TBM surrounding rock grade prediction method based on double-layer cyclic simulated annealing stacking model

Abstract

The invention discloses a TBM surrounding rock grade prediction method based on a double-layer cyclic simulated annealing stacking model. The TBM surrounding rock grade prediction method comprises the following steps: step 1, collecting TBM engineering data; 2, preprocessing the data to obtain processed data; 3, constructing a prediction effect evaluation index; 4, constructing a model pool and a double-layer circulation simulated annealing algorithm optimization program to obtain a prediction model, and training to obtain a final surrounding rock grade prediction model; 5, according to the prediction effect evaluation index, surrounding rock grade prediction is carried out, and comparison verification of similar models is carried out; and step 6, calculating the feature contribution degree of the final surrounding rock grade prediction model by using the SHAP model. According to the TBM surrounding rock grade prediction method based on the double-layer cyclic simulated annealing stacking model, the problem that the flow pattern change and response of the cavitation in the dynamic change process of the navigation depth of the supercavitation vehicle cannot be predicted in the prior art is solved, and real-time surrounding rock grade identification and early warning support is provided for TBM construction.

Description

Technical Field

[0001] This invention belongs to the field of surrounding rock classification technology, specifically relating to a TBM surrounding rock grade prediction method based on a two-layer cyclic simulated annealing stacking model. Background Technology

[0002] Accurate identification of the surrounding rock grade during TBM tunneling is crucial for tunneling efficiency, equipment safety, and support design. Traditional rock classification methods, such as RMR, Q-system, and HC method, rely on field surveys and human experience, making real-time and dynamic judgment difficult. In recent years, machine learning methods, such as neural networks, support vector machines, and random forests, have been introduced for rock grade prediction. However, when faced with high-dimensional, imbalanced, and noisy TBM tunneling data, single models often suffer from insufficient generalization ability and poor interpretability. Stacked ensemble learning, by combining multiple base models and meta-models, can effectively improve model robustness and prediction accuracy, but its performance is highly dependent on model combination and hyperparameter configuration. Traditional methods often rely on empirical selection and lack systematic optimization. Furthermore, TBM tunneling data contains issues such as class imbalance and feature redundancy, further affecting model performance. Summary of the Invention

[0003] The purpose of this invention is to provide a TBM-based method for predicting the surrounding rock grade based on a two-layer cyclic simulated annealing stacking model, which solves the problems of insufficient generalization ability and poor interpretability in existing surrounding rock grade prediction technologies.

[0004] The technical solution adopted in this invention is a method for predicting the surrounding rock grade of TBM based on a two-layer cyclic simulated annealing stacking model, comprising the following steps: Step 1: Collect TBM project data; Step 2: Preprocess the TBM engineering data to obtain processed data; Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0005] The invention is further characterized by: In step 1, the TBM engineering data includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section.

[0006] Step 2 preprocessing includes unique thermal coding of surrounding rock grade, outlier detection, feature selection, and data balancing.

[0007] TBM tunneling parameter data includes cutterhead rotation speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, support shoe cylinder left stroke A6, support shoe cylinder right stroke A7, top shield cylinder left pressure A8, top shield cylinder left stroke A9, top shield cylinder right pressure A10, top shield cylinder right stroke A11, wedge cylinder left pressure A12, wedge cylinder right pressure A13, side shield cylinder left pressure A14, side shield cylinder left stroke A15, side shield cylinder right pressure A16, and side shield cylinder right stroke A17.

[0008] Step 2 is as follows: Step 2.1: Import the TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Step 2.2: Thermal coding of surrounding rock grades: Class II, IIIa, IIIb, IV and V rock masses are coded as 0, 1, 2, 3 and 4 respectively; Step 2.3: Use PCA principal component analysis to reduce the dimensionality of the TBM tunneling parameter data and map the dimensionality-reduced data to three-dimensional space; Step 2.4: Set the outlier threshold to 10% and use the iForest isolated forest algorithm to remove outliers; Step 2.5: For the data after removing outliers, the RF-RFE random forest-recursive feature removal method is used to further filter features; Step 2.6: Divide the data after feature filtering into training and test sets in a 9:1 ratio. Use the Borderline-SMOTE boundary-sensitive synthetic sampling algorithm to balance the data of each category in the training set to obtain the final data.

[0009] The performance evaluation metrics in step 3 include accuracy (ACC), recall (REC), precision (PRC), score (F1), true positive rate (TPR), and false positive rate (FPR). The expression for accuracy ACC is shown in equation (1); (1); The expression for recall rate REC is shown in equation (2); (2); The expression for accuracy PRC is shown in equation (3); (3); The expression for the fraction F1 is shown in equation (4); (4); The expression for the true rate of return (TPR) is shown in equation (5); (5); The expression for the false positive rate FPR is shown in equation (6); (6); in, TP A true positive is a sample that is actually a positive sample and is correctly classified as a positive sample. TN A true negative is a sample that is actually a negative class and has been correctly classified as a negative class. FP A false positive is a sample that is actually a negative class but is incorrectly classified as a positive class. FN False negatives are samples that are actually positive but are incorrectly classified as negative; ACC is precision; REC is recall; PRC is accuracy; F1 score is the true positive rate; TPR is the false positive rate; FPR is the false positive rate.

[0010] The model pool in step 4 includes single models, ensemble learning models, and neural network models; Single models include Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Naive Bayes (NB). Ensemble learning models include Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Class Gradient Boosting (CatBoost), and Adaptive Boosting (AdaBoost). Neural network models include the multilayer perceptron classifier (MLPClassifier) ​​model.

[0011] The process of constructing the optimization program for the double-loop simulated annealing algorithm in step 4 is as follows: S41. Initialization and Neighborhood Generation: A solution is randomly initialized using the simulated annealing algorithm (SA), which contains a set of base models and a meta-model. In each iteration, a new solution is generated through a neighborhood operation, which randomly replaces a base model or the meta-model with equal probability. S42. Fitness Evaluation: Select a set of base models and meta-model combinations. The system triggers the inner loop mechanism and uses the SA algorithm to search the hyperparameter space of the base learners and meta-learners under the current combination to find the hyperparameter combination that maximizes the model prediction performance. By evaluating the fitness f(xi) of the hyperparameter combination, it decides whether to accept the new solution. S43. Solution acceptance criterion: The Metropolis classical criterion is adopted. If the fitness of the current solution xi is better than that of the previous solution pi, then pi is updated; if pi is better than the global optimal solution g, then g is updated; if the solution is not improved, the temperature T is reduced through simulated annealing mechanism to continue generating new solutions until the next optimal combination C′ is found. S44. Temperature Update and Termination: The temperature decays according to T←αT. The optimization terminates when the temperature drops to Tmin or the maximum number of iterations is reached, and the final stacked model composition and hyperparameters are obtained.

[0012] The beneficial effects of this invention are: The present invention provides a TBM surrounding rock grade prediction method based on a double-layer cyclic simulated annealing stacked model. It adopts a double-layer cyclic simulated annealing algorithm to realize automatic selection of base model and meta-model and hyperparameter co-optimization in the stacked ensemble model, overcoming the limitation of traditional ensemble learning methods that rely on manual experience configuration. Through a two-layer optimization mechanism, the optimal matching relationship between model combinations and geological features is effectively explored. Validation on multiple engineering datasets shows that the overall accuracy of the model reaches 0.938, and the macro F1 score reaches 0.939, significantly outperforming single models and traditional stacking methods. Anomaly detection is performed using principal component analysis and isolated forest PCA-iForest, combined with random forest recursive features to eliminate RF-RFE and screen key features. Borderline-SMOTE boundary-sensitive oversampling technology is introduced to handle class imbalance, constructing a high-quality training dataset. This preprocessing process effectively solves the problems of high noise, high dimensionality, and uneven sample distribution in TBM tunneling data, improving the robustness and generalization performance of the model under complex geological conditions. In the surrounding rock grade prediction, the SHAP interpretability analysis framework is introduced, which can quantify the contribution of each tunneling parameter to the classification results and reveal the nonlinear mapping relationship between key parameters such as thrust, penetration depth, and support shoe travel and the surrounding rock grade. This not only enhances the transparency and credibility of the model but also provides theoretical basis and engineering guidance for TBM tunneling parameter optimization and support design. Attached Figure Description

[0013] Figure 1 This is a flowchart of the TBM surrounding rock grade prediction method based on a double-layer cyclic simulated annealing stacking model of the present invention; Figure 2 This is an anomaly score distribution map of the isolated forest iForest in Embodiment 6 of the present invention; Figure 3 This is a graph showing the RF-RFE feature selection performance of Embodiment 6 of the present invention; Figure 4 This is the confusion matrix diagram of the model in Embodiment 6 of the present invention on the test set; Figure 5 This is the ROC curve of the model in Embodiment 6 of the present invention on the test set; Figure 6 This is a comparison chart of the number of misclassifications in the model in Embodiment 6 of the present invention; Figure 7 This is an input feature importance analysis diagram based on SHAP values ​​in Embodiment 6 of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0015] Example 1 The TBM surrounding rock grade prediction method proposed in this embodiment is based on a two-layer cyclic simulated annealing stacking model. Figure 1 As shown, it includes the following steps: Step 1: Collect TBM project data; Step 2: Preprocess the TBM engineering data to obtain processed data; Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0016] Example 2 The TBM surrounding rock grade prediction method proposed in this embodiment is based on a two-layer cyclic simulated annealing stacking model. Figure 1 As shown, it includes the following steps: Step 1: Collect TBM project data; TBM engineering data includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Step 2: Preprocess the TBM engineering data to obtain processed data; Preprocessing includes unique thermal coding of surrounding rock grade, outlier detection, feature selection, and data balancing. Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0017] Example 3 The TBM surrounding rock grade prediction method proposed in this embodiment is based on a two-layer cyclic simulated annealing stacking model. Figure 1 As shown, it includes the following steps: Step 1: Collect TBM project data; TBM engineering data includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section; TBM tunneling parameter data includes cutterhead rotation speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, support shoe cylinder left stroke A6, support shoe cylinder right stroke A7, top shield cylinder left pressure A8, top shield cylinder left stroke A9, top shield cylinder right pressure A10, top shield cylinder right stroke A11, wedge cylinder left pressure A12, wedge cylinder right pressure A13, side shield cylinder left pressure A14, side shield cylinder left stroke A15, side shield cylinder right pressure A16, and side shield cylinder right stroke A17; Step 2: Preprocess the TBM engineering data to obtain processed data; Preprocessing includes unique thermal coding of surrounding rock grade, outlier detection, feature selection, and data balancing. Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0018] Example 4 The TBM surrounding rock grade prediction method proposed in this embodiment is based on a two-layer cyclic simulated annealing stacking model. Figure 1 As shown, it includes the following steps: Step 1: Collect TBM project data; TBM engineering data includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section; TBM tunneling parameter data includes cutterhead rotation speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, support shoe cylinder left stroke A6, support shoe cylinder right stroke A7, top shield cylinder left pressure A8, top shield cylinder left stroke A9, top shield cylinder right pressure A10, top shield cylinder right stroke A11, wedge cylinder left pressure A12, wedge cylinder right pressure A13, side shield cylinder left pressure A14, side shield cylinder left stroke A15, side shield cylinder right pressure A16, and side shield cylinder right stroke A17; Step 2: Preprocess the TBM engineering data to obtain processed data; Preprocessing includes unique thermal coding of surrounding rock grade, outlier detection, feature selection, and data balancing. Specifically: Step 2.1: Import the TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Step 2.2: Thermal coding of surrounding rock grades: Class II, IIIa, IIIb, IV and V rock masses are coded as 0, 1, 2, 3 and 4 respectively; Step 2.3: Use PCA principal component analysis to reduce the dimensionality of the TBM tunneling parameter data and map the dimensionality-reduced data to three-dimensional space; Step 2.4: Set the outlier threshold to 10% and use the iForest isolated forest algorithm to remove outliers; Step 2.5: For the data after removing outliers, the RF-RFE random forest-recursive feature removal method is used to further filter features; Step 2.6: Divide the data after feature filtering into training and test sets in a 9:1 ratio. Use the Borderline-SMOTE boundary-sensitive synthetic sampling algorithm to balance the data of each category in the training set to obtain the final data. Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0019] Example 5 The TBM surrounding rock grade prediction method proposed in this embodiment is based on a two-layer cyclic simulated annealing stacking model. Figure 1 As shown, it includes the following steps: Step 1: Collect TBM project data; TBM engineering data includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section; TBM tunneling parameter data includes cutterhead rotation speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, support shoe cylinder left stroke A6, support shoe cylinder right stroke A7, top shield cylinder left pressure A8, top shield cylinder left stroke A9, top shield cylinder right pressure A10, top shield cylinder right stroke A11, wedge cylinder left pressure A12, wedge cylinder right pressure A13, side shield cylinder left pressure A14, side shield cylinder left stroke A15, side shield cylinder right pressure A16, and side shield cylinder right stroke A17; Step 2: Preprocess the TBM engineering data to obtain processed data; Preprocessing includes unique thermal coding of surrounding rock grade, outlier detection, feature selection, and data balancing. Specifically: Step 2.1: Import the TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Step 2.2: Thermal coding of surrounding rock grades: Class II, IIIa, IIIb, IV and V rock masses are coded as 0, 1, 2, 3 and 4 respectively; Step 2.3: Use PCA principal component analysis to reduce the dimensionality of the TBM tunneling parameter data and map the dimensionality-reduced data to three-dimensional space; Step 2.4: Set the outlier threshold to 10% and use the iForest isolated forest algorithm to remove outliers; Step 2.5: For the data after removing outliers, the RF-RFE random forest-recursive feature removal method is used to further filter features; Step 2.6: Divide the data after feature filtering into training and test sets in a 9:1 ratio. Use the Borderline-SMOTE boundary-sensitive synthetic sampling algorithm to balance the data of each category in the training set to obtain the final data. Step 3: Construct evaluation indicators for prediction results; The performance evaluation metrics in step 3 include accuracy (ACC), recall (REC), precision (PRC), score (F1), true positive rate (TPR), and false positive rate (FPR). The expression for accuracy ACC is shown in equation (1); (1); The expression for recall rate REC is shown in equation (2); (2); The expression for accuracy PRC is shown in equation (3); (3); The expression for the fraction F1 is shown in equation (4); (4); The expression for the true rate of return (TPR) is shown in equation (5); (5); The expression for the false positive rate FPR is shown in equation (6); (6); in, TP A true positive is a sample that is actually a positive sample and is correctly classified as a positive sample. TN A true negative is a sample that is actually a negative class and has been correctly classified as a negative class. FP A false positive is a sample that is actually a negative class but is incorrectly classified as a positive class. FN False negatives are samples that are actually positive but were incorrectly classified as negative; ACC is precision; REC is recall; PRC is accuracy; F1 score is the true positive rate; TPR is the false positive rate; FPR is the false positive rate. Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. The model pool includes single models, ensemble learning models, and neural network models; Single models include Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Naive Bayes (NB). Ensemble learning models include Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Class Gradient Boosting (CatBoost), and Adaptive Boosting (AdaBoost). Neural network models include the multilayer perceptron classifier (MLPClassifier) ​​model; The process of constructing the optimization program for the double-loop simulated annealing algorithm is as follows: S41. Initialization and Neighborhood Generation: A solution is randomly initialized using the simulated annealing algorithm (SA), which contains a set of base models and a meta-model. In each iteration, a new solution is generated through a neighborhood operation, which randomly replaces a base model or the meta-model with equal probability. S42. Fitness Evaluation: Select a set of base models and meta-model combinations. The system triggers the inner loop mechanism and uses the SA algorithm to search the hyperparameter space of the base learners and meta-learners under the current combination to find the hyperparameter combination that maximizes the model prediction performance. By evaluating the fitness f(xi) of the hyperparameter combination, it decides whether to accept the new solution. S43. Solution acceptance criterion: The Metropolis classical criterion is adopted. If the fitness of the current solution xi is better than that of the previous solution pi, then pi is updated; if pi is better than the global optimal solution g, then g is updated; if the solution is not improved, the temperature T is reduced through simulated annealing mechanism to continue generating new solutions until the next optimal combination C′ is found. S44. Temperature Update and Termination: Temperature decays according to T←αT. Optimization terminates when the temperature drops to Tmin or the maximum number of iterations is reached, resulting in the final stacked model composition and hyperparameters. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

[0020] Example 6 The TBM surrounding rock grade prediction method based on a two-layer cyclic simulated annealing stacking model proposed in this embodiment includes the following steps: S1, TBM engineering data collection, including obtaining TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Data from a water conveyance tunnel project in Ji'an City, Jilin Province, Northeast China, was used. The tunnel has a total length of 49.56 km, with branch tunnels of 8.44 km and a main tunnel of 41.11 km. The main tunnel's TBM excavation section is 37.68 km long, with a circular cross-section, an excavation diameter of 8.5 m, and a maximum burial depth approaching 1000 m. TBM excavation was carried out using machinery manufactured by Guangzhou Herrenknecht Tunneling Machinery Co., Ltd. TBM excavation parameters include cutterhead speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, and support shoe cylinder left stroke. A6. Right stroke of the support cylinder; A7. Left pressure of the top shield cylinder; A8. Left stroke of the top shield cylinder; A9. Right pressure of the top shield cylinder; A10. Right stroke of the top shield cylinder; A11. Left pressure of the wedge cylinder; A12. Right pressure of the wedge cylinder; A13. Left pressure of the side shield cylinder; A14. Left stroke of the side shield cylinder; A15. Right pressure of the side shield cylinder; A16. Right stroke of the side shield cylinder; A17. According to the preliminary survey data of the tunnel section and the water conservancy and hydropower survey and design specifications, the surrounding rock grade of the tunnel is divided into five grades: Grade II, Grade IIIa, Grade IIIb, Grade IV, and Grade V. Among the five types of surrounding rock, Class II, Class IIIa, and Class IIIb rock account for a relatively large proportion, while Class IV and Class V rock have shorter construction sections. The research data was collected from the underground tunnel section from 35+013.4 to 47+162.5, and the data are the average values ​​of the stable section of the TBM tunneling cycle. A total of 7184 sets of tunneling parameters for the five types of surrounding rock were collected. S2. Data preprocessing, including unique thermal coding of surrounding rock grade, outlier detection, feature selection and data balancing. Specifically, the following steps are included: S21. Import the TBM tunneling parameter data and surrounding rock grade data for each tunneling section, resulting in a total of 7184 data samples; S22. The unique thermal coding process for the surrounding rock grade is applied, with rock masses of type II, IIIa, IIIb, IV, and V coded as 0, 1, 2, 3, and 4, respectively. S23. Principal component analysis (PCA) is used to reduce the dimensionality of the TBM tunneling parameter data and map it to a three-dimensional space. S24. Use the Isolation Forest algorithm (iForest) to remove outliers, with an outlier threshold of 10%. The outlier score distribution of the Isolation Forest (iForest) is as follows: Figure 2 As shown; S25. For the data after removing outliers, the Random Forest-Recursive Feature Elimination (RF-RFE) method is used to further filter features. The performance curve of RF-RFE feature selection is shown below. Figure 3 As shown; S26. The extracted features are normalized using the z-score method. S27. Divide the data after feature filtering into training and test sets in a 9:1 ratio. Use the Borderline-SMOTE algorithm to balance the data of each category in the training set to obtain the final data. The final number of samples in each category in the training set is 2156, and the number of samples in categories II, IIIa, IIIb, IV, and V in the test set are 176, 194, 244, 12, and 21, respectively. S3, Prediction effect evaluation indicators; In step S3, the performance evaluation metrics include accuracy (ACC), recall (REC), precision (PRC), score (F1), true positive rate (TPR), and false positive rate (FPR). The expression for accuracy ACC is shown in equation (1); (1); The expression for recall rate REC is shown in equation (2); (2); The expression for accuracy PRC is shown in equation (3); (3); The expression for the fraction F1 is shown in equation (4); (4); The expression for the true rate of return (TPR) is shown in equation (5); (5); The expression for the false positive rate FPR is shown in equation (6); (6); in, TP A true positive is a sample that is actually a positive sample and is correctly classified as a positive sample. TN A true negative is a sample that is actually a negative class and has been correctly classified as a negative class. FP A false positive is a sample that is actually a negative class but is incorrectly classified as a positive class. FN False negatives are samples that are actually positive but were incorrectly classified as negative; ACC is precision; REC is recall; PRC is accuracy; F1 score is the true positive rate; TPR is the false positive rate; FPR is the false positive rate. AUC refers to the area under the ROC curve plotted with TPR and FPR as the x and y axes; S4. Prediction model construction, model pool construction, construction of a two-layer cyclic simulated annealing algorithm optimization program, and obtaining the final surrounding rock grade prediction model; The model pool contains 11 models, consisting of single models, ensemble learning models, and neural network models. Single models include Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Naive Bayes (NB). Ensemble learning models include Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Class Gradient Boosting (CatBoost), and Adaptive Boosting (AdaBoost). Neural network models include Multilayer Perceptron Classifier (MLPClassifier). The process of constructing the optimization program for the double-loop simulated annealing algorithm is as follows: S41. Initialization and Neighborhood Generation: Simulated Annealing (SA) algorithm first randomly initializes a solution, containing a set of base models and a meta-model. In each iteration, a new solution is generated through a neighborhood operation, which randomly replaces either a base model or the meta-model with equal probability.

[0021] S42. Fitness evaluation: Whenever a combination of base models and meta-models is selected, the system immediately triggers the inner loop mechanism. The SA algorithm is used to search the hyperparameter space of the base learners and meta-learners under the current combination to find the hyperparameter combination that can maximize the model prediction performance, i.e., inner optimization. The system then evaluates the fitness f(xi), i.e. based on the prediction accuracy ACC, to decide whether to accept the new solution. S43. Solution acceptance criterion: The Metropolis classic criterion is adopted. If the fitness of the current solution xi is better than that of the previous solution pi, then pi is updated; if pi is better than the global optimal solution g, then g is updated. If the solution is not improved, the temperature T is reduced through simulated annealing mechanism, i.e., T←αT, to continue generating new solutions until the next optimal combination C′ is found, i.e., outer optimization; S44. Temperature update and termination: The temperature decays according to T←αT. The optimization terminates when the temperature drops to Tmin or the maximum number of iterations is reached, and the final stacked model composition and hyperparameters are obtained. The hyperparameters of the stacked ensemble model are shown in Table 1. Table 1 Hyperparameters of the stacked ensemble model

[0022] S5. Surrounding rock grade prediction and comparison, including using a trained model to predict the surrounding rock grade and comparing and verifying it with similar models. The comparative validation with similar models includes all models in the model pool. The evaluation metrics for the double-layer cyclic simulated annealing stacked ensemble model in this application are summarized in Table 2. The confusion matrix of the model on the test set is shown below. Figure 4 As shown, the ROC curve of the model on the test set is as follows: Figure 5 As shown, the number of misclassifications by the model is as follows: Figure 6 As shown.

[0023] Table 2 Evaluation metrics for the double-layer cyclic simulated annealing stack integration model

[0024] According to the results in Table 1, the classification performance of Class IIIb and Class V surrounding rocks was the most outstanding. For Class IIIb surrounding rocks, the model demonstrated stable and excellent predictive performance, with all evaluation metrics remaining at a high level. Class V surrounding rocks performed perfectly at Precision=1.000, with an F1 score as high as 0.950, but its overall performance was slightly inferior to Class IIIb due to a slightly lower Recall=0.905. The classification performance of Class II and IIIa surrounding rocks was also quite ideal. Although the accuracy and other metrics fluctuated slightly, it still indicated that the model had good adaptability to these categories. In contrast, the classification performance of Class IV surrounding rocks was relatively poor. Although the accuracy and recall were relatively high, the decrease in the F1 score indicated that the classification was more difficult. Overall, the model's average accuracy, precision, and recall were all 0.939 across all categories, showing relatively stable performance. Furthermore, the analysis of the confusion matrix showed that the double-layer cyclic simulated annealing stacked ensemble model had a higher misclassification rate in Class IV surrounding rocks, while the misclassification rate for other categories was relatively low. Meanwhile, the area under the ROC curve (AUC) of the model is greater than 0.9 for all surrounding rock categories, further demonstrating that the two-layer cyclic simulated annealing stacked ensemble model has strong generalization ability and superior predictive performance. It is worth noting that the two-layer cyclic simulated annealing stacked ensemble model only takes 0.291 seconds on the test set, achieving a near-real-time prediction effect in engineering applications.

[0025] By comparing the number of misclassifications of different machine learning models on the test set, the two-layer cyclic simulated annealing stacked ensemble model showed significant advantages in all rock mass categories. Especially in the more challenging Class IV and V rock masses, the two-layer cyclic simulated annealing stacked ensemble model only had 2 misclassifications, demonstrating excellent prediction accuracy. In Class II, IIIa, and IIIb rock masses, the number of misclassifications remained low, further validating the model's strong generalization ability across different categories. This result indicates that single models have inherent limitations when handling rock mass classification tasks with complex characteristics, while the two-layer cyclic simulated annealing stacked ensemble model, through its two-stage ensemble architecture, effectively integrates the complementary features of different base models, thereby achieving higher overall classification reliability and robustness. This demonstrates the applicability and advancement of this modeling strategy in complex geological engineering scenarios. S6. Model interpretation, including using SHAP interpretability analysis to reveal the model's feature contribution; Input feature importance analysis based on SHAP values, such as Figure 7 As shown; Analysis shows that the dominant sensitivity characteristics of different surrounding rock types exhibit significant spatial differentiation: in Class II surrounding rock, the right stroke of the support cylinder (0.103) and the pressure of the support cylinder (0.102) contribute the most, reflecting the sensitivity of hard rock formations to the mechanical state of the support system; in Class IIIa surrounding rock, the synergistic effect of the penetration depth (0.048) and the right pressure of the side shield cylinder (0.062) is highlighted, suggesting the response characteristics of medium-strength rock masses to tunneling efficiency parameters. It is worth noting that in Class IIIb surrounding rock, the right stroke (0.089) and left stroke (0.067) of the hydraulic cylinder together constitute the extreme value of characteristic contribution, which may be related to the dynamic support requirements of complex fractured rock masses. In Class IV surrounding rock, the hydraulic cylinder pressure (0.048) shows the highest contribution. In Class V surrounding rock, the hydraulic cylinder pressure (0.052) and left stroke (0.043) exhibit a synergistic dominant effect. Both of these extremely soft rock types show a significant dependence on the support system parameters, revealing the strong correlation between weak rock masses and dynamic support mechanisms. The differences in the sensitivity of characteristics among different rock types reflect the nonlinear dynamic balance between rock mass conditions and mechanical parameters in the TBM-surrounding rock system. Intelligent identification of surrounding rock is essentially a nonlinear mapping problem in a high-dimensional feature space, and a single parameter cannot fully characterize the complex rock-machine interaction mechanism. The multidimensional decision boundary formed by the dynamic coupling of various tunneling parameters reveals the implicit correlation between geological conditions and construction parameters. The model in this application effectively captures the complex nonlinear relationship between different rock mass mechanical properties and construction parameters.

Claims

1. A method for predicting the surrounding rock grade of TBM based on a two-layer cyclic simulated annealing stacking model, characterized in that, Includes the following steps: Step 1: Collect TBM project data; Step 2: Preprocess the TBM engineering data to obtain processed data; Step 3: Construct evaluation indicators for prediction results; Step 4: Construct a model pool and optimize the double-layer cyclic simulated annealing algorithm to obtain a prediction model. Use the processed data to train the prediction model and obtain the final surrounding rock grade prediction model. Step 5: Based on the prediction effect evaluation index, use the final surrounding rock grade prediction model to predict the surrounding rock grade, and compare and verify with similar models based on the model pool. Step 6: Calculate the feature contribution of the final surrounding rock grade prediction model using the SHAP interpretability analysis model.

2. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 1, characterized in that, The TBM engineering data mentioned in step 1 includes TBM tunneling parameter data and surrounding rock grade data for each tunneling section.

3. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 1, characterized in that, The preprocessing described in step 2 includes unique thermal coding of the surrounding rock grade, outlier detection, feature selection, and data balancing.

4. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 2, characterized in that, The TBM tunneling parameter data includes cutterhead rotation speed A1, cutterhead torque A2, cutterhead thrust A3, penetration depth A4, support shoe cylinder pressure A5, support shoe cylinder left stroke A6, support shoe cylinder right stroke A7, top shield cylinder left pressure A8, top shield cylinder left stroke A9, top shield cylinder right pressure A10, top shield cylinder right stroke A11, wedge cylinder left pressure A12, wedge cylinder right pressure A13, side shield cylinder left pressure A14, side shield cylinder left stroke A15, side shield cylinder right pressure A16, and side shield cylinder right stroke A17.

5. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 3, characterized in that, Step 2 specifically involves: Step 2.1: Import the TBM tunneling parameter data and surrounding rock grade data for each tunneling section; Step 2.2: Thermal coding of surrounding rock grades: Class II, IIIa, IIIb, IV and V rock masses are coded as 0, 1, 2, 3 and 4 respectively; Step 2.3: Use PCA principal component analysis to reduce the dimensionality of the TBM tunneling parameter data and map the dimensionality-reduced data to three-dimensional space; Step 2.4: Set the outlier threshold to 10% and use the iForest isolated forest algorithm to remove outliers; Step 2.5: For the data after removing outliers, the RF-RFE random forest-recursive feature removal method is used to further filter features; Step 2.6: Divide the data after feature filtering into training and test sets in a 9:1 ratio. Use the Borderline-SMOTE boundary-sensitive synthetic sampling algorithm to balance the data of each category in the training set to obtain the final data.

6. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 3, characterized in that, The performance evaluation metrics mentioned in step 3 include accuracy (ACC), recall (REC), precision (PRC), score (F1), true positive rate (TPR), and false positive rate (FPR). The expression for the accuracy ACC is shown in equation (1); （1）； The expression for the recall rate REC is shown in equation (2); （2）； The expression for the accuracy rate PRC is shown in equation (3); （3）； The expression for the fraction F1 is shown in equation (4); （4）； The expression for the true rate TPR is shown in equation (5); （5）； The expression for the false positive rate FPR is shown in equation (6); （6）； in, TP A true positive is a sample that is actually a positive sample and is correctly classified as a positive sample. TN A true negative is a sample that is actually a negative class and has been correctly classified as a negative class. FP A false positive is a sample that is actually a negative class but is incorrectly classified as a positive class. FN False negatives are samples that are actually positive but are incorrectly classified as negative; ACC is precision; REC is recall; PRC is accuracy; F1 score is the true positive rate; TPR is the false positive rate; FPR is the false positive rate.

7. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 3, characterized in that, The model pool mentioned in step 4 includes single models, ensemble learning models, and neural network models; The single model includes the Support Vector Machine (SVM) model, the K-Nearest Neighbors (KNN) model, the Decision Tree (DT) model, and the Naive Bayes (NB) model. The ensemble learning models include Random Forest (RF) model, Gradient Boosting Machine (GBM) model, Extreme Gradient Boosting (XGBoost) model, Lightweight Gradient Boosting Machine (LightGBM) model, Class Gradient Boosting (CatBoost) model, and Adaptive Boosting (AdaBoost) model. The neural network model includes the multilayer perceptron classifier (MLPClassifier) ​​model.

8. The method for predicting the surrounding rock grade of TBM based on a double-layer cyclic simulated annealing stacking model according to claim 3, characterized in that, The construction process of the double-layer cyclic simulated annealing algorithm optimization program in step 4 is as follows: S41. Initialization and Neighborhood Generation: A solution is randomly initialized using the simulated annealing algorithm (SA), which contains a set of base models and a meta-model. In each iteration, a new solution is generated through a neighborhood operation, which randomly replaces a base model or the meta-model with equal probability. S42. Fitness Evaluation: Select a set of base models and meta-model combinations. The system triggers the inner loop mechanism and uses the SA algorithm to search the hyperparameter space of the base learners and meta-learners under the current combination to find the hyperparameter combination that maximizes the model prediction performance. By evaluating the fitness f(xi) of the hyperparameter combination, it decides whether to accept the new solution. S43. Solution acceptance criterion: The Metropolis classic criterion is adopted. If the fitness of the current solution xi is better than that of the previous solution pi, then pi is updated; if pi is better than the global optimal solution g, then g is updated. If the solution is not improved, the temperature T is lowered by simulated annealing to continue generating new solutions until the next optimal combination C′ is found. S44. Temperature Update and Termination: The temperature decays according to T←αT. The optimization terminates when the temperature drops to Tmin or the maximum number of iterations is reached, and the final stacked model composition and hyperparameters are obtained.

Images