Tunnel excavation ground surface settlement prediction method and system based on machine learning hybrid model

Abstract

The invention provides a tunnel excavation ground surface settlement prediction method and system based on a machine learning hybrid model, and relates to the technical field of tunnel engineering and machine learning crossing, and the method comprises the steps: obtaining the multi-source heterogeneous information of a target tunnel, and constructing a ground surface settlement data set; a Transform-BiLSTM hybrid model is constructed, the robustness of the algorithm in a noise environment is enhanced based on a VMD (variational mode decomposition) algorithm, hyper-parameters are adaptively adjusted and optimized by using a PSO (particle swarm optimization) algorithm based on a ground surface settlement data set, the model prediction precision is maximized, and a ground surface settlement prediction model is obtained; and analyzing decision logic of the ground surface settlement prediction model through the SHAP value, and outputting interpretable engineering guidance suggestions. By constructing a machine learning hybrid model, high-precision and real-time prediction of ground surface settlement in the whole process of tunnel excavation is realized. The precision and generalization ability of the model are improved, the characterization ability of complex spatial-temporal characteristics is enhanced, and overfitting is avoided; and the interpretability is optimized, and the influence of key parameters on a prediction result is quantified, so that construction parameter adjustment is guided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel engineering and machine learning, and particularly relates to a tunnel excavation ground surface settlement prediction method and system based on a machine learning hybrid model. BACKGROUND

[0002] The shield method has high safety, good automation and fast tunneling speed, and becomes the preferred method for urban subway tunnel construction. However, the shield tunnel construction process inevitably disturbs the surrounding stratum, causing the soil to bear complex mechanical behaviors such as extrusion, shear and twisting, changing the original stress balance of the stratum, causing stress release and redistribution of the stratum, and causing the ground surface to deform, causing varying degrees of damage.

[0003] The prior art has the following main defects in ground surface settlement prediction: The traditional empirical formula method relies on idealized assumptions and cannot reflect the asymmetric settlement characteristics in actual engineering; the numerical simulation method is complex to model and has low calculation efficiency, and cannot meet the real-time prediction requirements; a single machine learning model is sensitive to small sample data and prone to overfitting, and the provided data samples are limited, and the reliability and generalization ability in actual application are insufficient. The existing technology uses linear interpolation to complete, but cannot effectively handle nonlinear time series characteristics; the fusion ability of multi-source heterogeneous data such as geological parameters, construction parameters and monitoring data is insufficient. The model has poor interpretability, and the existing prediction model is a "black box" and cannot quantify the contribution of each parameter, making it difficult to guide engineering decision-making.

[0004] Therefore, there is an urgent need for a new ground surface settlement prediction method to solve the above technical problems. SUMMARY

[0005] The present application provides a tunnel excavation ground surface settlement prediction method and system based on a machine learning hybrid model, which integrates multi-source heterogeneous data and adaptively captures the spatio-temporal evolution law to at least partially solve the above problems.

[0006] The first aspect of the present application provides a tunnel excavation ground surface settlement prediction method based on a machine learning hybrid model, the method comprising: obtaining multi-source heterogeneous information of a target tunnel, constructing a ground surface settlement dataset, the multi-source heterogeneous information including stratum parameters, shield parameters, geometric parameters and ground surface settlement monitoring data; constructing a Transformer-BiLSTM hybrid model, adaptively tuning hyperparameters based on the ground surface settlement dataset using a PSO algorithm, maximizing model prediction accuracy, and obtaining a ground surface settlement prediction model; analyzing the decision logic of the ground surface settlement prediction model through SHAP values, and outputting an interpretable engineering guidance suggestion.

[0007] Optionally, multi-source heterogeneous information of the target tunnel is acquired, and a ground settlement dataset is constructed, including: In the case that the target tunnel has a missing section of ground settlement monitoring, a three-dimensional shield tunnel model corresponding to the target tunnel is established based on finite element analysis software, the whole construction process of the target tunnel is simulated, the ground mechanics parameters and the tunnel depth are adjusted to match the construction conditions of the missing section of monitoring, and after consistency verification of the simulation results and the field measured data, the missing data of ground settlement monitoring is completed based on the simulation results.

[0008] Optionally, multi-source heterogeneous information of the target tunnel is acquired, and a ground settlement dataset is constructed, including: Outliers of each parameter in the ground settlement dataset are removed through a box plot, and correlation coefficient analysis is performed on the processed dataset; Based on the correlation coefficient analysis results, a linear interpolation method is used to extend the time series of the ground settlement dataset, and the effectiveness of the extended ground settlement dataset is determined by comparing and verifying the scatter pair diagonal matrix diagram and the box plot, and the heat map results.

[0009] Optionally, a Transformer-BiLSTM hybrid model is constructed, and the PSO algorithm is used to adaptively optimize the hyperparameters based on the ground settlement dataset, to maximize the model prediction accuracy, and to obtain a ground settlement prediction model, including: The sequence information in the input ground settlement dataset is encoded through the encoder part of the Transformer, the multi-head attention mechanism is used to calculate the dependency relationship between different monitoring points or different time points in parallel, the query, key, and value matrices are calculated to assign different attention weights to each monitoring point or time point, and the feedforward neural network performs nonlinear transformation on the output of the attention mechanism to enhance the model's representation ability of ground settlement features, so that the model can learn complex settlement patterns; The feature sequence output by the Transformer encoder is input into the BiLSTM module, the ground settlement sequence information is processed through the gating mechanism of the traditional recurrent neural network, and the ground settlement prediction value is output.

[0010] Optionally, a Transformer-BiLSTM hybrid model is constructed, and the PSO algorithm is used to adaptively optimize the hyperparameters based on the ground settlement dataset, to maximize the model prediction accuracy, and to obtain a ground settlement prediction model, including: The ground settlement dataset is input into the VMD algorithm, and the original signal is adaptively decomposed into multiple intrinsic mode functions by iteratively solving the variational optimization problem; According to the modal frequency characteristics, high-frequency noise components are identified and removed, and low-frequency effective signals reflecting the long-term deformation law of the stratum are retained. The intrinsic mode function components after screening are superimposed to reconstruct the signal and generate the denoised ground subsidence time series data as the model input data.

[0011] Optionally, a Transformer-BiLSTM hybrid model is constructed, and PSO algorithm is used to adaptively optimize the hyperparameters based on the ground subsidence data set, so that the model prediction accuracy is maximized, and a ground subsidence prediction model is obtained. A parameter space is defined, and the search range of the key hyperparameters of the model is set. The root mean square error (RMSE) of the test set is used as the optimization target to quantify the prediction performance of the model and evaluate the fitness of the model. The particle swarm is initialized, and the hyperparameter combination is randomly generated; the parameter space is explored through the particle velocity and position updating mechanism; the search direction is dynamically adjusted according to the fitness value to gradually approach the global optimal solution; and the optimal hyperparameter combination is output after convergence for the final training of the model.

[0012] The second aspect of the present application provides a tunnel excavation ground subsidence prediction system based on a machine learning hybrid model, which comprises: A data acquisition module is configured to acquire multi-source heterogeneous information of a target tunnel, construct a ground subsidence data set, and the multi-source heterogeneous information comprises stratum parameters, shield parameters, geometric parameters and ground subsidence monitoring data. A model training module is configured to construct a Transformer-BiLSTM hybrid model, use PSO algorithm to adaptively optimize the hyperparameters based on the ground subsidence data set, maximize the model prediction accuracy, and obtain a ground subsidence prediction model. A suggestion module is configured to analyze the decision logic of the ground subsidence prediction model through SHAP value and output an interpretable engineering guidance suggestion.

[0013] The third aspect of the present application provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the tunnel excavation ground subsidence prediction method based on the machine learning hybrid model as described in the first aspect of the present application.

[0014] The fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the tunnel excavation ground subsidence prediction method based on the machine learning hybrid model as described in the first aspect of the present application.

[0015] The fifth aspect of the present application provides a computer program product comprising computer programs / instructions for implementing the steps of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model according to the first aspect of the present application by a processor.

[0016] In the embodiments of the present application, the missing monitoring data is completed by a numerical simulation method, outliers are removed by a box plot, and the data set is expanded by a linear interpolation method, which significantly improves the data quality and provides a reliable data basis for model training.

[0017] The (VMD) PSO-Transformer-BiLSTM hybrid model constructed in the embodiments of the present application performs well on the training set, the validation set and the test set, and significantly improves the prediction accuracy. The hybrid model combines the advantages of Transformer and BiLSTM, can better capture the time sequence characteristics and long-term dependence of the data, and enhances the generalization ability of the model.

[0018] The prediction method provided by the embodiments of the present application has high computing efficiency and can realize real-time prediction to provide timely and effective early warning for tunnel engineering construction safety. By introducing advanced machine learning algorithms and hyperparameter optimization strategies, the method of the present application can better adapt to complex and variable engineering conditions, improve model adaptability, and realize accurate prediction. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the description of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 is a step flow chart of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the present application; Figure 2 is a detailed flowchart of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the present application; Figure 3 is a shield tunnel finite element model corresponding to an exemplary case of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the present application; Figure 4 is a brief schematic diagram of the data processing process of the Transformer-BiLSTM hybrid model of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the present application; Figure 5is a comparison result of model prediction value and measured value of an exemplary case of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the application; Figure 6 is a SHAP scatter plot of input feature parameters of an exemplary case of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the application. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below in combination with the drawings and specific embodiments.

[0022] As shown in Figure 1 , it shows a step flow chart of a tunnel excavation ground surface settlement prediction method based on a machine learning hybrid model provided by an embodiment of the application, specifically, the method comprises the following steps: S101, obtaining multi-source heterogeneous information of a target tunnel, and constructing a ground surface settlement data set.

[0023] Among them, the multi-source heterogeneous information includes stratum parameters, shield parameters, geometric parameters and ground surface settlement monitoring data.

[0024] In the embodiment of the application, multi-source data such as stratum parameters, shield construction parameters, ground surface settlement monitoring data and numerical simulation results can be integrated to construct a double-line shield tunnel ground surface settlement data set. Outliers of each parameter in the data set are identified and replaced by box plots. The correlation between each parameter and the ground surface settlement is quantified by Pearson correlation, and two parameters with weak correlation are removed. Based on this, the data set is expanded by using linear interpolation method, and the distribution form of the data is verified by comparing the corresponding diagonal matrix graph with the box plot, and the correlation of the parameters is verified by comparing the Pearson thermodynamic diagram, and the effectiveness of the data expansion method is confirmed.

[0025] In the embodiment of the application, the stratum parameters include 8 parameters of compression modulus, cohesion, internal friction angle, natural water content, specific gravity, void ratio, compressive strength and lateral pressure coefficient, which are all derived from the engineering geological investigation report. The shield parameters include 6 construction parameters of upper soil bin pressure, grouting amount, cutter head speed, advancing speed, cutter head torque and total thrust, which are all collected from the real-time monitoring automatic collection system of the shield machine; the geometric parameters include tunnel burial depth and double-line tunnel spacing; the monitoring data includes measured ground surface settlement value and pore water pressure, which are derived from the field monitoring double-line tunnel center line ground surface settlement value. In the case of data missing, the monitoring data can also include numerical simulation completion data.

[0026] In the embodiment of the application, after the above-mentioned parameters are standardized, a 13-dimensional feature vector is formed, the output target is the ground surface settlement value at the center line of the shield tunnel, and the ground surface settlement data set is constructed.

[0027] In the embodiment of the present application, during the monitoring data collection process, some input parameters have data missing problems due to factors such as on-site construction operation and equipment failure. Based on this, the embodiment of the present application proposes: In the case of missing sections of surface settlement monitoring in the target tunnel, a three-dimensional shield tunnel model corresponding to the target tunnel is established based on finite element analysis software, the whole construction process of the target tunnel is simulated, the stratum mechanical parameters and tunnel burial depth are adjusted to match the construction conditions of the missing sections of monitoring, and after consistency verification of the simulation results and the field measured data, the missing data of surface settlement monitoring is completed based on the simulation results.

[0028] Specifically, as shown in Figure 2 , a detailed flowchart of the tunnel excavation surface settlement prediction method based on the machine learning hybrid model provided by the embodiment of the present application is shown. Figure 2

[0029] In the embodiment of the present application, a three-dimensional numerical model can be established based on actual engineering conditions to analyze the surface deformation characteristics. Further based on the comparison between the simulation data and the measured data, the spatial distribution characteristics and the time development characteristics of the surface settlement data are verified to verify the three-dimensional numerical model. After verifying the reliability of the three-dimensional numerical model, the missing values of the settlement monitoring can be completed based on the numerical simulation data, and the effectiveness of the completed model is further verified. After verification, the missing values of the surface settlement are completed based on the simulation data.

[0030] Specifically, in the embodiment of the present application, based on the finite element analysis software, the whole construction process of the tunnel is simulated, the stratum mechanical characteristics and the construction dynamic parameters are considered, and a three-dimensional shield tunnel model is established. The simulation results are compared with the field measured data to ensure that the settlement error of key monitoring points such as the tunnel axis and the surface center line is less than 15%, and the engineering reliability of the model is verified. For the missing sections of monitoring caused by construction disturbance, the model parameters are adjusted to generate numerical simulation data matching the working conditions of the missing sections, and the missing surface settlement values of the sections are completed.

[0031] Specifically, a three-dimensional finite element model (size 120 m x 80 m x 76 m) can be established by ABAQUS, as shown in Figure 3 ​As shown, it shows a shield tunnel finite element model corresponding to an exemplary case, first, the consistency of the known point data is verified for the basic model, then the stratum mechanics parameters and the tunnel depth are adjusted, the numerical model of the missing section is constructed, and the missing data of the ground settlement monitoring is completed. The consistency verification can include: verifying whether the numerical simulation results and the measured data of the known monitoring points are in good consistency in terms of settlement curve shape and numerical size, to verify the reliability of the numerical model for simulating the missing section. It can also include: verifying whether the numerical simulation results and the measured data present consistent time variation trends. Specifically, the settlement development process of each section monitoring point shows similar stage characteristics, which can be mainly divided into three stages of uplift, rapid settlement and final stability, to further verify the reliability of the numerical model for simulating the missing monitoring section.

[0032] In an exemplary case in the embodiment of the application, the feasibility of the model is verified by comparing and analyzing the measured data and the numerical data. The numerical results show that the ground settlement caused by the construction of the double-line shield tunnel has typical spatio-temporal evolution characteristics. In terms of spatial distribution, the settlement trough presents a significant "W" type double-peak feature, and the maximum settlement value appears in the tunnel center line area. In terms of time evolution, the settlement process presents a three-stage development law of "uplift-settlement-stability", and the ground deformation of different positions presents significant difference characteristics. At the left line tunnel axis ground: the ground appears slight uplift before the shield arrives, rapidly settles when the shield passes, and the settlement gradually converges after the shield tail passes; when the right line is constructed, the measuring point again presents similar three-stage deformation characteristics, and has an inhibitory effect on the overall settlement development; at the center ground settlement of the double-line tunnel: affected by the construction of the left and right lines, the settlement development presents a two-time "uplift-settlement-stability" development law; at the right line tunnel axis ground: affected by the construction of the left line, it first appears uplift, and then presents a typical three-stage deformation process of "uplift-settlement-stability" after the right line is connected. The numerical simulation section settlement and the settlement development curve have high fitting degree with the measured values.

[0033] In actual engineering, the collected ground settlement data sample size is small, and the monitoring missing point distribution is relatively concentrated, which seriously affects the integrity of the small sample data. Therefore, based on the verified basic model, the stratum mechanics parameters, tunnel depth and other key parameters of the corresponding section are adjusted to establish a numerical model consistent with the actual engineering conditions, and the simulation results are verified with the measured data of some intact known monitoring points to ensure the reliability of the completed data.

[0034] In the embodiment of the application, the collected ground settlement data set can also be preprocessed, specifically including: S1011, outliers of each parameter in the ground settlement data set are removed by box plot, and the processed data set is analyzed for correlation coefficient to remove weakly correlated parameters; S1012, the linear interpolation method is used for time series expansion of the ground settlement data set, and the scatter point diagonal matrix diagram and box type diagram and the thermal map result of the expanded ground settlement data set are compared and verified to determine the effectiveness of the expanded data.

[0035] Specifically, outliers can be identified using the box plot method, where Q1, Q2, and Q3 are quartiles that describe the central distribution of the data, and IQR = Q3-Q1 is the interquartile range, which can be used to measure the dispersion of the data. The upper and lower limits of the effective value are defined as Q1-1.5IQR and Q3+1.5IQR, and the values outside the range are replaced with boundary values. The shield parameters, stratum parameters and geometric parameters are corrected to ensure that the data distribution conforms to the actual working conditions.

[0036] Outliers are abnormal observations in the database that deviate significantly from the main distribution. As an effective data visualization tool, the box plot can intuitively display the distribution characteristics of the data and effectively identify outliers.

[0037] In the machine learning training process, outliers will significantly affect the prediction performance of the model. These extreme value samples will force the model parameters to adjust excessively to adapt to extreme features, resulting in a higher fitting effect of the training set and a damaged generalization ability, which will show obvious overfitting on the test set. In addition, outliers will interfere with the estimation of the true distribution of the model, distort the inherent relevance between variables, destroy the original rules and structural characteristics of the data, and weaken the ability of the prediction model to capture normal data features. Therefore, in the embodiment of the present application, the outliers identified by the box plot-based outlier processing method are replaced by the upper and lower boundary values of the corresponding variables (Q1-1.5IQR and Q3+1.5IQR).

[0038] Pearson correlation coefficient is a classic statistical indicator for measuring the linear correlation between two continuous variables. Therefore, in the embodiment of the present application, the Pearson correlation coefficient can be calculated to obtain all parameter heat maps. Parameters with low absolute value of correlation coefficient, such as lateral pressure coefficient (-0.13) and compressive strength (-0.11), have little effect on ground settlement, and weakly correlated parameters are excluded. Significant correlation parameters with high correlation coefficient, such as thrust speed (0.55), water content (0.58), void ratio (0.67), grouting amount (-0.44), and soil bin pressure (-0.47), are retained.

[0039] In the embodiment of the present application, the data set is expanded using linear interpolation method. For time series data, it is assumed that there are monitoring data at time points and , and and , and A new time point t is inserted between t and t, and the continuous construction parameters are interpolated with a time step Δt = 0.5h. The corresponding interpolation data can be calculated by a linear formula .

[0040] In this way, the original data is expanded, and the scatter plot matrix and box plot are drawn to verify the consistency of the distribution of the expanded data and the original data.

[0041] In the process of machine learning modeling, the size of the data will directly affect the prediction performance of the model, and small sample data sets are prone to overfitting problems. Limited by the field monitoring conditions, only a small amount of cross-section monitoring data may be collected in actual engineering. In order to fully reflect the complex nonlinear relationship between each parameter and ground settlement, the linear interpolation method is used to expand the original data in the embodiment of the application. The linear interpolation method has the advantages of high calculation efficiency, simple implementation, strong result interpretability, etc., can maintain the local change trend of the data and generate smooth transition values, effectively solve the small sample data problem, and provide a more comprehensive data basis for machine learning model training.

[0042] The result visualization of the scatter plot matrix can reveal the distribution of the parameters and the relationship between the parameters. In the diagonal position, the column chart of each parameter intuitively shows the data distribution characteristics; in the non-diagonal line area, the scatter plot and the red fitting curve show the correlation pattern between variables. When the fitting curve shows a left-down to right-up trend, there is a positive correlation between the variables; while a left-up to right-down trend reflects a negative correlation; if the scatter points are randomly distributed and the fitting curve is approximately horizontal, it indicates that the correlation between the variables is weak.

[0043] By comparing the scatter plot matrix and box plot of the ground settlement data set after data expansion with the scatter plot matrix and box plot of the original data set, and comparing the heat map results, the effectiveness of the linear interpolation method in maintaining the original data structure can be verified, as well as the reliability of the data expansion results.

[0044] In the embodiment of the present application, by integrating geological survey, construction parameters, settlement monitoring and numerical simulation results and other multi-source data, a machine learning data set (surface settlement data set) is constructed with the surface settlement at the center axis of the double tunnel as the prediction target. In an exemplary case, through analysis of the data composition of geometric parameters, shield parameters and stratum parameters, it is identified that the shield machine has abnormal working conditions such as idling and shutdown, and the engineering stratum is characterized by alternating distribution of silty clay and sand. The analysis results are consistent with the actual engineering situation. Based on the correlation coefficient analysis, it is shown that the advance speed, natural moisture content and void ratio are significantly positively correlated with the surface settlement, and the grouting amount, upper soil chamber pressure, internal friction angle and cohesion show significant negative correlation, and accordingly the side pressure coefficient and compressive strength parameters with weak correlation are excluded. After expanding the original data using linear interpolation method, through comparison and verification of parameter scatter pair diagonal matrix, box plot and Pearson heat map, it is confirmed that the data expansion retains the statistical properties and structural characteristics of the original data.

[0045] S102, a Transformer-BiLSTM hybrid model is constructed, and the PSO algorithm is used to adaptively optimize the hyperparameters based on the surface settlement data set, to maximize the prediction accuracy of the model, and obtain a surface settlement prediction model.

[0046] In the embodiment of the present application, the Transformer-BiLSTM hybrid model is composed of a Transformer encoder layer and a BiLSTM network layer, wherein the number of Transformer encoder layers is 2, the network dimension is not less than 128, the number of multi-head attention heads is not less than 32, the key-value dimension is not less than 64, and the dropout rate is configured to be 0.1-0.5, to increase the generalization ability of the model. The feedforward network is set to two fully connected layers, with 512 hidden units and ReLU activation function. Then residual connection and layer normalization are performed to accelerate the training convergence of the model, reduce problems such as gradient vanishing or explosion, and make the model more stable.

[0047] The BiLSTM is a double-layer structure, the forward and reverse LSTM outputs are spliced, and the double-layer structure is configured, with the number of hidden neurons in each layer not less than 64 and the dropout rate of 0.1-0.3; the output layer uses a linear activation function to adapt to the regression task.

[0048] Specifically, Figure 4A brief schematic diagram of the data processing process of the Transformer-BiLSTM hybrid model is shown, specifically, the input sequence is first encoded by the encoder part of the Transformer, the multi-head attention mechanism calculates the dependency between different monitoring points or different time points in parallel, and the feedforward neural network further nonlinearly transforms the output of the attention mechanism to enhance the model's representation ability of the ground settlement characteristics, so that the model can learn complex settlement patterns. Subsequently, the feature sequence output by the Transformer encoder is input into the BiLSTM module as the input of the recurrent neural network module BiLSTM, and the ground settlement sequence information is processed through the gating mechanism of the traditional recurrent neural network to output the ground settlement prediction value.

[0049] In the embodiment of the application, the Transformer encoder adopts a multi-layer self-attention mechanism to capture the global correlation between the construction parameters and the settlement values, especially the nonlinear relationship that cannot be obtained by traditional methods, and the position encoding technology is used to retain the time sequence information, solving the problem that the traditional attention mechanism is not sensitive to the order. The bidirectional long short-term memory network (BiLSTM) extracts time sequence features from the forward and reverse directions respectively, establishes the dynamic evolution law of phenomena such as the cumulative disturbance effect of shield tunneling on the stratum during the construction process, and introduces a dropout layer (Dropout) to suppress overfitting and improve the generalization ability of the model.

[0050] In the output layer, the time sequence features output by the BiLSTM are mapped to the settlement prediction value, and the activation function uses a linear function to adapt to the requirements of the regression task.

[0051] In the embodiment of the application, the Transformer model has significant advantages in shield tunnel ground settlement prediction due to its unique self-attention mechanism. The self-attention mechanism of the Transformer model maps the settlement data and related parameters in the time sequence into three vector spaces: query (Query), key (Key) and value (Value), dynamically weights the value based on the similarity between the query and the key, and efficiently captures long-distance dependencies. This design not only allows the model to focus directly on data at any time in the entire time sequence, but also allows the calculation process to be parallelized, greatly improving the efficiency and accuracy of the prediction. In addition, the model can also calculate the attention of multiple different representation subspaces in parallel through the multi-head attention mechanism, further enriching the model's understanding of data and more comprehensively mining the deep information contained in the data. At the same time, the application of layer normalization and residual connection makes the model training more stable and efficient.

[0052] From the data and task characteristics, the ground settlement data caused by shield tunnel construction as typical time series data has significant autoregressive characteristics, that is, the current settlement value is closely related to the past data and related factors. Therefore, in the embodiment of the present application, only the encoder layer of the Transformer is selected. First, the autoregressive method of the encoder layer is used to input the historical prediction results and related information, which conforms to the internal time series law of the ground settlement data and can more targetedly solve the prediction problem. Secondly, from the efficiency and stability of model training, only using the encoder layer can greatly reduce the number of parameters and the amount of calculation of the model, reduce the memory occupation, significantly improve the training efficiency and shorten the training time when processing the settlement data. Then, only using the encoder layer avoids the complex interaction and coordination problem of the encoder and the decoder, effectively enhances the stability of the training process, and is more easily converged to the ideal result when facing the noise and uncertainty that may exist in the settlement data. In addition, the encoder layer has unique advantages in information fusion and prediction performance improvement. In the ground settlement prediction, prior knowledge and auxiliary data such as geometric parameters, geological parameters and tunneling parameters are crucial, and the encoder layer can integrate these external information into the model, add related auxiliary features in each time step input, better capture the relationship between them and the settlement, and improve the prediction accuracy. The self-attention mechanism of the encoder layer can also strengthen the modeling ability of the local dependence relationship of the ground settlement data, pay more attention to the recent data, and improve the short-term prediction accuracy. At the same time, its relatively simple structure enhances the model interpretability, and the key data and features that the model focuses on during prediction can be intuitively understood, providing more valuable reference basis for actual engineering decision-making.

[0053] The BiLSTM (Bidirectional Long Short-Term Memory) model is an extension based on the LSTM model, which significantly improves the time series modeling capability through the bidirectional processing mechanism. The model not only retains the advantages of the LSTM in processing time series data through the gating mechanism, but also can process time series from both forward and reverse directions. The forward layer processes the input sequence in time order, gradually accumulating historical information; the reverse layer processes the sequence in reverse, effectively extracting future context features. The hidden states of the two LSTM layers capture data feature information from two directions respectively, and then they are spliced at each time step as the final output. When processing settlement data prediction of time series tasks, the bidirectional processing mechanism enables the model to make full use of past and future information, effectively capturing long-distance dependence relationships.

[0054] The BiLSTM model can not only accurately capture long-term trends and short-term fluctuations in the data like the LSTM, filter out irrelevant information and noise, improve the accuracy and stability of the prediction, but also can more comprehensively mine the potential law in the data through the bidirectional processing mechanism, and provide strong support for accurately predicting the ground settlement.

[0055] In the embodiment of the application, before the training of the Transformer-BiLSTM hybrid model, the ground settlement data set can also be optimized, specifically including: S11, inputting the ground settlement data set into the VMD algorithm, and adaptively decomposing the original signal into a plurality of intrinsic mode functions by iteratively solving the variational optimization problem.

[0056] S12, according to the modal frequency characteristics, identifying and removing the high-frequency noise component, and retaining the low-frequency effective signal reflecting the long-term deformation law of the stratum.

[0057] S13, superimposing the screened intrinsic mode function components to reconstruct the signal and generate the denoised ground settlement time series data as the model input data.

[0058] In the embodiment of the application, the VMD variational mode decomposition algorithm adaptively decomposes the complex non-stationary signal x(t) into K intrinsic mode functions (IMF) with a certain center frequency by constructing a constrained variational problem, effectively avoids the modal aliasing problem through strict frequency domain constraint conditions, and enhances the robustness of the algorithm in a noisy environment.

[0059] The parameter setting of the VMD algorithm is set as: the modal decomposition number K≥5, the penalty factor α≥1000, and the tolerance error ε≤1×10⁻ 6 The ground settlement time series data x(t) is decomposed into 8 IMF components, the high-frequency noise is removed, the low-frequency effective signal is retained, and the data optimization is realized.

[0060] In the embodiment of the application, in the model training process, the hyperparameter optimization method selects the PSO particle swarm algorithm, specifically, the method step S102 includes: S1021, defining a parameter space and setting the search range of the key hyperparameters of the model.

[0061] S1022, taking the root mean square error (RMSE) of the test set as the optimization target, quantifying the prediction performance of the model, and evaluating the fitness of the model.

[0062] S1023, initializing the particle swarm, randomly generating a combination of hyperparameters, exploring the parameter space through the particle velocity and position updating mechanism, dynamically adjusting the search direction according to the fitness value, gradually approaching the global optimal solution, and outputting the optimal hyperparameter combination after convergence for the final training of the model.

[0063] In the embodiment of the application, the initialized particle swarm size is not less than 30, the learning factors c1 and c2 are both set to 2, the inertia factor ω is set to 0.5-1.0, and the root mean square error of the test set is taken as the fitness function for iterative optimization.

[0064] Specifically, the particle swarm is initialized, the particle number is set to 50, and the parameter combination is randomly generated. In each iteration, the particle updates itself by tracking the two “extreme values” of pbest and gbest. The particle velocity and position are updated by using the formula:

[0065] wherein i =1,2,...,N, N is the total number of particles in the swarm; v i is the velocity of the particle; rand() is a random number between 0 and 1; x i is the current position of the particle; c 1 and c 2 are learning factors, usually c 1= c 2=2; ω is an inertia factor, which is non-negative, and the larger the value is, the stronger the global optimization ability is and the weaker the local optimization ability is.

[0066] After 42 iterations, the optimal parameters are: the position encoding dimension is 2, 64 attention heads are used, the key-value dimension is set to 128, the number of hidden layer neurons is 256, the dropout rate is set to 0.5, and the learning rate is configured to 0.001.

[0067] The optimized machine learning hybrid model has very small errors between the predicted value and the measured value of the ground surface settlement in the construction of the shield tunnel, such as Figure 5 , Figure 5 The comparison result of the predicted value and the measured value of an exemplary case model of the embodiment of the application is shown, and it can be seen that the error of the model prediction result is basically maintained within 1 mm (>90%), and only the ground surface settlement prediction result error of 4 settlement section numbers exceeds 2 mm, fully illustrating the accuracy and feasibility of the method.

[0068] In the embodiment of the application, the expanded ground surface settlement data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1. The finally constructed (VMD) PSO-Transformer-BiLSTM hybrid model exhibits excellent prediction ability on the training set, the validation set and the test set.

[0069] S103, outputting an interpretable engineering guidance suggestion by analyzing the decision logic of the ground settlement prediction model through the SHAP value.

[0070] SHAP (SHapley Additive exPlanations) is a model interpretability method based on Shapley Value in game theory. This method compares the prediction process of a machine learning model to a cooperative game problem, where input features are considered as game participants, and model prediction results are considered as game benefits. The SHAP value of each feature represents its marginal contribution in all possible feature combinations. The SHAP method quantifies the specific influence of each feature on a single prediction result by systematically evaluating its marginal contribution in all possible subsets. This method is based on strict mathematical foundations, requiring that the allocation of feature contributions must satisfy two key axioms: additivity, which ensures that the sum of all feature contributions is equal to the deviation of the model prediction from the baseline value, and consistency, which ensures that the ranking of feature importance remains consistent with the trend of model output changes. It is this strict mathematical framework based on game theory that makes the SHAP method a reliable and theoretically sound method for feature importance explanation for complex machine learning models, making it one of the most theoretically sound methods in current model interpretability research.

[0071] In the embodiments of the present application, the SHAP (Shapley Additive Explanations) framework based on the Shapley Value in game theory is used to analyze the marginal contribution of each input feature to the prediction result and quantify its importance.

[0072] The SHAP values of the output parameters are sorted to obtain the importance ranking of each parameter, and the key influencing factors are determined.

[0073] Finally, according to the analysis results, construction parameter optimization strategies can be developed to dynamically adjust the grouting amount, control the advancing speed, etc., and guide the engineering risk prevention and control.

[0074] Specifically, based on the additive explanation model of game theory, the marginal contribution of each feature to the prediction result is calculated:

[0075] where, φ i is the SHAP value of feature i , reflecting its contribution to the prediction; F is the set of all features, S is the subset without feature i .

[0076] Feature analysis of the prediction model is of great significance for understanding the mechanism of each influencing factor, improving the credibility of the model, and guiding engineering decision-making. In the embodiment of the present application, the SHAP method is used to analyze the feature importance of the (VMD) PSO-Transformer-BiLSTM model, and the training data of the PSO-Transformer-BiLSTM model comes from an exemplary case. The results of SHAP value calculation are as follows Figure 6 , Figure 6 The SHAP scatter plot of the input feature parameters of an exemplary case of the embodiment of the present application is shown, and in the analysis result, the positive and negative directions of the SHAP value reflect the promotion or inhibition effect of the feature on the prediction result, and the absolute value size represents the influence intensity. Through the combined distribution of the feature point color (blue represents a smaller value, and red represents a larger value) and the SHAP value interval, the nonlinear relationship between the feature and the prediction value can be clearly revealed.

[0077] Among all the input features of the model, the grouting amount (average SHAP value 0.48), the natural water content (0.41), and the upper soil chamber pressure (0.38) are identified as the most critical influencing factors. The high SHAP value of the grouting amount and its concentrated distribution in the positive interval indicate that increasing the grouting amount can significantly inhibit the prediction result of the ground settlement. The influence of the natural water content presents obvious asymmetry, and the low water content sample is mainly distributed in the SHAP negative value area, indicating that its inhibition effect on the prediction result of the ground settlement is weak. The influence mode of the upper soil chamber pressure is relatively complex, and both high and low values will have a significant impact on the prediction result. The medium important features include the construction parameters such as the void ratio (0.27), the advancing speed (0.22), and the cutter head speed (0.22), and the soil property parameters such as the compression modulus (0.12). It is worth noting that the influence degree of the tunnel depth (0.02) is the lowest, which is consistent with the engineering practice. In the design of the shield tunnel, a relatively conservative depth scheme is usually adopted to control the risk of ground settlement.

[0078] In the embodiment of the present application, the (VMD) PSO-Transformer-BiLSTM hybrid model adopts VMD for data preprocessing and combines the PSO algorithm to optimize the hyperparameters, and it exhibits excellent prediction performance on the training set, the validation set, and the test set. Through SHAP analysis of the influence mechanism of each feature parameter, in an exemplary case, the grouting amount, the natural water content, and the upper soil chamber pressure are identified as the most critical influencing factors, which provides a reliable theoretical basis and technical support for the prediction of ground settlement of the shield tunnel. In the embodiment of the present application, for different target tunnels, the corresponding tunnel multi-source heterogeneous information can be used to construct the ground settlement dataset and perform model training, and the most critical influencing factors for the target tunnel can be analyzed.

[0079] Based on the same inventive concept, the application also provides a tunnel excavation ground surface settlement prediction system based on a machine learning hybrid model, which comprises: A data acquisition module is configured to acquire multi-source heterogeneous information of a target tunnel, construct a ground surface settlement dataset, and the multi-source heterogeneous information comprises stratum parameters, shield parameters, geometric parameters and ground surface settlement monitoring data; A model training module is configured to construct a Transformer-BiLSTM hybrid model, use a PSO algorithm to adaptively optimize hyperparameters based on the ground surface settlement dataset, maximize model prediction accuracy, and obtain a ground surface settlement prediction model. A suggestion module is configured to analyze the decision logic of the ground surface settlement prediction model through SHAP values and output an interpretable engineering guidance suggestion.

[0080] Optionally, the data acquisition module is configured to: In the case that there is a missing section of ground surface settlement monitoring in the target tunnel, a three-dimensional shield tunnel model corresponding to the target tunnel is established based on a finite element analysis software, the construction of the whole process of the target tunnel is simulated, the stratum mechanical parameters and the tunnel burial depth are adjusted to match the construction conditions of the monitoring missing section, after the simulation results are consistent with the field measured data, the missing data of the ground surface settlement monitoring is completed based on the simulation results.

[0081] Optionally, the data acquisition module is configured to: Outliers of each parameter in the ground surface settlement dataset are removed through a box plot, and weakly correlated parameters are removed from the processed dataset through correlation coefficient analysis. The ground surface settlement dataset is expanded in time series by using a linear interpolation method, and the effectiveness of the expanded ground surface settlement dataset is determined by comparing and verifying the scatter plot diagonal matrix, the box plot and the heat map results.

[0082] Optionally, the model training module is configured to: The sequence information in the input ground surface settlement dataset is encoded through the encoder part of the Transformer, the multi-head attention mechanism is used to calculate the dependency relationship between different monitoring points or different time points, different attention weights are assigned to each monitoring point or time point through the calculation of the query, key and value matrices, and the output of the attention mechanism is nonlinearly transformed by the feedforward neural network to enhance the representation ability of the model to the ground surface settlement features, so that the model can learn complex settlement patterns. The feature sequence output by the Transformer encoder is used as the input of the BiLSTM module, the ground surface settlement sequence information is processed through the gating mechanism of the traditional recurrent neural network, and the ground surface settlement prediction value is output.

[0083] Optionally, the model training module is configured to: input the ground surface settlement data set into a VMD algorithm, and adaptively decompose an original signal into a plurality of intrinsic mode function components by iteratively solving a variational optimization problem; identify and remove high-frequency noise components according to modal frequency characteristics, and retain low-frequency effective signals reflecting long-term deformation rules of the stratum; superimpose the screened intrinsic mode function components, reconstruct a signal, and generate denoised ground surface settlement time series data as model input data.

[0084] Optionally, the model training module is configured to: define a parameter space, and set a search range of key hyperparameters of the model; quantify model prediction performance and evaluate model fitness by taking root mean square error of a test set as an optimization target; initialize a particle swarm, randomly generate a hyperparameter combination, explore the parameter space through a particle speed and position updating mechanism, dynamically adjust a search direction according to a fitness value, gradually approach a global optimal solution, and output an optimal hyperparameter combination for final training of the model after convergence.

[0085] Based on the same inventive concept, the present application also provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model according to any one of the above embodiments.

[0086] Based on the same inventive concept, the present application also provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the steps of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model according to any one of the above embodiments.

[0087] Based on the same inventive concept, the present application provides a computer program product comprising computer programs / instructions, wherein the computer programs / instructions are executed by a processor to implement the steps of the tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model according to any one of the above embodiments.

[0088] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referred to each other.

[0089] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by those skilled in the art be considered as within the scope of the present application. The embodiments of the present application will be described with reference to the attached drawings, wherein:

[0090] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, terminal device (apparatus) and computer program product according to the present application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 means for carrying out each of the one or more functions specified in the flowchart and / or block diagram block or blocks.

[0091] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 means for carrying out each of the one or more functions specified in the flowchart and / or block diagram block or blocks.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 means for carrying out each of the one or more functions specified in the flowchart and / or block diagram block or blocks.

[0093] While the preferred embodiments of the application have been described, additional variations and modifications can be made by those skilled in the art once they have the benefit of the present disclosure. Therefore, it is intended that the present application not be limited to the preferred and disclosed embodiments, but include all such variations and modifications as fall within the scope of the present application.

[0094] Finally, it needs to be pointed out that in the present application, the relational terms such as first and second and the like are used merely to differentiate one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles or terminal devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or further include elements inherent to such processes, methods, articles or terminal devices. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, method, article or terminal device including the element.

[0095] The tunnel excavation ground surface settlement prediction method based on the machine learning hybrid model provided by the present application is described in detail above, the principle and implementation mode of the present application are described by applying specific examples, the above example description is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation mode and application range, and the above description should not be understood as a limitation on the present application.

Claims

1. A method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model, characterized in that: The method comprises: Acquire multi-source heterogeneous information of the target tunnel and construct a surface settlement dataset. The multi-source heterogeneous information includes formation parameters, shield parameters, geometric parameters, and surface settlement monitoring data. A Transformer-BiLSTM hybrid model was constructed, and the PSO algorithm was used to adaptively tune hyperparameters based on the surface subsidence dataset to maximize the model prediction accuracy and obtain a surface subsidence prediction model. The decision logic of the surface subsidence prediction model is analyzed through SHAP values, and interpretable engineering guidance suggestions are output.

2. The method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to claim 1, characterized in that: Obtain multi-source heterogeneous information of the target tunnel and construct a surface settlement dataset, including: When there is a missing section for surface settlement monitoring in the target tunnel, a three-dimensional shield tunnel model corresponding to the target tunnel is established based on finite element analysis software, the entire construction process of the target tunnel is simulated, the stratum mechanical parameters and tunnel burial depth are adjusted to match the construction conditions of the missing section for monitoring, and after the simulation results are verified for consistency with the on-site measured data, the missing data for surface settlement monitoring are supplemented based on the simulation results.

3. The method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to claim 1, characterized in that: Obtain multi-source heterogeneous information of the target tunnel and construct a surface settlement dataset, including: Outliers of each parameter in the surface settlement data set were removed through box plots, and correlation coefficient analysis was performed on the processed data set to remove weakly correlated parameters; The linear interpolation method was used to expand the time series of the surface subsidence dataset. The scatter diagonal matrix diagram of the expanded surface subsidence dataset was compared with the box plot and heat map results to determine the validity of the expanded data.

4. The method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to claim 1, characterized in that: A Transformer-BiLSTM hybrid model was constructed. Based on the surface subsidence dataset, the PSO algorithm was used to adaptively tune hyperparameters to maximize the model prediction accuracy. The surface subsidence prediction model was obtained, including: The encoder part of the Transformer encodes the sequence information in the input surface subsidence dataset. The multi-head attention mechanism parallelizes the dependencies between different monitoring points or different time points. By calculating the query, key, and value matrix, different attention weights are assigned to each monitoring point or time point. The feedforward neural network performs nonlinear transformations on the output of the attention mechanism, enhancing the model's ability to represent surface subsidence features and enabling the model to learn complex subsidence patterns. The feature sequence output by the Transformer encoder is used as the input of the recurrent neural network module BiLSTM. The surface subsidence sequence information is processed through the gating mechanism of the traditional recurrent neural network to output the surface subsidence prediction value.

5. The method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to claim 1, characterized in that: A Transformer-BiLSTM hybrid model was constructed. Based on the surface subsidence dataset, the PSO algorithm was used to adaptively tune hyperparameters to maximize the model prediction accuracy. The surface subsidence prediction model was obtained, including: Inputting the surface settlement data set into the VMD algorithm, adaptively decomposing the original signal into multiple intrinsic mode functions by iteratively solving the variational optimization problem; According to the modal frequency characteristics, high-frequency noise components are identified and eliminated, retaining the low-frequency effective signals reflecting the long-term deformation law of the formation; The filtered intrinsic mode function components are superimposed to reconstruct the signal and generate the noise-reduced surface settlement time series data as the model input data.

6. The method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to claim 1, characterized in that: A Transformer-BiLSTM hybrid model was constructed. Based on the surface subsidence dataset, the PSO algorithm was used to adaptively tune hyperparameters to maximize the model prediction accuracy. The surface subsidence prediction model was obtained, including: Define the parameter space and set the search range for key hyperparameters of the model; The root mean square error of the test set is used as the optimization target to quantify the model prediction performance and evaluate the model fitness; Initialize the particle swarm and randomly generate hyperparameter combinations; explore the parameter space through the particle velocity and position update mechanism; dynamically adjust the search direction according to the fitness value, gradually approaching the global optimal solution; after convergence, output the optimal hyperparameter combination for final model training.

7. A tunnel excavation surface settlement prediction system based on a machine learning hybrid model, characterized in that: The system comprises: A data acquisition module is used to acquire multi-source heterogeneous information of the target tunnel and construct a surface settlement dataset. The multi-source heterogeneous information includes formation parameters, shield parameters, geometric parameters, and surface settlement monitoring data. A model training module is used to construct a Transformer-BiLSTM hybrid model, and adaptively tune hyperparameters using the PSO algorithm based on the surface subsidence dataset to maximize the model prediction accuracy and obtain a surface subsidence prediction model; The suggestion module is used to analyze the decision logic of the surface subsidence prediction model through SHAP values ​​and output interpretable engineering guidance suggestions.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for predicting surface settlement during tunnel excavation based on a machine learning hybrid model are implemented.

Images