Early warning method, system and equipment for ground disaster, storage medium and program product

Abstract

The invention discloses a ground disaster early warning method, system and device, a storage medium and a program product, and belongs to the technical field of geological disaster early warning, the method comprises the steps that multi-modal time sequence data of a target monitoring area is acquired in real time and preprocessed, and the multi-modal time sequence data comprises ground disaster environment data, image data and meteorological data; the preprocessed multi-modal time series data are input into a pre-trained ground disaster prediction model, the ground disaster prediction model carries out multi-modal feature fusion on the multi-modal time series data and dynamically adjusts the space-time weight, and the disaster risk probability in the target monitoring area is output; and triggering a corresponding disaster early warning or checking action according to the disaster risk probability. According to the invention, the occurrence probability of the geological disaster can be accurately predicted, and accurate prediction of the geological disaster is realized.

Description

Technical Field

[0001] This application belongs to the field of geological disaster early warning technology, and specifically relates to a geological disaster early warning method, system, equipment, storage medium and program product. Background Technology

[0002] In recent years, with climate change, geological disasters have occurred frequently. Therefore, improving the scientific, information-based, standardized and visualized nature of geological disaster monitoring and early warning work, providing data support for disaster reduction and prevention decisions, and ensuring the quality, efficiency and management of monitoring and early warning work have become key tasks in geological disaster monitoring and early warning.

[0003] However, existing geological disaster early warning methods have the following limitations: (1) Static threshold alarms that rely on a single monitoring indicator have a high false alarm and false negative rate due to insufficient data dimensions; (2) Although traditional machine learning models can learn historical patterns, they are static and cannot adapt to the dynamic evolution of the geological environment, resulting in a decrease in prediction accuracy over time.

[0004] Therefore, there is an urgent need for a disaster early warning method that can deeply integrate multi-source data and dynamically capture the spatiotemporal evolution of disasters, in order to overcome the shortcomings of existing solutions in terms of accuracy and adaptability. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a geological disaster early warning method, system, device, storage medium, and program product that can accurately predict the probability of geological disasters occurring, thereby achieving precise prediction of geological disasters.

[0006] To achieve the above objectives, this application provides the following technical solution: In a first aspect, embodiments of this application provide a geological disaster early warning method, the method comprising: Real-time acquisition and preprocessing of multimodal time-series data of the target monitoring area, including geological disaster environmental data, image data and meteorological data; The preprocessed multimodal time series data is input into a pre-trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk within the target monitoring area. Based on the stated probability of disaster risk, corresponding disaster warnings or verification actions are triggered.

[0007] Furthermore, the real-time acquisition and preprocessing of multimodal time-series data of the target monitoring area includes: Multiple geological disaster monitoring devices are deployed within the target monitoring area, and geological disaster environmental data of the target monitoring area are acquired in real time through these devices. Simultaneously, image data of the target monitoring area provided by the remote sensing platform is acquired in real time, and meteorological data of the target monitoring area is acquired in real time through the meteorological platform. The spatiotemporal difference algorithm is used to unify geological disaster environmental data, image data and meteorological data with different spatiotemporal resolutions into a monitoring grid of a preset size for timestamp alignment, so as to obtain time-aligned multimodal time series data. Normalization and data cleaning are performed on the time-aligned multimodal time series data to obtain preprocessed multimodal time series data.

[0008] Furthermore, the geological disaster monitoring equipment includes a GNSS receiver, a rain gauge, a crack gauge, an inclination accelerometer, a moisture meter, and a mud level gauge; the geological disaster environmental data includes surface displacement, rainfall, crack changes, rock strata tilt angle, soil moisture content, and groundwater level data. The spatial resolution of the image data is not greater than a first predetermined value; The meteorological data includes soil moisture and wind speed, and the time resolution of the meteorological data is a predetermined time.

[0009] Furthermore, the method also includes constructing and training the geological disaster prediction model, including: Based on different monitoring areas and their historical multimodal time series data, a spatiotemporal map corresponding to each monitoring area is constructed, and a multimodal attention network model is constructed according to the structure of each spatiotemporal map; The multimodal attention network model is trained using the training set to obtain the trained multimodal attention network model; By employing knowledge distillation techniques to deploy a lightweight multimodal attention network model after training, a geological disaster prediction model for the target monitoring area can be obtained.

[0010] Furthermore, the process of constructing a spatiotemporal map corresponding to each monitoring area based on different monitoring areas and their historical multimodal time-series data, and constructing a multimodal attention network model according to the structure of each spatiotemporal map, includes: Each monitoring area is divided into several grid units, each grid unit serves as a node in the corresponding spatiotemporal map, and each node is assigned static attributes and dynamic temporal characteristics; Based on the physical mechanism model and data-driven method, spatiotemporal edges are constructed and connected between nodes to obtain corresponding physical mechanism edges and data-driven edges, so as to form a spatiotemporal map of each monitoring area; A multimodal attention network model is constructed based on the nodes, physical mechanism edges, and data-driven edges in the spatiotemporal graphs of different monitoring areas. The static attributes include elevation, slope, soil type, and vegetation coverage. The dynamic time series features are obtained through preprocessed historical multimodal time series data, including surface displacement rate, groundwater level and cumulative rainfall over a past period.

[0011] Furthermore, during training, a hybrid loss function is used to optimize the multimodal attention network model; The hybrid loss function is defined using the focal loss and Kl divergence constraint model as follows: , In the formula, L represents the mixture loss function. and This represents the preset loss coefficient. Represents the focus loss function. and KLdiv represents the predicted values ​​and ground truth labels of a multimodal attention network model, and KLdiv represents the KL divergence constraint model, used to constrain the probability distribution of the predictions. Probability distribution derived from physical simulation model Maintain consistency.

[0012] Furthermore, the lightweight deployment of the trained multimodal attention network model using knowledge distillation technology to obtain a geological disaster prediction model for the target monitoring area includes: The trained multimodal attention network model is deployed on a central server and used as a general teacher model; For the target monitoring area to be deployed, a dedicated spatiotemporal map is constructed based on the multimodal time series data of the target monitoring area; A dedicated student model is constructed based on a dedicated spatiotemporal graph, and the knowledge of the teacher model is transferred to the lightweight dedicated student model through knowledge distillation technology. The dedicated student model is trained to obtain the trained dedicated student model, that is, the geological disaster prediction model; the geological disaster prediction model is deployed to the local edge computing device corresponding to the target monitoring area to perform real-time geological disaster prediction inference in the target monitoring area.

[0013] Furthermore, the disaster prediction model fuses multimodal features from multimodal time-series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk within the target monitoring area, including: The geological disaster prediction model extracts multimodal features from the input geological disaster environmental data, image data, and meteorological data to obtain corresponding geological disaster time series features, image time series features, and meteorological time series features. Cross-modal attention fusion is performed on geological disaster time-series features, image time-series features, and meteorological time-series features to obtain multimodal fused features; Based on the multimodal fusion features, information is propagated through a graph attention network on the dedicated spatiotemporal map corresponding to the target monitoring area. The spatial dependencies between nodes are dynamically learned, and the dynamic temporal features in the node features are analyzed to capture the temporal evolution law, thereby obtaining the predicted attention coefficient and temporal sensitive features. The time-sensitive features are input into the output layer of the geological disaster prediction model, and the disaster risk probability is calculated by combining the prediction attention coefficient and the preset hybrid loss function.

[0014] Furthermore, the cross-modal attention fusion of geological disaster time-series features, image time-series features, and meteorological time-series features to obtain multimodal fusion features includes: Using a cross-modal attention fusion mechanism, the correlation weight between any two modal features is calculated as follows: , In the formula, This represents the correlation weight between the i-th modal feature and the j-th modal feature. This represents the query vector for the i-th modal feature. The key vector representing the j-th modal feature. Indicates the dimension scaling factor. Represents the transpose of a matrix; The correlation weights are used to fuse all modal features to obtain a unified multimodal fusion feature: , In the formula, Indicates multimodal fusion features, This represents the total number of modal features. , , ; This represents the value vector of the j-th modal feature.

[0015] Furthermore, based on the aforementioned multimodal fusion features, information propagation is performed on the dedicated spatiotemporal map corresponding to the target monitoring area using a graph attention network. This dynamically learns the spatial dependencies between nodes and analyzes the dynamic temporal features within the node characteristics to capture temporal evolution patterns, thereby obtaining predicted attention coefficients and temporally sensitive features, including: The multimodal fusion features are injected into the dedicated spatiotemporal map corresponding to the target monitoring area; Spatiotemporal graph attention propagation is performed using graph attention networks. Predicted attention coefficients between nodes and their neighbors in a specific spatiotemporal graph are dynamically calculated, and information about the neighbors is aggregated based on these predicted attention coefficients to update node features. Based on the updated node features, a node feature sequence is obtained over multiple consecutive time steps, and the node feature sequence is input into a time-series decoder to generate time-sensitive features that reflect the time-series change patterns. The predicted attention coefficient is calculated using the following formula: , In the formula, This represents the predicted attention coefficient from the target node u to its neighbor node k, i.e., the predicted attention coefficient of the edge uv; u represents the target node, and k represents a neighbor node directly connected to the target node u. Represents the set of all neighboring nodes of node u; and Let represent the feature vectors of node u and its neighbor node k, respectively; || denotes the vector concatenation operation; and W represents the learnable weight matrix. This represents a learnable attention vector. This represents the activation function. This represents an exponential function.

[0016] Secondly, embodiments of this application provide a geological disaster early warning system, the system comprising: The real-time acquisition unit is used to acquire and preprocess multimodal time-series data of the target monitoring area in real time. The multimodal time-series data includes geological disaster environmental data, image data, and meteorological data. The prediction unit is used to input the preprocessed multimodal time series data into the trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk in the target monitoring area. The early warning unit is used to trigger corresponding disaster early warning or verification actions based on the disaster risk probability.

[0017] Furthermore, the system also includes a construction and training unit; The construction and training unit is used to construct and train the geological disaster prediction model; Specifically, the construction and training unit is used for: Based on different monitoring areas and their historical multimodal time series data, a spatiotemporal map corresponding to each monitoring area is constructed, and the multimodal attention network model is constructed according to the structure of each spatiotemporal map; The multimodal attention network model is trained using the training set to obtain the trained multimodal attention network model; By employing knowledge distillation techniques to deploy a lightweight multimodal attention network model after training, a geological disaster prediction model for the target monitoring area can be obtained.

[0018] Thirdly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program or instructions to implement the aforementioned disaster early warning method.

[0019] Fourthly, embodiments of this application also provide a computer storage medium storing a computer program or instructions, which, when executed by a processor, implement the aforementioned geological disaster early warning method.

[0020] Fifthly, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the aforementioned geological disaster early warning method.

[0021] Compared with existing technologies, this application has the following advantages: 1. This application achieves a significant improvement in early warning accuracy and reliability by constructing a multimodal spatiotemporal map attention network model; the model utilizes a cross-modal attention mechanism to deeply integrate multi-source information such as geological sensors, remote sensing images, and meteorological data, overcoming the shortcomings of insufficient dimensions of a single data source; at the same time, by introducing physical mechanisms to construct spatiotemporal correlations, the model combines data-driven capabilities with geophysical consistency, significantly reducing false alarm and false negative rates; in addition, the dynamic spatiotemporal attention mechanism enables the model to adapt to changes in the geological environment, overcoming the problem of the accuracy of traditional static models decaying over time; 2. This application also achieves an optimized balance between early warning efficiency and deployment cost through knowledge distillation technology; it transforms the complex headquarters "teacher model" into a lightweight local "student model", enabling edge devices to have real-time reasoning capabilities at the second level, which greatly improves the early warning response speed, while reducing the dependence on high-performance hardware and the cost of large-scale deployment; finally, the continuous risk probability output by the model supports hierarchical early warning and intelligent verification processes, and through a dynamic threshold optimization mechanism, it makes early warning decisions more refined and intelligent.

[0022] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart of a geological disaster early warning method according to an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of a geological disaster early warning system according to an embodiment of this application is shown; Figure 3 A schematic diagram of the structure of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] To address the shortcomings of existing technologies, this application discloses a method for early warning of geological disasters, such as... Figure 1 As shown, the method includes the following steps: Step S1: Acquire multimodal time-series data of the target monitoring area and perform preprocessing. The multimodal time-series data includes geological disaster environmental data, image data, and meteorological data. Step S2: Input the preprocessed multimodal time series data into the pre-trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk within the target monitoring area. Step S3: Trigger corresponding disaster warning or verification actions based on the disaster risk probability.

[0027] In some specific embodiments, the method further includes step S0: constructing and training a geological disaster prediction model, including the following: (1) Obtain the corresponding historical multimodal time series data from different monitoring areas and preprocess the historical multimodal time series data.

[0028] It should be noted that the acquisition and preprocessing processes in this step are consistent with the multimodal time series data acquisition and preprocessing processes in step S1.

[0029] (2) Based on different monitoring areas and their historical multimodal time series data, construct spatiotemporal maps corresponding to each monitoring area, including: Define the graph structure: model each monitoring area as a spatiotemporal graph. .

[0030] Define node V: Divide each monitoring area into several grid units, each grid unit serves as a node in the corresponding spatiotemporal graph, and assign static attributes and dynamic temporal characteristics to each node.

[0031] The static attributes include elevation, slope, soil type, and vegetation coverage. The dynamic time series features are obtained through preprocessed historical multimodal time series data, including surface displacement rate, groundwater level and cumulative rainfall over a past period.

[0032] Define edge E: Based on the physical mechanism model and data-driven method, construct and connect the spatiotemporal edges between nodes to obtain the corresponding physical mechanism edges and data-driven edges, so as to form the spatiotemporal map of each monitoring area.

[0033] The physical mechanism edges include groundwater diffusion connection edges constructed according to Darcy's law and rock stress transmission edges constructed based on an elasticity model. For example, groundwater diffusion connection edges are constructed between adjacent grids with a hydraulic gradient > 0.01 according to Darcy's law; and rock stress transmission edges are established between grids with a stress propagation distance ≤ 3 km according to the elasticity model.

[0034] The data-driven edges are established by calculating the dynamic time-normalized distance of the surface displacement sequence between grid cells. For example, the dynamic time-normalized distance of the surface displacement sequence between grid cells is calculated, and edges are established between nodes whose dynamic time-normalized distance is less than a preset threshold.

[0035] (3) Construct a multimodal attention network model based on the nodes, physical mechanism edges and data-driven edges in the spatiotemporal graphs of different monitoring areas.

[0036] (4) Training a multimodal attention network model: Labeled training samples are obtained from historical multimodal time-series data to construct a training set; the training set is then used to train a multimodal attention network model based on a spatiotemporal graph structure.

[0037] The mixed loss function is defined using the focal loss and Kl divergence constraint model as follows: , In the formula, L represents the mixture loss function. and This represents the preset loss coefficient, used to balance the weights of the two losses. ; This represents the focus loss function, used to address the imbalance problem between positive and negative samples (those that have experienced a geological disaster vs. those that have not). and represents the predicted values ​​and ground truth labels of a multimodal attention network model; KLdiv represents the KL divergence constraint model, used to constrain the probability distribution of the predictions. Probability distribution derived from physical simulation model Maintain consistency and incorporate prior physical knowledge to improve model interpretability and reliability.

[0038] Based on the hybrid loss function, the parameters of the multimodal attention network model (such as the prediction attention coefficient and the weight of the loss function) are continuously adjusted through the backpropagation algorithm. The multimodal attention network model is iteratively trained until the predicted disaster risk probability meets the convergence condition, thus obtaining the trained multimodal attention network model.

[0039] It should be noted that the model training process in this step is consistent with the prediction process of the geological disaster prediction model described in step S2.

[0040] (5) The trained multimodal attention network model is deployed in a lightweight manner using knowledge distillation technology to obtain a geological disaster prediction model for the target monitoring area, specifically including: The trained multimodal attention network model is deployed on a central server and used as a general teacher model. For the target monitoring area to be deployed, a dedicated spatiotemporal graph is constructed based on the multimodal time series data of the target monitoring area. A dedicated student model is constructed based on the dedicated spatiotemporal graph, and the essential knowledge of the teacher model (such as spatiotemporal attention weights as soft labels) is transferred to the lightweight dedicated student model through knowledge distillation technology.

[0041] The geological disaster prediction model (referring to the trained dedicated student model) is obtained by training a dedicated student model, and the geological disaster prediction model is deployed to the local edge computing device corresponding to the target monitoring area to perform real-time geological disaster prediction inference in the target monitoring area. It can quickly predict and output geological disasters and achieve risk assessment in seconds.

[0042] Among them, the geological disaster prediction model is a multimodal attention network model based on a spatiotemporal graph structure; the teacher model uses the spatiotemporal attention weights output by the multimodal spatiotemporal attention network model as soft labels; the student model uses cross-entropy loss to fit the soft labels and retains high-weight connections.

[0043] It should be noted that the geological disaster prediction model is used to perform real-time geological disaster prediction inference for the target monitoring area. The target monitoring area can be different target monitoring areas. As long as a local deployment is carried out for a certain target monitoring area and a dedicated spatiotemporal map is constructed, real-time prediction can be performed for that target monitoring area.

[0044] In some specific embodiments, step S1: acquiring multimodal time-series data of the target monitoring area and performing preprocessing, including the following: (1) Acquire multimodal time series data: Target monitoring areas are determined based on potential geological hazard sites.

[0045] Multiple geological disaster monitoring devices are deployed within the target monitoring area, and geological disaster environmental data of the target monitoring area are acquired in real time through these devices. The geological disaster monitoring devices include GNSS receivers, rain gauges, crack gauges, tilt accelerometers, moisture meters, and mud level gauges. The geological disaster environmental data includes surface displacement, rainfall, crack changes, rock strata tilt angle, soil moisture content, and groundwater level data.

[0046] Real-time acquisition of image data of the target monitoring area provided by the remote sensing platform; wherein, the image data includes surface deformation data or optical image data, and the spatial resolution of the image data is not greater than a first predetermined value (e.g., 10 meters).

[0047] Meteorological data of the target monitoring area is acquired in real time through a meteorological platform; wherein the meteorological data includes soil moisture and wind speed, and the time resolution of the meteorological data is a predetermined time (e.g., 5 minutes).

[0048] The specific acquisition accuracy indicators of the multimodal time series data are as follows: Image data: Spatial resolution ≤ 10 meters; Geological disaster environmental data: GNSS receivers collect three-dimensional coordinate changes with an accuracy within ±10mm; Subsurface inclinometer: measures the dip angle of rock strata (0.01° resolution); Rain gauge (using tipping bucket rain gauge): rainfall measurement resolution is 0.1 mm, rainfall intensity measurement range is 0~4 mm / min; Slurry level gauge: Monitoring range 0.6–40 m; resolution ±0.1% F·S; Crack gauge: Monitoring range is 0~50 / 100 / 200 / 500cm, and crack monitoring accuracy is ±0.1%F·S; Moisture meter: Monitoring range is dry soil to saturated soil, and the monitoring accuracy is ±4%.

[0049] (2) Preprocessing of multimodal time series data: The spatiotemporal difference algorithm is used to unify multimodal time series data with different spatiotemporal resolutions into a monitoring grid of a preset size (e.g., 5km×5km), and the timestamps are aligned to the UTC standard timestamps to obtain time-aligned multimodal time series data.

[0050] The time-aligned multimodal time series data is normalized using the following formula to eliminate dimensional differences, resulting in the normalized multimodal time series data: , In the formula, z represents the normalized multimodal time series data, x represents the time-aligned multimodal time series data, and μ represents the mean of the historical multimodal time series data. It represents the standard deviation.

[0051] Data cleaning is performed on the normalized multimodal time series data, including: detecting and removing outliers in the normalized multimodal time series data based on the Isolation Forest algorithm, and completing the missing data using the spatiotemporal kriging interpolation method to ensure the continuity of the data.

[0052] In some specific embodiments, step S2: The preprocessed multimodal time series data is input into a pre-trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights, outputting the disaster risk probability within the target monitoring area, including the following: (1) Feature extraction: The preprocessed multimodal time-series data is input into the geological disaster prediction model. The geological disaster prediction model extracts multimodal features from the input geological disaster environmental data, image data, and meteorological data to obtain corresponding geological disaster time-series features, image time-series features, and meteorological time-series features.

[0053] (2) The geological disaster prediction model performs cross-modal attention fusion on geological disaster time series features, image time series features and meteorological time series features to obtain multimodal fusion features, including: Using a cross-modal attention fusion mechanism, the correlation weight between any two modal features is calculated as follows: , In the formula, This represents the correlation weight between the i-th modal feature and the j-th modal feature. This represents the query vector for the i-th modal feature. The key vector representing the j-th modal feature. Indicates the dimension scaling factor. This represents the transpose of a matrix.

[0054] The correlation weights are used to fuse all modal features (referring to geological disaster time series features, image time series features, and meteorological time series features) to obtain a unified multimodal fusion feature: , In the formula, Indicates multimodal fusion features, This represents the total number of modal features. , , ; This represents the value vector of the j-th modal feature.

[0055] (3) Based on the multimodal fusion features, information is propagated through a graph attention network on the dedicated spatiotemporal map corresponding to the target monitoring area, the spatial dependencies between nodes are dynamically learned, and the dynamic temporal features in the node features are analyzed to capture the temporal evolution law, thereby obtaining the predicted attention coefficient and temporal sensitive features, specifically including: The multimodal fusion features are injected into the dedicated spatiotemporal map corresponding to the target monitoring area; Using a graph attention network for spatiotemporal graph attention propagation, the predicted attention coefficients between nodes and their neighboring nodes in the specific spatiotemporal graph are dynamically calculated as follows: , In the formula, This represents the predicted attention coefficient from the target node u to its neighbor node v, i.e., the predicted attention coefficient of the edge uv; u represents the target node, and v represents a neighbor node directly connected to the target node u. Represents the set of all neighboring nodes of node u; and Let represent the feature vectors of node u and its neighbor node v, respectively; || denotes the vector concatenation operation; and W represents the learnable weight matrix. This represents a learnable attention vector. This represents the activation function. Represents an exponential function; Based on the predicted attention coefficients, the information of neighboring nodes is aggregated to update the node features in the dedicated spatiotemporal graph, resulting in the updated node features: , In the formula, This represents the updated feature vector of node u. Represents a non-linear activation function. Let represent the predicted attention coefficients from the target node u to its neighboring node v, and W represent the learnable weight matrix. This represents the feature vector of the neighbor node v. Represents the set of all neighboring nodes of node u; Based on the updated node features, the node feature sequence at multiple consecutive time steps is obtained as follows: The node feature sequence is then input into a time-series decoder to generate time-sensitive features that reflect the time-series change patterns; where, Indicates a time step. This indicates the total number of time steps.

[0056] (4) Input the time-sensitive features into the output layer of the geological disaster prediction model, and combine the prediction attention coefficient and the preset hybrid loss function to calculate the probability of occurrence of each disaster in the target monitoring area, that is, obtain the disaster risk probability.

[0057] In some specific embodiments, step S3: triggering corresponding disaster warning or verification actions based on the disaster risk probability includes the following: If the probability of disaster risk is greater than or equal to a preset probability threshold, then the drone will be triggered to conduct a verification. Otherwise, different red, orange, and yellow warnings will be set based on other corresponding probability thresholds.

[0058] The probability threshold can be dynamically adjusted using the Pareto algorithm, and the probability threshold setting can be continuously optimized in practice to reduce the probability of false alarms and missed alarms. At the same time, different geological hazard points have different geological conditions, which leads to different types and magnitudes of geological hazards. Therefore, the preset probability threshold of the target monitoring area corresponding to different geological hazard points cannot be uniformly set.

[0059] Based on the same inventive concept, embodiments of this application disclose a geological disaster early warning system, such as... Figure 2 As shown, it includes: The real-time acquisition unit is used to acquire and preprocess multimodal time-series data of the target monitoring area in real time. The multimodal time-series data includes geological disaster environmental data, image data, and meteorological data. The prediction unit is used to input the preprocessed multimodal time series data into the trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk in the target monitoring area. The early warning unit is used to trigger corresponding disaster early warning or verification actions based on the disaster risk probability.

[0060] In some specific embodiments, the system further includes a construction and training unit; The construction and training unit is used to construct and train the geological disaster prediction model; Specifically, the construction and training unit is used for: Based on different monitoring areas and their historical multimodal time series data, a spatiotemporal map corresponding to each monitoring area is constructed, and the multimodal attention network model is constructed according to the structure of each spatiotemporal map; The multimodal attention network model is trained using the training set to obtain the trained multimodal attention network model; The multimodal attention network model trained by knowledge distillation is deployed in a lightweight manner to obtain a geological disaster prediction model for the target monitoring area. The geological disaster prediction model is a multimodal attention network model based on a spatiotemporal graph structure.

[0061] Regarding the system in the above embodiments, the specific manner in which each unit module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0062] Based on the same inventive concept, embodiments of this application also provide an electronic device, the structure of which is as follows: Figure 3 As shown, it includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program or instructions to implement the aforementioned geological disaster early warning method.

[0063] Based on the same inventive concept, this application also provides a computer storage medium storing a computer program or instructions, which, when executed by a processor, implements the aforementioned geological disaster early warning method.

[0064] Based on the same inventive concept, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the aforementioned geological disaster early warning method.

[0065] This application constructs a novel multimodal spatiotemporal graph attention network model by utilizing various existing known neural network models. This model possesses spatiotemporal graph analysis capabilities and incorporates an attention mechanism to analyze the geological disaster status of the monitoring area from multiple dimensions using multimodal geological disaster monitoring data. Through the classification mechanism within the multimodal spatiotemporal graph attention network model, it can calculate and output the probability of each geological disaster occurring. Then, by displaying the probability of each geological disaster (i.e., geological disaster risk probability) on the geological disaster monitoring equipment platform, and issuing an early warning to staff when the probability of a certain geological disaster risk reaches a set value, staff can promptly know the probability of a geological disaster occurring in the monitoring area, enabling them to formulate response strategies more quickly for impending geological disasters and ensuring timely and effective response to geological disasters.

[0066] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for early warning of geological disasters, characterized in that, The method includes: Real-time acquisition and preprocessing of multimodal time-series data of the target monitoring area, including geological disaster environmental data, image data and meteorological data; The preprocessed multimodal time series data is input into a pre-trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk within the target monitoring area. Based on the stated probability of disaster risk, corresponding disaster warnings or verification actions are triggered.

2. The geological disaster early warning method according to claim 1, characterized in that, The real-time acquisition and preprocessing of multimodal time-series data of the target monitoring area includes: Multiple geological disaster monitoring devices are deployed within the target monitoring area, and geological disaster environmental data of the target monitoring area are acquired in real time through these devices. Simultaneously, image data of the target monitoring area provided by the remote sensing platform is acquired in real time, and meteorological data of the target monitoring area is acquired in real time through the meteorological platform. The spatiotemporal difference algorithm is used to unify geological disaster environmental data, image data and meteorological data with different spatiotemporal resolutions into a monitoring grid of a preset size for timestamp alignment, so as to obtain time-aligned multimodal time series data. Normalization and data cleaning are performed on the time-aligned multimodal time series data to obtain preprocessed multimodal time series data.

3. The geological disaster early warning method according to claim 2, characterized in that, The geological disaster monitoring equipment includes a GNSS receiver, rain gauge, crack gauge, tilt accelerometer, moisture meter, and mud level gauge. The geological disaster environmental data includes surface displacement, rainfall, crack changes, rock strata tilt angle, soil moisture content, and groundwater level data. The spatial resolution of the image data is not greater than a first predetermined value; The meteorological data includes soil moisture and wind speed, and the time resolution of the meteorological data is a predetermined time.

4. A geological disaster early warning method according to claim 1 or 3, characterized in that, The method further includes constructing and training the geological disaster prediction model, including: Based on different monitoring areas and their historical multimodal time series data, a spatiotemporal map corresponding to each monitoring area is constructed, and a multimodal attention network model is constructed according to the structure of each spatiotemporal map; The multimodal attention network model is trained using the training set to obtain the trained multimodal attention network model; By employing knowledge distillation techniques to deploy a lightweight multimodal attention network model after training, a geological disaster prediction model for the target monitoring area can be obtained.

5. A geological disaster early warning method according to claim 4, characterized in that, Based on different monitoring areas and their historical multimodal time-series data, a spatiotemporal map corresponding to each monitoring area is constructed, and a multimodal attention network model is constructed according to the structure of each spatiotemporal map, including: Each monitoring area is divided into several grid units, each grid unit serves as a node in the corresponding spatiotemporal map, and each node is assigned static attributes and dynamic temporal characteristics; Based on the physical mechanism model and data-driven method, spatiotemporal edges are constructed and connected between nodes to obtain corresponding physical mechanism edges and data-driven edges, so as to form a spatiotemporal map of each monitoring area; A multimodal attention network model is constructed based on the nodes, physical mechanism edges, and data-driven edges in the spatiotemporal graphs of different monitoring areas. The static attributes include elevation, slope, soil type, and vegetation coverage. The dynamic time series features are obtained through preprocessed historical multimodal time series data, including surface displacement rate, groundwater level and cumulative rainfall over a past period.

6. A geological disaster early warning method according to claim 4, characterized in that, During training, a hybrid loss function is used to optimize the multimodal attention network model; The hybrid loss function is defined using the focal loss and Kl divergence constraint model as follows: ， In the formula, L represents the mixture loss function. and This represents the preset loss coefficient. Represents the focus loss function. and KLdiv represents the predicted values ​​and ground truth labels of a multimodal attention network model, and KLdiv represents the KL divergence constraint model, used to constrain the probability distribution of the predictions. Probability distribution derived from physical simulation model Maintain consistency.

7. A geological disaster early warning method according to claim 4, characterized in that, The process of lightweightly deploying the trained multimodal attention network model using knowledge distillation technology to obtain a geological disaster prediction model for the target monitoring area includes: The trained multimodal attention network model is deployed on a central server and used as a general teacher model; For the target monitoring area to be deployed, a dedicated spatiotemporal map is constructed based on the multimodal time series data of the target monitoring area; A dedicated student model is constructed based on a dedicated spatiotemporal graph, and the knowledge of the teacher model is transferred to the lightweight dedicated student model through knowledge distillation technology. The dedicated student model is trained to obtain the trained dedicated student model, that is, the geological disaster prediction model; the geological disaster prediction model is deployed to the local edge computing device corresponding to the target monitoring area to perform real-time geological disaster prediction inference in the target monitoring area.

8. A geological disaster early warning method according to claim 1 or 3, characterized in that, The geological disaster prediction model fuses multimodal features from multimodal time-series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk within the target monitoring area, including: The geological disaster prediction model extracts multimodal features from the input geological disaster environmental data, image data, and meteorological data to obtain corresponding geological disaster time series features, image time series features, and meteorological time series features. Cross-modal attention fusion is performed on geological disaster time-series features, image time-series features, and meteorological time-series features to obtain multimodal fused features; Based on the multimodal fusion features, information is propagated through a graph attention network on the dedicated spatiotemporal map corresponding to the target monitoring area. The spatial dependencies between nodes are dynamically learned, and the dynamic temporal features in the node features are analyzed to capture the temporal evolution law, thereby obtaining the predicted attention coefficient and temporal sensitive features. The time-sensitive features are input into the output layer of the geological disaster prediction model, and the disaster risk probability is calculated by combining the prediction attention coefficient and the preset hybrid loss function.

9. A geological disaster early warning method according to claim 8, characterized in that, The method of performing cross-modal attention fusion on geological disaster time-series features, image time-series features, and meteorological time-series features to obtain multimodal fused features includes: Using a cross-modal attention fusion mechanism, the correlation weight between any two modal features is calculated as follows: ， In the formula, This represents the correlation weight between the i-th modal feature and the j-th modal feature. This represents the query vector for the i-th modal feature. The key vector representing the j-th modal feature. Indicates the dimension scaling factor. Represents the transpose of a matrix; The correlation weights are used to fuse all modal features to obtain a unified multimodal fusion feature: ， In the formula, Indicates multimodal fusion features, This represents the total number of modal features. , , ; This represents the value vector of the j-th modal feature.

10. A geological disaster early warning method according to claim 8, characterized in that, Based on the aforementioned multimodal fusion features, information propagation is performed through a graph attention network on the dedicated spatiotemporal map corresponding to the target monitoring area. This dynamically learns the spatial dependencies between nodes and analyzes the dynamic temporal features within the node features to capture temporal evolution patterns, thereby obtaining predicted attention coefficients and temporally sensitive features, including: The multimodal fusion features are injected into the dedicated spatiotemporal map corresponding to the target monitoring area; Spatiotemporal graph attention propagation is performed using graph attention networks. Predicted attention coefficients between nodes and their neighbors in a specific spatiotemporal graph are dynamically calculated, and information about the neighbors is aggregated based on these predicted attention coefficients to update node features. Based on the updated node features, a node feature sequence is obtained over multiple consecutive time steps, and the node feature sequence is input into a time-series decoder to generate time-sensitive features that reflect the time-series change patterns. The predicted attention coefficient is calculated using the following formula: ， In the formula, This represents the predicted attention coefficient from the target node u to its neighbor node k, i.e., the predicted attention coefficient of the edge uv; u represents the target node, and k represents a neighbor node directly connected to the target node u. Represents the set of all neighboring nodes of node u; and Let represent the feature vectors of node u and its neighbor node k, respectively; || denotes the vector concatenation operation; and W represents the learnable weight matrix. This represents a learnable attention vector. This represents the activation function. This represents an exponential function.

11. A geological disaster early warning system, characterized in that, The system includes: The real-time acquisition unit is used to acquire and preprocess multimodal time-series data of the target monitoring area in real time. The multimodal time-series data includes geological disaster environmental data, image data, and meteorological data. The prediction unit is used to input the preprocessed multimodal time series data into the trained geological disaster prediction model. The geological disaster prediction model fuses the multimodal features of the multimodal time series data and dynamically adjusts the spatiotemporal weights to output the probability of disaster risk in the target monitoring area. The early warning unit is used to trigger corresponding disaster early warning or verification actions based on the disaster risk probability.

12. A geological disaster early warning system according to claim 11, characterized in that, The system also includes a construction and training unit; The construction and training unit is used to construct and train the geological disaster prediction model; Specifically, the construction and training unit is used for: Based on different monitoring areas and their historical multimodal time series data, a spatiotemporal map corresponding to each monitoring area is constructed, and the multimodal attention network model is constructed according to the structure of each spatiotemporal map; The multimodal attention network model is trained using the training set to obtain the trained multimodal attention network model; By employing knowledge distillation techniques to deploy a lightweight multimodal attention network model after training, a geological disaster prediction model for the target monitoring area can be obtained.

13. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program or instructions to implement a geological disaster early warning method according to any one of claims 1-10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement a geological disaster early warning method according to any one of claims 1-10.

15. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the geological disaster early warning method according to any one of claims 1-10.

Images