An Adaptive Adjustment Method for EPB Construction Parameters for Settlement Pre-control

Abstract

This invention relates to the field of intelligent construction technology for underground engineering, and provides an adaptive adjustment method for EPB construction parameters for settlement pre-control. The method includes: developing a bidirectional long short-term memory neural network prediction model (MHSA-Bi-LSTM model) with enhanced multi-head self-attention mechanism to capture the temporal and spatial correlations of shield-induced settlement; constructing a multi-classification control strategy for shield parameters for settlement pre-control; constructing an optimization mechanism for shield operation parameters based on optimization and composite inversion; and achieving parameter recommendation and real-time control through the developed shield parameter optimization system. This invention solves the problems of lag in parameter adjustment and insufficient settlement control accuracy in traditional shield construction, as well as the "one-size-fits-all" problem in traditional optimization methods, through parameter interaction decoupling and intelligent control, providing a scientific basis for the automation of shield construction.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology for underground engineering, and in particular to an adaptive adjustment method for EPB construction parameters for settlement pre-control. Background Technology

[0002] With the acceleration of urbanization, underground space development has become an important means of alleviating urban traffic congestion, among which the shield tunneling method is widely used due to its high efficiency and safety. However, surface settlement caused by soil disturbance during shield tunneling remains a key challenge in the engineering field. Traditional settlement control methods mainly rely on manual experience to adjust parameters or grouting, which have significant time lags and can easily lead to irreversible deformation, such as building cracking. Settlement control during shield tunneling has become an urgent problem to be solved in current underground transportation construction.

[0003] With the increasing integration of technologies such as digital twins and big data analytics with various industries, innovative thinking in industrial integration has yielded significant results in intelligent construction of civil engineering projects. Machine learning methods, driven by data, find mathematical relationships between various factors and prediction targets, providing prediction results in a shorter time, making them crucial in predicting settlement caused by shield tunnel construction. Traditional numerical simulations or single algorithms struggle to capture the spatiotemporal settlement patterns under complex geological conditions; Long Short-Term Memory (LSTM) possesses stronger memory retention and selection capabilities, and can consider the sequential effects of shield tunnel deformation and settlement. Attention mechanisms, introduced in recent years, excel at learning the spatiotemporal correlations between data by assigning attention weights to each step, demonstrating superior ability to capture long-term dependencies between steps, key data information, and internal data autocorrelation. Although the Multi-Head Self-Attention (MHSA) mechanism can extract richer data features from different subspaces from multiple perspectives compared to the self-attention mechanism, research on combining it with existing models to predict ground settlement caused by shield tunnel construction is very limited.

[0004] Existing research has shown a close relationship between geological parameters, tunneling parameters, and settlement deformation during tunnel boring machine (TBM) construction. However, when surface settlement control is used as a separate evaluation criterion, current methods primarily rely on optimization algorithms to find the optimal TBM parameters. The strong coupling between these parameters makes precise control difficult through such a single adjustment. Furthermore, parameter inversion by explicitly dividing samples into superior and inferior subsets inevitably has limitations in application. How to optimize TBM parameters, guide the TBM to operate in its optimal configuration, and achieve pre-control of settlement remains a challenging engineering problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to propose an adaptive adjustment method for EPB construction parameters for settlement pre-control, which avoids parameter conflicts caused by single optimization in traditional methods and improves the accuracy of prediction and recommendation.

[0006] An adaptive adjustment method for EPB construction parameters for settlement pre-control, the method comprising:

[0007] Step 1: Develop a bidirectional long short-term memory neural network prediction model with enhanced multi-head self-attention mechanism to capture the temporal and spatial correlation of shield tunneling-induced settlement, namely the MHSA-Bi-LSTM model.

[0008] Step 2: Construct a multi-classification control strategy for shield tunneling parameters for settlement pre-control; classify shield tunneling parameters into Class A (high sensitivity), Class B (low sensitivity and uncontrollable), and Class C (low sensitivity and controllable) parameters based on shield-induced settlement sensitivity and human controllability.

[0009] Step 3: Construct a shield tunneling parameter optimization mechanism based on optimization and composite inversion; use the particle swarm optimization (PSO) algorithm to optimize type A parameters to minimize settlement, and combine it with the random forest (RF) algorithm to invert type B and type C parameters in sequence;

[0010] Step 4: Implement parameter recommendation and real-time control through the developed shield tunneling parameter optimization system (SPOS).

[0011] Preferably, the development of the prediction model using a multi-head self-attention mechanism-enhanced bidirectional long short-term memory neural network (MHSA-Bi-LSTM) that simultaneously captures the temporal and spatial correlations of shield-induced settlement is as follows:

[0012] Step 1.1: Define the variable form, with a sampling unit of 7.5 meters (5 rings). The input parameters include geometric parameters, geological parameters, shield parameters and historical settlement. The output parameter is the shield-induced settlement.

[0013] The depth of soil cover was determined to be the only geometric parameter to be considered, and the measurement result was denoted as... Dc ;

[0014] Geological parameters are weighted according to the thickness of test samples at each survey location, and the resulting compressive modulus is calculated by interpolation and correlated with the shield tunneling location. This calculated modulus is denoted as... Mc Cohesion is denoted as Co The internal friction angle is denoted as A The lateral pressure coefficient is denoted as K Groundwater level is recorded as Wt ;

[0015] Shield tunneling parameters should prioritize those that have direct or indirect interaction with the soil; the synchronous grouting volume should be selected accordingly. VgTotal thrust Th Cutter head torque Tc Cutter head speed Rc Support and pressure Sp Speed ​​of advancement Sa Screw conveyor speed Rsc Screw conveyor torque Tsc and the volume of excavated earth Ve As parameters for tunnel boring machines;

[0016] Settlement data includes historical settlement and shield-induced settlement. It needs to be synchronized with shield parameters. Measurements were taken twice daily along the tunnel's longitudinal axis at the same 7.5-meter (5-ring) intervals, with the monitoring frequency increased to 3 times per hour within 15 meters before and after the shield excavation location. Surface settlement at 7.5 meters and 0 meters before the shield's arrival were respectively considered historical settlement, denoted as... Sh 7.5 , Sh 0; The surface settlement 10 days after the completion of the annular tunneling is considered as shield-induced settlement and is denoted as... Ss ;

[0017] Step 1.2: Construct the MHSA-Bi-LSTM model architecture, which adopts a configuration of Bi-LSTM layer + self-attention layer + Dropout layer + Bi-LSTM layer, as shown below:

[0018] The first layer is the Sequence Input Layer, which serves as the network's entry point and is responsible for receiving input data.

[0019] The second layer is the flatten layer;

[0020] The third layer is the Bi-LSTM-1 layer, which effectively processes long sequence data, solves the gradient vanishing or gradient explosion problem of traditional RNNs, and learns long-term dependencies in the sequence.

[0021] The fourth layer is the Self Attention Layer, which projects its output into multiple subspaces (heads) to compute context weights;

[0022] The fifth layer is the Dropout Layer, which enhances regularization, reduces data noise, and prevents overfitting that may reduce the generalization ability of the model.

[0023] The sixth layer is another Bi-LSTM layer, Bi-LSTM-2 Layer, which has the same function as the previous Bi-LSTM layers, further learning complex patterns and long-term dependencies in the sequence;

[0024] The seventh layer is a fully connected layer, which performs a linear transformation on the output of the previous Bi-LSTM layers, mapping the features to a specified dimension;

[0025] The eighth layer is the Regression Output Layer, used for regression tasks. It calculates the loss between the predicted and the true values ​​(usually mean squared error) and performs backpropagation during training to update the network parameters.

[0026] The optimizer is configured as an adaptive moment estimator (Adam). To mitigate instability during training, a segmented scheduling strategy is used to modulate the learning rate, starting from an initial learning rate of 0.001, with a descent period of 400 and a descent factor of 0.15.

[0027] Step 1.3: Train the model.

[0028] The training data is processed by min-max normalization to convert it into the interval [0,1].

[0029] The model hyperparameters were determined using a grid search method, including stride [1:1:5], number of hidden units in the Bi-LSTM layer [10:10:500], dropout rate in the Dropout layer [0.1:0.05:0.6], number of heads in the self-attention layer [2:1:10], and number of key channels [1-10 times the number of heads]. Five-fold time series cross-validation (5-TCV) was used to determine the optimal architecture of the proposed model.

[0030] The implicit form of the final MHSA-Bi-LSTM settlement prediction model is:

[0031] Formula 1: ;

[0032] In the formula, X For the input vector, X =[ Dc , Mc , Co , A , K , Wt , Sa , Rc , Tc , Sp , Vg , Th , Tsc , Rsc , Ve , Sh 0, Sh 7.5 ]; To conceal the state and enhance the ability to remember the sedimentation sequence;

[0033] Preferably, the content of constructing the multi-classification control strategy for shield tunneling parameters for settlement pre-control is as follows:

[0034] Step 2.1: Perform sensitivity analysis using the Sopol method:

[0035] From Formula 1, we can see that the assumed proxy model takes the following form:

[0036] Formula 2: ;

[0037] In the formula, For the output of the surrogate model, the unconditional variance is... It can be decomposed into partial variance accordingly:

[0038] Formula 3: ;

[0039] In the formula, and These represent the partial variance related to a single input variable and the partial variance related to the interaction between two input variables, respectively. and The variance; For all The partial variance resulting from the combined effect of the input variables;

[0040] Using these biases, the first-order, second-order, and total Sopol sensitivity indices are defined according to the total variance, as shown in Formulas 4-6:

[0041] Formula 4: ;

[0042] Formula 5: ;

[0043] Formula 6: ;

[0044] In the formula, MSI is the principal sensitivity index, or the first-order index. S i , indicating a single input X i The degree of interference to the output when the value fluctuates;

[0045] S ij It is a second-order exponent, or... X i and X j The second-order interaction effect index represents the degree to which the interaction between two inputs interferes with the output;

[0046] V ~i ExceptX i The variance of the other indicators;

[0047] TSI stands for Total Sensitivity Index, or... ST i , characterizing when input index X i When the value fluctuates, the total degree of interference to the output caused by all its effects is the sum of the sensitivity indices of each order of the input index;

[0048] Based on the range of parameter variables, Latin hypercube sampling is performed, and a proxy model is applied to calculate the correlation to determine the MSI and TSI of each parameter under different sample sizes. When the change in the results is less than 1% as the sample size increases, the results are considered to have converged, and these are the final MSI and TSI.

[0049] Step 2.2: Perform multiple classifications of shield tunneling parameters for settlement pre-control:

[0050] A classification control strategy is proposed, which divides the shield tunneling parameters into three categories: A, B, and C.

[0051] Step 2.2.1: Classification Requirements: The cumulative TSI of Class A parameters on shield-induced settlement is ≥90% to demonstrate their dominant role in settlement. Parameters with higher TSI rankings are selected first. The remaining low-sensitivity parameters are classified into Class B and Class C parameters according to their controllability during operation.

[0052] Step 2.2.2: The classified A-class parameters are verified through the parameter interaction matrix. The Pearson correlation coefficient is required to be ≤0.2 to ensure that the strong relationship between the parameters that have been forced to change has not yet become prominent, and to avoid destroying the strong correlation between parameters.

[0053] Step 2.2.3: If the verification requirements in step 2.2.2 are not met, remove the strongly related Class A parameters and return to step 2.2.1 to re-screen.

[0054] Preferably, the steps for constructing the shield tunneling parameter optimization mechanism based on optimization-composite inversion include:

[0055] Step 3.1, Optimize the search for Class A parameters:

[0056] Step 3.1.1, PSO particle initialization:

[0057] Set the size of the particle swarm N Each particle is randomly initialized with a position and velocity; the position of each particle represents a set of A-class parameters (A1, A2, ..., A...). n The values ​​of ) are combinations, while velocity represents the speed at which the particle moves in the parameter space, where the particle i Location ,speed ;

[0058] Based on the actual project conditions, the value range for each Class A parameter should be clearly defined. ;

[0059] Step 3.1.2, Fitness Calculation:

[0060] The shield-induced settlement value predicted by the MHSA-Bi-LSTM surrogate model is used as the fitness function, and the goal is to minimize this settlement value.

[0061] For each particle i , place its position The remaining parameters of the current ring location (geometric and geological parameters use the original conditions, while parameters of types B and C use the statistical mean) are input into the surrogate model to calculate the corresponding predicted values ​​of shield-induced settlement. This value is the fitness of the particle;

[0062] Step 3.1.3, Individual Optimal and Global Optimal Updates:

[0063] For each particle i Compare its current fitness Compared with the previously recorded individual optimal fitness pbest i ;like < pbest i Then update the individual's optimal position. pbest posi = and individual optimal fitness pbest i = ;

[0064] Find the minimum value among all the individual optimal fitnesss of the particles, and let the corresponding particle be denoted as . j ;like pbest j Less than the currently recorded global optimal fitness gbest Then update the global optimal position. gbest pos = pbest posj and global optimal fitness gbest = pbest j ;

[0065] Step 3.1.4, Velocity and Position Update:

[0066] Update the velocity and position of each particle according to formulas 7-8:

[0067] Speed ​​Update: Formula 7: ;

[0068] In the formula, This is the inertial weight, used to balance the global and local search capabilities of particles, and is generally set to 0.7-1.2. and It is the acceleration constant, typically taken as 1.5; r 1 and r 2 is a random number in the interval [0,1].

[0069] Position Update: Formula 8: ;

[0070] After updating the position, it is necessary to check whether the position of each particle exceeds the range of parameter values. If it exceeds the range, its position is adjusted to the boundary value.

[0071] Step 3.1.5, Termination condition determination:

[0072] Check if the termination condition is met. The termination condition is reaching the preset maximum number of iterations. T max If the global optimal fitness remains unchanged for 200 consecutive iterations, the algorithm terminates and outputs the global optimal position if the termination condition is met. gbest pos The optimized A-type parameter values ​​are used; if the termination condition is not met, return to step 3.1.2 to continue the iteration;

[0073] Step 3.1.6, PSO algorithm output:

[0074] After the algorithm finishes, it outputs the globally optimal position. gbest pos The value of the A-type parameter corresponding to this position is the result obtained by the PSO algorithm optimization, which can be used to guide the recommendation of shield parameters for this ring position; the optimal fitness is output, which corresponds to the minimum shield-induced settlement predicted by the surrogate model;

[0075] Step 3.1.7: Optimize Class A parameters for all predetermined loop positions;

[0076] Monitor whether the termination conditions of the predetermined ring position are met. If they are met, output the recommended shield parameters for all predetermined ring positions. If not, return to step 3.1.1 and continue the optimization operation for the next ring position.

[0077] Step 3.2: Establish the B-type parameter inversion algorithm flow based on the RF model:

[0078] The input data includes geometric parameters, geological parameters, initial type A parameters, and settlement data (historical settlement and shield-induced settlement), with type B parameters as the output.

[0079] Determine the optimal combination of hyperparameters for RF using a grid search method:

[0080] Number of decision trees tr ∈[50:5:200]; Minimum number of leaf node samples lf ∈[1:1:5];

[0081] A 5-fold cross-validation approach was adopted, with mean squared error as the objective function;

[0082] By inputting geometric parameters, geological parameters, optimized Class A parameters, and settlement data (historical settlement and predicted minimum shield-induced settlement), new Class B parameters are derived as recommendations for shield parameters.

[0083] Step 3.3: Establish the C-type parameter inversion algorithm flow, based on the RF model:

[0084] The input data includes geometric parameters, geological parameters, initial A and B type parameters, settlement data (historical settlement and shield-induced settlement), and C type parameters as output;

[0085] Determine the optimal combination of hyperparameters for RF using a grid search method:

[0086] Number of decision trees tr ∈[50:10:200]; Minimum number of leaf node samples lf ∈[1:1:5];

[0087] A 5-fold cross-validation approach was adopted, with mean squared error as the objective function;

[0088] By inputting geometric parameters, geological parameters, optimized Class A parameters, inverted Class B parameters, and settlement data (historical settlement and predicted minimum shield-induced settlement), new Class C parameters are inverted to serve as recommendations for shield parameters.

[0089] Among them, hierarchical inversion increases the dimension of input parameters, preserves more parameter interaction relationships, and improves the inversion prediction accuracy of actual controllable operating parameters.

[0090] Preferably, the steps for parameter recommendation and real-time control through the developed shield tunneling parameter optimization system (SPOS) include:

[0091] Step 4.1, Basic Data Collection; the specific process is as follows:

[0092] SPOS collects and receives multi-source data for the next construction phase, including geometric parameters, geological parameters, shield parameters, and settlement data as described in step 1.1.

[0093] The collected multi-source data is aligned and preprocessed to eliminate data errors and noise;

[0094] Step 4.2, Model Iterative Optimization:

[0095] The new sample data generated during the tunnel boring process are continuously added to the settlement prediction proxy model and the RF inversion model in steps 3.2 and 3.3. By continuously updating the model parameters and structure through incremental learning, the model can better adapt to the dynamic changes during the construction process.

[0096] Through continuous iterative optimization, the new sample data in step 4.1 is incorporated into the model, which is able to learn this new information, thereby improving the accuracy of predicting future results.

[0097] Step 4.3, Parameter Recommendation and Dynamic Update:

[0098] Based on the optimized model using new sample data, SPOS generates recommended shield tunneling parameters in real time for the current construction situation. The optimized shield tunneling parameters are transmitted to the shield machine operation interface in real time to achieve dynamic calibration of the parameters. As geological conditions change during construction, SPOS dynamically updates the recommended parameter scheme based on new sample data and model feedback.

[0099] The beneficial effects achieved by this invention are as follows:

[0100] In the hierarchical optimization strategy of this invention, Class A parameters are globally optimized through PSO, while Class B / C parameters maintain their interaction relationship through RF inversion, avoiding parameter conflicts caused by single optimization in traditional methods. By closely integrating model prediction, parameter adjustment, and actual settlement feedback, a closed-loop feedback mechanism is formed. This addresses the accuracy degradation problem of traditional models due to geological parameter variability, making it more suitable for the dynamic environment of long-term tunnel construction. Attached Figure Description

[0101] Figure 1 This is a schematic diagram of the settlement pre-control process for adaptive adjustment of EPB construction parameters in an embodiment of the present invention.

[0102] Figure 2 This is a network architecture diagram of the MHSA-Bi-LSTM shield tunnel settlement prediction model in an embodiment of the present invention.

[0103] Figure 3 This is a comparison chart of settlement prediction results for the test set in an embodiment of the present invention.

[0104] Figure 4 This is a heatmap correlation matrix of shield tunneling parameters in an embodiment of the present invention.

[0105] Figure 5 This is a flowchart of shield tunneling parameter inversion and prediction according to an embodiment of the present invention.

[0106] Figure 6 This is a diagram showing the settlement results of the 745-840 ring shield tunnel under system pre-control in an embodiment of the present invention. Detailed Implementation

[0107] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0108] Please see Figure 1 This invention provides an adaptive adjustment method for EPB construction parameters for settlement pre-control, and applies it to the Changchun Metro project as a case study to verify the proposed pre-control method. The analysis section is the upstream shield tunneling section from Feiyue Square Station to Eurasia Mall Station in Section 2 of Changchun Metro Line 6. The method includes:

[0109] Step 1: Develop a bidirectional long short-term memory neural network prediction model with enhanced multi-head self-attention mechanism to capture the temporal and spatial correlation of shield tunneling-induced settlement, namely the MHSA-Bi-LSTM model.

[0110] Step 1.1: Define the variable form, with a sampling unit of 7.5 meters (5 rings). The input parameters include geometric parameters, geological parameters, shield parameters and historical settlement. The output parameter is the shield-induced settlement.

[0111] Data was collected from ring 170 (ZK17+321.735 on 2022 / 1 / 3) to ring 665 (ZK16+579.235 on 2022 / 5 / 21) in the upward shield tunneling section from Feiyue Square Station to Eurasia Mall Station of Changchun Metro Line 6, with an interval of 5 rings, for a total of 100 sets of samples.

[0112] Geometric parameters typically include the overburden ratio (the ratio of overburden depth to tunnel diameter). Since tunnel dimensions are usually kept constant in a single engineering study, the overburden depth is determined to be the only geometric parameter to consider, and the measurement result is denoted as... Dc ;

[0113] Key mechanical parameters in the geological parameters include compressibility modulus, cohesion, internal friction angle, and lateral pressure coefficient. The hydrogeological parameter is the groundwater level. Initial results for these geological parameters were obtained through separate tests at each survey location. The fundamental cause of ground settlement during tunnel boring machine (TBM) excavation is the disturbance of the surrounding soil matrix by construction activities, leading to a redistribution of soil position and stress. Specifically, the overburden above the TBM tunnel is directly affected by the construction and is thus most significantly impacted. Therefore, key mechanical parameters are weighted according to the thickness of different test samples and used as geological strength parameters for each survey location, corresponding to the TBM location through interpolation. The calculated compressibility modulus is denoted as... Mc Cohesion is denoted as Co The internal friction angle is denoted as A The lateral pressure coefficient is denoted as K Groundwater level is recorded as Wt ;

[0114] Shield tunneling parameters should prioritize those that have direct or indirect interaction with the soil; the synchronous grouting volume should be selected accordingly. Vg Total thrust Th Cutter head torque Tc Cutter head speed Rc Support and pressure Sp Speed ​​of advancement Sa Screw conveyor speed Rsc Screw conveyor torque Tsc and the volume of excavated earth Ve As shield tunneling parameters, previous studies have shown that these parameters have a significant impact on the stability of the excavation face and surface settlement. The system monitors the shield tunneling parameters in real time and uses parameter multiplication to remove data from the downtime phase. Given the discontinuous nature of the construction process, the selection of specific data is refined to minimize the impact of noise and outliers. Sp It is determined by the average pressure at the excavation face. Sa , Rc These parameters are determined by averaging the values ​​at the corresponding ring positions;

[0115] Settlement data includes historical settlement and shield-induced settlement. It needs to be synchronized with shield parameters. Measurements were taken twice daily along the tunnel's longitudinal axis at the same 7.5-meter (5-ring) intervals, with the monitoring frequency increased to 3 times per hour within 15 meters before and after the shield excavation location. Surface settlement at 7.5 meters and 0 meters before the shield's arrival were respectively considered historical settlement, denoted as... Sh 7.5 , Sh 0; The surface settlement 10 days after the completion of the annular tunneling is considered as shield-induced settlement and is denoted as... Ss ;

[0116] The first 80% of the dataset (80 sets) was used for model development, and the remaining 20% ​​of the data constituted the test dataset;

[0117] Step 1.2: Construct the MHSA-Bi-LSTM model architecture. Please refer to [link / reference]. Figure 2 Preferably, to support the interpretability of the model, a configuration of Bi-LSTM layer + self-attention layer + Dropout layer + Bi-LSTM layer is adopted, as follows:

[0118] The first layer is the Sequence Input Layer, which serves as the network's entry point and is responsible for receiving input data.

[0119] The second layer is the flatten layer;

[0120] The third layer is the Bi-LSTM-1 layer, which effectively processes long sequence data, solves the gradient vanishing or gradient explosion problem of traditional RNNs, and learns long-term dependencies in the sequence.

[0121] The fourth layer is the Self Attention Layer, which projects its output into multiple subspaces (heads) to compute context weights;

[0122] The fifth layer is the Dropout Layer, which enhances regularization, reduces data noise, and prevents overfitting that may reduce the generalization ability of the model.

[0123] The sixth layer is another Bi-LSTM layer, Bi-LSTM-2 Layer, which has the same function as the previous LSTM layers, further learning complex patterns and long-term dependencies in the sequence;

[0124] The seventh layer is a fully connected layer, which performs a linear transformation on the output of the previous LSTM layers, mapping the features to a specified dimension;

[0125] The eighth layer is the Regression Output Layer, used for regression tasks. It calculates the loss between the predicted and the true values ​​(usually mean squared error) and performs backpropagation during training to update the network parameters.

[0126] The optimizer is configured as an adaptive moment estimator (Adam). To mitigate instability during training, a segmented scheduling strategy is used to modulate the learning rate, starting from an initial learning rate of 0.001, with a descent period of 400 and a descent factor of 0.15.

[0127] Step 1.3: Train the model.

[0128] The training data is processed by min-max normalization to convert it into the interval [0,1].

[0129] The model hyperparameters were determined using a grid search method, including stride [1:1:5], number of hidden units in the Bi-LSTM layer [10:10:500], dropout rate in the Dropout layer [0.1:0.05:0.6], number of heads in the self-attention layer [2:1:10], and number of key channels [1-10 times the number of heads]. Five-fold time series cross-validation (5-TCV) was used to determine the optimal architecture of the proposed model.

[0130] The implicit form of the final MHSA-Bi-LSTM settlement prediction model is:

[0131] Formula 1: ;

[0132] In the formula, X For the input vector, X =[ Dc , Mc , Co , A , K , Wt , Sa , Rc , Tc , Sp , Vg , Th , Tsc , Rsc , Ve , Sh 0, Sh 7.5 ]; To conceal the state and enhance the ability to remember the sedimentation sequence;

[0133] Compared with widely used BPNN, RF, GRU, and Bi-LSTM models, the accuracy of the training set is verified, such as... Figure 3 As shown.

[0134] Step 2: Construct a multi-classification control strategy for shield tunneling parameters for settlement pre-control; classify shield tunneling parameters into Class A (high sensitivity), Class B (low sensitivity and uncontrollable), and Class C (low sensitivity and controllable) parameters based on shield-induced settlement sensitivity and human controllability.

[0135] Step 2.1: Perform sensitivity analysis using the Sopol method:

[0136] From Formula 1, we can see that the assumed proxy model takes the following form:

[0137] Formula 2: ;

[0138] In the formula, For the output of the surrogate model, the unconditional variance is... It can be decomposed into partial variance accordingly:

[0139] Formula 3: ;

[0140] In the formula, and These represent the partial variance related to a single input variable and the partial variance related to the interaction between two input variables, respectively. and The variance; For all The partial variance resulting from the combined effect of the input variables;

[0141] Using these biases, the first-order, second-order, and total Sopol sensitivity indices are defined according to the total variance, as shown in Formulas 4-6:

[0142] Formula 4: ;

[0143] Formula 5: ;

[0144] Formula 6: ;

[0145] In the formula, MSI is the principal sensitivity index, or the first-order index. S i , indicating a single input X i The degree of interference to the output when the value fluctuates;

[0146] S ij It is a second-order exponent, or... X i and X j The second-order interaction effect index represents the degree to which the interaction between two inputs interferes with the output;

[0147] V ~i Except X i The variance of the other indicators;

[0148] TSI stands for Total Sensitivity Index, or... ST i , characterizing when input index X i When the value fluctuates, the total degree of interference to the output caused by all its effects is the sum of the sensitivity indices of each order of the input index;

[0149] Based on the range of parameter variables, Latin hypercube sampling is performed, and a proxy model is applied to calculate the correlation to determine the MSI and TSI of each parameter under different sample sizes. When the change in the results is less than 1% as the sample size increases, the results are considered to have converged, and these are the final MSI and TSI.

[0150] Step 2.2: Perform multiple classifications of shield tunneling parameters for settlement pre-control:

[0151] A classification control strategy is proposed, dividing shield tunneling parameters into three categories: A, B, and C. Category A parameters mainly affect shield-induced settlement and are preferentially optimized using search methods, ignoring the interaction between parameters. Category B parameters have a relatively small impact on shield-induced settlement, and are only considered to be affected by other parameters, but are not subject to human control. Category C parameters have a relatively small impact on shield-induced settlement, and are only considered to be affected by other parameters, while being subject to human control and intervention.

[0152] Step 2.2.1: Classification Requirements: The cumulative TSI of Class A parameters on shield-induced settlement must be ≥90% to demonstrate their dominant role in settlement. Parameters with higher TSI rankings are prioritized, and Class A parameters (A1: ...) are ultimately selected. Ve A2: Sp A3: Rc A4: Tc The remaining low-sensitivity parameters are classified into categories B and C based on their controllability during operation (B1: Th B2: Tsc B3: Vg C1: Rsc C2: Sa );

[0153] Step 2.2.2: The classified A-class parameters are verified through the parameter interaction matrix. Please refer to Figure 4. The Pearson correlation coefficient is required to be ≤0.2 to ensure that the strong relationship between the parameters that have been forced to change has not yet become prominent, and to avoid destroying the strong correlation between parameters.

[0154] Step 2.2.3: If the verification requirements in Step 2.2.2 are not met, remove the strongly related Class A parameters and return to Step 2.2.1 to re-screen. In this case, the partitioning scheme in Step 2.2.1 meets the requirements of Step 2.2.2 and does not need to be re-partitioned.

[0155] Step 3: Construct a shield tunneling parameter optimization mechanism based on optimization and composite inversion; use the particle swarm optimization (PSO) algorithm to optimize type A parameters to minimize shield-induced settlement, and combine it with the random forest (RF) algorithm to invert type B and type C parameters in sequence;

[0156] Step 3.1, Optimize the search for Class A parameters:

[0157] Step 3.1.1, PSO particle initialization:

[0158] Set the size of the particle swarm N Each particle is randomly initialized with a position and velocity; the position of each particle represents a set of A-class parameters (A1, A2, ..., A...). n The values ​​of ) are combinations, while velocity represents the speed at which the particle moves in the parameter space, where the particle i Location ,speed ;

[0159] Based on the actual project conditions, the value range for each Class A parameter should be clearly defined. ;

[0160] Step 3.1.2, Fitness Calculation:

[0161] The shield-induced settlement value predicted by the MHSA-Bi-LSTM surrogate model is used as the fitness function, and the goal is to minimize this settlement value.

[0162] For each particle i , place its position The remaining parameters of the current ring location (geometric and geological parameters use the original conditions, while parameters of types B and C use the statistical mean) are input into the surrogate model to calculate the corresponding predicted values ​​of shield-induced settlement. This value is the fitness of the particle;

[0163] Step 3.1.3, Individual Optimal and Global Optimal Updates:

[0164] For each particle i Compare its current fitness Compared with the previously recorded individual optimal fitness pbest i ;like < pbest i Then update the individual's optimal position. pbest posi = and individual optimal fitness pbest i = ;

[0165] Find the minimum value among all the individual optimal fitnesss of the particles, and let the corresponding particle be denoted as . j ;like pbest j Less than the currently recorded global optimal fitness gbest Then update the global optimal position. gbest pos = pbestposj and global optimal fitness gbest = pbest j ;

[0166] Step 3.1.4, Velocity and Position Update:

[0167] Update the velocity and position of each particle according to formulas 7-8:

[0168] Speed ​​Update: Formula 7: ;

[0169] In the formula, This is the inertial weight, used to balance the global and local search capabilities of particles, and is generally set to 0.7-1.2. and It is the acceleration constant, typically taken as 1.5; r 1 and r 2 is a random number in the interval [0,1].

[0170] Position Update: Formula 8: ;

[0171] After updating the position, it is necessary to check whether the position of each particle exceeds the range of parameter values. If it exceeds the range, its position is adjusted to the boundary value.

[0172] Step 3.1.5, Termination condition determination:

[0173] Check if the termination condition is met. The termination condition is reaching the preset maximum number of iterations. T max If the global optimal fitness remains unchanged for 200 consecutive iterations, the algorithm terminates and outputs the global optimal position if the termination condition is met. gbest pos The optimized A-type parameter values ​​are used; if the termination condition is not met, return to step 3.1.2 to continue the iteration;

[0174] Step 3.1.6, PSO algorithm output:

[0175] After the algorithm finishes, it outputs the globally optimal position. gbest pos The value of the A-type parameter corresponding to this position is the result obtained by the PSO algorithm optimization, which can be used to guide the recommendation of shield parameters for this ring position; the optimal fitness is output, which corresponds to the minimum shield-induced settlement predicted by the surrogate model;

[0176] Step 3.1.7: Optimize Class A parameters for all predetermined loop positions;

[0177] Monitor whether the termination conditions of the predetermined ring position are met. If they are met, output the recommended shield parameters for all predetermined ring positions. If not, return to step 3.1.1 and continue the optimization operation for the next ring position.

[0178] Step 3.2: Establish the B-type parameter inversion algorithm flow based on the RF model:

[0179] The process for establishing and applying the type B parameter inversion model is as follows: Figure 5 As shown, the training input data for inversion model 2 includes geometric parameters, geological parameters, initial type A parameters, settlement data (historical settlement and shield-induced settlement), and type B parameters as output.

[0180] Determine the optimal combination of hyperparameters for RF using a grid search method:

[0181] Number of decision trees tr ∈[50:5:200]; Minimum number of leaf node samples lf ∈[1:1:5];

[0182] A 5-fold cross-validation approach was adopted, with mean squared error as the objective function, and the accuracy was verified using a test set.

[0183] By inputting geometric geological parameters ( Dc , Mc , Co , A , K , Wt ), optimized Class A parameters (after PSO optimization) Ve', Sp', Rc', Tc' ), settlement data (historical settlement) Sh 0、 Sh 7.5 and the predicted minimum shield-induced settlement Ss' This process is used to deduce new Class B parameters as recommended shield tunneling parameters; the recommended Class B parameters are then output. Th' , Tsc' , Vg' ;

[0184] Step 3.3: Establish the C-type parameter inversion algorithm flow, based on the RF model:

[0185] The process of establishing and applying the C-type parameter inversion model is as follows: Figure 5 As shown, the input data for inversion model 3 includes geometric parameters, geological parameters, initial A and B type parameters, settlement data (historical settlement and shield-induced settlement), and C type parameters as output;

[0186] Determine the optimal combination of hyperparameters for RF using a grid search method:

[0187] Number of decision trees tr∈[50:10:200]; Minimum number of leaf node samples lf ∈[1:1:5];

[0188] A 5-fold cross-validation approach was adopted, with mean squared error as the objective function, and the accuracy was verified using a test set.

[0189] By inputting geometric geological parameters ( Dc , Mc , Co , A , K , Wt ), optimized Class A parameters (after PSO optimization) Ve', Sp', Rc', Tc' ), and the inverted type B parameters (inverted by inversion model 2) Th' , Tsc' , Vg' ), settlement data (historical settlement) Sh 0、 Sh 7.5 Predicted minimum shield-induced settlement Ss' This process is used to deduce new Class C parameters as recommended shield tunneling parameters, and the recommended Class C parameters are output: Rsc' , Sa' ;

[0190] Among them, hierarchical inversion increases the dimension of input parameters, preserves more parameter interaction relationships, and improves the inversion prediction accuracy of actual controllable operating parameters.

[0191] Step 4: Implement parameter recommendation and real-time control through the developed shield tunneling parameter optimization system (SPOS).

[0192] Preferably, the steps for parameter recommendation and real-time control through the developed shield tunneling parameter optimization system (SPOS) include:

[0193] Step 4.1, Basic Data Collection; the specific process is as follows:

[0194] SPOS collects and receives multi-source data for the next construction phase, including geometric parameters, geological parameters, shield parameters, and settlement data as described in step 1.1.

[0195] The collected multi-source data is aligned and preprocessed to eliminate data errors and noise;

[0196] Step 4.2, Model Iterative Optimization:

[0197] As the tunnel boring machine (TBM) progresses, new sample data (settlement data and parameters collected every 5 rings) are continuously added to the SPOS training database, expanding the sample size. Through incremental learning, the prediction error of the MHSA-Bi-LSTM model is gradually reduced.

[0198] The new sample data generated during the tunnel boring process are continuously added to the settlement prediction proxy model and the RF inversion model in steps 3.2 and 3.3. By continuously updating the model parameters and structure through incremental learning, the model can better adapt to the dynamic changes during the construction process.

[0199] Through continuous iteration and optimization, new sample data is incorporated into the model, which is able to learn this new information, thereby improving the accuracy of predicting future outcomes.

[0200] Step 4.3, Parameter Recommendation and Dynamic Update:

[0201] Based on new sample data and optimized models, SPOS generates recommended shield tunneling parameters in real time for the current construction situation; the optimized shield tunneling parameters are transmitted to the shield machine operation interface in real time to realize dynamic calibration of parameters. As geological conditions and other factors change during construction, SPOS dynamically updates the recommended parameter scheme based on real-time data and model feedback.

[0202] In the upstream shield tunneling section of Changchun Metro Line 6, from Feiyue Square Station to Eurasia Mall Station, rings 170 (ZK17+321.735 on January 3, 2022) to 665 (ZK16+579.235 on May 21, 2022) constitute the initial data acquisition segment. The shield tunneling did not stop during the data recording phase. Rings 665-740 constitute the initial closed-loop feedback segment. SPOS application began at ring 745. Subsequent shield settlement results from rings 745-840, pre-controlled by the system, are as follows: Figure 6 As shown.

[0203] The technological advantage and innovation of this part lies in the close integration of model prediction, parameter adjustment, and actual settlement feedback, forming a closed-loop feedback mechanism. This allows SPOS to continuously self-adjust and optimize to adapt to different construction conditions and changes. This adaptability and robustness solves the problem of accuracy degradation caused by the variability of geological parameters in traditional models, making it suitable for the dynamic environment of long-term tunnel construction.

[0204] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made using the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An adaptive adjustment method for EPB construction parameters for settlement pre-control, characterized in that, The method includes: Step 1: Develop a bidirectional long short-term memory neural network prediction model with enhanced multi-head self-attention mechanism to capture the temporal and spatial correlation of shield-induced settlement, namely the MHSA-Bi-LSTM model. Step 2: Construct a multi-classification control strategy for shield tunneling parameters for settlement pre-control; classify shield tunneling parameters into categories A, B, and C based on shield-induced settlement sensitivity and human controllability; Step 3: Construct a shield tunneling parameter optimization mechanism based on optimization and composite inversion; use the particle swarm optimization algorithm to optimize type A parameters to minimize settlement, and combine it with the random forest algorithm to invert type B and type C parameters in sequence; Step 4: Implement parameter recommendation and real-time control through the developed shield tunneling parameter optimization system; The steps for achieving parameter recommendation and real-time control through the developed shield tunneling parameter optimization system include: Step 4.1, Basic Data Collection: The developed shield tunneling parameter optimization system SPOS collects and receives multi-source data for the next construction phase, including geometric parameters, geological parameters, shield tunneling parameters, and settlement data. The collected multi-source data is aligned and preprocessed to eliminate data errors and noise; Step 4.2, Model Iterative Optimization: The new sample data generated during the tunnel boring process are continuously added to the settlement prediction surrogate model and the RF inversion model with B-type and C-type parameters, and the model parameters and structure are continuously updated through incremental learning. Step 4.3, Parameter Recommendation and Dynamic Update: Based on the optimized model using new sample data, SPOS generates recommended shield tunneling parameters in real time for the current construction situation. The optimized shield tunneling parameters are transmitted to the shield machine operation interface in real time to achieve dynamic calibration of the parameters. As geological conditions change during construction, SPOS dynamically updates the recommended parameter scheme based on new sample data and model feedback.

2. The adaptive adjustment method for EPB construction parameters for settlement pre-control according to claim 1, characterized in that, The content of the development of a regularized long short-term memory neural network prediction model that simultaneously captures the temporal and spatial correlation of shield tunneling-induced settlement is as follows: Step 1.1: Define the variable form, with a sampling unit of 7.5 meters. The input parameters include geometric parameters, geological parameters, shield parameters and historical settlement, and the output parameter is shield-induced settlement. The depth of soil cover was determined to be the only geometric parameter that needed to be considered, and the measurement result was denoted as Dc. Geological parameters are weighted according to the thickness of test samples at each survey location, and are corresponding to the shield location through interpolation. The calculated compression modulus is denoted as Mc, cohesion as Co, internal friction angle as A, lateral pressure coefficient as K, and groundwater level as Wt. The shield tunneling parameters should prioritize those that have direct or indirect interaction with the soil; the synchronous grouting volume Vg, total thrust Th, cutterhead torque Tc, cutterhead speed Rc, support pressure Sp, propulsion speed Sa, screw conveyor speed Rsc, screw conveyor torque Tsc, and excavation volume Ve are selected as shield tunneling parameters. Settlement data includes historical settlement and shield-induced settlement. It needs to be synchronized with shield parameters. Measurements were taken twice daily along the tunnel's longitudinal axis at the same 7.5-meter intervals, with the monitoring frequency increased to 3 times per hour within 15 meters before and after the shield excavation location. Surface settlement at 7.5 meters and 0 meters before the shield's arrival were respectively considered historical settlement, denoted as Sh. 7.5 Sh0; The surface settlement 10 days after the completion of the annular tunneling is taken as the shield-induced settlement and is denoted as Ss; Step 1.2: Construct the MHSA-Bi-LSTM model architecture, which adopts the configuration of Bi-LSTM layer + self-attention layer + Dropout layer + Bi-LSTM layer; Step 1.3: Train the model. The training data is processed by min-max normalization to convert it into the interval [0,1]. The model hyperparameters were determined using a grid search method, including stride [1:1:5], number of hidden units in the Bi-LSTM layer [10:10:500], dropout rate in the Dropout layer [0.1:0.05:0.6], number of heads in the self-attention layer [2:1:10], and number of key channels [1-10 times the number of heads]. The optimal architecture of the proposed model was determined using a five-fold time series cross-validation method. The implicit form of the final MHSA-Bi-LSTM settlement prediction model is: Formula 1: ; In the formula, X is the input vector, X=[Dc,Mc,Co,A,K,Wt,Sa,Rc,Tc,Sp,Vg,Th,Tsc,Rsc,Ve,Sh0,Sh 7.5 ]; To conceal the state, enhance the ability to remember the sedimentation sequence.

3. The adaptive adjustment method for EPB construction parameters for settlement pre-control according to claim 1, characterized in that, The MHSA-Bi-LSTM model architecture is constructed using a configuration of Bi-LSTM layer + Self Attention layer + Dropout layer + Bi-LSTM layer, as shown below: The first layer is the Sequence Input Layer, which serves as the network's entry point and is responsible for receiving input data. The second layer is the flatten layer; The third layer is the Bi-LSTM-1 layer; The fourth layer is the Self Attention Layer, which projects its output onto multiple subspaces to compute context weights; The fifth layer is the Dropout Layer, which enhances regularization; The sixth layer is another Bi-LSTM layer, Bi-LSTM-2 Layer, which has the same function as the previous Bi-LSTM layers; The seventh layer is a fully connected layer, which performs a linear transformation on the output of the previous Bi-LSTM layers, mapping the features to a specified dimension; The eighth layer is the Regression Output Layer, which is used for regression tasks. It calculates the loss between the predicted value and the true value and performs backpropagation during training to update the network parameters. The optimizer is configured as Adam adaptive moment estimation. To mitigate instability during training, a segmented scheduling strategy is used to modulate the learning rate, starting from an initial learning rate of 0.001, with a descent period of 400 and a descent factor of 0.

15.

4. The adaptive adjustment method for EPB construction parameters for settlement pre-control according to claim 2, characterized in that, The content of the multi-classification control strategy for shield tunneling parameters for settlement pre-control is as follows: Step 2.1: Perform sensitivity analysis using the Sopol method: From Formula 1, we can see that the assumed proxy model takes the following form: Formula 2: ; In the formula, For the output of the surrogate model, the unconditional variance is... It can be decomposed into partial variance accordingly: Formula 3: ; In the formula, and These represent the partial variance related to a single input variable and the partial variance related to the interaction between two input variables, respectively. and The variance; For all The partial variance resulting from the combined effect of the input variables; Using these biases, the first-order, second-order, and total Sopol sensitivity indices are defined according to the total variance, as shown in Formulas 4-6: Formula 4: ; Formula 5: ; Formula 6: ; In the formula, MSI is the principal sensitivity index, or the first-order index S. i , indicating a single input X i The degree of interference to the output when the value fluctuates; S ij It is a second-order exponent, or X i and X j The second-order interaction effect index represents the degree to which the interaction between two inputs interferes with the output; V ~i It is divided by X i The variance of the other indicators; TSI stands for Total Sensitivity Index, or ST. i , representing when the input index X i When the value fluctuates, the total degree of interference to the output caused by all its effects is the sum of the sensitivity indices of each order of the input index; Based on the range of parameter variables, Latin hypercube sampling is performed, and a proxy model is applied to calculate the correlation to determine the MSI and TSI of each parameter under different sample sizes. When the change in the results is less than 1% as the sample size increases, the results are considered to have converged, and these are the final MSI and TSI. Step 2.2: Perform multiple classifications of shield tunneling parameters for settlement pre-control: A classification control strategy is proposed, which divides the shield tunneling parameters into three categories: A, B, and C. Step 2.2.1: Classification Requirements: The cumulative TSI of Class A parameters on shield-induced settlement is ≥90% to demonstrate their dominant role in settlement. Parameters with higher TSI rankings are selected first. The remaining low-sensitivity parameters are classified into Class B and Class C parameters according to their controllability during operation. Step 2.2.2: The classified A-class parameters are validated using the parameter interaction matrix, requiring a Pearson correlation coefficient ≤ 0.2; Step 2.2.3: If the verification requirements in step 2.2.2 are not met, remove the strongly related Class A parameters and return to step 2.2.1 to re-screen.

5. The adaptive adjustment method for EPB construction parameters for settlement pre-control according to claim 1, characterized in that, The steps for constructing a shield tunneling parameter optimization mechanism based on optimization-composite inversion include: Step 3.1, Optimize the search for Class A parameters: Step 3.1.1, PSO particle initialization: The particle swarm size N is set, and the position and velocity of each particle are randomly initialized. The position of each particle represents a combination of values ​​of a set of A-type parameters (A1, A2, ..., An), while the velocity represents the speed at which the particle moves in the parameter space. The position of particle i is... ,speed ; Based on the actual project conditions, the value range for each Class A parameter should be clearly defined. ; Step 3.1.2, Fitness Calculation: The shield-induced settlement value predicted by the MHSA-Bi-LSTM surrogate model is used as the fitness function, and the goal is to minimize this settlement value. For each particle i, determine its position. The remaining parameters of the current ring position are input into the surrogate model to calculate the corresponding predicted value of shield-induced settlement. This value is the fitness of the particle; Step 3.1.3, Individual Optimal and Global Optimal Updates: For each particle i, compare its current fitness. pbest fitness of the previously recorded individual i ;like <pbest i Then update the individual's best position pbest. posi = and individual optimal fitness pbest i = ; Find the minimum value among all the individual optimal fitnesss of the particles, and let the corresponding particle be j; if pbest j If the global best fitness (gbest) is less than the currently recorded global best fitness, then update the global best position (gbest). pos =pbest posj And the global optimal fitness gbest=pbest j ; Step 3.1.4, Velocity and Position Update: Update the velocity and position of each particle according to formulas 7-8: Speed ​​Update: Formula 7: ; In the formula, Inertial weights are used to balance the global and local search capabilities of particles. and It is the acceleration constant; r1 and r2 are random numbers in the interval [0,1]. Position Update: Formula 8: ; After updating the position, it is necessary to check whether the position of each particle exceeds the range of parameter values. If it exceeds the range, its position is adjusted to the boundary value. Step 3.1.5, Termination condition determination: Check if the termination condition is met. The termination condition is reaching the preset maximum number of iterations T. max If the global optimal fitness remains unchanged for 200 consecutive iterations, the algorithm terminates and outputs the global optimal position gbest, provided the termination condition is met. pos The optimized A-type parameter values ​​are used; if the termination condition is not met, return to step 3.1.2 to continue the iteration; Step 3.1.6, PSO algorithm output: After the algorithm finishes, it outputs the globally optimal position gbest. pos The value of the A-type parameter corresponding to this position is the result obtained by the PSO algorithm optimization, which can be used to guide the recommendation of shield parameters for this ring position; the optimal fitness is output, which corresponds to the minimum shield-induced settlement predicted by the surrogate model; Step 3.1.7: Optimize Class A parameters for all predetermined loop positions; Monitor whether the termination conditions of the predetermined ring position are met. If they are met, output the recommended shield parameters for all predetermined ring positions. If not, return to step 3.1.1 and continue the optimization operation for the next ring position. Step 3.2: Establish the B-type parameter inversion algorithm flow based on the RF model: The input data includes geometric parameters, geological parameters, initial type A parameters, historical settlement and shield-induced settlement, and type B parameters are output. Determine the optimal combination of hyperparameters for RF using a grid search method: The number of decision trees tr∈[50:5:200]; the minimum number of leaf node samples lf∈[1:1:5]; A 5-fold cross-validation approach was adopted, with mean squared error as the objective function; By inputting geometric parameters, geological parameters, optimized Class A parameters, historical settlement, and predicted minimum shield-induced settlement, new Class B parameters are derived as recommendations for shield parameters. Step 3.3: Establish the C-type parameter inversion algorithm flow, based on the RF model: The input data includes geometric parameters, geological parameters, initial A and B type parameters, historical settlement and shield-induced settlement, and C type parameters are output. Determine the optimal combination of hyperparameters for RF using a grid search method: The number of decision trees tr∈[50:10:200]; the minimum number of leaf node samples lf∈[1:1:5]; A 5-fold cross-validation approach was adopted, with mean squared error as the objective function; By inputting geometric parameters, geological parameters, optimized type A parameters, inverted type B parameters, historical settlement, and predicted minimum shield-induced settlement, new type C parameters are inverted as recommendations for shield parameters. Among them, hierarchical inversion increases the dimension of input parameters, preserves more parameter interaction relationships, and improves the inversion prediction accuracy of actual controllable operating parameters.

Images