Space-time intelligent prediction method and system for full-space grouting of sandy gravel stratum

Abstract

The invention relates to the technical field of grouting prediction, in particular to a sandy gravel stratum full-space grouting space-time intelligent prediction method and system.The method comprises the steps that the predicted geological condition of each drill hole is obtained; according to the predicted geological condition of each drill hole, whether the current drill hole is a non-grouting area or not is determined; grout is injected into the grouting holes in the non-grouting area, and after the grouting operation is completed, the predicted geological condition of each drilled hole is obtained; judging whether the current drill hole is a grouted area or a non-grouted area according to the obtained predicted geological condition of each drill hole, if the current drill hole is the non-grouted area, predicting a predicted demand index of the non-grouted area by adopting a dynamic updating and predicting model of the non-grouted area, and if the current drill hole is the grouted area, predicting the predicted geological condition of the current drill hole according to the predicted geological condition of the current drill hole. And inputting the result into the grouted area fitting prediction model, and stopping grouting until a preset requirement is met. The timeliness and accuracy of prediction are improved, and real-time guidance is provided for grouting construction.

Description

Technical Field

[0001] The invention relates to the technical field of grouting prediction, and in particular to a method and system for spatiotemporal intelligent prediction of full-space grouting in a sandy and gravel stratum. Background Art

[0002] As a special heterogeneous stratum, the sand and gravel stratum has a unique pore structure and permeability characteristics, which puts forward higher technical requirements for grouting engineering. The spatial distribution of porosity and permeability in the sand and gravel stratum is significantly non-uniform, which makes it difficult to accurately predict the diffusion law of grouting slurry in the stratum, and the flow control and diffusion range regulation during the grouting construction process face great uncertainty. At the same time, the sand and gravel stratum is often accompanied by water-rich characteristics and complex hydrogeological conditions, which further increases the difficulty of interpretation of the coupling between slurry and groundwater. How to effectively predict the diffusion behavior of slurry during grouting and its effect on the transformation of stratum permeability has become a technical problem that needs to be solved urgently in the engineering community.

[0003] Traditional grouting diffusion prediction methods mainly rely on empirical formulas and numerical simulation techniques. However, empirical formulas usually ignore the impact of formation heterogeneity on grouting diffusion and are difficult to adapt to complex sand and gravel formations; while numerical simulation methods are limited by modeling and computational efficiency and require a large number of simplifications of formation physical parameters, resulting in low prediction accuracy. In addition, existing grouting monitoring systems rely more on discrete point data (such as drilling test data) for a posteriori judgment of the diffusion range, lack the ability to predict the grouting process in full space in real time, and are difficult to meet the timeliness requirements in actual engineering.

[0004] In recent years, with the development of deep learning technology, data-driven methods have gradually been applied to the solution of complex engineering problems. In grouting engineering, deep learning models can capture the nonlinear laws in the slurry diffusion process through the fusion training of multi-source data, and provide technical support for the intelligent prediction of the grouting process. However, in the grouting prediction of sand and gravel formations, how to combine the drilling data before and after grouting and the dynamic change data generated during the grouting process to build a time-space full-space prediction model is still a key difficulty in current research. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the present invention provides a spatiotemporal intelligent prediction method and system for full-space grouting in sandy and gravel formations;

[0006] On the one hand, a spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel formations is provided, including:

[0007] (1): Obtain the predicted geological conditions of each borehole; determine whether the current borehole is an un-grouted area based on the predicted geological conditions of each borehole;

[0008] (2): Inject grout into the grouting holes in the ungrouted area. After the grouting operation is completed, obtain the predicted geological conditions of each borehole again;

[0009] (3): Based on the predicted geological conditions of each borehole, determine whether the current borehole is a grouted area or an ungrouted area. If it is an ungrouted area, the dynamic update and prediction model of the ungrouted area is used to predict the prediction demand index of the ungrouted area. If the index is greater than the set threshold, return to (2); if it is a grouted area, the predicted geological conditions of the current borehole are input into the grouted area fitting prediction model to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to (2); if it has met the preset requirements, stop grouting.

[0010] On the other hand, a spatiotemporal intelligent prediction system for full-space grouting in sand and gravel formations is provided, including:

[0011] An acquisition module is configured to: acquire the predicted geological conditions of each borehole; determine whether the current borehole is an un-grouted area according to the predicted geological conditions of each borehole;

[0012] A grouting module, which is configured to: inject grout into the grouting holes in the un-grouted area, and after the grouting operation is completed, obtain the predicted geological conditions of each borehole again;

[0013] The prediction module is configured as follows: according to the predicted geological conditions of each borehole, determine whether the current borehole is a grouted area or an ungrouted area; if it is an ungrouted area, use the dynamic update and prediction model of the ungrouted area to predict the prediction demand index of the ungrouted area; if the index is greater than the set threshold, return to the grouting module; if it is a grouted area, input the predicted geological conditions of the current borehole into the fitting prediction model of the grouted area to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to the grouting module; if it has met the preset requirements, stop grouting.

[0014] The above technical solution has the following advantages or beneficial effects:

[0015] (1) This technical solution realizes dynamic prediction and real-time update of drilling data by constructing a spatiotemporal graph convolutional network. By using the drilling location and attribute information as nodes and combining them with time series data for modeling, the dynamic changes in the grouting process can be accurately captured. This method greatly improves the timeliness and accuracy of the prediction, avoids the prediction deviation caused by data lag in traditional static models, and provides real-time guidance for grouting construction.

[0016] (2) Multiple loss functions are introduced, including prediction error loss, physical consistency loss, regularization loss, and robustness loss, to ensure that the prediction results are not only highly accurate but also conform to geophysical laws. For example, the physical consistency loss is used to constrain the physical meaning of the prediction results in terms of porosity, fracture density, and other attributes, thus avoiding the occurrence of outliers or unreasonable results and providing reliable guarantees for practical engineering applications.

[0017] (3) By establishing a nonlinear mapping relationship between multiple attributes such as crack density, porosity, and elastic modulus and the grouting demand index, the grouting demand index of the un-grouted area is accurately evaluated, and the grouting strategy is dynamically adjusted. By giving priority to the prediction and optimal allocation of high-demand areas, the efficiency of grouting construction is significantly improved, while the waste of slurry materials is reduced, achieving the optimal allocation of resources.

[0018] (4) Combining deep learning with online learning mechanisms, the system can update model parameters in real time during the construction process and dynamically adjust the prediction results, forming an intelligent grouting prediction and optimization system. With the real-time spatiotemporal visualization function, the system can intuitively display the attribute distribution and demand index of the grouting area, providing clear guidance for construction personnel, thereby achieving precise construction, intelligent regulation and scientific decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 This is a flow chart of the method of embodiment 1.

[0021] Figure 2 Schematic diagram of the internal structure of the spatiotemporal graph convolutional network of Example 1.

[0022] Figure 3 Schematic diagram of the tunnel face, fractures, boreholes and aquifers in Example 1. DETAILED DESCRIPTION

[0023] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0024] Embodiment 1, as Figure 1 As shown, this embodiment provides a spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel formations, including:

[0025] S101: Obtaining the predicted geological conditions of each borehole; determining whether the current borehole is an un-grouted area according to the predicted geological conditions of each borehole;

[0026] S102: injecting grout into the grouting holes in the un-grouted area, and after the grouting operation is completed, obtaining the predicted geological conditions of each borehole again;

[0027] S103: According to the predicted geological conditions of each borehole obtained in S102, determine whether the current borehole is a grouted area or an ungrouted area. If it is an ungrouted area, the dynamic update and prediction model of the ungrouted area is used to predict the prediction demand index of the ungrouted area. If the index is greater than the set threshold, return to S102; if it is a grouted area, the predicted geological conditions of the current borehole are input into the fitting prediction model of the grouted area to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to S102; if it has met the preset requirements, stop grouting.

[0028] Further, the obtaining of the predicted geological conditions of each borehole includes: drilling a tunnel face to be grouted, the number of boreholes being n, obtaining geophysical information corresponding to each borehole, and inputting the geophysical information corresponding to each borehole into the trained spatiotemporal graph convolutional network in sequence according to the drilling time sequence, to obtain the predicted geological conditions of each borehole;

[0029] The method of obtaining the predicted geological conditions of each borehole again includes: drilling n boreholes in sequence to obtain geophysical information corresponding to each borehole, and inputting the geophysical information corresponding to each borehole into the trained spatiotemporal graph convolutional network in sequence according to the drilling time sequence to obtain the predicted geological conditions of each borehole.

[0030] Furthermore, S101: drilling the tunnel face to be grouted, the number of holes drilled is n, and geophysical information corresponding to each hole is obtained, and the geophysical information is collected by a sensor in front of the drill bit. The geophysical information includes: fracture density, porosity, permeability, etc.

[0031] Furthermore, the geophysical information corresponding to each borehole is sequentially input into the trained spatiotemporal graph convolutional network in the order of drilling time to obtain the predicted geological conditions of each borehole. The training process of the trained spatiotemporal graph convolutional network includes:

[0032] Constructing a training set, wherein the training set is geophysical information corresponding to each borehole whose geological conditions are known;

[0033] The training set is input into the spatiotemporal graph convolutional network, and the spatiotemporal graph convolutional network is trained. When the loss function value of the spatiotemporal graph convolutional network no longer decreases, or the number of iterations exceeds the set number, the training is stopped to obtain the trained spatiotemporal graph convolutional network.

[0034] Furthermore, the loss function L of the spatiotemporal graph convolutional network is total , its formula expression is:

[0035] L total =α 1 ·L prediction +α 2 ·L consistency +α 3 ·L regularization +α 4 ·L robustness ;

[0036] Among them, α 1 ,α 2 ,α 3 ,α 4 is a hyperparameter used to balance the weight of each part of the loss.

[0037] Among them, L prediction The formula expression is:

[0038]

[0039] Among them, L prediction Represents the prediction error loss, which is used to measure the prediction value and the true value y i The error, dynamic weight ω i , used to increase the attention to high deviation samples, β represents a dynamic adjustment parameter, which is used to control the influence of outliers. Represents the mean of the data set; the purpose of using prediction error loss is to dynamically adjust the focus on different samples and enhance the prediction ability of high deviation areas.

[0040] Furthermore, the L consistency The formula expression is:

[0041]

[0042] Among them, L consistency represents the physical consistency loss function, L consistency Used to constrain the prediction results to conform to the laws of geophysics; R min Represents R injection The physical lower limit, R max Represents R injection The physical upper limit of R injection,iRepresents the predicted grouting demand index; the purpose of using the physical consistency loss function is to prevent the predicted value from being too large or too small and to ensure that the physical meaning is reasonable.

[0043] Furthermore, the L regularization The formula expression is:

[0044]

[0045] Among them, λ 1 Indicates the size of the control weight to prevent overfitting; λ 2 It means controlling the balance of weights and encouraging uniform distribution. The purpose of using weight regularization loss function is to avoid overfitting and unbalanced distribution caused by excessive weights.

[0046] Among them, L robustness The specific expression is:

[0047]

[0048] Among them, δ represents the outlier threshold, which controls the degree of penalty for excessive errors, and L robustness Represents the robustness loss function, which is used to enhance the robustness of the model to outliers. φ(z) represents the kernel function of the loss function, which is used to process the error between the model prediction value and the true value. z represents the difference (error) between the prediction value and the true value, and N represents the total number of data samples.

[0049] Furthermore, if Figure 2 As shown, the trained spatiotemporal graph convolutional network includes:

[0050] Spatial feature input layer, adjacency matrix input layer and temporal feature input layer;

[0051] The output of the spatial feature input layer and the output of the adjacency matrix input layer are connected to the input of the graph convolution layer, the output of the graph convolution layer is connected to the input of the multi-layer graph convolution module, and the output of the multi-layer graph convolution module is connected to the input of the fusion layer;

[0052] The output of the temporal feature input layer and the output of the adjacency matrix input layer are both connected to the input of the temporal convolution layer, the output of the temporal convolution layer is connected to the input of the temporal recursive layer, and the output of the temporal recursive layer is connected to the input of the fusion layer;

[0053] The input end of the fusion layer is also connected to the output end of the attention mechanism layer; the output end of the fusion layer is respectively connected to the input end of the dynamic update module and the input end of the grouting demand index prediction layer, and the dynamic update module and the grouting demand index prediction layer are interconnected; the output end of the dynamic update module is also connected to the grouting demand index prediction layer through the attribute prediction layer.

[0054] Furthermore, the spatial feature input layer is used to input the static geological attribute feature matrix of each node in the space-time graph, with a shape of N×F s , where N is the number of nodes and F s It is the attribute dimension of each node (such as crack density, porosity, etc.). The space-time diagram is a node connection diagram with both time dimension and space dimension; the time dimension means that the information collected by each node at different times is different; the space dimension means that each node is at different coordinates in the tunnel; each node corresponds to a borehole or geological observation point.

[0055] The static geological attribute characteristic matrix refers to a quantitative information matrix that describes the geological characteristics of the borehole or observation point (i.e., node) at a specific time point or before the grouting operation. These characteristics are usually basic geological attributes that are relatively stable in time and do not change significantly with the grouting operation. The static geological attribute characteristic matrix is ​​one of the inputs of the spatiotemporal model and is used to describe the basic geological conditions of the node.

[0056] The static geological attribute characteristic matrix is ​​expressed as follows:

[0057]

[0058] Where N is the number of nodes (number of holes), F s Represents the attribute dimension of each node, that is, the number of types of static geological attributes. i,j Represents the j-th geological attribute value of node i.

[0059] Furthermore, the time feature input layer is used to input the time series dynamic features of each node, and the shape is N×T×F t , where T is the number of time steps and F t It is the temporal characteristic dimension (such as the temporal variation of crack density).

[0060] The dynamic characteristics of time series refer to the changes in the geological properties of the borehole or node (such as fracture density, porosity, permeability, water content, etc.) in the time dimension over time. It reflects the dynamic evolution of the grouting process or geological conditions over time.

[0061] Furthermore, the adjacency matrix input layer is used to input an adjacency matrix describing the spatial relationship between nodes, which has a shape of N×N and whose weights reflect the spatial connection strength between nodes.

[0062] The adjacency matrix is ​​a mathematical representation used to describe the connection relationship between nodes in a graph. It reflects the association between boreholes (or nodes) in geological space, such as distance, attribute similarity or connectivity. The adjacency matrix is ​​an important component of the spatiotemporal graph convolutional network and guides the calculation of graph convolution.

[0063] The adjacency matrix is ​​expressed as follows:

[0064]

[0065] Where N is the number of nodes (number of holes), a ij Represents the connection strength or relationship weight between node i and node j.

[0066] Assume that the spatial coordinates of each node i and j are x i and x j , the connection strength a can be calculated based on the distance ij :

[0067]

[0068] Among them, ||x i -x j || represents the Euclidean distance between nodes i and j, and σ represents the distance scale parameter, which is used to control the decay rate of the connection strength.

[0069] Furthermore, the graph convolution layer (GCL) uses graph convolution to extract the spatial features of nodes. The formula is:

[0070]

[0071] Among them, H (l+1) is the node feature matrix of all nodes in the graph at the l+1 layer, D is the degree matrix, D ii =∑ j A ij , W (l) is the weight matrix of the lth layer; σ is the ReLU activation function, A is the adjacency matrix of the graph, and A represents the spatial relationship of the drilling nodes. In the spatiotemporal graph convolutional network, the drilling holes are regarded as nodes of the graph, and the adjacent relationship between the drilling holes is represented by edges.

[0072] Node and edge update: The re-drilled borehole location is added to the space-time graph as a new node and connected to the adjacent boreholes. The newly added time node reflects the newly revealed grouting information, allowing the entire space-time graph to be dynamically expanded. Attribute update: The grouting attribute information of the new node is propagated in the space-time graph through spatial interpolation and temporal recursive update methods to achieve the status update of the grouted and ungrouted areas.

[0073] Furthermore, the multi-layer graph convolution module, including several graph convolution layers connected in sequence, captures higher-order node association features by stacking multiple layers of GCL. The output of each layer serves as the input of the next layer. The multi-layer graph convolution module outputs the spatial feature matrix obtained after L layers of graph convolution operations.

[0074] Furthermore, the temporal convolution layer (TCN) uses temporal convolution or LSTM to extract features in the time dimension to capture the characteristics of temporal changes. To extract the dynamic change characteristics of time series data, a one-dimensional convolution operation is used:

[0075]

[0076] Among them, j represents the convolution kernel position index in the time series, is the output feature after temporal convolution; θ j are the parameters of the convolution kernel; represents the input data at time rj, k is the width of the convolution kernel, and b is the bias. The output of the temporal convolution layer is the temporal feature matrix The shape of the time feature matrix is ​​N×F (L) ; where N represents the number of nodes (number of holes), F (L) Represents the feature dimension of the last layer.

[0077] Furthermore, the Gate Recurrent Unit (GRU) captures long-term dependencies in the time dimension, processes the time series features through the recurrent neural network GRU, and outputs the updated time feature matrix

[0078] If the LSTM network is used, its core formula is:

[0079] f t =σ(W f ·[h t-1 ,x t ]+b f );i t =σ(W i ·[h t-1 ,x t ]+b i );

[0080]

[0081] o t =σ(W o ·[h t-1 ,x t ]+b o );h t =o t *tanh(C t );

[0082] Among them, f t ,it ,o t are forget, input and output gates respectively, C t is the unit state, h t Hidden state

[0083] Furthermore, the fusion layer transforms the spatial feature matrix and the time feature matrix Perform feature fusion to generate a comprehensive spatiotemporal feature representation F:

[0084] F = concat(H spatial ,H temporal );

[0085] Furthermore, the attention mechanism layer assigns weights to different features and optimizes the feature fusion effect through the attention mechanism:

[0086]

[0087] Among them, α i represents the attention weight, e i It represents the score obtained by weighting the feature by the attention weight matrix, e j represents the scores of all features used for normalization, W attention It represents the weight matrix learned during training and is used to optimize the score calculation process. i The representation is a feature of the input, fusing spatial and temporal information.

[0088] Furthermore, the dynamic update module includes: an adjacency matrix update unit, a node feature update unit and an online learning mechanism unit; the adjacency matrix update unit includes:

[0089] When new drilling data is added, the adjacency matrix A is dynamically adjusted:

[0090] A new =A old +ΔA;

[0091] Among them, A new represents the updated adjacency matrix, including the newly added drilling node and its spatial relationship with other nodes, A old represents the adjacency matrix before updating, describing the spatial relationship between existing drilling nodes, and ΔA represents the adjacency matrix before updating, describing the spatial relationship between existing drilling nodes.

[0092] The node feature updating unit includes: interpolating or dynamically updating the features of the newly added nodes:

[0093] H new =H old +ΔH;

[0094] Among them, H new represents the updated node feature matrix, including the features of the newly added nodes and the adjusted results of the existing node features. old It represents the node feature matrix before updating, including the geological attribute features of the existing drilling nodes, and ΔH represents the incremental or updated information of the newly added node features.

[0095] The online learning mechanism unit includes: dynamically adjusting model parameters and optimizing weights through incremental training:

[0096]

[0097] Among them, θ t+1 represents the updated model parameters, including the weights and bias parameters of the graph convolution layer, temporal convolution layer, or fully connected layer, θ t represents the model parameters before the update, which is the parameter set used by the model at the current moment. η represents the learning rate, which controls the step size of the model parameter update and is usually a small positive value. t represents the predicted value of the model at the current time t, that is, the attribute prediction result of the newly added drilling node, Y true Indicates the true attribute value of the newly added drilling node, which is used to calculate the prediction error. It represents the gradient of the loss function L to the model parameter θ, reflecting the influence of the current parameter on the error.

[0098] Furthermore, the grouting requirement index prediction layer includes: calculating the grouting requirement index based on the predicted attribute value:

[0099]

[0100] Among them, R injection (x) represents the grouting demand index, which reflects the comprehensive evaluation value of the grouting demand in the current drilling area, w i represents the weight of the i-th attribute in the grouting demand, f i (y i ) represents the i-th attribute value y i A nonlinear mapping function, n represents the total number of attributes, usually including multiple geological attributes such as fracture density, porosity, permeability, elastic modulus, etc.

[0101] Furthermore, the attribute prediction layer predicts the attributes of the drilling nodes at future moments through the fully connected layer (MLP):

[0102] Y t+1 =MLP(F);

[0103] Among them, Y t+1It represents the predicted value of the borehole node attribute at the future time t+1, including fracture density, porosity, etc. F represents the comprehensive feature representation, which integrates the information of spatial features and temporal features.

[0104] Furthermore, the method of determining whether the current borehole is an un-grouted area based on the predicted geological conditions of each borehole includes: using sensor monitoring or re-drilling to reveal data to detect whether there are traces of slurry in the current borehole. If there are traces of slurry, it indicates that the current borehole is a grouted area; if there are no traces of slurry, it indicates that the current borehole is an un-grouted area.

[0105] Further, S102: injecting slurry into the grouting holes in the un-grouted area, wherein the difference between grouting holes and boreholes is that grouting holes are holes specifically used to inject slurry to improve the properties of the formation or control permeability. Grouting holes are usually designed and optimized to ensure that the slurry can be evenly diffused in the target area to achieve the expected reinforcement or water plugging effect. Boreholes are holes excavated in the formation by drilling technology, which are used to survey geological information, reveal formation characteristics or provide conditions for subsequent construction (such as grouting, testing, etc.).

[0106] Furthermore, S103: according to the predicted geological conditions of each borehole obtained in S102, determine whether the current borehole is a grouted area or an ungrouted area. If there is slurry in the borehole, it means that the current borehole is a grouted area; if there is no slurry in the borehole, it means that the current borehole is an ungrouted area.

[0107] Further, in S103: if it is an un-grouted area, the predicted demand index of the un-grouted area is predicted by using the dynamic update and prediction model of the un-grouted area, and if the index is greater than the set threshold, return to S102; wherein the dynamic update and prediction model of the un-grouted area refers to:

[0108] Getting the predicted attribute value Finally, the grouting demand index is calculated by combining the nonlinear weighting function;

[0109]

[0110] Among them, R injection (x) represents the grouting demand index, f i (·) is a nonlinear mapping function, such as an exponential or logarithmic function, which reflects the influence of various geological attributes on the grouting requirements; w i is the weight of each attribute, obtained through training or expert experience.

[0111] in, A nonlinear mapping that represents crack properties; the width and density of cracks are usually positively correlated with the grouting requirement. The more developed the cracks are, the greater the grouting requirement is. Therefore, a piecewise nonlinear mapping can be selected:

[0112]

[0113] Among them, k 1 ,k 2 is the coefficient, T 1 is the threshold value of the crack attribute.

[0114]

[0115] in, Represents a nonlinear mapping of water content. The effect of water content on grouting demand may be nonlinearly increasing or decreasing, depending on whether the water richness of the formation hinders or promotes the diffusion of slurry.

[0116]

[0117] in, Represents the nonlinear mapping of porosity. The effect of porosity on grouting demand is usually positively correlated, but the effect may tend to saturation at high porosity. The nonlinear mapping of porosity takes the form of a saturated function.

[0118]

[0119] in, It represents the nonlinear mapping of elastic modulus. The elastic modulus is negatively correlated with the grouting demand. The greater the rigidity, the lower the grouting demand. The nonlinear mapping of elastic modulus uses a logarithmic decreasing function.

[0120] Further, in S103: if it is a grouting area, the predicted geological conditions of the current borehole are input into the fitting prediction model of the grouting area to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to S102; if it has met the preset requirements, stop grouting; wherein, the fitting prediction model of the grouting area refers to:

[0121] The radial basis function (RBF) or Kriging interpolation method is used to fit the grouting distribution to obtain the spatial variation of the properties of the area. For the prediction of the grouting distribution at point x, the interpolation formula is as follows:

[0122]

[0123] Among them, α i is the weight coefficient, obtained by the fitting process; φ(||xx i ||) is the radial basis function, usually a Gaussian kernel or a polynomial kernel function; x i represents the known drilling position. f(x) represents the slurry distribution fitting function.

[0124] Gradient fitting optimization: In the grouting area, the refined interpolation of the control area is carried out by gradient fitting, using the following gradient formula:

[0125]

[0126] in, represents the gradient of the radial basis function, represents the gradient of the fitting function f(x), α i Represents the weight coefficient.

[0127] The gradient information is used to optimize the distribution of interpolation and improve the fitting accuracy of data in the grouting area.

[0128] According to the engineering design requirements, define the reference values ​​of the grouting effect, such as the target permeability reduction value, slurry coverage radius, crack density reduction degree, etc. Set a threshold range with clear physical meaning for each reference indicator, for example: the permeability reduction value must be greater than or equal to a certain target value; the slurry coverage range must be greater than or equal to a certain distance range; the crack density reduction ratio must be greater than or equal to a certain percentage. Figure 3 Schematic diagram of the tunnel face, fractures, boreholes and aquifers in Example 1.

[0129] The newly added boreholes are regarded as the time of the space-time series, and the attribute information is regarded as the information to be predicted. The dynamic space-time graph update and attribute information prediction are realized through the deep learning model.

[0130] 1. Construction of spatiotemporal data and generation of initial model: Initial borehole data integration: Use existing borehole data to obtain attribute information (such as fracture density, porosity, permeability, etc.), build an initial three-dimensional spatiotemporal grid model, standardize these data and map them into a unified three-dimensional coordinate system.

[0131] Creation of spatiotemporal sequence: The attribute value of each borehole data is used as a data point in the time series, and arranged in sequence according to the exposure time or construction time of the borehole to form a spatiotemporal sequence.

[0132] 2. Deep learning modeling of space-time graph. The definition of space-time graph: the borehole information and spatial position are used as the nodes of the graph, and the relative spatial distance between boreholes is used as the edge to construct a dynamic space-time graph. The attribute information of each node (such as crack density, permeability, etc.) is used as the feature to be predicted.

[0133] Deep learning model selection: Combine spatiotemporal graph neural networks (such as ST-GCN, ASTGCN) or spatiotemporal convolutional neural networks (such as T-GCN) to model spatiotemporal graphs. This model updates and predicts attribute information through node features and time relationships.

[0134] 3. Data preprocessing and feature engineering, data standardization: Standardize all newly added borehole data to make the data collected at different time points consistent in dimension so that the model can effectively identify features in different time series. Fusion of time series and spatial features: Extract time series features (such as changes in crack density and water seepage rate at different time points) and spatial features (such as spatial relationships between boreholes and differences in adjacent attributes), and input them into the model for feature extraction.

[0135] 4. Model training and prediction, dynamic update and online training: When new drilling data is added, the space-time diagram is updated in real time, and the new data points are used as new time nodes. The model is trained using online learning, so that the model can continuously learn new data and improve the accuracy of the prediction.

[0136] Predicting attribute information: The model predicts the spatiotemporal sequence of newly added drilling nodes, outputs the attribute values ​​of each node at future moments, such as fracture density, porosity, permeability, etc., and constructs dynamic prediction results. Data increment and model iteration: Through incremental training, new drilling data and real-time monitoring information are introduced, and model parameters are updated to enable it to have the ability to predict spatiotemporal changes.

[0137] 5. Intelligent grouting demand assessment based on prediction, nonlinear mapping of attribute information and grouting demand: By constructing a nonlinear relationship function between attribute information (such as crack density, porosity, etc.) and grouting demand, the attribute information predicted by the model is converted into a grouting demand index. Dynamic adjustment of demand in spatiotemporal prediction: With the update of drilling data and real-time monitoring, the grouting demand threshold is dynamically adjusted to identify newly emerging high-demand areas.

[0138] 6. Dynamic visualization and model verification, space-time visualization: Real-time display of predicted space-time attribute graphs, showing attribute changes at each location and regional distribution of grouting requirements, supporting decision-making during construction.

[0139] Model validation and physical consistency check: Use historical data and borehole monitoring data to validate the model's predictions and ensure the physical consistency of the predictions (e.g., the relationship between permeability and porosity, and the relationship between fracture density and grouting requirements).

[0140] 7. After grouting, dynamic update and real-time guidance based on re-drilling information, dynamic update of the grouted area: In the post-grouting stage, according to the slurry distribution revealed by the new drilling, the attribute information of the grouted area (such as reduction in fracture density, change in porosity, etc.) is updated, and these new data are added to the space-time map.

[0141] Fitting prediction update: Through radial basis function interpolation or Kriging interpolation, the attribute distribution of the grouting area is fitted and predicted to capture the geological changes after grouting and adjust the characteristic values ​​of each node.

[0142] Time series update of re-drilling nodes: The new borehole data generated by re-drilling are regarded as new time nodes in the space-time diagram. These data are continuously added to the space-time diagram to dynamically reflect the changes in the grouting area and the un-grouting area.

[0143] Adjustment of demand forecast for ungrouted areas: Real-time forecast update of demand for ungrouted areas. Through time series convolution and spatiotemporal graph prediction methods, the demand index of ungrouted areas is adjusted using new data, and areas that require priority grouting are dynamically divided.

[0144] Consistency and Visualization: Verify the physical consistency of the model (such as the porosity and permeability relationship) and display the demand index of grouted and ungrouted areas in real time in a space-time diagram to help engineering decision makers clearly understand the current demand.

[0145] The spatio-temporal prediction model is constructed based on graph convolutional neural network (GCN) and time series neural network (such as LSTM, GRU). The spatio-temporal graph convolutional network (ST-GCN) structure is used here to integrate the node (drilling data) information and time series features of the spatio-temporal graph.

[0146] Spatial Convolution Layer: The spatial features of each borehole node are extracted through the graph convolution layer, and the geological attributes between boreholes (such as fracture density and porosity) and the correlation between adjacent boreholes are processed. Temporal Convolution Layer: Temporal convolution or LSTM processing is performed on the changes of nodes in the time dimension, and the time series features are extracted and the evolution law of node characteristics over time is modeled. Fusion Layer: The features of spatial convolution and temporal convolution are integrated to generate a comprehensive spatiotemporal feature representation.

[0147] Spatial feature input: The geological attributes of each drilling node (such as fracture density, porosity, permeability, elastic modulus), which can be expressed as a feature vector X = [x 1 ,x 2 ,…,x n ].

[0148] Time feature input: The attribute changes observed in each borehole in the time series, expressed as a time series matrix T = [t 1 ,t 2 ,…,t m ], the attributes of each time point include the geological characteristics of the drilling area and dynamically updated information. Model output:

[0149] Prediction time tn+1 Attribute information of each drill hole These predicted values ​​are the attribute information of each drilling node at future moments, such as fracture density, porosity, permeability, etc.

[0150] The final output grouting demand index R injection (x), a comprehensive grouting demand index is generated based on the predicted attribute information of each node.

[0151] As new drilling data is added, the model is dynamically adjusted during the training process through online learning or recursive iterative updates. If recursive updates are used, the following formula can be used:

[0152]

[0153] Among them, θ represents the model parameters; is the predicted value and the true value Y t The loss function between them; η is the learning rate, which controls the update step size.

[0154] Dynamic update of spatiotemporal data and construction of time series after grouting, initial modeling: In the pre-grouting stage, a three-dimensional spatial model is established based on the existing drilling data, and the information such as cracks, porosity, and permeability are standardized to construct a basic attribute distribution model.

[0155] Post-grouting area update: After grouting, the attribute data of the grouting area (such as porosity and permeability changes) is updated according to the revealed grouting slurry information, and marked as a "grouted" node in the space-time diagram.

[0156] Drilling re-drilling and time series update: The re-drilling of the borehole is regarded as a new node in the time series, and an incremental time series is constructed. The grouting data added by the new node is integrated with the updated time-space diagram of the time series to achieve dynamic adjustment of the grouting area and the un-grouting area.

[0157] Update of nodes and edges after grouting: The re-drilled borehole location is added to the space-time graph as a new node, connected to the adjacent boreholes. The newly added time nodes reflect the newly revealed grouting information, allowing the entire space-time graph to be dynamically expanded.

[0158] Post-grouting property update: The grouting property information of the new node is propagated in the space-time graph through spatial interpolation and temporal recursive update methods to achieve the status update of the grouting area and the un-grouting area.

[0159] Each time a new node is added, the state of the graph node is dynamically updated using the following formula:

[0160]

[0161] Where: H(l) is the feature matrix of all nodes in the graph at layer l; A is the adjacency matrix of the graph, representing the spatial neighborhood relationship; W (l) is the weight matrix of the lth layer, which updates the feature information of the graph; σ is the activation function, which is used to introduce nonlinear transformation.

[0162] After grouting, the demand prediction of the un-grouted area is carried out: the un-grouted area is predicted based on the grouting information on the time-space diagram. The grouting demand of each area (such as crack density and porosity change) is used as the weight to comprehensively calculate its grouting demand index R injection (x).

[0163]

[0164] Time series prediction of new drilling nodes: Using time series neural networks such as LSTM, input the current time series data T = [t 1 ,t 2 ,…,t n ], predict the next time node The property value of the grouting area is updated in real time, and the demand forecast of the un-grouted area is updated in real time. The update formula is as follows:

[0165] h t+1 =σ(W·[h t ,x t ]+b)

[0166] Dynamic Optimization and Consistency Check: Physical consistency check: To ensure the rationality of the prediction of the grouting area, the physical properties such as porosity and permeability are verified for consistency. The consistency relationship between the properties is evaluated by the following formula:

[0167] k fracture =k 0 ·φ m ;

[0168] Where: k fracture is the permeability of the fracture, k 0 is the benchmark permeability, φ is the porosity, m is the empirical index, and the dynamic optimization criterion is: the change rate of each attribute A gradient descent optimization was performed to ensure a smooth transition between the grouted and ungrouted areas.

[0169] Spatial-temporal visualization: Real-time display of the fitted prediction of the grouted area and the demand index of the ungrouted area, marking the grouting areas that need attention, and supporting real-time decision-making during the construction process.

[0170] Engineering application of dynamic prediction: through real-time updated space-time diagram, guide future drilling layout and grouting operations. According to the time series changes of re-drilling node data, risk assessment and construction optimization of un-grouted areas can be carried out. Through this method, continuous tracking and prediction of grouted areas and un-grouted areas can be achieved, and engineering layout and grouting decisions can be guided based on real-time data during the grouting process.

[0171] Embodiment 2

[0172] This embodiment provides a spatiotemporal intelligent prediction system for full-space grouting in sand and gravel formations, including:

[0173] An acquisition module is configured to: acquire the predicted geological conditions of each borehole; determine whether the current borehole is an un-grouted area according to the predicted geological conditions of each borehole;

[0174] A grouting module, which is configured to: inject grout into the grouting holes in the un-grouted area, and after the grouting operation is completed, obtain the predicted geological conditions of each borehole again;

[0175] The prediction module is configured as follows: according to the predicted geological conditions of each borehole, determine whether the current borehole is a grouted area or an ungrouted area; if it is an ungrouted area, use the dynamic update and prediction model of the ungrouted area to predict the prediction demand index of the ungrouted area; if the index is greater than the set threshold, return to the grouting module; if it is a grouted area, input the predicted geological conditions of the current borehole into the fitting prediction model of the grouted area to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to the grouting module; if it has met the preset requirements, stop grouting.

[0176] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel formations, characterized by: include: (1): Obtain the predicted geological conditions of each borehole; determine whether the current borehole is an un-grouted area based on the predicted geological conditions of each borehole; (2): Inject grout into the grouting holes in the ungrouted area. After the grouting operation is completed, obtain the predicted geological conditions of each borehole again; (3): Based on the predicted geological conditions of each borehole, determine whether the current borehole is a grouted area or an ungrouted area. If it is an ungrouted area, use the dynamic update and prediction model of the ungrouted area to predict the prediction demand index of the ungrouted area. If the index is greater than the set threshold, return to (2); if it is a grouted area, input the predicted geological conditions of the current borehole into the grouted area fitting prediction model to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; If the preset requirement is not met, return to (2); if the preset requirement is met, stop grouting.

2. The spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel strata according to claim 1 is characterized in that: The method of obtaining the predicted geological conditions of each borehole includes: drilling a tunnel face to be grouted, the number of boreholes being n, obtaining geophysical information corresponding to each borehole, and inputting the geophysical information corresponding to each borehole into the trained spatiotemporal graph convolutional network in sequence according to the drilling time sequence, so as to obtain the predicted geological conditions of each borehole; The method of obtaining the predicted geological conditions of each borehole again includes: drilling n boreholes in sequence to obtain geophysical information corresponding to each borehole, and inputting the geophysical information corresponding to each borehole into the trained spatiotemporal graph convolutional network in sequence according to the drilling time sequence to obtain the predicted geological conditions of each borehole.

3. The spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel strata according to claim 2 is characterized in that: The geophysical information corresponding to each borehole is sequentially input into the trained spatiotemporal graph convolutional network in the order of drilling time to obtain the predicted geological conditions of each borehole. The training process of the trained spatiotemporal graph convolutional network includes: Constructing a training set, wherein the training set is geophysical information corresponding to each borehole whose geological conditions are known; Input the training set into the spatiotemporal graph convolutional network, train the spatiotemporal graph convolutional network, and stop training when the loss function value of the spatiotemporal graph convolutional network no longer decreases, or when the number of iterations exceeds the set number, to obtain the trained spatiotemporal graph convolutional network; The loss function L of the spatiotemporal graph convolutional network is total , its formula expression is: L total =α1·L prediction +α2·L consistency +α3·L regularization +α4·L robustness ; Among them, α1, α2, α3, and α4 are hyperparameters used to balance the weights of the losses in each part; Among them, L prediction The formula expression is: Among them, L prediction Represents the prediction error loss, which is used to measure the prediction value and the true value y i The error, ω i represents the dynamic weight, β represents the dynamic adjustment parameter, Represents the mean of the data set; The L consistency The formula expression is: Among them, L consistency represents the physical consistency loss function, L consistency Used to constrain the prediction results to conform to the laws of geophysics; R min Represents R injection The physical lower limit, R max Represents R injection The physical upper limit of R injection,i represents the predicted grouting demand index; The L regularization The formula expression is: Among them, λ1 represents the size of the control weight to prevent overfitting; λ2 represents the balance of the control weight; Among them, L robustness The specific expression is: Among them, δ represents the outlier threshold, which controls the degree of penalty for excessive errors, and L robustness represents the robustness loss function, which is used to enhance the robustness of the model to outliers; φ(z) represents the kernel function of the loss function, which is used to process the error between the model predicted value and the true value, z represents the difference between the predicted value and the true value, and N represents the total number of data samples.

4. The spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel strata according to claim 2, characterized in that: The trained spatiotemporal graph convolutional network comprises: Spatial feature input layer, adjacency matrix input layer and temporal feature input layer; The output of the spatial feature input layer and the output of the adjacency matrix input layer are connected to the input of the graph convolution layer, the output of the graph convolution layer is connected to the input of the multi-layer graph convolution module, and the output of the multi-layer graph convolution module is connected to the input of the fusion layer; The output of the temporal feature input layer and the output of the adjacency matrix input layer are both connected to the input of the temporal convolution layer, the output of the temporal convolution layer is connected to the input of the temporal recursive layer, and the output of the temporal recursive layer is connected to the input of the fusion layer; The input end of the fusion layer is also connected to the output end of the attention mechanism layer; the output end of the fusion layer is respectively connected to the input end of the dynamic update module and the input end of the grouting demand index prediction layer, and the dynamic update module and the grouting demand index prediction layer are connected to each other; the output end of the dynamic update module is also connected to the grouting demand index prediction layer through the attribute prediction layer; The spatial feature input layer is used to input the static geological attribute feature matrix of each node in the space-time graph, with a shape of N×F s , where N is the number of nodes and F s is the attribute dimension of each node; The time feature input layer is used to input the time series dynamic features of each node, with a shape of N×T×F t , where T is the number of time steps and F t It is the time characteristic dimension; The adjacency matrix input layer is used to input an adjacency matrix describing the spatial relationship between nodes, which has a shape of N×N and whose weight reflects the spatial connection strength between nodes.

5. The method for spatiotemporal intelligent prediction of full-space grouting in sandy and gravel strata according to claim 4, characterized in that: Convolutional layer, using graph convolution to extract the spatial features of nodes, the formula is: Among them, H (l+1) is the node feature matrix of all nodes in the graph at the l+1 layer, D is the degree matrix, D ii =∑ j A ij , W (l) is the weight matrix of the lth layer; σ is the ReLU activation function, A is the adjacency matrix of the graph, and A represents the spatial relationship of the drilling nodes; The temporal convolution layer uses time series convolution or LSTM to extract features in the time dimension to capture the time series change characteristics; it extracts the dynamic change characteristics of time series data and uses a one-dimensional convolution operation: Among them, j represents the convolution kernel position index in the time series, is the output feature after temporal convolution; θ j are the parameters of the convolution kernel; Represents the input data at time tj, k is the convolution kernel width, and b is the bias; the output of the temporal convolution layer is the temporal feature matrix The shape of the time feature matrix is ​​N×F (L) ; where N represents the number of nodes, F (L) Represents the feature dimension of the last layer.

6. The method for spatiotemporal intelligent prediction of full-space grouting in sandy and gravel strata according to claim 4, characterized in that: The fusion layer transforms the spatial feature matrix and the time feature matrix Perform feature fusion to generate a comprehensive spatiotemporal feature representation F: F=concat(H spatial ,H temporal ); The attention mechanism layer assigns weights to different features and optimizes the feature fusion effect through the attention mechanism: Among them, α i represents the attention weight, e i It represents the score obtained by weighting the feature by the attention weight matrix, e j represents the scores of all features used for normalization, W attention It represents the weight matrix learned during training and is used to optimize the score calculation process. i The representation is the feature of the input, fusing spatial and temporal information.

7. The spatiotemporal intelligent prediction method for full-space grouting in sandy and gravel strata according to claim 4 is characterized in that: The dynamic update module includes: an adjacency matrix update unit, a node feature update unit and an online learning mechanism unit; the adjacency matrix update unit includes: When new drilling data is added, the adjacency matrix A is dynamically adjusted: IN new = Yes old +ΔA; Among them, A new represents the updated adjacency matrix, including the newly added drilling node and its spatial relationship with other nodes, A old represents the adjacency matrix before updating, describing the spatial relationship between existing drilling nodes, ΔA represents the adjacency matrix before updating, describing the spatial relationship between existing drilling nodes; The node feature updating unit includes: interpolating or dynamically updating the features of the newly added nodes: H new =H old +ΔH; Among them, H new represents the updated node feature matrix, including the features of the newly added nodes and the adjusted results of the existing node features. old It represents the node feature matrix before updating, including the geological attribute features of the existing drilling nodes, and ΔH represents the increment or update information of the newly added node features; The online learning mechanism unit includes: dynamically adjusting model parameters and optimizing weights through incremental training: Among them, θ t+1 represents the updated model parameters, including the weights and bias parameters of the graph convolution layer, temporal convolution layer, or fully connected layer, θ t represents the model parameters before the update, which is the parameter set used by the model at the current moment, η represents the learning rate, Y t Represents the predicted value of the model at the current time t, and the attribute prediction result of the newly added drilling node, Y true Indicates the true attribute value of the newly added drilling node, which is used to calculate the prediction error. It represents the gradient of the loss function L to the model parameter θ, reflecting the influence of the current parameter on the error; The grouting demand index prediction layer includes: calculating the grouting demand index based on the predicted attribute values: Among them, R injection (x) represents the grouting demand index, which reflects the comprehensive evaluation value of the grouting demand in the current drilling area, w i represents the weight of the i-th attribute in the grouting demand, f i (y i ) represents the i-th attribute value y i A nonlinear mapping function, n represents the total number of attributes; The attribute prediction layer predicts the attributes of drilling nodes in the future through the fully connected layer: Y t+1 =MLP(F); Among them, Y t+1 It represents the predicted value of the borehole node attribute at the future time t+1, including fracture density and porosity. F represents the comprehensive feature representation.

8. The method for spatiotemporal intelligent prediction of full-space grouting in sandy and gravel strata according to claim 1, characterized in that: The dynamic update and prediction model of the un-grouted area refers to: Getting the predicted attribute value Finally, the grouting demand index is calculated by combining the nonlinear weighting function; Among them, R injection (x) represents the grouting demand index, f i (·) is a nonlinear mapping function, w i is the weight of each attribute; in, Nonlinear mapping representing fracture properties; Among them, k1, k2 are coefficients, T1 is the threshold of crack attributes; in, A nonlinear map representing water content; in, A nonlinear map representing porosity; in, Represents a nonlinear map of elastic moduli.

9. The method for spatiotemporal intelligent prediction of full-space grouting in sandy and gravel strata according to claim 1, characterized in that: The fitting prediction model of the grouting area refers to: The radial basis function or Kriging interpolation method is used to fit the grouting distribution to obtain the spatial variation of the regional attributes; for the prediction of the grouting distribution at point x, the interpolation formula is as follows: Among them, α i is the weight coefficient, obtained by the fitting process; φ(∥xx i ∥) is the radial basis function, usually a Gaussian kernel or a polynomial kernel function; x i represents the known drilling position; f(x) represents the slurry distribution fitting function; Gradient fitting optimization: In the grouting area, the refined interpolation of the control area is carried out by gradient fitting, using the following gradient formula: in, represents the gradient of the radial basis function, represents the gradient of the fitting function f(x), α i Represents the weight coefficient.

10. The spatiotemporal intelligent prediction system for full-space grouting in sand and gravel formations is characterized by: include: An acquisition module is configured to: acquire the predicted geological conditions of each borehole; determine whether the current borehole is an un-grouted area according to the predicted geological conditions of each borehole; A grouting module, which is configured to: inject grout into the grouting holes in the un-grouted area, and after the grouting operation is completed, obtain the predicted geological conditions of each borehole again; The prediction module is configured as follows: based on the predicted geological conditions of each borehole, determine whether the current borehole is a grouted area or an ungrouted area; if it is an ungrouted area, use the dynamic update and prediction model of the ungrouted area to predict the predicted demand index of the ungrouted area; if the index is greater than the set threshold, return to the grouting module; if it is a grouted area, input the predicted geological conditions of the current borehole into the fitting prediction model of the grouted area to obtain the result of whether the grouting conditions of the current borehole meet the preset requirements; if it does not meet the preset requirements, return to the grouting module; if it has met the preset requirements, stop grouting.

Images