Stratum deformation prediction method and system

Abstract

The invention relates to the field of tunnel and underground engineering construction, in particular to a stratum deformation prediction method and system. The method comprises the following steps: optimizing an initial topological structure and an initial adjacent matrix of a stratum to be monitored based on a deformation value of each monitoring point at each historical moment; obtaining a target adjacent matrix corresponding to each historical moment through a first attention mechanism module; respectively inputting the feature matrix, the degree matrix and the target adjacent matrix corresponding to each historical moment into a corresponding graph convolution model for processing, and outputting a predicted value of a preset monitoring point at each historical moment; the predicted values of the preset monitoring point at all historical moments are input to the time convolutional network layer through the full connection layer, and the initial deformation predicted value of the preset monitoring point at the preset moment is output; and sequentially passing through a second attention mechanism module and a full connection layer, and outputting a deformation prediction value of the preset monitoring point at a preset moment. According to the method, the stratum deformation domain characteristics are comprehensively considered, and the stratum deformation prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of tunnel and underground engineering construction, and in particular to a stratum deformation prediction method and system. Background Art

[0002] With the rapid advancement of urban construction in China, urban underground transportation systems have also achieved remarkable development. Due to its high efficiency, safety and economy, shield machines have become widely used equipment in various tunnel projects. Shield machine construction will inevitably cause ground deformation, which may cause safety hazards to surrounding buildings. Excessive deformation may even cause building collapse. In addition, maintaining ground stability is crucial to ensuring the construction efficiency of shield machines. Accurately predicting ground deformation helps engineers understand the construction situation in a timely manner, thereby effectively managing and controlling the quality of shield tunnel construction.

[0003] Traditional ground deformation prediction methods mainly include four categories: empirical formula, theoretical analysis, numerical simulation and model test methods. The empirical formula method was first used to predict ground deformation caused by tunnel excavation. The numerical simulation method is now one of the main research tools for studying ground deformation caused by tunnel excavation in complex environments. Scholars have also proposed some model tests to study the laws of ground deformation during shield tunnel construction. Model tests can restore the true ontological relationship of the soil, but the cost is high and the control requirements for test accuracy are also high. In traditional ground deformation prediction methods, although empirical formulas, theoretical analysis, numerical simulation and model tests each have their own unique advantages and application scenarios, they have certain limitations in dealing with complex situations and dynamic data. In order to overcome these limitations, machine learning (ML) and deep learning (DL) methods have gradually attracted the attention of researchers in recent years. Machine learning (ML) methods mainly use various models to train data to capture the inherent laws of the data and then make predictions. Compared with traditional methods, machine learning (ML) methods have better performance because they have strong learning ability and can better extract data features. However, the performance of these machine learning (ML) methods is limited because they cannot effectively process time series data with different lengths and large data volumes. In order to further improve the prediction performance, deep learning (DL) has been used to predict ground deformation because deep learning (DL) can solve nonlinear complex problems, and many experimental results have also shown that the prediction performance of DL-based methods is more powerful than traditional ML methods.

[0004] However, formation deformation is a complex and dynamic interactive process. Existing machine learning (ML) and deep learning (DL) methods only process grid-based data and divide the space into discrete units, such as quadrilaterals, while ignoring the graph structure of monitoring points, resulting in poor prediction accuracy of formation deformation. In addition, existing models mainly focus on the mathematical characteristics of formation deformation data, while ignoring the field characteristics of formation deformation, which makes the existing models less interpretable. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the defects in the prior art that the graph structure of the monitoring points is ignored, resulting in poor accuracy in prediction of stratum deformation and ignoring the field characteristics of stratum deformation, making the existing model less interpretable.

[0006] In order to solve the above technical problems, the present invention provides a method for predicting stratum deformation, comprising the following steps:

[0007] Acquire a monitoring data set of the stratum to be monitored, the monitoring data set including: a monitoring point layout diagram, deformation values ​​of each monitoring point at different historical moments, and stratum deformation parameter data;

[0008] The monitoring data set of the stratum to be monitored is input into the stratum deformation prediction model, the stratum deformation value of the stratum to be monitored is predicted, and the deformation prediction value of the preset monitoring point at the preset time is output, including:

[0009] Based on the monitoring point layout diagram, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed;

[0010] Based on the deformation value of each monitoring point at each historical moment, the initial topological structure and the initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment. Based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained.

[0011] Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment;

[0012] Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained;

[0013] The feature matrix, degree matrix, and target adjacency matrix corresponding to each historical moment are input into the corresponding graph convolution model respectively, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output;

[0014] After the predicted values ​​of the preset monitoring points at all historical moments pass through the fully connected layer, they are input into the temporal convolutional network layer to capture the patterns and dependencies in the sequence and output the initial deformation prediction value of the preset monitoring points at the preset time.

[0015] The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

[0016] Preferably, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed based on the monitoring point arrangement diagram, including:

[0017] Based on the monitoring point layout diagram, each monitoring point is regarded as a node, and each node is assumed to be connected to form an edge between the nodes;

[0018] Based on each node and the edges between nodes, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed.

[0019] Preferably, the initial topological structure and the initial adjacency matrix are optimized based on the deformation value of each monitoring point at each historical moment to obtain the optimized topological structure and adjacency matrix at each historical moment, including:

[0020] Based on the deformation value of each monitoring point at each historical moment, determine the deformation influence range of each monitoring point at each historical moment;

[0021] Based on the deformation influence range of each monitoring point at each historical moment, the connection relationship between the nodes is confirmed, and the edges between the nodes in the initial topological structure of the monitored stratum are deleted to obtain the optimized topological structure and its adjacency matrix at each historical moment.

[0022] Preferably, based on the deformation value of each monitoring point at each historical moment, the deformation influence range of each monitoring point at each historical moment is calculated respectively by the Peck formula.

[0023] Preferably, the step of obtaining the degree matrix corresponding to each historical moment based on the optimized topological structure at each historical moment includes:

[0024] Construct a×a zero matrix corresponding to each historical moment, where a is the total number of monitoring points of the stratum to be monitored;

[0025] Based on the optimized topological structure at each historical moment, the number of edges associated with each node at each historical moment and each other node is obtained, and the number of edges associated with each node at each historical moment and each other node is used as the elements of the main diagonal of the a×a zero matrix corresponding to each historical moment to obtain the degree matrix corresponding to each historical moment.

[0026] Preferably, the graph convolution model includes multiple graph convolutional network layers and bidirectional long short-term memory network layers.

[0027] Preferably, the feature matrix, degree matrix and target adjacency matrix corresponding to the current historical moment are respectively input into the corresponding graph convolution model, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at the current historical moment are output, including the following steps:

[0028] The feature matrix, degree matrix, and target adjacency matrix corresponding to the current historical moment are combined to obtain the graph structure corresponding to the current historical moment;

[0029] The graph structure corresponding to the current historical moment passes through the first graph convolutional network layer to extract the spatial features of the deformed data and obtain the output features of the first graph convolutional network layer;

[0030] The output of the i-th graph convolutional network layer is linearly transformed through the ReLu activation function, and then input into the i-th bidirectional long short-term memory network layer to update the weight parameters, and obtain the output features of the i-th bidirectional long short-term memory network layer;

[0031] The output features of the i-th bidirectional long short-term memory network layer are used as the input of the i+1-th graph convolutional network layer, where i=1, 2…N-1, and N represents the total number of graph convolutional network layers;

[0032] The output features of the Nth graph convolutional network layer pass through the ReLu activation function to output the predicted value of the preset monitoring point at the current historical moment.

[0033] Preferably, the step of obtaining a characteristic matrix corresponding to each historical moment based on the formation deformation parameter data at each historical moment includes:

[0034] The stratum deformation parameter data at each historical moment is converted into a tensor structure through the tensor function to obtain the feature matrix corresponding to each historical moment.

[0035] Preferably, the stratum deformation parameter data include: geological parameters, shield parameters, structural parameters and tunnel parameters.

[0036] The present invention also provides a stratum deformation prediction system, comprising:

[0037] A data acquisition module is used to acquire a monitoring data set of the stratum to be monitored, wherein the monitoring data set includes: a monitoring point layout diagram, a deformation value of each monitoring point at different historical moments, and stratum deformation parameter data;

[0038] The prediction module is used to input the monitoring data set of the stratum to be monitored into the stratum deformation prediction model, predict the stratum deformation value of the stratum to be monitored, and output the deformation prediction value of the preset monitoring point at the preset time, including:

[0039] Based on the monitoring point layout diagram, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed;

[0040] Based on the deformation value of each monitoring point at each historical moment, the initial topological structure and the initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment. Based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained.

[0041] Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment;

[0042] Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained;

[0043] The feature matrix, degree matrix, and target adjacency matrix corresponding to each historical moment are input into the corresponding graph convolution model respectively, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output;

[0044] After the predicted values ​​of the preset monitoring points at all historical moments pass through the fully connected layer, they are input into the temporal convolutional network layer to capture the patterns and dependencies in the sequence and output the initial deformation prediction value of the preset monitoring points at the preset time.

[0045] The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

[0046] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0047] The present invention discloses a method and system for predicting stratum deformation. The present invention not only considers the mathematical characteristics of stratum deformation data, but also fully considers the characteristics of stratum deformation field. Based on the monitoring point layout diagram, the initial topological structure and initial adjacency matrix of the stratum to be monitored are constructed; based on the deformation value of each monitoring point at each historical moment, the initial topological structure and initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment; based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained; the optimized adjacency matrix at each historical moment is respectively input into the first attention mechanism module, and different weights are assigned to each edge in the optimized topological structure at each historical moment to obtain the degree matrix corresponding to each historical moment. The corresponding target adjacency matrix; based on the stratum deformation parameter data at each historical moment, the feature matrix corresponding to each historical moment is obtained; the feature matrix, degree matrix and target adjacency matrix corresponding to each historical moment are respectively input into the corresponding graph convolution model for processing, the spatial characteristics of the monitoring data are mined, and the predicted value of the preset monitoring point at each historical moment is output; after the predicted value of the preset monitoring point at all historical moments passes through the fully connected layer, it is input into the time convolution network layer to capture the pattern and dependency in the sequence, and the initial deformation prediction value of the preset monitoring point at the preset moment is output; the initial deformation prediction value of the preset monitoring point at the preset moment passes through the second attention mechanism module and the fully connected layer in sequence, and the deformation prediction value of the preset monitoring point at the preset moment is output. The present invention makes full use of the graph structure of the monitoring point, so that the stratum deformation prediction model is more interpretable, the stratum deformation prediction is more comprehensive and accurate, and the graph convolution model and the time convolution network layer are adopted. The graph convolution model and the time convolution network layer can effectively capture the dynamic spatiotemporal characteristics and domain characteristics in the ground deformation data, and improve the accuracy of stratum deformation prediction.

[0048] In addition, a bidirectional long short-term memory layer is introduced into the graph convolution model, and the weight parameters of the graph structure of the monitored stratum at each historical moment are dynamically updated by the bidirectional long short-term memory layer. The bidirectional long short-term memory layer adopts a dynamic adjacency matrix calculation method, which can describe the dynamic spatial changes of the graph structure and better capture the characteristics and laws of stratum deformation, thereby improving the accuracy and effectiveness of the stratum deformation prediction model in predicting stratum deformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0050] Figure 1 It is a structural diagram of the stratum deformation prediction model.

[0051] Figure 2 It is a schematic diagram of the deformation curve.

[0052] Figure 3 It is a schematic diagram of the impact range caused by the deformation of the monitoring point.

[0053] Figure 4 It is a schematic diagram of the optimization process of the initial topology structure and the initial adjacency matrix.

[0054] Figure 5 It is a schematic diagram of the weight update of the edge of the attention mechanism module.

[0055] Figure 6 It is a diagram of the graph convolution model structure.

[0056] Figure 7 Schematic diagram of the computation of causal convolutions in a temporal convolutional network layer dilation.

[0057] Figure 8 It is the engineering plan view in the second embodiment of the present invention.

[0058] Fig. 9 It is a cross-sectional view of the project in the second embodiment of the present invention.

[0059] Fig.10 This is a monitoring point arrangement diagram in the second embodiment of the present invention.

[0060] Fig.11 The present invention is a flow chart of a method for predicting formation deformation.

[0061] Fig.12 Schematic diagram of the prediction results of the single-step prediction model for stratum deformation.

[0062] Fig.13 Schematic diagram of the prediction results of the T-GCN model under single-step prediction.

[0063] Fig.14 This is a schematic diagram of the prediction results of the ConvLSTM model under single-step prediction. DETAILED DESCRIPTION

[0064] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0065] Reference Fig.11 As shown, the first embodiment of the present invention provides a method for predicting formation deformation, comprising the following steps:

[0066] Acquire a monitoring data set of the stratum to be monitored, the monitoring data set including: a monitoring point layout diagram, deformation values ​​of each monitoring point at different historical moments, and stratum deformation parameter data;

[0067] In this embodiment, the positions of the monitoring points for measuring the formation deformation are determined, and instruments are set up to perform real-time monitoring and recording, which are recorded as It represents the deformation value of the rth monitoring point at the nth moment, r represents the rth monitoring point, and n represents the moment. Similarly, the stratum deformation parameter data in the shield tunneling system are collected and sorted according to the time series, and the data are summarized and sorted into an Excel table.

[0068] In this embodiment, preferably, the stratum deformation parameter data includes: geological parameters, shield parameters, structural parameters and tunnel parameters, such as thrust, torque, soil bin pressure, etc., and geological parameters collected and sorted according to geological survey reports, design data, etc.

[0069] like Figure 1 As shown, Figure 1 It is a structural diagram of the stratum deformation prediction model. The monitoring data set is input into the stratum deformation prediction model, the deformation prediction value is predicted, and the deformation prediction value of the preset monitoring point at the preset time is output.

[0070] Based on the monitoring point layout diagram, the initial topological structure G′ of the stratum to be monitored is constructed n and its initial adjacency matrix A′ n ;

[0071] In this embodiment, preferably, the initial topological structure G′ of the stratum to be monitored is constructed based on the monitoring point arrangement diagram. n and its initial adjacency matrix A′ n ,include:

[0072] Based on the monitoring point layout diagram, each monitoring point is regarded as a node, and each node is assumed to be connected to form an edge between the nodes;

[0073] Based on each node and the edges between nodes, the initial topological structure G′ of the stratum to be monitored is constructed n and its initial adjacency matrix A′ n .

[0074] Based on the deformation value of each monitoring point at each historical moment, the initial topological structure G′ of the monitored stratum is n and its initial adjacency matrix A′ n Optimize and obtain the optimized topological structure G of the stratum to be monitored at each historical moment n and its adjacency matrix A n , based on the optimized topological structure at each historical moment, obtain the degree matrix corresponding to each historical moment;

[0075] In this embodiment, preferably, Figure 4 As shown, Figure 4 is a schematic diagram of the optimization process of the initial topological structure and the initial adjacency matrix. In this embodiment, preferably, the initial topological structure G′ of the monitored stratum is determined based on the deformation value of each monitoring point at each historical moment. nand its initial adjacency matrix A′ n Optimize and obtain the optimized topological structure G of the stratum to be monitored at each historical moment n and its adjacency matrix A n ,include:

[0076] Based on the deformation value of each monitoring point at each historical moment, determine the deformation influence range of each monitoring point at each historical moment;

[0077] Based on the deformation influence range of each monitoring point at each historical moment, the connection relationship between the nodes is confirmed, and the edges between the nodes in the initial topological structure of the monitored stratum are deleted to obtain the optimized topological structure and its adjacency matrix at each historical moment.

[0078] G′ n represents the initial topological structure of the stratum to be monitored at the nth moment, A′ n represents the initial adjacency matrix of the stratum to be monitored at the nth moment, G n represents the topological structure of the stratum to be monitored after optimization at the nth moment, A n represents the adjacency matrix of the stratum to be monitored after optimization at the nth moment, n = 1, 2, ... T, T represents the total number of historical moments;

[0079] In this embodiment, preferably, Figure 2 , Figure 3 As shown, Figure 2 is a schematic diagram of the deformation curve. Figure 3 It is a schematic diagram of the influence range caused by the deformation of the monitoring point. In this embodiment, preferably, based on the deformation value of each monitoring point at each historical moment, the deformation influence range of each monitoring point at each historical moment is calculated by the Peck formula. The Peck formula can provide a more accurate estimation of the deformation influence range based on actual monitoring data, combined with geological and soil characteristics, and improve the prediction accuracy of the stratum deformation prediction model.

[0080] In this embodiment, preferably, the step of obtaining the degree matrix corresponding to each historical moment based on the optimized topological structure at each historical moment includes:

[0081] Construct a×a zero matrix corresponding to each historical moment, where a is the total number of monitoring points of the stratum to be monitored;

[0082] Based on the optimized topological structure at each historical moment, the number of edges associated with each node at each historical moment and each other node is obtained, and the number of edges associated with each node at each historical moment and each other node is used as the elements of the main diagonal of the a×a zero matrix corresponding to each historical moment to obtain the degree matrix corresponding to each historical moment.

[0083] like Figure 5 As shown, Figure 5 This is a schematic diagram of the weight update of the edge of the attention mechanism module. Considering that in actual situations, the data from different monitoring points also have spatial correlation, that is, the data of a certain monitoring point is affected by the spatial distance of other monitoring points, and the degree of influence varies with the change of spatial distance, the attention mechanism is introduced.

[0084] Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment;

[0085] Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained;

[0086] In this embodiment, preferably, the step of obtaining a characteristic matrix corresponding to each historical moment based on the formation deformation parameter data at each historical moment includes:

[0087] The stratum deformation parameter data at each historical moment is converted into a tensor structure through the tensor function to obtain the feature matrix corresponding to each historical moment.

[0088] The graph structure corresponding to each historical moment is obtained by combining the feature matrix, degree matrix, and target adjacency matrix corresponding to the optimized topological structure at each historical moment.

[0089] like Figure 6 As shown, Figure 6 The graph convolution model structure diagram shows that the feature matrix, degree matrix and target adjacency matrix corresponding to each historical moment are respectively input into the corresponding graph convolution model (GCN model), the spatial characteristics of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output;

[0090] In this embodiment, preferably, the graph convolution model includes multiple graph convolutional network layers (GCN layers) and bidirectional long short-term memory network layers (BiLSTM layers).

[0091] In this embodiment, preferably, the feature matrix, degree matrix and target adjacency matrix corresponding to the current historical moment are respectively input into the corresponding graph convolution model, the spatial features of the monitoring data are mined, and the predicted value of the preset monitoring point at the current historical moment is output, including the following steps:

[0092] The feature matrix, degree matrix, and target adjacency matrix corresponding to the current historical moment are combined to obtain the graph structure corresponding to the current historical moment;

[0093] The graph structure corresponding to the current historical moment passes through the first graph convolutional network layer to extract the spatial features of the deformed data and obtain the output features of the first graph convolutional network layer;

[0094] The output of the i-th graph convolutional network layer is linearly transformed through the ReLu activation function, and then input into the i-th bidirectional long short-term memory network layer to update the weight parameters, and obtain the output features of the i-th bidirectional long short-term memory network layer;

[0095] The output features of the i-th bidirectional long short-term memory network layer are used as the input of the i+1-th graph convolutional network layer, where i=1, 2…N-1, and N represents the total number of graph convolutional network layers;

[0096] The output features of the Nth graph convolutional network layer pass through the ReLu activation function to output the predicted value of the preset monitoring point at the current historical moment.

[0097] like Figure 7 As shown, Figure 7 This is a schematic diagram of the expanded causal convolution calculation of the temporal convolutional network layer. The predicted values ​​of the preset monitoring points at all historical moments are passed through the fully connected layer and then input into the temporal convolutional network layer (TCN layer) to capture the patterns and dependencies in the sequence, and output the initial deformation prediction value of the preset monitoring point at the preset moment. The temporal convolutional network layer (TCN layer) captures the patterns and dependencies in the sequence through convolution operations. Compared with recurrent neural networks, the temporal convolutional network layer (TCN layer) can adopt parallel computing, which can effectively shorten the calculation time.

[0098] The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

[0099] The temporal convolutional network layer (TCN layer) gives the same weight to the input time series features during the training process, which may lead to insufficient attention to some important features. In order to solve this problem and improve the accuracy of the formation deformation prediction model, the present invention introduces an attention mechanism to highlight the importance of features at different time nodes. Assuming that the output hidden state of the TCN layer is Y = [Y1, Y2, ..., Y p ]. p represents the number of output hidden states of the TCN layer, and the attention score calculation formula is:

[0100] u j =tanh(W j Y j +b j )

[0101]

[0102] Among them, Wj represents the weight of the jth element in the output hidden state of the TCN layer, b j represents the bias vector of the jth element in the output hidden state of the TCN layer, α j Indicates the weight coefficient corresponding to the j-th element attention mechanism in the output hidden state of the TCN layer, u j represents the attention score of the jth element in the output hidden state of the TCN layer, Y j represents the jth element in the output hidden state of the TCN layer, and S represents the output of the attention mechanism module.

[0103] like Figure 8 , 10 As shown, Fig.10 This is the engineering plan view in the second embodiment of the present invention. Fig.10 This is a diagram of the arrangement of monitoring points in the second embodiment of the present invention. The second embodiment of the present invention applies the above-mentioned ground deformation prediction method to an engineering example of a subway tunnel section project of a certain urban rail transit.

[0104] Subway Line 12 is a backbone line in the Beijing subway network, spanning Beijing from east to west, with a total length of 26.55 kilometers. The new subway tunnel is excavated using the shield method, and the shield segment size is 6400mm in outer diameter, 300mm in thickness, and 1200mm in width. When the shield is constructed along the North Third Ring Road East from Sanyuanqiao Station to Xibahe Station, it is necessary to pass under the existing Xibahe Bridge. Both the left and right line tunnels are excavated from Sanyuanqiao Station to Xibahe Station, from Figure 7 It can be seen that the left tunnel needs to pass through two piles with a diameter of 1.0m and 1.2m respectively, and the right tunnel needs to pass through four piles with a diameter of 1.2m and eight piles with a diameter of 1.0m. The tunnel needs to pass through a total of 16 reinforced concrete piles, and the diameter of the steel bars contained in the piles is 20-28mm.

[0105] like Fig. 9 As shown, Fig. 9 This is a cross-sectional view of the project in Example 2 of the present invention. From the surface downward, there are miscellaneous fill (first layer), clay (second layer), silty clay (second layer), silt (second layer), silty clay (third layer), sandy silt (third layer), rounded gravel (third layer), silt (fourth layer), fine sand (fifth layer), silty clay (sixth layer), and clay (sixth layer). The miscellaneous fill soil is uneven in hardness and softness, and most of them have loose structures, high compressibility, and low bearing capacity. When there is an external load, its ability to resist deformation is poor and it is easy to sink or even collapse. In this project, the shield machine mainly passes through fine sand (fifth layer), silty clay (sixth layer) and clay (sixth layer).

[0106] In this embodiment, relevant data of geological parameters, shield parameters, structural parameters and tunnel parameters are obtained through monitoring instruments, shield information acquisition systems, geological survey reports, design reports and other materials;

[0107] Through correlation analysis, appropriate input parameters are selected, data preprocessing is performed, and the ratio of the test set to the training set is determined. Then, the optimal hyperparameters of the formation deformation prediction model are determined through experiments on the training set. For example, according to the number of iterations, learning rate, loss rate, number of neural units, etc., the deviation between the deformation prediction value and the actual value of the preset monitoring point at the preset time is used to optimize the loss function. Finally, the trained formation deformation prediction model is verified through the validation set, and the performance of the trained formation deformation prediction model is evaluated using evaluation indicators to obtain the target formation deformation prediction model.

[0108] The performance of the trained formation deformation prediction model is evaluated by using evaluation indicators, which include root mean square error, mean absolute error and determination coefficient. The calculation formulas are:

[0109]

[0110] Among them, α RMSE represents the root mean square error, α MAE represents the mean absolute error, R 2 represents the coefficient of determination, y q represents the actual value of the formation deformation at the preset monitoring point at the preset time of the qth validation set sample, Y q represents the deformation prediction value of the preset monitoring point at the preset time of the qth validation set sample, It represents the average value of the actual value of the formation deformation at the preset monitoring point at the preset time of m validation set samples, and m represents the total number of samples in the validation set.

[0111] In the process of predicting stratum deformation, the present invention takes into account the domain knowledge of data, overcomes the shortcoming of poor model interpretability caused by pure data-driven traditional prediction models, and proposes a dynamic adjacency matrix calculation method based on stratum deformation domain knowledge and attention mechanism, which can describe the dynamic spatial changes of the graph structure of the stratum to be monitored.

[0112] In order to further verify the effectiveness of the formation deformation prediction model proposed in this invention, the existing convolutional LSTM neural network and temporal graph convolutional network with good prediction performance were selected for comparison, and two prediction tasks were designed to reflect the prediction effect of the prediction model under different prediction step sizes.

[0113] As shown in Table 1, Table 1 shows the prediction results of different models under single-step prediction. The evaluation indicators include: root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ).

[0114] Table 1 Prediction results of different models under single-step prediction

[0115]

[0116] As shown in Table 2, Table 2 shows the prediction results of different models under multi-step prediction. The evaluation indicators include: root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ).

[0117] Table 2 Prediction results of the model under multi-step prediction

[0118]

[0119] like Fig.12 , 13 , 14, Fig.12 This is a schematic diagram of the prediction results of the single-step prediction model for ground deformation. Fig.13 This is a schematic diagram of the prediction results of the T-GCN model under single-step prediction. Fig.14 This is a schematic diagram of the prediction results of the convolutional long short-term memory network (ConvLSTM model) under single-step prediction. Tables 1 and 2 are similar to Fig.12 , 13 ,14 The results show that in all prediction tasks, the performance of the stratum deformation prediction model proposed in the present invention is better than the existing convolutional long short-term memory network (ConvLSTM model) and temporal graph convolutional network (T-GCN) with better prediction performance, and a smaller error value is obtained on the test set. It can be concluded that the stratum deformation prediction model provided by the present invention has better stability and higher prediction accuracy in the face of the characteristics of large number, high dimension and time-space interaction of stratum data, and is more suitable for dynamic prediction of stratum deformation.

[0120] The third embodiment also provides a system for predicting formation deformation, including:

[0121] A data acquisition module is used to acquire a monitoring data set of the stratum to be monitored, wherein the monitoring data set includes: a monitoring point layout diagram, a deformation value of each monitoring point at different historical moments, and stratum deformation parameter data;

[0122] The prediction module is used to input the monitoring data set of the stratum to be monitored into the stratum deformation prediction model, predict the stratum deformation value of the stratum to be monitored, and output the deformation prediction value of the preset monitoring point at the preset time, including:

[0123] Based on the monitoring point layout diagram, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed;

[0124] Based on the deformation value of each monitoring point at each historical moment, the initial topological structure and the initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment. Based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained.

[0125] Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment;

[0126] Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained;

[0127] The feature matrix, degree matrix, and target adjacency matrix corresponding to each historical moment are input into the corresponding graph convolution model respectively, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output;

[0128] After the predicted values ​​of the preset monitoring points at all historical moments pass through the fully connected layer, they are input into the temporal convolutional network layer to capture the patterns and dependencies in the sequence and output the initial deformation prediction value of the preset monitoring points at the preset time.

[0129] The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

[0130] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0131] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0134] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for predicting stratum deformation, characterized in that: The following steps are involved: Acquire a monitoring data set of the stratum to be monitored, the monitoring data set including: a monitoring point layout diagram, deformation values ​​of each monitoring point at different historical moments, and stratum deformation parameter data; The monitoring data set of the stratum to be monitored is input into the stratum deformation prediction model, the stratum deformation value of the stratum to be monitored is predicted, and the deformation prediction value of the preset monitoring point at the preset time is output, including: Based on the monitoring point layout diagram, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed; Based on the deformation value of each monitoring point at each historical moment, the initial topological structure and the initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment. Based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained. Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment; Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained; The feature matrix, degree matrix, and target adjacency matrix corresponding to each historical moment are input into the corresponding graph convolution model respectively, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output; After the predicted values ​​of the preset monitoring points at all historical moments pass through the fully connected layer, they are input into the temporal convolutional network layer to capture the patterns and dependencies in the sequence and output the initial deformation prediction value of the preset monitoring points at the preset time. The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

2. A method for predicting stratum deformation according to claim 1, characterized in that: The initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed based on the monitoring point arrangement diagram, including: Based on the monitoring point layout diagram, each monitoring point is regarded as a node, and each node is assumed to be connected to form an edge between the nodes; Based on each node and the edges between nodes, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed.

3. A method for predicting stratum deformation according to claim 2, characterized in that: The initial topological structure and the initial adjacency matrix are optimized based on the deformation value of each monitoring point at each historical moment to obtain the optimized topological structure and adjacency matrix at each historical moment, including: Based on the deformation value of each monitoring point at each historical moment, determine the deformation influence range of each monitoring point at each historical moment; Based on the deformation influence range of each monitoring point at each historical moment, the connection relationship between the nodes is confirmed, and the edges between the nodes in the initial topological structure of the monitored stratum are deleted to obtain the optimized topological structure and its adjacency matrix at each historical moment.

4. A method for predicting stratum deformation according to claim 3, characterized in that: Based on the deformation value of each monitoring point at each historical moment, the deformation influence range of each monitoring point at each historical moment is calculated by the Peck formula.

5. A method for predicting stratum deformation according to claim 3, characterized in that: The step of obtaining the degree matrix corresponding to each historical moment based on the optimized topological structure at each historical moment includes: Construct a×a zero matrix corresponding to each historical moment, where a is the total number of monitoring points of the stratum to be monitored; Based on the optimized topological structure at each historical moment, the number of edges associated with each node at each historical moment and each other node is obtained, and the number of edges associated with each node at each historical moment and each other node is used as the elements of the main diagonal of the a×a zero matrix corresponding to each historical moment to obtain the degree matrix corresponding to each historical moment.

6. A method for predicting stratum deformation according to claim 1, characterized in that: The graph convolution model includes multiple graph convolutional network layers and bidirectional long short-term memory network layers.

7. A method for predicting stratum deformation according to claim 6, characterized in that: The feature matrix, degree matrix, and target adjacency matrix corresponding to the current historical moment are respectively input into the corresponding graph convolution model to mine the spatial features of the monitoring data and output the predicted values ​​of the preset monitoring points at the current historical moment, including the following steps: The feature matrix, degree matrix, and target adjacency matrix corresponding to the current historical moment are combined to obtain the graph structure corresponding to the current historical moment; The graph structure corresponding to the current historical moment passes through the first graph convolutional network layer to extract the spatial features of the deformed data and obtain the output features of the first graph convolutional network layer; The output of the i-th graph convolutional network layer is linearly transformed through the ReLu activation function, and then input into the i-th bidirectional long short-term memory network layer to update the weight parameters, and obtain the output features of the i-th bidirectional long short-term memory network layer; The output features of the i-th bidirectional long short-term memory network layer are used as the input of the i+1-th graph convolutional network layer, where i=1, 2…N-1, and N represents the total number of graph convolutional network layers; The output features of the Nth graph convolutional network layer pass through the ReLu activation function to output the predicted value of the preset monitoring point at the current historical moment.

8. A method for predicting stratum deformation according to claim 1, characterized in that: The method of obtaining a characteristic matrix corresponding to each historical moment based on the stratum deformation parameter data at each historical moment includes: The stratum deformation parameter data at each historical moment is converted into a tensor structure through the tensor function to obtain the feature matrix corresponding to each historical moment.

9. A method for predicting stratum deformation according to claim 1, characterized in that: The stratum deformation parameter data include: geological parameters, shield parameters, structural parameters and tunnel parameters.

10. A system for predicting formation deformation, characterized in that: include: A data acquisition module is used to acquire a monitoring data set of the stratum to be monitored, wherein the monitoring data set includes: a monitoring point layout diagram, a deformation value of each monitoring point at different historical moments, and stratum deformation parameter data; The prediction module is used to input the monitoring data set of the stratum to be monitored into the stratum deformation prediction model, predict the stratum deformation value of the stratum to be monitored, and output the deformation prediction value of the preset monitoring point at the preset time, including: Based on the monitoring point layout diagram, the initial topological structure of the stratum to be monitored and its initial adjacency matrix are constructed; Based on the deformation value of each monitoring point at each historical moment, the initial topological structure and the initial adjacency matrix are optimized to obtain the optimized topological structure and adjacency matrix at each historical moment. Based on the optimized topological structure at each historical moment, the degree matrix corresponding to each historical moment is obtained. Input the optimized adjacency matrix of each historical moment into the first attention mechanism module respectively, assign different weights to each edge in the optimized topological structure of each historical moment, and obtain the target adjacency matrix corresponding to each historical moment; Based on the stratum deformation parameter data at each historical moment, the characteristic matrix corresponding to each historical moment is obtained; The feature matrix, degree matrix, and target adjacency matrix corresponding to each historical moment are input into the corresponding graph convolution model respectively, the spatial features of the monitoring data are mined, and the predicted values ​​of the preset monitoring points at each historical moment are output; After the predicted values ​​of the preset monitoring points at all historical moments pass through the fully connected layer, they are input into the temporal convolutional network layer to capture the patterns and dependencies in the sequence and output the initial deformation prediction value of the preset monitoring points at the preset time. The initial deformation prediction value of the preset monitoring point at the preset time is passed through the second attention mechanism module and the fully connected layer in sequence to output the deformation prediction value of the preset monitoring point at the preset time.

Images