Geological disaster early warning method, system, device and medium based on artificial intelligence

Abstract

The application relates to a geological disaster early warning method, system, device and medium based on artificial intelligence. The method comprises the following steps: acquiring multi-source geological disaster monitoring original data, inputting the multi-source geological disaster monitoring original data into a geological disaster feature extraction model, performing parallel feature extraction, and obtaining a multi-modal geological disaster feature vector set; performing risk probability prediction based on the multi-modal geological disaster feature vector set through a dynamic geological disaster space-time graph model, generating geological disaster risk probability features and uncertainty metrics; performing feature fusion on short-impending forecast information, the geological disaster risk probability features and the uncertainty metrics, and obtaining comprehensive risk features; and performing early warning decision-making based on a dynamic cost matrix and the comprehensive risk features, and generating a geological disaster early warning instruction. The method can introduce uncertainty, realize adaptive early warning and release, and improve the accuracy of early warning.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence early warning, and in particular relates to methods, systems, equipment and media for geological disaster early warning based on artificial intelligence. Background Technology

[0002] The rapid development of geographic information and sensing technologies has spurred the emergence of automated geological disaster monitoring technologies. These technologies enable continuous data collection and remote transmission of key parameters such as terrain changes, rainfall, and displacement, significantly expanding the scope of human perception. Consequently, real-time monitoring based on sensor networks combined with manual inspections has become the mainstream disaster early warning method.

[0003] In traditional technologies, early warning decisions primarily rely on threshold judgments for single or limited monitoring indicators. Experts analyze data such as rainfall and displacement rates, combining historical experience with regional geological knowledge to set fixed warning thresholds. Once the monitored data exceeds the preset standard, the system triggers an early warning. The entire process is highly dependent on the expert's understanding of the geological conditions of the specific area and the experience with threshold setting.

[0004] However, current threshold-based early warning methods have inherent limitations. Geological disasters are the product of complex coupling effects of multiple factors such as geology, meteorology, and hydrology. Single or static thresholds are difficult to effectively characterize their nonlinear evolution process, and are prone to failure, especially under rare or extreme conditions. This method is essentially a passive perception-response mode, lacking a deep understanding of the disaster's formation mechanism and spatial correlation, and cannot achieve advanced risk assessment and precise location. In the face of vast mountainous areas and complex terrain, traditional monitoring networks are costly to deploy and have many blind spots, resulting in significant shortcomings in the spatial coverage and timeliness of the early warning system. Summary of the Invention

[0005] Therefore, it is necessary to provide an artificial intelligence-based geological disaster early warning method, system, equipment, and medium that can accurately locate risks and make adaptive decision-making and early warning in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a geological disaster early warning method based on artificial intelligence, including:

[0007] The raw data of multi-source geological disaster monitoring are obtained and input into the geological disaster feature extraction model for parallel feature extraction to obtain a set of multi-modal geological disaster feature vectors.

[0008] By using a dynamic spatiotemporal map model of geological hazards, and based on a set of multimodal geological hazard feature vectors, risk probability prediction is performed, generating geological hazard risk probability characteristics and uncertainty measures.

[0009] A comprehensive risk characteristic is obtained by fusing short-term forecast information, geological hazard risk probability characteristics, and uncertainty measures.

[0010] Based on the dynamic cost matrix and comprehensive risk characteristics, early warning decisions are made and geological disaster early warning instructions are generated; among them, geological disaster early warning instructions are used to instruct forecast and early warning actions.

[0011] Furthermore, the method also includes:

[0012] Using each monitoring unit as a node and the geological hazard feature vector corresponding to the monitoring unit as the node attribute, entity nodes are constructed, and the relationships between entity nodes are extracted to obtain the edges of the geological hazard spatiotemporal map model; among which, the relationships include spatial adjacency relationships, geological structure relationships, and hydrological path relationships;

[0013] The node attributes are input into the dynamic weight model to calculate the dynamic weights of the edges between entity nodes.

[0014] Based on entity nodes, edges, and dynamic weights, a primary spatiotemporal map model of geological hazards is constructed.

[0015] Based on weight thresholds and dynamic weights, redundant edges are pruned from the primary geological hazard map model to obtain a dynamic geological hazard spatiotemporal map model.

[0016] Furthermore, through a dynamic spatiotemporal model of geological hazards, based on a multimodal set of geological hazard feature vectors, risk probability prediction is performed, generating geological hazard risk probability characteristics and uncertainty measures, including:

[0017] Based on the multimodal geological hazard feature vector set, the node attributes of the dynamic geological hazard spatiotemporal map model are updated, and the node information of the dynamic geological hazard spatiotemporal map model is aggregated to obtain a high-order node feature vector set;

[0018] The set of feature vectors of high-order nodes is input into a Bayesian neural network to perform risk probability prediction, and obtain the disaster occurrence probability and uncertainty measure for each node.

[0019] Based on the probability of disaster occurrence and the dynamic weights in the dynamic geological disaster spatiotemporal map model, spatial risk propagation simulation is performed to obtain the node risk probability value;

[0020] By mapping the node risk probability value and uncertainty measure to the risk level corresponding to the risk level rule, a risk level partitioning map is obtained;

[0021] Summary features are extracted from the risk level zoning map, and geological hazard risk probability features are obtained based on the summary features.

[0022] Furthermore, feature fusion is performed on short-term forecast information, geological hazard risk probability characteristics, and uncertainty measures to obtain comprehensive risk characteristics, including:

[0023] Based on the type of short-term forecast information, the short-term forecast information is encoded to obtain independent short-term disaster-causing signal vectors and short-term uncertainty measures;

[0024] A linear transformation is performed on the probability characteristics and uncertainty measure of geological disaster risk to obtain a query vector. A linear projection is performed on the independent short-term disaster-causing signal vector to obtain a key vector and a value vector.

[0025] Calculate the relevance score between the query vector and each key vector to obtain the initial relevance score, and adjust the initial relevance score based on the short-term uncertainty measure to obtain the attention weight distribution list;

[0026] Based on the attention weight distribution list, a weighted sum of all value vectors is performed to obtain the short-term neighbor signal context vector;

[0027] By integrating the context vector of short-term signals and the probability characteristics of geological disaster risks, a comprehensive risk characteristic is obtained.

[0028] Furthermore, based on the dynamic cost matrix and comprehensive risk characteristics, early warning decisions are made to obtain geological disaster early warning instructions, including:

[0029] Based on comprehensive risk characteristics, scenario analysis is performed to obtain scenario features, and based on these scenario features, the basic cost variables are adjusted to obtain a dynamic cost matrix.

[0030] Based on the dynamic cost matrix, the expected cost of early warning actions is calculated; the elements in the dynamic cost matrix include the predicted probability of geological disasters, the cost of false alarms, and the cost of missed alarms.

[0031] The formula for calculating the expected cost is as follows:

[0032]

[0033]

[0034] in, The expected cost of issuing an early warning, The expected cost of not issuing a warning, The predicted probability of geological disasters. To pay the price for false alarms The cost of underreporting;

[0035] Based on the expected cost, a binary decision-making process for early warning is conducted to obtain binary decision instructions, and a comprehensive uncertainty measure is extracted from the comprehensive risk characteristics.

[0036] Based on binary decision instructions and comprehensive uncertainty measurement, early warning level decisions are made to obtain geological disaster early warning instructions.

[0037] Furthermore, the attention weights in the attention weight distribution list are calculated using the following formula:

[0038]

[0039] in, For attention weights, The corrected score for the i-th short-term forecast signal is given by [the following]. For temperature parameters, Let i be the corrected score of the j-th short-term forecast signal, where i and j are the indices of the short-term forecast signal.

[0040] The formula for calculating the corrected score of the short-term forecast signal is as follows:

[0041]

[0042] in, Let Q be the corrected score for the i-th short-term forecast signal, and let Q be the query vector. Let be the key vector of the i-th short-term forecast signal. Let be the dimension of the key vector. The uncertainty penalty coefficient, Let be the short-term uncertainty measure of the i-th short-term forecast signal.

[0043] Secondly, this application also provides an artificial intelligence-based geological disaster early warning system, including:

[0044] The feature module is used to acquire raw data of multi-source geological disaster monitoring and input the raw data of multi-source geological disaster monitoring into the geological disaster feature extraction model for parallel feature extraction to obtain a set of multimodal geological disaster feature vectors;

[0045] The prediction module is used to predict the risk probability based on a multimodal geological hazard feature vector set using a dynamic geological hazard spatiotemporal map model, and to generate geological hazard risk probability characteristics and uncertainty measures.

[0046] The fusion module is used to fuse short-term forecast information, geological hazard risk probability characteristics, and uncertainty measures to obtain comprehensive risk characteristics.

[0047] The decision-making module is used to make early warning decisions based on the dynamic cost matrix and comprehensive risk characteristics, and generate geological disaster early warning instructions; among them, the geological disaster early warning instructions are used to instruct the forecasting and early warning actions.

[0048] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the method provided in the first aspect of this application.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0050] The aforementioned AI-based geological disaster early warning method, system, equipment, and media acquire multi-source geological disaster monitoring raw data and input it into a geological disaster feature extraction model for parallel feature extraction, resulting in a multimodal geological disaster feature vector set. Using a dynamic geological disaster spatiotemporal map model, based on the multimodal geological disaster feature vector set, risk probability prediction is performed, generating geological disaster risk probability features and uncertainty measures. Short-term forecast information, geological disaster risk probability features, and uncertainty measures are fused to obtain comprehensive risk features. Based on the dynamic cost matrix and comprehensive risk features, early warning decisions are made, generating geological disaster early warning instructions. These instructions are used to instruct on forecasting and early warning actions. The system can introduce machine learning and spatiotemporal prediction models to explore the complex nonlinear relationship between disaster-causing factors and disaster occurrence, achieving a shift from static threshold judgment to dynamic probability prediction, dynamically updating risk probabilities, and accurately capturing risks. It introduces uncertainty, providing confidence assessments for each risk prediction value, and employs risk-based decision theory to achieve adaptive and interpretable early warning issuance, replacing rigid fixed thresholds. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram illustrating the process of an artificial intelligence-based geological disaster early warning method according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based geological disaster early warning system provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based geological disaster early warning method is provided. This embodiment illustrates the application of this method to a geological disaster early warning terminal. It is understood that this method can also be applied to a geological disaster early warning server, and further to a geological disaster early warning system including both a geological disaster early warning terminal and a geological disaster early warning server, and is implemented through the interaction between the two. In this embodiment, the method includes the following steps:

[0056] Step 101: Obtain the original data of multi-source geological disaster monitoring, and input the original data of multi-source geological disaster monitoring into the geological disaster feature extraction model to perform parallel feature extraction and obtain a set of multi-modal geological disaster feature vectors.

[0057] Optionally, raw data for multi-source geological disaster monitoring refers to raw observation data about geological disasters obtained from various types of monitoring equipment or channels that have not yet undergone in-depth processing.

[0058] Optionally, the raw data for multi-source geological hazard monitoring can originate from surface displacement monitoring stations, rain gauges, groundwater level sensors, remote sensing images, etc. The geological hazard feature extraction model is a specially designed computational model that can automatically identify and extract key information patterns that characterize the formation and occurrence process of geological hazards from the raw data. The multimodal geological hazard feature vector set refers to a set of structured data obtained after feature extraction. Multimodal features originate from different types of raw data; a feature vector can represent the key information of each monitoring unit or each data mode using a set of values; the multimodal geological hazard feature vector set is represented as the sum of feature vectors from all monitoring units or all modes.

[0059] Optionally, the geological disaster early warning terminal can input the raw data into a pre-trained geological disaster feature extraction model. This model can simultaneously analyze and process data from different sources and of different types, extracting more abstract and favorable signals for subsequent risk assessment from massive amounts of raw data through its internal computational structure. The signals are then regularized into numerical vectors, outputting a comprehensive set containing feature vectors from all monitoring points and all data modalities. Alternatively, based on the type of raw data from multi-source geological disaster monitoring, different feature extraction models can be used for feature extraction.

[0060] Step 102: Using a dynamic geological disaster spatiotemporal map model, based on a multimodal geological disaster feature vector set, risk probability prediction is performed to generate geological disaster risk probability characteristics and uncertainty measures.

[0061] Specifically, a dynamic spatiotemporal geological hazard model can be a mathematical model used to describe the spatial and temporal evolution of geological hazard risks. Each monitoring point is modeled as a node in the graph, and the geological, hydrological, and spatial relationships between monitoring points are modeled as edges. The weights of these edges can dynamically change based on real-time data, thereby simulating the propagation and interaction of risks in space. Geological hazard risk probability characteristics are comprehensive feature representations that can be used to quantify the likelihood of a geological hazard occurring in a specific area within a specific time period. These characteristics can include risk probability information after spatial propagation simulation and aggregation. Uncertainty measures can be numerical indicators used to quantify the credibility or confidence level of risk prediction results. They indicate how confident the model is in its risk probability estimates; a higher uncertainty measure value generally indicates greater uncertainty and lower credibility in the prediction results.

[0062] Optionally, the geological disaster early warning terminal can input a set of multimodal feature vectors into a dynamic geological disaster spatiotemporal map model. The dynamic geological disaster spatiotemporal map model can update the state attributes of each node in the map based on the new features. Along the defined edges, multiple rounds of information transmission and aggregation are carried out between nodes, so that each node can not only perceive its own state, but also perceive the risk status of its associated area, thereby generating high-order node features containing broader spatial information. The high-order node features are fed into a Bayesian neural network that can simultaneously output the predicted value and its uncertainty, calculate the initial probability and uncertainty of disaster occurrence at each node, and simulate the propagation of the initial probability in the spatial network based on the dynamic weights of the edges in the map, calculate the final stable risk probability value of each node, and abstract it together with the uncertainty measure to obtain the geological disaster risk probability feature.

[0063] Step 103: Perform feature fusion on short-term forecast information, geological disaster risk probability characteristics, and uncertainty measurement to obtain comprehensive risk characteristics.

[0064] Specifically, short-term forecast information refers to forecast information on severe weather or other external events that may trigger geological disasters in the very short term.

[0065] For example, short-term forecast information can be refined rainstorm forecasts, strong earthquake early warnings, etc. Comprehensive risk characteristics are feature representations that integrate medium- and long-term geological background risk characteristics with short-term disaster-causing signals. They are risk status descriptions that are more timely and targeted, formed by embedding short-term forecast information on the basis of risk probability characteristics.

[0066] Optionally, the geological disaster early warning terminal can encode short-term forecast information, converting it into machine-processable vector signals and assessing its own forecast uncertainty. The geological disaster early warning terminal converts the geological disaster risk probability features into query vectors, converts the short-term forecast signal vectors into key vectors and value vectors respectively, calculates the correlation between the query vector and the key vector of each short-term signal, corrects the correlation score based on the uncertainty measure of each short-term signal, normalizes all corrected scores to obtain a set of attention weights, which can be used to characterize the importance of each short-term signal during fusion, and uses the attention weights to perform a weighted summation of the value vectors of all short-term signals to generate a context vector that condenses all key short-term information. The geological disaster early warning terminal fuses the context vector with the original geological risk probability features to obtain comprehensive risk features.

[0067] Step 104: Based on the dynamic cost matrix and comprehensive risk characteristics, make early warning decisions and generate geological disaster early warning instructions; wherein, the geological disaster early warning instructions are used to instruct forecast and early warning actions.

[0068] Optionally, a dynamic cost matrix is ​​a mathematical tool used to quantify the consequences of early warning decisions. The dynamic cost matrix includes the potential costs of issuing or not issuing an early warning under different scenarios. These mainly include the cost of false alarms (the socio-economic costs of issuing an early warning but the disaster not occurring) and the cost of underreporting (the casualties and losses caused by a disaster occurring but no early warning being issued). The costs in the matrix can be dynamically adjusted according to the specific scenario revealed by the current comprehensive risk characteristics. Geological disaster early warning instructions are specific instructions that can be directly executed by lower-level early warning systems. They at least include a binary decision of whether to issue an early warning, and may also include specific early warning levels, which indicate the corresponding early warning information release and emergency response actions.

[0069] Optionally, the geological disaster early warning terminal can analyze the current specific situation based on comprehensive risk characteristics, adjust the general benchmark values ​​of false alarm cost and false alarm cost, and form a dynamic cost matrix adapted to the current time and current region. The geological disaster early warning terminal combines the geological disaster occurrence prediction probability extracted from the comprehensive risk characteristics and uses the cost matrix to calculate the expected cost of issuing an early warning and not issuing an early warning respectively. The geological disaster early warning terminal compares the two expected costs and selects the decision with the smaller expected cost as the initial binary decision instruction. Under the premise of deciding to issue an early warning, the geological disaster early warning terminal further combines information such as the uncertainty measure in the comprehensive risk characteristics to refine and determine the final early warning level and generate a complete geological disaster early warning instruction.

[0070] This embodiment provides an AI-based geological disaster early warning method. It acquires raw data from multi-source geological disaster monitoring and inputs this data into a geological disaster feature extraction model for parallel feature extraction, resulting in a multimodal geological disaster feature vector set. Using a dynamic geological disaster spatiotemporal map model, based on the multimodal geological disaster feature vector set, it performs risk probability prediction, generating geological disaster risk probability features and uncertainty metrics. It then fuses short-term forecast information, geological disaster risk probability features, and uncertainty metrics to obtain comprehensive risk features. Based on the dynamic cost matrix and comprehensive risk features, it makes early warning decisions and generates geological disaster early warning instructions. These instructions are used to instruct on forecasting and early warning actions. By introducing machine learning and spatiotemporal prediction models, it can uncover the complex nonlinear relationship between disaster-causing factors and disaster occurrence, achieving a shift from static threshold judgment to dynamic probability prediction, dynamically updating risk probabilities, and accurately capturing risks. By introducing uncertainty, it provides confidence assessments for each risk prediction value and employs risk-based decision theory to achieve adaptive and interpretable early warning issuance, replacing rigid fixed thresholds.

[0071] In an optional embodiment of this application, the artificial intelligence-based geological disaster early warning method may further include:

[0072] Step 201: Using each monitoring unit as a node and the geological hazard feature vector corresponding to the monitoring unit as the node attribute, construct entity nodes and extract the relationships between entity nodes to obtain the edges of the geological hazard spatiotemporal graph model; wherein, the relationships include spatial adjacency relationships, geological structure relationships and hydrological path relationships.

[0073] Optionally, a monitoring unit refers to a basic spatial unit or location that is independently monitored within a geological disaster monitoring network.

[0074] Optionally, a monitoring unit can be a specific landslide monitoring point, a slope unit, or a grid area. An entity node can be a basic element in a graph model representing a monitoring unit in the real world. In this embodiment, each monitoring unit can be modeled as a node. Node attributes can be assigned to each entity node. Node attributes can be characteristic data used to describe its state. In this embodiment, node attributes can be defined as the geological hazard feature vector corresponding to the monitoring unit, i.e., the feature representation of the unit in the multimodal feature vector set. An edge can be an element used in a graph model to connect two nodes. An edge can be an element representing a relationship between nodes. In this embodiment, it specifically refers to the edges connecting different monitoring units in the spatiotemporal graph model of geological hazards. Spatial adjacency can refer to whether two monitoring units are directly adjacent in geographic space or located in the same slope unit, etc. Geological structure relationship can refer to the correlation or similarity between the geological bodies to which two monitoring units belong in terms of structure, lithology, strata, etc. Hydrological path relationship can refer to the connection between two monitoring units on the surface or groundwater flow path, such as being in the same confluence path or having hydraulic connections.

[0075] Optionally, the geological disaster early warning terminal can instantiate each monitoring unit in the monitoring network as an entity node in the graph model. The geological disaster early warning terminal can assign the feature vector corresponding to each monitoring unit in the multimodal feature vector set to the entity node representing the monitoring unit as the node attribute of the node, thereby completing the initialization of the entity node. According to the predefined relationship rules, the terminal can extract and establish connection relationships among all entity nodes.

[0076] For example, a geological disaster early warning terminal can analyze whether there is a spatial adjacency relationship between node pairs, such as whether they share a boundary; geological structure relationship, such as whether the properties of the soil and rock are similar or belong to the same structure; and hydrological path relationship, such as whether the water flow direction is connected. As long as any of these relationships are satisfied, an edge will be established between the two corresponding entity nodes.

[0077] Step 202: Input the node attributes into the dynamic weight model to calculate the dynamic weights of the edges between entity nodes.

[0078] Specifically, a dynamic weight model can be a specialized machine learning or computational model. It can be used to calculate the weight of an edge connecting two nodes based on the input node attributes, so that the edge weight is no longer a fixed value but can be updated in real time as the node state changes. Dynamic weight refers to the numerical value assigned to each edge in the graph, calculated by the dynamic weight model. Dynamic weight can quantify the strength of the mutual influence or the ease of risk transmission between the two nodes connected by the edge, and it changes dynamically as the node attributes are updated.

[0079] Optionally, the geological disaster early warning terminal can input the node attributes of two connected nodes into the dynamic weight model. The dynamic weight model analyzes and compares the two feature vectors through its internal algorithm, learns the association pattern, and outputs a scalar value, which is set as the dynamic weight of the edge connecting the two nodes.

[0080] Step 203: Based on entity nodes, edges, and dynamic weights, a primary spatiotemporal map model of geological disasters is constructed.

[0081] Specifically, the primary geological hazard spatiotemporal map model is a preliminary, unfiltered, and unoptimized geological hazard risk propagation network map.

[0082] The primary geological hazard spatiotemporal map model is a complete data structure that includes all identified monitoring units, all identified relationships, and the real-time impact intensity value calculated for each edge. The primary geological hazard spatiotemporal map model is a preliminary digital mapping of the real monitoring network and its complex interrelationships.

[0083] The geological disaster early warning terminal can integrate all entity nodes, all edges, and the specific dynamic weights calculated for each edge to create a graph data structure. The geological disaster early warning terminal binds the node set, edge list, and the dynamic weight values ​​attached to each edge together to form a unified primary geological disaster spatiotemporal graph model that can be used for mathematical operations and graph calculations. The primary geological disaster spatiotemporal graph model fully retains all connections derived from the original data and relational rules.

[0084] Step 204: Based on the weight threshold and dynamic weight, redundant edges are pruned on the primary geological hazard map model to obtain the dynamic geological hazard spatiotemporal map model.

[0085] Specifically, the weight threshold can be a preset numerical threshold used to determine whether the dynamic weight of an edge is significant. Edges with weights higher than this threshold are considered important connections that need to be retained; edges with weights lower than this threshold are considered weak or unimportant. Redundant edge pruning refers to the operation of removing edges considered unimportant from the graph model. In this embodiment, redundant edges specifically refer to edges with dynamic weights lower than the preset weight threshold. The dynamic geological disaster spatiotemporal graph model can refer to the graph model obtained after redundant edge pruning of the primary graph model, retaining core, high-impact risk propagation paths, resulting in a more streamlined model structure, and its edge weights have dynamically changing characteristics.

[0086] The geological disaster early warning terminal can traverse every edge in the primary geological disaster spatiotemporal graph model, comparing the dynamic weight values ​​associated with the model with preset weight thresholds one by one. For any edge, if its dynamic weight value is strictly less than the weight threshold, the edge is determined to be a weak connection and permanently removed from the edge set of the graph model. This process continues until the weights of all edges have been compared. After pruning, only edges with dynamic weights greater than or equal to the threshold and the nodes connected by such edges are retained in the graph, generating a sparser and cleaner graph network.

[0087] This embodiment, by constructing a graph model, can provide a complete topological foundation for risk assessment, ensuring that no possible risk interaction channels are missed in subsequent analysis; by eliminating noisy connections, the structure of the spatiotemporal graph model of geological disasters can be optimized, allowing computational resources to be focused on key risk interactions, improving the computational efficiency of risk simulation and prediction, reducing model complexity, and enhancing the interpretability of the model.

[0088] In one embodiment, a dynamic geological hazard spatiotemporal map model is used to predict risk probability based on a multimodal geological hazard feature vector set, generating geological hazard risk probability characteristics and uncertainty measures, including:

[0089] Step 301: Based on the multimodal geological hazard feature vector set, update the node attributes of the dynamic geological hazard spatiotemporal map model, and aggregate the node information of the dynamic geological hazard spatiotemporal map model to obtain a high-order node feature vector set.

[0090] Optionally, the multimodal geological hazard feature vector set can be a set of feature vectors extracted in parallel from monitoring data, representing the current state of each geological monitoring unit. The dynamic geological hazard spatiotemporal graph model can be a graph network model composed of entity nodes, edges, and dynamic weights, used to simulate risk propagation. Node attributes can be feature data carried by each entity node in the graph, describing its own state. In this embodiment, node attributes can be updated by a new set of feature vectors. Node information aggregation can be an operation in graph computing, referring to the process by which each entity node receives information from its neighboring nodes and integrates this external information with its own information. The higher-order node feature vector set can refer to the new feature vectors possessed by each entity node after information aggregation. It is called a higher-order feature vector because the higher-order feature vector contains the original information of the entity node itself and integrates the related information of its neighborhood and even a larger range, forming a richer and more global representation of the local risk environment.

[0091] Optionally, the geological disaster early warning terminal can input the latest obtained multimodal geological disaster feature vector set into the dynamic geological disaster spatiotemporal graph model. Each feature vector in the newly obtained multimodal geological disaster feature vector set can be used to replace or update the node attributes of the corresponding entity nodes in the dynamic geological disaster spatiotemporal graph model, so that the state data of each point in the dynamic geological disaster spatiotemporal graph model can be kept up-to-date. The geological disaster early warning terminal initiates node information aggregation operation. Each entity node in the dynamic geological disaster spatiotemporal graph model can send its own state information to all its direct neighbor nodes through the connected edges. The entity node also receives state information from all neighbor nodes. After collecting the information from the neighbors, each entity node will use an aggregation function to fuse and calculate the external information with its own information, thereby generating a brand-new feature vector that can reflect the comprehensive state of itself and its surrounding environment. After completing this operation for all entity nodes, a high-order node feature vector set is obtained.

[0092] For example, aggregation functions can be summation, averaging, or constructed using neural networks.

[0093] Step 302: Input the set of high-order node feature vectors into the Bayesian neural network to perform risk probability prediction and obtain the disaster occurrence probability and uncertainty measure for each node.

[0094] Specifically, the set of high-order node feature vectors can be a set of node features rich in spatial correlation information, obtained after aggregating information from the dynamic geological disaster spatiotemporal map model. A Bayesian neural network is a special type of neural network that can output predicted values ​​and estimate the uncertainty of those values. It treats the parameters of the dynamic geological disaster spatiotemporal map model as probability distributions rather than fixed values, thus quantifying the confidence level of the prediction results. The probability of a disaster occurring is the likelihood of a geological disaster occurring in the geological unit represented by each entity node within a certain future period, calculated by the dynamic geological disaster spatiotemporal map model; it is a value between 0 and 1. The uncertainty measure is an index synchronously output by the Bayesian neural network, used to quantify the reliability or confidence level of the corresponding disaster occurrence probability prediction value. It reflects the degree of doubt about the prediction results due to data noise, model limitations, and other reasons.

[0095] Optionally, the geological disaster early warning terminal can take the set of high-order node feature vectors as input and input them into a Bayesian neural network. The Bayesian neural network performs forward propagation calculation on the high-order feature vectors of each node and outputs the probability of disaster occurrence for each node. The probability of disaster occurrence is the likelihood estimate and uncertainty measure of the occurrence of disaster at that point, which represents the fluctuation range or credibility of the probability value.

[0096] For example, a probability of 0.7 and low uncertainty indicates high risk but high certainty; a probability of 0.7 but high uncertainty indicates that the risk seems high but the judgment is not very certain.

[0097] Step 303: Based on the probability of disaster occurrence and the dynamic weights in the dynamic geological disaster spatiotemporal map model, spatial risk propagation simulation is performed to obtain the node risk probability value.

[0098] Specifically, the probability of disaster occurrence can be the initial local disaster occurrence probability predicted by a Bayesian neural network for each node. Dynamic weights can be values ​​attached to each edge in a dynamic geological disaster spatiotemporal graph model, representing the intensity of influence between entity nodes and changing in real time. Spatial risk propagation simulation can be a simulation process used to propagate and influence risks, represented in probabilistic form, along the edges of the graph based on the edge weights, between entity nodes. The node risk probability value can refer to the stable risk probability value of each entity node after spatial risk propagation simulation, taking into account the risk transmission and superposition effects from its upstream or neighboring nodes.

[0099] Optionally, the geological disaster early warning terminal can use the disaster occurrence probability of each entity node as the initial state, and the dynamic weights in the graph as the transmission coefficients for risk transmission between entity nodes. The simulation process is carried out in an iterative manner. In each iteration, each entity node will receive the risk probability from its upstream neighbor nodes proportionally according to the dynamic weights of all edges pointing to itself. At the same time, the entity node will also transmit part of its own risk to the downstream nodes. After multiple iterations, the risk probability is propagated and redistributed in the entire dynamic geological disaster spatiotemporal graph model until the risk distribution of the entire dynamic geological disaster spatiotemporal graph model reaches a stable state, and the probability value of each entity node is updated to the node risk probability value.

[0100] Step 304: Map the node risk probability value and uncertainty measure to the risk level corresponding to the risk level rule to obtain the risk level partition map.

[0101] Optionally, the node risk probability value can be the final, comprehensive probability of disaster occurrence for each entity node after spatial propagation simulation. The uncertainty measure can be the uncertainty value output by a Bayesian neural network corresponding to the initial probability of each entity node. The risk level rules can be a predefined set of standards that map continuous probability values ​​and uncertainty measures to discrete risk levels. The risk level zoning map can be a visual map product in which each geographic unit is assigned a specific risk level based on its node risk probability value and uncertainty measure, thus clearly showing the distribution of different risk levels in geographic space.

[0102] Optionally, the geological disaster early warning terminal can transform numerical predictions into intuitive visualizations and classification results. The geological disaster early warning terminal compares the node risk probability value and uncertainty measure of each entity node with predefined risk level rules. The risk level rules are judgment logics that include multiple conditions. Each entity node is classified into a certain risk level matched in the rules according to its specific probability and uncertainty values. The geological disaster early warning terminal can render a color zoning map, i.e., a risk level zoning map, by combining all entity nodes and their assigned risk levels with the geographical location information of the entity nodes. Different colors or filled areas in the risk level zoning map can represent different levels of risk.

[0103] Step 305: Extract summary features from the risk level zoning map, and obtain geological disaster risk probability features based on the summary features.

[0104] Optionally, the risk level zoning map can be a visual map showing the risk levels of different geographical areas. Summary features can be statistical or structural features extracted and calculated from the risk level zoning map that can highly summarize the overall risk situation.

[0105] Optionally, the summary features include the total area of ​​the red alert zone, the number of connected areas at the highest risk level, and the total length of the boundaries of areas at different risk levels. The geological hazard risk probability feature can be a comprehensive feature representation. The geological hazard risk probability feature can be set based on the summary features to provide a general description of the current macroscopic geological risk situation of the entire assessment area. The geological hazard risk probability feature can be one of the inputs for fusion with short-term forecast information.

[0106] Optionally, the geological disaster early warning terminal can extract macro-decision features, analyze the risk level zoning map, and calculate a series of statistical quantities or structural features from a global perspective, namely summary features. For example, the geological disaster early warning terminal calculates the area ratio of red areas, the number of patches in orange areas, and the geometric center coordinates of high-risk areas in the map. It condenses the core information of the zoning map from different dimensions, integrates and sets the summary features to form a structured data object, and obtains the geological disaster risk probability features. The geological disaster risk probability features are a set of machine-readable data that can represent the global risk situation.

[0107] This embodiment uses highly generalized and abstract geological disaster risk probability characteristics, discarding specific graphic details and retaining the most critical macro-risk information. This allows the macro-risk information to serve as an effective input, enabling efficient fusion with signals from other modalities such as short-term meteorological forecasts at the feature level. This provides information-intensive risk representation for early warning decisions. The risk level zoning map can greatly improve the readability and operability of risk information, facilitating rapid regional risk assessment and emergency resource pre-positioning.

[0108] In one embodiment, short-term forecast information, geological hazard risk probability characteristics, and uncertainty measures are fused to obtain comprehensive risk characteristics, including:

[0109] Step 401: Based on the type of short-term forecast information, encode the short-term forecast information to obtain an independent short-term disaster-causing signal vector and a short-term uncertainty measure.

[0110] Optionally, short-term forecast information can refer to forecasts or warnings issued by meteorological, seismological, and other departments, targeting sudden disaster-causing events that may trigger or exacerbate geological disasters in the near future.

[0111] Optionally, common types of geological disaster risk probability characteristics include refined rainstorm forecasts, severe convective weather warnings, and earthquake rapid reporting information. The type of short-term forecast information can refer to the specific category of forecast information; different types of information typically have different data structures and physical meanings. Independent short-term disaster-causing signal vectors can refer to machine-readable numerical vectors obtained after independently encoding each type of short-term forecast information. Each independent short-term disaster-causing signal vector represents the intensity, spatial location, and other information of the disaster-causing signal carried by that type of forecast. Short-term uncertainty measures can refer to numerical values ​​generated synchronously during the encoding process to quantify the reliability of the corresponding short-term forecast signal itself. Independent short-term disaster-causing signals reflect the uncertainty of the forecast itself, such as the error range of rainfall forecasts or the accuracy of earthquake location.

[0112] Optionally, the geological disaster early warning terminal can perform standardized preprocessing on externally input short-term forecast information of various formats. The geological disaster early warning terminal identifies the type of each input short-term forecast information, and calls the corresponding encoder according to the type of short-term forecast information to convert the short-term forecast information into a fixed-dimensional numerical vector, that is, an independent short-term disaster-causing signal vector. The encoding process can estimate the numerical value that characterizes its reliability based on the original attributes of the forecast information, that is, the short-term uncertainty measure. Each type of forecast will independently generate a pair (signal vector, uncertainty measure).

[0113] Step 402: Perform a linear transformation on the probability characteristics and uncertainty measure of geological disaster risk to obtain the query vector, and perform a linear projection on the independent short-term disaster-causing signal vector to obtain the key vector and value vector.

[0114] Specifically, the probability characteristics of geological disaster risk can be a comprehensive feature vector characterizing the regional geological background risk, obtained from long-term monitoring and analysis of dynamic geological disaster spatiotemporal map models. Uncertainty measures can be indicators that quantify the reliability of geological background risk predictions, accompanying the probability characteristics of geological disaster risk. Linear transformation and linear projection are basic mathematical operations. They can use a learnable weight matrix to perform matrix multiplication on the input vector, mapping it to another vector space. The purpose is to extract or transform features to adapt to subsequent calculations. In the attention mechanism, the query vector represents the feature representation of the initiating party. The query vector is transformed from the geological background risk characteristics and is used to inquire which information in the short-term signals is relevant to itself. In the attention mechanism, the key vector represents the feature representation of the retrieved party. The key vector is transformed from the short-term disaster-causing signal vector and is used for matching calculations with the query vector to measure relevance. In the attention mechanism, the value vector represents the feature representation of the actual information carried. The value vector is transformed from the short-term disaster-causing signal vector and represents the information itself to be aggregated.

[0115] Optionally, the geological disaster early warning terminal can splice or combine the probability characteristics of geological disaster risks and their uncertainty measures that represent geological background risks, and process them through a linear transformation layer to generate a query vector representing the current geological risk status. The geological disaster early warning terminal performs two different linear projections on each independent short-term disaster-causing signal vector, thereby generating a corresponding key vector and a value vector for each short-term signal. The key vector can be used to calculate the correlation with the query vector, and the value vector can carry the detailed information of the forecast signal.

[0116] Step 403: Calculate the relevance score between the query vector and each key vector to obtain the initial relevance score, and adjust the initial relevance score based on the short-term uncertainty measure to obtain the attention weight distribution list.

[0117] Specifically, the relevance score can be a scalar value. It measures the degree of association or similarity between a query vector and a key vector. A higher relevance score indicates a closer match and greater relevance between the short-term forecast signal and the current geological risk background. The initial relevance score can refer to the raw relevance score calculated directly through mathematical operations before considering uncertainty adjustments. The attention weight distribution list can be a list where each element is a weight value between 0 and 1, corresponding to a short-term forecast signal. The sum of all weights in the attention weight distribution list is 1, representing the proportion of importance each short-term forecast signal should possess during fusion.

[0118] Optionally, the geological disaster early warning terminal can calculate the dot product between the query vector and the key vector of each short-term forecast signal to obtain a series of initial correlation scores. The correlation scores reflect the matching degree between the signal and the background without considering the forecast reliability. The geological disaster early warning terminal can introduce uncertainty correction. The short-term uncertainty measure of each short-term forecast signal is used as a penalty term. The geological disaster early warning terminal can multiply the uncertainty measure of the signal by a preset uncertainty penalty coefficient, and subtract this product from the initial correlation score of the signal to obtain the corrected score. The higher the uncertainty, the more points are deducted, thereby reducing its influence. The geological disaster early warning terminal can apply the Softmax function to the corrected scores of all short-term signals to perform normalization exponential operation, converting them into a probability distribution, i.e., an attention weight distribution list. Each weight value in the list combines the correlation of the signal and its own reliability.

[0119] Step 404: Based on the attention weight distribution list, perform a weighted summation of all value vectors to obtain the short-term context vector.

[0120] Optionally, weighted summation can be a mathematical operation. In this embodiment, weighted summation specifically refers to linearly combining the value vectors using attention weights, where each vector is first multiplied by its corresponding weight, and then all the resulting vectors are summed. The short-term signal context vector can be a comprehensive feature vector obtained by weighted summation of the value vectors of all short-term signals. The short-term signal context vector condenses the key information from all short-term forecast signals after weighted filtering, and represents the overall impact of short-term disaster-causing factors in the current context.

[0121] Optionally, the geological disaster early warning terminal can multiply the value vector of each short-term signal by its corresponding attention weight, add all the weighted vectors together, and synthesize a new vector, namely the short-term signal context vector.

[0122] Step 405: Integrate the short-term signal context vector and the geological hazard risk probability features to obtain comprehensive risk features.

[0123] Optionally, fusion can refer to the process of combining two or more feature vectors into a new, more representative feature vector through mathematical operations. Comprehensive risk features can be a comprehensive feature representation used for early warning decision-making, simultaneously including long-term risk probability features from the geological background and short-term disaster-causing signal context from meteorological and other departments. The short-term disaster-causing signal context is the most comprehensive quantitative description of the current overall disaster risk situation.

[0124] Optionally, the geological disaster early warning terminal can combine the short-term signal context vector with the geological disaster risk probability features. The combination can be vector concatenation or nonlinear combination by feeding it into a fully connected neural network layer. The geological disaster early warning terminal superimposes the short-term urgent disaster-causing signal onto the long-term geological risk base to generate a unified comprehensive risk feature.

[0125] This embodiment integrates risk information from different time scales, enabling early warning decisions based on this feature to achieve scenario-coupled risk assessment, which can greatly improve the accuracy and timeliness of early warnings.

[0126] In one embodiment, based on a dynamic cost matrix and comprehensive risk characteristics, an early warning decision is made to obtain a geological disaster early warning instruction, including:

[0127] Step 501: Based on the comprehensive risk characteristics, perform scenario analysis to obtain scenario characteristics, and adjust the basic cost variables based on the scenario characteristics to obtain the dynamic cost matrix.

[0128] Optionally, the comprehensive risk characteristics can be a comprehensive feature vector that integrates long-term geological background risks and short-term disaster-causing signals, comprehensively representing the overall risk status of the current region. Context analysis can be the process of analyzing and understanding the specific environment and conditions in which the current decision is made. In this embodiment, context analysis specifically refers to interpreting specific contextual factors that may affect the cost of the decision from the comprehensive risk characteristics. Contextual characteristics can be specific, quantifiable environmental parameters obtained through context analysis, and they can describe the specific background at the time of decision-making.

[0129] Optionally, scenario characteristics include population density, land type, and current time in the risk area. The basic cost variables can be predefined benchmark values ​​for two core decision costs: false alarm cost (the losses such as evacuation costs, social panic, and decreased public trust resulting from issuing a warning but the disaster not occurring) and underreporting cost (the consequences of casualties and property damage resulting from a disaster occurring but no warning being issued). False alarm and underreporting costs are the initial parameters for cost calculation. The dynamic cost matrix can be a mathematical structure used to quantify different decision consequences. It can contain all the key parameters needed to calculate the expected cost of a decision in the current specific scenario, most importantly the actual false alarm and underreporting costs after scenario adjustment. The dynamic cost matrix is ​​dynamic, meaning that the cost values ​​within it will adjust as scenario characteristics change.

[0130] Optionally, the geological disaster early warning terminal can perform scenario analysis on comprehensive risk characteristics. Through an analysis module, it can extract key environmental parameters affecting costs and generate scenario features. For example, the geological disaster early warning terminal analyzes that the current time is nighttime and the risk area type is a residential area. Based on the scenario features, it adjusts the pre-set basic cost variables. In the nighttime and residential area scenarios, missed reports may cause greater casualties, so the cost of missed reports will be increased. At the same time, false reports may cause more serious sleep disruption and confusion, so the cost of false reports will also be appropriately increased. Through a set of predefined adjustment rules or functions, the basic cost variables are corrected, thereby forming a dynamic cost matrix that fits the current actual situation.

[0131] Step 502: Calculate the expected cost of the early warning action based on the dynamic cost matrix; the elements in the dynamic cost matrix include the predicted probability of geological disaster occurrence, the cost of false alarm, and the cost of missed alarm.

[0132] The formula for calculating the expected cost is as follows:

[0133]

[0134]

[0135] Optionally, The expected cost of issuing an early warning, The expected cost of not issuing a warning, The predicted probability of geological disasters. To pay the price for false alarms The cost of underreporting.

[0136] Specifically, the dynamic cost matrix can be a matrix containing parameters such as the false alarm cost and the missed alarm cost after scenario adjustment. Early warning actions can refer to two basic decision-making behaviors: issuing an early warning and not issuing an early warning. Expected cost can be the average cost that might be incurred in making a decision under uncertainty conditions; it is an expected value calculated after comprehensively considering the probability of disaster occurrence and the corresponding consequences of the decision. The predicted probability of geological disaster occurrence can be an estimate of the probability of a geological disaster about to occur in the currently assessed area, extracted from comprehensive risk characteristics. The false alarm cost can be the scenario-adjusted cost in the dynamic cost matrix corresponding to the issuance of an early warning but the disaster not occurring. The missed alarm cost can be the scenario-adjusted cost in the dynamic cost matrix corresponding to the occurrence of a disaster without an early warning.

[0137] Optionally, the geological disaster early warning terminal can extract the predicted probability of geological disaster occurrence from comprehensive risk characteristics, and obtain the adjusted false alarm cost and missed alarm cost from the dynamic cost matrix. The geological disaster early warning terminal can perform calculations using the provided expected cost calculation formula. If an early warning is issued, a cost will only be incurred if the disaster does not actually occur. If no early warning is issued, a cost will only be incurred if the disaster actually occurs.

[0138] Step 503: Based on the expected cost, perform a binary decision-making process for early warning, obtain a binary decision-making instruction, and extract a comprehensive uncertainty measure from the comprehensive risk characteristics.

[0139] Specifically, the expected cost can be the expected cost of issuing an early warning and the expected cost of not issuing an early warning. The binary decision-making process for early warning can be a decision-making process that chooses between the two basic options of issuing an early warning and not issuing one. The binary decision instruction can be the result of the binary decision-making process for early warning; it can be an explicit Boolean instruction, and its value can be either issuing an early warning or not issuing one. The comprehensive uncertainty measure can be an indicator extracted from comprehensive risk characteristics that quantifies the uncertainty of the current overall risk assessment. The comprehensive uncertainty measure can reflect the degree of confidence that the dynamic geological disaster spatiotemporal map model has in the currently predicted risk probability.

[0140] Optionally, the geological disaster early warning terminal can directly compare two expected costs. If the expected cost of issuing an early warning is less than the expected cost of not issuing an early warning, then it decides to issue an early warning; otherwise, it decides not to issue an early warning. The comparison result is output as a clear binary decision instruction. At the same time, the geological disaster early warning terminal can extract a comprehensive uncertainty measure from the comprehensive risk characteristics in parallel. The comprehensive uncertainty measure value does not participate in this binary comparison, but provides key input for the next level decision.

[0141] Step 504: Based on the binary decision instruction and the comprehensive uncertainty measure, make a decision on the early warning level to obtain the geological disaster early warning instruction.

[0142] Specifically, a binary decision instruction can be a basic instruction indicating whether to issue an early warning. A comprehensive uncertainty measure can be a numerical value extracted from comprehensive risk characteristics that characterizes the overall predictive uncertainty. An early warning level decision can be a decision-making process that further determines the early warning level, given that an early warning has already been decided upon; different early warning levels correspond to different emergency response intensities. A geological disaster early warning instruction can be a complete, directly executable instruction, including whether to issue an early warning and the level of warning.

[0143] Optionally, the geological disaster early warning terminal can check the binary decision instruction. If the instruction is not to issue an early warning, the terminal will directly generate an instruction not to issue an early warning. If the instruction is to issue an early warning, the terminal will enter the early warning level decision sub-process. In the sub-process, the geological disaster early warning terminal can combine the comprehensive uncertainty measure extracted from the comprehensive risk characteristics and other relevant risk intensity information to determine the specific level according to a set of predefined level classification rules.

[0144] For example, the decision-making rules could be: if the risk probability is extremely high and the uncertainty is low, then a red alert is issued; if the risk probability is high but the uncertainty is moderate, then an orange alert is issued; if the risk probability is relatively high but the uncertainty is high, then a yellow alert is issued, generating a geological disaster warning instruction containing specific level information.

[0145] This embodiment uses a two-level decision-making structure to ensure that the early warning instructions are both economically reasonable and operationally accurate, forming a complete, thorough, and interpretable automated early warning decision-making closed loop.

[0146] In one embodiment, the attention weights in the attention weight distribution list are calculated using the following formula:

[0147]

[0148] in, For attention weights, The corrected score for the i-th short-term forecast signal is given by [the following]. For temperature parameters, Let be the corrected score of the j-th short-term forecast signal, where i and j are the indices of the short-term forecast signal.

[0149] The formula for calculating the corrected score of the short-term forecast signal is as follows:

[0150]

[0151] in, Let Q be the corrected score for the i-th short-term forecast signal, and let Q be the query vector. Let be the key vector of the i-th short-term forecast signal. Let be the dimension of the key vector. The uncertainty penalty coefficient, Let be the short-term uncertainty measure of the i-th short-term forecast signal.

[0152] Specifically, attention weight can be the weight value assigned to the i-th short-term forecast signal in the attention mechanism. Attention weight is a scalar between 0 and 1, and the sum of the attention weights of all signals is 1. The attention weight directly determines the proportion of that signal in the final fusion result. The corrected score is the correlation score between the i-th short-term forecast signal and the current geological risk background after considering its own uncertainty penalty. The corrected score is the direct input for calculating the attention weight. The temperature parameter can be an adjustable hyperparameter, used to control the sharpness or smoothness of the attention weight distribution. The smaller the temperature parameter value, the sharper the weight distribution, with higher scores receiving higher weights and lower scores receiving lower weights; the larger the temperature parameter value, the smoother the weight distribution and the more even the weight allocation. The query vector can be a feature vector converted from the probability characteristics of geological disaster risk, representing a query about the current geological background risk status. The query vector can be used to match with various short-term signals to find relevant information. The key vector can be a feature vector converted from the i-th short-term forecast signal, representing the index or keyword of that signal. The key vector can be used to calculate the similarity with the query vector. The dimension of the key vector can be its length, i.e., the number of elements it contains. This dimension can be used to scale the dot product result to prevent it from becoming too large, thus stabilizing the training process. The uncertainty penalty coefficient can be a preset coefficient. It adjusts the severity of the penalty imposed on the influence of the uncertainty of the short-term forecast signal itself. A larger uncertainty penalty coefficient results in a greater penalty for signals with high uncertainty. The short-term uncertainty metric can be a numerical value characterizing the reliability of the i-th short-term forecast signal. A larger short-term uncertainty metric value indicates higher uncertainty and lower reliability of the forecast signal.

[0153] This embodiment introduces uncertainty penalty, dynamically assigning importance weights to each short-term forecast signal based on the correlation between the signal and the current geological risk and the reliability of the signal itself. Signals with high correlation and high certainty will receive the highest attention weight and contribute the most in subsequent fusion; while the weights of signals with low correlation or high uncertainty will be reduced, significantly improving the robustness and scientific nature of multi-source information fusion, and making decision-making more dependent on credible and relevant evidence.

[0154] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0155] Based on the same inventive concept, this application also provides an AI-based geological disaster early warning system for implementing the AI-based geological disaster early warning method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more AI-based geological disaster early warning system embodiments provided below can be found in the limitations of the AI-based geological disaster early warning method described above, and will not be repeated here.

[0156] In one exemplary embodiment, such as Figure 2 As shown, an artificial intelligence-based geological disaster early warning system 700 is provided, comprising:

[0157] Feature module 701 is used to acquire raw data of multi-source geological disaster monitoring and input the raw data of multi-source geological disaster monitoring into the geological disaster feature extraction model for parallel feature extraction to obtain a set of multi-modal geological disaster feature vectors;

[0158] The prediction module 702 is used to predict the risk probability based on a multimodal geological hazard feature vector set using a dynamic geological hazard spatiotemporal map model, and to generate geological hazard risk probability characteristics and uncertainty measures.

[0159] The fusion module 703 is used to perform feature fusion on short-term forecast information, geological hazard risk probability characteristics and uncertainty measures to obtain comprehensive risk characteristics;

[0160] The decision module 704 is used to make early warning decisions based on the dynamic cost matrix and comprehensive risk characteristics, and generate geological disaster early warning instructions; wherein, the geological disaster early warning instructions are used to instruct the forecasting and early warning actions.

[0161] Furthermore, the device also includes a model module for:

[0162] Using each monitoring unit as a node and the geological hazard feature vector corresponding to the monitoring unit as the node attribute, entity nodes are constructed, and the relationships between entity nodes are extracted to obtain the edges of the geological hazard spatiotemporal map model; among which, the relationships include spatial adjacency relationships, geological structure relationships, and hydrological path relationships;

[0163] The node attributes are input into the dynamic weight model to calculate the dynamic weights of the edges between entity nodes.

[0164] Based on entity nodes, edges, and dynamic weights, a primary spatiotemporal map model of geological hazards is constructed.

[0165] Based on weight thresholds and dynamic weights, redundant edges are pruned from the primary geological hazard map model to obtain a dynamic geological hazard spatiotemporal map model.

[0166] Furthermore, the prediction module 702 is also used for:

[0167] Based on the multimodal geological hazard feature vector set, the node attributes of the dynamic geological hazard spatiotemporal map model are updated, and the node information of the dynamic geological hazard spatiotemporal map model is aggregated to obtain a high-order node feature vector set;

[0168] The set of feature vectors of high-order nodes is input into a Bayesian neural network to perform risk probability prediction, and obtain the disaster occurrence probability and uncertainty measure for each node.

[0169] Based on the probability of disaster occurrence and the dynamic weights in the dynamic geological disaster spatiotemporal map model, spatial risk propagation simulation is performed to obtain the node risk probability value;

[0170] By mapping the node risk probability value and uncertainty measure to the risk level corresponding to the risk level rule, a risk level partitioning map is obtained;

[0171] Summary features are extracted from the risk level zoning map, and geological hazard risk probability features are obtained based on the summary features.

[0172] Furthermore, the fusion module 703 is also used for:

[0173] Based on the type of short-term forecast information, the short-term forecast information is encoded to obtain independent short-term disaster-causing signal vectors and short-term uncertainty measures;

[0174] A linear transformation is performed on the probability characteristics and uncertainty measure of geological disaster risk to obtain a query vector. A linear projection is performed on the independent short-term disaster-causing signal vector to obtain a key vector and a value vector.

[0175] Calculate the relevance score between the query vector and each key vector to obtain the initial relevance score, and adjust the initial relevance score based on the short-term uncertainty measure to obtain the attention weight distribution list;

[0176] Based on the attention weight distribution list, a weighted sum of all value vectors is performed to obtain the short-term neighbor signal context vector;

[0177] By integrating the context vector of short-term signals and the probability characteristics of geological disaster risks, a comprehensive risk characteristic is obtained.

[0178] Furthermore, decision module 704 is also used for:

[0179] Based on comprehensive risk characteristics, scenario analysis is performed to obtain scenario features, and based on these scenario features, the basic cost variables are adjusted to obtain a dynamic cost matrix.

[0180] Based on the dynamic cost matrix, the expected cost of early warning actions is calculated; the elements in the dynamic cost matrix include the predicted probability of geological disasters, the cost of false alarms, and the cost of missed alarms.

[0181] The formula for calculating the expected cost is as follows:

[0182]

[0183]

[0184] in, The expected cost of issuing an early warning, The expected cost of not issuing a warning, The predicted probability of geological disasters. To pay the price for false alarms The cost of underreporting;

[0185] Based on the expected cost, a binary decision-making process for early warning is conducted to obtain binary decision instructions, and a comprehensive uncertainty measure is extracted from the comprehensive risk characteristics.

[0186] Based on binary decision instructions and comprehensive uncertainty measurement, early warning level decisions are made to obtain geological disaster early warning instructions.

[0187] Furthermore, the attention weights in the attention weight distribution list are calculated using the following formula:

[0188]

[0189] in, For attention weights, The corrected score for the i-th short-term forecast signal is given by [the following]. For temperature parameters, Let i be the corrected score of the j-th short-term forecast signal, where i and j are the indices of the short-term forecast signal.

[0190] The formula for calculating the corrected score of the short-term forecast signal is as follows:

[0191]

[0192] in, Let Q be the corrected score for the i-th short-term forecast signal, and let Q be the query vector. Let be the key vector of the i-th short-term forecast signal. Let be the dimension of the key vector. The uncertainty penalty coefficient, Let be the short-term uncertainty measure of the i-th short-term forecast signal.

[0193] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the artificial intelligence-based geological disaster early warning method as described above.

[0194] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0195] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts 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 disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0196] The embodiments described above are merely examples of several implementation methods of the embodiments of this application. While the descriptions are specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A geological disaster early warning method based on artificial intelligence, characterized in that, The method includes: Obtain raw data from multi-source geological hazard monitoring, and input the raw data into a geological hazard feature extraction model for parallel feature extraction to obtain a set of multimodal geological hazard feature vectors; Using a dynamic spatiotemporal geological disaster map model, and based on the multimodal geological disaster feature vector set, risk probability prediction is performed, generating geological disaster risk probability characteristics and uncertainty measures. A comprehensive risk characteristic is obtained by fusing short-term forecast information, the geological hazard risk probability characteristics, and the uncertainty measure. Based on the dynamic cost matrix and the comprehensive risk characteristics, early warning decisions are made, and geological disaster early warning instructions are generated; wherein, the geological disaster early warning instructions are used to instruct forecasting and early warning actions.

2. The method according to claim 1, characterized in that, The method further includes: Using each monitoring unit as a node and the geological hazard feature vector corresponding to the monitoring unit as the node attribute, entity nodes are constructed, and the relationships between the entity nodes are extracted to obtain the edges of the geological hazard spatiotemporal map model; wherein, the relationships include spatial adjacency relationships, geological structure relationships, and hydrological path relationships; The node attributes are input into the dynamic weight model to calculate the dynamic weight of the edges between the entity nodes; Based on the entity nodes, the edges, and the dynamic weights, a primary spatiotemporal map model of geological hazards is constructed. Based on the weight threshold and the dynamic weight, redundant edges are pruned on the primary geological hazard map model to obtain the dynamic geological hazard spatiotemporal map model.

3. The method according to claim 2, characterized in that, The process involves using a dynamic spatiotemporal geological hazard model, based on the multimodal geological hazard feature vector set, to predict risk probability and generate geological hazard risk probability features and uncertainty measures, including: Based on the multimodal geological hazard feature vector set, the node attributes of the dynamic geological hazard spatiotemporal map model are updated, and the node information of the dynamic geological hazard spatiotemporal map model is aggregated to obtain a high-order node feature vector set. The set of high-order node feature vectors is input into a Bayesian neural network to perform risk probability prediction, thereby obtaining the disaster occurrence probability and uncertainty measure for each node. Based on the probability of disaster occurrence and the dynamic weights in the dynamic geological disaster spatiotemporal map model, spatial risk propagation simulation is performed to obtain node risk probability values; The node risk probability value and the uncertainty measure are mapped to the risk level corresponding to the risk level rule to obtain the risk level partitioning map; The summary features are extracted from the risk level zoning map, and the geological disaster risk probability features are obtained based on the summary features.

4. The method according to claim 1, characterized in that, The feature fusion of short-term forecast information, geological hazard risk probability characteristics, and uncertainty measures yields comprehensive risk characteristics, including: Based on the type of the short-term forecast information, the short-term forecast information is encoded to obtain an independent short-term disaster-causing signal vector and a short-term uncertainty measure; A query vector is obtained by performing a linear transformation on the geological disaster risk probability characteristics and the uncertainty measure, and a key vector and a value vector are obtained by performing a linear projection on the independent short-term disaster-causing signal vector. Calculate the relevance score between the query vector and each of the key vectors to obtain an initial relevance score, and adjust the initial relevance score based on the short-term uncertainty measure to obtain an attention weight distribution list; Based on the attention weight distribution list, all the value vectors are weighted and summed to obtain the short-term context vector; The comprehensive risk characteristics are obtained by fusing the short-term signal context vector and the geological hazard risk probability characteristics.

5. The method according to claim 4, characterized in that, The process of making early warning decisions based on the dynamic cost matrix and the comprehensive risk characteristics to obtain geological disaster early warning instructions includes: Based on the comprehensive risk characteristics, scenario analysis is performed to obtain scenario features, and based on the scenario features, the basic cost variables are adjusted to obtain the dynamic cost matrix; Based on the dynamic cost matrix, the expected cost of the early warning action is calculated; the elements in the dynamic cost matrix include the predicted probability of geological disaster occurrence, the cost of false alarm, and the cost of missed alarm. The formula for calculating the expected cost is as follows: in, The expected cost of issuing an early warning, The expected cost of not issuing a warning, The predicted probability of the occurrence of the geological disaster. The cost of the false alarm, The cost of the aforementioned underreporting; Based on the expected cost, a binary decision-making process for early warning is performed to obtain a binary decision instruction, and a comprehensive uncertainty measure is extracted from the comprehensive risk characteristics. Based on the binary decision instruction and the comprehensive uncertainty measure, a warning level decision is made to obtain the geological disaster warning instruction.

6. The method according to claim 4, characterized in that, The attention weights in the attention weight distribution list are calculated using the following formula: in, For attention weights, The corrected score for the i-th short-term forecast signal is given by [the following]. For temperature parameters, Let i be the corrected score of the j-th short-term forecast signal, where i and j are the indices of the short-term forecast signal. The formula for calculating the corrected score of the short-term forecast signal is as follows: in, Let Q be the corrected score of the i-th short-term forecast signal, and let Q be the query vector. Let the key vector be the i-th short-term forecast signal. Let be the dimension of the key vector. The uncertainty penalty coefficient, Let be the short-term uncertainty measure of the i-th short-term forecast signal.

7. A geological disaster early warning system based on artificial intelligence, characterized in that, The system includes: The feature module is used to acquire raw data of multi-source geological disaster monitoring and input the raw data of multi-source geological disaster monitoring into the geological disaster feature extraction model for parallel feature extraction to obtain a set of multi-modal geological disaster feature vectors; The prediction module is used to predict the risk probability based on the multimodal geological hazard feature vector set using a dynamic geological hazard spatiotemporal map model, and to generate geological hazard risk probability features and uncertainty measures. The fusion module is used to perform feature fusion on short-term forecast information, the geological hazard risk probability characteristics, and the uncertainty measure to obtain comprehensive risk characteristics; The decision-making module is used to make early warning decisions based on the dynamic cost matrix and the comprehensive risk characteristics, and generate geological disaster early warning instructions; wherein, the geological disaster early warning instructions are used to instruct forecast and early warning actions.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images