A method for predicting shield tunneling parameters in composite strata

Abstract

The present invention discloses a method for predicting shield tunneling parameters in composite strata, comprising the following steps: collecting and collating historical shield tunneling data in composite strata, including geological parameters and construction parameters, removing outliers, filling missing values, and standardizing the data; combining geological parameters with geometric parameters according to tunneling ring numbers to generate a quantized geological matrix for each ring; using 1-5 coding for the geometric parameters to represent the position of the soil layer relative to the tunnel excavation face and ensuring input dimension uniformity through zero-padding alignment; constructing and training a CNN-Attention-SVR model; inputting new geological data into the trained model to predict tunneling parameters and guide construction adjustments. The present invention provides reliable technical support and decision-making basis for tunnel construction in composite strata, thereby improving the overall safety and economy of shield construction.

Description

Technical Field

[0001] The present invention relates to the technical field of shield construction, and in particular to a method for predicting shield tunneling parameters under composite strata. Background Art

[0002] With the acceleration of urbanization, shield tunneling has gained widespread application in tunnel construction due to its advantages such as high mechanization, safety, and efficiency. However, the complex nonlinear relationship between tunneling parameters and the geological environment makes the selection of appropriate tunneling parameters particularly important under complex geological conditions. Although the selection of tunneling parameters currently relies primarily on manual experience, the rapid development of artificial intelligence technology has led to the development of intelligent decision-making and unmanned shield tunneling.

[0003] Machine learning algorithms, such as random forests, K-nearest neighbor algorithms, and neural networks, have been widely used in tunneling parameter prediction due to their powerful nonlinear modeling capabilities. Existing research focuses on two main areas: one is to construct intelligent prediction models based on the temporal and correlation characteristics of tunneling parameters, without considering geological parameters; the other is to construct intelligent prediction models based on geological characteristic parameters. However, due to the complexity of geological conditions, especially the large variations in geological characteristic parameters within complex strata, both methods have limitations in quantifying complex geological conditions and extracting features. Summary of the Invention

[0004] Purpose of the Invention: This invention provides a method for predicting shield tunneling parameters in complex strata. This method segments different soil layers into a matrix, statistically analyzing the geological parameters of each stratum. Geometric parameters are then introduced through unit coding to create a quantitative matrix of the geological conditions. A convolutional neural network combined with an attention mechanism is used to automatically extract features from the quantitative matrix. Finally, a regression support vector machine is used to intelligently predict tunneling parameters.

[0005] Technical solution: The present invention provides a method for predicting shield tunneling parameters in composite strata, comprising the following steps:

[0006] (1) Data preprocessing: Collect and organize historical data of shield tunneling under composite strata, including geological parameters and construction parameters, and perform outlier removal, missing value filling, and data standardization;

[0007] (2) Constructing a quantitative geological matrix: The geological parameters are combined with the geometric parameters according to the tunneling ring number to generate a quantitative geological matrix for each ring; the geometric parameters are coded from 1 to 5 to represent the position of the soil layer relative to the tunnel excavation surface and zero-padding is used to ensure the uniformity of the input dimensions;

[0008] (3) Constructing a CNN-Attention-SVR model: The input layer receives the quantized geological matrix; extracting local geological features through a convolutional neural network (CNN), including two convolutional layers and a pooling layer to introduce an attention mechanism to dynamically calculate feature weights and highlight key areas; using support vector regression (SVR) combined with a radial basis function (RBF) kernel, nonlinear regression is performed on the weighted features to output the predicted values of the excavation parameters;

[0009] (4) Model training: The Adam optimizer is used to train the model, and cross-validation is used to prevent overfitting. The loss function is the mean square error (MSE).

[0010] (5) Parameter prediction: New geological data are input into the trained model to predict excavation parameters and guide construction adjustments.

[0011] Furthermore, in step (1), the geological parameters include: gravity, internal friction angle, and cohesion; and the construction parameters include: penetration, excavation speed, total propulsion force, cutterhead speed, cutterhead torque, total grouting volume, HBW grease dosage, and EP2 grease dosage.

[0012] Furthermore, in step (1), the 3σ standard deviation method is used to eliminate the non-stationary tunneling section data, and the moving window smoothing polynomial is used to reduce the noise of the tunneling parameters.

[0013] Furthermore, in step (2), the dimension of the quantized geological matrix is 18 × 5, with each row corresponding to a soil layer unit, and contains three geological parameters and two geometric parameters. Zero padding is used to align the input matrices to ensure that the dimensions of all rings are consistent; the geometric parameters are soil layer thickness and position code.

[0014] Furthermore, in step (2), the codes 1-5 are as follows: 1 is the upper part of the horizontal plane above the excavation surface, 2 is the lower part of the horizontal plane above the excavation surface, 3 is the upper part of the excavation surface, 4 is the lower part of the excavation surface, and 5 is the lower part of the excavation surface.

[0015] Furthermore, in step (3), CNN adopts the LeNet architecture, which includes two convolutional layers with a convolution kernel size of 3×3 and two maximum pooling layers with a pooling window of 2×2; the attention mechanism is implemented through self-attention calculation, and the formula is:

[0016] ;

[0017] in, is the dimension of the key; Q, K, V are the linear transformations of the input feature matrix.

[0018] Furthermore, in step (3), the loss function of SVR is:

[0019] ;

[0020] in, is the weight vector, is the regularization parameter, is a slack variable.

[0021] Furthermore, in step (4), the ratio of the training set, validation set, and test set is 8:1:1, and the training termination condition is that the loss value does not decrease significantly after 30 consecutive iterations.

[0022] Furthermore, in step (5), the prediction results are restored to the actual parameter values through denormalization. The denormalization formula is:

[0023] ;

[0024] in, is the original data; is the normalized data; 、 are the maximum and minimum values of the corresponding parameters respectively.

[0025] The present invention provides a shield tunneling parameter prediction system under composite strata, comprising:

[0026] Data preprocessing module: used to collect and organize historical data of shield tunneling under complex strata, including geological parameters and construction parameters, to eliminate outliers, fill in missing values and standardize data;

[0027] Build a quantitative geological matrix module: This module combines geological parameters with geometric parameters by excavation ring number to generate a quantitative geological matrix for each ring. Geometric parameters are coded 1-5 to represent the position of the soil layer relative to the tunnel excavation face, and zero padding is used to align the input dimensions to ensure uniformity.

[0028] CNN-Attention-SVR module: This module receives the quantized geological matrix at the input layer. It extracts local geological features through a convolutional neural network (CNN), which includes two convolutional layers and a pooling layer. It introduces an attention mechanism to dynamically calculate feature weights and highlight key areas. It uses support vector regression (SVR) combined with a radial basis function (RBF) kernel to perform nonlinear regression on the weighted features and output predicted values for tunneling parameters.

[0029] Model training module: used to train the model using the Adam optimizer, preventing overfitting through cross-validation, and the loss function is the mean square error (MSE).

[0030] Parameter prediction module: used to input new geological data into the trained model, predict excavation parameters and guide construction adjustments.

[0031] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention comprehensively considers the complexity and uncertainty of shield construction under complex stratum conditions, and constructs a shield tunneling parameter prediction model based on the CNN-Attention-SVR architecture; the model is based on geological survey data and historical construction data, and through the establishment of a quantitative geological matrix, it fully considers various influencing factors including geological type, stratum hardness, water content, etc., combined with key parameters such as propulsion force, torque, propulsion speed recorded during shield construction, and the model can accurately predict the construction parameter requirements of shield tunneling; the present invention has formed a complete set of shield tunneling parameter prediction and guidance schemes through a large amount of model training and optimization, which can not only improve the accuracy and efficiency of construction, but also significantly reduce construction risks, and provide reliable technical support and decision-making basis for tunnel construction under complex stratum conditions, thereby improving the overall safety and economy of shield construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the present invention;

[0033] Figure 2 It is a schematic diagram of the quantitative geological feature matrix of the present invention;

[0034] Figure 3 Schematic diagram of the CNN-Attention-SVR model of the present invention. DETAILED DESCRIPTION

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] like Figure 1-Figure 3 As shown, an embodiment of the present invention provides a method for predicting shield tunneling parameters in composite strata, comprising the following steps:

[0037] Step S1, collecting and arranging historical data of shield tunneling under composite strata and performing data preprocessing, including:

[0038] Step S11, collecting geological data related to the composite stratum and construction data during the shield tunneling process from completed shield projects, the geological data including degree, internal friction angle, and cohesion, and the construction data including penetration, tunneling speed, total propulsion force, cutterhead speed, cutterhead torque, total grouting volume, HBW grease usage, and EP2 grease usage;

[0039] In step S12, the collected data is cleaned and processed to remove outliers and fill in missing data to ensure data integrity and accuracy. First, the data within the ring is integrated into the inter-ring data. Records that are not in a state of advancement or are not stable are excluded. This includes data from when the shield machine stopped tunneling due to maintenance, segment assembly, or failure, as well as data from the shield machine's start-up to stable tunneling. The 3σ standard deviation method is used to distinguish data from the initial and stable tunneling phases, retaining data from the stable phase. Data denoising is performed on tunneling parameters using a moving window smoothing polynomial method to reduce the impact of noise on data analysis.

[0040] Step S2, such as Figure 2 As shown, the model input parameters are constructed and a quantitative geological matrix is established by ring. The matrix includes geological parameters and geometric parameters, including:

[0041] Step S21, based on the processed geological data and shield tunneling parameters, group the data by tunneling ring number to generate a quantitative geological matrix for each ring;

[0042] In step S22, geometric parameters are added to the quantitative geological matrix, including soil layer thickness and relative position of the tunnel face. These parameters can significantly affect the shield tunneling status and parameters. Next, parameter setting and initialization are performed. The data processing work in this process needs to ensure the accuracy and consistency of the data.

[0043] In step S23, the dimensionality of the quantitative geological matrix is adjusted to ensure that all input data is of the same dimension, and the matrix is further standardized. The geometric parameters in the quantitative geological matrix are represented using a coding method of 1 to 5. 1 and 2 represent the horizontal plane above and below the tunnel excavation surface, respectively; 3 and 4 both represent the position within the tunnel excavation surface, with 3 representing the upper half of the excavation surface and 4 representing the lower half; and number 5 represents the soil layer below the tunnel excavation surface. The geological parameters and geometric parameters are arranged in sequence to form the quantized row vectors of the corresponding soil layer. These are then combined from top to bottom according to the depth of the soil layer to form the geological feature quantization matrix corresponding to the ring number. To ensure that the geological quantization matrix of each ring has the same dimension for input into the prediction model, all matrices are padded with zeros and aligned.

[0044] Step S3, such as Figure 3 As shown in the figure, a multi-layer convolutional neural network model is constructed, which takes the quantified geological matrix as input and outputs the predicted values of shield tunneling parameters. It includes:

[0045] Step S31: Design a convolutional neural network model. The input layer receives the quantized geological matrix, and the initial layer includes multiple convolution kernels to extract local features of geological properties and geometric parameters. The convolution layer is the core component of the CNN, and its main function is to extract data features from multiple angles. The convolution layer essentially performs discrete convolution operations to calculate the convolution result of each receptive field matrix.

[0046] ;

[0047] Where: W is the convolution kernel matrix; b is the bias term; is the receptive field matrix.

[0048] In step S32, a pooling layer is added after the convolutional layer. This downsampling operation reduces the size of the feature matrix, enhancing the model's generalization ability and preventing overfitting. The main function of the pooling layer is to compress the data, thereby reducing the number of parameters and reducing the risk of model overfitting. Common pooling operations include max pooling and average pooling. By reducing the spatial size of the data, the pooling layer reduces computational complexity and memory consumption, thereby helping to build more efficient and stable deep learning models. The calculation formula for the pooling layer is:

[0049] ;

[0050] Where: z is the result after pooling; is the pooling window matrix of the previous layer output matrix; l is the window shape.

[0051] The present invention adopts the classic LeNet architecture, which mainly includes two convolutional layers, two pooling layers and two fully connected layers. This structure effectively controls the complexity of the model while ensuring the model training effect, thereby reducing the risk of overfitting.

[0052] Step S33: Set up an attention mechanism and train the attention weights through the model to highlight the feature areas that are more important for the prediction of tunneling parameters. Then, these weighted feature maps are used for the regression task of tunneling parameters. Specifically, the self-attention mechanism is used to dynamically calculate the weights of the features. Given an input feature matrix, a query matrix is generated through linear transformation. , key matrix Sum Matrix , the formula is as follows:

[0053] ;

[0054] Where: is a learnable weight matrix;

[0055] Then, the softmax activation function is used to calculate the similarity between the query and the key to obtain the attention weight. The specific calculation formula is as follows:

[0056] ;

[0057] in, is the dimension of the key; Q, K, V are the linear transformations of the input feature matrix.

[0058] SVR is used for regression tasks, and the features extracted through the convolutional layer and attention mechanism are input into the SVR regression model. SVR will use these feature maps to predict tunneling parameters, including eight key tunneling parameters: penetration, total thrust, cutterhead torque, tunneling speed, cutterhead speed, total grouting volume, HBW grease usage, and EP2 grease usage. SVR uses the radial basis function (RBF) kernel to handle nonlinear relationships, mapping the input feature space to a high-dimensional feature space to make the data linearly separable. This local sensitivity sample has a greater impact on decision-making, thereby better adapting to the local structure of the data and having strong generalization capabilities. The specific loss function is as follows:

[0059] ;

[0060] in, is the weight vector, is the regularization parameter, is a slack variable.

[0061] Step S4, training the convolutional neural network model using historical data to optimize model parameters, including:

[0062] In step S41, the historical dataset is divided into a training set and a validation set. The model is initially trained using the training set, and the network weights are adjusted by minimizing the prediction error. The data volume ratios for the training set, validation set, and test set are 0.8, 0.1, and 0.1, respectively. The optimizer uses the backpropagation algorithm to minimize the loss function (MSE) to adjust the model parameters. The formula is as follows:

[0063] ;

[0064] Where: is the true value; is the predicted value; n is the number of samples.

[0065] Step S42: Use cross-validation during the training process to evaluate the performance of the model on different data sets to avoid overfitting of the model on the training set.

[0066] In step S43, the optimization algorithm (Adam) is used to further adjust the learning rate and parameters of the model to ensure that the model quickly converges to the global optimum during training. The initial learning rate is set to 0.001, and the model is trained for a maximum of 100 iterations. At the same time, to prevent overfitting of the model, training will be terminated early if the loss value does not improve significantly within 30 iterations. Before training, the input and output data are normalized to the range of 0 to 1, and the maximum and minimum values of the output parameters are recorded so that they can be denormalized after the prediction results. The formula is as follows:

[0067] ;

[0068] in, is the original data; is the normalized data; 、 are the maximum and minimum values of the corresponding parameters respectively.

[0069] Step S5, using the trained model to perform prediction and output parameter values for shield tunneling under composite strata, said step S5 includes:

[0070] Step S51: Input the new geological data into the trained CNN-Attention-SVR model to predict the key parameter values during the excavation process;

[0071] Step S52: Analyze and verify the prediction results, and evaluate the prediction accuracy of the model by comparing them with the parameters in actual construction;

[0072] Step S53: Adjust shield tunneling construction parameters according to the prediction results, optimize the construction plan, and improve tunneling efficiency and safety.

[0073] The method of this invention aims to comprehensively consider geological conditions and construction efficiency, optimizing the prediction of shield tunneling parameters through artificial intelligence deep learning methods to ensure stable and efficient tunneling in complex strata. This method can improve construction accuracy, reduce construction risks, and provide technical support for tunnel construction in complex geological conditions.

Claims

1. A method for predicting shield tunneling parameters in composite strata, characterized in that: The following steps are involved: (1) Data preprocessing: Collect and organize historical data of shield tunneling under composite strata, including geological parameters and construction parameters, and perform outlier removal, missing value filling, and data standardization; (2) Constructing a quantitative geological matrix: The geological parameters are combined with the geometric parameters according to the tunneling ring number to generate a quantitative geological matrix for each ring; the geometric parameters are coded from 1 to 5 to represent the position of the soil layer relative to the tunnel excavation surface and zero-padding is used to ensure the uniformity of the input dimensions; (3) Constructing a CNN-Attention-SVR model: the input layer receives the quantized geological matrix; extracting local geological features through a convolutional neural network (CNN), including two convolutional layers and a pooling layer to introduce an attention mechanism to dynamically calculate feature weights and highlight key areas; using support vector regression (SVR) combined with a radial basis function (RBF) kernel, performing nonlinear regression on the weighted features and outputting the predicted values of the excavation parameters; CNN adopts a LeNet architecture, including two convolutional layers with a convolution kernel size of 3×3 and two maximum pooling layers with a pooling window of 2×2; the attention mechanism is implemented through self-attention calculation, and the formula is: ； in, is the dimension of the key; Q, K, V are the linear transformations of the input feature matrix; The loss function of SVR is: ； in, is the weight vector, is the regularization parameter, is a slack variable; (4) Model training: The Adam optimizer is used to train the model, and cross-validation is used to prevent overfitting. The loss function is the mean square error (MSE). (5) Parameter prediction: New geological data are input into the trained model to predict excavation parameters and guide construction adjustments.

2. The method for predicting shield tunneling parameters under composite strata according to claim 1, characterized in that: In step (1), the geological parameters include: gravity, internal friction angle, and cohesion; the construction parameters include: penetration, excavation speed, total propulsion force, cutterhead speed, cutterhead torque, total grouting volume, HBW grease dosage, and EP2 grease dosage.

3. The method for predicting shield tunneling parameters under composite strata according to claim 1, characterized in that: In step (1), the 3σ standard deviation method is used to eliminate the non-stationary tunneling section data, and the moving window smoothing polynomial is used to reduce the noise of the tunneling parameters.

4. The method for predicting shield tunneling parameters under composite strata according to claim 1, characterized in that: In step (2), the dimension of the quantized geological matrix is 18 × 5, and each row corresponds to a soil layer unit. It contains three geological parameters and two geometric parameters. Zero padding is used to align the input matrices to ensure that the dimensions of all rings are consistent. The geometric parameters are soil layer thickness and position code.

5. The method for predicting shield tunneling parameters in composite strata according to claim 1, characterized in that: In step (2), the codes 1-5 are as follows: 1 is the upper part of the horizontal plane above the excavation surface, 2 is the lower part of the horizontal plane above the excavation surface, 3 is the upper part of the excavation surface, 4 is the lower part of the excavation surface, and 5 is the lower part of the excavation surface.

6. The method for predicting shield tunneling parameters in composite strata according to claim 1, characterized in that: In step (4), the ratio of training set, validation set and test set is 8:1:1, and the training termination condition is that the loss value does not decrease significantly after 30 consecutive iterations.

7. The method for predicting shield tunneling parameters in composite strata according to claim 1, characterized in that: In step (5), the prediction results are restored to the actual parameter values through denormalization. The denormalization formula is: ； in, is the original data; is the normalized data; 、 are the maximum and minimum values of the corresponding parameters respectively.

8. A shield tunneling parameter prediction system under composite strata, characterized in that: include: Data preprocessing module: used to collect and organize historical data of shield tunneling under complex strata, including geological parameters and construction parameters, to eliminate outliers, fill in missing values and standardize data; Build a quantitative geological matrix module: This module combines geological parameters with geometric parameters by excavation ring number to generate a quantitative geological matrix for each ring. Geometric parameters are coded 1-5 to represent the position of the soil layer relative to the tunnel excavation face, and zero padding is used to align the input dimensions to ensure uniformity. CNN-Attention-SVR module: This module receives the quantized geological matrix at the input layer. It extracts local geological features through a convolutional neural network (CNN), which includes two convolutional layers and a pooling layer. It introduces an attention mechanism to dynamically calculate feature weights and highlight key areas. It uses support vector regression (SVR) combined with a radial basis function (RBF) kernel to perform nonlinear regression on the weighted features and output predicted values for tunneling parameters. Model training module: used to train the model using the Adam optimizer, preventing overfitting through cross-validation, and the loss function is the mean square error (MSE). Parameter prediction module: used to input new geological data into the trained model, predict excavation parameters and guide construction adjustments.

Images