Emergency situation intelligent decision-making method, device and system based on deep learning

Abstract

This invention discloses an intelligent emergency situation decision-making method, device, and system based on deep learning. First, multimodal data from disaster-stricken areas is collected and fused into a unified embedding vector. Then, a spatiotemporal perception neural network is used to extract temporal and spatial features, constructing a state sequence containing spatiotemporal evolution characteristics. Next, a dynamic causal graph is constructed based on this sequence to identify potential causal relationships, thereby predicting future emergency situation development trends. Finally, combining historical response experience with current resource constraints, an emergency decision-making strategy accurately adapted to the actual situation of the disaster-stricken area is generated. This invention enables accurate perception of disaster situations, evolution prediction, and intelligent decision generation under resource constraints.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an emergency situation intelligent decision-making method, device and system based on deep learning. Background Technology

[0002] In the field of emergency management, a typical approach in existing technologies is to construct disaster situation awareness systems based on data from single sources such as remote sensing images, ground sensors, or social media text. For example, some systems characterize the disaster situation by manually designing feature extraction rules (such as pixel change rate of disaster areas and frequency statistics of public opinion keywords) and generate response instructions by combining them with preset logical judgment conditions. Another approach uses traditional machine learning models (such as support vector machines or shallow neural networks) to predict resource allocation schemes by performing pattern matching on the static features of historical disaster cases. In addition, recent studies have attempted to introduce time series models (such as LSTM) or spatial graph networks to model disaster states, but due to the lack of deep fusion capabilities for multimodal heterogeneous data, it is difficult to achieve dynamic coupling analysis of the evolution process of disaster events. At the same time, existing systems usually rely on expert knowledge bases or static causal graphs in the causal reasoning stage, and cannot dynamically correct the correlation of disaster chain reactions based on real-time situation data. This results in low reliability of prediction results for future emergency scenarios, and the generation of final decision-making strategies often ignores the synergistic optimization of historical handling experience and resource constraints, making it difficult to meet the dynamic response needs in complex disaster scenarios.

[0003] Therefore, existing technologies are insufficient to achieve accurate perception of disaster situations, prediction of their evolution, and intelligent decision-making under resource constraints, thus affecting the effectiveness of emergency management. Summary of the Invention

[0004] This invention provides a method, device, and system for intelligent decision-making in emergency situations based on deep learning, which can achieve accurate perception of disaster situations, prediction of their evolution, and intelligent decision generation under resource constraints, effectively improving the effectiveness of emergency management.

[0005] One embodiment of the present invention provides an intelligent emergency situation decision-making method based on deep learning, comprising:

[0006] Multimodal data from disaster-stricken areas are fused and processed to generate a unified representation of the multimodal emergency situation embedding vector;

[0007] Feature extraction is performed on the multimodal emergency situation embedding vector using a spatiotemporal awareness neural network model. Modeling is performed in both the time series dimension and the spatial correlation dimension, and the output is an emergency situation state sequence containing spatiotemporal evolution characteristics.

[0008] A dynamic causal graph is constructed using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors;

[0009] Based on potential causal relationships, the future development trend of the emergency situation in the disaster-stricken area is predicted, and the causal evolution result of the emergency situation in the disaster-stricken area is generated.

[0010] Based on the causal evolution results of the emergency situation, and combined with historical response experience data and current resource constraints, emergency decision-making strategies for the disaster-stricken areas are generated.

[0011] As an improvement to the above scheme, the step of fusing multimodal data from disaster-stricken areas to generate a unified representation of the multimodal emergency situation embedding vector includes the following sub-steps:

[0012] Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data.

[0013] A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix.

[0014] Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing.

[0015] As an improvement to the above scheme, the step of extracting features from the multimodal emergency situation embedding vector using a spatiotemporal awareness neural network model, modeling in both the time series dimension and the spatial correlation dimension, and outputting an emergency situation state sequence containing spatiotemporal evolution characteristics includes the following sub-steps:

[0016] The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate;

[0017] A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network;

[0018] The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector.

[0019] The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster.

[0020] As an improvement to the above scheme, the step of constructing a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors includes the following sub-steps:

[0021] Map each dimension of the emergency situation state sequence to a node of a causal graph, and initialize the causal strength parameters between the nodes;

[0022] Based on the causal strength parameters between nodes, the causal relationships between nodes in the causal graph are analyzed, and causal edges are determined.

[0023] A dynamic causal graph is generated based on each node and the causal edges between nodes;

[0024] Based on dynamic causal graphs, potential causal relationships between disaster events and influencing factors are identified.

[0025] As an improvement to the above scheme, the step of predicting the future emergency situation development trend of the disaster-stricken area based on potential causal relationships and generating the causal evolution result of the emergency situation in the disaster-stricken area includes the following sub-steps:

[0026] Extract the node causal association matrix based on the potential causal relationship as prior causal knowledge;

[0027] The emergency situation state sequence and the causal correlation matrix are input into a pre-trained time series prediction model to perform multi-step long trend prediction and obtain multi-step trend prediction results.

[0028] Based on the multi-step long-term trend prediction results, the situation evolution path of the disaster-stricken area under different emergency measures is simulated to generate the causal evolution results of the emergency situation of the disaster-stricken area.

[0029] As an improvement to the above scheme, the step of generating an emergency decision-making strategy for the disaster-stricken area based on the causal evolution results of the emergency situation, combined with historical response experience data and current resource constraints, includes the following sub-steps:

[0030] Acquire historical emergency response experience data, including historical emergency plans, resource consumption data, and response effects;

[0031] Transform the current available rescue resources, team locations, and transportation network status into resource constraints.

[0032] Using a pre-trained deep learning model, the initial emergency decision-making strategy is generated by taking the causal evolution results of the emergency situation, the historical handling experience data, and the resource constraints as inputs.

[0033] Based on the causal logic of the dynamic causal graph, the potential secondary disaster risks caused by the initial decision-making strategy are verified, and the final emergency decision-making strategy for the disaster-stricken area is output.

[0034] Another embodiment of the present invention provides an emergency situation intelligent decision-making device based on deep learning, comprising:

[0035] The data fusion module is used to fuse multimodal data from disaster-stricken areas and generate a unified representation of the multimodal emergency situation embedding vector.

[0036] The feature extraction module is used to extract features from the multimodal emergency situation embedding vector through a spatiotemporal perception neural network model, modeling in the time series dimension and spatial correlation dimension respectively, and outputting an emergency situation state sequence containing spatiotemporal evolution characteristics.

[0037] The causal identification module is used to construct a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors;

[0038] The prediction module is used to predict the future development trend of the emergency situation in the disaster-stricken area based on potential causal relationships, and generate the causal evolution results of the emergency situation in the disaster-stricken area.

[0039] The strategy generation module is used to generate emergency decision-making strategies for the disaster-stricken area based on the causal evolution results of the emergency situation, combined with historical handling experience data and current resource constraints.

[0040] As an improvement to the above solution, the data fusion module is specifically used for:

[0041] Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data.

[0042] A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix.

[0043] Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing.

[0044] As an improvement to the above scheme, the feature extraction module is specifically used for:

[0045] The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate;

[0046] A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network;

[0047] The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector.

[0048] The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster.

[0049] Another embodiment of the present invention provides an emergency situation intelligent decision-making system based on deep learning, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the emergency situation intelligent decision-making method based on deep learning described in the above embodiment of the invention.

[0050] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0051] Multimodal data from disaster-stricken areas is collected using multi-source devices, and fusion technology is employed to integrate image, text, and voice data into a unified multimodal emergency situation embedding vector, eliminating data noise and redundancy and enabling the data to more comprehensively and accurately reflect the emergency situation. Using a spatiotemporal awareness neural network model, feature extraction and modeling are performed on the embedding vector in both the time series and spatial correlation dimensions to deeply explore the spatiotemporal evolution patterns of the disaster situation, thereby outputting an emergency situation state sequence containing spatiotemporal characteristics. This state sequence is used to construct a dynamic causal graph to identify potential causal relationships between disaster events and influencing factors, providing a basis for prediction. Based on these causal relationships, future emergency situation development trends are predicted, generating causal evolution results. Finally, combined with historical response experience data and current resource constraints, emergency decision-making strategies are generated. In this process, multimodal data fusion provides a comprehensive data foundation for subsequent analysis, spatiotemporal feature extraction accurately captures the spatiotemporal changes of the disaster, dynamic causal graph construction enables precise identification of potential causal relationships, and the decision-making strategies generated based on multi-factor integration are more aligned with actual needs. In summary, the embodiments of the present invention achieve deep fusion and analysis of emergency situation data through deep fusion, dynamic modeling and causal reasoning mechanisms, accurately grasp the spatiotemporal characteristics and causal relationships of disasters, thereby enabling accurate perception of disaster situations, evolution prediction and intelligent decision generation under resource constraints. Attached Figure Description

[0052] Figure 1This is a flowchart illustrating an intelligent emergency situation decision-making method based on deep learning, provided in an embodiment of the present invention.

[0053] Figure 2 This is a schematic diagram of the structure of an emergency situation intelligent decision-making device based on deep learning, provided in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the structure of an emergency situation intelligent decision-making system based on deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] See Figure 1 This is a flowchart illustrating an intelligent emergency situation decision-making method based on deep learning, provided in an embodiment of the present invention. The intelligent emergency situation decision-making method based on deep learning includes the following steps:

[0057] S10, fuse and process multimodal data from disaster-stricken areas to generate a unified representation of multimodal emergency situation embedding vector;

[0058] S11, the feature extraction of the multimodal emergency situation embedding vector is performed through a spatiotemporal perception neural network model, and models are performed in the time series dimension and spatial correlation dimension respectively, and the emergency situation state sequence containing spatiotemporal evolution characteristics is output.

[0059] S12, construct a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors;

[0060] S13, predict the future development trend of the emergency situation in the disaster-stricken area based on potential causal relationships, and generate the causal evolution result of the emergency situation in the disaster-stricken area;

[0061] S14. Based on the causal evolution results of the emergency situation and combined with historical handling experience data and current resource constraints, an emergency decision-making strategy for the disaster-stricken area is generated.

[0062] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0063] Multimodal data from disaster-stricken areas is collected using multi-source devices, and fusion technology is employed to integrate image, text, and voice data into a unified multimodal emergency situation embedding vector, eliminating data noise and redundancy and enabling the data to more comprehensively and accurately reflect the emergency situation. Using a spatiotemporal awareness neural network model, feature extraction and modeling are performed on the embedding vector in both the time series and spatial correlation dimensions to deeply explore the spatiotemporal evolution patterns of the disaster situation, thereby outputting an emergency situation state sequence containing spatiotemporal characteristics. This state sequence is used to construct a dynamic causal graph to identify potential causal relationships between disaster events and influencing factors, providing a basis for prediction. Based on these causal relationships, future emergency situation development trends are predicted, generating causal evolution results. Finally, combined with historical response experience data and current resource constraints, emergency decision-making strategies are generated. In this process, multimodal data fusion provides a comprehensive data foundation for subsequent analysis, spatiotemporal feature extraction accurately captures the spatiotemporal changes of the disaster, dynamic causal graph construction enables precise identification of potential causal relationships, and the decision-making strategies generated based on multi-factor integration are more aligned with actual needs. In summary, the embodiments of the present invention achieve deep fusion and analysis of emergency situation data through deep fusion, dynamic modeling and causal reasoning mechanisms, accurately grasp the spatiotemporal characteristics and causal relationships of disasters, thereby enabling accurate perception of disaster situations, evolution prediction and intelligent decision generation under resource constraints, effectively improving the effectiveness of emergency management.

[0064] As one example, the process of fusing multimodal data from disaster-stricken areas to generate a unified representation of a multimodal emergency situation embedding vector includes the following sub-steps:

[0065] Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data.

[0066] A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix.

[0067] Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing.

[0068] In this embodiment, for the diverse and widely sourced multimodal data (satellite imagery, meteorological data, and public opinion text data) from disaster-stricken areas, spatiotemporal alignment technology is first used to eliminate differences in time and space dimensions, generating a time-synchronized multimodal data set. This lays the foundation for subsequent analysis and ensures data consistency and comparability. Next, a cross-modal attention mechanism is used to deeply explore the potential connections between the spatial features of satellite imagery data, the variation characteristics of meteorological data, and the semantic features of public opinion text data. By calculating association weights, an intermodal association matrix is ​​generated, enabling a quantitative assessment of the importance of features from different modalities and fully capturing the complex relationships between data. Finally, based on the intermodal association matrix, various features are weighted and fused. Through dimensionality reduction, the high-dimensional and complex multimodal data is transformed into a unified representation of a multimodal emergency situation embedding vector. This reduces data dimensionality while retaining key information, improving data processing efficiency and the accuracy of subsequent analysis. Therefore, this embodiment can effectively solve the problems of data heterogeneity, difficulty in mining feature correlations, and difficulty in processing high data dimensionality in the process of multimodal data fusion. It can achieve efficient integration and deep fusion of multimodal data in disaster-stricken areas, generate more representative and usable emergency situation data representation, provide high-quality data support for subsequent deep learning-based spatiotemporal feature extraction, causal relationship analysis, and emergency decision-making strategy generation, significantly improve the comprehensiveness and accuracy of emergency situation analysis, and thus optimize the scientificity and effectiveness of emergency decision-making.

[0069] Specifically, the working process of this embodiment is as follows:

[0070] (1) Feature extraction from multimodal data:

[0071] For satellite imagery data, convolutional neural network (CNN) technology, such as ResNet, is used. The satellite imagery data is input into the network and processed through multiple convolutional and pooling layers to extract spatial features containing information such as the geographical outline of the disaster area, building damage, and road interruptions. For example, by sliding the convolutional kernel across the image, the feature response of local pixels is calculated, and spatial features at different scales are gradually extracted.

[0072] For meteorological data, the LSTM (Long Short-Term Memory) network in the time series analysis algorithm is used to input meteorological data such as wind speed, rainfall, and temperature over a period of time into the network in chronological order. The gating mechanism is used to memorize and update the information, and the change characteristics of meteorological data in the time dimension are extracted, such as the changing trend of meteorological elements and abrupt change points.

[0073] For public opinion text data, the BERT (Encoder of Bidirectional Transformer) model in the field of natural language processing is adopted. The text data is segmented and vectorized before being input into the model. The semantic relationship between words in the text is learned through the self-attention mechanism, and semantic features are extracted, such as the needs of disaster-stricken people and the description of the disaster site.

[0074] (2) Cross-modal attention mechanism to calculate association weights:

[0075] Let the spatial features extracted from satellite imagery data be S = [s1, s2, ..., s...]. m ], where m is the feature dimension; the variation characteristics of meteorological data are W = [w1, w2, ..., w n ], where n is the feature dimension; the semantic features of the public opinion text data are T = [t1, t2, ..., t] p ], where p is the feature dimension.

[0076] First, the features of the three modalities are mapped to the same dimension through linear transformation, resulting in S′=WS+b1, where W is a learnable weight matrix and b1 is a bias vector; W′=VW+b2, where V is a learnable weight matrix and b2 is a bias vector; and T′=UT+b3, where U is a learnable weight matrix and b3 is a bias vector.

[0077] Then, the attention scores between modes are calculated. For satellite imagery and meteorological data, the attention score A is... SW The calculation formula is: Where d k This refers to the mapped feature dimension. The softmax function is used to normalize the scores into a probability distribution, representing the degree of correlation between satellite imagery features and meteorological features. Similarly, the attention scores for satellite imagery and public opinion text data... Attention scores of meteorological data and public opinion text data

[0078] Finally, the three attention scores are combined into an intermodal correlation matrix M: Where A WS A TS A TW They are A SW A ST A WT The transpose of , where the 1 on the diagonal indicates that the correlation degree of its own mode is 1.

[0079] (3) Weighted fusion and dimensionality reduction:

[0080] The spatial features S of satellite imagery, the variation features W of meteorological data, the semantic features T of public opinion text data, and the intermodal correlation matrix M are weighted and fused. The formula for calculating the fused feature F is F = M. 11 S+M 12 W+M 13 T+M 21 S+M 22 W+M 23 T+M 31 S+M 32 W+M 33 T, here M ij This represents the element in the i-th row and j-th column of the intermodal correlation matrix M.

[0081] Principal component analysis (PCA) is used to reduce the dimensionality of the fused features F. By calculating the covariance matrix, eigenvalues, and eigenvectors, the main feature components are selected, redundant information is removed, and a unified representation of the multimodal emergency situation embedding vector is generated. This vector contains key information from the three modalities of data and is used for subsequent emergency situation analysis.

[0082] As one example, the step of extracting features from the multimodal emergency situation embedding vector using a spatiotemporal awareness neural network model, modeling in both the time series dimension and the spatial correlation dimension, and outputting an emergency situation state sequence containing spatiotemporal evolution characteristics includes the following sub-steps:

[0083] The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate;

[0084] A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network;

[0085] The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector.

[0086] The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster.

[0087] In this embodiment, targeted feature extraction and fusion are performed on the multimodal emergency situation embedding vectors, focusing on the spatiotemporal characteristics of disaster emergency situations. First, a temporal convolutional network with a causal convolutional structure and memory gates is used to process the embedding vectors, capturing time-series features such as disaster spread rate and resource consumption rate. The memory gates preserve historical information, enhancing the ability to capture dynamic changes in the time series. Next, a spatial topology map is constructed based on the geographical grid of the disaster-stricken area. A dual-scale graph attention network is used to mine the spatial correlation features of disaster impacts between regions, taking into account both local and global node relationships. Then, the time-series features and spatial correlation feature tensors are concatenated, and a weighted fusion is used to generate a spatiotemporal joint feature vector, balancing the importance of spatiotemporal features. Finally, the spatiotemporal joint feature vector is serialized and recombined, and the emergency situation state sequence is output through a fully connected network, presenting the spatiotemporal evolution law of the disaster. In summary, this embodiment effectively separates and extracts the spatiotemporal features of disasters, overcoming the problem of insufficient handling of spatiotemporal heterogeneity in traditional methods; the dual-scale attention mechanism enhances the capture of disaster propagation paths; weighted fusion and serialization recombination allow the model to adapt to different disaster scenarios, and the output state sequence provides strong support for disaster causal analysis, situation prediction and emergency decision-making, improving the scientific nature and timeliness of emergency response.

[0088] Specifically, the working process of this embodiment is as follows:

[0089] 3. Example of Feature Extraction from Spatiotemporal Aware Neural Network Model

[0090] 3.1 Time Series Feature Extraction: The multimodal emergency situation embedding vector generated above is used as the input data for this step. This embedding vector is a feature representation obtained by fusing satellite imagery data, meteorological data, and public opinion text data. It exists in matrix form, denoted as... Where T represents the time step, i.e., the number of time points set during data acquisition, reflecting the observation length of the disaster situation in the time series; D is the embedding vector dimension, representing the number of features in the fused data, which contain multifaceted disaster situation information. This step uses an existing Temporal Convolutional Network (TCN) to extract time series features. TCN is built based on the principle of causal convolution. By setting different dilation factors, it effectively expands the network's receptive field without significantly increasing the number of parameters, thereby capturing long-range dependencies in the time series data. In practice, the multimodal emergency situation embedding vector is input into the TCN along the time dimension. The network automatically learns time series features such as disaster spread rate and resource consumption rate through multiple causal convolutional layers and residual connection structures. After training and computation, the TCN stably outputs a time series feature matrix. Where D T As a time series feature dimension, this matrix lays the foundation for subsequent processing.

[0091] 3.2 Spatial Association Feature Extraction: The disaster-stricken area is uniformly divided into N×N grids in geographic space, and a spatial topology map is constructed accordingly. Among them, the node set Each node corresponds to a geographic raster; the edge set ε represents the connection relationship between adjacent rasters, that is, if two rasters are geographically adjacent, there is an edge between their corresponding nodes. Each node v i initial feature vector D is composed of the multimodal emergency situation embedding vector features corresponding to the grid position. S This represents the initial dimension of the spatial features.

[0092] An improved Graph Attention Network (GAT) is employed to extract spatial correlation features of disaster impacts between regions. Traditional GAT struggles to capture complex long-distance impacts when processing geospatial data. To address this issue, a multi-level propagation mechanism is designed.

[0093] First, calculate the basic attention weights between nodes. For node i and its neighbor node j, the basic attention weight β is... ij The calculation is as follows:

[0094]

[0095] Where W is a dimension of The weight matrix, D′, is used to perform a linear transformation on the initial features of the nodes. S Here are the transformed feature dimensions; 'a' is a feature with dimension 1. Attention parameter vector; [Wf i ||Wf j The symbol ] represents concatenating the transformed feature vectors of nodes i and j; LeakyReLU is the activation function. It is the set of neighboring nodes of node i.

[0096] Next, a multi-level propagation mechanism is introduced, with the propagation level set to P. At the p-th propagation level, the feature update formula for node i is:

[0097]

[0098] in, It is the feature vector of node i after propagation at level p. For initial features; W p σ is the weight matrix for the p-th propagation stage; σ is the activation function, such as ReLU. Through multi-stage propagation, nodes gradually aggregate the influence information of distant nodes.

[0099] Finally, by integrating the node features obtained from the last stage of propagation, a spatial correlation feature matrix is ​​obtained. Complete spatial feature extraction.

[0100] 3.3 Spatiotemporal Joint Feature Vector Generation: In acquiring time series features Spatial correlation features Subsequently, existing tensor concatenation techniques are used for feature fusion. Due to the differences in temporal and spatial feature dimensions, the time-series feature T is first copied and expanded along the spatial dimension, and the spatially correlated feature S is copied and expanded along the temporal dimension to match their dimensions. Then, standard tensor concatenation operations (such as the `torch.cat()` function in the PyTorch framework) are used to concatenate the features along the feature dimensions, generating a spatiotemporal joint feature vector. This operation can directly integrate feature information from both temporal and spatial dimensions, providing a complete feature representation for subsequent analysis.

[0101] 3.4 Emergency Situation Status Sequence Output: After obtaining the spatiotemporal joint feature vector Next, it needs to be converted into an emergency situation state sequence. First, the three-dimensional spatiotemporal joint feature vector is flattened into a two-dimensional matrix in sequence. This is to facilitate subsequent processing by the fully connected network. Then, the flattened matrix is ​​sequentially input into two fully connected network layers. The first fully connected network layer performs preliminary transformations on the input features through matrix multiplication and activation function operations, extracting basic semantic features; the second fully connected network layer further abstracts the features, outputting a feature vector with higher semantic expressive power. The operation of the two fully connected network layers can be represented as: H1 = ReLU(W1·J flat H2 = ReLU(W2·H1+b2), where W1 and W2 are the weight matrices of the two fully connected layers, b1 and b2 are the corresponding bias vectors, and ReLU is the activation function used to introduce nonlinearity and enhance the network's expressive power. Finally, the feature vector H2 output by the second fully connected layer is reshaped to convert it into an emergency situation state sequence. This sequence fully contains information on the evolution of disasters in both time and space, and can be directly used for subsequent operations such as constructing dynamic causal graphs and predicting the development trend of emergency situations, providing core data support for emergency decision-making.

[0102] As one example, the step of constructing a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors includes the following sub-steps:

[0103] Map each dimension of the emergency situation state sequence to a node of a causal graph, and initialize the causal strength parameters between the nodes;

[0104] Based on the causal strength parameters between nodes, the causal relationships between nodes in the causal graph are analyzed, and causal edges are determined.

[0105] A dynamic causal graph is generated based on each node and the causal edges between nodes;

[0106] Based on dynamic causal graphs, potential causal relationships between disaster events and influencing factors are identified.

[0107] In this embodiment, firstly, multi-dimensional observation variables of the emergency situation (such as disaster intensity, environmental parameters, emergency resources, etc.) are mapped as nodes in a causal graph. The causal strength parameters between nodes are initialized based on data-driven methods (such as temporal correlation, information entropy, or neural networks). Then, significant causal edges are screened and spurious associations are eliminated using causal inference algorithms (such as Granger causality tests, Bayesian structural learning, or reinforcement learning strategies). Next, a dynamically updatable causal graph structure is generated by combining a temporal sliding window or state transition model to capture the causal interactions of various factors in the disaster evolution in real time. Finally, through reverse reasoning, path search, or intervention analysis of the causal graph, the key driving factors and chain reaction paths of the disaster event are located. As can be seen from the above analysis, this embodiment achieves dynamic and interpretable modeling of disaster causal relationships. It can extract temporally dependent causal networks from complex multi-dimensional data and adapt to the abrupt changes in disaster scenarios through dynamic adjustment mechanisms. This provides a scientific basis for emergency decision-making that ensures "traceable causal relationships, quantifiable key paths, and predictable intervention effects," thereby improving the accuracy of disaster response.

[0108] Specifically, the working process of this embodiment is as follows:

[0109] 1. Map each dimension of the emergency situation state sequence to a node in a causal graph, and initialize the causal strength parameters between nodes: First, obtain the emergency situation state sequence output above. This sequence is a feature representation containing the spatiotemporal evolution law of the disaster, denoted as... Where T represents the time step, reflecting the changes of the disaster over time; N is the sequence dimension, corresponding to various disaster-related features, such as disaster intensity, resource status, and geographical environmental parameters.

[0110] Each dimension of sequence Z is mapped to a node in a causal graph, thus constructing a node set. For example, if a dimension represents rainfall, the corresponding node represents the rainfall factor; if a dimension represents the number of rescue teams, the corresponding node represents the rescue resource factor.

[0111] An improved initialization method based on mutual information and spatiotemporal distance is used when initializing the causal strength parameters between nodes. For any two nodes v i and v j Initialize the causal strength parameter ω ij The calculation formula is as follows:

[0112]

[0113] Wherein, I(v) i ;v j ) represents node v i and v j Mutual information, used to measure the dependence between features of two nodes, is derived by calculating the relationship between the joint entropy of their probability distributions and their individual entropies. A higher mutual information value indicates a stronger correlation between the two nodes. Specifically, the calculation first involves statistically analyzing node v... i and v j The value distribution in the emergency situation state sequence is then calculated using the mutual information formula I(X;Y)=H(X)+H(Y)-H(X,Y) (where H(X) is the entropy of variable X, and H(X,Y) is the joint entropy of variables X and Y). k，l I(v l ;v l ω is the maximum value among all node pairs' mutual information, used to normalize the mutual information so that ω ij The value of d ranges from [0, 1]. ij Represents node v i and v j The comprehensive distance metric corresponding to the features in the spatiotemporal dimensions. If the node corresponds to a geospatial feature, then d ij Euclidean distance is calculated based on geographic coordinates; if a time dimension is involved, a weighted calculation is performed considering differences in time steps. For example, when v i and v j When d represents the disaster intensity at adjacent time steps in different geographical locations, ij The distance and time interval between the two locations will be taken into account. max This is the maximum spatiotemporal distance between all node pairs, used to normalize the spatiotemporal distance so that the exponent term takes a value between [0,1]. Through this calculation process, both the intrinsic correlation between nodes and the spatiotemporal distance factor are considered, making the initialized causal strength parameter more consistent with the causal relationship characteristics in actual disaster scenarios.

[0114] 2. Based on the causal strength parameters between nodes, analyze the causal relationships between nodes in the causal graph and determine causal edges: After initializing the causal strength parameters, analyze the causal relationships between nodes based on these parameters to determine causal edges. This step uses an improved dynamic temporal causality test algorithm. For each pair of nodes v... i and v j Calculate its dynamic time causality score S ij The formula is as follows: Where T is the time step of the emergency situation sequence; τ is the set time lag step, used to account for the possible time delay in causal relationships. For example, in flood disasters, changes in rainfall may not immediately cause water levels to rise, requiring a certain amount of time; τ is used to capture this delay. Δy j,t+τ Represents node v j The change in eigenvalue at time t+τ; Δx i，t Represents node v i The change in eigenvalues ​​at time t. ω ij The node v initialized in the above process i and v j The causal strength parameter between nodes v. In general, by calculating node v... i Changes in v j The degree of influence after a lag time τ is calculated and weighted in conjunction with the initial causal strength parameter to obtain a dynamic temporal causality score. A threshold ε is set, when S... ij When ε > ε, determine the node v. i to node v j There exists a causal edge with weight S. ij In this way, not only is the static correlation strength between nodes considered, but also the causal influence over time is taken into account, thus more accurately determining the causal relationships between nodes.

[0115] 3. Generate a dynamic causal graph based on the nodes and the causal edges between them: This is based on the node set determined in the first two steps. Generate a dynamic causal graph from the set of causal edges ε. To reflect the dynamic change of causal relationships over time, a time decay factor γ∈(0,1) is introduced. For a causal edge (v) determined at time t... i ,v j The formula for updating the weight of ) at subsequent time t′ (t′>t) is: w ij (t′)=w ij (t)×γ t′-t , where w ij (t) is the causal edge (v) at time t. i v j The weights of causal relationships are assigned as follows: As time progresses, if no new evidence emerges to suggest the causal relationship persists or strengthens, its weight gradually decreases, reflecting the dynamic nature of causal relationships in disaster scenarios. Nodes and edges from different times are integrated and stored and represented using existing graph data structures (such as adjacency matrices or adjacency lists), ultimately generating a dynamic causal graph. This dynamic cause-effect diagram can reflect the changes in the causal relationship between disaster events and influencing factors over time in real time.

[0116] 4. Identifying Potential Causal Relationships Between Disaster Events and Influencing Factors Based on Dynamic Causal Graphs: Based on the generated dynamic causal graphs, an improved causal path search and analysis algorithm is used to identify potential causal relationships. First, the target disaster event nodes (such as nodes corresponding to "severe casualties" or "huge property losses") and the set of possible influencing factor nodes are determined. Then, starting from the influencing factor nodes, depth-first search (DFS) or breadth-first search (BFS) algorithms are used, combined with the weights of causal edges, to search for all possible paths to the target disaster event nodes. For each searched causal path... The causal path strength CPS(P) is calculated using the following formula: in are adjacent nodes in path P and The weights of causal edges are determined. A higher CPS(P) value indicates a stronger causal relationship along the path, making it more likely to be a key causal path leading to the target disaster event. By ranking and analyzing the strength of all causal paths, paths with higher strength are selected, thereby identifying potential causal relationships between disaster events and influencing factors. Simultaneously, the identification results are verified and corrected using practical disaster domain knowledge to ensure the rationality and reliability of potential causal relationships, providing a scientific basis for subsequent emergency decision-making.

[0117] As one example, the process of predicting the future emergency situation development trend of the disaster-stricken area based on potential causal relationships and generating the causal evolution result of the emergency situation in the disaster-stricken area includes the following sub-steps:

[0118] Extract the node causal association matrix based on the potential causal relationship as prior causal knowledge;

[0119] The emergency situation state sequence and the causal correlation matrix are input into a pre-trained time series prediction model to perform multi-step long trend prediction and obtain multi-step trend prediction results.

[0120] Based on the multi-step long-term trend prediction results, the situation evolution path of the disaster-stricken area under different emergency measures is simulated to generate the causal evolution results of the emergency situation of the disaster-stricken area.

[0121] In this embodiment, a complete chain from causal knowledge extraction to situation prediction and simulation is built based on potential causal relationships. First, a node causal correlation matrix is ​​extracted from the identified potential causal relationships, transforming abstract causal logic into a computable matrix form, serving as prior knowledge for subsequent predictions and injecting causal constraints into the prediction process. Next, the emergency situation state sequence and the causal correlation matrix are input into a pre-trained time series prediction model. Leveraging the model's ability to process time series data and combining it with the causal correlation matrix, a multi-step long-term trend prediction of the emergency situation in the disaster-stricken area is achieved, ensuring that the prediction results include both the changing patterns over time and conform to causal logic. Finally, based on the multi-step long-term trend prediction results, the situation evolution path of the disaster-stricken area under different emergency intervention measures is simulated. By changing the relevant parameters in the causal correlation matrix, the impact of different emergency measures on the situation is simulated, thereby generating the causal evolution results of the emergency situation. In summary, on the one hand, this embodiment overcomes the problem of lack of causal interpretability in traditional time series forecasting models by introducing a causal correlation matrix as prior knowledge, thereby improving the reliability and interpretability of the forecast results. On the other hand, this embodiment can effectively capture the long-term changing trend of the emergency situation through multi-step long-term trend forecasting, allowing sufficient time for emergency decision-making. In addition, simulating the evolution path of the situation under different emergency measures can intuitively present the impact of various measures on the development of the emergency situation, helping decision-makers to quickly assess the advantages and disadvantages of different options.

[0122] Specifically, the working process of this embodiment is as follows:

[0123] 1. After completing the construction of the dynamic cause-effect graph and the identification of potential causal relationships, extract the causal association information between nodes and generate a node causal association matrix. Where N is the total number of nodes in the causal graph, corresponding to the number of dimensions in the emergency situation state sequence. The elements C in the matrix... ij Represents node v i For node v j The strength of the causal relationship is determined by the weights of the edges in the dynamic causal graph. Specifically, during the extraction process, if node v... i to node v j There exists a causal edge with weight w. ij Then C ij =w ij If no causal edge exists, then C ij =0. To enhance the reliability of causal relationships, the matrix is ​​normalized: In this way, the sum of the elements in each row of the matrix is ​​1, which is convenient for use as weights in the subsequent prediction model, thereby transforming the complex causal relationships in the dynamic causal graph into a structured prior causal knowledge matrix.

[0124] 2. The emergency situation status sequence output from the above process. The node causal correlation matrix C obtained from the above process is used as input and fed into the improved time series prediction model. This model is based on the Transformer architecture and introduces a causal attention mechanism to fully utilize prior causal knowledge to guide the prediction process.

[0125] Model input processing: The emergency situation state sequence Z is input sequentially at time steps T, and the feature vector at each time step has a dimension of N. The node causal correlation matrix C is introduced as global context information during the model's attention calculation.

[0126] Causal attention mechanism calculation: Improve the attention calculation method in the multi-head attention module of Transformer. For query vectors key vector Sum value vector (where d) k and d v (These are the dimensions of keys and values, respectively). A causal correlation matrix C is introduced to adjust the attention weights, calculated as follows: Here, α is a hyperparameter used to adjust the influence of the causal association matrix on the attention weights. Through this calculation process, the model, when calculating attention, not only considers the feature associations of the input sequence, but also enhances the attention weights between nodes with causal relationships based on prior causal knowledge, thereby more accurately capturing the causal dependencies between emergency situation features.

[0127] Multi-step prediction: During the training phase, the model uses historical emergency situation state sequences as input to predict the situation characteristics for multiple future time steps. A teacher forcing strategy is employed to gradually guide the model to learn long-sequence prediction capabilities during training. In the prediction phase, based on the current input emergency situation state sequence, the model's forward propagation calculates and outputs multi-step trend prediction results for the next L time steps. Where L is the preset prediction step size.

[0128] 3. Multi-step long-term trend prediction results obtained based on the above process Construct an emergency response simulation system to generate the evolution path of the disaster-stricken area under different emergency response measures.

[0129] Emergency response modeling involves abstracting various emergency measures into intervention operations on the causal correlation matrix C and the emergency situation state sequence. For example, increasing the input of relief supplies can be seen as enhancing the negative causal impact of the "quantity of relief supplies" node on the "degree of disaster loss" node, which is simulated by reducing the value of the corresponding element in C; opening new relief channels can be seen as changing the strength of the causal correlation between geographical and transportation-related nodes.

[0130] Furthermore, situational evolution simulation: employing a simulation algorithm based on causal reasoning, from the prediction results... Starting from this point, and combining the modified causal relationship matrix C′, the situation evolution is simulated through iterative calculation. For each time step t, the update formula for the node state is calculated based on the causal relationship as follows: in, Represents node v j The predicted state value at time t, where f(·) is the activation function (such as ReLU) used to introduce nonlinearity. This formula indicates that the state change of each node is influenced by the state of all its causally related nodes at the previous time step. In this way, the situation evolution path under different emergency measures over a period of time is gradually generated.

[0131] The simulated situation evolution path is output as a visual chart (such as dynamic curves and heat maps) and structured data (such as node state matrices at different time steps), forming the causal evolution results of the emergency situation in the disaster-stricken area. This result intuitively demonstrates the changing trends and causal transmission processes of various emergency situation indicators in the disaster-stricken area under different emergency intervention measures, providing a comprehensive reference for emergency decision-making.

[0132] As one example, the process of generating an emergency decision-making strategy for the disaster-stricken area based on the causal evolution results of the emergency situation, combined with historical response experience data and current resource constraints, includes the following sub-steps:

[0133] Acquire historical emergency response experience data, including historical emergency plans, resource consumption data, and response effects;

[0134] Transform the current available rescue resources, team locations, and transportation network status into resource constraints.

[0135] Using a pre-trained deep learning model, the initial emergency decision-making strategy is generated by taking the causal evolution results of the emergency situation, the historical handling experience data, and the resource constraints as inputs.

[0136] Based on the causal logic of the dynamic causal graph, the potential secondary disaster risks caused by the initial decision-making strategy are verified, and the final emergency decision-making strategy for the disaster-stricken area is output.

[0137] In this embodiment, emergency response plans, resource consumption, and response effectiveness data under multiple scenarios are first extracted from historical databases to construct an experience knowledge base based on case-based reasoning. Simultaneously, real-time rescue resources (such as material inventory and team locations) and traffic network status are encoded into constrained mathematical programming conditions (such as a system of linear inequalities) to ensure decision feasibility. Then, causal evolution prediction results (such as disaster spread paths and resource demand timelines) are input into a pre-trained deep reinforcement learning model along with the aforementioned data. An initial decision (such as material allocation plans and evacuation routes) that satisfies resource constraints is generated through a policy network. Finally, based on the causal logic of a dynamic causal graph, counterfactual reasoning and risk propagation algorithms are used to simulate the intervention effect of the initial strategy on key nodes (such as transportation hubs and secondary disaster sources), quantify the probability of secondary risks (such as road congestion causing rescue delays), and optimize and adjust the strategy weights through game theory to output a final decision that balances timeliness, safety, and resource efficiency. Therefore, this embodiment realizes a closed loop of "causal interpretability-driven decision generation and dynamic risk chain proactive defense". It improves the scientific nature of decision-making through historical experience and causal inference, and reduces the risk of secondary disasters through causal verification mechanism, thereby enhancing the robustness and feasibility of decision-making in complex disaster scenarios.

[0138] Specifically, the working process of this embodiment is as follows:

[0139] 1. Obtain historical emergency response experience data, including historical emergency plans, resource consumption data, and response effects: Establish a historical emergency response experience data collection and storage system to obtain relevant data from multiple channels. Collect complete information on historical disaster events by establishing data interfaces with emergency management department archives, rescue team work record systems, and disaster statistics databases.

[0140] Historical emergency response plans include detailed rescue operation plans, such as textual and graphical documentation on rescue team deployment, material distribution point setup, and emergency evacuation route planning. Resource consumption data records in detail the quantity and destination of various resources consumed during each disaster response, such as medical supplies, food, and machinery. Response effectiveness data is reflected in post-disaster assessment reports from affected areas, including quantitative indicators such as the reduction in casualties, the degree of control over economic losses, and the restoration of social order. After structuring this data, it is stored in a historical response experience database, forming a standardized historical response experience dataset H. Each row corresponds to a historical disaster response event, and each column represents different data attributes, such as emergency response plan code, resource category, and response effectiveness score.

[0141] 2. Transform the current availability of rescue resources, team distribution locations, and traffic network status into resource constraints: Utilize IoT devices and Geographic Information Systems (GIS) to collect data in real time. Sensors deployed in supply warehouses acquire information on the current availability of rescue resources, such as the quantity of medicines and tents in stock. Using positioning terminals carried by rescue teams, combined with GIS technology, the precise distribution locations of teams are determined, forming a dataset containing latitude and longitude coordinates, team personnel, and equipment configuration information. Data from traffic cameras, road sensors, and navigation platforms is used to analyze the current traffic network status, obtaining information such as road capacity, congested road sections, and bridge and tunnel conditions.

[0142] The above data is integrated and transformed into a resource constraint matrix R. The rows of the matrix correspond to different types of resource constraints, such as "maximum quantity of medical supplies," "maximum number of rescue teams that can be deployed," and "traffic restrictions on key road sections"; the columns correspond to different disaster-stricken areas or rescue task nodes. For example, R... ij This represents the upper limit or limitation of the availability of resource type i in the j-th disaster-stricken area. If R... 1j If the value represents the upper limit of the quantity of medical supplies, then its value is the total amount of medical supplies that can be allocated to the j-th disaster-stricken area.

[0143] 3. Using a pre-trained deep learning model, and taking the causal evolution results of the emergency situation, the historical handling experience data, and the resource constraints as input, an initial emergency decision-making strategy is generated: an improved deep learning model based on the Transformer architecture is constructed and trained. This model introduces a gating mechanism on top of the traditional Transformer to better integrate different types of input data.

[0144] The input data processing involves organizing the causal evolution result E of the emergency situation into a tensor containing multiple future time steps and multiple situation indicators. Each element E... t，k This represents the predicted value of the k-th situation indicator (such as the affected area or the number of people trapped) at the t-th time step in the future. Historical response experience data H, after feature extraction and encoding, is transformed into a fixed-length vector representation, retaining key information. The resource constraint matrix R, after normalization, is merged with other data.

[0145] Gating mechanism calculation: A gating unit is added between the input layer and the attention layer of the model. For the input emergency situation causal evolution result vector e, historical response experience data vector h, and resource constraint vector r, the gating weights g1, g2, and g3 are calculated: g1 = σ(W g1 ·e+b g1 g2=σ(W) g2 ·h+b g2 g3=σ(W) g3·r+b g3 ), where σ is the sigmoid activation function, which maps the output to the (0,1) interval; W g1 W g2 W g3 Let b be the weight matrix. g1 b g2 b g3 Here are the bias vectors. These parameters are learned through model training. The final fused input vector x is: x = g1⊙e + g2⊙h + g3⊙r, where ⊙ denotes element-wise multiplication.

[0146] Model Training and Prediction: Historical disaster data and corresponding effective emergency decision-making strategies are used as training samples. The model is trained using the cross-entropy loss function and the Adam optimizer. After training, the current causal evolution of the emergency situation, historical response experience data, and resource constraints are input into the model. Through feature extraction and decision inference using a multi-layer Transformer module, the initial emergency decision-making strategy D is output. init This strategy encompasses specific aspects such as the allocation of rescue resources, the scheduling of rescue operations, and personnel deployment plans. For details on the training process, please refer to existing model training techniques.

[0147] 4. Based on the causal logic of the dynamic cause-effect graph, verify the potential secondary disaster risks caused by the initial decision-making strategy, and output the final emergency decision-making strategy for the disaster-stricken area: The initial emergency decision-making strategy D... init This translates into intervention operations on nodes in a dynamic causal graph. For example, increasing the deployment of relief supplies to a certain area corresponds to adjusting the state value of the "supply quantity" node, and changing the route of the rescue team corresponds to modifying the causal relationship strength between geographical and transportation-related nodes.

[0148] Based on the causal logic of dynamic causal graphs, a causal inference algorithm is used to calculate the potential secondary disaster risk caused by each intervention operation. For each node v in the dynamic causal graph... i After implementing the decision-making strategy, its state changes Δv i The calculation method is as follows, influenced by the nodes connected by relevant causal edges: Where ε is the set of edges in the dynamic causal graph, (v j v i ) indicates from node v j to node v i causal edge, w ji Let Δv be the weight of the causal edge. j For node v j The state changes of all nodes are calculated iteratively to assess the likelihood and impact of secondary disasters.

[0149] A risk threshold τ is set. If the calculated secondary disaster risk value is higher than the threshold, the initial emergency decision-making strategy is adjusted, such as re-planning resource allocation or modifying the order of rescue operations. If the risk value is lower than the threshold, the initial emergency decision-making strategy is determined as the final emergency decision-making strategy D for the disaster-stricken area. final It outputs complete decision-making documents, including detailed rescue plans and risk assessment reports, to guide actual emergency rescue work.

[0150] See Figure 2 This is a schematic diagram of an emergency situation intelligent decision-making device based on deep learning, provided in an embodiment of the present invention. The deep learning-based emergency situation intelligent decision-making device includes:

[0151] Data fusion module 10 is used to fuse multimodal data from disaster-stricken areas and generate a unified representation of multimodal emergency situation embedding vectors;

[0152] Feature extraction module 11 is used to extract features from the multimodal emergency situation embedding vector through a spatiotemporal perception neural network model, model in the time series dimension and spatial correlation dimension respectively, and output an emergency situation state sequence containing spatiotemporal evolution characteristics.

[0153] Causal identification module 12 is used to construct a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors;

[0154] Prediction module 13 is used to predict the future development trend of the emergency situation in the disaster-stricken area based on potential causal relationships, and generate the causal evolution result of the emergency situation in the disaster-stricken area.

[0155] The strategy generation module 14 is used to generate emergency decision-making strategies for the disaster-stricken area based on the causal evolution results of the emergency situation and in combination with historical handling experience data and current resource constraints.

[0156] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0157] Multimodal data from disaster-stricken areas is collected using multiple sources of equipment. Fusion technology integrates image, text, and voice data into a unified multimodal emergency situation embedding vector, eliminating data noise and redundancy and enabling a more comprehensive and accurate reflection of the emergency situation. A spatiotemporal awareness neural network model is used to extract and model features from the embedding vector in both the time series and spatial correlation dimensions, deeply exploring the spatiotemporal evolution of the disaster situation to output an emergency situation state sequence containing spatiotemporal characteristics. This state sequence is then used to construct a dynamic causal graph to identify potential causal relationships between disaster events and influencing factors, providing a basis for prediction. Based on these causal relationships, future emergency situation development trends are predicted, generating causal evolution results. Finally, combined with historical response experience and current resource constraints, emergency decision-making strategies are generated. In this process, multimodal data fusion provides a comprehensive data foundation for subsequent analysis, spatiotemporal feature extraction accurately captures the spatiotemporal changes of the disaster, dynamic causal graph construction enables precise identification of potential causal relationships, and the decision-making strategies generated based on multi-factor synthesis are more aligned with actual needs. In summary, the embodiments of the present invention achieve deep fusion and analysis of emergency situation data through deep fusion, dynamic modeling and causal reasoning mechanisms, accurately grasp the spatiotemporal characteristics and causal relationships of disasters, thereby enabling accurate perception of disaster situations, evolution prediction and intelligent decision generation under resource constraints, effectively improving the effectiveness of emergency management.

[0158] As one example, the data fusion module is specifically used for:

[0159] Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data.

[0160] A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix.

[0161] Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing.

[0162] As one example, the feature extraction module is specifically used for:

[0163] The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate;

[0164] A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network;

[0165] The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector.

[0166] The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster.

[0167] It is understood that the content of the above-described embodiment of the deep learning-based emergency situation intelligent decision-making device can be referred to the content of the above-described embodiment of the deep learning-based emergency situation intelligent decision-making method, and will not be repeated here.

[0168] See Figure 3 This is a schematic diagram of an emergency situation intelligent decision-making system based on deep learning according to an embodiment of the present invention. The deep learning-based emergency situation intelligent decision-making system of this embodiment includes: a processor 100, a memory 101, and a computer program stored in the memory 101 and executable on the processor 100, such as a deep learning-based emergency situation intelligent decision-making program. When the processor 100 executes the computer program, it implements the steps in the above-described embodiments of the deep learning-based emergency situation intelligent decision-making methods. Alternatively, when the processor 100 executes the computer program, it implements the functions of each module / unit in the above-described device embodiments.

[0169] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the deep learning-based emergency situation intelligent decision-making system.

[0170] The deep learning-based intelligent decision-making system for emergency situations can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The system may include, but is not limited to, processors and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a deep learning-based intelligent decision-making system for emergency situations and does not constitute a limitation on such a system. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the deep learning-based intelligent decision-making system for emergency situations may also include input / output devices, network access devices, buses, etc.

[0171] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the deep learning-based emergency situation intelligent decision-making system, connecting all parts of the system via various interfaces and lines.

[0172] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the deep learning-based emergency situation intelligent decision-making system by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0173] The modules / units integrated into the deep learning-based emergency situation intelligent decision-making system, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted from the content as required by the legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium may not include electrical carrier signals and telecommunication signals.

[0174] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0175] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A deep learning-based emergency situation intelligent decision-making method, characterized in that, Includes the following steps: Multimodal data from disaster-stricken areas are fused and processed to generate a unified representation of the multimodal emergency situation embedding vector; Feature extraction is performed on the multimodal emergency situation embedding vector using a spatiotemporal awareness neural network model. Modeling is performed in both the time series dimension and the spatial correlation dimension, and the output is an emergency situation state sequence containing spatiotemporal evolution characteristics. A dynamic causal graph is constructed using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors; Based on potential causal relationships, the future development trend of the emergency situation in the disaster-stricken area is predicted, and the causal evolution result of the emergency situation in the disaster-stricken area is generated. Based on the causal evolution results of the emergency situation, and combined with historical response experience data and current resource constraints, an emergency decision-making strategy for the disaster-stricken area is generated. The step of constructing a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors includes: Obtain the emergency situation status sequence, which is a characteristic representation containing the spatiotemporal evolution law of the disaster, denoted as... ,in It represents the time step, reflecting the changes of the disaster over time; The sequence dimension corresponds to various disaster-related features; the sequence Each dimension is mapped to a node in the causal graph, thus constructing a node set; and the causal strength parameters between nodes are initialized using the following initialization method based on mutual information and spatiotemporal distance: for any two nodes and Initialize causal strength parameters The calculation formula is as follows: , Represents a node and Mutual information between nodes, where the nodes are counted first. and The value distribution in the emergency situation state sequence is then calculated using the mutual information formula: ,in For variables entropy, For variables and joint entropy, It is the maximum value among all node pairs' mutual information, making The value of is in the range [0,1]. Represents a node and A comprehensive distance metric for corresponding features across the spatiotemporal dimensions; It is the maximum spatiotemporal distance between all node pairs; An improved dynamic temporal causality test algorithm is adopted for each pair of nodes. and Calculate its dynamic time causality score The formula is as follows: ,in, The time step of the emergency situation status sequence; It is the set time lag step. Represents a node exist The change in eigenvalues ​​at time t; Represents a node exist The change in eigenvalues ​​at time t, The nodes initialized in the above process and Causal strength parameters between nodes; calculated by nodes Changes Lag The degree of impact over time is calculated and weighted in conjunction with the initial causal strength parameters to obtain a dynamic time-based causal score; a threshold is set. ,when At that time, determine the node To the node There exists a causal edge with a weight of . ; A dynamic causal graph is generated based on the nodes and the causal edges between them. The time decay factor is introduced. For the moment Determined causal edge Its weight in subsequent moments The update formula is: ,in, , It is a moment Causal edge The weights are determined; a graph data structure is used for storage and representation, ultimately generating a dynamic causal graph. ; Based on the generated dynamic causal graph, an improved causal path search and analysis algorithm is used to identify potential causal relationships: First, the target disaster event node and the set of possible influencing factor nodes are determined. Starting from the influencing factor nodes, and combining the weights of causal edges, all possible paths to the target disaster event node are searched; for each searched causal path... Calculate its causal path strength The formula is as follows: ,in It is a path Middle adjacent nodes and The weights of causal edges between events; by ranking and analyzing the strength of all causal paths, paths with higher strength are selected, thereby identifying the potential causal relationship between disaster events and influencing factors; The process of generating emergency decision-making strategies for the disaster-stricken area based on the causal evolution results of the emergency situation, combined with historical response experience data and current resource constraints, includes the following sub-steps: Acquire historical emergency response experience data, including historical emergency plans, resource consumption data, and response effects; Transform the current available rescue resources, team locations, and transportation network status into resource constraints. Using a pre-trained deep learning model, the initial emergency decision-making strategy is generated by taking the causal evolution results of the emergency situation, the historical handling experience data, and the resource constraints as inputs. The initial emergency decision-making strategy is transformed into intervention operations on nodes in a dynamic causal graph. Based on the causal logic of the dynamic causal graph, a causal inference algorithm is used to calculate the potential secondary disaster risk caused by each intervention operation. For each node in the dynamic causal graph... After the initial emergency decision-making strategy is implemented, its state changes. The calculation method is as follows, influenced by the nodes connected by relevant causal edges: ,in, For a dynamic causal graph, the set of edges. Indicates from node To the node The causal edge, The weight of this causal edge. For nodes The state changes are calculated iteratively to obtain the state changes of all nodes, thereby assessing the likelihood and impact of secondary disasters; risk thresholds are set. If the calculated secondary disaster risk value is higher than the threshold, the initial emergency decision-making strategy will be adjusted; if the secondary disaster risk value is lower than the threshold, the initial emergency decision-making strategy will be determined as the final emergency decision-making strategy for the disaster-stricken area. The final emergency decision-making strategy for the disaster-stricken area will be output. 2.The deep learning based emergency situation intelligent decision method of claim 1, wherein, The process of fusing multimodal data from disaster-stricken areas to generate a unified representation of the multimodal emergency situation embedding vector includes the following sub-steps: Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data. A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix. Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing. 3.The deep learning based emergency situation intelligent decision method of claim 2, wherein, The step of extracting features from the multimodal emergency situation embedding vector using a spatiotemporal awareness neural network model, modeling in both the time series dimension and the spatial correlation dimension, and outputting an emergency situation state sequence containing spatiotemporal evolution characteristics includes the following sub-steps: The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate; A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network; The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector. The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster. 4.The deep learning based emergency situation intelligent decision method of claim 3, wherein, The process of predicting the future emergency situation development trend of the disaster-stricken area based on potential causal relationships and generating the causal evolution result of the emergency situation in the disaster-stricken area includes the following sub-steps: Extract the node causal association matrix based on the potential causal relationship as prior causal knowledge; The emergency situation state sequence and the causal correlation matrix are input into a pre-trained time series prediction model to perform multi-step long trend prediction and obtain multi-step trend prediction results. Based on the multi-step long-term trend prediction results, the situation evolution path of the disaster-stricken area under different emergency measures is simulated to generate the causal evolution results of the emergency situation of the disaster-stricken area.

5. An emergency situation intelligent decision-making device based on deep learning, characterized in that, include: The data fusion module is used to fuse multimodal data from disaster-stricken areas and generate a unified representation of the multimodal emergency situation embedding vector. The feature extraction module is used to extract features from the multimodal emergency situation embedding vector through a spatiotemporal perception neural network model, modeling in the time series dimension and spatial correlation dimension respectively, and outputting an emergency situation state sequence containing spatiotemporal evolution characteristics. The causal identification module is used to construct a dynamic causal graph using the emergency situation state sequence to identify potential causal relationships between disaster events and influencing factors; The prediction module is used to predict the future development trend of the emergency situation in the disaster-stricken area based on potential causal relationships, and generate the causal evolution results of the emergency situation in the disaster-stricken area. The strategy generation module is used to generate emergency decision-making strategies for the disaster-stricken area based on the causal evolution results of the emergency situation, combined with historical handling experience data and current resource constraints. Specifically, the causal identification module is used for: Obtain the emergency situation status sequence, which is a characteristic representation containing the spatiotemporal evolution law of the disaster, denoted as... ,in It represents the time step, reflecting the changes of the disaster over time; The sequence dimension corresponds to various disaster-related features; the sequence Each dimension is mapped to a node in the causal graph, thus constructing a node set; and the causal strength parameters between nodes are initialized using the following initialization method based on mutual information and spatiotemporal distance: for any two nodes and Initialize causal strength parameters The calculation formula is as follows: , Represents a node and Mutual information between nodes, where the nodes are counted first. and The value distribution in the emergency situation state sequence is then calculated using the mutual information formula: ,in For variables entropy, For variables and joint entropy, It is the maximum value among all node pairs' mutual information, making The value of is in the range [0,1]. Represents a node and A comprehensive distance metric for corresponding features across the spatiotemporal dimensions; It is the maximum spatiotemporal distance between all node pairs; An improved dynamic temporal causality test algorithm is adopted for each pair of nodes. and Calculate its dynamic temporal causality score The formula is as follows: ,in, The time step of the emergency situation status sequence; It is the set time lag step. Represents a node exist The change in eigenvalues ​​at time t; Represents a node exist The change in eigenvalues ​​at time t, The nodes initialized in the above process and Causal strength parameters between nodes; calculated by nodes Changes Lag The degree of impact over time is calculated and weighted in conjunction with the initial causal strength parameters to obtain a dynamic time-based causal score; a threshold is set. ,when At that time, determine the node To the node There exists a causal edge with a weight of . ; A dynamic causal graph is generated based on the nodes and the causal edges between them. The time decay factor is introduced. For the moment Determined causal edge Its weight in subsequent moments The update formula is: ,in, , It is a moment Causal edge The weights are determined; a graph data structure is used for storage and representation, ultimately generating a dynamic causal graph. ; Based on the generated dynamic causal graph, an improved causal path search and analysis algorithm is used to identify potential causal relationships: First, the target disaster event node and the set of possible influencing factor nodes are determined. Starting from the influencing factor nodes, and combining the weights of causal edges, all possible paths to the target disaster event node are searched; for each searched causal path... Calculate its causal path strength The formula is as follows: ,in It is a path Middle adjacent nodes and The weights of causal edges between events; by ranking and analyzing the strength of all causal paths, paths with higher strength are selected, thereby identifying the potential causal relationship between disaster events and influencing factors; The strategy generation module is specifically used for: Acquire historical emergency response experience data, including historical emergency plans, resource consumption data, and response effects; Transform the current available rescue resources, team locations, and transportation network status into resource constraints. Using a pre-trained deep learning model, the initial emergency decision-making strategy is generated by taking the causal evolution results of the emergency situation, the historical handling experience data, and the resource constraints as inputs. The initial emergency decision-making strategy is transformed into intervention operations on nodes in a dynamic causal graph. Based on the causal logic of the dynamic causal graph, a causal inference algorithm is used to calculate the potential secondary disaster risk caused by each intervention operation. For each node in the dynamic causal graph... After the initial emergency decision-making strategy is implemented, its state changes. The calculation method is as follows, influenced by the nodes connected by relevant causal edges: ,in, For a dynamic causal graph, the set of edges. Indicates from node To the node The causal edge, The weight of this causal edge. For nodes The state changes are calculated iteratively to obtain the state changes of all nodes, thereby assessing the likelihood and impact of secondary disasters; risk thresholds are set. If the calculated secondary disaster risk value is higher than the threshold, the initial emergency decision-making strategy will be adjusted; if the secondary disaster risk value is lower than the threshold, the initial emergency decision-making strategy will be determined as the final emergency decision-making strategy for the disaster-stricken area. The final emergency decision-making strategy for the disaster-stricken area will be output.

6. The emergency situation intelligent decision-making device based on deep learning as described in claim 5, characterized in that, The data fusion module is specifically used for: Feature extraction is performed on satellite imagery data, meteorological data, and public opinion text data from multimodal data of disaster-stricken areas to obtain the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data. A cross-modal attention mechanism is used to calculate the correlation weights between the spatial features of the satellite imagery data, the variation features of the meteorological data, and the semantic features of the public opinion text data, thereby generating an intermodal correlation matrix. Based on the intermodal correlation matrix, the spatial features, the change features, and the semantic features are weighted and fused, and a unified representation of the multimodal emergency situation embedding vector is generated through dimensionality reduction processing.

7. The emergency situation intelligent decision-making device based on deep learning as described in claim 6, characterized in that, The feature extraction module is specifically used for: The multimodal emergency situation is embedded into a vector input temporal convolutional network to extract time-series features of disaster spread rate and resource consumption rate; A spatial topology map is constructed based on the geographical grid division of the disaster-stricken area, and spatial correlation features of disaster impact between regions are extracted through graph attention network; The time series features and spatial correlation features are concatenated using tensors to generate a spatiotemporal joint feature vector. The spatiotemporal joint feature vector is serialized and recombined to output an emergency situation state sequence that characterizes the spatiotemporal evolution of the disaster.

8. An intelligent decision-making system for emergency situations based on deep learning, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the deep learning-based intelligent decision-making method for emergency situations as described in any one of claims 1 to 4.

Images