic time warping (DTW) algorithm is used to calculate the similarity between two time series. In this step, extract the part related to the response delay time from the disaster response time series features to form a time series \(T1\), and the response delay time distribution of a certain case in the time series feature library of pre-stored disaster cases is the time series \(T2\).

[0110] The specific steps of the DTW algorithm are as follows: Create a \((m + 1)\times(m + 1)\) matrix \(D\) to store the intermediate calculation results. Initialize \(D[0, 0]=0\), for \(i>0\), \(D[i, 0]=\infty\), and for \(j>0\), \(D[0, j]=\infty\).

[0111] Then, for i from 1 to m and j from 1 to m, calculate the value of D[i, j]. The calculation formula for 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]).

[0112] Finally, the DTW distance is D[m, m]. To obtain the similarity, divide 1 by (1 + DTW distance) to get a similarity value.

[0113] Perform DTW calculations on 100 cases in the time series feature library in sequence to obtain a set of time series matching coefficients containing 100 similarity values, such as [0.7, 0.65, 0.8, …].

[0114] Step S1314: Input the geographical topological distribution feature into the topological feature matching channel of the multi-modal dynamic decision model, compare the road network resilience score and the shortest path distance of residential areas for each case in the topological feature library of pre-stored disaster cases, and generate a set of topological matching coefficients.

[0115] Assume that the geographical topological distribution feature is represented by a 30-dimensional vector, which contains information such as the road network resilience score and the shortest path distance of residential areas. There are 100 cases in the topological feature library of pre-stored disaster cases, and each case also contains the road network resilience score and the shortest path distance of residential areas.

[0116] For the road network resilience score R1 in the geographical topological distribution feature and the road network resilience score R2 of a certain case in the topological feature library, calculate the absolute value of their difference |R1 - R2|. For the shortest path distance D1 of residential areas in the geographical topological distribution feature and the shortest path distance D2 of residential areas of a certain case in the topological feature library, also calculate the absolute value of their difference |D1 - D2|.

[0117] Then, assign weights to the road network resilience score and the shortest path distance of residential areas. Assume that the weight of the road network resilience score is 0.6 and the weight of the shortest path distance of residential areas is 0.4.

[0118] Comprehensive score = 0.6 × |R1 - R2| + 0.4 × |D1 - D2|.

[0119] To obtain the similarity, divide 1 by (1 + comprehensive score) to get a similarity value.

[0120] Perform the above calculations on 100 cases in the topological feature library in sequence to obtain a set of topological matching coefficients containing 100 similarity values, such as [0.85, 0.9, 0.8, …].

[0121] Step S132: Perform normalized weighted summation on the image matching coefficient set, environment matching coefficient set, time series matching coefficient set, and topology matching coefficient set to generate the global matching degree score for each pre-stored disaster case.

[0122] First, perform normalization on the image matching coefficient set, environment matching coefficient set, time series matching coefficient set, and topology matching coefficient set. Taking the image matching coefficient set as an example, find the maximum value max1 and minimum value min1 in the set. For each element x in the set, the normalized element x'=(x - min1) / (max1 - min1). Similarly, perform normalization on the environment matching coefficient set, time series matching coefficient set, and topology matching coefficient set.

[0123] Suppose the weights of the image matching coefficient set, environment matching coefficient set, time series matching coefficient set, and topology matching coefficient set are 0.3, 0.2, 0.2, and 0.3 respectively.

[0124] For the first case in the pre-stored disaster cases, its normalized image matching coefficient is 0.8, normalized environment matching coefficient is 0.7, normalized time series matching coefficient is 0.6, and normalized topology matching coefficient is 0.8.

[0125] Global matching degree score = 0.3×0.8 + 0.2×0.7 + 0.2×0.6 + 0.3×0.8 = 0.24 + 0.14 + 0.12 + 0.24 = 0.74.

[0126] Calculate for the 100 cases in the pre-stored disaster cases in sequence to obtain the global matching degree score for each case, forming a set containing 100 score values, such as [0.74, 0.78, 0.72,...].

[0127] Step S133: Arrange the pre-stored disaster cases in descending order according to the global matching degree score, select the cases with global matching degree scores higher than the set score threshold as the similar case set, and extract the corresponding historical response strategies and their loss reduction rates for each case in the similar case set.

[0128] Suppose the set score threshold is 0.7. Arrange the global matching degree score set in descending order, for example, the arranged set is [0.85, 0.8, 0.78, 0.76, 0.74, 0.72,...].

[0129] Select the cases with scores higher than 0.7. Suppose 20 cases are obtained after screening, and these cases form the similar case set.

[0130] For each case in the set of similar cases, extract its corresponding historical response strategies and loss reduction rates from the pre-stored database. Historical response strategies may include resource scheduling plans, personnel evacuation plans, etc. The loss reduction rate refers to the proportion by which the disaster losses are reduced after adopting the historical response strategy compared to when no strategy is adopted. For example, the historical response strategy for the first case is to prioritize evacuating residents to safe areas, with a loss reduction rate of 30%; the historical response strategy for the second case is to quickly allocate relief supplies, with a loss reduction rate of 25%, and so on.

[0131] Step S134: Perform an adaptive adjustment of the strategy parameters and the current disaster environment parameters for the historical response strategy, 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 optimized strategies.

[0132] For each historical response strategy in the set of similar cases, perform parameter adjustment.

[0133] Taking the adjustment of the starting coordinates of the resource scheduling path as an example, assume that the starting coordinates of the resource scheduling path for a certain historical response strategy are (x1, y1). In the current disaster environment, according to the location of available resources and road conditions, it is found that the coordinates (x2, y2) which are closer to the starting point and more suitable as the starting point, then adjust the starting coordinates of the resource scheduling path to (x2, y2).

[0134] For the adjustment of the evacuation route density coefficient, it is carried out according to factors such as the population density and road carrying capacity of the current disaster area. Assume that the evacuation route density coefficient in the historical response strategy is 0.8. The population density in the current disaster area is higher than that in the historical case, and the road carrying capacity is limited. After evaluation, adjust the evacuation route density coefficient to 0.7.

[0135] For the adjustment of the equipment deployment time window, consider the development stage of the current disaster and the availability of the equipment. Assume that the historical response strategy plans to deploy a certain equipment from 10 am to 11 am, but the current disaster is developing rapidly, and the equipment is available from 9 am to 10 am. Adjust the equipment deployment time window to 9 am to 10 am.

[0136] Perform the above parameter adjustments for 20 historical response strategies in the set of similar cases to obtain 20 adjusted strategies, and these strategies form a set of candidate optimized strategies. [[ID=,19]]

[0137] Step S135: Calculate the execution priority score for each candidate optimized strategy based on the loss reduction rate and the adjusted strategy parameters, and use the set of candidate optimized strategies sorted by the execution priority score as the set of disaster response optimized strategies, and use the global matching degree score as the disaster scenario matching degree score.

[0138] Weights are assigned to the loss reduction rate and the adjusted policy 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 policy parameters is 0.4.

[0139] For each policy in the set of candidate optimization policies, the comprehensive evaluation score of the adjusted policy parameters can be obtained by evaluating factors such as the rationality of the resource scheduling path, the feasibility of the evacuation route, and the timeliness of equipment deployment time. Assume that the loss reduction rate of the first candidate optimization policy is 30%, and the comprehensive evaluation score of the adjusted policy parameters is 0.8.

[0140] Execution priority score = 0.6×30% + 0.4×0.8 = 0.18 + 0.32 = 0.5.

[0141] Calculate for the 20 policies in the set of candidate optimization policies in turn to obtain the execution priority score of each policy. Sort the set of candidate optimization policies in descending order according to the execution priority score. The sorted set of candidate optimization policies is the set of disaster response optimization policies. At the same time, use the previously calculated global matching degree score as the disaster scenario matching degree score. For example, for the first candidate optimization policy, the corresponding disaster scenario matching degree score is the previously calculated global matching degree score of this case, which is 0.85.

[0142] Step S140: Based on the disaster scenario matching degree score, screen the target optimization policies from the set of disaster response optimization policies, generate a set of disaster emergency decision instructions, and send the set of disaster emergency decision instructions to the disaster response terminal.

[0143] In this embodiment, after obtaining the disaster scenario matching degree score and the set of disaster response optimization policies, it is necessary to screen out the target optimization policies according to the disaster scenario matching degree score, generate the corresponding set of disaster emergency decision instructions, and send them to the disaster response terminal. The specific steps are as follows: Step S141: Determine the weight allocation ratio of each candidate optimization policy in the set of disaster response optimization policies according to the difference between the disaster scenario matching degree score and the set scoring threshold.

[0144] Suppose the set scoring threshold is set to 0.7, and the disaster scenario matching degree scores have different values. For each candidate optimization strategy in the set of disaster response optimization strategies, the weight allocation ratio is determined according to the difference between the disaster scenario matching degree score and the set scoring threshold. For example, when the disaster scenario matching degree score is 0.8, the difference is 0.8 - 0.7 = 0.1. For the first candidate optimization strategy, assume that according to the difference and the preset weight allocation rule, its weight allocation ratio is determined to be 0.3; the weight allocation ratio of the second candidate optimization strategy is 0.25; the weight allocation ratio of the third candidate optimization strategy is 0.2; the weight allocation ratio of the fourth candidate optimization strategy is 0.15; the weight allocation ratio of the fifth candidate optimization strategy is 0.1. The weight allocation rule here can be linear allocation according to the difference size or allocation through a more complex non-linear function, depending on the actual application requirements and model training results.

[0145] Step S142: Extract the starting coordinates of the resource scheduling path, the evacuation route density coefficient, and the device deployment time window for each strategy in the set of candidate optimization strategies, and perform conflict detection with the set of available resource coordinates, the real-time traffic density data, and the device status data of the current disaster response terminal.

[0146] Extract the key parameters of each strategy from the set of candidate optimization strategies, such as the starting coordinates of the resource scheduling path, the evacuation route density coefficient, and the device 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 of the first candidate optimization strategy are (110, 210), the second is (120, 220), etc. At the same time, obtain the set of available resource coordinates of the current disaster response terminal. Assume that there are three available resource points in the set of available resource coordinates, and the coordinates are (105, 205), (115, 215), (125, 225) respectively. For the evacuation route density coefficient, each candidate optimization strategy has a corresponding value, such as the first being 0.9, the second being 0.85, etc. The real-time traffic density data can be obtained through traffic monitoring devices. For example, the real-time traffic density of a certain road is 50 vehicles per kilometer. The device status data includes information such as whether the device is available and the available time. For example, a certain rescue device is available from 11 am to 2 pm. Compare these parameters to detect whether there are conflicts. For example, if the starting coordinates of the resource scheduling path of a certain candidate optimization strategy are far from all the available resource coordinates, 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 this candidate optimization strategy, there is also a conflict; if the device deployment time window does not match the available time in the device status data, there is also a conflict.

[0147] Step S143: For the candidate optimization strategies with resource coordinate conflicts, re-plan the path starting coordinate to the nearest available resource coordinate, and update the path length and estimated arrival time.

[0148] When it is detected that a certain candidate optimization strategy has a resource coordinate conflict, the path starting coordinate needs to be re-planned. Assume that the original resource scheduling path starting coordinate of the first candidate optimization strategy is (110, 210). After comparing with the set of available resource coordinates, it is found that the nearest available resource coordinate is (115, 215). Then update the path starting coordinate of this candidate optimization strategy to (115, 215). Next, update the path length and estimated arrival time. The path length can be calculated by the Geographic Information System (GIS) based on the new starting coordinate and the coordinates of the target disaster area. Assume that the original path length is 20 kilometers. After re-planning, the new path length calculated by the GIS is 18 kilometers. The estimated arrival time can be calculated based on the path length and the average speed of the transportation vehicle. Assume that the average speed of the transportation vehicle is 60 kilometers per hour. Then the original estimated arrival time is 20÷60×60 = 20 minutes, and the updated estimated arrival time is 18÷60×60 = 18 minutes.

[0149] Step S144: For the evacuation routes with traffic density exceeding the limit, divide the route into multiple sub-sections according to the density coefficient threshold, and assign a backup route identifier and a turning instruction to each sub-section.

[0150] For the evacuation routes with traffic density exceeding the limit, process them 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, exceeding the threshold. Divide this evacuation route into multiple sub-sections according to certain rules, such as dividing it according to the road milestone or key intersections. Assume that it is divided into three sub-sections. Assign a backup route identifier and a turning instruction to each sub-section. The backup route identifier can be a unique number. For example, the backup route identifier of sub-section 1 is R001, that of sub-section 2 is R002, and that of sub-section ₃ is R003. The turning instruction can be determined based on the Geographic Information System and traffic monitoring data. For example, at a certain intersection in sub-section 1, a left turn is required to enter the backup route R001.

[0151] Step S145: For the strategies with mismatched device deployment time windows and device status, re-assign the deployment sequence according to the device available time, and generate device deployment instructions with timestamp alignment.

[0152] When the device deployment time window of a certain candidate optimization strategy does not match the device status, it is necessary to reallocate the deployment sequence. Suppose a certain candidate optimization strategy plans to deploy a device at 10:00 am, but the available time of this device is from 11:00 am to 2:00 pm. According to the available time of the device, adjust the deployment time of this device to 11:00 am, and rearrange the deployment sequences of other devices to ensure the rationality of the entire deployment process. Generate device deployment instructions with timestamp alignment, such as clearly indicating in the device deployment instructions that Device 1 starts to be deployed at 11:00 am, Device 2 starts to be deployed at 11:30 am, etc.

[0153] Step S146: Sort the updated candidate optimization strategies according to the execution priority score, and select the top N strategies to generate the set of disaster emergency decision instructions, where the disaster emergency decision instructions include the adjusted resource scheduling path coordinate sequence, evacuation sub-section turning instructions, and device deployment timestamps.

[0154] Sort the candidate optimization strategies after the above conflict handling and adjustment according to the execution priority score. Suppose the execution priority scores from high to low are the first candidate optimization strategy 0.5, the second 0.45, the third 0.4, etc. Select the top N strategies. Here, assume N is 3, that is, select the three strategies with the highest execution priority scores. Extract the adjusted resource scheduling path coordinate sequence, evacuation sub-section turning 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 of the first strategy is (115, 215), (120, 220), (125, 225); the evacuation sub-section turning instruction is to turn left at intersection A; the device deployment timestamp is 11:00 am. Integrate this information into the set of disaster emergency decision instructions.

[0155] Step S147: Package the set of disaster emergency decision instructions into the format of the disaster response terminal protocol, add the execution time limit identifier and retry times limit parameter for each disaster emergency decision instruction, and send it to the corresponding disaster response terminal.

[0156] Package the generated set of disaster emergency decision instructions according to the format of the disaster response terminal protocol. The format of the disaster response terminal protocol stipulates the transmission format and data structure of the instructions to ensure that the disaster response terminal can correctly parse and execute these instructions. Add the execution time limit identifier for each disaster emergency decision instruction. For example, it is stipulated that a certain instruction must be executed before 3:00 pm. At the same time, add the retry times limit parameter. Suppose the retry times limit for a certain instruction is 3 times. If this instruction fails during execution, it can be retried at most 3 times. Finally, send the packaged set of disaster emergency decision instructions to the corresponding disaster response terminals, which can be rescue vehicles, handheld devices of rescue personnel, etc.

[0157] Step S150: Obtain the decision execution feedback data returned by the disaster response terminal, and perform incremental parameter optimization processing on the multi-modal dynamic decision-making model based on the decision execution feedback data.

[0158] In this embodiment, after sending the disaster emergency decision instruction set to the disaster response terminal, it is necessary to obtain the decision execution feedback data and perform incremental parameter optimization processing on the multi-modal dynamic decision-making model. The specific steps are as follows: Step S151: Receive the instruction execution status parameters uploaded by the disaster response terminal. 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 start-up delay time of the equipment deployment.

[0159] 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, assume that the expected arrival time of a certain resource scheduling instruction is 11 am, and the actual arrival time is 11:10 am. 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 a certain evacuation sub-section is 0.8, indicating a relatively high degree of congestion on this section. The actual start-up delay time of the equipment deployment refers to the difference between the actual start-up time of the equipment and the start-up time specified in the instruction. Assume that the instruction stipulates that the equipment starts at 11 am, and it actually starts at 11:15 am. Then the actual start-up delay time is 15 minutes.

[0160] Step S152: Calculate the absolute error between the actual arrival time deviation value and the instruction expected time, the ratio of the real-time congestion coefficient to the preset threshold, and the difference between the actual start-up delay time and the deployment timestamp, and generate a set of policy execution deviation indicators.

[0161] For the absolute error between the actual arrival time deviation value and the instruction expected time, taking the previous example, the expected arrival time is 11 am, that is, 660 minutes (starting from 0 o'clock), and the actual arrival time is 11:10 am, that is, 670 minutes. The absolute error is 670 - 660 = 10 minutes. For the ratio of the real-time congestion coefficient to the preset threshold, assume that the preset threshold is 0.6, and the real-time congestion coefficient of a certain evacuation sub-section is 0.8. Then the ratio is 0.8 ÷ 0.6 = 1.33. For the difference between the actual start-up delay time and the deployment timestamp, assume that the deployment timestamp is 11 am, that is, 660 minutes, and the actual start-up time is 11:15 am, that is, 675 minutes. The difference is 675 - 660 = 15 minutes. Combine these calculation results to generate a set of policy execution deviation indicators, such as [10, 1.33, 15].

[0162] Step S153: Collect the 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 state update parameters.

[0163] Collect the real-time environmental data of the current disaster area. The latest sampling value of the wind speed fluctuation sequence can be obtained through a wind speed sensor. Suppose the latest sampling value is 5 m / s. The current slope of the temperature change gradient can be obtained by analyzing the temperature change data over a period of time. For example, in the past hour, the temperature has risen from 20°C to 22°C, so the current slope of the temperature change gradient is (22 - 20) ÷ 1 = 2 °C / hour. The peak displacement of the geological vibration frequency spectrum can be measured by devices such as seismographs. Suppose the peak displacement is 0.5 cm. Combine these data to generate environmental state update parameters, such as [5, 2, 0.5].

[0164] Step S154: Input the set of policy execution deviation metrics and the environmental state update parameters into the feedback learning channel of the multi-modal dynamic decision-making model, and calculate the correction amount of the historical matching degree score for each candidate optimization policy. The correction amount of the historical matching degree score is a linear combination of the deviation metric and the environmental parameter.

[0165] Input the set of policy execution deviation metrics and the environmental state update parameters into the feedback learning channel of the multi-modal dynamic decision-making model. Suppose the set of policy execution deviation metrics is [10, 1.33, 15], and the environmental state update parameters are [5, 2, 0.5]. Assign corresponding weights to each parameter. For example, the weights of the set of policy execution deviation metrics are 0.4, 0.3, and 0.3 respectively, and the weights of the environmental state update parameters are 0.2, 0.3, and 0.5 respectively. For a certain candidate optimization policy, calculate the correction amount of its historical matching degree score through a linear combination, that is, 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 correction amount of the historical matching degree score for each candidate optimization policy in turn.

[0166] 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 multi-modal dynamic decision-making model according to the correction amount of the historical matching degree score.

[0167] Adjust the weights of each channel in the multi-modal dynamic decision-making model according to the calculated historical matching degree score correction amount. Assume that the historical matching degree score correction amount is 10.749. For the pixel gradient similarity weight of the image feature matching channel, which was originally 0.3, according to a certain adjustment rule, it is adjusted to 0.3 + 10.749×0.01 = 0.40749. For the covariance matrix spectral norm weight of the environmental feature matching channel, which was originally 0.2, it is adjusted to 0.2 + 10.749×0.005 = 0.253745. For the dynamic time warping alignment weight of the temporal feature matching channel, which was originally 0.2, it is adjusted to 0.2 + 10.749×0.008 = 0.285992. For the invulnerability score weight of the topological feature matching channel, which was originally 0.3, it is adjusted to 0.3 + 10.749×0.012 = 0.428988. The adjustment rule here can be set according to the training and actual application of the model to ensure that the model can better adapt to new disaster scenarios and decision feedback.

[0168] Step S156: Update the image feature library, environmental feature library, temporal feature library, and topological feature library of the pre-stored disaster cases by using a sliding window mechanism, and store the current disaster feature set and the adjusted policy parameters as new cases into the corresponding feature libraries.

[0169] Step S1561: Set the maximum number of stored cases in the feature libraries of the pre-stored disaster cases. When the total number exceeds the limit due to new cases, start the sliding window elimination mechanism.

[0170] Set the maximum number of stored cases in the image feature library, environmental feature library, temporal feature library, and topological feature library of the pre-stored disaster cases. Assume that the maximum number of stored cases in each feature library is 100. When the current disaster feature set and the adjusted policy parameters are to be stored as new cases in the feature library, if the total number exceeds 100 after storage, start the sliding window elimination mechanism.

[0171] Step S1562: Calculate the product of the most recent access timestamp and the global matching degree score of each historical case as the case activity index.

[0172] For each historical case of the pre-stored disaster cases, calculate the product of its most recent access timestamp and the global matching degree score as the case activity index. Assume that the most recent access timestamp of a certain historical case is 1000 (calculated with a certain starting time as 0), and the global matching degree score is 0.8. Then the case activity index of this case is 1000×0.8 = 800. Calculate the case activity index of each historical case in turn.

[0173] Step S1563: Arrange the 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.

[0174] Arrange all historical cases in ascending order according to the case activity index. Assume M is 5, that is, eliminate the 5 cases with the lowest case activity index. For example, after sorting, the indexes of the 5 cases with the lowest case activity index are 100, 200, 300, 400, and 500 respectively. Delete these 5 cases from the corresponding feature library to free up storage space.

[0175] Step S1564: Encode the disaster image dynamic feature, disaster environment association feature, disaster response time series feature, and geographical topology distribution feature of the current disaster feature set into standardized feature vectors respectively.

[0176] Encode each type of feature in the current disaster feature set into a standardized feature vector. Taking the disaster image dynamic feature as an example, assume that the disaster image dynamic feature includes the landform deformation gradient feature, the thermal diffusion direction vector, and the water body coverage area change rate. Convert these features into standardized feature vectors according to a certain encoding rule. For example, the landform deformation gradient feature is 0.1 in the horizontal direction and 0.2 in the vertical direction, the thermal diffusion direction vector is (0.3, 0.4), and the water body coverage area change rate is 0.1. Combine and standardize these values to obtain a standardized feature vector [0.1, 0.2, 0.3, 0.4, 0.1]. Similarly, encode the disaster environment association feature, disaster response time series feature, and geographical topology distribution feature to obtain the corresponding standardized feature vectors.

[0177] Step S1565: Associate the standardized feature vector with the corresponding adjusted policy parameter and execution deviation index to generate a new case data block.

[0178] Associate the encoded standardized feature vector with the corresponding adjusted policy parameter and execution deviation index. For example, for the standardized feature vector [0.1, 0.2, 0.3, 0.4, 0.1] of the disaster image dynamic feature, the corresponding adjusted policy parameter is the starting coordinate of the resource scheduling path (115, 215), and the execution deviation index is the actual arrival time deviation value of 10 minutes. Associate this information together to generate a new case data block. Similarly, perform similar processing on the standardized feature vectors of other features.

[0179] Step S1566: Append the new case data block to the end of the image feature library, environment feature library, time series feature library, and topology feature library of the pre-stored disaster cases according to the feature type respectively, and update the index identifiers of the corresponding feature libraries.

[0180] Append the generated new case data blocks to the end of the corresponding feature libraries according to the feature types respectively. For example, append the new case data block of the dynamic features of disaster images to the end of the image feature library, append the new case data block of the associated features of the disaster environment to the end of the environmental feature library, append the new case data block of the sequential features of disaster response to the end of the sequential feature library, and append the new case data block of the geographical topological distribution features to the end of the topological feature library. After the appending is completed, update the index identifiers of the corresponding feature libraries to ensure that these new case data can be accurately located and retrieved. For example, the original index identifiers of the image feature library range from 1 to 95, and after appending the new case data block, the index identifiers are updated to range from 1 to 96. In this way, the update of the feature libraries of the pre-stored disaster cases is completed, enabling the model to utilize the latest disaster information for subsequent decision-making analysis.

[0181] During the entire emergency disaster dynamic decision-making process, various disaster situations can be more accurately addressed, improving the efficiency and accuracy of disaster emergency decision-making and reducing the losses caused by disasters. For example, when acquiring real-time disaster monitoring data, through the comprehensive collection of multiple data sources and timestamp alignment processing, the integrity and accuracy of the data are ensured. During the feature extraction process, different extraction methods are adopted for different types of data, resulting in a multi-modal disaster feature set that can reflect the disaster scenario. When performing the disaster scenario matching degree analysis, through multi-channel matching calculations and weighted summation, similar cases are accurately found, and an optimized strategy is generated. In the decision execution and feedback optimization phase, the model is adjusted according to the actual execution situation, enabling the model to continuously adapt to new disaster scenarios and improving the reliability of the decision-making.

[0182] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an emergency disaster dynamic decision-making system 100 based on a multi-modal AI large model that can implement the idea of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the emergency disaster dynamic decision-making system 100 based on the multi-modal AI large model and is used to execute the functions in the present application.

[0183] The emergency disaster dynamic decision-making system 100 based on the multi-modal 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 the multi-modal AI large model of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0184] 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 a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, 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-transitory storage media, or any combination thereof. The methods 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 further includes an I / O interface 150 between the computer and other input / output devices.

[0185] For ease of explanation, only one processor is described in the emergency disaster dynamic decision-making system 100 based on the multimodal AI large model. However, it should be noted that the emergency disaster dynamic decision-making system 100 based on the multimodal AI large model in the present application may also include multiple processors. Therefore, the steps executed by one processor described in the present application can also be jointly executed or separately executed by multiple processors. For example, if the processor of the emergency disaster dynamic decision-making system 100 based on the multimodal AI large model executes step A and step B, it should be understood that step A and step B can also be jointly executed by two different processors or separately executed 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 jointly execute steps A and B.

[0186] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. 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.

[0187] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An emergency disaster dynamic decision-making method based on a multi-modal AI large model, characterized in that, The method includes: Obtaining a real-time disaster monitoring data set of a target disaster area, where 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; Performing feature extraction on the real-time disaster monitoring data set to obtain a multi-modal disaster feature set of the disaster scene, where the multi-modal disaster feature set includes disaster image dynamic features, disaster environment association features, disaster response time series features, and geographical topology distribution features; Invoking a pre-trained multi-modal dynamic decision-making model to perform disaster scene matching degree analysis processing on the multi-modal disaster feature set, generating a disaster scene matching degree score of the target disaster area and a disaster response optimization strategy set; Based on the disaster scene matching degree score, screening a target optimization strategy from the disaster response optimization strategy set, 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 multi-modal dynamic decision-making model based on the decision execution feedback data.

2. The emergency disaster dynamic decision-making method based on the multi-modal AI large model according to claim 1, characterized in that The obtaining of the real-time disaster monitoring data set of the target disaster area includes: Invoking satellite remote sensing equipment to collect disaster image acquisition data of the target disaster area, where the disaster image acquisition data includes a landform change pixel matrix of the disaster area, a thermal distribution gradient information, and a water body coverage change trajectory; Obtaining disaster environment sensing data uploaded in real time by a plurality of environment sensing nodes deployed in the target disaster area, where 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 disaster historical response record data of the target disaster area from a disaster emergency database, where the disaster historical response record data includes historical disaster type labels, historical emergency response timestamp sequences, historical resource scheduling paths, and historical loss assessment reports; Invoking a geographic information system to analyze the geographical coordinate boundary of the target disaster area, generating disaster geographical distribution topology data, where the disaster geographical distribution topology data includes an elevation gradient distribution map, a road connectivity network, and a set of coordinates of residential aggregation areas; Performing timestamp alignment processing on the disaster image acquisition data, the disaster environment sensing data, the disaster historical response record data, and the disaster geographical distribution topology data to obtain the real-time disaster monitoring data set.

3. The emergency disaster dynamic decision-making method based on the multi-modal AI large model according to claim 1, characterized in that, The performing of feature extraction on the real-time disaster monitoring data set to obtain a multi-modal disaster feature set of the disaster scene includes: Invoking a pre-trained disaster image encoder to perform convolutional feature extraction processing on the disaster image acquisition data to obtain disaster image dynamic features, where the disaster image dynamic features include landform deformation gradient features, thermal diffusion direction vectors, and water body coverage area change rates; Performing time series analysis processing on the disaster environment sensing data to extract disaster environment association features, where the disaster environment association features include a covariance matrix of wind speed and temperature, a frequency domain correlation degree of humidity and geological vibration, and a time lag correlation coefficient of multi-sensing data; Input the disaster history response record data into the time series feature extraction network to generate disaster response time series features, where 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 loss and response time; Perform connectivity analysis processing on the disaster geographical distribution topology data through a graph neural network to generate geographical topology distribution features, where the geographical topology distribution features include a road network invulnerability score, the shortest path distance between a residential area and the disaster center, and a mapping relationship between elevation gradient and disaster diffusion speed; Perform feature dimension alignment and splicing processing on the disaster image dynamic features, disaster environment association features, disaster response time series features, and geographical topology distribution features to generate the multi-modal disaster feature set.

4. The emergency disaster dynamic decision-making method based on the multi-modal AI large model according to claim 1, characterized in that Call the pre-trained multi-modal dynamic decision-making model to perform disaster scenario matching degree analysis processing on the multi-modal disaster feature set, generating a disaster scenario matching degree score for the target disaster area and a set of disaster response optimization strategies, including: Input the multi-modal disaster feature set into the multi-modal dynamic decision-making model, and calculate an image matching coefficient set, an environment matching coefficient set, a time series matching coefficient set, and a topology matching coefficient set between the multi-modal disaster feature set and pre-stored disaster cases; Perform 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 degree score for each pre-stored disaster case; Arrange the pre-stored disaster cases in descending order according to the global matching degree score, select the cases with a global matching degree score higher than the set score threshold as the similar case set, and extract the historical response strategies and their loss reduction rates corresponding to each case in the similar case set; Perform adaptive adjustment of the strategy parameters and the current disaster environment parameters for the historical response strategies, 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; Calculate the execution priority score of each candidate optimization strategy based on the loss reduction rate and the adjusted strategy parameters, take the set of candidate optimization strategies sorted by the execution priority score as the set of disaster response optimization strategies, and take the global matching degree score as the disaster scenario matching degree score.

5. The emergency disaster dynamic decision-making method based on a multi-modal AI large model according to claim 4, characterized in that, The step of inputting the multi-modal disaster feature set into the multi-modal dynamic decision-making model and calculating the image matching coefficient set, the environment matching coefficient set, the time series matching coefficient set, and the topology matching coefficient set between the multi-modal disaster feature set and pre-stored disaster cases includes: Input the multi-modal disaster feature set into the image feature matching channel of the multi-modal dynamic decision-making model, calculate the pixel gradient similarity between the disaster image dynamic features and each case in the image feature library of the pre-stored disaster cases, and generate an image matching coefficient set; Input the disaster environment - related features into the environmental feature matching channel of the multi - modal dynamic decision - making model, extract the environmental parameter covariance matrices of each case in the environmental feature library of pre - stored disaster cases, calculate the spectral norm ratios of the covariance matrices, and generate a set of environmental matching coefficients; Input the disaster response time - series features into the time - series feature matching channel of the multi - modal dynamic decision - making model, perform dynamic time warping alignment with the response delay time distributions of each case in the time - series feature library of pre - stored disaster cases, and generate a set of time - series matching coefficients; Input the geographical topological distribution features into the topological feature matching channel of the multi - modal dynamic decision - making model, compare the road network invulnerability scores and the shortest path distances of residential areas of each case in the topological feature library of pre - stored disaster cases, and generate a set of topological matching coefficients.

6. The emergency disaster dynamic decision-making method based on the multi-modal AI large model according to claim 4, wherein, Screen the target optimization strategies from the set of disaster response optimization strategies based on the disaster scenario matching degree score, generate a set of disaster emergency decision instructions, and send the set of disaster emergency decision instructions to the disaster response terminal, including: Determine the weight allocation ratios of each candidate optimization strategy in the set of disaster response optimization strategies according to the difference between the disaster scenario matching degree score and the set score threshold; Extract the starting point coordinates of the resource scheduling path, the evacuation route density coefficient, and the equipment deployment time window of each strategy in the candidate optimization strategy set, and perform conflict detection with the set of available resource coordinates, real - time traffic density data, and equipment status data of the current disaster response terminal; For the candidate optimization strategies with resource coordinate conflicts, re - plan the starting point coordinates of the path to the nearest available resource coordinates, and update the path length and estimated arrival time; For the evacuation routes with traffic density exceeding the limit, divide the route into multiple sub - sections according to the density coefficient threshold, and assign backup route identifiers and turning instructions to each sub - section; For the strategies with mismatched equipment deployment time windows and equipment status, re - allocate the deployment sequence according to the available time of the equipment, and generate equipment deployment instructions with timestamp alignment; Sort the updated candidate optimization strategies according to the execution priority score, select the top N strategies to generate the set of disaster emergency decision instructions, where the disaster emergency decision instructions include the adjusted resource scheduling path coordinate sequence, evacuation sub - section turning instructions, and equipment deployment timestamps; Encapsulate the set of disaster emergency decision instructions into the protocol format of the disaster response terminal, add the execution time limit identifier and retry times limit parameter of each disaster emergency decision instruction, and send them to the corresponding disaster response terminal.

7. The emergency disaster dynamic decision-making method based on the multi-modal AI large model according to claim 4, wherein, Obtain the decision execution feedback data returned by the disaster response terminal, and perform incremental parameter optimization processing on the multi - modal dynamic decision - making model based on the decision execution feedback data, including: Receive the instruction execution status parameters uploaded by the disaster response terminal, where 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 start delay time of the equipment deployment. Calculate the absolute error between the actual arrival time deviation value and the instruction expected 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, and generate a set of policy execution deviation metrics; Collect the 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 state update parameters; Input the set of policy execution deviation metrics and the environmental state update parameters into the feedback learning channel of the multi-modal dynamic decision-making model, and calculate the correction amount of the historical matching degree score for each candidate optimization policy. The correction amount of the historical matching degree score is a linear combination of the deviation metrics and the environmental parameters; 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 multi-modal dynamic decision-making model according to the correction amount of the historical matching degree score; Update the image feature library, environmental feature library, temporal feature library, and topological feature library of the pre-stored disaster cases by using a sliding window mechanism, and store the current disaster feature set and the adjusted policy parameters as new cases in the corresponding feature libraries.

8. The emergency disaster dynamic decision-making method based on a multi-modal AI large model according to claim 7, wherein, The step of updating the image feature library, environmental feature library, temporal feature library, and topological feature library of the pre-stored disaster cases by using a sliding window mechanism, and storing the current disaster feature set and the adjusted policy parameters as new cases in the corresponding feature libraries includes: Set the maximum number of stored cases in the feature library of the pre-stored disaster cases. When the total number exceeds the limit due to new cases, start the sliding window elimination mechanism; Calculate the product of the most recent access timestamp and the global matching degree score of each historical case as the case activity index; Arrange the historical cases in ascending order of the case activity index, and eliminate the M cases with the lowest case activity index to free up storage space; Encode the disaster image dynamic features, disaster environment correlation features, disaster response temporal features, and geographical topological distribution features of the current disaster feature set into standardized feature vectors respectively; Associate the standardized feature vectors with the corresponding adjusted policy parameters and execution deviation metrics to generate a new case data block; 9. The emergency disaster dynamic decision-making method based on a multi-modal AI large model according to claim 3, wherein 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 the feature type respectively, and update the index identifiers of the corresponding feature libraries. The step of inputting the disaster historical response record data into the temporal feature extraction network to generate disaster response temporal features includes: Perform time slicing processing on the disaster historical response record data to generate a response action sequence within an equally spaced time window; Extract the response action type distribution vector, the resource scheduling quantity change curve, and the loss accumulation rate within each time window; Call a long short-term memory network to model the temporal dependence relationship of the response action sequence and generate a hidden state vector; Calculate the correlation weights between the hidden state vectors of different time windows through the self-attention mechanism to generate global temporal attention features; Perform weighted fusion of the global temporal attention features and the hidden state vectors of each time window to obtain the disaster response temporal features.

10. An emergency disaster dynamic decision-making system based on a multi-modal AI large model, characterized in that, 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 multi-modal AI large model described in any one of claims 1-9 above.

Images