Geological deformation early warning method and system based on AI vision

Abstract

The application relates to the technical field of geological disaster early warning, and discloses a geological deformation early warning method and system based on AI vision, wherein the geological deformation early warning method based on AI vision comprises the following steps: acquiring ground surface micro-deformation features and constructing a geological map structure network by using an AI vision system; fusing a nonlinear geomechanics equation into a graph neural network to realize feature extraction driven by physical constraints; creating a multi-dimensional graph structure model representing different depth layers, realizing cross-layer information transmission through a physical information guided graph attention mechanism; deploying a graph neural network of a Bayesian probability framework to model geological stress field distribution and quantify prediction uncertainty; and finally generating intelligent early warning decision information with risk level division; the application combines AI vision and a physically constrained graph neural network, realizes accurate inference of an underground stress field, can quantify prediction uncertainty, and improves the reliability and timeliness of geological disaster early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster early warning, more specifically, it relates to a geological deformation early warning method and system based on AI vision. BACKGROUND

[0002] Geological deformation early warning is a key link in geological disaster prevention and control, and it is of great significance to protect people's life and property safety. With the development of science and technology, geological deformation monitoring technology has experienced the evolution process from traditional manual observation to automatic monitoring and then to intelligent monitoring.

[0003] Early geological deformation monitoring mainly relies on manual patrol and simple measuring tools such as level, theodolite, etc., to judge the geological stability by periodically measuring the ground displacement. This method is limited by human resources and measurement frequency, and it is difficult to achieve continuous monitoring, and only limited surface information can be obtained.

[0004] With the advancement of sensor technology, automatic monitoring systems have gradually been applied to the field of geological deformation monitoring. This kind of system realizes the continuous monitoring of ground deformation by deploying displacement meters, tiltmeters, strain gauges and other sensing devices. However, these methods are still limited to surface or shallow monitoring, and it is difficult to reveal the stress changes and potential risks in the deep underground.

[0005] In recent years, the rapid development of computer vision technology has provided new technical means for geological deformation monitoring. Through high-precision cameras, laser scanners and other equipment, it can capture the characteristics of small deformation of the ground. At the same time, the application of artificial intelligence and deep learning technology makes it possible to extract effective information from massive visual data.

[0006] However, geological deformation is a complex three-dimensional process, and it is difficult to accurately infer the stress changes and potential risks in the deep underground by relying only on surface observation data. In addition, the complexity and uncertainty of the geological environment also bring challenges to early warning decision-making. Therefore, how to combine advanced AI vision technology with geomechanics theory to realize accurate inference and reliable early warning of the underground stress field has become a key problem to be solved in the current geological disaster early warning field. SUMMARY

[0007] The present application provides a geological deformation early warning method and system based on AI vision, which solves the technical problems in the related art that it is difficult to accurately infer the distribution of the deep underground stress field by relying only on surface observation data, the timeliness of early warning is insufficient, the prediction results lack uncertainty quantification, and the early warning decision support is imperfect.

[0008] The present application provides a geological deformation early warning method based on AI vision, comprising the following steps:

[0009] Utilize high-precision AI vision system to obtain surface micro-deformation features and automatically construct a geological map structure network with topological relationships;

[0010] Based on the constructed geological map structure network, nonlinear geomechanics equations are integrated into the deep graph neural network architecture to realize physically constrained graph convolution operations and feature extraction;

[0011] Using the features extracted by physically constrained graph convolution, a multi-dimensional graph structure model representing different depth layers underground is created, and an innovative physical information guided graph attention mechanism is used to achieve accurate cross-layer geological information transmission;

[0012] Based on the multi-level graph structure model, a graph neural network based on the Bayesian probability framework is deployed to model the geological stress field distribution, and the variational inference technique is used to quantify the uncertainty of the prediction results;

[0013] According to the inferred geological stress field distribution and uncertainty evaluation results, the stress field distribution features and uncertainty evaluation indicators are integrated to generate intelligent early warning decision information with risk level division.

[0014] In a preferred embodiment, the step of utilizing a high-precision AI vision system to obtain surface micro-deformation features and automatically constructing a geological map structure network with topological relationships comprises:

[0015] Collecting multi-dimensional surface deformation data;

[0016] Performing noise reduction, registration, and enhancement processing on the collected raw visual data to extract key deformation features;

[0017] Representing the monitoring area as a graph structure, where nodes represent key monitoring points and edges represent the physical constraint relationship between nodes.

[0018] In a preferred embodiment, the step of realizing physically constrained graph convolution operations and feature extraction further comprises:

[0019] Encoding geomechanics equations as constraint parameters to guide the message passing process of the graph neural network;

[0020] Updating node features based on received messages;

[0021] Introducing a physical consistency loss function to ensure that the network inference results comply with geomechanics laws.

[0022] In a preferred embodiment, the step of creating a multi-dimensional graph structure model representing different depth layers underground comprises:

[0023] Dividing the underground structure of the monitoring area into multiple depth layers, each layer represented as a graph structure;

[0024] Each layer of the graph structure is connected to each other through a vertical connection edge to form a three-dimensional graph structure.

[0025] Implementing physical information guided graph attention computation, where a physical constraint matrix is introduced to guide the allocation of attention weights.

[0026] In a preferred embodiment, the step of deploying a graph neural network of Bayesian probability framework for geological stress field distribution modeling comprises:

[0027] Converting deterministic parameters of the graph neural network into random variables and introducing a prior distribution;

[0028] Estimating the posterior distribution of parameters through variational inference;

[0029] Introducing uncertainty-aware graph attention computation;

[0030] Estimating the predictive uncertainty of the stress field distribution through Monte Carlo sampling.

[0031] In a preferred embodiment, the step of generating intelligent early warning decision information with risk level division comprises:

[0032] Considering the stress field distribution, stress gradient, and uncertainty level factors, constructing a risk assessment index;

[0033] Focusing on monitoring areas with high predictive uncertainty;

[0034] Based on the risk assessment index and the preset threshold, dividing the monitoring area into different risk levels and triggering the corresponding level of early warning signal;

[0035] Combining the geomechanical model, based on the inferred stress field distribution, predicting the location and development trend of potential fault zones.

[0036] In a preferred embodiment, the step of introducing uncertainty-aware graph attention computation adjusts the attention weight by introducing an edge uncertainty factor, which is calculated based on the variance of the attention weight, so that the influence of the edge connection in the message passing process is appropriately reduced when the uncertainty is high.

[0037] In a preferred embodiment, predicting the location of potential fault zones is based on the principal stress difference and the Mohr-Coulomb failure criterion, by comparing the maximum and minimum principal stress difference with the parameters related to the rock cohesion and internal friction angle, to determine the area where the fault may occur.

[0038] In a preferred embodiment, variational inference is achieved by minimizing the KL divergence between the true posterior distribution and the variational approximation, with the optimization objective being to maximize the lower bound of evidence, thereby obtaining the approximate posterior distribution of network parameters.

[0039] In a preferred embodiment, an AI vision-based geological deformation early warning system is used to perform an AI vision-based geological deformation early warning method, comprising:

[0040] An AI vision data acquisition and graph structure construction module is used to acquire surface micro-deformation features using a high-precision AI vision system and automatically construct a geological graph structure network with topological relationships;

[0041] A physically constrained graph convolution module is used to fuse nonlinear geomechanics equations into a deep graph neural network architecture based on the constructed geological graph structure network, to realize physically constrained graph convolution operations and feature extraction;

[0042] A multi-level graph structure inference module is used to create a multi-dimensional graph structure model representing different depth layers in the subsurface using the features extracted by the physically constrained graph convolution, and to realize accurate cross-layer geological information transfer through an innovative physically information-guided graph attention mechanism;

[0043] A Bayesian variational inference module is used to deploy a graph neural network based on the Bayesian probability framework to model the geological stress field distribution based on the multi-level graph structure model, and to quantify the uncertainty of the prediction results using variational inference techniques;

[0044] A risk-aware early warning decision module is used to integrate stress field distribution features and uncertainty assessment indicators based on the inferred geological stress field distribution and its uncertainty assessment results, to generate intelligent early warning decision information with risk level classification.

[0045] The beneficial effects of the present application are:

[0046] Improved early warning timeliness: by combining AI vision technology with physically constrained graph neural networks, the present application can infer subsurface stress field distribution from surface micro-deformation features, significantly advancing the discovery of geological disaster precursors, providing longer early warning lead time than traditional methods, and gaining valuable time for disaster prevention and mitigation.

[0047] Enhanced prediction accuracy: physically constrained graph convolution operations that incorporate geomechanics equations ensure that model inference results conform to geomechanics laws, significantly improving prediction accuracy, especially in predicting potential fault zone locations.

[0048] Realize deep inference: the multi-level graph structure model can represent geological conditions at different depths, and the physically information-guided graph attention mechanism enables accurate cross-layer information transfer, breaking through the limitations of traditional methods that are limited to surface monitoring, and enabling effective inference of deep subsurface stress fields.

[0049] Quantifying uncertainty: The Bayesian graph neural network framework can quantify the uncertainty of the prediction results, providing reliable risk assessment basis for early warning decision-making, avoiding the misjudgment risk that may be brought by traditional deterministic methods.

[0050] Adaptability: The method of the present invention can adapt to different geological environment characteristics and show excellent early warning effect in various scenarios such as mountain landslide, mining area slope, engineering construction area, underground space and earthquake activity area.

[0051] Efficient use of resources: Compared with the traditional method which needs a large number of drilling and physical detection, the present invention mainly relies on non-contact AI vision system to obtain data, which greatly reduces the monitoring cost and improves the resource utilization efficiency.

[0052] Perfect decision support: The hierarchical early warning decision information generated based on stress field distribution and uncertainty assessment provides scientific and intuitive decision basis for management departments, which helps to develop more accurate disaster prevention and mitigation measures. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of a geological deformation early warning method based on AI vision of the present invention;

[0054] Figure 2 is a column chart of the early warning time comparison of different early warning methods of the present invention;

[0055] Figure 3 is a line chart of the stress field inference accuracy of different depths of the present invention varying with monitoring time;

[0056] Figure 4 is a radar chart of the performance comprehensive evaluation of the technical solution of the present invention;

[0057] Figure 5 is a scatter plot of the prediction fracture zone position of the present invention and the actual verification coincidence degree. DETAILED DESCRIPTION

[0058] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can include changes in the function and arrangement of elements discussed without departing from the scope of the disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate or desired. Also, features described in relation to one example can be combined in other examples.

[0059] In at least one embodiment of the present invention, a geological deformation early warning method based on AI vision is disclosed, as shown in Figure 1 including the following steps:

[0060] Step 1: Use high-precision AI vision system to obtain surface micro-deformation features and automatically construct a geological map structure network with topological relationships;

[0061] Specifically, the following steps are included:

[0062] Step 1.1: Collect multi-dimensional surface deformation data;

[0063] 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 can be collected.

[0064] The collection frequency is dynamically adjusted according to the risk level of the monitoring area. High-risk areas can reach once an hour, and low-risk areas can be once a day.

[0065] In some embodiments, unmanned aerial photography systems and ground fixed camera arrays can be used in combination to form a multi-angle, multi-scale monitoring network, improving the comprehensiveness and reliability of data collection.

[0066] Step 1.2: Data preprocessing and feature extraction;

[0067] The collected raw visual data is processed for noise reduction, registration, and enhancement to extract key deformation features, including displacement field, strain field, and crack distribution.

[0068] Under complex weather conditions, adaptive image enhancement algorithms can be used to process raw data, including de-fogging, de-raining, and de-snowing, to ensure the quality of visual data in adverse weather conditions.

[0069] Step 1.3: Construct a geological map structure;

[0070] The monitoring area is represented as a graph structure , where the node set represents key monitoring points, and the edge set represents the physical constraint relationship between nodes.

[0071] For node , its feature vector contains multi-dimensional observation data of the monitoring point;

[0072] For edge , its feature vector represents the physical relationship parameters between nodes and .

[0073] Step 1.4: Initialize graph node and edge attributes;

[0074] According to the geological prior knowledge and historical monitoring data, initial attribute values are assigned to the nodes and edges in the graph structure, and an initial geological state representation is established.

[0075] Step 2, based on the constructed geological graph network, the nonlinear geomechanics equation is integrated into the deep graph neural network architecture to realize the physical constraint driven graph convolution operation and feature extraction.

[0076] Specifically, the following steps are included:

[0077] Step 2.1, constructing a physical constraint encoder;

[0078] The geomechanics equation (such as Hooke's law, Mohr-Coulomb criterion, etc.) is encoded as a constraint parameter to guide the message passing process of the graph neural network. The constraint parameter is defined based on the physical relationship between nodes and , including elastic modulus, Poisson's ratio, internal friction angle, etc. geomechanics parameters.

[0079] For Hooke's law, the relationship between the stress and strain of an elastic body can be expressed as:

[0080] ;

[0081] where is the stress tensor, representing the internal force distribution acting on the material; is the strain tensor, representing 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 part of the edge constraint parameter to represent the physical constraint relationship between nodes in the graph neural network.

[0082] In some embodiments, according to the characteristics of different rock-soil bodies, more complex rock-soil body constitutive models such as Drucker-Prager yield criterion or Hoek-Brown criterion can also be introduced to more accurately describe the nonlinear deformation behavior.

[0083] Step 2.2, implementing a physical constraint graph convolution operation;

[0084] Modify the message passing process of the standard graph convolution operation and integrate the physical constraint information, and the message passing function is represented as:

[0085] ;

[0086] where represents the message passing from node to node message vector, containing the influence information of node on node ; is the parameterized message function, representing the message passing function from node to node , determined by neural network parameters ; is the feature vector of source node , containing multi-dimensional observation data of the monitoring point, such as displacement, strain, etc. is the feature vector of target node ; is the edge feature connecting node and node , representing the spatial relationship parameters (such as distance, direction, etc.) between the two monitoring points. is the physical constraint parameter, encoding the geomechanical constraint relationship (such as elastic modulus, Poisson's ratio, internal friction angle, etc. geomechanical parameters) between node and node .

[0087] The specific implementation of the message function 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:

[0088] ;

[0089] where is the weight matrix of the first linear transformation, used to map the input features to the hidden layer space. is the weight matrix of the second linear transformation, used to map the hidden layer features to the output space. is the bias vector of the first layer, 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, connecting node features , , edge features , and physical constraint parameters into a long vector as the network input. is the rectified linear unit activation function, used to introduce nonlinear transformation ability and enhance the model's ability to express complex geological relationships. is the parameterized message function, representing the message passing function from node to node , determined by neural network parameters ; is the feature vector of the source node ; is the feature vector of the target node ; is the edge feature between the connection node and the node , representing the spatial relationship parameters between the two monitoring points; is the physical constraint parameter, encoding the geomechanical constraint relationship between the node and the node .

[0090] Step 2.3, node feature update;

[0091] Update the node features based on the received messages, the update rule is:

[0092] ;

[0093] wherein, is the updated feature vector of the node at the th layer, representing the new state of the monitoring point after one graph convolution operation; is the current feature vector of the node at the th layer, containing the current multi-dimensional observation data of the monitoring point; is the neighbor set of the node , representing all monitoring points directly connected to the monitoring point ; is the message vector transmitted from the neighbor node to the node , containing the influence information of the neighbor node on the node ; represents aggregating all neighbor messages, i.e., performing summation operation on the messages of all neighbor nodes of the node ; is the update function, responsible for fusing the current node feature and the aggregated neighbor messages to generate the new feature representation of the node.

[0094] The specific implementation of the update function is the Gated Recurrent Unit (GRU), which fuses the current node feature and the aggregated messages:

[0095] ;

[0096] wherein, is the updated feature vector of the node at the th layerthe updated feature vector of the current monitoring point, representing the updated monitoring point state information; the first monitoring point layer node the updated current feature vector of the current monitoring point, representing the current monitoring point state information; the node the neighbor node set of the node , representing other monitoring points that have physical association with the current monitoring point; the message vector transmitted from the neighbor node to the node , containing the state information of the neighbor node and its relationship information with the center node; representing aggregating all messages from neighbor nodes;

[0097] Step 2.4, physical consistency check;

[0098] Introducing a physical consistency loss function to ensure that the network inference result conforms to the geomechanics law, which is defined as:

[0099] ;

[0100] wherein, is the physical consistency loss function, used to measure the degree of conformity between the model prediction result and the physical law; represents the summation of all edges in the graph structure, i.e., considering the physical constraint relationship between all connected node pairs; is a constraint function based on physical laws, with the input being the feature vectors of node and node , and the output being the predicted physical relationship that should be satisfied between the two nodes; and are the feature vectors of node and node , respectively, containing the multi-dimensional observation data of each monitoring point; is a predefined physical constraint parameter, encoding the geomechanics constraint relationship that should be satisfied between node and node ; represents the square of the Euclidean norm, used to calculate the mean square error between the predicted physical relationship and the predefined constraint.

[0101] The physical consistency loss and the prediction task loss together constitute the optimization objective of the model:

[0102] ;

[0103] wherein, is the total loss function, which is the final objective of model optimization; is the prediction task loss, which measures the deviation between the model prediction results and the true observation values, usually using loss functions such as mean square error or cross-entropy; is the physical consistency loss, which is used to ensure that the model inference results conform to the geomechanics laws; is the balance parameter, which is a positive scalar value, controlling the strength of physical constraints. A larger value will make the model pay more attention to meeting the physical constraint conditions, and a smaller value will allow the model to deviate from the physical laws to a certain extent to better fit the observations.

[0104] Step 3, using the features extracted by the physical constraint graph convolution, create a multi-dimensional graph structure model representing different depth layers underground, and realize accurate cross-layer geological information transmission through innovative physical information guided graph attention mechanism;

[0105] Specifically, the following steps are included:

[0106] Step 3.1, build a multi-level graph structure;

[0107] Divide the underground structure of the monitoring area into multiple depth layers, each layer represented as a graph wherein, represents the depth index, and represents the identifier of the specific depth layer; represents the node set in the graph layer with depth , representing all monitoring points or calculation points on that depth layer; represents the edge set in the graph layer with depth , representing the connection relationship between monitoring points within the same depth layer, reflecting the horizontal geological correlation.

[0108] Each layer of graph structure is connected to each other through vertical connection edges , wherein represents the vertical connection edge set between different depth layers, used to establish the association between nodes in different depth layers, reflecting the geomechanics transmission relationship in the vertical direction, thus forming a complete three-dimensional graph structure.

[0109] According to the geomechanics theory, the underground 0-50 meters range is divided into 5 depth layers, namely 0-10 meters, 10-20 meters, 20-30 meters, 30-40 meters and 40-50 meters. The node distribution of each depth layer is determined according to the surface monitoring point grid and geological structure characteristics, and the nodes between adjacent depth layers are connected through vertical connection edges.

[0110] ​In some embodiments, the division of depth layers can be non-uniform according to different geological environmental characteristics, for example, finer division (such as 0-5 meters, 5-10 meters, 10-15 meters, etc.) is adopted in the shallow part to obtain a more detailed shallow stress field distribution.

[0111] Step 3.2, implement physical information guided graph attention computation;

[0112] Introduce information transmission weights between attention computation nodes, and the attention computation formula is:

[0113] ;

[0114] wherein, represents a physical information guided attention computation function; is a query matrix, representing the feature representation of the current node, used to query the association strength with other nodes; is a key matrix, representing the feature representation of the target node, used to calculate the attention score with the query matrix; is a value matrix, containing the information content of the target node, used for information aggregation according to the attention weight; is the transpose matrix of the key matrix ; is a physical constraint matrix, encoding the physical association strength between nodes based on geomechanics theory, used to guide the attention mechanism to pay attention to nodes that are more physically relevant; is a balance parameter, controlling the influence degree of physical constraints in attention computation, a larger value will make the model more dependent on physical constraints, and a smaller value will depend more on data-driven feature similarity; is the feature dimension, representing the dimension of the key vector, used to scale the dot product result to avoid the problem of gradient disappearance; is a normalization function, which converts the attention score into a probability distribution, ensuring that the sum of all attention weights is 1.

[0115] The physical constraint matrix contains the physical association strength between nodes, which is calculated based on geomechanics theory:

[0116] ;

[0117] wherein, is an element in the physical constraint matrix representing the physical association strength between node and node ; is the spatial distance between node and node , used to represent the spatial position relationship of the two monitoring points; and are the spatial distance between node and the node The density of the rock-soil mass at the location of the node, reflecting the quality characteristics of the rock-soil mass; and respectively the node and the node The elastic modulus at the location of the node, indicating the ability of the rock-soil mass to resist elastic deformation; and respectively the node and the node The Poisson's ratio at the location of the node, indicating the ratio of deformation in the force direction to the deformation perpendicular to the force direction of the rock-soil mass; is a physical correlation function, which is a function designed based on the principles of geomechanics, used to convert various physical parameters into the correlation strength between nodes, and the larger the output value indicates the stronger the physical correlation between the two nodes.

[0118] Step 3.3, inter-layer information transmission and state update;

[0119] Inter-layer information transmission is carried out through vertical connection edges to realize the propagation of surface observation information to deep underground, and to update the node state of each depth layer.

[0120] The information transmission process adopts a bidirectional propagation mechanism from top to bottom and from bottom to top:

[0121] ;

[0122] wherein, is the updated state of the depth layer node at time , indicating the latest feature representation of the depth layer monitoring point; is the current state of the depth layer node at time , indicating the current feature representation of the depth layer monitoring point; is the state of the depth layer (lower layer) node at time , indicating the feature representation of the relevant monitoring point in the lower layer; is the state of the depth layer (upper layer) node at time , indicating the feature representation of the relevant monitoring point in the upper layer; is the information aggregation function from bottom to top, responsible for transmitting the information of the lower layer node to the current layer, capturing the influence of deep underground on shallow layers; is the information aggregation function from top to bottom, responsible for transmitting the information of the upper layer node to the current layer, capturing the influence of shallow surface layers on deep layers; The state update balance parameter has a value range of 1. This controls the ratio of retaining the original state information to integrating new information; a larger value indicates a greater reliance on the new aggregated information. The directional balance parameter has a range of values. This controls the fusion ratio of bottom-up and top-down information; a larger value indicates that more of the influence of lower-level information is considered.

[0123] Step 3.4, inference of stress field distribution;

[0124] Based on the updated multi-level graph structure, the stress field distribution within a depth of 10 to 50 meters underground is inferred, and the stress tensor field is output. Each spatial location This corresponds to a stress tensor.

[0125] The stress tensor contains 6 independent components ( , , , , , The final output layer of the graph neural network yields:

[0126] ;

[0127] in, For nodes The predicted stress tensor at that point contains six independent components describing the stress state at that point. The final layer node features represent the process after... After processing by layered graph neural networks, the nodes The feature vector encodes the comprehensive status information of the monitoring point; The output layer weight matrix is ​​a linear transformation parameter that maps node features to the stress tensor space. This is the output layer bias vector, used to adjust the baseline level of the predicted values ​​to ensure that the stress tensor output by the model is within a reasonable range.

[0128] Of the components of the stress tensor, , and They represent along , and Normal stress in the axial direction; , and They represent , and Shear stress on a plane.

[0129] The stress tensor at any position in space is obtained by interpolation:

[0130] ;

[0131] where, is the stress tensor at spatial position , representing the stress state at this position; is the set of nodes near position , containing all the computational nodes that are close to the position and have an impact on the stress state of the position; is the stress tensor at node , representing the stress state at the node position; is the distance-based weight function, determining the degree of influence of node on the stress state of position , which is inversely proportional to the distance from node to position , the closer the distance, the greater the weight; represents the weighted sum of all nodes near position , realizing the interpolation calculation of the stress state at any position in space.

[0132] Step 4, based on the multi-level graph structure model, deploy the graph neural network of the Bayesian probability framework to model the geological stress field distribution, and combine the variational inference technology to quantify the uncertainty of the prediction results;

[0133] Specifically, the following steps are included:

[0134] Step 4.1, construct a Bayesian graph neural network;

[0135] Convert the deterministic parameters of the graph neural network into random variables, introduce the prior distribution , and choose the standard normal distribution.

[0136] For each network parameter , define its prior distribution as:

[0137] ;

[0138] where, represents the set of weight parameters in the graph neural network, which is all the learnable weight parameters in the model; represents the weight parameter in the graph neural network, which is a single learnable network parameter; represents the prior probability distribution of parameter , describing the prior belief of a single parameter; represents the standard normal distribution with mean 0 and variance The Gaussian distribution of the parameters is the prior distribution form of the parameters; The prior variance controls the dispersion of the parameter distribution. A larger value allows the parameter to take a wider range of values, while a smaller value causes the parameter to concentrate around the mean, thus controlling the regularization strength of the parameter.

[0139] Step 4.2: Implement the variational inference algorithm;

[0140] Estimate the posterior distribution of the parameters through variational inference, and minimize the true posterior distribution. Variational approximation The KL divergence between them is optimized with the following objective:

[0141] Equivalent to maximizing the lower bound of evidence (ELBO):

[0142] ;

[0143] in, Given observation data (Graphic structure) and Network parameters under (node ​​characteristics) conditions The true posterior distribution represents the probability estimate of the model parameters after observing specific data; It is a parameterized variational distribution used to approximate the true posterior distribution; The set of parameters for the variational distribution; This is the Kullback-Leibler divergence, used to measure the difference between two probability distributions; the smaller the value, the more accurate the approximation. These are variational parameters, which are the set of target parameters that need to be adjusted during the optimization process; The lower bound for evidence is the lower bound for the logarithmic marginal likelihood. Maximizing the ELBO is equivalent to minimizing the KL divergence. In variational distribution Lower log-likelihood The expected value represents the degree to which the model fits the observed data; Represents the natural logarithm function; The KL divergence between the variational distribution and the prior distribution is used as a regularization term to prevent overfitting. The prior distribution of the parameters represents the prior assumptions about the model parameters before the observed data are collected. This represents the set of weight parameters in a graph neural network, which consists of all learnable weight parameters in the model. Represents the first in a graph neural network Each weight parameter is a single learnable network parameter.

[0144] Variational approximation The posterior distribution is decomposed into independent Gaussian distributions using mean-field approximation:

[0145] ;

[0146] where, is the parameterized variational distribution used to approximate the true posterior distribution; is the parameter set of the variational distribution; denotes the set of weight parameters in the graph neural network, which is all the learnable weight parameters in the model; denotes the product operation on all parameters ; denotes the i-th weight parameter in the graph neural network, which is a single learnable network parameter; is the mean of the i-th network parameter, representing the optimal estimate value of the parameter; is the variance of the i-th network parameter, representing the uncertainty size of the parameter estimate; together they constitute the learnable variational parameter , which is constantly adjusted through the optimization process to approximate the true posterior distribution. The optimization process uses the reparameterization trick to convert random sampling into a deterministic function plus noise:

[0147] ;

[0148] ;

[0149] where, denotes the i-th weight parameter in the graph neural network, which is a single learnable network parameter; denotes random noise sampled from a standard normal distribution, with mean 0 and variance 1; is the mean of the i-th network parameter, representing the optimal estimate value of the parameter; is the parameter standard deviation, controlling the fluctuation range of the sampling result; is the network parameter sample obtained after reparameterization, used for forward propagation calculation and gradient backpropagation. In some embodiments, to improve computational efficiency, more complex variational distribution families such as Normalizing Flow or Matrix Variate Gaussian can be used to improve the expressiveness of the posterior approximation.

[0150] Step 4.3, introduce uncertainty-aware graph attention calculation;

[0151] Step 4.3, introduce uncertainty-aware graph attention calculation;

[0152] ​​​The uncertainty of edge attention weight is calculated by parameter uncertainty propagation, and the attention calculation formula is:

[0153] ;

[0154] wherein, is the attention weight of node to neighbor node , indicating the influence degree of node on node during information transmission; is the attention vector, which is a learnable model parameter, used to calculate the correlation score between nodes; is the attention weight matrix, used for linear transformation of node features; and are the feature vectors of node and node , containing the state information of the node; represents the concatenation operation of the transformed feature vectors of node and node ; is a modified linear unit activation function with leakage, used to introduce nonlinearity and prevent gradient vanishing; represents the neighbor node set of node , containing all nodes directly connected to node ; represents the summation of all neighbor nodes of node , used for normalizing attention weight; is the edge uncertainty factor, reflecting the confidence of the model's importance judgment of edge , the higher the value, the lower the uncertainty of the model's importance judgment of the edge; represents the transpose of attention vector ; represents the exponential function.

[0155] Edge uncertainty factor is calculated based on the variance of attention weight:

[0156] ;

[0157] wherein, is the uncertainty factor of edge , indicating the confidence of the model's importance judgment of the edge, the larger the value, the lower the uncertainty; is the variance of attention weight estimated by Monte Carlo sampling, quantifying the fluctuation degree of the model's estimation of the edge attention weight, the larger the value, the higher the uncertainty; For scaling parameter, controlling the sensitivity of uncertainty to edge weight, larger value makes the model more sensitive to uncertainty, smaller value reduces the impact of uncertainty; For exponential function, converting negative weighted variance to a multiplicative factor in the range of 0 to 1, the larger the variance, the closer the uncertainty factor to 0, thus reducing the importance of the edge in information transmission.

[0158] Step 4.4, predict uncertainty quantification;

[0159] Estimate the prediction uncertainty of stress field distribution by Monte Carlo sampling, for spatial position stress tensor , the prediction mean and variance are respectively:

[0160] ;

[0161] ;

[0162] where, is the prediction mean of stress tensor at spatial position , representing the best estimate of the model for the stress state at this position; is the prediction variance of stress tensor at spatial position , quantifying the uncertainty of the prediction result; is the stress tensor prediction obtained by the th Monte Carlo sampling, representing the output of the model under a single network parameter sampling; is the number of Monte Carlo sampling, determining the sample size of uncertainty estimation.

[0163] Sampling frequency Choose 20-50, in practical application, need to balance between calculation efficiency and estimation accuracy. Prediction uncertainty can also be decomposed into cognitive uncertainty (model parameter uncertainty) and random uncertainty (intrinsic randomness of data):

[0164] ;

[0165] where, is the total prediction uncertainty, representing the overall variability of the prediction result; is the cognitive uncertainty, derived from the uncertainty of model parameters, reflecting the limitations of training data or model structure, which can be reduced by increasing training data; is the random uncertainty, derived from the inherent randomness of data and measurement noise, which cannot be eliminated by increasing training data, reflecting the inherent randomness of the system.

[0166] In some embodiments, a deep ensemble method can also 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 samples.

[0167] Step 5. Based on the inferred geologic stress field distribution and its uncertainty assessment results, integrate the stress field distribution characteristics and uncertainty assessment indicators to generate intelligent early warning decision information with risk level classification.

[0168] Specifically, the following steps are included:

[0169] Step 5.1. Establish a risk assessment index system.

[0170] Considering factors such as stress field distribution, stress gradient, and uncertainty level, construct risk assessment indicators to quantitatively represent potential geological disaster risks.

[0171] Risk assessment indicators The calculation formula is:

[0172] ;

[0173] Where, is the comprehensive risk assessment indicator, representing the degree of geological disaster risk in a specific area, and the higher the value, the greater the risk; is the stress level indicator, quantifying the proximity of the stress state in the region to the critical failure stress, reflecting the stress state of the region; is the stress gradient indicator, representing the rate of spatial variation of the stress field, and a larger stress gradient often means a stress concentration area, which is a potential source of damage; is the uncertainty indicator, reflecting the confidence of the model in predicting the stress field, and higher uncertainty indicates lower reliability of the prediction results, which requires additional attention; is the time evolution indicator, describing the trend and rate of change of the stress field over time, and areas with rapid changes usually have higher risks; , , , respectively represent the weight coefficients of the stress level indicator, stress gradient indicator, uncertainty indicator, and time evolution indicator, satisfying to ensure consistency and comparability of the assessment.

[0174] Step 5.2. Implement key area monitoring.

[0175] Key areas with high prediction uncertainty are monitored, data collection frequency is increased, and the blind area of the early warning system is reduced.

[0176] The identification of key monitoring areas is based on the uncertainty threshold:

[0177] ;

[0178] wherein, denotes the set of spatial regions that need to be monitored intensively, containing all spatial points that satisfy the uncertainty condition; denotes the coordinate point in three-dimensional space, used to locate the specific monitoring position; denotes the spatial position where the predicted variance of stress tensor is, quantifying the uncertainty of stress prediction at this position; is the uncertainty threshold, a preset critical value, used to determine whether the uncertainty is high enough to need intensive monitoring, dynamically adjusted according to risk tolerance; denotes the condition restriction in set definition, the left side is the set element, and the right side is the condition that the element needs to satisfy.

[0179] In some embodiments, based on the risk assessment indicators and historical data, an active learning strategy can be used to adaptively adjust the allocation of monitoring resources, allocating more monitoring resources to high-risk and high-uncertainty areas.

[0180] Step 5.3, generate hierarchical early warning decision;

[0181] Based on the risk assessment indicators and preset thresholds, the monitoring area is divided into different risk levels, and the corresponding level of early warning signal is triggered. Early warning classification usually includes four levels: normal, attention, warning and danger.

[0182] The determination rule of early warning level is:

[0183] Normal level: , indicating that the comprehensive risk assessment indicator is below the first threshold, the area is in a safe state and does not need special attention;

[0184] Attention level: , indicating that the comprehensive risk assessment indicator exceeds the first threshold but is below the second threshold, the area appears slight abnormalities and needs to strengthen the monitoring frequency;

[0185] Warning level: , indicating that the comprehensive risk assessment indicator exceeds the second threshold but is below the third threshold, the area has obvious risks, and emergency plans need to be started and personnel evacuation preparations need to be made;

[0186] Danger level: , indicating that the comprehensive risk assessment indicator exceeds the third threshold, the area is in a state of extremely high risk, and emergency plans need to be executed immediately and personnel need to be evacuated;

[0187] wherein, To comprehensively assess the risk evaluation index, the regional geological disaster risk degree is quantified; is the first risk threshold, marking the critical point from normal state to attention; is the second risk threshold, marking the critical point from attention state to warning; is the third risk threshold, marking the critical point from warning state to danger state. These thresholds are determined according to historical disaster data statistical analysis and geological expert experience knowledge, and can be adjusted according to different regional geological conditions.

[0188] Step 5.4, predict the location and development trend of potential fault zone;

[0189] Combined with the geomechanical model, based on the inferred stress field distribution, the location and development trend of potential fault zone are predicted to provide decision support for disaster prevention and mitigation.

[0190] The identification of fault zone location is based on the principal stress difference and Mohr-Coulomb failure criterion:

[0191] ;

[0192] where, represents the spatial region set of potential fault zone, containing all spatial points satisfying the failure criterion; represents the coordinate point in three-dimensional space, used to locate the specific fault location; is the maximum principal stress at spatial position , representing the maximum tensile or compressive stress at this point; is the minimum principal stress at spatial position , representing the minimum tensile or compressive stress at this point; is the principal stress difference, representing the size of shear stress; is the cohesion, representing the ability of material to resist shear failure under zero normal stress, usually in units of MPa; is the internal friction angle, representing the parameter of material internal friction characteristics, in degrees, reflecting the growth rate of material shear strength when normal stress increases; is the internal friction coefficient, representing the proportional relationship between normal stress and the increment of shear strength generated thereby; The symbol represents the condition limitation in set definition, with the left side being the set element and the right side being the condition that the element needs to satisfy.

[0193] The fault development direction is perpendicular to the direction of maximum principal stress, which is determined by the characteristic vector of the principal stress field.

[0194] Application examples of the embodiment:

[0195] In a mine slope monitoring application, this method has been practically verified. The mine is located in a complex geological structure area, and the slope stability is affected by multiple factors, making it difficult for traditional monitoring methods to accurately predict potential landslide risks.

[0196] The implementation process of this method is as follows:

[0197] Visual data collection: 20 high-definition cameras and 2 laser scanners were deployed around the slope to form a monitoring network. The camera acquisition frequency was once an hour, and the laser scanning was once every 6 hours. After 3 months of data collection, more than 2160 hours of monitoring data were accumulated.

[0198] Graph structure construction: Based on the geological characteristics of the monitoring area, the area was divided into 350 monitoring nodes, and the graph structure was constructed through physical constraint relationships. Each node contains 15-dimensional features representing displacement, strain, crack, etc.

[0199] Physical constraint graph convolution modeling: According to the rock and soil mechanics parameters of the mine area (average elastic modulus E=2.5GPa, 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 adopts a 3-layer graph convolution structure with a hidden layer dimension of 64.

[0200] Multi-level graph structure inference: The underground 0-50 meters range is divided into 5 layers, and a three-dimensional graph structure is constructed. Through the physical information guided graph attention model, interlayer information transmission is carried out. The model successfully inferred the underground stress field distribution, especially in the 10 to 30 meters depth range, identifying 3 potential high stress concentration areas.

[0201] Bayesian variational inference and uncertainty quantification: Bayesian graph neural networks are used to probabilistically model stress field distribution, and variational inference is used to quantify prediction uncertainty. The number of Monte Carlo samples is set to 30, successfully identifying 5 high-uncertainty areas, 3 of which overlap with high-stress concentration areas.

[0202] Risk-aware early warning decision: Based on stress field distribution and uncertainty assessment, risk assessment indicators are constructed to divide the monitoring area into different risk levels. The system identified a potential fault zone in the northern slope, with a risk assessment indicator reaching the warning level.

[0203] Technical effect verification:

[0204] Early warning capability improvement: The method successfully predicted the deformation acceleration of the northern slope in practical application, with an early warning time 42 days earlier than the traditional monitoring method (15 days for the traditional method and 57 days for the method), a proportion of 57 / 15=380%, verifying the technical effect of early warning time 30-60% earlier. The coincidence degree of the predicted location of the potential fracture zone and the subsequent drilling verification reached 82%, verifying the technical effect of a prediction accuracy of 80%.

[0205] Enhanced decision reliability: During the 3-month monitoring period, the traditional monitoring method produced 5 false positives and 2 false negatives, while the method produced only 2 false positives and no false negatives, with a false positive rate reduction of (5-2) / 5=60% and a false negative rate reduction of 100%, verifying the technical effects of a 40% reduction in false positive rate and a 50% reduction in false negative rate.

[0206] Optimized resource utilization: Compared with the traditional geophysical prospecting scheme that requires drilling 12 monitoring holes, the method only needs to deploy a visual monitoring system and 3 verification drill holes, with a cost reduction of (12-3) / 12=75%, verifying the technical effect of a 65% reduction in monitoring cost. The monitoring coverage area has expanded from 2.5 square kilometers in the traditional method to 6.8 square kilometers, an increase of 172%.

[0207] As shown in Figures 2 to 5 , respectively, are the early warning time comparison of different early warning methods; the accuracy of stress field inference at different depths changes with monitoring time; comprehensive performance evaluation of technical solutions; and the coincidence degree of the predicted fracture zone location and actual verification.

[0208] The above describes embodiments of the present application, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. An AI vision-based geological deformation early warning method, characterized in that, The method comprises the following steps: acquiring surface micro-deformation features using a high-precision AI vision system and automatically constructing a geological map structure network with topological relationships; based on the constructed geological map structure network, integrating nonlinear geomechanics equations into a deep graph neural network architecture to realize physically constrained graph convolution operations and feature extraction; using the features extracted by the physically constrained graph convolution, creating a multi-dimensional graph structure model representing different depth layers underground, and realizing accurate cross-layer geological information transmission through an innovative physically information-guided graph attention mechanism; based on the multi-dimensional graph structure model, deploying a graph neural network of a Bayesian probability framework to model the distribution of the geological stress field, and quantifying the prediction uncertainty using variational inference techniques; the step of deploying the graph neural network of the Bayesian probability framework to model the distribution of the geological stress field comprises: converting the deterministic parameters of the graph neural network into random variables and introducing a prior distribution; estimating the posterior distribution of the parameters through variational inference; introducing uncertainty-aware graph attention calculation; evaluating the prediction uncertainty of the stress field distribution through Monte Carlo sampling; integrating the stress field distribution features and uncertainty evaluation indicators based on the inferred geological stress field distribution and uncertainty evaluation results to generate intelligent early warning decision information with risk level classification. 2.The AI vision-based geological deformation early warning method of claim 1, wherein, The step of acquiring surface micro-deformation features using a high-precision AI vision system and automatically constructing a geological map structure network with topological relationships comprises: collecting multi-dimensional surface deformation data; performing noise reduction, registration, and enhancement processing on the collected raw visual data to extract key deformation features; representing the monitoring area as a graph structure, where nodes represent key monitoring points and edges represent the physical constraint relationships between nodes. 3.The AI vision-based geological deformation early warning method of claim 1, wherein, The step of realizing physically constrained graph convolution operations and feature extraction further comprises: encoding geomechanics equations as constraint parameters to guide the message passing process of the graph neural network; updating node features based on received messages; introducing a physical consistency loss function to ensure that the network inference results comply with geomechanics laws. 4.The AI vision-based geological deformation early warning method of claim 1, wherein, The step of creating a multi-dimensional graph structure model representing different depth layers underground comprises: dividing the underground structure of the monitoring area into multiple depth layers, with each layer represented as a graph structure; connecting the graph structures of each layer to each other through vertical connection edges to form a three-dimensional graph structure; implementing physically information-guided graph attention calculation, where a physical constraint matrix is introduced to guide the allocation of attention weights. 5.The AI vision-based geological deformation early warning method of claim 1, wherein, The step of generating intelligent early warning decision information with risk level classification comprises: considering the stress field distribution, stress gradient, and uncertainty level factors to construct a risk assessment indicator; focusing on monitoring areas with high prediction uncertainty; based on the risk assessment indicator and pre-set thresholds, dividing the monitoring area into different risk levels and triggering corresponding level warning signals; combining the geomechanics model and the inferred stress field distribution to predict the location and development trend of potential fault zones. 6.The AI vision-based geological deformation early warning method of claim 1, wherein, The step of introducing uncertainty-aware graph attention calculation comprises adjusting the attention weights by introducing an edge uncertainty factor, which is calculated based on the variance of the attention weights, so that the influence of edges with higher uncertainty in the message passing process is reduced.

7. The AI vision-based geological deformation early warning method according to claim 5, characterized in that, The potential fracture zone position is predicted based on the principal stress difference and the Mohr-Coulomb failure criterion, and by comparing the difference between the maximum and minimum principal stresses with the parameters related to the rock cohesion and internal friction angle to determine the area where the fracture is likely to occur. 8.The AI vision-based geological deformation early warning method of claim 1, wherein, Variational inference is achieved by minimizing the KL divergence between the true posterior distribution and the variational approximation, and the optimization objective is to maximize the lower bound of evidence, thereby obtaining the approximate posterior distribution of network parameters.

9. An AI vision-based geological deformation early warning system for performing the AI vision-based geological deformation early warning method of any one of claims 1-8. The method comprises the following steps: An AI vision data acquisition and graph structure construction module is used to acquire surface micro-deformation features by using a high-precision AI vision system and to automatically construct a geological graph structure network with topological relationships; A physically constrained graph convolution module is used to fuse nonlinear geomechanics equations into a deep graph neural network architecture based on the constructed geological graph structure network, so as to realize physically constrained graph convolution operation and feature extraction; A multi-dimensional graph structure inference module is used to create a multi-dimensional graph structure model representing different depth layers in the underground by using the features extracted by the physically constrained graph convolution, and to realize accurate transmission of cross-layer geological information by using an innovative physical information guided graph attention mechanism; A Bayesian variational inference module is used to deploy a graph neural network of a Bayesian probability framework to model the geological stress field distribution based on the multi-dimensional graph structure model, and to quantify the uncertainty of the prediction result by using a variational inference technique; A risk perception early warning decision module is used to integrate stress field distribution characteristics and uncertainty evaluation indexes to generate intelligent early warning decision information with risk level division according to the inferred geological stress field distribution and uncertainty evaluation results.

Images