Geological deformation early warning method and system based on AI vision

Abstract

The invention relates to the technical field of geological disaster early warning, and discloses a geological deformation early warning method and system based on AI vision, and the method comprises the steps: obtaining surface micro-deformation features through an AI vision system, and constructing a geological map structure network; fusing a nonlinear geomechanical equation in a graph neural network to realize physical constraint driven feature extraction; creating a multi-dimensional graph structure model representing different depth layers, and realizing cross-layer information transmission through a physical information guide type graph attention mechanism; deploying a graph neural network of a Bayesian probability framework to carry out geological stress field distribution modeling, and quantizing prediction uncertainty; and finally, generating intelligent early warning decision information with risk grade division. According to the method, the AI vision and the physical constraint graph neural network are combined, accurate inference of the underground stress field is realized, the uncertainty can be quantitatively predicted, and the reliability and timeliness of geological disaster early warning are improved.

Description

Technical Field

[0001] The present invention relates to the field of geological disaster early warning technology, and more specifically, to a geological deformation early warning method and system based on AI vision. Background Art

[0002] Geological deformation early warning is a key step in geological disaster prevention and control, and is of great significance to protecting people's lives and property. With the development of science and technology, geological deformation monitoring technology has evolved from traditional manual observation to automated monitoring and then to intelligent monitoring.

[0003] Early geological deformation monitoring relied primarily on manual inspections and simple surveying tools, such as levels and theodolites, to assess geological stability by regularly measuring surface displacement. This approach, limited by human resources and measurement frequency, struggled to achieve continuous monitoring and only provided limited surface information.

[0004] With advances in sensor technology, automated monitoring systems are increasingly being used in geological deformation monitoring. These systems deploy sensors such as displacement meters, inclinometers, and strain gauges to continuously monitor surface deformation. However, these methods are still limited to surface or shallow monitoring, making it difficult to reveal stress changes and potential risks deep underground.

[0005] In recent years, the rapid development of computer vision technology has provided new technical means for geological deformation monitoring. High-precision cameras, laser scanners, and other equipment can capture subtle surface deformation features. Furthermore, the application of artificial intelligence and deep learning technologies has made it possible to extract effective information from massive amounts of visual data.

[0006] However, geological deformation is a complex three-dimensional process, making it difficult to accurately infer deep-subsurface stress changes and potential risks based solely on surface observation data. Furthermore, the complexity and uncertainty of the geological environment pose challenges to early warning decision-making. Therefore, combining advanced AI vision technology with geomechanics theory to accurately infer underground stress fields and provide reliable early warnings has become a key challenge in the field of geological disaster early warning. Summary of the Invention

[0007] The present invention provides a geological deformation early warning method and system based on AI vision, which solves the technical problems in related technologies such as the difficulty in accurately inferring the distribution of deep underground stress fields by relying solely on surface observation data, insufficient early warning timeliness, lack of uncertainty quantification of prediction results, and imperfect early warning decision support.

[0008] The present invention provides a geological deformation early warning method based on AI vision, comprising the following steps: Use high-precision AI vision systems to obtain surface micro-deformation features and automatically construct a geological map structure network with topological relationships; Based on the constructed geological map structure network, nonlinear geomechanical equations are integrated into the deep graph neural network architecture to implement physical constraint-driven graph convolution operations and feature extraction; By leveraging features extracted from physically constrained graph convolutions, we create a multi-dimensional graph structure model representing different depths of the underground. This allows for precise cross-layer geological information transfer through an innovative, physically guided graph attention mechanism. Based on a multi-level graph structure model, a graph neural network with a Bayesian probability framework is deployed to model the geological stress field distribution, and variational inference technology is combined to quantify the uncertainty of the prediction results; Based on the inferred geological stress field distribution and its uncertainty assessment results, the stress field distribution characteristics and uncertainty assessment indicators are integrated to generate intelligent early warning decision-making information with risk level classification.

[0009] In a preferred embodiment, the step of using a high-precision AI visual system to obtain surface micro-deformation features and automatically constructing a geological map structure network with topological relationships includes: Collect multi-dimensional surface deformation data; Perform noise reduction, registration and enhancement processing on the collected raw visual data to extract key deformation features; The monitoring area is represented as a graph structure, where nodes represent key monitoring points and edges represent physical constraint relationships between nodes.

[0010] In a preferred embodiment, the step of implementing physical constraint-driven graph convolution operation and feature extraction further includes: Encoding geomechanical equations as constraint parameters to guide the message passing process of graph neural networks; updating node characteristics based on received messages; A physical consistency loss function is introduced to ensure that the network inference results conform to the laws of geomechanics.

[0011] In a preferred embodiment, the step of creating a multi-dimensional graph structure model representing different depth layers of the underground includes: The underground structure of the monitoring area is divided into multiple depth layers, and each layer is represented as a graph structure; Connect each layer of graph structure to each other through vertical connection edges to form a three-dimensional graph structure; Implement physical information-guided graph attention computation, where a physical constraint matrix is introduced to guide the allocation of attention weights.

[0012] In a preferred embodiment, the step of deploying a graph neural network with a Bayesian probability framework to perform geological stress field distribution modeling includes: Convert the deterministic parameters of the graph neural network into random variables and introduce prior distribution; Estimate parameter posterior distributions via variational inference; Introducing uncertainty-aware graph attention computation; Estimation of prediction uncertainty of stress field distribution via Monte Carlo sampling.

[0013] In a preferred embodiment, the step of generating intelligent early warning decision information with risk level classification includes: Comprehensively consider stress field distribution, stress gradient, and uncertainty level factors to construct risk assessment indicators; Focus monitoring on areas with higher forecast uncertainty; Based on risk assessment indicators and preset thresholds, the monitoring area is divided into different risk levels and early warning signals of corresponding levels are triggered; Combined with geomechanical models, the location and development trend of potential fault zones are predicted based on the inferred stress field distribution.

[0014] In a preferred embodiment, the step of introducing uncertainty-aware graph attention calculation adjusts the attention weight by introducing an edge uncertainty factor. The edge uncertainty factor is calculated based on the variance of the attention weight, so that the impact of edge connections with higher uncertainty in the message transmission process is appropriately reduced.

[0015] In a preferred embodiment, the potential fault zone location is predicted based on the principal stress difference and the Mohr-Coulomb failure criterion, and the area where fracture may occur is determined by comparing the maximum and minimum principal stress differences with parameters related to rock cohesion and internal friction angle.

[0016] In a preferred embodiment, variational inference is achieved by minimizing the KL divergence between the true posterior distribution and the variational approximation, and the optimization goal is to maximize the lower bound of evidence, thereby obtaining an approximate posterior distribution of the network parameters.

[0017] In a preferred embodiment, a geological deformation early warning system based on AI vision is used to implement a geological deformation early warning method based on AI vision, including: AI visual data acquisition and graph structure construction module, which uses a high-precision AI visual system to acquire surface micro-deformation features and automatically construct a geological graph structure network with topological relationships; The physical constraint graph convolution module, based on the constructed geological map structure network, integrates nonlinear geomechanical equations into the deep graph neural network architecture to implement physical constraint-driven graph convolution operations and feature extraction; A multi-level graph structure inference module, which uses features extracted by physically constrained graph convolution to create a multi-dimensional graph structure model representing different depths of the subsurface. This module also enables accurate cross-layer geological information transfer through an innovative physical information-guided graph attention mechanism. The Bayesian variational inference module, based on a multi-level graph structure model, deploys a graph neural network with a Bayesian probability framework to model the geological stress field distribution and combines variational inference technology to quantify the uncertainty of the prediction results; The risk perception warning decision-making module is used to integrate the stress field distribution characteristics and uncertainty assessment indicators based on the inferred geological stress field distribution and its uncertainty assessment results, and generate intelligent warning decision-making information with risk level classification.

[0018] The beneficial effects of the present invention are: Improve the timeliness of early warning: By combining AI vision technology with physical constraint graph neural networks, the present invention can infer the distribution of underground stress fields from tiny surface deformation features, significantly detecting precursors of geological disasters in advance. Compared with traditional methods, it has a longer early warning lead time, buying valuable time for disaster prevention and mitigation.

[0019] Enhanced prediction accuracy: The physical constraint graph convolution operation that integrates geomechanical equations ensures that the model inference results conform to the laws of geomechanics, significantly improving prediction accuracy, especially in predicting the location of potential fault zones.

[0020] Achieving deep inference: The multi-level graph structure model can characterize the geological conditions at different depths and achieve accurate cross-layer information transmission through the graph attention mechanism guided by physical information. This breaks through the limitation of traditional methods that are limited to surface monitoring and realizes effective inference of deep underground stress fields.

[0021] Quantifying uncertainty: The Bayesian graph neural network framework can quantify the uncertainty of prediction results, providing a reliable risk assessment basis for early warning decisions and avoiding the risk of misjudgment that may be brought about by traditional deterministic methods.

[0022] Strong adaptability: The method of the present invention can adapt to the characteristics of different geological environments and shows excellent early warning effects in various scenarios such as mountain landslides, mining slopes, engineering construction areas, underground spaces and seismic activity areas.

[0023] Efficient resource utilization: Compared with traditional methods that require a lot of drilling and physical detection, this invention mainly relies on a non-contact AI vision system to obtain data, which greatly reduces monitoring costs and improves resource utilization efficiency.

[0024] Improved decision support: The hierarchical early warning decision information generated based on stress field distribution and uncertainty assessment provides management departments with a scientific and intuitive decision-making basis, which helps to formulate more accurate disaster prevention and mitigation measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of a geological deformation early warning method based on AI vision of the present invention; Figure 2 It is a bar chart comparing the early warning lead time of different early warning methods of the present invention; Figure 3 It is a line graph showing the change of stress field inference accuracy at different depths versus monitoring time according to the present invention; Figure 4 It is a radar chart for comprehensive performance evaluation of the technical solution of the present invention; Figure 5 It is a scatter plot showing the degree of agreement between the predicted fault zone location of the present invention and the actual verification. DETAILED DESCRIPTION

[0026] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0027] At least one embodiment of the present invention discloses a geological deformation early warning method based on AI vision, such as Figure 1 As shown, the following steps are included: Step 1: Use a high-precision AI vision system to obtain surface micro-deformation characteristics and automatically construct a geological map structure network with topological relationships; The specific steps include: Step 1.1, collect multi-dimensional surface deformation data; By deploying high-precision camera arrays, laser scanners, InSAR and other AI vision systems, multi-dimensional feature data including surface displacement, crack development, and hydrological changes are collected.

[0028] The collection frequency is dynamically adjusted according to the risk level of the monitored area, and can be as high as once an hour in high-risk areas and once a day in low-risk areas.

[0029] In some embodiments, a drone aerial photography system and a ground-based fixed camera array can be used in combination to form a multi-angle, multi-scale monitoring network to improve the comprehensiveness and reliability of data collection.

[0030] Step 1.2, data preprocessing and feature extraction; The collected raw visual data is subjected to noise reduction, registration and enhancement processing to extract key deformation features, including displacement field, strain field, crack distribution, etc.

[0031] Under complex meteorological conditions, adaptive image enhancement algorithms can be used to process the raw data, including defogging, rain removal, snow removal, etc., to ensure the quality of visual data under severe weather conditions.

[0032] Step 1.3, construct the geological map structure; Representing the monitoring area as a graph structure , where the node set represents key monitoring points, edge sets Represents the physical constraint relationship between nodes.

[0033] For nodes , whose eigenvector Contains multidimensional observation data of the monitoring point; For the edge , whose eigenvector Representation node and The physical relationship parameters between them.

[0034] Step 1.4, initialize graph node and edge attributes; Based on geological prior knowledge and historical monitoring data, initial attribute values are assigned to nodes and edges in the graph structure to establish an initial geological state representation.

[0035] Step 2: Based on the constructed geological map structure network, nonlinear geomechanical equations are integrated into the deep graph neural network architecture to implement physical constraint-driven graph convolution operations and feature extraction; The specific steps include: Step 2.1, build a physical constraint encoder; Encoding geomechanical equations (such as Hooke's law, Mohr-Coulomb criterion, etc.) as constraint parameters , used to guide the message passing process of the graph neural network. Constraint parameters Node-based and The physical relationship between them is defined, including geomechanical parameters such as elastic modulus, Poisson's ratio, and internal friction angle.

[0036] For Hooke's law, the relationship between stress and strain in an elastic body can be expressed as: ; in, is the stress tensor, which represents the distribution of internal forces acting on the material; is the strain tensor, which represents the measure of material deformation; is the elastic constant tensor, also known as the stiffness tensor, which characterizes the mechanical properties of the material and determines the linear relationship between stress and strain. These physical relationships are encoded as edge constraint parameters Part of , used to represent the physical constraint relationship between nodes in graph neural networks.

[0037] In some embodiments, according to the characteristics of different rock and soil masses, more complex rock and soil constitutive models such as the Drucker-Prager yield criterion or the Hoek-Brown criterion may be introduced to more accurately describe the nonlinear deformation behavior.

[0038] Step 2.2, implement the physical constraint graph convolution operation; The message passing process of the standard graph convolution operation is modified to incorporate physical constraint information. The message passing function is expressed as: ; in, Represents a slave node Pass to node The message vector of the node For Node The influence information is the core of information transmission between nodes in graph neural networks; It is a parameterized message function, which represents the slave node To Node The message passing function is composed of the neural network parameters Sure; is the source node The characteristic vector of contains the multi-dimensional observation data of the monitoring point, such as displacement, strain, etc. Target node The eigenvector of To connect nodes and nodes The edge features represent the spatial relationship parameters (such as distance, direction, etc.) between two monitoring points; For physical constraint parameters, the node is encoded and nodes The geomechanical constraints between them (such as geomechanical parameters such as elastic modulus, Poisson's ratio, internal friction angle, etc.).

[0039] Message Function The specific implementation is a multi-layer perceptron (MLP), whose input is the concatenation of node features, edge features and physical constraint parameters, and the output is a message vector: ; in, is the weight matrix of the first layer linear transformation, which is used to map the input features to the hidden layer space; is the weight matrix of the second layer linear transformation, which is used to map the hidden layer features to the output space; is the bias vector of the first layer, which is used to adjust the flexibility of the first layer transformation; is the bias vector of the second layer, used to adjust the final output; Represents the vector concatenation operation, which converts the node features 、 , edge features and physical constraint parameters Concatenate into a long vector as network input; Modified linear unit activation function, used to introduce nonlinear transformation capabilities and enhance the model's ability to express complex geological relationships; It is a parameterized message function, which represents the slave node To Node The message passing function is composed of the neural network parameters Sure; Source node The eigenvector of Target node The eigenvector of To connect nodes and nodes The edge feature represents the spatial relationship parameter between two monitoring points; For physical constraint parameters, the node is encoded and nodes The geomechanical constraints between them.

[0040] Step 2.3, node feature update; Update node features based on received messages. The update rules are: ; in, For the Layer Node The updated feature vector represents the new state of the monitoring point after a graph convolution operation; For the Layer Node The current feature vector of contains the current multi-dimensional observation data of the monitoring point; For nodes The neighbor set of the monitoring point All directly connected monitoring points; From neighbor nodes Pass to node The message vector contains the neighbor nodes For Node Impact information; Indicates the aggregation of messages from all neighbors, that is, for nodes Sum the messages from all neighboring nodes; It is an update function responsible for fusing the current features of the node with the aggregated neighbor messages to generate a new feature representation of the node.

[0041] Update Function The specific implementation is the Gated Recurrent Unit (GRU), which fuses the current node features and aggregated messages: ; in, For the Layer Node The updated feature vector of represents the updated monitoring point status information; For the Layer Node The updated current feature vector represents the current monitoring point status information; For nodes The neighbor node set represents other monitoring points that are physically associated with the current monitoring point; From neighbor nodes Pass to node The message vector contains the status information of neighbor nodes and their relationship with the central node; Indicates the aggregation of messages from all neighboring nodes; It is a gated recurrent unit, a recursive neural network structure that controls the flow of information and effectively integrates the current node state with neighbor information to avoid long-term dependency problems.

[0042] Step 2.4, physical consistency check; Introducing physical consistency loss function , ensuring that the network reasoning results conform to the laws of geomechanics, is defined as: ; in, is the physical consistency loss function, which is used to measure the degree of conformity between the model prediction results and the physical laws; Represents all edges in the graph structure Perform summation, i.e. consider the physical constraints between all connected node pairs; is a constraint function based on physical laws, with input nodes and nodes The feature vector of , the output is the predicted value of the physical relationship that should be satisfied between the two nodes; and Node and nodes The characteristic vector of contains the multidimensional observation data of each monitoring point; For predefined physical constraint parameters, the nodes are encoded and nodes The geomechanical constraints that should be satisfied between them; Represents the square of the Euclidean norm, which is used to calculate the mean squared error between the predicted physical relationship and the predefined constraints.

[0043] The physical consistency loss and the prediction task loss together constitute the optimization objective of the model: ; in, is the total loss function, which is the ultimate goal of model optimization; The loss function is used to measure the deviation between the model prediction result and the actual observation value, usually using loss functions such as mean square error or cross entropy. This is the physical consistency loss, which is used to ensure that the model reasoning results conform to the laws of geomechanics; It is a balance parameter, a positive scalar value that controls the strength of the physical constraint. Larger values will make the model more focused on satisfying physical constraints, while smaller values allow the model to deviate from physical laws to a certain extent in order to better fit the observations.

[0044] Step 3: Using the features extracted by the physical constraint graph convolution, a multi-dimensional graph structure model is created to represent different depth layers of the underground. An innovative physical information-guided graph attention mechanism is used to achieve accurate cross-layer geological information transmission. The specific steps include: Step 3.1, construct a multi-level graph structure; The underground structure of the monitoring area is divided into multiple depth layers, each layer is represented as a graph ,in, Represents a depth index, which is an identifier of a specific depth layer; Indicates the depth The node set in the layer represents all monitoring points or calculation points on the depth layer; Indicates the depth The edge set in the layer represents the connection relationship between monitoring points in the same depth layer, reflecting the geological correlation in the horizontal direction.

[0045] Each layer of the graph structure is connected by vertical edges interconnected, where It represents the set of vertical connection edges between different depth layers, and is used to establish the association between nodes at different depth layers. It reflects the geomechanical conduction relationship in the vertical direction, thus forming a complete three-dimensional graph structure.

[0046] Based on geomechanics theory, the 0-50m underground range is divided into five depth layers: 0-10m, 10-20m, 20-30m, 30-40m, and 40-50m. The node distribution in each depth layer is determined by the surface monitoring point grid and geological structure characteristics. Nodes between adjacent depth layers are connected by vertical edges.

[0047] In some embodiments, the depth layer division may be non-uniform according to different geological environment characteristics, for example, finer divisions (such as 0 to 5 meters, 5 to 10 meters, 10 to 15 meters, etc.) may be used in the shallow part to obtain a more refined shallow stress field distribution.

[0048] Step 3.2: Implement graph attention calculation guided by physical information; The information transfer weight between attention calculation nodes is introduced, and the attention calculation formula is: ; in, Represents the attention calculation function guided by physical information; is the query matrix, which represents the feature representation of the current node and is used to query the association strength with other nodes; is the key matrix, which represents the feature representation of the target node and is used to calculate the attention score with the query matrix; is a value matrix containing the information content of the target node, which is used to aggregate information according to the attention weight; is the bond matrix The transposed matrix of The physical constraint matrix encodes the physical correlation strength between nodes based on geomechanics theory, which is used to guide the attention mechanism to focus on more physically relevant nodes. To balance the parameters and control the influence of physical constraints on attention calculation, larger values will make the model rely more on physical constraints, while smaller values will rely more on data-driven feature similarity; is the feature dimension, which represents the dimension of the key vector and is used to scale the dot product result to avoid the gradient disappearance problem; is a normalization function that converts the attention scores into probability distributions, ensuring that the sum of all attention weights is 1.

[0049] Physical Constraint Matrix Contains the physical connection strength between nodes, which is calculated based on geomechanics theory: ; in, Represents a node in the physical constraint matrix and nodes the strength of the physical connection between elements; For nodes and nodes The spatial distance between them is used to characterize the spatial position relationship between the two monitoring points; and Node and nodes The density of the rock and soil at the location reflects the quality characteristics of the rock and soil; and Node and nodes The elastic modulus at the location indicates the ability of the rock mass to resist elastic deformation; and Node and nodes The Poisson's ratio at the location represents the ratio of the deformation of the rock mass in the direction of force to the deformation perpendicular to the direction of force; It is a physical correlation function, which is designed based on the principles of geomechanics. It is used to convert various physical parameters into the correlation strength between nodes. The larger the output value, the stronger the physical correlation between the two nodes.

[0050] Step 3.3, inter-layer information transmission and status update; By transmitting inter-layer information through vertical connection edges, surface observation information can be transmitted to deep underground layers, and the node status of each depth layer can be updated.

[0051] The information transmission process adopts a top-down and bottom-up bidirectional propagation mechanism: ; in, For depth Layer Node In time The updated state represents the latest feature representation of the monitoring point in the depth layer; For depth Layer Node In time The current state of represents the current feature representation of the monitoring point in the depth layer; For depth Layer (lower layer) node In time The state of , which represents the characteristic representation of the related monitoring points in the lower layer; For depth Layer (upper layer) node In time The state of , which represents the feature representation of the upper layer related monitoring points; It is a bottom-up information aggregation function responsible for transferring the information of the lower layer nodes to the current layer, capturing the impact of the deep underground layer on the shallow layer; It is a top-down information aggregation function responsible for transferring the upper layer node information to the current layer and capturing the impact of the shallow surface layer on the deep layer; is the state update balance parameter, the value range is , controls the ratio of retaining the original state information and integrating the new information. A larger value means more reliance on the new aggregated information; is the direction balance parameter, and its value range is , controls the fusion ratio of bottom-up information and top-down information. A larger value means that the influence of lower-level information is considered more.

[0052] Step 3.4, stress field distribution inference; Based on the updated multi-level graph structure, the stress field distribution within the range of 10 to 50 meters underground is inferred and the stress tensor field is output. , where each spatial position corresponds to a stress tensor.

[0053] The stress tensor consists of 6 independent components ( , , , , , ), obtained through the final output layer of the graph neural network: ; in, For nodes The predicted value of the stress tensor at , which contains 6 independent components describing the stress state at this point; is the final layer node feature, indicating that Nodes after layer graph neural network processing The feature vector encodes the comprehensive status information of the monitoring point; is the output layer weight matrix, which is the linear transformation parameter that maps the node features to the stress tensor space; It is the output layer bias vector, which is used to adjust the baseline level of the predicted value to ensure that the stress tensor output by the model is within a reasonable range.

[0054] Among the components of the stress tensor, 、 and Respectively indicate along 、 and Normal stress in the axial direction; 、 and Respectively 、 and Shear stress on a plane.

[0055] The stress tensor at any location in space is obtained by interpolation: ; in, For spatial location The stress tensor at represents the stress state at that location; For location The nearby node set includes all computational nodes that are close to the location and have an impact on the stress state of the location; For nodes The stress tensor at represents the stress state at the node; is a distance-based weight function that determines the node Position The degree of influence of stress state and node To location The closer the distance, the greater the weight; Indicates position The weighted summation of all nearby nodes enables interpolation calculation of the stress state at any position in space.

[0056] Step 4: Based on the multi-level graph structure model, a graph neural network with a Bayesian probability framework is deployed to model the geological stress field distribution, and variational inference technology is combined to quantify the uncertainty of the prediction results; The specific steps include: Step 4.1, constructing a Bayesian graph neural network; The deterministic parameters of the graph neural network Convert to random variables and introduce prior distribution , select Standard normal distribution.

[0057] For each network parameter , and define its prior distribution as: ; in, Represents the set of weight parameters in the graph neural network, which are all learnable weight parameters in the model; Represents the first A weight parameter is a single learnable network parameter; Representation parameters The prior probability distribution of , describing the prior belief about a single parameter; The mean is 0 and the variance is The Gaussian distribution of is the prior distribution form of the parameter; It is the prior variance, which controls the degree of discreteness of the parameter distribution. A larger value allows the parameter to take a wider range of values, while a smaller value makes the parameters concentrated near the mean, thereby controlling the regularization strength of the parameters.

[0058] Step 4.2, implement the variational inference algorithm; Estimate the parameter posterior distribution through variational inference and minimize the true posterior distribution Variational approximation The KL divergence between , the optimization goal is: Equivalent to maximizing the evidence lower bound (ELBO): ; in, For a given observation data (graph structure) and Network parameters under (node characteristics) conditions The true posterior distribution of , which represents the probability estimate of the model parameters after observing specific data; is a parameterized variational distribution used to approximate the true posterior distribution; is the parameter set of the variational distribution; Kullback-Leibler divergence is used to measure the difference between two probability distributions. The smaller the value, the more accurate the approximation. is the variational parameter, which is the set of target parameters that need to be adjusted during the optimization process; is the lower bound of the evidence, which is the lower bound of the log marginal likelihood. Maximizing ELBO is equivalent to minimizing KL divergence; For the variational distribution Lower log-likelihood The expected value of , which indicates how well the model fits the observed data; represents the natural logarithm function; is the KL divergence between the variational distribution and the prior distribution, which serves as a regularization term to prevent overfitting; is the parameter prior distribution, which represents the prior assumptions about the model parameters before observing the data; Represents the set of weight parameters in the graph neural network, which are all learnable weight parameters in the model; Represents the first A weight parameter is a single learnable network parameter.

[0059] Variational approximation Using the mean field approximation, the posterior distribution is decomposed into independent Gaussian distributions: ; in, is a parameterized variational distribution used to approximate the true posterior distribution; is the parameter set of the variational distribution; Represents the set of weight parameters in the graph neural network, which are all learnable weight parameters in the model; Indicates that for all parameters Perform multiplication operations; Represents the first A weight parameter is a single learnable network parameter; For the The mean of the network parameters represents the optimal estimate of the parameters; For the The variance of the network parameters indicates the uncertainty of the parameter estimation; the two together constitute the learnable variational parameters , which is continuously adjusted through the optimization process to approach the true posterior distribution.

[0060] The optimization process uses the reparameterization technique to convert random samples into a deterministic function plus noise: ; in, Represents the first A weight parameter is a single learnable network parameter; represents random noise sampled from a standard normal distribution with a mean of 0 and a variance of 1; For the The mean of the network parameters represents the optimal estimate of the parameters; is the parameter standard deviation, which controls the fluctuation range of the sampling results; It is a sample of network parameters obtained after reparameterization, which is used for forward propagation calculation and gradient back propagation.

[0061] In some embodiments, to improve computational efficiency, a more complex family of variational distributions such as flow-based Normalizing Flow or Matrix Variate Gaussian can be used to improve the expressive power of the posterior approximation.

[0062] Step 4.3, introduce uncertainty-aware graph attention calculation; By propagating parameter uncertainty, the uncertainty of the edge attention weight is calculated, and the attention calculation formula is: ; in, For nodes For neighbor nodes The attention weight of the node in the process of information transmission For Node the extent of the impact; is the attention vector, which is a learnable model parameter used to calculate the relevance score between node pairs; is the attention weight matrix, which is used to perform linear transformation on node features; and Node and nodes The feature vector of contains the node status information; Indicates that the transformed node and nodes The feature vector concatenation operation; It is a leaky rectified linear unit activation function, which is used to introduce nonlinearity and prevent gradient disappearance; Representation node The neighbor node set of node All directly connected nodes; Indicates the node The sum of all neighbor nodes is used to normalize the attention weight; is the edge uncertainty factor, reflecting the model's uncertainty of the edge The confidence level of the importance judgment. A higher value indicates that the model has lower uncertainty about the importance judgment of the edge. Represents the attention vector The transpose of Represents the exponential function.

[0063] Edge uncertainty factor Variance calculation based on attention weights: ; in, For the edge The uncertainty factor indicates the credibility of the model's judgment of the importance of the edge. The larger the value, the lower the uncertainty. is the attention weight estimated by Monte Carlo sampling The variance of quantifies the degree of fluctuation of the model's estimate of the attention weight of this edge. A larger value indicates higher uncertainty. It is a scaling parameter that controls the sensitivity of uncertainty to edge weights. Larger values make the model more sensitive to uncertainty, while smaller values reduce the impact of uncertainty. is an exponential function that converts negative weighted variance into The larger the variance is, the closer the uncertainty factor is to 0, thus reducing the importance of the edge in information transmission.

[0064] Step 4.4, quantify the prediction uncertainty; The prediction uncertainty of the stress field distribution is estimated by Monte Carlo sampling, for the spatial location The stress tensor , whose predicted mean and variance are: ; ; in, For spatial location The predicted mean of the stress tensor at represents the model’s best estimate of the stress state at that location; For spatial location The predicted variance of the stress tensor at quantifies the uncertainty of the prediction results; For the The stress tensor prediction obtained by Monte Carlo sampling represents the output of the model under a single sampling of network parameters; is the number of Monte Carlo sampling, which determines the sample size of the uncertainty estimation.

[0065] Sampling times Choosing 20-50 requires a balance between computational efficiency and estimation accuracy in practical applications. Prediction uncertainty can also be decomposed into epistemic uncertainty (model parameter uncertainty) and random uncertainty (intrinsic randomness of the data): ; in, is the total forecast uncertainty, which indicates the overall variability of the forecast results; Epistemic uncertainty stems from the uncertainty of model parameters, reflecting the insufficiency of training data or the limitations of model structure, which can be reduced by increasing training data; It is random uncertainty, which comes from the inherent randomness of the data itself and measurement noise. It cannot be eliminated by increasing training data and reflects the inherent randomness of the system.

[0066] In some implementations, a deep ensemble method (DeepEnsemble) can be used as a supplement to the Bayesian method to further improve the reliability of uncertainty estimation by training multiple models with different initializations and data sampling.

[0067] Step 5: Based on the inferred geological stress field distribution and its uncertainty assessment results, the stress field distribution characteristics and uncertainty assessment indicators are integrated to generate intelligent early warning decision information with risk level classification; The specific steps include: Step 5.1, establish a risk assessment indicator system; Taking into account factors such as stress field distribution, stress gradient, and uncertainty level, a risk assessment index is constructed to quantitatively characterize the potential geological hazard risks.

[0068] Risk Assessment Indicators The calculation formula is: ; in, It is a comprehensive risk assessment indicator that indicates the degree of geological hazard risk in a specific area, with higher values indicating greater risk. It is a stress level index that quantifies the degree of proximity between the stress state in the region and the critical failure stress, reflecting the stress condition of the region; It is a stress gradient index that characterizes the spatial change rate of the stress field. A larger stress gradient often means a stress concentration area, which is a potential source of damage. It is an uncertainty index that reflects the confidence of the model in predicting the stress field. Higher uncertainty means that the prediction results are less reliable and require additional attention. It is a time evolution indicator that describes the trend and rate of change of the stress field over time. Areas with rapid changes usually have higher risks. 、 、 、 Represent the weight coefficients of stress level index, stress gradient index, uncertainty index and time evolution index respectively, satisfying To ensure consistency and comparability of assessments.

[0069] Step 5.2, implement key area monitoring; Focus on monitoring areas with higher forecast uncertainty, increase the frequency of data collection, and reduce blind spots in the early warning system.

[0070] The identification of key monitoring areas is based on uncertainty thresholds: ; in, Represents the set of spatial areas that need to be monitored, including all spatial points that meet the uncertainty conditions; Represents a coordinate point in three-dimensional space, used to locate a specific monitoring location; Indicates spatial location The predicted variance of the stress tensor at quantifies the uncertainty of the stress prediction at that location; The uncertainty threshold is a preset critical value used to determine whether the uncertainty is high enough to require focused monitoring and is dynamically adjusted according to risk tolerance; The symbol represents the conditional restrictions in the set definition. The left side is the set element, and the right side is the condition that the element needs to meet.

[0071] In some embodiments, active learning strategies can be used to adaptively adjust the allocation of monitoring resources based on risk assessment indicators and historical data, allocating more monitoring resources to high-risk and high-uncertainty areas.

[0072] Step 5.3, generate graded warning decisions; Based on risk assessment indicators and preset thresholds, the monitoring area is divided into different risk levels, and corresponding warning signals are triggered. Warning levels generally include normal, caution, warning, and danger.

[0073] The rules for determining the warning level are: Normal level: , indicating that the comprehensive risk assessment index is lower than the first threshold, the area is in a safe state and does not require special attention; Attention Level: , indicating that the comprehensive risk assessment index exceeds the first threshold but is lower than the second threshold, and the region has a slight abnormality, requiring increased monitoring frequency; Warning Level: , indicating that the comprehensive risk assessment index exceeds the second threshold but is lower than the third threshold, and there is a clear risk in the area, requiring the activation of the emergency plan and preparation for evacuation of personnel; Danger Level: , indicating that the comprehensive risk assessment index exceeds the third threshold, the area is in an extremely high-risk state, and the emergency plan needs to be implemented immediately and personnel evacuated; in, The regional geological hazard risk level was quantified for comprehensive risk assessment indicators; The first risk threshold marks the critical point from normal status to the point where attention is required; The second risk threshold marks the critical point from the attention state to the need for warning; The third risk threshold marks the critical point where a warning state turns into a dangerous state. These thresholds are determined based on statistical analysis of historical disaster data and the experience and knowledge of geological experts, and can be adjusted appropriately based on the geological conditions of different regions.

[0074] Step 5.4: predict the location and development trend of potential fault zones; Combined with geomechanical models, based on the inferred stress field distribution, the location and development trend of potential fault zones are predicted to provide decision support for disaster prevention and mitigation.

[0075] The identification of the fault zone location is based on the principal stress difference and the Mohr-Coulomb failure criterion: ; in, The set of spatial regions representing potential fault zones, including all spatial points that meet the failure criteria; Represents a coordinate point in three-dimensional space, used to locate the specific fracture location; For spatial location The maximum principal stress at a point represents the maximum tensile or compressive stress at that point; For spatial location The minimum principal stress at a point represents the minimum tensile or compressive stress at that point; is the principal stress difference, indicating the magnitude of the shear stress; The cohesion is the ability of the material to resist shear failure under zero normal stress, and the unit is usually MPa; is the internal friction angle, a parameter that represents the internal friction characteristics of the material, with the unit of degree, reflecting the growth rate of the shear strength of the material when the normal stress increases; is the internal friction coefficient, which expresses the proportional relationship between normal stress and the resulting increase in shear strength; The symbol represents the conditional restrictions in the set definition. The left side is the set element, and the right side is the condition that the element needs to meet.

[0076] The fracture development direction is perpendicular to the direction of the maximum principal stress and is determined by the eigenvector of the principal stress field.

[0077] Application examples of this implementation: This method has been validated in a slope monitoring application at a mining area. Located in a complex geological structure, the slope stability is affected by multiple factors, making it difficult for traditional monitoring methods to accurately predict potential landslide risks.

[0078] The implementation process of this method is as follows: Visual Data Collection: A monitoring network consisting of 20 high-definition cameras and two laser scanners was deployed around the slope. Cameras collected data once an hour, and laser scanners every six hours. After three months of data collection, over 2,160 hours of monitoring data were accumulated.

[0079] Graph Structure Construction: Based on the geological characteristics of the monitoring area, the area is divided into 350 monitoring nodes, and the graph structure is constructed through physical constraints. Each node contains 15-dimensional features, representing information such as displacement, strain, and cracks.

[0080] Physical Constraint Graph Convolutional Modeling: Based on the mechanical parameters of the mining area's geotechnical mass (average elastic modulus E = 2.5 GPa, Poisson's ratio ν = 0.28, internal friction angle φ = 30°), a physical constraint encoder was constructed to integrate physical constraints into the graph neural network. The model uses a three-layer graph convolutional architecture with a hidden layer dimension of 64.

[0081] Multi-layer graph structure inference: The 0-50-meter underground range is divided into five layers, constructing a three-dimensional graph structure. A physics-guided graph attention model is used to transfer information between layers. The model successfully inferred the underground stress field distribution, specifically identifying three potential high stress concentration areas within the 10-30-meter depth range.

[0082] Bayesian variational inference and uncertainty quantification: A Bayesian graphical neural network was used to probabilistically model the stress field distribution and quantify the prediction uncertainty through variational inference. Using a Monte Carlo sampling algorithm of 30, five regions of high uncertainty were successfully identified, three of which overlapped with areas of high stress concentration.

[0083] Risk Perception and Early Warning Decision-Making: Based on stress field distribution and uncertainty assessment, a risk assessment index is constructed to categorize the monitored area into different risk levels. The system identified a potential fault zone on the northern slope, and the risk assessment index reached the warning level.

[0084] Technical effect verification: Improved early warning capabilities: This method successfully predicted accelerated deformation of the northern slope 42 days earlier than traditional monitoring methods (57 days compared to 15 days for traditional methods). This represents a 380% improvement in early warning times, demonstrating the effectiveness of the technology in achieving a 30-60% improvement in early warning times. The predicted location of the potential fault zone matched subsequent drilling verification at a rate of 82%, confirming the effectiveness of the technology in achieving an 80% prediction accuracy.

[0085] Enhanced decision reliability: During the three-month monitoring period, the traditional monitoring method produced five false alarms and two missed alarms, while this method produced only two false alarms and no missed alarms. The false alarm rate was reduced by (5-2) / 5=60%, and the missed alarm rate was reduced by 100%, verifying the technical effect of reducing the false alarm rate by 40% and the missed alarm rate by 50%.

[0086] Optimized resource utilization: Compared to traditional geophysical exploration methods that require drilling 12 monitoring holes, this method only requires the deployment of a visual monitoring system and three verification holes, reducing costs by (12-3) / 12 = 75%, demonstrating the technical effectiveness of a 65% reduction in monitoring costs. The monitoring coverage area has been expanded from 2.5 square kilometers with traditional methods to 6.8 square kilometers, a 172% increase.

[0087] like Figures 2 to 5 As shown, there are comparisons of early warning lead times of different early warning methods; changes in stress field inference accuracy at different depths with monitoring time; comprehensive performance evaluation of technical solutions; and consistency between predicted fault zone locations and actual verification.

[0088] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A geological deformation early warning method based on AI vision, characterized in that: The following steps are involved: Use high-precision AI vision systems to obtain surface micro-deformation features and automatically construct a geological map structure network with topological relationships; Based on the constructed geological map structure network, nonlinear geomechanical equations are integrated into the deep graph neural network architecture to implement physical constraint-driven graph convolution operations and feature extraction; By leveraging features extracted from physically constrained graph convolutions, we create a multi-dimensional graph structure model representing different depths of the underground. This allows for precise cross-layer geological information transfer through an innovative, physically guided graph attention mechanism. Based on a multi-level graph structure model, a graph neural network with a Bayesian probability framework is deployed to model the geological stress field distribution, and variational inference technology is combined to quantify the uncertainty of the prediction results; Based on the inferred geological stress field distribution and its uncertainty assessment results, the stress field distribution characteristics and uncertainty assessment indicators are integrated to generate intelligent early warning decision-making information with risk level classification.

2. The geological deformation early warning method based on AI vision according to claim 1 is characterized in that: The steps of using a high-precision AI visual system to obtain surface micro-deformation features and automatically construct a geological map structure network with topological relationships include: Collect multi-dimensional surface deformation data; Perform noise reduction, registration and enhancement processing on the collected raw visual data to extract key deformation features; The monitoring area is represented as a graph structure, where nodes represent key monitoring points and edges represent physical constraint relationships between nodes.

3. The geological deformation early warning method based on AI vision according to claim 1 is characterized in that: The step of implementing physical constraint-driven graph convolution operation and feature extraction also includes: Encoding geomechanical equations as constraint parameters to guide the message passing process of graph neural networks; updating node characteristics based on received messages; A physical consistency loss function is introduced to ensure that the network inference results conform to the laws of geomechanics.

4. The geological deformation early warning method based on AI vision according to claim 1 is characterized in that: The step of creating a multi-dimensional graph structure model representing different depth layers of the underground includes: The underground structure of the monitoring area is divided into multiple depth layers, and each layer is represented as a graph structure; Connect each layer of graph structure to each other through vertical connection edges to form a three-dimensional graph structure; Implement physical information-guided graph attention computation, where a physical constraint matrix is introduced to guide the allocation of attention weights.

5. The geological deformation early warning method based on AI vision according to claim 1 is characterized in that: The steps of deploying the graph neural network of the Bayesian probability framework to perform geological stress field distribution modeling include: Convert the deterministic parameters of the graph neural network into random variables and introduce prior distribution; Estimate parameter posterior distributions via variational inference; Introducing uncertainty-aware graph attention computation; Estimation of prediction uncertainty of stress field distribution via Monte Carlo sampling.

6. The geological deformation early warning method based on AI vision according to claim 1, characterized in that: The step of generating intelligent early warning decision information with risk level classification includes: Comprehensively consider stress field distribution, stress gradient, and uncertainty level factors to construct risk assessment indicators; Focus monitoring on areas with higher forecast uncertainty; Based on risk assessment indicators and preset thresholds, the monitoring area is divided into different risk levels and early warning signals of corresponding levels are triggered; Combined with geomechanical models, the location and development trend of potential fault zones are predicted based on the inferred stress field distribution.

7. The geological deformation early warning method based on AI vision according to claim 5 is characterized in that: The step of introducing uncertainty-aware graph attention calculation adjusts the attention weight by introducing an edge uncertainty factor. The edge uncertainty factor is calculated based on the variance of the attention weight, so that the influence of edge connections with higher uncertainty in the message transmission process is appropriately reduced.

8. The geological deformation early warning method based on AI vision according to claim 6 is characterized in that: The prediction of potential fault zone locations is based on the principal stress difference and the Mohr-Coulomb failure criterion. By comparing the maximum and minimum principal stress differences with parameters related to rock cohesion and internal friction angle, the area where fracture may occur is determined.

9. The geological deformation early warning method based on AI vision according to claim 1, characterized in that: Variational inference is achieved by minimizing the KL divergence between the true posterior distribution and the variational approximation. The optimization goal is to maximize the lower bound of evidence to obtain an approximate posterior distribution of the network parameters.

10. A geological deformation early warning system based on AI vision, used to execute the geological deformation early warning method based on AI vision according to any one of claims 1 to 9, characterized in that: include: AI visual data acquisition and graph structure construction module, which uses a high-precision AI visual system to acquire surface micro-deformation features and automatically construct a geological graph structure network with topological relationships; The physical constraint graph convolution module, based on the constructed geological map structure network, integrates nonlinear geomechanical equations into the deep graph neural network architecture to implement physical constraint-driven graph convolution operations and feature extraction; A multi-level graph structure inference module, which uses features extracted by physically constrained graph convolution to create a multi-dimensional graph structure model representing different depths of the subsurface. This module also enables accurate cross-layer geological information transfer through an innovative physical information-guided graph attention mechanism. The Bayesian variational inference module, based on a multi-level graph structure model, deploys a graph neural network with a Bayesian probability framework to model the geological stress field distribution and combines variational inference technology to quantify the uncertainty of the prediction results; The risk perception warning decision module is used to integrate the stress field distribution characteristics and uncertainty assessment indicators based on the inferred geological stress field distribution and its uncertainty assessment results, and generate intelligent warning decision information with risk level classification.

Images