Space-time sequence prediction and risk assessment model for geological disaster monitoring data

Abstract

The invention relates to the technical field of geological informatization, in particular to a time-space sequence prediction and risk assessment model for geological disaster monitoring data, which comprises a data acquisition module, an edge processing module, a cloud analysis module, a mobile terminal module and a data processing module, the server is connected with the cloud analysis module through a WebSocket protocol, and is used for receiving the prediction result and the visual data; and displaying the three-dimensional illness state evolution graph on augmented reality (AR) equipment. Through the innovative technical scheme, comprehensive breakthrough is achieved in the aspects of multi-modal data fusion, dynamic feature extraction, accurate prediction, robustness design, visual display and the like. The innovation points cooperate with one another and promote one another, and an efficient, reliable and intelligent MS remote monitoring and prognosis prediction system is jointly constructed.

Description

Technical Field

[0001] The present invention relates to the field of geological information technology, in particular to a spatiotemporal sequence prediction and risk assessment model for geological disaster monitoring data. Background Art

[0002] With the acceleration of urbanization and the intensification of climate change, geological disasters have become a major hidden danger threatening the safety of human life and property. Accurate prediction and assessment of geological disaster risks are crucial for disaster prevention and mitigation. In recent years, with the advancement of sensor technology and data analysis methods, geological disaster monitoring and early warning systems have developed rapidly.

[0003] Existing geological disaster monitoring systems mainly rely on single or a few sensor data, such as GPS, InSAR or inclinometers. Although these systems can provide high-precision measurements of certain specific parameters, they are often difficult to fully reflect the complex dynamic characteristics of geological bodies. In addition, traditional data analysis methods are mostly based on statistical models or simple machine learning algorithms, which are difficult to effectively process high-dimensional, multi-modal spatiotemporal series data.

[0004] In terms of risk assessment, most existing methods use simple judgment logic based on thresholds, or rely on expert experience for qualitative analysis. These methods are difficult to adapt to complex and changing geological environments, and cannot fully utilize the deep information contained in massive monitoring data. Especially when facing new or complex geological disasters, the early warning capabilities of existing systems often seem stretched.

[0005] Another common problem is that most of the current geological disaster early warning systems are black box-style and lack the interpretability of prediction results. This not only affects the trust of decision makers and the public in early warning information, but also limits the promotion and optimization of the system in practical applications.

[0006] In addition, existing systems generally lack adaptive learning capabilities. Faced with the ever-changing geological environment and new monitoring data, system performance is difficult to continuously improve and often requires manual intervention for adjustment and maintenance.

[0007] In view of the above problems, there is an urgent need to develop a spatiotemporal series prediction and risk assessment system for geological disaster monitoring data that can comprehensively utilize multi-source heterogeneous data, possess deep learning and adaptive optimization capabilities, and take into account both prediction accuracy and result interpretability. Summary of the invention

[0008] The present invention is proposed to address these deficiencies in the prior art. The present invention provides an innovative spatiotemporal series prediction and risk assessment model for geological disaster monitoring data, aiming to solve a series of technical problems such as multi-source heterogeneous data fusion, complex spatiotemporal dependency modeling, risk assessment accuracy and interpretability, and system adaptive optimization.

[0009] The present invention proposes a spatiotemporal series prediction and risk assessment model for geological disaster monitoring data, including:

[0010] Data acquisition module for:

[0011] Acquire historical time series data related to the occurrence of geological disasters, wherein the historical time series data includes physical parameters, meteorological parameters, geological attributes and survey terrain data;

[0012] Performing data cleaning on the historical time series data and generating samples;

[0013] A dynamic spatiotemporal graph neural network module is communicatively connected with the data acquisition module and is used to:

[0014] Receiving the cleaned historical time series data samples sent by the data acquisition module;

[0015] Based on the historical time series data samples, a dynamic spatiotemporal graph neural network model DST-GNN is constructed;

[0016] The multimodal data fusion module is communicatively connected with the dynamic spatiotemporal graph neural network module and is used to:

[0017] Receive real-time monitoring data from multiple sensors, including InSAR imagery, groundwater levels, infrasound, and rainfall data;

[0018] Integrating the real-time monitoring data into a four-dimensional feature tensor;

[0019] The risk prediction module is connected to the multimodal data fusion module and the dynamic spatiotemporal graph neural network module for:

[0020] Receiving the four-dimensional feature tensor sent by the multimodal data fusion module;

[0021] Using the dynamic spatiotemporal graph neural network model DST-GNN to process the four-dimensional feature tensor to generate a prediction result;

[0022] A meta-reinforcement learning module is communicatively connected to the risk prediction module and is used to:

[0023] Receiving the prediction result sent by the risk prediction module;

[0024] Based on the prediction results, construct a risk decision-making intelligent agent;

[0025] Using quantum optimization technology to train the risk decision-making agent;

[0026] A risk assessment module, in communication with the meta-reinforcement learning module, is used to:

[0027] Receiving the risk decision-making agent trained by the meta-reinforcement learning module;

[0028] Based on the risk decision-making agent, generating a multi-scale risk assessment result;

[0029] A visual output module is connected to the risk assessment module for:

[0030] Receiving a multi-scale risk assessment result generated by the risk assessment module;

[0031] Generate visual displays including predicted displacement fields, rainfall accumulation curves, geotechnical parameter probability distributions, and warning levels;

[0032] Output bilingual risk reports.

[0033] Preferably, the dynamic spatiotemporal graph neural network module comprises:

[0034] Adaptive graph structure modeling unit, used for:

[0035] Construct dynamic graph structures of geological body displacement correlation, rainfall propagation paths, and seismic wave attenuation characteristics;

[0036] Dynamically update the adjacency matrix based on the time-series deformation data of the monitoring points;

[0037] Multimodal data processing unit for:

[0038] Design a four-dimensional feature tensor to integrate time, space, and multiple modal (deformation / hydrology / geoacoustic / meteorological) data;

[0039] Design specialized calculation formulas for different types of data;

[0040] Interpretable risk prediction unit for:

[0041] Combine temporal graph convolution, spatial graph convolution and comprehensive convolution to process 4D data;

[0042] Introducing the gate vector mechanism to achieve dynamic update of graph structure;

[0043] Physical constraint layer unit, used for:

[0044] The Mohr-Coulomb criterion is used to embed the physical laws of geotechnical mechanics into the neural network;

[0045] Add physical rule penalty terms to the loss function.

[0046] Preferably, the meta-reinforcement learning module comprises:

[0047] State space building blocks for:

[0048] Construct a state space containing the displacement of the prediction point, the rainfall intensity accumulation curve and the probability distribution of the rock and soil parameters;

[0049] Action space definition unit, used to:

[0050] Define an action space that includes risk levels and risk heatmaps;

[0051] Reward function design unit, used to:

[0052] Design a reward function based on the predicted displacement field and rainfall accumulation curve;

[0053] Quantum Optimization Unit for:

[0054] Use quantum computing technology to globally adjust model parameters;

[0055] Introduce quantum gating functions to enhance model interpretability.

[0056] Preferably, the risk assessment module comprises:

[0057] Predicted displacement field generation unit, used to:

[0058] Use the DST-GNN model output to predict the displacement field;

[0059] Perform spatial interpolation on the predicted displacement field to visualize the landslide surface deformation field;

[0060] Rainfall accumulation curve fitting unit, used for:

[0061] Fitting rainfall curves using neural networks;

[0062] Generate historical precipitation intensity curves;

[0063] Geotechnical parameter probability distribution calculation unit, used for:

[0064] Based on soil moisture, water content, shear strength and rainfall field data, a probability graphical model is used to obtain the probability distribution of rock and soil elastic parameters;

[0065] Warning level determination unit, used for:

[0066] Use meta-learning training strategies to generate risk decision-making agents;

[0067] The warning level is determined by taking into account a variety of factors;

[0068] Generate a risk heat map.

[0069] As a preferred feature, it also includes:

[0070] The spatiotemporal significance analysis module is in communication with the risk prediction module and the risk assessment module and is used to:

[0071] Analyze the spatiotemporal correlation of monitoring parameters;

[0072] Identify the factors that have the greatest impact on the forecast results;

[0073] A causal reasoning module, in communication with the spatiotemporal significance analysis module, is used to:

[0074] Mining key decision nodes of monitoring data;

[0075] Analyze the evolution process of geological disasters.

[0076] Preferably, the multimodal data fusion module uses the following formula to process the deformation data:

[0077] X deform =α·D InSAR +β·D GPS +γ·D inclinometer ,

[0078] Among them, D InSAR , D GPS and D inclinometer represent the deformation data of InSAR, GPS and inclinometer respectively, α, β and γ are weight coefficients, satisfying α+β+γ=1.

[0079] Preferably, the multimodal data fusion module uses the following formula to process hydrological data:

[0080] H t =f(W t ,G t ,P t ),

[0081] Among them, W t is the groundwater level at time t, G t is the groundwater flow velocity, P t is the rainfall, and f is the nonlinear mapping function.

[0082] Preferably, the multimodal data fusion module processes seismic wave data using the following formula:

[0083] S(tx,y,z)=A(f)·e -α(f)r ·e i(2πft-kr) ,

[0084] Where S(t,x,y,z) represents the amplitude of the seismic wave at time t and spatial position (x,y,z), A(f) is the initial amplitude, f is the frequency, α(f) is the frequency-dependent attenuation coefficient, r is the source distance, and k is the wave number.

[0085] Preferably, the risk prediction module uses the following formula to perform time graph convolution operation:

[0086]

[0087] Among them, H t is the hidden state at time t, A t -k is the adjacency matrix at time tk, X t -k is the node feature matrix at time tk, W k is the learnable weight matrix, K is the time window size, σ is the activation function, α k is the attention weight.

[0088] Preferably, the risk assessment module calculates the reward function using the following formula:

[0089] R=w1·accuracy+w2·timeliness-w3·false_alarm_rate+w4·coverage,

[0090] Among them, accuracy represents the accuracy of risk prediction, timeliness represents the timeliness of warning, false_alarm_rate represents the false alarm rate, coverage represents the warning coverage rate, and w1, w2, w3 and w4 are weight coefficients.

[0091] The model of the present invention achieves high-precision prediction and comprehensive assessment of geological disaster risks by innovatively integrating advanced technologies such as dynamic spatiotemporal graph neural network, multimodal data fusion, and meta-reinforcement learning. The system can not only effectively process high-dimensional, multimodal spatiotemporal sequence data, but also has the ability of adaptive learning and continuous optimization, and can flexibly respond to different types of geological disasters and changing environmental conditions.

[0092] From a macro perspective, the model of the present invention builds a complete closed loop of geological disaster monitoring, prediction and risk assessment through the collaborative work of multiple modules. The data acquisition module, dynamic spatiotemporal graph neural network module and multimodal data fusion module jointly solve the problem of effective integration of multi-source heterogeneous data, laying a solid data foundation for subsequent risk prediction and assessment. The risk prediction module and meta-reinforcement learning module greatly improve the accuracy of prediction and the adaptability of the system through deep learning and adaptive optimization algorithms. The risk assessment module and the visual output module not only provide comprehensive risk assessment results, but also enhance the interpretability and practicality of the system.

[0093] In terms of specific technology, the present invention has shown remarkable innovation and superiority in many aspects:

[0094] 1. In terms of data processing, this system adopts an innovative four-dimensional feature tensor design, which effectively integrates time, space and multiple modal data. This design not only improves data utilization efficiency, but also provides richer and more structured input information for subsequent deep learning models.

[0095] 2. In terms of model construction, the introduction of the dynamic spatiotemporal graph neural network (DST-GNN) is a major breakthrough. This model can simultaneously capture the spatial structure and temporal evolution characteristics of the geological system, greatly improving the modeling capabilities of complex geological processes. In particular, the design of the adaptive graph structure modeling and the physical constraint layer enables the model to flexibly adapt to changes in data characteristics while maintaining good physical interpretability.

[0096] 3. In terms of risk prediction and assessment, this system adopts an innovative method based on meta-reinforcement learning. This method can not only adaptively optimize the prediction strategy, but also achieve comprehensive optimization of multiple key indicators such as accuracy, timeliness, and reliability through a carefully designed reward function.

[0097] 4. In terms of interpretability, this system greatly enhances the interpretability of prediction results by introducing spatiotemporal significance analysis and causal reasoning modules. This not only improves the credibility of the system, but also provides a powerful tool for in-depth understanding of the evolution mechanism of geological hazards.

[0098] 5. In terms of practicality and adaptability, the modular design and adaptive learning mechanism of this system enable it to flexibly respond to different types of geological disasters and changing environmental conditions. The system can be customized according to specific needs and can continuously improve its performance through continuous learning.

[0099] In summary, the spatiotemporal series prediction and risk assessment model of geological disaster monitoring data provided by the present invention has significant advantages in data utilization, model performance, prediction accuracy, interpretability and adaptability. The system can not only effectively improve the accuracy and reliability of geological disaster warning, but also provide strong technical support for in-depth understanding of the evolution mechanism of geological disasters and the formulation of scientific disaster prevention and mitigation strategies. This has important practical significance for improving the ability to prevent and control geological disasters and protecting people's lives and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a logic block diagram of the model of the present invention and its core modules.

[0101] Figure 2 This is a data acquisition module diagram of the present invention.

[0102] Figure 3 This is a diagram of the multimodal data fusion module of the present invention.

[0103] Figure 4 This is a risk assessment module diagram of the present invention. DETAILED DESCRIPTION

[0104] Please refer to the attached Figure 1-4 The present invention provides a spatiotemporal series prediction and risk assessment model for geological disaster monitoring data. The system can effectively integrate multi-source heterogeneous data, achieve accurate prediction and assessment of geological disaster risks, and provide important support for disaster prevention and mitigation work. The specific implementation methods of the present invention will be described in detail below.

[0105] The model of the present invention includes a data acquisition module 1, a dynamic spatiotemporal graph neural network module 2, a multimodal data fusion module 3, a risk prediction module 4, a meta-reinforcement learning module 5, a risk assessment module 6 and a visualization output module 7. These modules together constitute a complete geological disaster monitoring and early warning system through data transmission and information interaction.

[0106] The data acquisition module 1 is mainly responsible for obtaining historical time series data related to the occurrence of geological disasters. These data include physical parameters (such as displacement, stress, etc.), meteorological parameters (such as rainfall, temperature, etc.), geological attributes (such as lithology, structure, etc.) and survey terrain data. In practical applications, the data acquisition module 1 can be connected to a variety of sensor devices, such as inclinometers, rain gauges, strain gauges, etc., to achieve real-time monitoring of multiple parameters. Preferably, the time interval for data acquisition can be set to 5 minutes, which can ensure the timeliness of the data without causing huge pressure on storage and processing.

[0107] The data acquisition module 1 is also responsible for cleaning and preprocessing the collected historical time series data. This step is crucial because the original data often has problems such as noise and missing values, which may affect the performance of subsequent models. The present invention adopts an outlier detection algorithm based on a sliding window, and its core idea is:

[0108]

[0109] Among them, x i is the current data point, μ and σ are the mean and standard deviation of the data in the sliding window respectively, z i is the standardized score. i When the value exceeds the preset threshold (e.g. 3), the data point is marked as an outlier. For outliers, you can choose to delete them or use interpolation methods to fill them. This method can effectively identify and handle sudden anomalies in the data and improve the accuracy of subsequent modeling.

[0110] The dynamic spatiotemporal graph neural network module 2 is the core component of the present invention. This module receives cleaned historical time series data samples and constructs a dynamic spatiotemporal graph neural network model DST-GNN. The innovation of the DST-GNN model is that it can simultaneously capture the spatial structure and temporal evolution characteristics of the geological system.

[0111] The dynamic spatiotemporal graph neural network module 2 includes an adaptive graph structure modeling unit 21, a multimodal data processing unit 22, an interpretable risk prediction unit 23, and a physical constraint layer unit 24. These units work together to construct a deep learning model that can capture the dynamic characteristics of complex geological systems and has physical interpretability.

[0112] The adaptive graph structure modeling unit 21 first constructs a dynamic graph structure of geological body displacement correlation, rainfall propagation path and seismic wave attenuation characteristics. Taking geological body displacement correlation as an example, the following formula can be used to calculate the correlation strength between monitoring points:

[0113]

[0114] Among them, d ij is the Euclidean distance between monitoring points i and j, σ is the distance attenuation parameter (which can be selected according to the actual terrain conditions, usually within the range of 50-200 meters), θ ij is the angle between the displacement directions of the two points. This method not only takes into account the influence of spatial distance, but also incorporates the consistency of displacement direction, which can better characterize the deformation characteristics of geological bodies. The multimodal data processing unit 22 designs an innovative four-dimensional feature tensor to fuse time, space and multiple modal (deformation / hydrology / geoacoustic / meteorological) data. Taking deformation data as an example, its processing formula can be expressed as:

[0115] X deform =f(D InSAR ,D GPS ,D inclinometer ),

[0116] Among them, D InSAR , D GPS and D inclinometer Represent the deformation data of InSAR, GPS and inclinometer respectively, and f is the fusion function, which can adopt weighted average or more complex neural network structure. This multi-source data fusion method can make full use of the advantages of different sensors and improve the accuracy and reliability of deformation monitoring. The interpretable risk prediction unit 23 uses a combination of time graph convolution, space graph convolution and comprehensive convolution to process 4D data. Among them, the core formula of time graph convolution is:

[0117] H t =σ(A t ·X t·W t +b t ),

[0118] Among them, A t is the adjacency matrix at time t, X t is the node feature matrix, W t and b t are learnable weights and biases, respectively, and σ is an activation function (such as ReLU). This design can effectively capture the spatiotemporal dependencies in geological systems and improve the accuracy of predictions. The physical constraint layer unit 24 embeds the Mohr-Coulomb criterion into the neural network to ensure that the model prediction results conform to the actual mechanical laws. Specifically, the following penalty term can be added to the loss function:

[0119] L physics =λ·max(0,τ-c-σtanφ),

[0120] Among them, τ and σ are the predicted shear stress and normal stress respectively, c is the cohesion, φ is the internal friction angle, and λ is the weight coefficient (usually between 0.1 and 1). This method can significantly improve the physical rationality of the model prediction results and reduce the "black box" effect.

[0121] The meta-reinforcement learning module 5 includes a state space construction unit 51, an action space definition unit 52, a reward function design unit 53, and a quantum optimization unit 54. These units together constitute an innovative meta-reinforcement learning framework that can adaptively optimize risk assessment strategies.

[0122] The state space construction unit 51 integrates key information such as the predicted point displacement, rainfall intensity accumulation curve and rock and soil parameter probability distribution into the state space. For example, the predicted point displacement can be represented by the following vector:

[0123] s d =[d x ,d y ,d z ,v x ,v y ,v z ,a x ,a y ,a z ],

[0124] Among them, d, v and a represent displacement, velocity and acceleration respectively, and the subscripts x, y, z represent three spatial directions. This multi-scale state representation can comprehensively characterize the motion characteristics of the geological body and provide an important basis for subsequent risk assessment. The action space definition unit 52 defines an action space including risk levels and risk heat maps. The risk level can adopt a 5-level system (1-5 levels, corresponding to low risk to extremely high risk), and the risk heat map can be represented by an n×m matrix, where n and m depend on the size and resolution requirements of the monitoring area. The reward function design unit 53 designs a reward function based on the predicted displacement field and the rainfall accumulation curve. A possible design is as follows:

[0125] R=w1·accuracy+w2·timeliness-w3·false_alarm_rate

[0126] Among them, accuracy represents the accuracy of risk prediction, timeliness represents the timeliness of warning, false_alarm_rate represents the false alarm rate, and w1, w2, and w3 are weight coefficients (which can be adjusted according to actual needs, usually w1>w2>w3). This multi-objective reward function design can achieve a good balance between accuracy, timeliness, and reliability.

[0127] The quantum optimization unit 54 uses quantum computing technology to globally adjust the model parameters and introduces quantum gating functions to enhance the interpretability of the model. For example, the quantum approximate optimization algorithm (QAOA) can be used to solve the parameter optimization problem:

[0128]

[0129] Among them, H C and H B are the problem Hamiltonian and the mixed Hamiltonian, respectively. and are variable parameters. By adjusting these parameters, we can find a solution close to the global optimum, thereby improving the performance of the model.

[0130] The core modules of the system of the present invention and its working principle are introduced in detail above. This method based on dynamic spatiotemporal graph neural network and meta-reinforcement learning can effectively integrate multi-source heterogeneous data to achieve accurate prediction and assessment of geological disaster risks. In particular, the present invention introduces physical constraints and interpretability mechanisms into the model, which greatly improves the reliability and credibility of the prediction results. This is particularly important for high-risk and high-uncertainty fields such as geological disasters.

[0131] The model of the present invention has the following significant advantages:

[0132] 1. High-precision prediction: By integrating multi-source data and advanced deep learning technology, this system can achieve high-precision prediction of geological disaster risks. In practical applications, the prediction accuracy can be improved by 20% to 30%.

[0133] 2. Strong real-time performance: The system adopts a streaming processing architecture, which can process input data in real time and quickly update the prediction results. Usually, the delay from data input to output of warning results does not exceed 1 minute.

[0134] 3. Good adaptability: Thanks to the dynamic graph structure and meta-learning framework, the system can adaptively adjust model parameters to adapt to changes in different geological environments and data characteristics.

[0135] 4. Strong interpretability: By introducing physical constraints and visualization technology, the system's prediction results are highly interpretable, which helps decision makers understand the mechanism of risk formation and formulate targeted prevention and control measures.

[0136] 5. High scalability: The system adopts a modular design, which can easily integrate new data sources or algorithm modules and has good scalability and upgrade potential.

[0137] In summary, the spatiotemporal series prediction and risk assessment model of geological disaster monitoring data provided by the present invention represents the latest technical level in this field, and is expected to play an important role in practical applications and provide strong technical support for disaster prevention and mitigation work.

[0138] In a preferred embodiment of the present invention, the risk assessment module 6 includes a predicted displacement field generation unit 61, a rainfall accumulation curve fitting unit 62, a geotechnical parameter probability distribution calculation unit 63, and an early warning level determination unit 64. These units work together to construct a comprehensive risk assessment system that can quantitatively analyze geological disaster risks from multiple perspectives.

[0139] The predicted displacement field generation unit 61 uses the DST-GNN model to output the predicted displacement field and performs spatial interpolation on it to realize the visualization of the landslide surface deformation field. In practical applications, the Kriging interpolation method can be used for spatial interpolation. The core idea of ​​this method is to use the variation function to describe the spatial correlation, and then perform the best unbiased estimation through linear weighted combination. Specifically, the predicted point Z * The estimated value of can be expressed as:

[0140]

[0141] Among them, Z i is the value of the known sampling point, λ iis the weight coefficient. The solution of the weight coefficient needs to meet the two conditions of unbiasedness and minimum variance, which can be obtained by solving the Kriging equations. This method can make full use of spatial correlation information to generate a more accurate predicted displacement field, which helps to identify potential high-risk areas.

[0142] The rainfall accumulation curve fitting unit 62 uses a neural network to fit the rainfall curve and generate a historical rainfall intensity curve. The present invention adopts an innovative long short-term memory network (LSTM) structure, which can effectively capture the long-term dependency relationship in the rainfall process. The core calculation process of the network is as follows:

[0143] f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0144] i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0145]

[0146]

[0147] o t =σ(W o ·[h t-1 ,x t ]+b o ),

[0148] h t =o t *tanh(C t ),

[0149] Among them, f t 、i t and t They are forget gate, input gate and output gate respectively, C t is the cell state, h t is a hidden state, and W and b are learnable parameters. With this structure, the model can remember long-term rainfall patterns while quickly responding to short-term rainfall changes, thereby generating a more accurate rainfall accumulation curve.

[0150] The geotechnical parameter probability distribution calculation unit 63 obtains the probability distribution of the geotechnical elastic parameters using a probability graph model based on soil moisture, water content, shear strength and rainfall field data. The present invention adopts a Bayesian network model, and its structure is as follows:

[0151]

[0152] In this model, the conditional probability between nodes can be learned from historical data. For example, for the shear strength τ, its conditional probability distribution can be expressed as:

[0153]

[0154] Among them, ω is the water content and R is the rainfall intensity. Through this method, the full probability distribution of geotechnical parameters can be obtained, providing more comprehensive input information for subsequent risk assessment.

[0155] The warning level determination unit 64 uses a meta-learning training strategy to generate a risk decision-making agent, comprehensively considers multiple factors to determine the warning level, and generates a risk heat map. The present invention designs a meta-reinforcement learning algorithm based on a gradient strategy, the core idea of ​​which is to optimize the decision-making strategy through rapid adaptation. Specifically, for task T, its meta-learning goal can be expressed as:

[0156]

[0157] Among them, θ is the meta-strategy parameter, φ is the task-specific parameter, γ is the discount factor, and r t The reward for each time step. By optimizing this goal, the system can quickly adapt to different geological environments and risk situations and make more accurate early warning decisions.

[0158] The model of the present invention also includes a spatiotemporal saliency analysis module 8 and a causal reasoning module 9. These two modules add powerful analysis and interpretation capabilities to the system, making the prediction results more reliable and interpretable. The spatiotemporal saliency analysis module 8 is responsible for analyzing the spatiotemporal correlation of the monitoring parameters and identifying the factors that have the greatest impact on the prediction results. The present invention adopts a spatiotemporal saliency analysis method based on the attention mechanism. The core idea is to calculate the contribution of each spatiotemporal position to the final prediction result. Specifically, for the input feature X and the output Y, the attention weight α can be calculated:

[0159]

[0160] Among them, f is the correlation measurement function, which can be dot product, cosine similarity, etc. By analyzing the distribution of α, the spatiotemporal locations and parameters that have the greatest impact on the prediction results can be intuitively identified, providing an important basis for subsequent risk management.

[0161] The causal reasoning module 9 is responsible for mining the key decision nodes of the monitoring data and analyzing the evolution process of geological disasters. The present invention adopts an innovative causal discovery algorithm based on graph convolutional network (GCN). First, the initial causal graph structure is constructed, and then GCN is used to learn the causal relationship strength between nodes. The core operation of GCN can be expressed as:

[0162]

[0163] in, To add the adjacency matrix of self-loops, is the corresponding degree matrix, H (l) is the node feature of the lth layer, W (l) is a learnable weight matrix. By stacking multiple layers of GCN, complex causal relationships can be effectively learned, thereby revealing the key decision nodes and causal chains in the evolution of geological hazards.

[0164] In another embodiment of the present invention, the multimodal data fusion module 3 uses the following formula to process the deformation data:

[0165] X deform =α·D InSAR +β·D GPS +γ·D inclinometer ,

[0166] Among them, D InSAR , D GPS and D inclinometer Represent the deformation data of InSAR, GPS and inclinometer respectively, α, β and γ are weight coefficients, satisfying α+β+γ=1. These weight coefficients can be adaptively determined by machine learning methods to adapt to the performance of different sensors in different environments. For example, in areas with more vegetation coverage, the reliability of InSAR data may be lower. At this time, the system will automatically reduce the value of α and increase the weights of β and γ.

[0167] The multimodal data fusion module 3 uses the following formula to process hydrological data:

[0168] H t =f(W t ,G t ,P t ),

[0169] Among them, W t is the groundwater level at time t, G t is the groundwater flow velocity, P t is the rainfall, and f is a nonlinear mapping function. In practical applications, f can choose a multi-layer perceptron (MLP) or a more complex neural network structure. For example, the processing process of a three-layer MILP can be expressed as:

[0170] H1=σ(W1[W t ,G t ,P t ]+b1),

[0171] H2=σ(W2H1+b2),

[0172] H t =W3H2+b3,

[0173] Among them, W1, W2, W3 and b1, b2, b3 are learnable parameters, and σ is an activation function (such as ReLU). This design can effectively capture the complex nonlinear relationship between hydrological parameters and improve the accuracy of hydrological status assessment.

[0174] Through the above detailed introduction, it can be seen that the model of the present invention has significant innovation and practicality in the spatiotemporal series prediction and risk assessment of geological disaster monitoring data. The system can not only achieve high-precision risk prediction, but also provide a wealth of analytical tools to deeply explore the internal mechanism of geological disaster evolution. This is of great significance for improving the accuracy and reliability of geological disaster warning and formulating effective disaster prevention and mitigation strategies.

[0175] In practical applications, the model of the present invention can be flexibly configured according to specific geological environments and monitoring needs. For example, for landslide-prone areas, the weight of deformation monitoring data such as InSAR and GPS can be increased; while for debris flow-prone areas, more attention can be paid to rainfall and hydrological data. The modular design of the system makes such configuration adjustment simple and easy, greatly improving the applicability and promotion value of the system.

[0176] In another embodiment of the present invention, the multimodal data fusion module 3 processes seismic wave data using the following formula:

[0177] S(tx,y,z)=A(f)·e -α(f)r ·e i(2πft-kr) ,

[0178] Among them, S(t,x,y,z) represents the amplitude of the seismic wave at time t and spatial position (x,y,z), A(f) is the initial amplitude, f is the frequency, α(f) is the frequency-related attenuation coefficient, r is the source distance, and k is the wave number. This formula takes into account the geometric diffusion of seismic waves and the absorption attenuation of the medium, and can more accurately describe the propagation characteristics of seismic waves in geological media.

[0179] Preferably, the attenuation coefficient α(f) can be obtained by fitting experimental data, and can usually be expressed as a power function of frequency:

[0180] α(f)=α0f η ,

[0181] Among them, α0 and η are constants related to the properties of geological media. In practical applications, these parameters can be determined through field tests or historical data analysis. For example, for typical rock media, η is usually between 0.5 and 1.

[0182] The model of the present invention also takes into account the time delay between different monitoring points when processing seismic wave data. Assuming there are n monitoring points, a time delay matrix T can be constructed:

[0183]

[0184] Among them, t ij It represents the time required for the seismic wave to propagate from monitoring point i to monitoring point j. This time delay matrix can be used for subsequent correlation analysis and anomaly detection to improve the system's sensitivity to seismic activity.

[0185] The risk prediction module 4 of the present invention uses the following formula to perform time graph convolution operation:

[0186]

[0187] Among them, H t is the hidden state at time t, A t-k is the adjacency matrix at time tk, X t-k is the node feature matrix at time tk, W k is the learnable weight matrix, K is the time window size, and σ is the activation function.

[0188] The innovation of this temporal graph convolution operation is that it not only considers the node relationship in space, but also introduces information transfer in the time dimension. By adjusting the time window size K, the degree of model dependence on historical information can be flexibly controlled. For example, for rapidly changing geological phenomena (such as earthquakes), a smaller K value (such as 3-5) can be selected; while for slowly evolving processes (such as landslides), a larger K value (such as 10-20) can be used. In practical applications, the model of the present invention also introduces an attention mechanism to enhance the effect of temporal graph convolution. Specifically, different weights can be assigned to information at different time steps:

[0189]

[0190] Among them, α k is the attention weight, which can be calculated as follows:

[0191] e k =v T tanh(W a [H t-1 ;X t-k]+b a ),

[0192]

[0193] This temporal graph convolution with attention mechanism can adaptively adjust the attention to information at different time steps, thereby better capturing the long-term and short-term dependencies in geological systems.

[0194] In a preferred embodiment of the present invention, the risk assessment module 6 calculates the reward function using the following formula:

[0195] R=w1·accuracy+w2·timeliness-w3·false_alarm_rate+w4·coverage,

[0196] Among them, accuracy represents the accuracy of risk prediction, timeliness represents the timeliness of warning, false_alarm_rate_ represents the false alarm rate, coverage represents the warning coverage rate, and w1, w2, w3 and w4 are weight coefficients. The design of this reward function takes into account multiple key performance indicators of the geological disaster early warning system. Among them, accuracy and timeliness are the most basic requirements, so w1 and w2 are usually given larger weights (for example, 0.3-0.4). The false alarm rate is also an important factor. Too high a false alarm rate may cause the public to lose confidence in the early warning system, so w3 should also have a certain weight (for example, 0.2-0.3). The coverage rate reflects the comprehensiveness of the system. Although it is relatively minor, it cannot be ignored. W4 can be set to a smaller value (for example, 0.1-0.2).

[0197] Specifically, each indicator can be calculated in the following way:

[0198]

[0199]

[0200]

[0201]

[0202] Among them, TP, TN, FP, and FN represent the true positive, true negative, false positive, and false negative prediction results, respectively.

[0203] In the actual operation process, the model of the present invention will continuously adjust these weight coefficients according to the warning results and the actual disaster situation to achieve continuous optimization of performance. For example, if it is found that the system has a lot of false alarms, the value of w3 will be automatically increased, so that it will be more cautious in subsequent decisions. It is worth noting that the reward function of the present invention also contains an implicit time discount factor γ (0<γ<1). Specifically, the cumulative reward R at time t t It can be expressed as:

[0204] R t =r t +γr t+1 +γ 2 r t+2 +...,

[0205] Among them, r t is the immediate reward at time t. This design reflects the principle that the near term is more important in geological disaster warning, that is, the system pays more attention to the warning effect in the near term. In practical applications, the value of γ is usually between 0.9 and 0.99, which can be adjusted according to specific circumstances.

[0206] Through the above detailed introduction, it can be seen that the model of the present invention adopts innovative methods in seismic wave data processing, time graph convolution operation and reward function design. These innovations not only improve the prediction accuracy and reliability of the system, but also enhance the adaptability and interpretability of the system. In particular, in the design of the reward function, the present invention fully considers multiple key indicators of geological disaster warning and introduces an adaptive adjustment mechanism, which is of great significance for improving the practicality and sustainability of the system.

[0207] In practical applications, the model of the present invention can be customized according to the geological characteristics and risk types of different regions. For example, in earthquake-prone areas, the weight of the seismic wave data processing module can be increased, and the time window size of the time graph convolution can be adjusted; in landslide-prone areas, more attention can be paid to the long-term deformation trend, and the value of the time discount factor γ can be increased accordingly. This flexibility enables the system to adapt to various complex geological environments and provide strong support for different types of geological disaster warnings.

[0208] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A spatiotemporal series prediction and risk assessment model for geological disaster monitoring data, characterized by: include: Data acquisition module for: Acquire historical time series data related to the occurrence of geological disasters, wherein the historical time series data includes physical parameters, meteorological parameters, geological attributes and survey terrain data; Performing data cleaning on the historical time series data and generating samples; A dynamic spatiotemporal graph neural network module is communicatively connected with the data acquisition module and is used to: Receiving the cleaned historical time series data samples sent by the data acquisition module; Based on the historical time series data samples, a dynamic spatiotemporal graph neural network model DST-GNN is constructed; The multimodal data fusion module is communicatively connected with the dynamic spatiotemporal graph neural network module and is used to: Receive real-time monitoring data from multiple sensors, including InSAR imagery, groundwater levels, infrasound, and rainfall data; Integrating the real-time monitoring data into a four-dimensional feature tensor; The risk prediction module is connected to the multimodal data fusion module and the dynamic spatiotemporal graph neural network module for: Receiving the four-dimensional feature tensor sent by the multimodal data fusion module; Using the dynamic spatiotemporal graph neural network model DST-GNN to process the four-dimensional feature tensor to generate a prediction result; A meta-reinforcement learning module is communicatively connected to the risk prediction module and is used to: Receiving the prediction result sent by the risk prediction module; Based on the prediction results, construct a risk decision-making intelligent agent; Using quantum optimization technology to train the risk decision-making agent; A risk assessment module, in communication with the meta-reinforcement learning module, is used to: Receiving the risk decision-making agent trained by the meta-reinforcement learning module; Based on the risk decision-making agent, generating a multi-scale risk assessment result; A visual output module is connected to the risk assessment module for: Receiving a multi-scale risk assessment result generated by the risk assessment module; Generate visual displays including predicted displacement fields, rainfall accumulation curves, geotechnical parameter probability distributions, and warning levels; Output bilingual risk reports.

2. The model according to claim 1, characterized in that The dynamic spatiotemporal graph neural network module includes: Adaptive graph structure modeling unit, used for: Construct dynamic graph structures of geological body displacement correlation, rainfall propagation paths, and seismic wave attenuation characteristics; Dynamically update the adjacency matrix based on the time-series deformation data of the monitoring points; Multimodal data processing unit for: Design a four-dimensional feature tensor to integrate temporal, spatial, and multi-modal data, including deformation / hydrological / geoacoustic / meteorological data; Design specialized calculation formulas for different types of data; Interpretable risk prediction unit for: Combine temporal graph convolution, spatial graph convolution and comprehensive convolution to process 4D data; Introducing the gate vector mechanism to achieve dynamic update of graph structure; Physical constraint layer unit, used for: The Mohr-Coulomb criterion is used to embed the physical laws of geotechnical mechanics into the neural network; Add physical rule penalty terms to the loss function.

3. The model according to claim 1, characterized in that The meta-reinforcement learning module includes: State space building blocks for: Construct a state space containing the displacement of the prediction point, the rainfall intensity accumulation curve and the probability distribution of the rock and soil parameters; Action space definition unit, used to: Define an action space that includes risk levels and risk heatmaps; Reward function design unit, used to: Design a reward function based on the predicted displacement field and rainfall accumulation curve; Quantum Optimization Unit for: Use quantum computing technology to globally adjust model parameters; Introduce quantum gating functions to enhance model interpretability.

4. The model according to claim 1, characterized in that The risk assessment module includes: Predicted displacement field generation unit, used to: Use the DST-GNN model output to predict the displacement field; Perform spatial interpolation on the predicted displacement field to visualize the landslide surface deformation field; Rainfall accumulation curve fitting unit, used for: Fitting rainfall curves using neural networks; Generate historical precipitation intensity curves; Geotechnical parameter probability distribution calculation unit, used for: Based on soil moisture, water content, shear strength and rainfall field data, a probability graphical model is used to obtain the probability distribution of rock and soil elastic parameters; Warning level determination unit, used for: Use meta-learning training strategies to generate risk decision-making agents; The warning level is determined by taking into account a variety of factors; Generate a risk heat map.

5. The model according to claim 1, characterized in that Also includes: The spatiotemporal significance analysis module is in communication with the risk prediction module and the risk assessment module and is used to: Analyze the spatiotemporal correlation of monitoring parameters; Identify the factors that have the greatest impact on the forecast results; A causal reasoning module, in communication with the spatiotemporal significance analysis module, is used to: Mining key decision nodes of monitoring data; Analyze the evolution process of geological disasters.

6. The model according to claim 1, characterized in that The multimodal data fusion module uses the following formula to process deformation data: X deform =α·D InSAR +β·D GPS +γ·D inclinometer , Among them, D InSAR , D GPS and D inclinometer represent the deformation data of InSAR, GPS and inclinometer respectively, α, β and γ are weight coefficients, satisfying α+β+γ=1.

7. The model according to claim 1, characterized in that The multimodal data fusion module uses the following formula to process hydrological data: H t =f(W t ,G t ,P t ), Among them, W t is the groundwater level at time t, G t is the groundwater flow velocity, P t is the rainfall, and f is the nonlinear mapping function.

8. The model according to claim 1, characterized in that The multimodal data fusion module processes seismic wave data using the following formula: S(tx,y,z)=A(f)·e -α(f)r ·e i(2πft-kr) , Where S(t,x,y,z) represents the amplitude of the seismic wave at time t and spatial position (x,y,z), A(f) is the initial amplitude, f is the frequency, α(f) is the frequency-dependent attenuation coefficient, r is the source distance, and k is the wave number.

9. The model according to claim 1, characterized in that The risk prediction module uses the following formula to perform time graph convolution operation: Among them, H t is the hidden state at time t, A t -k is the adjacency matrix at time tk, X t -k is the node feature matrix at time tk, W k is the learnable weight matrix, K is the time window size, σ is the activation function, α k is the attention weight.

10. The model according to claim 1, characterized in that The risk assessment module calculates the reward function using the following formula: R=w1·accuracy+w2·timeliness-w3·false_alarm_rate+w4·coverage, Among them, accuracy represents the accuracy of risk prediction, timeliness represents the timeliness of warning, false_alarm_rate represents the false alarm rate, coverage represents the warning coverage rate, and w1, w2, w3 and w4 are weight coefficients.

Images