A Deep Learning-Based High-Precision Measurement and Control Method and System for Shield Tail Gap
By combining Online PDNN and NSGA-II methods, high-precision measurement and control of the shield tail gap were achieved, solving the problem that existing shield tail gap measurement methods cannot effectively control the gap, thus improving the safety and efficiency of tunnel construction.
Patent Information
- Application Number
- CN202410266325.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-08
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-03-08
AI Technical Summary
Existing methods for measuring the shield tail gap cannot effectively control the size of the shield tail gap, and have high requirements for the quality and quantity of source data. Traditional deep learning methods have low robustness and lack physical interpretability.
A high-precision measurement and control method for shield tail gap based on deep learning is adopted. Combining Online PDNN and NSGA-II, the deep neural network is integrated through online learning and physical constraints to achieve accurate estimation and multi-objective optimization of shield tail gap.
It improved the accuracy and control of shield tail gap measurement, reduced the shield tail gap value, avoided contact and squeezing between the shield tail and the lining, and improved the safety and efficiency of tunnel construction.
Smart Images

Figure CN118223901B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tunnel boring machine safety technology, and in particular relates to a high-precision measurement and control method and system for the shield tail gap based on deep learning. Background Technology
[0002] Tunnel boring machines (TBMs), as advanced tunneling equipment, are widely used in subway tunnel excavation. However, during tunnel excavation, if the tail shield clearance is too small, excessive compression between the tail shield and the lining segments can reduce the excavation speed, accelerate the wear of the tail shield sealing brush, and even cause tunnel leakage or ground subsidence. Therefore, adopting effective methods to control the tail shield clearance is a crucial step in ensuring the safe construction of tunnel boring machines.
[0003] Currently, research on shield tail clearance mainly focuses on its measurement. Commonly used methods include theoretical analysis, manual measurement, contact measurement, and non-contact measurement. However, these methods fail to explore the potential causal relationships between shield tail clearance and other TBM operating parameters, thus hindering shield tail clearance optimization and control. To address this issue, it is necessary to explore deep learning methods based on causal analysis for shield tail clearance measurement and optimization. However, traditional deep learning methods require a large quantity and quality of training data and suffer from low robustness and lack of physical interpretability. Therefore, there is an urgent need to propose an advanced method for shield tail clearance measurement and control to meet these challenges.
[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: traditional gap measurement methods cannot be used to implement shield tail gap control operations and have high requirements for source data. Summary of the Invention
[0005] To address the problems existing in the prior art, this invention provides a high-precision measurement and control method and system for shield tail gap based on deep learning. It integrates Online PDNN and NSGA-II to measure and control the shield tail gap, aiming to achieve online prediction and real-time control of the shield tail gap during TBM excavation operations, and effectively avoid tunnel safety accidents caused by excessively small shield tail gaps.
[0006] This invention is implemented as follows: a high-precision measurement and control method for shield tail gap based on deep learning, comprising:
[0007] S1, Calculation of tail gap: Based on the theory of minimum tail gap calculation, assuming that the lining ring is assembled in an ideal state, there are six minimum tail gap calculation models.
[0008] S2. First, the derived equation for calculating the shield tail gap at any position is embedded as a physical constraint into the DNN framework to obtain the PDNN model. Then, the PDNN is trained online so that it can adaptively update the model hyperparameters according to the input observation stream, thereby achieving accurate estimation of the shield tail gap.
[0009] S3. Five statistical indicators are selected to evaluate the performance of the established Online PDNN model; the Shapley Additive exPlanations method is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features.
[0010] S4. First, a Pareto solution set for the MOO problem is searched using non-dominated sorting. Then, the crowding distance of each solution is calculated to maintain the diversity and uniform distribution of the solution set. Finally, a crowding comparison operator is used to compare two solutions, where the solution with the lower rank and the larger crowding distance are tended to be selected. The above steps are repeated until the maximum number of iterations is reached, ultimately generating the Pareto front.
[0011] Furthermore, S1 specifically includes: based on the minimum shield tail clearance calculation theory, assuming the lining ring assembly is in an ideal state, there are six minimum shield tail clearances (δ). min The calculation model is given by formula (1).
[0012]
[0013] In the formula, δ0 is the initial shield tail gap; δ1 is the change in shield tail gap caused by the change in pitch angle α; δ2 is the change in shield tail gap caused by the tunnel boring machine excavating along the route of deflection angle η and tunnel curve radius R. Their mathematical expressions are shown in formulas (2) and (3), respectively.
[0014] δ1=l tanα (2)
[0015]
[0016] In the formula, l is the length of the shield tail covering segment; α is the pitch angle; R is the radius of the tunnel curve; and D is the outer diameter of the lining ring.
[0017] The three concentric circles with center O represent a lining ring with radius D / 2 and a lining ring with radius (D / 2+δ). min ) and (D / 2+δ max The two auxiliary circles are O'; the ellipse centered at O' is the 1-1 section of the shield shell. F1 and F2 are the two foci of the ellipse; the major and minor axes of the ellipse are 2R / cosα and 2R, respectively. Based on the definition of an ellipse and the law of cosines, the tail clearance (δ) for any position can be derived. pThe physical equation for the calculation is formula (4);
[0018]
[0019] In the formula, β is the horizontal angle, which represents the position where the shield tail gap needs to be calculated; |OF1| and |OF2| can be calculated by formulas (5) and (6) respectively;
[0020]
[0021]
[0022] In the formula, |OF1| and |OF2| are the distances from the center O of the lining ring to the foci F1 and F2 of the ellipse, respectively.
[0023] Furthermore, the construction of the PDNN model in S2 specifically includes:
[0024] The derived equation for calculating the shield tail gap at any position is integrated as a physical law into the loss function of the DNN model to constrain the model's training process. This helps to compensate for the shortcomings of purely data-driven DNN models, which rely excessively on data and lack reasonable explanations of the internal mechanisms of the physical system. According to the PDNN framework used for shield tail gap measurement, the loss function of the PDNN model coupled with the physical mechanism consists of the observation loss MSE and the physical loss MSE. p It consists of two parts, and its mathematical expression is formula (7):
[0025]
[0026] In the formula, λ1 and λ2 represent MSE and MSE, respectively. p The weight; δ r δ p and These represent the observed, theoretical, and predicted values of the shield tail gap, respectively.
[0027] Furthermore, the online update mechanism of the PDNN model in S2 specifically includes:
[0028] To enhance the practicality and generalization of the PDNN model, an online learning approach is adopted to train the model. Specifically, the data used to develop the Online PDNN model is gradually updated as the excavation operation progresses, and the model's hyperparameters are dynamically adjusted as the dataset is updated. In the online prediction concept, it is assumed that the initial dataset contains N0 samples. Then, N1 samples are selected from the N0 samples as training data to train the model, and the remaining N2 samples are used as test data to verify the model's performance. The BO algorithm is used to automatically search for the optimal combination of hyperparameters for the Online PDNN model.
[0029] Furthermore, the evaluation of the Online PDNN model in S3 specifically includes:
[0030] To evaluate the performance of the established Online PDNN model, five commonly used statistical metrics were selected, including the coefficient of determination (R²). 2 R20, a20_index, variance interpretation (VAF), mean absolute error (MAE), and root mean square error (RMSE) are expressed mathematically in formulas (8)-(12). Typically, R20... 2 The closer the values of VAF and a20_index are to 1, and the closer the values of MAE and RMSE are to 0, the higher the prediction accuracy of the model.
[0031]
[0032]
[0033]
[0034]
[0035]
[0036] In the formula, n is the total number of samples; and y i These are the predicted value and the observed value, respectively. It is the average of all samples; m 20 It is the number of samples with an error not exceeding ±20% of the observed value.
[0037] Furthermore, the interpretability of the Online PDNN model in S3 specifically includes:
[0038] To understand the decision-making strategy of the developed Online PDNN model, this study uses the Shapley Additive exPlanations method to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features. The mathematical expression of the Shapley value is shown in formula (13):
[0039]
[0040] In the formula, Z is the set of all input features with dimension M; S is a subset of the input features with dimension |S|; f x (S)=E[f(x)|x S The mean of the samples is calculated based on the training subset S.
[0041] SHAP analysis, based on the additive eigenvalue attribution method, can simplify complex models into linear functions:
[0042]
[0043] In the formula, φ0 is the average of the estimated values; z' j ∈{0,1} M Describes whether feature i is observed, z' j =0 and z' j =1 indicates that the feature either does not exist or exists, respectively.
[0044] Furthermore, S4 specifically includes: In the minimum MOO problem, four components need to be considered, including a k-dimensional objective vector F(x), m inequality constraints, n equality constraints, and a p-dimensional decision variable x, the mathematical representation of which is shown in formulas (15)-(17):
[0045] Minimize F(x)=(f1(x),f2(x),...,f k (x)), k≥2 (15)
[0046]
[0047] x = (x1, x2, ..., x p )∈Ω (17)
[0048] In the formula, f i (x) is the i-th objective function; Ω is the decision space with p independent variables;
[0049] The main steps for optimizing the shield tail gap based on Online PDNN and NSGA-II are as follows: First, a Pareto solution set for the MOO problem is searched using non-dominated sorting. Then, the crowding distance of each solution is calculated to maintain the diversity and uniform distribution of the solution set. Finally, the crowding comparison operator is used to compare two solutions, where the solution with the lower rank and the solution with the larger crowding distance are often selected. The above steps are repeated until the maximum number of iterations is reached, and finally, the Pareto front is generated. In order to determine the optimal solution from the Pareto front, the distance between the Pareto optimal solution vector and the ideal solution vector is calculated using formula (18), and the solution with the smallest distance is selected as the optimal solution.
[0050]
[0051] In the formula, It is the value of the i-th objective function in the j-th Pareto solution; i is the ideal value of the i-th objective function; k is the number of all objective functions.
[0052] Another objective of this invention is to provide a deep learning-based high-precision measurement and control system for shield tail clearance, which implements the aforementioned deep learning-based high-precision measurement and control method, comprising:
[0053] The shield tail gap calculation module is used to calculate the minimum shield tail gap based on the minimum shield tail gap calculation theory, assuming that the lining ring is assembled in an ideal state, and there are six minimum shield tail gap calculation models.
[0054] The PDNN model acquisition module is used to embed the derived equation for calculating the shield tail gap at any position as a physical constraint into the DNN framework to obtain the PDNN model.
[0055] The PDNN model training module is used to train the PDNN in an online learning manner, enabling it to adaptively update the model hyperparameters based on the input observation stream, thereby achieving accurate estimation of the shield tail gap.
[0056] The PDNN model evaluation module is used to select five statistical indicators to evaluate model performance.
[0057] The PDNN model interpretability analysis module is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features using the Shapley Additive exPlanations method.
[0058] The shield tail gap MOO execution module is used to execute the shield tail gap MOO.
[0059] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, causing the processor to perform the steps of the deep learning-based high-precision measurement and control method for shield tail gap.
[0060] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based high-precision measurement and control method for shield tail gap.
[0061] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned deep learning-based high-precision measurement and control system for shield tail gap.
[0062] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:
[0063] First, addressing the shortcomings of existing shield tail gap measurement methods, such as the inability to effectively control the size of the shield tail gap and the high requirements for the quality and quantity of source data, this invention proposes a high-precision shield tail gap measurement and control method based on deep learning. The specific research and development process and the beneficial effects achieved are described below:
[0064] 1. Discovery of a Theory for Calculating Tail Clearance at Arbitrary Locations. During shield tunneling construction, on-site personnel often focus on the size of the tail clearance at different locations. However, existing theories for calculating the minimum tail clearance can only estimate the minimum tail clearance value and cannot provide the location where the minimum tail clearance occurs, thus failing to meet the needs of on-site operations. Therefore, this invention proposes an improved theory for calculating tail clearance, establishing the physical relationship between the tail clearance and parameters such as the pitch angle and horizontal angle, providing a basis for theoretically calculating the tail clearance at different locations.
[0065] 2. Development of Online PDNN Model. Since theoretical analysis methods cannot consider the influence of operating parameters such as cylinder stroke and shield tail deviation on the shield tail gap, and in addition, pure data-driven machine learning or deep learning models have black box attributes and cannot provide users with potential interpretable information about the physical mechanism, and the trained proxy model has low robustness. Therefore, this invention aims to develop a deep learning model guided by physical information. Considering that there is little field data available in the early stage of the project, this invention further coupled an online update mechanism on the basis of the PDNN model to construct an Online PDNN model. The specific development process is as follows: (1) Acquisition and preprocessing of source data. This invention selects monitoring data from a tunnel construction project example, including pitch angle, horizontal displacement and vertical displacement of shield tail and cut, corresponding displacement trends, total thrust, total extrusion pressure, tunneling speed, soil chamber pressure, cutterhead speed and torque, cylinder stroke, etc. The collected monitoring data often have duplicate values, missing values and outliers, so it is necessary to clean the data. Moreover, in order to eliminate the attribute differences between feature variables, the minimum-maximum normalization method is used to standardize the dataset. (2) The implementation of the Online PDNN model structure. The key to developing the Online PDNN model is: (a) embedding the derived theoretical model for calculating the shield tail gap at any position as a physical constraint into the framework of the DNN loss function. That is, the loss function of the Online PDNN model consists of the observation loss MSE and the physical loss MSE. pIt consists of two parts. (b) The available dataset for the Online PDNN model gradually expands as the number of tunnel excavation rings increases, and the optimal hyperparameters of the model are dynamically updated as the dataset is updated. That is, assuming the initial dataset contains N0 samples, then N1 samples are selected from the N0 samples as training data to train the model, and the remaining N2 samples are used as test data to verify the model performance. If the model's evaluation results meet the acceptable standard, the model is applied to the prediction of a new sample N3. If the prediction result is accurate, the model's hyperparameters are not updated; otherwise, the Online PDNN model will be retrained using the updated dataset (N0+N3). The optimal hyperparameter combination of the model is automatically searched and determined by a Bayesian optimization (BO) algorithm based on a Gaussian process as a surrogate model and the expected increment as the acquisition function. The optimal hyperparameter combination obtained based on BO can always minimize the loss function and maximize the model's prediction accuracy during training.
[0066] 3. Integration of Online PDNN Model and NSGA-II Optimization Algorithm. The main reason existing methods for measuring shield tail gap cannot control it is that they do not consider the causal relationship between the shield tail gap and other key operating parameters, thus preventing the application of optimization algorithms to solve the multi-objective optimization problem of the shield tail gap. However, controlling the shield tail gap is crucial for the smooth progress of tunnel excavation. Therefore, this invention integrates the NSGA-II optimization algorithm into the developed causal analysis-oriented Online PDNN model to achieve multi-objective optimization of the shield tail gap. Specifically, minimizing the shield tail gap at different locations is taken as the optimization objective, and the operating parameters of the tunnel boring machine (TBM) are used as adjustable decision variables. By calculating the overall optimization degree of the shield tail gap under different adjustment ranges of the decision variables, the optimal variation space of the adjustable parameters is determined, providing valuable guidance for controlling the shield tail gap during TBM tunneling.
[0067] 4. The beneficial effects of the developed Online PDNN and NSGA-II hybrid model. Taking the application of a TBM in the construction of a tunnel under construction as an example, the Online PDNN model, when predicting the shield tail gap at different locations, achieves R... 2The proposed OnlinePDNN model achieves a score of 0.9 or higher, a VAF of over 90%, an a20_index value close to 1, and RMSE and MSE both less than 4, indicating high prediction accuracy and strong robustness. When the hybrid model integrating Online PDNN and NSGA-II is used to find the optimal solution for the shield tail gap MOO, the optimal variation space of the adjustable parameters is found to be within the range of 20% of the independent values. Furthermore, under this constraint, the shield tail gap is reduced by an average of 30.87%. In conclusion, the hybrid model integrating Online PDNN and NSGA-II proposed in this invention can reliably and stably predict and control the shield tail gap, improving the dynamic safety of tunnel excavation.
[0068] Secondly, the high-precision measurement and control method and system for shield tail gap based on deep learning proposed in this invention does not rely on special equipment such as industrial cameras and ultrasonic rangefinders for the acquisition and measurement of shield tail gap data. This means that the method proposed in this invention will be more practical in actual tunnel engineering projects.
[0069] The present invention proposes a high-precision measurement and control method for shield tail clearance based on deep learning, which can achieve an R0.9 or higher. 2 The score accurately estimates the size of the tail gap. Furthermore, the optimal variation space of the adjustable parameters of the tunnel boring machine can be determined by the MOO (Mean Offset of Optimization) to be within 20% of the self-value. Under this constraint, the tail gap value can be reasonably and significantly reduced, avoiding contact compression between the lining and the shield shell and wear of the tail brush, thus improving the safety level of tunnel construction.
[0070] The present invention proposes a high-precision measurement and control method for shield tail gap based on deep learning, which can not only reduce the dependence of data-driven models on the size of the dataset and improve the interpretability and robustness of the model, but also provide effective suggestions for on-site personnel to control the shield tail gap through the decision scheme generated by MOO.
[0071] Third, the technical solution of this invention fills a technological gap in the industry both domestically and internationally: The technical solution of this invention fills a technological gap in the implementation of shield tail gap control work both domestically and internationally. Most existing research results on shield tail gap focus on its measurement, without considering its optimization and control. To address this research gap, this invention proposes a high-precision measurement and control method for shield tail gap based on deep learning. The innovation of this method can be summarized in two key aspects: (a) integrating the derived physical equations for calculating shield tail gap at any position and the online update mechanism into a DNN framework to generate an Online PDNN model capable of accurately predicting the shield tail gap; (b) combining the developed online PDNN model with the NSGA-II algorithm to achieve intelligent control of the shield tail gap at any position.
[0072] Fourth, the deep learning-based high-precision measurement and control method for shield tail gap provided by this invention demonstrates significant technological advancements, specifically in the following aspects:
[0073] 1. High-precision shield tail clearance calculation: This method is based on the minimum shield tail clearance calculation theory. By assuming that the lining ring assembly is in an ideal state, six minimum shield tail clearance calculation models are derived. The introduction of this theoretical model and calculation method improves the accuracy of shield tail clearance calculation, providing a more accurate data foundation for subsequent measurement and control.
[0074] 2. Combination of Deep Learning and Physical Constraints: This method embeds the derived equation for calculating the shield tail gap at any location as a physical constraint into a deep neural network (DNN) framework, resulting in a physical deep neural network (PDNN) model. This combination enables the model to adaptively update its hyperparameters based on the input observation flow, thereby achieving accurate estimation of the shield tail gap. This innovative method not only improves measurement accuracy but also enhances the model's adaptability and robustness.
[0075] 3. Model Performance Evaluation and Feature Importance Analysis: This method evaluates the performance of the established Online PDNN model by selecting five statistical indicators and uses the Shapley Additive exPlanations method to analyze the marginal contribution of each input feature to the model, determining the ranking of the importance of the input features. This evaluation and analysis method not only helps to understand the performance of the model but also provides valuable reference information for model optimization and improvement.
[0076] 4. Handling Multi-Objective Optimization Problems: In controlling the shield tail gap, multiple optimization objectives often need to be considered simultaneously. This method utilizes non-dominated sorting to search for the Pareto solution set of the multi-objective optimization (MOO) problem, and maintains the diversity and uniform distribution of the solution set by calculating the crowding distance of each solution. This approach can find a balance among multiple optimization objectives, making the control of the shield tail gap more comprehensive and effective.
[0077] The deep learning-based high-precision measurement and control method for shield tail gap provided by this invention achieves significant technological advancements through the introduction of high-precision computational models, the combination of deep learning and physical constraints, model performance evaluation and feature importance analysis, and the handling of multi-objective optimization problems. These advancements not only improve the measurement accuracy and control effect of shield tail gap but also provide valuable reference and inspiration for the technological development of related fields. Attached Figure Description
[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0079] Figure 1 This is a flowchart of a high-precision measurement and control method for shield tail gap based on deep learning provided in an embodiment of the present invention;
[0080] Figure 2 This is a structural diagram of a high-precision measurement and control system for shield tail gap based on deep learning provided in an embodiment of the present invention;
[0081] Figure 3 Here is a schematic diagram of the shield tail gap calculation model provided in this embodiment of the invention: (a) TBM excavates along a straight line, (b) TBM excavates in pitch, (c) TBM excavates along a curve, (d) TBM excavates upward and along a curve (δ1<δ2), (e) TBM excavates upward and along a curve (δ1>δ2), (f) TBM excavates downward along a curve.
[0082] Figure 4 This is a schematic diagram of shield tail clearance calculation provided in an embodiment of the present invention;
[0083] Figure 5 This is a schematic diagram of the PDNN algorithm framework for shield tail gap measurement provided in an embodiment of the present invention;
[0084] Figure 6 This is a schematic diagram of the development process of the Online PDNN model provided in the embodiments of the present invention;
[0085] Figure 7 This is a schematic diagram of the shield tail gap measurement deviation of the Online PDNN method provided in the embodiments of the present invention on different data; (a) training data, (b) test data;
[0086] Figure 8 This is a schematic diagram illustrating the performance evaluation of the Online PDNN model on different datasets after each dataset update, provided in this embodiment of the invention; (a) training data, (b) test data;
[0087] Figure 9 These are schematic diagrams of SHAP analysis of the Online PDNN model provided in this embodiment of the invention; (a) heatmap, (b) dependency graph;
[0088] Figure 10 These are schematic diagrams comparing optimized values with original values under different constraints provided in the embodiments of the present invention; (a) y1, (b) y2, (c) y3;
[0089] Figure 11 This is a schematic diagram of the Pareto front of MOO when the variation space of the adjustable decision variable is equal to the range above and below its own value, provided in the embodiments of the present invention; (a) 10%, (b) 20%, (c) 30%, (d) 40%, (e) 50%. Detailed Implementation
[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0091] To address the problems existing in the prior art, the present invention provides a method and system for high-precision measurement and control of shield tail gap based on deep learning. The present invention will be described in detail below with reference to the accompanying drawings.
[0092] S1: Calculation of shield tail clearance
[0093] Assuming the lining ring is assembled under ideal conditions, six calculation models for the minimum shield tail gap are proposed. These models calculate the shield tail gap based on factors such as the geometry of the lining ring, pitch angle, and tunnel curve radius.
[0094] S2: Physically Constrained Embedded Deep Neural Network (PDNN)
[0095] PDNN Model Construction: The shield tail gap calculation equation is embedded as a physical constraint into a deep neural network (DNN) framework to form a PDNN model. This model is trained online and can adaptively update the model hyperparameters to accurately estimate the shield tail gap.
[0096] S3: Model Performance Evaluation
[0097] Performance Evaluation: Model performance was evaluated using five statistical metrics. The Shapley Additive Explanations (SHAP) method was used to analyze the marginal contribution of each input feature to the model, thereby determining the ranking of feature importance.
[0098] S4: Multi-Objective Optimization (MOO)
[0099] Solution search: Search for Pareto solutions to the multi-objective optimization problem by non-dominated sorting, calculate the crowding distance of each solution to maintain the diversity and uniform distribution of the solution set, and select the best solution using the crowding comparison operator. Repeat this process until the maximum number of iterations is reached to generate the Pareto front.
[0100] Changes in tail clearance: δ0 is the initial tail clearance, and δ1 and δ2 are the changes in tail clearance caused by changes in pitch angle (α) and tunnel curve radius (R), respectively.
[0101] Physical equation: Using the definition of an ellipse and the law of cosines, the formula for calculating the shield tail gap (δp) at any position is derived, where β is the horizontal angle and represents the position where the shield tail gap needs to be calculated.
[0102] This method integrates physical models, deep learning technology, and multi-objective optimization strategies to achieve high-precision measurement and control of the shield tail gap during shield tunnel construction, thereby improving the safety and efficiency of tunnel construction.
[0103] like Figure 1 As shown in the embodiments of the present invention, the high-precision measurement and control method for shield tail gap based on deep learning includes four main steps: exploration of the physical mechanism of shield tail gap calculation, development of the Online PDNN model, evaluation and interpretation of the model, and MOO of the shield tail gap. The specific steps are as follows:
[0104] 1. Physical mechanism for calculating shield tail gap
[0105] Based on the theory of minimum shield tail clearance calculation, assuming the lining ring assembly is in an ideal state, there are six calculation models for the minimum shield tail clearance (such as...). Figure 3 (a)-(f) are shown), and their mathematical expressions are shown in formula (1).
[0106]
[0107] In the formula, δ0 is the initial shield tail gap; δ1 and δ2 are the changes in shield tail gap caused by changes in pitch angle and tunnel curve radius, respectively, and their mathematical expressions are shown in formulas (2) and (3).
[0108] δ1=l tanα (2)
[0109]
[0110] In the formula, l is the length of the shield tail covering segment; α is the pitch angle; R is the radius of the tunnel curve; and D is the outer diameter of the lining ring.
[0111] Yizi Figure 3 Taking (f) as an example, the basic principle of shield tail clearance calculation is as follows: Figure 4 As shown. The three concentric circles with center O represent a lining ring with radius D / 2 and a lining ring with radius (D / 2 + δ). min ) and (D / 2+δ maxTwo auxiliary circles are defined. The ellipse centered at O' is section 1-1 of the shield shell. F1 and F2 are the two foci of the ellipse. The major and minor axes of the ellipse are 2R / cosα and 2R, respectively. Based on the definition of an ellipse and the law of cosines, the tail clearance (δ) for any position can be derived. p The physical equation for the calculation is formula (4).
[0112]
[0113] In the formula, β is the horizontal angle, which represents the position where the shield tail gap needs to be calculated; |OF1| and |OF2| can be calculated by formulas (5) and (6) respectively.
[0114]
[0115]
[0116] In the formula, |OF1| and |OF2| are the distances from the center O of the lining ring to the foci F1 and F2 of the ellipse, respectively.
[0117] 2. Development of Online PDNN Model
[0118] The basic idea behind developing an online PDNN model is as follows: First, the derived equation for calculating the shield tail gap at any position is embedded as a physical constraint into a DNN framework to obtain the PDNN model. Then, the PDNN is trained online, enabling it to adaptively update the model's hyperparameters based on the input observation stream, thereby achieving accurate estimation of the shield tail gap. The specific steps are as follows:
[0119] (1) Construction of PDNN model
[0120] The PDNN model design concept adopted in this invention integrates the derived equation for calculating the shield tail gap at any position as a physical law into the loss function of the DNN model to constrain the model's training process. This helps to compensate for the shortcomings of purely data-driven DNN models, which rely excessively on data and lack reasonable explanations of the internal mechanisms of the physical system. Figure 5 As shown in the PDNN framework for shield tail gap measurement, the loss function of the PDNN model after coupling the physical mechanism consists of the observation loss (mean squared error, MSE) and the physical loss (mean squared error, MSE). p It consists of two parts, and its mathematical expression is formula (7):
[0121]
[0122] In the formula, λ1 and λ2 represent MSE and MSE, respectively. p The weight; δ r δ p and These represent the observed, theoretical, and predicted values of the shield tail gap, respectively.
[0123] (2) Online update mechanism of PDNN model
[0124] To enhance the practicality and generalization ability of the PDNN model, this study employs online learning to train the model. Specifically, the data used to develop the Online PDNN model is gradually updated as the excavation operation progresses, and the model's hyperparameters are dynamically adjusted as the dataset is updated. Figure 6 The development process of an Online PDNN model is presented. In the online prediction concept, it is assumed that the initial dataset contains N0 samples. Then, N1 samples are selected from the N0 samples as training data to train the model, and the remaining N2 samples are used as test data to verify the model performance. If the loss function can be minimized on the currently set hyperparameter combination θ, it indicates that the current hyperparameter combination θ is optimal. Otherwise, a hyperparameter optimization algorithm is needed to find the optimal combination of hyperparameters. Bayesian optimization (BO) considers prior knowledge to select the parameter configuration for the next iteration, which has significant advantages in handling complex global optimization problems. Therefore, this study uses the BO algorithm to automatically search for the optimal hyperparameter combination of the Online PDNN model.
[0125] 3. Model Evaluation and Interpretability
[0126] (3) Evaluation of Online PDNN Model
[0127] To evaluate the performance of the established Online PDNN model, five commonly used statistical metrics were selected, including the coefficient of determination (R²). 2 R20, a20_index, variance interpretation (VAF), mean absolute error (MAE), and root mean square error (RMSE) are expressed mathematically in formulas (8)-(12). Typically, R20... 2 The closer the values of VAF and a20_index are to 1, and the closer the values of MAE and RMSE are to 0, the higher the prediction accuracy of the model.
[0128]
[0129]
[0130]
[0131]
[0132]
[0133] In the formula, n is the total number of samples; and y i These are the predicted value and the observed value, respectively. It is the average of all samples; m 20 It is the number of samples with an error not exceeding ±20% of the observed value.
[0134] (4) Interpretability of Online PDNN Model
[0135] To understand the decision-making strategy of the developed Online PDNN model, this study uses the Shapley Additive exPlanations (SHAP) method to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features. The mathematical expression for the Shapley value is given in formula (13):
[0136]
[0137] In the formula, Z is the set of all input features with dimension M; S is a subset of the input features with dimension |S|; f x (S)=E[f(x)|x S ] is the sample mean calculated based on the training subset S.
[0138] SHAP analysis, based on the additive eigenvalue attribution method, can simplify complex models into linear functions:
[0139]
[0140] In the formula, φ0 is the average of the estimated values; z' j ∈{0,1} M Describes whether feature i is observed, z' j =0 and z' j =1
[0141] These indicate whether the feature is absent or present, respectively.
[0142] 4. MOO at the shield tail gap
[0143] To avoid excessive compression between the shield shell and the concrete lining, this invention proposes a novel method integrating OnlinePDNN and NSGA-II for performing the Minimum Motion Omission (MOO) of the shield tail gap. The Minimum Motion Omission problem requires consideration of four components: a k-dimensional objective vector F(x), m inequality constraints, n equality constraints, and a p-dimensional decision variable x, mathematically represented by equations (15)-(17):
[0144] Minimize F(x)=(f1(x),f2(x),...,fk (x)), k≥2 (15)
[0145]
[0146] x = (x1, x2, ..., x p )∈Ω (17)
[0147] In the formula, f i (x) is the i-th objective function; Ω is the decision space of p independent variables.
[0148] The main steps for optimizing the shield tail gap based on Online PDNN and NSGA-II are as follows: First, a Pareto solution set for the MOO problem is searched using non-dominated sorting. Then, the crowding distance of each solution is calculated to maintain the diversity and uniform distribution of the solution set. Finally, the crowding comparison operator is used to compare two solutions, where the solution with the lower rank (when the ranks of the two solutions are different) and the solution with the larger crowding distance (when the two solutions are located on the same front) are often selected. The above steps are repeated until the maximum number of iterations is reached, and finally, a Pareto front is generated. In order to determine the optimal solution from the Pareto front, the distance between the Pareto optimal solution vector and the ideal solution vector (composed of the optimal solutions of all objective functions) is calculated using formula (18), and the solution with the smallest distance is selected as the optimal solution.
[0149]
[0150] In the formula, It is the value of the i-th objective function in the j-th Pareto solution; i is the ideal value of the i-th objective function; k is the number of all objective functions.
[0151] like Figure 2 As shown, the deep learning-based high-precision measurement and control system for shield tail clearance includes:
[0152] The shield tail gap calculation module is used to calculate the minimum shield tail gap based on the minimum shield tail gap calculation theory, assuming that the lining ring is assembled in an ideal state, and there are six minimum shield tail gap calculation models.
[0153] The PDNN model acquisition module is used to embed the derived equations for calculating the shield tail gap at any position as physical constraints into the DNN framework to obtain the PDNN model.
[0154] The PDNN model training module is used to train the PDNN in an online learning manner, enabling it to adaptively update the model hyperparameters based on the input observation stream, thereby achieving accurate estimation of the shield tail gap.
[0155] The PDNN model evaluation module is used to select five statistical indicators to evaluate model performance.
[0156] The PDNN model interpretability analysis module is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features using the Shapley Additive exPlanations method.
[0157] The shield tail gap MOO execution module is used to execute the shield tail gap MOO.
[0158] Reference Figure 3 As shown, this embodiment provides a shield tail gap measurement and control technology based on Online PDNN and NSGA-II. The case study is a TBM application in a tunnel construction project. The main steps include:
[0159] Step 1: Selection and Preprocessing of Model Variables. To accurately predict the tail clearance size, 25 TBM mechanical parameters closely related to the tail clearance were used as feature variables of the Online PDNN model. These include: pitch angle (x1), horizontal displacement trend (x2), cutterhead horizontal deviation (x3), tail clearance horizontal deviation (x4), cutterhead vertical deviation (x5), tail clearance vertical deviation (x6), vertical displacement trend (x7), soil chamber pressure (x8), tunneling speed (x9), cutterhead rotation speed (x10), cutterhead torque (x11), total thrust (x12), total extrusion pressure (x13), and cylinder stroke (x14-x25). In the application case selected in this invention, the tail clearance was measured at three locations: directly above, to the upper left (β = 135°), and to the upper right (β = 45°) of 553 lining rings. Therefore, these were all used as output targets of the Online PDNN model, denoted as y1, y2, and y3, respectively. Detailed information on the feature variables and output targets is shown in Table 1. Because the values and dimensions of each feature variable are different, the model is difficult to converge. Therefore, before constructing the prediction model, the prepared dataset is preprocessed using the Min-Max normalization method. That is, without changing the original data distribution, the original data is mapped to the interval [0,1] using Equation (19). In addition, considering that the shield attitude of the current ring will be affected by the shield attitude of the adjacent ring, when evaluating the shield tail gap of the current ring, not only the feature data of the current ring should be considered, but also the features and target data of the adjacent rings should be input into the model as feature variables. Therefore, the final number of samples used to construct the model is 552, and the data composition of each sample consists of 50 feature variables and 3 target variables.
[0160]
[0161] In the formula, x pFor normalized values, min(x) and max(x) represent the minimum and maximum values of variable x in the training dataset, respectively.
[0162] Table 1. Data Explanation of Selected Tunneling Machine Parameters and Performance Indicators
[0163]
[0164]
[0165] Step 2: Construction of the shield tail gap prediction model. According to formula (4), the size of the shield tail gap is related to the parameter β representing the azimuth. Therefore, the Online PDNN models used to predict the shield tail gap at different locations have different loss functions, which means that separate prediction models need to be designed for y1, y2 and y3 respectively. The specific development process is as follows: take the first 501 samples from the 552 preprocessed samples, randomly select 498 of them as training data, and use the remaining samples as test data. After the initial model training is completed, the model is used to predict new samples with ring numbers 502 to 504. If the model performs well, the hyperparameters are not updated; otherwise, the new samples need to be integrated into the initial dataset and divided into training data and test data, and the model is retrained and evaluated. Repeat the above process until no new samples are available. Considering the complexity of the input features and the size of the sample set, the initial parameters of the online PDNN model were set as follows: two hidden layers, the number of neurons in each hidden layer, and the number of iterations after each dataset update were determined using the Boolean algorithm. The learning rate was 0.001, the physical loss weight was 0.1, and the ReLU function was used to activate the neurons. Furthermore, the Adam optimizer was employed to optimize the coefficients of the neural network, improving the model's convergence speed and performance.
[0166] Step 3: Evaluation and interpretation of the shield tail gap prediction model. The Online PDNN model of this invention demonstrates high performance in predicting shield tail gaps in multiple azimuths. As shown in Table 2, the R² of the online prediction of shield tail gaps in three azimuths... 2 The means were 0.93, 0.90, and 0.90, respectively. According to... Figure 7 As shown in the shield tail gap prediction deviation, the developed Online PDNN model performs better in predicting the shield tail gap directly above, with the prediction deviation controlled within 5mm in 94.44% of the test samples. Figure 8 As shown, after each dataset update, the VAF, a20_index, and R of the Online PDNN model... 2The values of y1 and y2 all remained at or above 90%, 1, and 0.9, respectively. Therefore, the Online PDNN model exhibits strong robustness. All calculation results demonstrate that the developed Online PDNN model provides reliable estimation performance for the measurement of the shield tail gap. Furthermore, to understand the effect of incorporating physical laws into the algorithm, the SHAP method was used to analyze the contributions of various features, especially the influence of the pitch angle on the model output y1. Figure 9 As shown in (a), among the 51 feature variables that contribute to the output y1, the pitch angle ranks second. According to Figure 9 (b) It can be seen that, while keeping other characteristics stable, the Shap value corresponding to the pitch angle increases with the increase of the pitch angle characteristic value. Therefore, the pitch angle has a positive impact on the prediction results of the shield tail clearance.
[0167] Table 2. Performance Evaluation of the OnlinePDNN Model
[0168]
[0169] Step 4: MOO of the shield tail gap. The shield tail gap measurement and control technology proposed in this invention based on Online PDNN and NSGA-II performs well in searching for the optimal solution of the shield tail gap MOO, and the optimal adjustment space of the adjustable parameter found is within 20% above and below the self value. Figure 10 and Figure 11 The table shows the optimized target values and Pareto fronts obtained under different constraints with adjustable parameter values fluctuating within 10%, 20%, 30%, 40%, and 50%. It can be seen that as the range of adjustable decision variables increases, the tail gap gradually decreases, and the Pareto solution set is more evenly distributed throughout the non-dominated solution space. As shown in Table 3, under the 50% constraint, the overall optimization percentage of the tail gap increased by 58.37% compared to the original data, and the optimized tail gaps did not exceed their warning value of 10 mm, indicating that the optimization scheme obtained based on Online PDNN and NSGA-II for MOO is reasonable. Considering the increased safety risks associated with numerous adjustments to TBM mechanical parameters during tunneling, the optimal variation space of the adjustable parameters used in the tail gap MOO is limited to within 20% of their values.
[0170] Table 3. Statistical analysis of data between original values and optimized values under different constraints
[0171]
[0172] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a deep learning-based high-precision measurement and control method for shield tail gap.
[0173] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of a deep learning-based high-precision measurement and control method for shield tail gap.
[0174] An application embodiment of the present invention provides an information data processing terminal, which is used to realize a high-precision measurement and control system for shield tail gap based on deep learning.
[0175] The proposed deep learning-based high-precision measurement and control method and system for shield tail gap is loaded onto the TBM control platform. By utilizing multi-source information obtained from TBM sensors, online measurement and real-time control of the shield tail gap can be achieved, ensuring the construction safety of tunnel excavation.
[0176] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.
[0177] (1) Figures 7-11 Tables 2 and 3 illustrate the application effects of the embodiments of the present invention in a tunnel construction project example. The results show that the embodiments of the present invention, when predicting the shield tail gap in different azimuths, have an R... 2 Scores can reach 0.90 and above. The prediction error for the shield tail gap directly above 94.44% of the test samples can be controlled within 5mm. Even with continuous dataset updates, the developed Online PDNN model maintains superior performance. Specifically, VAF, a20_index, and R... 2 The values of are all maintained at 90%, 1, and 0.9 or above, respectively. This indicates that the embodiments of the present invention can reliably estimate the shield tail clearance. Furthermore, under different constraints where the adjustable parameter values fluctuate within the range of 10%, 20%, 30%, 40%, and 50%, the MOO (Modular Oscillator) of the shield tail clearance yielded overall optimization degrees of 16.53%, 30.87%, 43.38%, 52.22%, and 58.37%, respectively. Moreover, the optimized shield tail clearances did not exceed the warning value of 10 mm. This demonstrates that the shield tail clearance optimization scheme obtained based on the embodiments of the present invention is reasonable and provides theoretical support for field personnel to control the shield tail clearance.
[0178] (2) Compared with existing SluM and ENZAN shield tail gap measurement systems, the embodiments of the present invention do not require the use of specialized equipment such as ultrasonic sensors and industrial cameras to collect data for shield tail gap measurement. This simplifies the shield tail gap measurement process and reduces the cost of shield tail gap measurement. Furthermore, compared with the published shield tail gap measurement accuracy records of SluM, the error range of the shield tail gap predicted by the embodiments of the present invention is no greater than that of the SluM system (millimeters of accuracy).
[0179] The deep learning-based high-precision measurement and control method for shield tail clearance proposed in this invention can be illustrated with two specific application examples to demonstrate the application of this technology in practical engineering:
[0180] Example 1: Construction of Urban Underground Tunnels
[0181] In urban underground tunnel construction projects, tunnel boring machines (TBMs) traverse areas with complex geology and dense underground pipelines. Precise control of the tail clearance is crucial in this process to prevent ground subsidence and protect underground infrastructure.
[0182] 1. Tail gap calculation: Based on the specific geological conditions of the project (such as soil type and water level) and design parameters (such as tunnel diameter and tunnel boring machine model), select a suitable minimum tail gap calculation model and calculate the initial tail gap δ0.
[0183] 2. PDNN Model Application: In the early stages of construction, preliminary construction data (such as pitch angle, tunnel curve radius changes, etc.) are collected, and a PDNN model is trained based on this data. As the project progresses, real-time data is continuously collected and the PDNN model is updated online to accurately estimate the shield tail clearance in real time.
[0184] 3. Performance Evaluation and Feature Analysis: Using actual construction data, the SHAP method is used to analyze the impact of each feature on the accuracy of shield tail gap estimation, thereby optimizing the input feature set and improving the accuracy and reliability of the model.
[0185] 4. Multi-objective optimization application: Based on real-time construction data, multi-objective optimization methods are applied to adjust construction parameters (such as tunnel boring machine advance speed and lining ring installation speed) to optimize construction efficiency and cost while ensuring construction safety.
[0186] Example 2: Mountain Tunnel Crossing
[0187] In mountain tunnel projects, the challenges faced during construction include complex geological structures and large elevation changes, which require more precise control of the shield tail gap to adapt to the terrain and reduce environmental impact.
[0188] 1. Precise calculation of shield tail gap: In view of the unique geological and topographical conditions in mountainous areas, a highly adaptable shield tail gap calculation model is selected, taking into account the influence of large elevation differences and curve radii on the shield tail gap, and calculating the initial and changing shield tail gap.
[0189] 2. Online PDNN model adjustment: During construction, data such as pitch angle changes and geological hardness are collected in real time, and the PDNN model is continuously adjusted to adapt to the ever-changing construction environment and achieve precise control of the shield tail gap.
[0190] 3. Performance Evaluation and Adjustment: Regularly evaluate the performance of the PDNN model. By analyzing the model's dependence on various features, adjust the model structure or input features to ensure the accuracy and practicality of the model's predictions.
[0191] 4. Practical application of multi-objective optimization: Combining the special needs of mountain tunnels, multi-objective optimization technology is used to balance construction speed, cost and environmental impact, so as to achieve the dual goals of economic benefits and ecological protection.
[0192] These two examples demonstrate how to apply deep learning-based high-precision measurement and control methods for shield tail gap under different construction environments and engineering requirements, and achieve efficient and safe tunnel construction through continuous data collection, model training and optimization.
[0193] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.
[0194] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A high-precision measurement and control method for shield tail clearance based on deep learning, characterized in that, include: S1, Calculation of tail gap: Based on the theory of minimum tail gap calculation, assuming that the lining ring is assembled in an ideal state, there are six minimum tail gap calculation models. S2. First, the derived equation for calculating the shield tail gap at any position is embedded into the DNN framework as a physical constraint to obtain the PDNN model. Then, the PDNN is trained in an online learning manner so that it can adaptively update the model hyperparameters according to the input observation stream, thereby achieving accurate estimation of the shield tail gap. S3. Five statistical indicators are selected to evaluate the performance of the established Online PDNN model; the Shapley Additive exPlanations method is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features. S4. First, use non-dominated sorting to search for the Pareto solution set of the MOO problem; then, calculate the crowding distance of each solution to maintain the diversity and uniform distribution of the solution set; finally, use the crowding comparison operator to compare two solutions, where the solution with the lower rank and the solution with the larger crowding distance are selected; repeat the above steps until the maximum number of iterations, and finally generate the Pareto front. S1 specifically includes: based on the minimum shield tail gap calculation theory, assuming that the lining ring assembly is in an ideal state, there are six minimum shield tail gap calculation models, and their mathematical expressions are shown in formula (1). (1) In the formula, δ0 is the initial shield tail clearance; δ1 is the change in shield tail clearance caused by the change in pitch angle α. δ2 is the change in the shield tail gap caused by the tunnel boring machine excavating along the route of the deflection angle η and the radius of the tunnel curve R. Its mathematical expressions are shown in formulas (2) and (3), respectively. (2) (3) In the formula, l is the length of the shield tail covering segment; α is the pitch angle; R is the radius of the tunnel curve; and D is the outer diameter of the lining ring. The three concentric circles with center O represent a lining ring with radius D / 2 and a lining ring with radius (D / 2+δ). min ) and (D / 2+δ max Two auxiliary circles; the ellipse centered at O' is the 1-1 section of the shield shell; F1 and F2 are the two foci of the ellipse; the major and minor axes of the ellipse are 2R / cosα and 2R, respectively; based on the definition of an ellipse and the law of cosines, the tail clearance (δ) for any position is derived. p The physical equation for the calculation is formula (4); (4) In the formula, β is the horizontal angle, representing the position where the shield tail gap needs to be calculated; and Calculate using formulas (5) and (6) respectively; (5) (6) In the formula, and These are the distances from the center O of the lining ring to the foci F1 and F2 of the ellipse, respectively.
2. The high-precision measurement and control method for shield tail gap based on deep learning as described in claim 1, characterized in that, The construction of the PDNN model in S2 specifically includes: The derived equation for calculating the shield tail gap at any position is integrated as a physical law into the loss function of the DNN model to constrain the model's training process. This helps to compensate for the shortcomings of purely data-driven DNN models, which rely excessively on data and lack reasonable explanations of the internal mechanisms of the physical system. According to the PDNN framework used for shield tail gap measurement, the loss function of the PDNN model coupled with the physical mechanism consists of the observation loss MSE and the physical loss. It consists of two parts, and its mathematical expression is formula (7): (7) In the formula, and They represent MSE and The weights; , and These represent the observed, theoretical, and predicted values of the shield tail gap, respectively.
3. The high-precision measurement and control method for shield tail gap based on deep learning as described in claim 1, characterized in that, The online update mechanism of the PDNN model in S2 specifically includes: To enhance the practicality and generalization of the PDNN model, an online learning approach is adopted to train the model. Specifically, the data used to develop the Online PDNN model is gradually updated as the excavation operation progresses, and the model's hyperparameters are dynamically adjusted as the dataset is updated. In the online prediction concept, it is assumed that the initial dataset contains N0 samples. Then, N1 samples are selected from the N0 samples as training data to train the model, and the remaining N2 samples are used as test data to verify the model's performance. The BO algorithm is used to automatically search for the optimal combination of hyperparameters for the Online PDNN model.
4. The high-precision measurement and control method for shield tail gap based on deep learning as described in claim 1, characterized in that, The evaluation of the Online PDNN model in S3 specifically includes: To evaluate the performance of the established Online PDNN model, five commonly used statistical metrics were selected, including the coefficient of determination (C). 2 a20_index, Variance Explained Variable Variable (VAF), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) are expressed mathematically in formulas (8)-(12); typically, a20_index, Variance Explained Variable Variable (VAF), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) are expressed mathematically in formulas (8)-(12); 2 The closer the values of VAF and a20_index are to 1, and the closer the values of MAE and RMSE are to 0, the higher the prediction accuracy of the model. (8) (9) (10) (11) (12) In the formula, 𝑛 is the total number of samples; and These are the predicted value and the observed value, respectively. It is the average of all samples; 20 It is the number of samples with an error not exceeding ±20% of the observed value.
5. The high-precision measurement and control method for shield tail gap based on deep learning as described in claim 1, characterized in that, The interpretability of the Online PDNN model in S3 specifically includes: To understand the decision-making strategy of the developed Online PDNN model, the Shapley Additive Explanations method is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features. The mathematical expression of the Shapley value is shown in formula (13): (13) In the formula, Z is the set of all input features with dimension M; 𝑆 is a subset of the input features with dimension M. ; 𝑥 (k) = 𝐸[k(k)|k 𝑆 ] is the sample mean calculated based on the training subset 𝑆; SHAP analysis, based on the additive eigenvalue attribution method, can simplify complex models into linear functions: (14) In the formula, The average of the estimated values; Describe whether feature i has been observed. and These indicate whether the feature is absent or present, respectively.
6. The high-precision measurement and control method for shield tail gap based on deep learning as described in claim 1, characterized in that, S4 specifically includes: In the minimum MOO problem, four components need to be considered, including the k-dimensional objective vector. The mathematical representations of m inequality constraints, n equality constraints, and p-dimensional decision variables x are shown in formulas (15)-(17): (15) (16) (17) In the formula, f i (x) is the i-th objective function; Ω is the decision space with p independent variables; The main steps for optimizing the shield tail gap based on Online PDNN and NSGA-II are as follows: First, a Pareto solution set for the MOO problem is searched using non-dominated sorting; then, the crowding distance of each solution is calculated to maintain the diversity and uniform distribution of the solution set; finally, the crowding comparison operator is used to compare two solutions, where the solution with the lower rank and the solution with the larger crowding distance are selected; the above steps are repeated until the maximum number of iterations is reached, and finally, the Pareto front is generated; in order to determine the optimal solution from the Pareto front, the distance between the Pareto optimal solution vector and the ideal solution vector is calculated using formula (18), and the solution with the smallest distance is selected as the optimal solution: (18) In the formula, It is the value of the i-th objective function in the j-th Pareto solution; i is the ideal value of the i-th objective function; k is the number of all objective functions.
7. A deep learning-based high-precision measurement and control system for shield tail clearance, implementing the deep learning-based high-precision measurement and control method for shield tail clearance as described in any one of claims 1 to 6, characterized in that, include: The shield tail gap calculation module is used to calculate the minimum shield tail gap based on the minimum shield tail gap calculation theory, assuming that the lining ring is assembled in an ideal state, and there are six minimum shield tail gap calculation models. The PDNN model acquisition module is used to embed the derived equations for calculating the shield tail gap at any position as physical constraints into the DNN framework to obtain the PDNN model. The PDNN model training module is used to train the PDNN in an online learning manner, enabling it to adaptively update the model hyperparameters based on the input observation stream, thereby achieving accurate estimation of the shield tail gap. The PDNN model evaluation module is used to select five statistical indicators to evaluate model performance. The PDNN model interpretability analysis module is used to analyze the marginal contribution of each input feature to the model and determine the importance ranking of the input features using the Shapley Additive exPlanations method. The shield tail gap MOO execution module is used to execute the shield tail gap MOO.
8. A computer device comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the deep learning-based high-precision measurement and control method for shield tail gap as described in any one of claims 1 to 6.
9. An information data processing terminal, used to implement the deep learning-based high-precision measurement and control system for shield tail gap as described in claim 7.
Citation Information
Patent Citations
VR-based shield segment virtual assembling method and system, and data processing terminal
CN113345104A
Calculation method for shield tail gap distribution characteristics of shield tunneling machine
CN115614041A
