A deep learning shield posture real-time prediction method fusing physical indexes

Abstract

This invention discloses a deep learning-based real-time shield tunneling attitude prediction method integrating physical indicators, including data acquisition and preprocessing, physical indicator calculation, data fusion and normalization, dataset partitioning, model construction, and model training. It constructs multi-dimensional input features by integrating geological parameters, tunnel geometric parameters, and shield parameters, preprocesses the data, and supplements it with physical knowledge acquired from field-collected data. After normalizing the above data, the dataset is partitioned, and the shield tunneling attitude prediction model constructed by the deep learning algorithm is trained on the training set. The method inputs the shield tunneling construction, attitude parameters, and physical knowledge of the previous tunnel segment, as well as the geological parameters of the next tunnel segment, to predict the shield tunneling attitude parameters of the next tunnel segment. The trained shield tunneling attitude prediction model is then tested on a test set to select the best-performing model. This invention can better handle time-series data, and the model can better adapt to different geological conditions.

Description

Technical Field

[0001] This invention belongs to the field of tunnel boring machine (TBM) construction technology, and more particularly to the field of TBM attitude prediction technology. Specifically, it relates to a deep learning method for real-time prediction of TBM attitude that integrates physical indicators. Background Technology

[0002] Attitude control during tunnel boring machine (TBM) construction in rock strata is crucial for ensuring construction quality and safety. Changes in TBM attitude directly affect the tunnel's axial deviation, thus impacting structural stability and construction efficiency. Traditional TBM attitude prediction methods rely on complex geological analysis and expert experience, making them ill-suited for varying geological conditions. In recent years, with the development of deep learning technology, data-driven prediction methods have become a research hotspot. These methods can learn from vast amounts of historical data about geological characteristics and TBM behavior patterns, thereby improving prediction accuracy and efficiency.

[0003] Existing deep learning-based methods for predicting tunnel boring machine (TBM) attitude still have many limitations. Most methods rely solely on data-driven approaches and fail to effectively integrate the correlation between geological physical properties and construction parameters, resulting in poor model interpretability and an inability to reflect the essential impact of geological conditions on attitude. Current technologies also exhibit limited accuracy in outlier detection and missing value imputation for multivariate time series. Some methods struggle to capture the long-term dependencies in time series during TBM construction, and their characterization of dynamic attitude evolution is insufficient. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a deep learning-based real-time shield tunneling attitude prediction method that integrates physical indicators. This method constructs multi-dimensional input features by comprehensively considering geological parameters, tunnel geometric parameters, and shield tunneling parameters. It preprocesses the data and obtains physical indicators based on field-collected data to supplement the model's input parameters. After normalizing the data, the dataset is divided, and the shield tunneling attitude prediction model constructed by the deep learning algorithm is trained on the training set. The model is then trained by inputting the shield tunneling construction, attitude parameters, and physical indicators of the previous tunnel segment, as well as the geological parameters of the next tunnel segment, to predict the shield tunneling attitude parameters of the next tunnel segment. Finally, the trained shield tunneling attitude prediction model is tested on a test set to select the best-performing model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based real-time prediction method for tunnel boring machine (TBM) attitude integrating physical indicators, characterized by the following steps:

[0007] S1. Data Acquisition and Preprocessing: Collect historical data from the shield tunneling site as the raw dataset X1, where X1 = [A, B, C]. TWhere A is the geological parameter set, B is the shield tunneling construction parameter set, and C is the shield attitude parameter set; abnormal data preprocessing and missing data preprocessing are performed on the dataset to obtain the processed dataset X1';

[0008] S2. Calculation of physical indicators; Based on the data in shield construction parameter set B and shield attitude parameter set C, physical indicators are obtained, with tunneling specific energy as the core quantitative indicator of stratum excavability. The calculation formula is as follows: ;

[0009] In the formula, S The specific energy for tunneling is expressed in kW·h / m. 3 ; T The torque of the cutter head is expressed in kN·m. N The rotational speed of the cutter head is expressed in r·min. -1 ; F Total propulsion force, unit: kN; v The propulsion speed is expressed in mm·min. -1 ; D The diameter of the tunnel boring machine cutterhead is in meters (m).

[0010] All the calculated S indicators are combined into a set Q;

[0011] S3. Data Fusion and Normalization: Merge set Q with the preprocessed dataset X1' to construct an enhanced dataset X = [X1', Q]. T ;

[0012] The min-max normalization method is used to map all parameters to [0,1], as shown in the following formula: ;

[0013] in, x * It is normalized data. x The original data, x min The minimum value of the data. x max The maximum value of the data is used to obtain the normalized dataset X. * ;

[0014] S4. Dataset partitioning; data reconstruction using the sliding window method, dividing the normalized dataset X... * The data in the dataset is read continuously to generate a time series dataset Y; the dataset Y is then divided into a training set and a test set.

[0015] S5. Model Construction; The inputs to the prediction model include the shield tunneling construction, attitude parameters, and physical indicators of the previous tunnel segment, as well as the geological parameters of the next tunnel segment. The output is the shield attitude parameters of the next tunnel segment. The prediction model includes: an input layer, responsible for receiving input data; a Bi-LSTM layer, which finds the mapping relationship between input and output data; a Dropout layer, which prevents the model from overfitting; a fully connected layer, which maps the output to the target dimension; and an output layer, which gives the final prediction result.

[0016] S6. Model Training: The shield tunneling attitude prediction model is trained using a training set. The hyperparameters of the machine learning algorithm are selected through trial and error, and the mean squared error (MSE) is used as the loss value. ;

[0017] In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield attitude parameters. When the mean square error (MSE) of the loss function value of the shield attitude prediction model no longer decreases, training is stopped, and the optimal combination of hyperparameters of the shield attitude prediction model after preliminary training is obtained.

[0018] The test set is input into the initially trained shield attitude prediction model. Training stops when the prediction accuracy exceeds a set threshold; if the prediction accuracy is less than the set threshold, the training set is changed and training is repeated. Root mean square error (RMSE) and goodness-of-fit (R²) are used as metrics. 2 To evaluate the accuracy of the model:

[0019] ;

[0020] ;

[0021] In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield tunneling attitude parameters. It is the average of the measured values ​​of the shield tunneling attitude parameters; the smaller the RMSE result, the better the R... 2 The closer the result is to 1, the higher the prediction accuracy of the model.

[0022] Furthermore, in step S1, the geological parameter set A is obtained by processing the geological survey report before tunnel excavation, including: the natural unit weight, cohesion, internal friction angle, and compression modulus of the overlying soil layer, and the thickness of various overlying soil layers along the excavation direction; the shield construction parameter set B is obtained by the shield machine's automatic data acquisition system, including: soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, and total thrust; the shield attitude parameter set C includes: shield cutterhead diameter, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation.

[0023] Furthermore, in step S1, the abnormal data preprocessing of the dataset involves using the Multivariate Time Series Anomaly Detection Algorithm (USAD) to process outliers in the collected data. Specifically, this includes constructing two autoencoders, AE1 and AE2, training AE2 to distinguish the authenticity of the output data from AE1, and calculating anomaly scores: "AnomalyScore = α⋅‖". x 1-AE1( x 1)" "‖" _"2" "+(1- α )⋅‖ x 1-AE2( x 1) "‖" _"2", where, x 1 represents the original data. α is The weighting parameter has a value range of [0, 1]; a threshold is set to identify and remove data with abnormal scores that exceed the threshold.

[0024] Furthermore, in step S1, the missing data preprocessing involves imputing missing values ​​using Missing Forest MF: specifically including: for complete samples... Where N is the number of samples and M is the number of features, a random forest is trained to predict samples X containing missing values ​​based on non-missing features. miss Missing values; define the iteration termination condition γ=10 -5 The iteration is controlled by the difference formula Δ for continuous or discrete data until Δ < γ.

[0025] Furthermore, in step S4, the dataset Y is divided into two parts: the first 80% of the dataset Y is used as the training set for model training, and the last 20% of the dataset Y is used as the test set for verifying generalization ability.

[0026] Furthermore, in step S5, the prediction model includes:

[0027] An input layer that receives input data;

[0028] Two Bi-LSTM layers are used to find the mapping relationship between input and output data. The forward LSTM learns the feature dependency of the sequence t=0→t=W; the backward LSTM learns the feature dependency of the sequence t=W→t=0; the forward and backward outputs are concatenated into a 2H-dimensional vector, where H is the dimension of a single LSTM hidden layer, which is 64.

[0029] The formula is: ;

[0030] A Dropout layer to prevent the model from overfitting;

[0031] A fully connected layer maps the output to the target dimension, and uses the ReLU activation function to compress the feature dimension, as shown in the formula: ;

[0032] An output layer that predicts the shield's attitude parameters through linear activation: The final prediction result will be given.

[0033] Furthermore, during the training process in step S6, the input data for the model includes the natural unit weight, cohesion, internal friction angle, compression modulus, thickness along the tunneling direction, soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, total thrust, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation within s time periods. The output data of the model includes the shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation within the predicted time period t. The test set is then input into the pre-trained shield attitude prediction model. When the prediction accuracy exceeds the set threshold, the pre-trained shield attitude prediction model is considered the final trained shield tunneling attitude prediction model. If the prediction accuracy is less than the set threshold, the training set is changed and retraining is performed.

[0034] The shield tunneling machine attitude prediction system constructed in this invention includes:

[0035] The data acquisition module uses a sensor interface to connect with the ground-penetrating radar, the tunnel boring machine controller, and the gyroscope to obtain geological parameters, tunnel boring machine construction parameters, and tunnel boring machine attitude parameters.

[0036] The data preprocessing module calculates a potential index based on the collected data and incorporates it as a supplementary input parameter into the model. Subsequently, all data is normalized and proportionally divided into training and test sets.

[0037] The attitude prediction module obtains the predicted values ​​of the tunnel boring machine's attitude in real time through the trained prediction model.

[0038] The present invention has the following beneficial effects:

[0039] The method of this invention integrates physical indicators, which enhances the model's adaptability to complex geological conditions and significantly improves prediction accuracy compared to traditional physical models. It uses a Bi-LSTM network to capture the bidirectional dependencies of time series and combines it with sliding window technology to make full use of time series features and improve dynamic prediction capabilities.

[0040] This invention employs Bi-LSTM as the core algorithm of its prediction model, enabling better processing of time-series data and capturing the dynamic characteristics of shield tunneling attitude changes. By integrating physical indicators, the model can better adapt to different geological conditions, enhancing its generalization ability and adaptability. Verification through practical engineering cases demonstrates that this invention can effectively guide shield tunneling construction, reduce engineering risks and costs caused by attitude deviations, and provide scientific decision support for shield tunneling construction. Attached Figure Description

[0041] Figure 1 This is a flowchart of the shield tunneling attitude prediction method of the present invention; Figure 2 This is a comparison chart of the model prediction results and measured values ​​corresponding to the horizontal deviation of the shield head in this invention; Figure 3 This is a comparison chart of the model prediction results and measured values ​​corresponding to the vertical deviation of the shield head in this invention; Figure 4 This is a comparison chart of the model prediction results and measured values ​​corresponding to the horizontal deviation of the shield tail in this invention; Figure 5 This is a comparison chart of the model prediction results and measured values ​​corresponding to the vertical deviation of the shield tail in this invention; Figure 6 This is a comparison chart of the prediction results of various models corresponding to the horizontal deviation of the shield head in this invention; Figure 7 This is a comparison chart of the prediction results of various models corresponding to the vertical deviation of the shield head in this invention; Figure 8 This is a comparison chart of the prediction results of various models corresponding to the horizontal deviation of the shield tail in this invention; Figure 9 This is a comparison chart of the prediction results of various models corresponding to the vertical deviation of the shield tail in this invention; Figure 10 This invention relates to a shield tunneling machine attitude prediction system. Detailed Implementation

[0042] The present invention will be further described below with reference to specific embodiments. These specific embodiments are further explanations of the principles of the present invention and are not intended to limit the present invention in any way. Any technology that is the same as or similar to the present invention does not exceed the scope of protection of the present invention.

[0043] Example:

[0044] As shown in the figure, this embodiment relates to a deep learning-based real-time prediction method for tunnel boring machine (TBM) attitude that integrates physical indicators. The prediction method specifically includes the following steps:

[0045] S1. Collect historical data from the shield tunneling site as the raw dataset X1, where X1 = [A, B, C]. THere, A represents the geological parameter set, B represents the shield tunneling construction parameter set, and C represents the shield attitude parameter set. Geological parameter set A is processed based on the geological survey report before tunnel excavation and mainly includes the natural unit weight, cohesion, internal friction angle, and compression modulus of the overlying soil layers, as well as the thickness of various overlying soil layers along the excavation direction. Shield tunneling construction parameter set B is obtained from the shield machine's automatic data acquisition system and includes soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, and total thrust. Shield attitude parameter set C includes the cutterhead diameter, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation.

[0046] Furthermore, due to sensor malfunctions, outliers in operating parameters often occur during tunneling. The Multivariate Time Series Anomaly Detection Algorithm (USAD) is employed to handle outliers in the collected data: two autoencoders, AE1 and AE2, are constructed. AE2 is trained to distinguish between genuine and non-genuine output data from AE1, and an anomaly score is calculated: "AnomalyScore = α⋅‖". x 1-AE1( x 1)" "‖" _"2" "+(1- α )⋅‖ x 1-AE2( x 1) "‖" _"2", where, x 1 represents the original data. α is The weighting parameter has a value range of [0, 1]; a threshold is set to identify and remove data with abnormal scores that exceed the threshold.

[0047] Furthermore, missing values ​​are imputed using a missing forest (MF) model: for complete samples... Where N is the number of samples and M is the number of features, a random forest is trained to predict samples X containing missing values ​​based on non-missing features. miss Missing values; define the iteration termination condition γ=10 -5 The iteration is controlled by the difference formula Δ for continuous or discrete data until Δ < γ.

[0048] The dataset X1' is obtained after preprocessing.

[0049] S2. Substitute the cutterhead torque (T), cutterhead rotation speed (N), total propulsion force (F), propulsion speed (v), and shield cutterhead diameter (D) from construction parameter set B and shield parameter set C into the calculation formula for tunneling specific energy: The tunneling specific energy corresponding to each time point is obtained, and the calculated values ​​of the tunneling specific energy are constructed into a set Q.

[0050] The excavability of a stratum is quantified by this index, reflecting the stratum lithology, frictional properties, and resistance to tunneling by the tunnel boring machine (TBM). A higher excavation energy indicates greater difficulty in tunneling and a greater impact on the TBM's attitude.

[0051] S3. Merge set Q with the preprocessed dataset X1' to construct the augmented dataset X = [X1', Q] T Furthermore, the data in X is normalized, including:

[0052] The min-max normalization method is used to map all parameters to [0,1], as shown in the following formula: ;

[0053] in, x * It is normalized data. x The original data, x min The minimum value of the data. x max The maximum value of the data is used to obtain the normalized dataset X. * .

[0054] S4. Data reconstruction is performed using the sliding window method; the step size of the shield tunneling attitude prediction model is determined. s and the predicted time period t For dataset X * Perform continuous readings to generate time series data Y. Divide the dataset Y into two parts: the first 80% is used as the training set, and the last 20% is used as the test set.

[0055] S5. Construct a shield tunneling attitude prediction model. The model's inputs include the shield tunneling construction, attitude parameters, and physical indicators of the previous tunnel segment, as well as the geological parameters of the next tunnel segment. The output is the shield tunneling attitude parameters for the next tunnel segment. The framework of the prediction model includes:

[0056] An input layer that receives input data.

[0057] Two Bi-LSTM layers are used to find the mapping relationship between the input and output data. The forward LSTM learns the feature dependencies of the sequence t=0→t=W, such as the impact of previous thrust changes on the current attitude; the backward LSTM learns the feature dependencies of the sequence t=W→t=0, such as the constraint of future ground hardness on the current attitude. The forward and backward outputs are concatenated into a 2H-dimensional vector, where H is the dimension of a single LSTM hidden layer, which is set to 64. The formula is: .

[0058] A Dropout layer to prevent the model from overfitting.

[0059] A fully connected layer maps the output to the target dimension, and uses the ReLU activation function to compress the feature dimension, as shown in the formula: .

[0060] An output layer that predicts the shield's attitude parameters through linear activation: The final prediction result will be given.

[0061] S6. Model Training: The shield tunneling attitude prediction model is trained using a training set. The hyperparameters of the machine learning algorithm are selected through trial and error. The mean squared error (MSE) is used as the loss value, and the specific formula is as follows: ;

[0062] In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield tunnel attitude parameters. When the mean square error (MSE) of the loss function value of the shield attitude prediction model no longer decreases, training is stopped, and the optimal hyperparameter combination of the shield attitude prediction model after preliminary training is obtained, as shown in Table 1 below.

[0063]

[0064] During training, the model's input data includes the natural unit weight, cohesion, internal friction angle, compression modulus, thickness along the tunneling direction, soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, total thrust, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation over s time intervals. The output data includes the shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation over the prediction time interval t. The test set is then input into the pre-trained shield attitude prediction model. When the prediction accuracy exceeds a set threshold, the pre-trained shield attitude prediction model is considered the final trained shield tunneling attitude prediction model. If the prediction accuracy is less than the set threshold, the training set is changed and retraining is performed.

[0065] Prediction accuracy is measured using root mean square error (RMSE) and goodness of fit (R²). 2 The evaluation is conducted using the following formula: ;

[0066] ;

[0067] In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield tunneling attitude parameters. This is the average of the measured values ​​of the shield tunneling attitude parameters. The smaller the RMSE result, the better the R... 2The closer the result is to 1, the higher the prediction accuracy of the model. The R-squared value of the prediction results of the model in this invention... 2 The values ​​were 0.94, 0.94, 0.90, and 0.87, respectively.

[0068] The comparison between the predicted and actual values ​​of the shield head horizontal attitude, shield head vertical attitude, shield tail horizontal attitude, and shield tail vertical attitude of the model in this invention can be found by referring to... Figures 2 to 5 As shown, the comparison of the prediction results of each model for the shield's horizontal attitude, vertical attitude, horizontal attitude, and vertical attitude can be found in the reference. Figures 6 to 9 The evaluation metrics for the prediction results of each model for the horizontal attitude of the shield head, the vertical attitude of the shield head, the horizontal attitude of the shield tail, and the vertical attitude of the shield tail are shown in Table 2 below.

[0069]

[0070] The model proposed in this invention exhibits high agreement between predicted and measured values ​​for the shield's horizontal, vertical, horizontal, and vertical attitudes of the shield head and tail, with highly consistent curve trends. The RMSE values ​​of this invention's model are 0.05, 0.04, 0.08, and 0.08, respectively, while the RMSE values ​​of the Bi-LSTM model are 0.13, 0.16, 0.16, and 0.14, respectively. The RMSE values ​​of the Bi-LSTM model considering physical parameters are 0.18, 0.18, 0.21, and 0.21, respectively. The RMSE values ​​of this invention's model are significantly lower than other mainstream models, indicating that the average deviation between the predicted and measured values ​​is minimal. The R-value of this invention's model... 2 The values ​​are 0.94, 0.94, 0.90, and 0.87, respectively, and the R-values ​​of the Bi-LSTM model are... 2 The values ​​are 0.83, 0.82, 0.80, and 0.81, respectively. Considering the physical parameters, the R-values ​​of the Bi-LSTM model are... 2 The values ​​are 0.88, 0.86, 0.85, and 0.85, respectively. The R-value of the model in this invention is... 2 A value closer to 1 indicates a stronger ability to explain the changes in the shield tunnel's attitude.

[0071] like Figure 10 As shown, this embodiment employs a tunnel boring machine attitude prediction system, including:

[0072] Data Acquisition Module: The data acquisition module is the foundation of the entire system. It acquires the tunnel boring machine's (TBM) working data by connecting to hardware devices such as ground-penetrating radar, TBM controller, and gyroscopes. Geological parameters include the natural unit weight, cohesion, internal friction angle, and compression modulus of the overlying soil. TBM construction parameters include soil chamber pressure, advance speed, cutterhead rotation speed, cutterhead torque, and total thrust, all of which affect the TBM's working attitude. TBM attitude parameters include cutterhead diameter, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation.

[0073] Data Preprocessing Module: The main task of the data preprocessing module is to transform the raw data collected on-site into time-series data usable by the model. First, using the formula for calculating tunneling specific energy, its calculated values ​​at various time points are obtained and used as one of the input parameters. Then, all input and output parameters are normalized to eliminate dimensional differences. Finally, the data is divided into time series using a sliding window method, providing standardized input for subsequent model training.

[0074] Attitude Prediction Module: This module uses a trained Bi-LSTM prediction model to predict the attitude over time. Its core task is to predict the attitude parameters of the tunnel boring machine (TBM) based on preprocessed time-series data. First, the input time-series data is fed into the deep learning model, which models the temporal dependency between tunneling conditions and attitude changes. Then, the model outputs predicted attitude values ​​for future moments, which are compared with actual measurements to evaluate prediction accuracy. This module can identify TBM attitude change trends in advance, supporting automated control and construction safety during the tunneling process.

[0075] This invention can accurately identify and replace outliers in multivariate time series data, ensuring the quality and consistency of input data and further improving the stability and reliability of the model. It innovatively integrates physical indicators with deep learning models, effectively supplementing the shortcomings of traditional data-driven models in expressing physical mechanisms, and constructing an efficient, accurate and highly interpretable real-time prediction method for tunnel boring machine attitude.

Claims

1. A deep learning-based real-time prediction method for tunnel boring machine (TBM) attitude integrating physical indicators, characterized in that... Includes the following steps: S1. Data Acquisition and Preprocessing: Collect historical data from the shield tunneling site as the raw dataset X1, where X1 = [A, B, C]. T Where A is the geological parameter set, B is the shield tunneling construction parameter set, and C is the shield attitude parameter set; abnormal data preprocessing and missing data preprocessing are performed on the dataset to obtain the processed dataset X1'; S2. Calculation of physical indicators; Based on the data from shield construction parameter set B and shield attitude parameter set C, physical knowledge is obtained, and the tunneling specific energy is used as the core quantitative indicator of the excavability of the strata. The calculation formula is as follows: ; In the formula, S The specific energy for tunneling is expressed in kW·h / m. 3 ; T The torque of the cutter head is expressed in kN·m. N The rotational speed of the cutter head is expressed in r·min. -1 ; F Total propulsion force, unit: kN; v The propulsion speed is expressed in mm·min. -1 ; D The diameter of the tunnel boring machine cutterhead is in meters (m). All the calculated S indicators are combined into a set Q; S3. Data Fusion and Normalization: Merge set Q with the preprocessed dataset X1' to construct an enhanced dataset X = [X1', Q]. T ; The min-max normalization method is used to map all parameters to [0,1], as shown in the following formula: ; in, x * It is normalized data. x The original data, x min The minimum value of the data. x max The maximum value of the data is used to obtain the normalized dataset X. * ; S4. Dataset partitioning; data reconstruction using the sliding window method, dividing the normalized dataset X... * The data in the dataset is read continuously to generate a time series dataset Y; the dataset Y is then divided into a training set and a test set. S5. Model Construction; The inputs to the prediction model include the shield tunneling construction, attitude parameters, and physical knowledge of the previous tunnel segment, as well as the geological parameters of the next tunnel segment. The output is the shield attitude parameters of the next tunnel segment. The prediction model includes: an input layer, responsible for receiving input data; a Bi-LSTM layer, which finds the mapping relationship between input and output data; a Dropout layer, which prevents the model from overfitting; a fully connected layer, which maps the output to the target dimension; and an output layer, which gives the final prediction result. S6. Model Training: The shield tunneling attitude prediction model is trained using a training set. The hyperparameters of the machine learning algorithm are selected through trial and error, and the mean squared error (MSE) is used as the loss value. ； In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield attitude parameters. When the mean square error (MSE) of the loss function value of the shield attitude prediction model no longer decreases, training is stopped, and the optimal combination of hyperparameters of the shield attitude prediction model after preliminary training is obtained. The test set is input into the initially trained shield attitude prediction model. Training stops when the prediction accuracy exceeds a set threshold; if the prediction accuracy is less than the set threshold, the training set is changed and training is repeated. Root mean square error (RMSE) and goodness-of-fit (R²) are used as metrics. 2 The accuracy of the model is evaluated: ； ； In the formula, These are the measured values ​​of the shield tunneling machine's attitude parameters. These are the predicted values ​​of the shield tunneling attitude parameters. It is the average of the measured values ​​of the shield tunneling attitude parameters; the smaller the RMSE result, the better the R... 2 The closer the result is to 1, the higher the prediction accuracy of the model.

2. The deep learning-based real-time prediction method for tunnel boring machine attitude based on fused physical indicators according to claim 1, characterized in that: In step S1, the geological parameter set A is obtained by processing the geological survey report before tunnel excavation, including: the natural unit weight, cohesion, internal friction angle, and compression modulus of the overlying soil layer, and the thickness of various overlying soil layers along the excavation direction; the shield construction parameter set B is obtained by the shield machine's automatic data acquisition system, including: soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, and total thrust; the shield attitude parameter set C includes: shield cutterhead diameter, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation.

3. The deep learning-based real-time prediction method for tunnel boring machine attitude based on fused physical indicators according to claim 1, characterized in that: In step S1, the abnormal data preprocessing of the dataset involves using the Multivariate Time Series Anomaly Detection Algorithm (USAD) to process outliers in the collected data. Specifically, this includes: constructing two autoencoders, AE1 and AE2; training AE2 to distinguish between the authenticity of AE1's output data; and calculating the anomaly score: "AnomalyScore = α⋅‖". x 1-AE1( x 1)" "‖" _"2" "+(1- α )⋅‖ x 1-AE2( x 1) "‖" _"2", where, x 1 represents the original data. α is The weighting parameter has a value range of [0,1]; a threshold is set to identify and remove data with abnormal scores that exceed the threshold.

4. The deep learning-based real-time prediction method for tunnel boring machine attitude based on fused physical indicators according to claim 1, characterized in that: In step S1, the missing data preprocessing involves imputing missing values ​​using Missing Forest MF: specifically, this includes: processing complete samples... Where N is the number of samples and M is the number of features, a random forest is trained to predict samples X containing missing values ​​based on non-missing features. miss Missing values; define the iteration termination condition γ=10 -5 The iteration is controlled by the difference formula Δ for continuous or discrete data until Δ < γ.

5. The deep learning-based real-time prediction method for shield tunnel attitude based on fused physical indicators according to claim 1, characterized in that: In step S4, the dataset Y is divided into two parts: the first 80% of the dataset Y is used as the training set for model training, and the last 20% of the dataset Y is used as the test set for verifying generalization ability.

6. The deep learning-based real-time prediction method for tunnel boring machine attitude based on fused physical indices according to claim 1, characterized in that: In step S5, the prediction model includes: An input layer that receives input data; Two Bi-LSTM layers are used to find the mapping relationship between input and output data. The forward LSTM learns the feature dependency of the sequence t=0→t=W; the backward LSTM learns the feature dependency of the sequence t=W→t=0; the forward and backward outputs are concatenated into a 2H-dimensional vector, where H is the dimension of a single LSTM hidden layer, which is 64. The formula is: ; A Dropout layer to prevent the model from overfitting; A fully connected layer maps the output to the target dimension, and uses the ReLU activation function to compress the feature dimension, as shown in the formula: ; An output layer that predicts the shield's attitude parameters through linear activation: The final prediction result will be given.

7. The deep learning-based real-time prediction method for tunnel boring machine attitude based on fused physical indicators according to claim 1, characterized in that: In step S6, during training, the model's input data includes the natural unit weight, cohesion, internal friction angle, compression modulus, thickness along the tunneling direction, soil chamber pressure, propulsion speed, cutterhead rotation speed, cutterhead torque, total thrust, shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation over s time periods. The model's output data includes the shield head horizontal deviation, shield tail horizontal deviation, shield head vertical deviation, and shield tail vertical deviation over the predicted time period t. The test set is then input into the pre-trained shield attitude prediction model. When the prediction accuracy exceeds the set threshold, the pre-trained shield attitude prediction model is considered the final trained shield tunneling attitude prediction model. If the prediction accuracy is less than the set threshold, the training set is changed and retraining is performed.

8. The deep learning-based real-time prediction method for shield tunnel attitude based on the fusion of physical indices according to any one of claims 1 to 7, characterized in that: The prediction method employs a tunnel boring machine attitude prediction system constructed from the following modules: The data acquisition module uses a sensor interface to connect with the ground-penetrating radar, the tunnel boring machine controller, and the gyroscope to obtain geological parameters, tunnel boring machine construction parameters, and tunnel boring machine attitude parameters. The data preprocessing module calculates the exploitability index based on the collected data and incorporates it into the model as a supplementary input parameter; subsequently, all data is normalized and divided into training and test sets according to a certain ratio. The attitude prediction module obtains the predicted values ​​of the tunnel boring machine's attitude in real time through the trained prediction model.

Images