Method and system for controlling settlement of building group induced by TBM (Tunnel Boring Machine) tunneling through data physical dual-drive

Abstract

The invention discloses a method and a system for controlling settlement of a building group induced by TBM (Tunnel Boring Machine) tunneling through data and physics dual drive, which are used for controlling settlement of the building group induced by TBM construction of a shield tunneling machine. The method comprises the steps that a settlement calculation formula under the complex stratum condition is proposed, the formula serves as a physical loss function to be embedded into machine learning, and a physical information machine learning PIMLMS model considering the complex stratum is formed; and introducing the established PIMLMS prediction model as an objective function into a multi-objective optimization algorithm, taking the settlement value of the building group as an optimization objective, and reducing the settlement of the building group caused by TBM tunneling to the greatest extent by controlling TBM parameters. According to the method, accurate prediction of building group settlement induced by TBM tunneling in the complex stratum is achieved, and meanwhile a feasible scheme is provided for reducing the settlement problem of the building group.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel boring machine construction, and in particular to a data- and physics-based dual-driven method and system for controlling the settlement of buildings induced by TBM excavation. Background Art

[0002] With the advancement of urbanization, urban population and building densities are increasing, necessitating the urgent need to improve underground transportation infrastructure. Tunnel construction methods primarily include open-cut tunneling, the New Austrian Tunneling Method (NATM), and the shield tunneling machine (TBM). The TBM has been widely used in urban tunnel construction in recent years due to its operational safety, high degree of mechanization, and minimal impact on the surrounding environment. However, during construction, soil disturbance and longitudinal displacement are inevitable due to variations in TBM performance, worker experience, and the complexity of the strata. This can lead to surface subsidence along the tunnel. When surface subsidence reaches a certain level, it can significantly impact the surrounding environment, buildings, and underground pipelines, raising a series of safety concerns. Metro planning areas are primarily concentrated in densely populated and built-up urban agglomerations, placing high demands on controlling ground subsidence. Furthermore, tunnels often traverse diverse regional geological environments with varying historical origins and complex geological structures. This poses significant challenges in controlling tunnel-induced subsidence.

[0003] While existing research has considered optimizing TBM parameters, few have explicitly considered building settlement as a TBM operation objective. Furthermore, previous studies investigating control methods often treat complex geological conditions as mere input variables, ignoring the underlying physical principles. In particular, the relationship between geological data and TBM operating parameters is often a key factor influencing settlement. Therefore, developing a rational and effective settlement control method is essential. Summary of the Invention

[0004] In order to overcome the above problems, the purpose of the present invention is to propose a data- and physics-driven method and system for controlling the settlement of buildings induced by TBM excavation.

[0005] According to one aspect of the present invention, a data-physics dual-driven method for controlling TBM-induced building settlement is disclosed. A settlement calculation formula that takes complex stratum conditions into account is proposed. This formula is embedded into machine learning as a physical loss function. Combining complex stratum characteristics with physical constraints, a physical information machine learning PIMLMS model that considers complex strata is established. This method can reasonably and reliably simulate building settlement caused by TBM construction under complex stratum conditions.

[0006] The physical information machine learning algorithm PIMLMS, which considers complex strata, is integrated into the multi-objective optimization algorithm MOO to form a hybrid algorithm of the physical information machine learning algorithm PIMLMS and the multi-objective optimization algorithm MOO. A settlement control system based on dynamic adjustment of TBM parameters is established to reduce the impact of TBM construction on building settlement by controlling TBM parameters.

[0007] Furthermore, the hybrid algorithm aims to minimize the building settlement caused by TBM excavation;

[0008] Firstly, the building settlement prediction model was established using the PIMLMS model;

[0009] Then, the prediction model is introduced as the objective function, and a multi-objective optimization model based on the non-dominated sorting genetic algorithm IINSGA-II and TOPSIS is established;

[0010] Finally, the multi-objective optimization algorithm iteratively generates the optimal result under the control of the TBM's six operating parameters (total thrust, propulsion speed, cutterhead speed, cutterhead torque, screw conveyor speed, and screw conveyor pressure).

[0011] According to another aspect of the present invention, a data- and physics-driven TBM excavation-induced building complex settlement control system is disclosed, which adopts the above-mentioned data- and physics-driven TBM excavation-induced building complex settlement control method.

[0012] The present invention minimizes the impact of tunnel boring machine (TBM) parameters on building settlement by adjusting them. Compared to traditional machine learning models, this method improved the prediction accuracy of PIMLMS by 8.41% and reduced building settlement by 40.90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of the method for controlling settlement of buildings within the influence range of tunnel construction provided by the present invention.

[0014] Figure 2 Schematic diagram of ground subsidence in the double tunnels provided by the present invention.

[0015] Figure 3 This is a schematic diagram of the stratigraphic distribution of the building complex locations in the embodiment provided by the present invention.

[0016] Figure 4 This is a schematic diagram of the PIMLMS structure provided by the present invention. DETAILED DESCRIPTION

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] like Figure 1-4A data-physics dual-driven method for controlling building settlement induced by tunnel construction is proposed. A settlement calculation formula under complex stratum conditions is proposed, and the formula is embedded in machine learning as a physical loss function. Combined with the characteristics of complex strata and physical constraints, a physical information machine learning PIMLMS model considering complex strata is established to reasonably and reliably simulate the building settlement caused by TBM construction under complex stratum conditions.

[0019] The physical information machine learning (PIMLMS) that considers complex strata is integrated into the multi-objective optimization algorithm (MOO) to form a hybrid algorithm of PIMLMS and MOO. A settlement control system based on dynamic adjustment of TBM parameters is established to minimize the impact of TBM construction on building settlement by controlling TBM parameters.

[0020] The hybrid algorithm of PIMLMS and MOO is designed to minimize the building settlement caused by TBM excavation.

[0021] Firstly, the PIMLMS model is used to establish a building settlement prediction model.

[0022] Then, the prediction model is introduced as the objective function, and a multi-objective optimization model based on non-dominated sorting genetic algorithm IINSGA-II and TOPSIS is established.

[0023] Finally, the multi-objective optimization algorithm MOO iteratively generates the optimal result under the control of the six operating parameters of the TBM (total thrust, propulsion speed, cutterhead speed, cutterhead torque, screw conveyor speed, and screw conveyor pressure).

[0024] The specific steps are as follows:

[0025] Step 101, deriving a calculation formula for building settlement caused by excavation of a parallel double-track tunnel in a complex stratum;

[0026] The settlement curve caused by tunnel excavation is roughly similar to the normal distribution curve. For the excavated tunnel, the maximum settlement value appears at the tunnel centerline, e.g. Figure 2 The two settlement curves are superimposed to form the settlement curve of the two-lane tunnel. The calculation formula is: For the excavated tunnel, the maximum settlement value occurs at the tunnel centerline:

[0027]

[0028] Where: S(x,y,z) is the settlement at the coordinate (x,y,z); x is the coordinate along the tunnel excavation direction; y is the horizontal coordinate perpendicular to the tunnel excavation direction; z is the vertical coordinate perpendicular to the tunnel excavation direction; L is the distance between the tunnel axes; S max is the maximum settlement caused by single-line tunnel excavation, and the calculation formula is:

[0029]

[0030] Where i is the width coefficient of the sedimentation tank:

[0031]

[0032] V S is the soil loss rate per unit length of the tunnel:

[0033] V S =ηπR 2 (4)

[0034] Where H is the depth of the tunnel axis from the ground; is the friction angle of the soil; η is the soil loss rate; R is the radius of the tunnel.

[0035] By combining the above formulas, the calculation formula for the settlement value at point (x, y, z) is:

[0036]

[0037] The geological conditions during TBM excavation often involve complex strata. Traditional machine learning uses geological drilling data as input to establish an input-output mapping relationship between geological conditions and settlement. However, this purely data-driven approach lacks interpretability, and the complexity and variability of soil strata result in poor generalization capabilities of machine learning. This paper provides a settlement calculation method considering complex strata:

[0038]

[0039] Among them, S all (x,y,z) represents the total settlement value at point (x,y,z); n is the total number of strata at point (x,y,z); S i (x, y, z) represents the settlement value of the i-th layer at point (x, y, z), and its calculation formula is:

[0040]

[0041] Among them, z i is the thickness of the i-th layer, z all is the total thickness of the soil. Substituting formula (5) and formula (7) into formula (6), the sum of the settlement values ​​of each soil layer at point (x, y, z) can be obtained, as shown in formula (8):

[0042]

[0043] Step 102: embed the settlement calculation formula into machine learning as a physical loss function to form a data-physics dual-driven settlement prediction model;

[0044] In physical information machine learning (PIML), employing efficient and high-performance algorithms is crucial. The gradient boosting framework (LightGBM), an improved algorithm based on the gradient boosting decision tree (GBDT), has become the preferred algorithm in the PIML field due to its efficient computational speed and excellent learning results. Physical formulas often have complex nonlinear relationships between variables. LightGBM can capture these complex relationships and gradually approximate the true value by constructing multiple decision trees, thereby improving the interpretability of the model.

[0045] LightGBM significantly improves the training efficiency and prediction performance of the model by introducing technologies such as histogram algorithm, gradient-based single-side sampling (GOSS) and mutually exclusive feature bundling (EFB). GOSS efficiently extracts significant data features by prioritizing high-gradient samples, while randomly sampling low-gradient samples to retain information-rich data. The specific process is as follows: First, the training instances are sorted in descending order and divided into two subsets according to the absolute gradient value. In this method, subset A contains data samples with larger gradients, and their gradient values ​​account for a×100%; subset B consists of the remaining low-gradient samples A. S The random sampling composition is b×|A S |, the gradient value accounts for (1-a)×100%. Then, the algorithm will calculate the estimated variance gain of feature j in the subset A∪B by formula (9) And based on this, the decision of feature splitting point is made:

[0046]

[0047] Among them A l =[x i ∈A:x ij <d],A r =[x i ∈A:x ij >d]; B l =[x i ∈B:x ij <d];B r =[x i ∈B:x ij >d],g i is the negative gradient of the loss function.

[0048] In order to accurately predict building settlement based on actual data such as TBM parameters and soil layers and physical laws based on equation settlement formulas, a hybrid physical information driven settlement prediction model is constructed, such as Figure 4As shown in Figure 2. LightGBM is a decision tree-based gradient boosting framework whose core concept is to improve model performance by iteratively optimizing the loss function. The first and second derivatives of the loss function are used to guide the update of model parameters. Unlike traditional GBDT, LightGBM not only relies on the first-order derivative for optimization but also incorporates information from the second-order derivative to more accurately describe the curvature of the loss function, making model updates more stable and efficient. Combining the physical formula of sedimentation with the loss function of the LightGBM model imposes certain physical constraints on the training process and the extraction of data associations.

[0049] These constraints are used to build more accurate and scientific machine learning models. The custom loss function of the model is defined as formula (10):

[0050]

[0051] The present invention uses mean square error (MSE) as the data-driven component of the loss function, as shown in formula (11):

[0052]

[0053] Similarly, the definition of physical loss can be expressed by formula (12):

[0054]

[0055] Where N is the total number of test sets; y i is the label value; y phy,i is the predicted value calculated by the physical equation; Represents the predicted value obtained by the data-driven method; Loss phy Represents the residual between the physical formula value and the label; Loss pre Represents the residual between the model loss function value and the ground truth value. λ represents the proportion of physical loss to total loss. By adjusting the value of λ, loss functions with different deviations can be obtained. The present invention tried different values ​​and found that this is an effective choice.

[0056] The first-order and second-order derivatives of the loss function are shown in formulas (13) and (14):

[0057]

[0058] Step 103: Use model-based sequential optimization to fine-tune the model hyperparameters and establish performance evaluation indicators for the prediction model:

[0059] In order to measure the performance of the prediction model, the present invention uses three evaluation indicators, namely, mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2The closer the MAE and RMSE values ​​are to zero, the closer the model's predicted values ​​are to the actual settlement values. 2 The closer the value is to 1, the higher the correlation between the settlement value and other input parameters (such as TBM parameters). 2 The calculation of is shown in formulas (15) to (17):

[0060]

[0061] Where N represents the total number of test samples, f i (x) is the predicted value of the model, y i is the true value of the sample. During the PIML model training process, the choice of hyperparameters has a decisive influence on the model performance. As an efficient gradient boosting framework, LightGBM is very sensitive to the configuration of hyperparameters, which requires us to perform fine-grained tuning to obtain the best performance. The present invention adopts the model-based sequential optimization (SMBO) technology and implements hyperparameter optimization through the Parsons Estimator Tree (TPE) algorithm. Among them, the optimization objective function adopts the expected improvement (EI) criterion, and its mathematical expression is shown in formula (18):

[0062]

[0063] in Represents the expected improvement relative to the current best objective value, and by maximizing this value, new sampling points can be selected to find a better hyperparameter configuration. * is the lowest observed objective function value, used as a benchmark to determine whether a new sampling point produces better results. x is the current sampling point being evaluated, which represents a specific set of hyperparameters. y is the possible value x of the objective function at that sampling point. Because Bayesian optimization makes predictions about the function f(x) based on a probabilistic model, the observed value y is treated as a random variable. p(y|x) represents the probability density function of the objective function value y at a given point x, quantifying the likelihood of y taking different values ​​when the model makes a prediction for x.

[0064] Step 104: introduce the established building settlement prediction model as the objective function into the multi-objective optimization algorithm MOO, take the settlement value of the building complex as the optimization target, and establish a multi-objective optimization model;

[0065] The MOO method has been widely used for parameter optimization in the design, construction, operation, and maintenance phases of civil engineering projects. This paper uses the settlement values ​​of three buildings along the TBM excavation route as the three target variables and selects six TBM operating parameters as the independent variables. The MOO model can be expressed as follows:

[0066] min F(x i )=[f1(xi ),f2(x i ),f3(x i )] (i=1,2,…,6) (19)

[0067] The constraints are as follows:

[0068] g j (xi)≤0 (j=1,2,3) (20)

[0069]

[0070] Where F(x i ) is based on three target variables f1(x i ),f2(x i ),f3(x i ) to construct a multi-objective optimization function; g j (x) is an inequality constraint; and and are respectively x i The lower and upper limits of the value.

[0071] Step 105 , by controlling TBM parameters, generating the Pareto front of the multi-objective model, and using the TOPSIS method to generate the optimal solution, ultimately achieving the goal of minimizing the settlement of the building complex caused by TBM excavation.

[0072] The Pareto front of the MOO algorithm in the present invention is generated using the NSGA-II algorithm based on the geatpy library. NSGA-II is an evolutionary algorithm for multi-objective optimization. The process starts with generating an initial population of size P, where each individual represents a potential solution. Individuals are divided into different Pareto fronts through non-dominated sorting. Within each front, the crowding distance (CD) is calculated to evaluate the diversity of solutions. Selection is then made based on the Pareto front and crowding distance, with priority given to lower fronts and higher CD values. The selected individuals undergo genetic operations such as crossover and mutation to create a new offspring population. The current population is merged with the offspring, and the process of non-dominated sorting and crowding distance calculation is repeated. This iterative process will continue until the algorithm reaches the termination criterion, and eventually an approximate Pareto front with both optimality and diversity balance is generated.

[0073] To select the optimal solution from the Pareto front, the TOPSIS method is used for analysis. The TOPSIS method ranks candidate solutions based on their Euclidean distance to the ideal solution and the most negative solution. The score of each solution (denoted as ) is calculated using these distances. It represents the Euclidean distance to the most ideal solution and the distance to the most negative solution. The score ranges from 0 to 1, and the solution with the highest score is considered to be the optimal solution in the Pareto front.

[0074] Example studies:

[0075] This method was validated using Wuhan Metro Line 19 as an example. Data was collected from a tunnel stretching approximately 1 km between the West Square of Wuhan Railway Station and Wudong Station. Building settlement values ​​and corresponding TBM parameters were collected over a 21-day period.

[0076] The raw data on which this embodiment is based were manually recorded by staff and organized into a series of PDF files by time. Each file records in detail the settlement data of each monitoring point on that day. In this embodiment, a total of 21 sets of PDF data were collected, totaling 88 TBM excavation rings, which lasted 21 days, with the TBM excavating an average of 4 rings per day. The collected raw data is in units of days and cannot be used directly. The finite element method was introduced to model the data, which plays an important role in subsequent processing. Using the finite element method, drilling measurements and ground survey reports were incorporated into the numerical model to construct a geological structure model. We identified three buildings that were most affected by TBM excavation as our training samples. Each building has 4 monitoring points, and each monitoring point has 21 sets of data, so 84 sets of raw data were collected for each building. On this basis, the finite element method was used to supplement the potential settlement data of each building during the TBM excavation process. After finite element simulation, another 268 sets of data were generated and grouped into the training set.

[0077] For model features, the present invention screens and combines the operating parameters of the TBM, and only selects the average value of the TBM during the stable operation stage as the input feature. The y and z coordinates of the observation points are further combined to form a 9-dimensional feature space. The geological data is not directly used as model features, but indirectly affects the model prediction through the physical loss function. Finally, we obtained three arrays from the measurement data, each containing 84 rows and 10 columns, for model testing. Through the finite element method, we obtained three arrays, each containing 268 rows and 10 columns, for model training. In summary, the data of the training set and the test set account for 76% and 24% of the total data, respectively.

[0078] The prediction model trained by the present invention needs to be further combined with the MOO algorithm. Therefore, the hybrid model has three basic responsibilities: (1) describing the building settlement caused by TBM excavation. (2) embedding physical laws in the training process and introducing geological conditions into the model. (3) optimizing TBM parameters to achieve the overall optimal value of building settlement of the building complex. In order to achieve the second point, the geological conditions of the case need to be added to the data set. In this case, the strata are divided into five types (as shown in Table 1). In the present invention, six key TBM operating parameters (x1 to x6) are selected as the input data of the training model (as shown in Table 2), and the optimization targets S1, S2 and S3 are set at the same time. Table 3 shows the distribution characteristics of TBM parameters.

[0079] Table 1

[0080]

[0081] Table 2

[0082]

[0083] Table 3

[0084]

[0085] The PIMLMS model requires the following key hyperparameters: the number of leaves N l : The maximum number of leaf nodes in a single tree, controlling the model complexity; learning rate η; data sampling ratio F b ; Feature sampling ratio F f ; Maximum tree depth D max ; The maximum level allowed for a single decision tree to grow; The minimum number of leaf samples L min : The minimum amount of data required for a single leaf node. Bayesian optimization has good performance in hyperparameter search. Within an appropriate range, Bayesian optimization obtains a large number of hyperparameters to build a model, which can effectively optimize hyperparameters. The hyperparameter search space and the optimal hyperparameter results are shown in Table 4. In addition, the present invention sets hyperparameters for NSGA-II, as shown in Table 5. This set of hyperparameters is designed to achieve efficient search and fast convergence of the algorithm, and is optimized based on problem characteristics, algorithm design, and computing resources.

[0086] Table 4

[0087]

[0088] Table 5

[0089]

[0090] Result analysis:

[0091] In this paper, we use PIMLMS combined with the NSGA-II algorithm to obtain the optimal control solution for a TBM. To evaluate the performance of this hybrid algorithm, we conducted various analyses, including comparisons with other models and in-depth analysis of the model's output. The specific results are as follows:

[0092] (1) The proposed PIMLMS outperforms other ML prediction models in terms of prediction accuracy. Table 6 provides the corresponding performance indicators (MAE, RMSE and R 2 ). The model comparison results in Table 6 show that the proposed PIMLMS model outperforms other considered ML models for three different targets (s1 and S3). Specifically, from the determination coefficient R 2 Evaluation index analysis shows that the PIMLMS model shows excellent fitting ability. The R 2 The values ​​are 0.943, 0.918 and 0.946 respectively, which are significantly better than other comparison models and show its leading position in accuracy. Taking target S1 as an example, the mean absolute error MAE of the PIMLMS model is 2.857 (E-04) and the RMSE is 3.746 (E-04). Compared with the closer Decision Tree model, the latter has slightly higher MAE and RMSE, which are 2.794 (E-04) and 3.941 (E-04) respectively, indicating that PIMLMS is more effective in reducing prediction errors. Compared with PIMLMS, PIML has a higher R 2 Slightly lower than PIMLMS, although it also shows a higher level of fit, and also has a slight disadvantage in MAE and RMSE. Compared with LightGBM and XGBoost models, PIMLMS and PIML show better performance. The coefficient of determination R of LightGBM and XGBoost on all datasets is 2 The values ​​are relatively low, especially on the S3 target metric, and XGBoost’s R 2 The value only reaches 0.740. This shows that the above two models have a large deviation from the true value and the prediction accuracy is limited. The prediction accuracy of AdaBoost is between LightGBM and XGBoost. Compared with data-driven LightGBM, R 2 The accuracy increased by 7.93%-8.71% (8.41% on average) and the RMSE decreased by 40.30%-55.87% (50.35% on average). Overall, the proposed PIMLMS model showed strong predictive ability in all the performance indicators examined.

[0093] Table 6

[0094]

[0095] (2) The MOO algorithm effectively controlled the ground settlement of the building. The obtained solution set was optimized using the TOPSIS method. The calculation results of combining PIMLMS with different MOO models are shown in Table 7. The five multi-objective optimization algorithms are adaptive weighted genetic algorithm (AWGA), non-dominated sorting genetic algorithm III (NSGA-III), multi-objective evolutionary algorithm based on decomposition (MOEA / D) and reference vector guided evolutionary algorithm (RVEA). The experimental results show that the NSGA-II algorithm presents the most significant optimization effect under all working conditions. Especially in S3, NSGA-II achieved the highest optimization rate of 40.90%. In contrast, NSGA-III and RVEA algorithms performed poorly in S1 working condition, with optimization rates of only 14.07% and 13.11% respectively, which are the lowest levels among all algorithms. The optimization effect of NSGA-II in S1 is also significant, with an optimization rate of 30.78%. In contrast, the optimization effect of S2 is relatively small, with only a decrease of 7.20%. Although its standard deviation is significantly reduced (by 62.3%). This improvement shows that the optimized S2 exhibits higher stability. These results fully demonstrate the effectiveness of the optimization algorithm in controlling building settlement and provide an important reference for subsequent engineering practice.

[0096] Table 7

[0097]

[0098] Note: μ represents the mean, and σ represents the standard deviation.

[0099] The present invention proposes a data-physical dual-driven TBM excavation-induced building complex settlement control system, which adopts the above-mentioned data-physical dual-driven TBM excavation-induced building complex settlement control method.

[0100] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention may be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present invention is defined by the appended claims, not the foregoing description, and is intended to encompass all variations within the meaning and scope of the accompanying claims. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

Claims

1. A data- and physics-driven method for controlling building settlement induced by TBM excavation, characterized by: This method proposes a settlement calculation formula that takes into account complex geological conditions. The formula is embedded into machine learning as a physical loss function to form a physical information machine learning PIMLMS model that takes into account complex geological conditions. This model can reasonably and reliably simulate building settlement caused by TBM construction under complex geological conditions. The physical information machine learning (PIMLMS) algorithm that considers complex strata is integrated into the multi-objective optimization algorithm (MOO) to form a hybrid algorithm of physical information machine learning and multi-objective optimization algorithm. A settlement control system based on dynamic adjustment of TBM parameters is established to reduce the impact of TBM construction on building settlement by controlling TBM parameters.

2. The control method according to claim 1, wherein: The hybrid algorithm aims to minimize the building settlement caused by TBM excavation; Firstly, the building settlement prediction model was established using the PIMLMS model; Then, the prediction model is introduced as the objective function, and a multi-objective optimization model based on the non-dominated sorting genetic algorithm IINSGA-II and TOPSIS is established; Finally, the multi-objective optimization algorithm MOO iteratively generates the optimal result under the control of TBM operating parameters.

3. The control method according to claim 2, characterized in that: The specific steps are as follows: Step 101, deriving a calculation formula for building settlement caused by excavation of a parallel double-track tunnel in a complex stratum; Step 102: embed the settlement calculation formula into machine learning as a physical loss function to form a data-physics dual-driven settlement prediction model; Step 103: Use model-based sequential optimization to fine-tune model hyperparameters and establish performance evaluation indicators for the prediction model; Step 104: introduce the established building settlement prediction model as the objective function into the multi-objective optimization algorithm MOO, take the settlement value of the building complex as the optimization target, and establish a multi-objective optimization model; Step 105 , by controlling TBM parameters, generating the Pareto front of the multi-objective model, and using the TOPSIS method to generate the optimal solution, ultimately achieving the goal of minimizing the settlement of the building complex caused by TBM excavation.

4. The control method according to claim 3, wherein: In step 101, the settlement curve caused by tunnel excavation is roughly similar to the normal distribution curve. For the excavated tunnel, the maximum settlement value occurs on the tunnel centerline. The two settlement curves are superimposed to form the settlement curve of the double-track tunnel. The calculation formula is: Where: S(x,y,z) is the settlement at the coordinate (x,y,z); x is the coordinate along the tunnel excavation direction; y is the horizontal coordinate perpendicular to the tunnel excavation direction; z is the vertical coordinate perpendicular to the tunnel excavation direction; L is the distance between the tunnel axes; S max is the maximum settlement caused by single-line tunnel excavation, and the calculation formula is: Where i is the width coefficient of the sedimentation tank: V S is the soil loss rate per unit length of the tunnel: V S =ηπR 2 (4) Where H is the depth of the tunnel axis from the ground; is the friction angle of the soil; η is the soil loss rate; R is the radius of the tunnel; By combining the above formulas, the calculation formula for the settlement value at point (x, y, z) is:

5. The control method according to claim 4, characterized in that: In step 101, the settlement calculation method considering complex strata is given: Among them, S all (x,y,z) represents the total settlement value at point (x,y,z); n is the total number of strata at point (x,y,z); S i (x, y, z) represents the settlement value of the i-th layer at point (x, y, z), and its calculation formula is: Among them, z i is the thickness of the i-th layer, z all is the total thickness of the soil; by substituting formula (5) and formula (7) into formula (6), the sum of the settlement values ​​of each soil layer at point (x, y, z) can be obtained, as shown in formula (8):

6. The control method according to claim 3, wherein: In step 102, building settlement is accurately predicted based on TBM parameters, actual soil layer data, and physical laws based on equation settlement formulas; a settlement prediction model driven by hybrid physical information is constructed; LightGBM is a gradient enhancement framework based on decision trees, the core idea of ​​which is to improve model performance by iteratively optimizing the loss function; physical conditions are introduced by customizing the loss function; the first and second derivatives of the loss function are used to guide the update of model parameters; LightGBM relies on the first-order derivative for optimization and combines the information of the second-order derivative to accurately describe the curvature of the loss function, making the model update more stable and efficient; The sedimentation physics formula is introduced into LightGBM as a physical loss function to impose certain physical constraints on the process of model training and extracting data associations. In this method, subset A contains data samples with large gradients, whose gradient values ​​account for a×100%; subset B consists of the remaining low-gradient samples A. S The random sampling composition is b×|A S |, the gradient value accounts for (1-a)×100%; Then, the algorithm calculates the estimated variance gain of feature j in the subset A∪B using formula (9): And based on this, the decision of feature splitting point is made: Among them, A l =[x i ∈A:x ij <d],A r =[x i ∈A:x ij >d]; B l =[x i ∈B:x ij <d];B r =[x i ∈B:x ij >d],g i is the negative gradient of the loss function.

7. The control method according to claim 6, characterized in that: In step 102, these constraints are used to build a more accurate and scientific machine learning model; the custom loss function of the model is defined as formula (10): The mean square error (MSE) is used as the data-driven component of the loss function, as shown in formula (11): Similarly, the definition of physical loss can be expressed by formula (12): Where N is the total number of test sets; y i is the label value; y phy,i is the predicted value calculated by the physical equation; Represents the predicted value obtained by the data-driven method; Loss phy Represents the residual between the physical formula value and the label; Loss pre Represents the residual between the model loss function value and the ground truth value; λ represents the proportion of physical loss to total loss; by adjusting the value, we can obtain loss functions with different deviations; The first-order and second-order derivatives of the loss function are shown in formulas (13) and (14):

8. The control method according to claim 3, wherein: In step 103, three evaluation indicators are used to measure the performance of the prediction model, namely, mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R) 2 ; The closer the MAE and RMSE values ​​are to zero, the closer the model's predicted value is to the actual settlement value; when R 2 The closer the value is to 1, the higher the correlation between the settlement value and other input parameters; MAE, RMSE, and R 2 The calculation of is shown in formulas (15) to (17): Where N represents the total number of test samples, f i (x) is the predicted value of the model, y i is the true value of the sample; in the PIMLMS model training process, the selection of hyperparameters has a decisive influence on the model performance; LightGBM is an efficient gradient enhancement framework, and its results are very sensitive to the configuration of hyperparameters, which requires the present invention to perform fine-grained tuning to obtain the best performance; the present invention adopts the model-based sequential optimization SMBO technology and realizes hyperparameter optimization through the Parson estimator tree TPE algorithm; wherein, the optimization objective function adopts the expected improvement EI criterion, and its mathematical expression is shown in formula (18): in represents the expected improvement relative to the current best objective value. By maximizing this value, new sampling points can be selected to find a better hyperparameter configuration; * is the lowest observed objective function value, which is used as a benchmark to determine whether a new sampling point will produce better results; x is the current sampling point being evaluated, which represents a specific set of hyperparameters; y is the possible value x of the objective function at the sampling point; since Bayesian optimization predicts the function f(x) based on a probability model, the observed value y is considered a random variable; p(y|x) represents the probability density function of the objective function value y at a given point x, which quantifies the likelihood of y taking different values ​​when the model predicts x.

9. The control method according to claim 3, wherein: In step 104, the multi-objective optimization algorithm MOO for tunnel-induced building settlement uses the settlement values ​​of three buildings along the TBM excavation route as three objective variables and selects six operating parameters of the TBM as independent variables. The MOO model can be expressed by the following formula: minF(x i )=[f1(x i ),f2(x i ),f3(x i )](i=1,2,…,6) (19) The constraints are as follows: g j (x i )≤0(j=1,2,3) (20) Where F(x i ) is based on three target variables f1(x i ),f2(x i ),f3(x i ) to construct a multi-objective optimization function; g j (x) is an inequality constraint; and and are respectively x i The lower and upper limits of the value.

10. The control method according to claim 3, characterized in that: In step 105, the Pareto front of the MOO algorithm is generated using the NSGA-II algorithm based on the geatpy library. NSGA-II is an evolutionary algorithm for multi-objective optimization. The process begins by generating an initial population of size P, where each individual represents a potential solution. Individuals are sorted into different Pareto fronts through non-dominated sorting. Within each front, the crowding distance CD is calculated to assess the diversity of solutions. Selection is then performed based on the Pareto front and crowding distance, with lower fronts and higher CD values ​​being prioritized. The selected individuals undergo genetic manipulation to create a new offspring population. The current population is merged with the offspring, and the process of non-dominated sorting and crowding distance calculation is repeated. This iterative process will continue until the algorithm reaches the termination criterion, and eventually generate an approximate Pareto front that has both optimality and diversity balance characteristics; To filter the optimal solution from the Pareto front, the TOPSIS method is used for analysis; the TOPSIS method ranks the candidate solutions based on their Euclidean distance to the ideal solution and the most negative solution; a score for each solution is calculated using these distances, and the solution with the highest score is considered the optimal solution in the Pareto front.

11. A data- and physics-driven TBM excavation-induced building settlement control system, characterized by: The control system adopts the control method described in any one of claims 1-10.

Images