Method for evaluating vulnerability of earthquake and secondary disasters thereof in combination with multiple agents and deep learning

Abstract

The invention discloses an earthquake and secondary disaster vulnerability assessment method in combination with multiple agents and deep learning, and the method comprises the following steps: (1), collecting and preprocessing data related to an earthquake and an infrasound disaster thereof, and obtaining multi-source static data; key features are extracted from the multi-source static data; step (2), constructing a hierarchical structure index system of vulnerability assessment by utilizing expert knowledge, constructing a judgment matrix by adopting an analytic hierarchy process, and calculating the weight of each index to obtain an AHP weight vector; step (3), multi-agent system construction and disaster propagation dynamic simulation; step (4), designing a GCN + Transform deep learning model, and combining the multi-source static data with the dynamic evolution data to construct a spatial-temporal feature tensor as the input of the GCN + Transform deep learning model; and step (5), outputting a vulnerability evaluation result. According to the method, the problems of the existing earthquake and secondary disaster vulnerability assessment precision and timeliness can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural disaster vulnerability assessment. Specifically, it is a method for earthquake and secondary disaster vulnerability assessment combining multi-agent and deep learning. BACKGROUND

[0002] Earthquakes and their secondary disasters (such as landslides, debris flows, fires, etc.) have caused great threats to human society and the natural environment. Accurate assessment of the vulnerability of earthquakes and their secondary disasters is of great significance for disaster prevention, emergency response, and post-disaster recovery. Through effective vulnerability assessment, governments and relevant departments can develop scientific and reasonable disaster prevention and mitigation strategies, optimize resource allocation, reduce disaster losses, and protect people's lives and property safety.

[0003] Currently, the vulnerability assessment of earthquakes and their secondary disasters mainly relies on traditional empirical models, statistical analysis methods, and numerical simulation methods based on physical processes. Although these methods can provide preliminary assessment of disaster vulnerability to some extent, they have significant limitations in many aspects. First, traditional methods usually rely on limited historical data and empirical formulas, which are difficult to accurately reflect the complexity and nonlinearity of disaster systems. Second, these methods can only perform static assessment and cannot effectively capture the dynamic changes and complex interactions during the disaster occurrence process. In addition, traditional assessment methods face many challenges in handling multi-source heterogeneous data (such as seismic intensity data, geological structure data, meteorological data, geographic information data, etc.), and often cannot fully utilize the rich information in these data, resulting in limited assessment accuracy.

[0004] In recent years, with the rapid development of deep learning technology, its powerful data processing and feature extraction capabilities have provided new solutions for disaster vulnerability assessment. However, existing deep learning methods still face many challenges when applied to disaster vulnerability assessment. First, the occurrence process of earthquakes and secondary disasters involves complex spatiotemporal dynamic changes, and traditional deep learning models (such as convolutional neural networks (CNN), recurrent neural networks (RNN), etc.) have limitations in synchronously processing spatiotemporal dependencies. Second, disaster vulnerability assessment needs to consider multiple influencing factors and their interactions, and existing methods often have difficulty in effectively integrating data from different sources and types, resulting in limited accuracy and reliability of the assessment results.

[0005] To address the above challenges, graph convolutional neural networks (GCN) and Transformer architectures have shown great potential. GCN is good at capturing the spatial topological relationship between data and can effectively model the mutual influence of disasters in geographical space. While the Transformer architecture, with its self-attention mechanism, performs well in processing long sequence data and capturing temporal dependencies. Combining GCN with Transformer can theoretically achieve collaborative modeling of spatiotemporal dynamic characteristics of disasters, but how to apply this combination to earthquake and secondary disaster vulnerability assessment and integrate expert knowledge to enhance the model's interpretability is a difficult problem that has not been solved in current research. SUMMARY

[0006] To this end, the technical problem to be solved by the present application is to provide a method for earthquake and secondary disaster vulnerability assessment combining multi-agent and deep learning, to solve the existing earthquake and secondary disaster vulnerability assessment accuracy and timeliness.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] A method for earthquake and secondary disaster vulnerability assessment combining multi-agent and deep learning, comprising the following steps:

[0009] Step (1), collect data related to earthquakes and secondary disasters and preprocess to obtain multi-source static data; extract key features from the multi-source static data;

[0010] Step (2), construct an evaluation model and determine the weights using the analytic hierarchy process (AHP) to obtain an AHP weight vector, and perform decision analysis and comprehensive scoring and evaluation based on the fuzzy comprehensive evaluation method;

[0011] Step (3), construct a multi-agent system and give each agent autonomous decision-making ability; design rules for local information exchange and collaboration between agents to dynamically simulate the propagation path and spatiotemporal evolution process of disasters, and obtain dynamic evolution data;

[0012] Step (4), design a GCN+Transformer deep learning model, and combine the multi-source static data and dynamic evolution data to construct a spatiotemporal feature tensor as the input of the GCN+Transformer deep learning model for model training and evaluation;

[0013] Step (5), use the trained GCN+Transformer deep learning model to evaluate the vulnerability of earthquakes and secondary disasters, and output a vulnerability level index, and fuse the vulnerability level index with social and economic data and / or spatial geographic data to generate a disaster vulnerability distribution map.

[0014] The present application provides a seismic and secondary disaster vulnerability assessment method combining multi-agent system, analytic hierarchy process (AHP), graph convolutional neural network (GCN) and Transformer architecture. The method not only effectively fuses multi-source heterogeneous data, but also significantly improves the accuracy and timeliness of disaster vulnerability assessment through spatio-temporal modeling and complex interaction simulation, providing stronger technical support for earthquake and secondary disaster warning, emergency response and post-disaster recovery.

[0015] The above-mentioned seismic and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (1), the data related to the earthquake and secondary disaster include seismic data, topographic data, meteorological data, geological structure data, human engineering activity data and social economic data;

[0016] The preprocessing of multi-source data includes coordinate conversion, resampling to unified spatio-temporal resolution and noise removal; coordinate conversion is unified using WGS84 geodetic coordinate system or specific projection coordinate system; according to the spatial resolution requirement, bilinear interpolation, nearest neighbor interpolation or cubic convolution interpolation is selected for resampling; Gaussian filter or median filter is used to remove noise in meteorological data grid data;

[0017] The extracted key features include seismic intensity, distance from historical earthquake point and distance from agent in seismic data, slope, slope height and distance from water system in topographic data, precipitation in meteorological data, stratum lithology and distance from fault in geological structure data, distance from road and land use type in human engineering activity data, population density and economic density in social economic data.

[0018] The above-mentioned seismic and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (2), the hierarchical index system of vulnerability assessment includes disaster-causing factors and disaster-inducing factors; disaster-causing factors include slope, slope height, stratum, fault, seismic intensity, precipitation and earthquake point; disaster-inducing factors include multi-agent, population density, economic density, land use, road and water system;

[0019] Step (2) specifically includes the following steps:

[0020] Step (2-1), the hierarchical index system of vulnerability assessment is constructed by expert knowledge, including target layer, criterion layer and sub-criterion layer;

[0021] Step (2-1), the judgment matrix is constructed by analytic hierarchy process: in each level, the importance of each element is judged by pairwise comparison method; the relative importance between elements is evaluated by 1-9 scale method; suppose the element set in the hierarchical structure is X1, X2, …, X nThen, for each pair of elements X i and X j , a judgment matrix A = [a ij ] needs to be constructed, where a ij represents the relative importance of element X i with respect to element X j , and satisfies symmetry;

[0022] Step (2-3), calculating the AHP weight vector and normalizing: the relative weight of each element is calculated by the judgment matrix using the eigenvector method; the calculation method of the eigenvector is: A·W = λ max ·W; in the formula, W is the eigenvector, representing the weight of each element, and λ max is the largest eigenvalue; the sum of the weights after normalization is 1;

[0023] Step (2-4), consistency check: if the consistency ratio is greater than or equal to 0.1, it means that the judgment matrix is inconsistent and needs to be adjusted; the calculation method of the consistency ratio CR is: In the formula, CI is the consistency index, RI is the random consistency index, and n is the order of the judgment matrix;

[0024] Step (2-5), comprehensive evaluation and decision-making: the weights of each level are synthesized by weighting to make a comprehensive evaluation; for each scheme, the final score is obtained by weighting and summing the weights and scores under each criterion; for each scheme, the final score is obtained by weighting and summing the weights and scores under each criterion.

[0025] In the above method of earthquake and secondary disaster vulnerability assessment combining multi-agent and deep learning, in step (3), the agent system construction parameters: each agent represents a geographic unit, and the number of agents is 100-5000;

[0026] In the dynamic simulation of the propagation path and spatio-temporal evolution process of the disaster: the information exchange frequency is set to be once every hour or every 6 hours according to the disaster evolution speed; the agent decision-making model is a rule-based or reinforcement learning-based method to simulate agent behavior; the information sharing mechanism is set to be local information sharing, i.e. agents only exchange information with their neighboring agents;

[0027] Disaster propagation simulation parameters: according to the geological conditions and the type of disaster, the propagation speed of secondary disasters is set, and the rate of debris flow is set to be 2-10 meters / hour; the influence range of disaster propagation is estimated based on historical disaster data or physical models, and the radius of landslide-prone areas is 500-1000 meters.

[0028] The earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (4), the GCN+Transformer deep learning model architecture is designed as:

[0029] (4-1), graph construction: an adjacency matrix is constructed using geographical adjacency relationship to represent the spatial dependency relationship between agent units; each graph node represents an agent unit;

[0030] (4-2), GCN layer: 2-3 layers of graph convolution layers are set to aggregate the spatial features of the neighborhood nodes and capture the spatial dependency relationship between agents;

[0031] (4-3), Transformer layer; 4-6 encoder layers are set to process the time series data generated by the multi-agent system and capture the dynamic characteristics of disaster evolution.

[0032] The earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (4), the model training is guided by expert knowledge: the AHP weight vector is used as part of the loss function, i.e. an AHP consistency regularization term LAHP is introduced; the key features are weighted at the input end to guide the model to focus on important features.

[0033] The earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (4), the Adam optimizer is used, the initial learning rate is set to 1e-3, and the learning rate decay strategy is configured, i.e. every 10 epochs are decayed by 0.5; the training batch is set to 32-64; the number of training rounds is set to 100-200 rounds, or until the model converges on the validation set; the mean square error and / or root mean square error are used as the main loss function, and the AHP consistency regularization term is combined to form a composite loss function.

[0034] The earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (4), the model is trained on the historical labeled data set, and the performance is evaluated on an independent test set; the vulnerability assessment parameters include the prediction time domain and the output; according to the disaster type, the prediction time domain is set to dynamic vulnerability prediction for 1 hour to 24 hours in the future; the output is the vulnerability index of each agent unit at different time points in the future, as well as the disaster risk prediction report containing risk level, area and intensity.

[0035] The above-mentioned earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (5), the vulnerability index output by the model is spatially superimposed with social and economic data to calculate the risk loss; the vulnerability index is combined with spatial geographic data to generate a vulnerability level distribution map with a spatial resolution of 30 meters to 1 kilometer; and the disaster risk area is displayed in the form of a heat map and / or a hierarchical color map.

[0036] The above-mentioned earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning, in step (5), further includes emergency response optimization and decision support: including resource scheduling optimization and decision support system;

[0037] Resource scheduling optimization: based on the disaster impact range and vulnerability level, a genetic algorithm or linear programming optimization algorithm is used to generate an optimal allocation scheme of emergency resources;

[0038] Decision support system: the trained GCN+Transformer deep learning model and algorithm are integrated to build a visual decision support system to provide pre-disaster warning, disaster response and post-disaster recovery decision support for disaster management departments.

[0039] The technical scheme of the present application has the following beneficial technical effects:

[0040] The earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning of the present application can significantly improve the accuracy, interpretability and timeliness of disaster vulnerability assessment by effectively fusing multi-source heterogeneous data, simulating the complex interaction and dynamic change in the disaster occurrence process. This method is suitable for vulnerability assessment and risk management of earthquakes and secondary disasters (such as landslides, debris flows, fires, etc.). The innovation points are:

[0041] (1) Deep fusion of expert knowledge and data-driven: for the first time, the present application organically fuses the expert prior knowledge of analytic hierarchy process (AHP) and the GCN+Transformer deep learning model in a multi-task learning framework, guides the model to focus on key influencing factors by taking the AHP weight as an auxiliary loss function, effectively solves the "black box" problem of traditional deep models, and significantly improves the interpretability and reliability of the evaluation results.

[0042] (2) Spatio-temporal integration of multi-source heterogeneous data: the present application integrates seismic data, topographic and geomorphic data, meteorological data, geological structure data, human engineering activity data and social and economic data, and combines dynamic evolution data simulated by a multi-agent system to construct a comprehensive spatio-temporal feature set. This provides a richer information base for vulnerability assessment and can more accurately reflect the disaster influencing factors and their complex spatio-temporal interactions.

[0043] (3) The innovative application of GCN+Transformer collaborative modeling: The invention innovatively applies the GCN+Transformer hybrid architecture, fully leveraging the respective advantages of GCN in capturing spatial dependencies and Transformer in handling temporal sequence dynamics; the two work together to achieve simultaneous modeling of the complex spatiotemporal processes of earthquakes and their secondary disasters, significantly improving the accuracy and efficiency of the assessment.

[0044] (4) Multi-agent based dynamic simulation of disaster processes: Through multi-agent system simulation of the dynamic interaction and evolution process between different geographical units or disaster factors in the disaster system, the propagation path and influence range of the disaster can be reflected in real time. This dynamic simulation enables real-time updating of the assessment results, making them more realistic and providing timely and accurate decision support for disaster warning and emergency management. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flowchart of the earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning in the embodiment of the invention;

[0046] Figure 2 The earthquake and secondary disaster vulnerability assessment factor diagram in the embodiment of the invention;

[0047] Figure 3 The sorting weight diagram of elements in the intermediate layer to the decision target in the embodiment of the invention;

[0048] Figure 4 The sorting weight diagram of elements in the disaster-causing factor layer to the decision target in the embodiment of the invention;

[0049] Figure 5 The sorting weight diagram of elements in the disaster-inducing factor layer to the decision target in the embodiment of the invention. DETAILED DESCRIPTION

[0050] As shown in the embodiment, the earthquake and secondary disaster vulnerability assessment method combining multi-agent and deep learning includes the following steps: Figure 1 1. Multi-source data acquisition, data preprocessing and feature engineering

[0051] (1) Multi-source data acquisition

[0052]

[0053] ​The data related to earthquakes and secondary disasters is collected from multiple data sources. The types of data collected include earthquake data (seismic intensity, distance from historical earthquake points, distance from intelligent agents, etc.), topographic data (slope, slope height, distance from water system, etc.), meteorological data (precipitation, etc.), geological structure data (stratum lithology, distance from fault, etc.), human engineering activity data (distance from road, land use type, etc.), and social and economic data (population density, economic density, etc.).

[0054] (2) Data preprocessing

[0055] The collected multi-source data is standardized, normalized, and preprocessed, including coordinate conversion, resampling to a unified spatio-temporal resolution, and noise removal. In this embodiment, the coordinate conversion is unified using the WGS84 geodetic coordinate system or a specific projection coordinate system. Resampling is performed according to the spatial resolution requirement (e.g., 30m), and a suitable resampling method (such as bilinear interpolation, nearest neighbor interpolation, or cubic convolution interpolation) is selected to ensure that the spatial resolution of all data layers is consistent. Noise removal is performed using a Gaussian filter or a median filter to remove noise in grid data such as meteorological data. After preprocessing, the collected multi-source data is obtained as multi-source static data.

[0056] (3) Feature engineering

[0057] Feature extraction is performed on the multi-source static data to ensure the uniformity and usability of multi-source data from different sources, and key features related to earthquake and secondary disaster vulnerability are extracted. The features extracted in this embodiment include seismic intensity, distance from historical earthquake points, distance from intelligent agents, slope, slope height, distance from water system, precipitation, stratum lithology, distance from fault, distance from road, land use type, population density, and economic density.

[0058] In this embodiment, the earthquake data is shown in Tables 1-3, the topographic data is shown in Tables 4-6, the meteorological data is shown in Table 7, the geological structure data is shown in Tables 8 and 9, the human engineering activity data is shown in Tables 10 and 11, and the social and economic data is shown in Tables 12 and 13.

[0059] Table 1 Seismic intensity

[0060] Intensity Classification Assignment V 1 0.5 VI 2 2.5 VII 3 3.5 VIII 4 4.5 IX 5 5

[0061] Table 2 Distance from historical earthquake points

[0062]

[0063]

[0064] Table 3 Distance from intelligent agents

[0065] Agent (buffer, km) Classification Assignment 1 1 5 2 2 4 3 3 3 4 4 2 5 5 1

[0066] Table 4 Slope

[0067] Slope (degree) Classification Assignment ＜1 1 0 1-10 2 2 10-15 3 3 15-20 4 5 20-25 5 3.5 25-30 6 4 30-35 7 2.5 35-40 8 1.5 ＞40 9 1

[0068] Table 5 Elevation

[0069]

[0070]

[0071] Table 6 Distance to water

[0072] Water system (buffer, km) Classification Assignment 0.5 1 5 1 2 4.5 1.5 3 3.5 2 4 3 2.5 5 2.5 3 6 2 3.5 7 1.5 4 8 1 5 9 0.5

[0073] Table 7 Precipitation

[0074] Rainfall (mm) Classification Assignment <200 1 0.5 200-300 2 1 300-400 3 1.5 400-500 4 2 500-600 5 2.5 600-700 6 3 700-800 7 3.5 800-900 8 4 900-1000 9 4.5 >1000 10 5

[0075] Table 8 Formation

[0076]

[0077]

[0078] Table 9 Distance to fault

[0079] Fault (buffer, km) Classification Assignment 0.5 1 5 1 2 4.5 1.5 3 4 2 4 3.5 2.5 5 3 3 6 2.5 3.5 7 2 4 8 1.5 5 9 1

[0080] Table 10 Distance to road

[0081] Road (buffer, km) Classification Assignment 0.5 1 5 1 2 4 1.5 3 3.5 2 4 3 2.5 5 2.5 3 6 2 3.5 7 1.5 4 8 1 5 9 0.5

[0082] Table 11 Land use

[0083] Land use type Assignment Unused land 1 Shrub-grass 2 Forest land 3 Farmland 4 Building land 5

[0084] Table 12 Population density

[0085]

[0086]

[0087] Table 13 Economic density

[0088] Economic density (ten thousand yuan / km2) Assignment 0-1 0 1-100 1 100-500 2 500-800 3 800-1000 4 >1000 5

[0089] 2、AHP (Analytic Hierarchy Process) index modeling and prior weight extraction

[0090] The hierarchical index system of vulnerability assessment is constructed by using expert knowledge, the judgment matrix is constructed by using analytic hierarchy process (AHP), and the weight of each index is calculated by consistency check, and the weight vector is obtained. The weight vector will be used as the prior knowledge of experts to guide the training of the subsequent deep learning model. The analytic hierarchy process can be divided into the following main steps:

[0091] (1) Construct a hierarchical model: target layer (the final goal of the decision problem), criterion layer (the main criteria or factors affecting the decision), and sub-criterion layer (specific sub-criteria under the criteria).

[0092] (2) Construct a judgment matrix: In each level, use the method of pairwise comparison to judge the importance of each element. Use the 1-9 scale method to evaluate the relative importance between elements. Suppose the element set in the hierarchical structure is X1, X2, …, Xn, then for each pair of elements X n and X i , a judgment matrix A = [a j ] needs to be constructed, where: ij

[0093] a ij represents the relative importance of element X i relative to element X j , which satisfies:

[0094]

[0095] i.e. symmetry.

[0096] (3) Calculate the weight vector: Calculate the relative weight of each element through the judgment matrix. The commonly used method is the eigenvalue method:

[0097] Calculate the eigenvector: for the judgment matrix A, solve the eigenvalue problem:

[0098] A·W = λ max ·W

[0099] where W is the eigenvector, representing the weight of each element, and λ max is the maximum eigenvalue.

[0100] Normalize the weight: the calculated eigenvector W needs to be normalized so that the weight sum is 1.

[0101] (4) Consistency check: the analytic hierarchy process needs to check the consistency of the judgment matrix. If the consistency ratio (CR) is too large (greater than or equal to 0.1), the judgment matrix is inconsistent and needs to be adjusted:

[0102] Calculate the consistency index CI:

[0103] ​

[0104] where n is the order of the judgment matrix.

[0105] Calculate the consistency ratio (CR): calculated by the ratio of the consistency index CI and the random consistency index RI (Table 14):

[0106]

[0107] Table 14 RI lookup table

[0108] Order n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 RI 0 0 0.58 0.9 1.12 1.24 1.32 1.41 1.45 1.49 1.51 1.48 1.56 1.57 1.59

[0109] (5) Comprehensive evaluation: the weights of each level are synthesized by weighting to evaluate comprehensively. For each scheme, the final score is obtained by weighted summation according to its weight and score under each criterion.

[0110] (6) Select the optimal scheme: according to the comprehensive score, the scheme with the highest score is selected as the final decision.

[0111] As shown in Figure 2 , in the post-earthquake vulnerability evaluation process, the analytic hierarchy process is used to divide the extracted key features into two categories: disaster-causing factors and disaster-inducing factors. The disaster-causing factors include slope, slope height, stratum, fault, seismic intensity, precipitation and earthquake point; the disaster-inducing factors include multi-agent, population density, economic density, land use, road and water system.

[0112] In this embodiment, the calculation results of the weight of each index are shown in Table 15 and Table 16.

[0113] Table 15 Order weight of elements in scheme layer to decision target

[0114] Alternative Weight Seismic intensity 0.1764 Seismic point 0.1764 Multi-agent 0.1312 Fault 0.0873 Stratum 0.0873 Precipitation 0.0648 Road 0.0496 Land use 0.0496 Water system 0.0496 Slope 0.0482 Population density 0.0266 Economic density 0.0266 Slope height 0.0262

[0115] Table 16 Order weight of elements in intermediate layer to decision target

[0116] Intermediate layer element Weight Disaster-causing factor 0.6667 Disaster-inducing factor 0.3333

[0117] The weight vector constructed in this embodiment is shown in Table 17 and Table 18.

[0118] Disaster-causing factor consistency ratio: 0.0082, weight of "post-earthquake vulnerability evaluation": 0.6667; λ max : 7.0668.

[0119] Table 17

[0120]

[0121] Disaster factor consistency ratio: 0.0044; weight for "post-earthquake vulnerability assessment": 0.3333; λ max : 6.0275.

[0122] Table 18

[0123]

[0124] 3. Multi-agent system construction and disaster propagation dynamic simulation

[0125] A multi-agent system is constructed, with each agent representing a geographic unit or a disaster factor (earthquake wave propagation unit, landslide prone area, debris flow path, etc.); each agent is given autonomous decision-making ability and can exchange information with other agents; by designing local information exchange and coordination rules between agents, the propagation path and spatio-temporal evolution process of disasters are dynamically simulated, and dynamic evolution data are obtained to provide dynamic input data for deep learning models.

[0126] (1) Multi-agent system construction parameters

[0127] Agent definition: Each agent represents a geographic unit, such as a "1 km x 1 km" grid area. The number of agents is set according to the size and resolution of the study area, which can be hundreds to thousands.

[0128] (2) Interaction mechanism in dynamic simulation process

[0129] ① Information exchange frequency: set to once every hour or every 6 hours according to the disaster evolution speed.

[0130] ② Agent decision-making model: use rule-based or reinforcement learning (such as Q-learning) to simulate agent behavior.

[0131] ③ Information sharing mechanism: set to local information sharing, i.e. agents only exchange information with their neighboring agents, such as vibration intensity, landslide state, etc.

[0132] (3) Disaster propagation simulation parameters

[0133] ① Propagation speed: set the propagation speed of secondary disasters according to geological conditions and disaster types (e.g. debris flow rate set to 2-10 meters / hour).

[0134] ② Influence range: based on historical disaster data or physical model estimation, for example, the radius of landslide prone area may be 500-1000 meters.

[0135] 4. GCN+Transformer deep learning model training and evaluation

[0136] A GCN+Transformer hybrid deep learning model is designed. The GCN layer is used to capture the spatial dependency between agents (geographical units), while the Transformer layer uses its self-attention mechanism to process time series data generated by the multi-agent system and capture the dynamic characteristics of disaster evolution.

[0137] The pre-processed multi-source static data (such as geological and topographical data) is combined with the dynamic evolution data generated by the multi-agent system to construct a spatio-temporal feature tensor, which serves as the input to the GCN+Transformer model.

[0138] The Adam optimizer and learning rate decay strategy are used to improve the convergence speed and generalization ability of the model. A regularization term based on AHP weights is introduced into the loss function to build a multi-task learning framework, allowing the model to fit vulnerability labels while aligning with expert knowledge, thereby enhancing the model's explainability and decision-making credibility.

[0139] The model is trained on historical labeled data sets and evaluated on independent test sets. The mean squared error (MSE), root mean squared error (RMSE), and other indicators are used to measure the prediction performance, while the model's fitting effect and interpretability under AHP guidance are evaluated to ensure the model's accuracy and expert consistency.

[0140] (1) Model architecture design

[0141] ① Graph construction: An adjacency matrix is constructed using geographical adjacency relationships to represent the spatial dependency between agent units. Each graph node represents an agent unit.

[0142] ② GCN layer: 2-3 layers of graph convolution layers are set up to aggregate the spatial features of neighboring nodes.

[0143] ③ Transformer layer: 4-6 encoder layers are set up to process time series data generated by the multi-agent system and capture temporal dependencies.

[0144] (2) AHP weight embedding

[0145] AHP weights are embedded into the model in the following ways: first, as part of the loss function, i.e., introducing AHP consistency regularization term LAHP; second, weighting the features at the input end to guide the model to focus on important features.

[0146] (3) Optimization algorithm and model training parameters

[0147] ① Optimizer: The Adam optimizer is used with an initial learning rate of 1e-3, and a learning rate decay strategy can be configured, such as decaying by 0.5 every 10 epochs.

[0148] ②Training batch: Batch size is set to 32-64, adjusted according to GPU memory.

[0149] ③Training rounds: Training rounds are set to 100-200 rounds, or until the model converges on the validation set.

[0150] ④Loss function: Mean Squared Error (MSE) is used as the main loss function, combined with AHP consistency regularization term to form a composite loss function.

[0151] (4) Vulnerability assessment parameters

[0152] ①Prediction time domain: According to the type of disaster, dynamic vulnerability prediction for 1 hour to 24 hours in the future is set.

[0153] ②Output: The output is the vulnerability index of each agent unit at different time points in the future, as well as a disaster risk prediction report containing risk levels, regions, and intensity.

[0154] 5、Vulnerability assessment results and decision support

[0155] (1) GIS data fusion and visualization display

[0156] Fuse the vulnerability level output by the GCN+Transformer model with spatial geographic data to generate high-precision disaster vulnerability distribution maps. The results can be used for visualization platform display and form fine spatial risk zoning to provide intuitive geographic support for urban or mountain disaster management.

[0157] ①Result fusion: Spatially superimpose the vulnerability index output by the model with social and economic data (population, GDP) to calculate risk loss.

[0158] ②Mapping: Combine vulnerability scores with spatial information to generate vulnerability level distribution maps with spatial resolution of 30 meters to 1 kilometer.

[0159] ③Visualization: Use heat maps, hierarchical color maps, and other methods to display disaster risk areas.

[0160] (2) Emergency response optimization and decision support

[0161] Integrate disaster emergency response optimization algorithms with vulnerability assessment results to help relevant departments develop efficient emergency response plans; establish a decision support system to provide scientific and accurate decision-making basis for governments and relevant departments, and improve post-disaster recovery and emergency management efficiency.

[0162] ①Resource scheduling optimization: Based on disaster impact range and vulnerability level, use optimization algorithms (such as genetic algorithm or linear programming) to generate optimal allocation schemes for emergency resources (such as rescue teams, medical supplies).

[0163] ② Decision Support System: Integrate the above models and algorithms to build a visual decision support system, providing disaster management departments with pre-disaster warning, disaster response, and post-disaster recovery decision support.

[0164] The implementation of this embodiment can achieve high-precision dynamic assessment and prediction of earthquake and secondary disaster vulnerability. The combination of multi-agent system and GCN+Transformer architecture effectively simulates the complex dynamic changes during the disaster occurrence, fully excavates and utilizes the rich information in multi-source data, thereby significantly improving the precision and efficiency of vulnerability assessment. In addition, the deep integration and visualization of geographic information system (GIS) data provide intuitive and scientific decision support for disaster warning and emergency management, further enhancing the precision and timeliness of disaster response.

Claims

1. A method for assessing the vulnerability of earthquakes and their secondary disasters by combining multi-agent and deep learning, characterized in that, Includes the following steps: Step (1): Collect data related to earthquakes and their infrasound hazards and preprocess the data to obtain multi-source static data; extract key features from the multi-source static data; Step (2): The analytic hierarchy process (AHP) is used to construct an evaluation model and determine the weights to obtain the AHP weight vector. The comprehensive score and evaluation of the decision analysis are then performed based on the fuzzy comprehensive evaluation method. Step (3): Construct a multi-agent system and endow each agent with autonomous decision-making capabilities; by designing local information exchange and collaborative action rules among agents, dynamically simulate the propagation path and spatiotemporal evolution of disasters to obtain dynamic evolution data; Step (4): Design the GCN+Transformer deep learning model, and combine multi-source static data with dynamic evolution data to construct a spatiotemporal feature tensor as the input of the GCN+Transformer deep learning model for model training and evaluation. Step (5): Use the trained GCN+Transformer deep learning model to assess the vulnerability of earthquakes and their secondary disasters, and output the vulnerability level index. Then, integrate the vulnerability level index with socio-economic data and / or spatial geographic data to generate a disaster vulnerability distribution map.

2. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 1, characterized in that, In step (1), the data related to earthquakes and their infrasound hazards include earthquake data, topographic data, meteorological data, geological structure data, human engineering activity data, and socio-economic data; Preprocessing of multi-source data includes coordinate transformation, resampling to a uniform spatiotemporal resolution, and noise removal; coordinate transformation uniformly adopts the WGS84 geodetic coordinate system or a specific projected coordinate system; resampling is performed by selecting bilinear interpolation, nearest neighbor interpolation, or cubic convolution interpolation according to spatial resolution requirements; and noise in the raster data of meteorological data is removed using Gaussian filters or median filters. The key features extracted include seismic intensity, distance from historical earthquake points, and distance from the agent in seismic data; slope, slope height, and distance from water systems in topographic data; precipitation in meteorological data; lithology and distance from faults in geological structure data; distance from roads and land use type in human engineering activity data; and population density and economic density in socioeconomic data.

3. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 2, characterized in that, In step (2), the hierarchical index system of vulnerability assessment includes disaster-causing factors and disaster-inducing factors; disaster-causing factors include slope, slope height, strata, faults, seismic intensity, precipitation and earthquake points; disaster-inducing factors include multi-agent, population density, economic density, land use, roads and water systems; Step (2) specifically includes the following steps: Step (2-1): Construct a hierarchical index system for vulnerability assessment using expert knowledge, including the target layer, criterion layer, and sub-criterion layer; Step (2-1): Construct a judgment matrix using the analytic hierarchy process (AHP): At each level, use pairwise comparisons to determine the importance of each element; use a 1-9 scale to assess the relative importance of elements; assume the set of elements in the hierarchical structure is X1, X2, ..., X... n Then for each pair of elements X i and X j We need to construct a judgment matrix A = [a ij ], a ij Representing element X i Relative to element X j The relative importance of , and satisfying symmetry; Step (2-3): Calculate and normalize the AHP weight vector: Calculate the relative weight of each element using the eigenvector method based on the judgment matrix; the eigenvector is calculated as: A·W=λ max ·W; where W is the eigenvector, representing the weight of each element, and λ max It is the largest eigenvalue; The weights sum to 1 after normalization; Step (2-4), Consistency Check: If the consistency ratio is greater than or equal to 0.1, it indicates that the judgment matrix is ​​inconsistent and needs to be readjusted; the consistency ratio CR is calculated as follows: In the formula, CI is the consistency index, RI is the random consistency index, and n is the order of the judgment matrix. Steps (2-5), comprehensive evaluation and decision-making: The weights of each level are combined and evaluated comprehensively. For each solution, a weighted sum is calculated based on its weight and score under each criterion to obtain the final score.

4. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 3, characterized in that, In step (3), the intelligent agent system construction parameters are: each intelligent agent represents 1 geographical unit, and the number of intelligent agents is 100-5000; In the dynamic simulation of the propagation path and spatiotemporal evolution of disasters: the information exchange frequency is set to once per hour or once every 6 hours according to the speed of disaster evolution; the agent decision-making model simulates agent behavior in a rule-based or reinforcement learning-based manner; the information sharing mechanism is set to local information sharing, that is, the agent only exchanges information with its neighboring agents; Disaster propagation simulation parameters: Based on geological conditions and disaster type, the propagation speed of secondary disasters is set, with debris flow rate set at 2-10 m / h; the impact range of disaster propagation is estimated based on historical disaster data or physical models, with the radius of landslide-prone areas being 500-1000 meters.

5. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 4, characterized in that, In step (4), the GCN+Transformer deep learning model architecture is designed as follows: (4-1) Graph construction: Construct an adjacency matrix using geographical adjacency relationships to represent the spatial dependencies between intelligent agents; each graph node represents an intelligent agent. (4-2) GCN layer: Set 2-3 graph convolutional layers to aggregate the spatial features of neighboring nodes and capture the spatial dependencies between agents; (4-3) Transformer layer; set up 4-6 encoder layers to process time series data generated by multi-agent systems and capture the dynamic characteristics of disaster evolution.

6. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 5, characterized in that, In step (4), expert knowledge is used to guide model training: the AHP weight vector is used as part of the loss function, that is, the AHP consistency regularization term LAHP is introduced; the key features are weighted at the input end to guide the model to focus on important features.

7. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 6, characterized in that, In step (4), the Adam optimizer is used, the initial learning rate is set to 1e-3, and the learning rate decay strategy is configured, that is, the learning rate decays by 0.5 every 10 epochs of training; the training batch is set to 32-64; the number of training epochs is set to 100-200 epochs, or until the model converges on the validation set; the mean square error and / or root mean square error are used as the main loss function, and combined with the AHP consistency regularization term to form a composite loss function.

8. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 7, characterized in that, In step (4), the model is trained on a historical labeled dataset and its performance is evaluated on an independent test set; the vulnerability assessment parameters include the prediction time domain and the output; the prediction time domain is set to the dynamic vulnerability prediction for the next 1 hour to 24 hours according to the disaster type. The output includes vulnerability indices for each intelligent agent at different points in the future, as well as disaster risk prediction reports containing risk level, region, and intensity.

9. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 8, characterized in that, In step (5), the vulnerability index output by the model is spatially overlaid with socioeconomic data to calculate risk loss; the vulnerability index is combined with spatial geographic data to generate a vulnerability level distribution map with a spatial resolution of 30 meters to 1 kilometer. Disaster risk areas can be displayed using heat maps and / or graded color maps.

10. The method for assessing earthquake and secondary disaster vulnerability by combining multi-agent and deep learning according to claim 9, characterized in that, Step (5) also includes emergency response optimization and decision support: including resource scheduling optimization and decision support systems; Resource scheduling optimization: Based on the scope of disaster impact and vulnerability level, the optimal allocation scheme of emergency resources is generated using genetic algorithms or linear programming optimization algorithms; Decision Support System: Integrates the trained GCN+Transformer deep learning model and algorithm to build a visualized decision support system, providing disaster management departments with full-process decision support for pre-disaster early warning, in-disaster response and post-disaster recovery.

Images