Intelligent posture deviation correction control method and system for continuous tunneling shield tunneling machine

Abstract

The invention belongs to the field of shield tunneling machines, and discloses an intelligent posture deviation correction control method and system for a continuous tunneling shield tunneling machine, and the method comprises the steps: carrying out a pushing and splicing synchronous construction experiment to collect data; integrating the BiLSTM and the KAN, and predicting the posture of the continuous tunneling shield tunneling machine; in combination with an MOGWO method, an optimal proportional-integral-differential (PID) parameter is determined, and attitude deviation is minimized. The invention provides a shield attitude intelligent control method integrating deep learning, physical modeling and an intelligent optimization algorithm. The key innovation point is that the KAN model and the BiLSTM model are combined to accurately predict the attitude of the continuous tunneling shield machine in various assembling hydraulic oil cylinder missing scenes. And then combining the improved BiLSTM model with an MOGWO algorithm to optimize the response performance of the propulsion hydraulic system, and finally improving the efficiency and quality of tunnel construction. Verification research is carried out by utilizing actual measurement data of a continuous tunneling shield tunneling machine test bed, and the effectiveness of the method is verified.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the technical field of shield machines, and in particular relates to an intelligent deviation correction control method and system for a continuous tunneling shield machine. Background Art

[0002] In recent decades, shield machines (TBMs) have become the mainstream equipment for long-distance, deep tunnel construction due to their safety, efficiency, and minimal environmental impact. To shorten the traditional "excavation-stop-assembly" cycle, the industry has developed a simultaneous push-and-assemble technology, enabling cutterhead excavation and segment installation to proceed simultaneously. This technology is said to improve construction efficiency by 30%–50%.

[0003] Currently, there are three main types of push-and-assemble synchronization technologies: the lattice tunneling method relies on multiple sets of telescopic cylinders to construct a sliding front shield, which can accurately control the cutterhead posture, but has extremely high requirements for installation space and control accuracy; the F-Navi method uses a ball seat-cylinder swing mechanism to adjust the posture and synchronize operations. The manufacturing cost remains high due to the complexity of the ball seat sealing and manual control; the dual hydraulic cylinder method uses a sliding middle shield of the push cylinder and the assembly cylinder to complete the synchronous assembly by using its reaction force. It has a simple design and high reliability, and is most widely used in China and Japan.

[0004] However, the dual-hydraulic-cylinder method of synchronous pushing and splicing can weaken the shield's attitude stability and increase the risk of tunnel misalignment. Accidents at home and abroad have demonstrated that attitude loss can lead to significant economic losses and project delays. Existing research on attitude control primarily focuses on traditional shield machines, lacking systematic analysis and optimization of shield controllability under synchronous pushing and splicing conditions, particularly when assembly cylinders are missing.

[0005] With the development of machine learning and reinforcement learning, data-driven models have been used to predict and control shield machine attitude. However, existing methods still have shortcomings in time series feature extraction, reward function design, and adaptability to uncertain environments. Furthermore, they have not studied the cylinder-missing condition during synchronous pushing and assembly. Therefore, it is urgent to develop an intelligent attitude control method that can robustly operate under conditions of partially missing assembly cylinders and highly uncertain operating conditions, thereby reducing the risk of shield machine snaking and ensuring safety during long-distance tunnel construction. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention provides an intelligent deviation correction control method for the posture of a continuous tunneling shield machine.

[0007] The present invention is achieved by providing an intelligent deviation correction control method for a continuous tunneling shield machine, characterized in that the intelligent deviation correction control method for a continuous tunneling shield machine specifically comprises:

[0008] S1: Conduct push-and-splice synchronous construction experiments to collect data;

[0009] S2: Integrate BiLSTM and KAN networks to predict the posture of continuous tunneling shield machines;

[0010] S3: Combined with the MOGWO method, the optimal proportional-integral-derivative (PID) parameters are determined to minimize the attitude deviation.

[0011] Furthermore, the specific experimental steps of S1 are as follows:

[0012] (1) Initialize all experimental parameters, system variables and the operating status of each component of the continuous tunneling shield machine;

[0013] (2) Operator adjusts parameters;

[0014] (3) When the space for installing the first lining ring is prepared, the push-and-spin synchronous construction mode is activated;

[0015] (4) According to the segment installation sequence, retract the assembly hydraulic cylinder group corresponding to the segment to be installed, and then use the segment installation machine to transport the segment to the designated location, where the installation is completed. Repeat this process until the entire lining ring is installed;

[0016] (5) Repeat steps (1)-(4) until the end of the experiment.

[0017] Furthermore, in S2, the input layer of the BiLSTM receives raw data for feature extraction and processing, including geological parameters, the pressure difference of the propulsion hydraulic cylinder group, and the pressure of the assembly hydraulic cylinder group; then, the data passes through the BiLSTM hidden layer to fully capture and learn the feature information in the data; thereafter, the output from the BiLSTM hidden layer is transmitted to the input layer of the KAN; subsequently, the complex relationship between the variables is further refined and optimized in the hidden layer of the KAN to obtain intermediate variables; finally, the output layer of the KAN generates a prediction result, that is, the shield posture represented by the stroke difference of the propulsion hydraulic cylinder group.

[0018] Furthermore, in S3, a multi-objective optimization framework for intelligent deviation correction of the continuous tunneling shield machine's posture was developed to determine the optimal values of the PID controller parameters. Specifically, this framework combines an improved BiLSTM model, the MOGWO algorithm, and a physical simulation model built in the MATLAB Simulink environment. After using this integrated framework to generate the Pareto front, the TOPSIS method was applied to select the optimal solution, thereby achieving intelligent deviation correction of the continuous tunneling shield machine's posture.

[0019] The present invention also provides a continuous tunneling shield machine posture intelligent deviation correction control system, the system comprising:

[0020] The experimental data collection subsystem is used to perform intelligent deviation correction experiments on the shield machine in the push-and-assemble synchronous construction mode, collecting time series data including the pressure difference of the propulsion hydraulic cylinder group, the pressure of the assembly hydraulic cylinder group, and geological parameters;

[0021] Bidirectional long short-term memory network module (BiLSTM), used to extract time series features from raw data and construct intermediate state feature vector sequences;

[0022] Kolmogorov–Arnold network module (KAN), which receives BiLSTM output, models the coupling relationship between nonlinear features, and outputs the shield posture prediction (expressed as the stroke difference of the propulsion hydraulic cylinder group);

[0023] The multi-objective optimization parameter tuning module performs global search and performance convergence on proportional-integral-derivative (PID) controller parameters based on the multi-objective Grey Wolf Optimization Algorithm (MOGWO) and physical simulation models.

[0024] The decision integration module selects the optimal control strategy on the Pareto front through the TOPSIS method for attitude closed-loop control.

[0025] Furthermore, the experimental acquisition subsystem supports a dynamic sampling mode based on the lining ring installation cycle, including:

[0026] Initialize the excavation status and various parameter variables;

[0027] Activate the push-and-splice synchronous construction mode and record the feedback parameters of each hydraulic system during the operation;

[0028] The system clock is used to drive the sampling mechanism to form a time-aligned data stream for subsequent model training and verification.

[0029] Furthermore, the input layer of the BiLSTM module receives the feature sequence after preprocessing (outlier detection) and encodes the forward and reverse dependencies through a bidirectional gating mechanism, thereby improving the posture prediction model's ability to perceive historical and future trends in the advancement process. The KAN module uses a learnable activation function mechanism to approximate the objective function, thereby enhancing the interactive expression ability between features.

[0030] Furthermore, the MOGWO algorithm integrated in the optimization and parameter adjustment module is built on the basis of the digital twin model in the MATLAB Simulink environment. The three performance indicators of overshoot (O), adjustment time (T), and time integral of absolute error product (ITAE) are simultaneously optimized to form a Pareto optimal solution set. Then, the optimal PID controller parameter combination is selected through the TOPSIS algorithm to realize intelligent deviation correction control of the shield machine posture.

[0031] Furthermore, the posture prediction result output by the KAN module is represented by the stroke difference of the propulsion hydraulic cylinder group, which is used to accurately reflect the posture change trend of the shield machine, such as pitch.

[0032] Furthermore, the multiple objectives in the optimization and parameter adjustment module consider an adjustable weight mechanism, and users can assign different control objective weights according to tunnel working conditions to adapt to the attitude control requirements under different strata and lining ring installation methods.

[0033] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0034] This paper proposes a shield machine posture control method that integrates deep learning, physical modeling, and intelligent optimization. The key innovation lies in combining the KAN model with the BiLSTM model to accurately predict the shield machine's posture under different scenarios where the hydraulic cylinder assembly is missing. The improved BiLSTM model is then combined with the MOGWO algorithm to optimize the response performance of the propulsion hydraulic system, ultimately improving the efficiency and quality of tunnel construction. A validation study using measured data from a continuous tunneling shield machine test bench confirmed the effectiveness of this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a framework of a method for intelligent control of the attitude of a continuous tunneling shield machine provided by an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of a continuous tunneling shield machine provided by an embodiment of the present invention;

[0037] Figure 3 This is the installation process of the lining ring provided by the embodiment of the present invention: (a)-(c) standard blocks; (d)-(e) adjacent blocks; (f) capping block;

[0038] Figure 4 This is the structure of the improved BiLSTM model provided by an embodiment of the present invention;

[0039] Figure 5 This is a continuous tunneling shield machine attitude control simulation model provided by an embodiment of the present invention;

[0040] Figure 6 This is the MOGWO algorithm process provided by an embodiment of the present invention;

[0041] Figure 7 This is a continuous tunneling shield machine test bench provided by an embodiment of the present invention: (a) overall principle diagram; (b) assembly hydraulic cylinder group entity; (c) reaction and propulsion hydraulic cylinder group entity;

[0042] Figure 8The layout of the propulsion hydraulic cylinder group and the assembly hydraulic cylinder group provided in the embodiment of the present invention;

[0043] Figure 9 It is the correlation analysis between different variables provided by the embodiment of the present invention;

[0044] Figure 10 The hidden layer structure optimization results provided by the embodiments of the present invention are: the MSE of models with different BiLSTM hidden layer structures on (a) the training set and (b) the test set, and the MSE of models with different KAN hidden layer structures on (c) the training set and (d) the test set;

[0045] Figure 11 is the training loss of the improved BiLSTM model provided in the embodiment of the present invention;

[0046] Figure 12 This is a comparison of the construction efficiency of a conventional shield machine and a continuous tunneling shield machine provided by an embodiment of the present invention;

[0047] Figure 13 is the prediction deviation of the improved BiLSTM model provided in the embodiment of the present invention;

[0048] Figure 14 is a relationship diagram between the predicted posture and the measured posture provided by an embodiment of the present invention;

[0049] Figure 15 It is the feature importance ranking provided by the embodiment of the present invention;

[0050] Figure 16 PDP and ICE curves of (a) x5, (b) x21, and (c) x4 provided in embodiments of the present invention;

[0051] Figure 17 2DPDP for (a) x5 and x21, (b) x5 and x4, and (c) x4 and x21, provided in an embodiment of the present invention;

[0052] Figure 18 Comparison results of (a) T, (b) O, and (c) ITAE distribution curves obtained by the method provided by the embodiment of the present invention and the manual control method under comprehensive consideration of all scenarios;

[0053] Figure 19 It is the shield attitude curve under different PID parameter tuning methods provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] like Figure 1 As shown, an embodiment of the present invention provides a method for intelligent deviation correction control of a continuous tunneling shield machine, which specifically includes:

[0056] S1: Conduct push-and-splice synchronous construction experiments to collect data;

[0057] S2: Integrate BiLSTM and KAN networks to predict the posture of continuous tunneling shield machines;

[0058] S3: Combined with the MOGWO method, the optimal proportional-integral-derivative (PID) parameters are determined to minimize the attitude deviation.

[0059] Continuous tunneling shield machine equipped with dual hydraulic system Figure 2 As shown in the figure, its main components include the cutterhead, front shield, propulsion hydraulic system, middle shield, assembly hydraulic system, and segment installation machine. Unlike the "excavation-stop-installation" construction process used by traditional shield machines, the continuous tunneling shield machine can perform soil excavation and segment installation operations simultaneously. Although this uninterrupted excavation method significantly improves tunnel construction efficiency, it also increases the continuous tunneling shield machine's sensitivity to the front soil pressure and the reaction force of the rear assembly hydraulic cylinder group. This high sensitivity poses a huge challenge to accurately controlling the continuous tunneling shield machine's posture in complex dynamic environments. This is mainly because, in actual construction, the control of the shield machine's posture relies heavily on the PID controller to adjust the pressure difference between the propulsion hydraulic cylinder groups. Therefore, exploring a continuous tunneling shield machine posture control method that comprehensively considers soil characteristics and the lack of assembly hydraulic cylinder groups has important academic value and engineering significance.

[0060] As is well known, the working principle of a PID controller for posture control is to minimize the deviation between the current posture and the target posture by setting appropriate control parameter values. These control parameter values include the proportional coefficient, the integral coefficient, and the differential coefficient. The mathematical description of this process is shown in formula (1). In the present invention, the differential coefficient is set to zero.

[0061]

[0062] Where t represents time; P(t) is the regulated pressure difference; E(t) represents the deviation between the current posture and the target posture at time t, and its value changes with time.

[0063] The specific experimental steps of S1 are as follows:

[0064] (1) Initialize all experimental parameters, system variables, and the operating status of each component of the continuous tunneling shield machine. For example, before the experiment begins, the propulsion hydraulic cylinder group and the assembly hydraulic cylinder group are both in the locked state, and the experimental platform is in the reset state.

[0065] (2) The operator adjusts parameters, such as the oil pressure opening of each set of thrust hydraulic cylinders, to switch the continuous tunneling shield machine to tunneling mode, causing the cutterhead to move forward and cut the soil.

[0066] (3) When the space for installing the first lining ring is prepared, the push-and-spin synchronous construction mode is activated. Specifically, the propulsion hydraulic system in the front shield continues to drive the cutterhead to cut the soil, while the assembly hydraulic system in the middle shield starts to coordinate with the segment installation machine to install the lining segments.

[0067] (4) According to the segment installation sequence, retract the assembly hydraulic cylinder group corresponding to the segment to be installed. Then, use the segment installation machine to transport the segment to the designated location and complete the installation there. Repeat this process until the entire liner ring is installed.

[0068] (5) Repeat step 14 until the experiment is completed.

[0069] Numerous sensors installed on the continuous tunneling shield machine test bench automatically monitor and record operational data during the experiment. This collected data is then transmitted to the control terminal and stored in a CSV file format using information conversion technology. Furthermore, the shield machine's control terminal is capable of communicating and transmitting data with an external computing device (a Dell G15 laptop) via a Modbus module, ensuring efficient management and utilization of experimental data.

[0070] According to step 4, when installing segments at different locations, the assembly hydraulic cylinder groups at these locations need to be retracted. Therefore, when the lining ring is divided into different blocks, the retraction scheme of the assembly hydraulic cylinder group is also different. For example, Figure 3 The figure shows the installation process of a lining ring using a "1+5" segmentation pattern. It can be observed that the number and location of the missing assembly hydraulic cylinder groups are significantly affected by the position of the segments to be assembled. In other words, the pressure distribution of the assembly hydraulic cylinder groups varies significantly when installing segments in different positions. This indicates that the propulsion hydraulic system, influenced by the assembly hydraulic system, will also exhibit different thrust distributions, thus affecting the tunneling attitude of the shield machine.

[0071] Therefore, to ensure the shield machine can tunnel smoothly according to the predetermined posture, it is necessary to explore corresponding shield posture adjustment strategies for various possible scenarios where the assembly hydraulic cylinder is missing. However, the lack of a clear mathematical relationship between pressure difference and shield posture, coupled with the potential influence of other factors such as soil characteristics, makes it significantly challenging to directly construct an analytical model.

[0072] Deep learning, a powerful data-driven approach, can automatically extract complex nonlinear features, effectively process high-dimensional data, and model implicit relationships between inputs and outputs without requiring prior knowledge. With this in mind, this paper utilizes deep learning to reveal the underlying nonlinear relationship between pressure differential and shield attitude. Furthermore, the paper considers the effects of soil properties and the absence of assembly cylinders on shield attitude adjustment. This approach is expected to improve the accuracy of continuous tunneling shield machine attitude prediction.

[0073] The S2, improved BiLSTM neural network architecture is as follows Figure 4 As shown in Figure 2, the BiLSTM input layer receives raw data for feature extraction and processing, including geological parameters, pressure differences between propulsion hydraulic cylinder groups, and pressures within each assembly hydraulic cylinder group. The data then passes through the BiLSTM hidden layer, fully capturing and learning the data's feature information. The output from the BiLSTM hidden layer is then transmitted to the KAN input layer. The complex relationships between the variables are then further refined and optimized in the KAN hidden layer to obtain intermediate variables. Finally, the KAN output layer generates a predicted result, namely the shield posture represented by the stroke difference between the propulsion hydraulic cylinder groups.

[0074] In order to comprehensively evaluate the performance of the improved BiLSTM model, the present invention uses three evaluation indicators, including root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 These indicators quantify the deviation between the predicted results and the measured data from different perspectives. The specific calculation method is shown in the following formula.

[0075]

[0076] where N pre represents the number of samples; represents the mean of all observations; is the predicted value of the i-th sample; and y (i) is the observed value of the i-th sample.

[0077] The adjustment of the S3 shield machine's posture depends on the precise control of the PID controller on the pressure of each group of propulsion hydraulic cylinders. With this goal in mind, the present invention develops a multi-objective optimization framework for intelligent deviation correction of the posture of a continuous tunneling shield machine to determine the optimal values of the PID controller parameters. In particular, the framework combines an improved BiLSTM model, a MOGWO algorithm, and a physical simulation model built in the MATLAB Simulink environment. After using the integrated framework to generate the Pareto frontier, the TOPSIS method is applied to select the most ideal solution to adjust the posture of the continuous tunneling shield machine. The structure of the integrated framework and the detailed description of the TOPSIS method are as follows.

[0078] The virtual model used to control the vertical posture of the continuous tunneling shield machine is as follows: Figure 5 As shown. The main components include input signals, PID controllers, delay modules, actuators and output signals. Specifically, the input signals correspond to target working postures, including: horizontal straight line excavation, upward posture excavation and downward posture excavation. The signal values of these three postures are set to 0mm, 15mm and -15mm, respectively. The total response time of each posture is 30 seconds, including adjustment and steady-state stages. Due to the different scenarios of missing assembly hydraulic cylinder groups in the push-and-splice synchronous construction process. Therefore, for each scenario, the three shield postures must be adjusted independently. In addition, since the differential module of the PID controller will amplify the noise in the measurement data, the present invention only uses proportional and integral modules. In this case, the decision variable of MOO is the proportional coefficient K p and the integral coefficient K i . The delay module is designed to reflect the time delay effect in the response of the propulsion hydraulic system. The actuator is an improved BiLSTM model developed, which establishes a mapping relationship between the shield posture and geological parameters, the pressure of the assembly hydraulic cylinder group, and the pressure difference between the propulsion hydraulic cylinder groups after PID adjustment. The output signal represents the adjusted shield posture, and its value can be observed on the oscilloscope. In short, during the posture adjustment using PID, the virtual model can simultaneously consider the geological parameters and the working status of the assembly hydraulic cylinder group.

[0079] In many control system tuning tasks, operators are typically concerned with three performance metrics: settling time, overshoot, and the integral of the time multiplied by the absolute error. Given this, the present invention uses these metrics as objective functions and solves them using the MOGWO algorithm. Compared to GWO, MOGWO significantly improves upon GWO by adding an archiving strategy and an updated leader selection mechanism. The archiving strategy enables MOGWO to preserve nondominated solutions by creating an external archive. Figure 6 The algorithm flow of MOGWO is shown. The process starts with population initialization, where the initial position of the gray wolf population is set by generating a set of random solutions. The performance of each solution is then evaluated based on a predefined objective function. Subsequently, the individuals in the group are sorted according to their individual objective function values, and the positions of the gray wolf leaders (α, β, and δ) are updated. Next, the hunting behavior of the gray wolves is used to update the position of each wolf in the population to generate new candidate solutions. Through continuous iteration, the quality of the solution is gradually optimized. Finally, the new solution is evaluated and compared with the current optimal solution to construct the final Pareto frontier.

[0080] The present invention uses the TOPSIS method to determine the most ideal solution from the Pareto front. The core principle is to determine the score of each candidate solution by calculating the Euclidean distance between each candidate solution and the optimal solution and the worst solution, and then sort these solutions accordingly.

[0081] The effectiveness of the proposed shield attitude control method is verified by simulating field experimental data during shield construction.

[0082] (1) Data resources

[0083] As we all know, deep learning relies on the quantity and quality of existing data. Therefore, the present invention conducted a field experiment of the synchronous construction of a continuous tunneling shield machine, aiming to collect a large amount of reliable observation data for the development of a shield attitude prediction model. Specifically, the experiment was carried out on a 1:1 restored continuous tunneling shield machine test bench with a diameter of 6m and a length of 10m. Figure 7 Unlike actual continuous tunneling shield excavation, this test rig uses a reaction hydraulic system to simulate the various complex geological conditions the shield machine may traverse. However, since varying soil properties are simulated by adjusting the pressure of the reaction hydraulic cylinders, this precludes the installation of a cutterhead at the front of the continuous tunneling shield. Furthermore, since the shield shell is not surrounded by soil, simulation of shield tail grouting is also limited.

[0084] The reaction hydraulic system is divided into four reaction hydraulic cylinder groups: upper, lower, left and right. The upper and lower reaction hydraulic cylinder groups are composed of three reaction hydraulic cylinders with a diameter of 0.22 meters and a stroke of 1.65 meters, and the left and right reaction hydraulic cylinder groups are composed of four identical reaction hydraulic cylinders. In general, these 14 reaction hydraulic cylinders are evenly distributed along the circumference. In view of the fact that soft homogeneous strata and upper hard and lower soft composite strata are more detrimental to the safety of tunnel construction, the present invention focuses on these two types of strata. Under the working conditions of soft homogeneous strata, the pressure distribution of the four groups of reaction hydraulic cylinders is relatively balanced, with an average pressure of 100 bar; while in the upper hard and lower soft composite strata, the pressure of the cylinder groups shows significant differences: the average pressure of the left, right and lower groups of cylinders is 136.56 bar, and the average pressure of the upper group of cylinders is as high as 219.92 bar.

[0085] Figure 8(a) describes the layout of the hydraulic cylinders in the propulsion hydraulic system. It can be seen that the propulsion hydraulic system is also composed of 4 groups (a total of 14) of cylinders with a diameter of 0.22m and a stroke of 1.65m. The upper and lower propulsion hydraulic cylinder groups control the vertical posture, and the left and right propulsion hydraulic cylinder groups control the horizontal posture. In essence, the shield posture can be represented by the stroke difference between the propulsion hydraulic cylinder groups in relative positions. In contrast, the assembly hydraulic system of the continuous tunneling shield machine includes 16 groups (a total of 32) of hydraulic cylinders with a diameter of 0.18m and a stroke of 2.25m. These hydraulic cylinders are evenly distributed along the circumference, as shown in Figure 2. Figure 8 (b) shown.

[0086] With the exception of the hydraulic cylinder dimensions, the hydraulic system circuit and maximum pressure and flow rate conform to the specifications commonly used in actual shield machines. The front and middle shields are connected via an articulated joint, facilitating the transition between locking and reset operations. Other components of the continuous tunneling shield and their testing procedures are consistent with actual construction.

[0087] also, Figure 8 (b) shows the block pattern of the lining ring adopted in this study, which follows the configuration of "3 standard blocks (B1, B2, B3) + 2 adjacent blocks (L1, L2) + 1 capping block (F)". Except for the capping block which has a central angle of 22.5°, the other five segments each correspond to a central angle of 67.5°. Figure 3 The assembly process of the lining ring in the "1+5" segmentation model shown in Figure 2 illustrates six scenarios involving missing assembly cylinders, as shown in Table 2. Specifically, when assembling segment B1, assembly cylinders numbered 9, 10, 11, and 12 must be retracted, while the remaining assembly cylinders remain extended. In scenario 2, assembly of segment B2 is completed by retracting assembly cylinders numbered 6, 7, 8, and 9. Compared to B1 and B2, assembly of segment B3 requires the retraction of five groups of assembly cylinders: 6, 13, 14, 15, and 16. Similarly, the installation of adjacent segment L1 also involves the retraction of five groups of assembly cylinders: 3, 4, 5, 6, and 16. When installing another adjacent segment, L2, scenario 5 indicates that assembly cylinders 1, 2, 3, and 16 must be retracted. Finally, when installing capping segment K, only the third group of assembly cylinders needs to be retracted.

[0088] Table 2 Working status of each assembly hydraulic cylinder when installing different segments

[0089]

[0090] The target posture is mainly achieved by adjusting the pressure difference between the propulsion hydraulic cylinder groups by the PID controller. However, these pressures are affected by the combined influence of geological conditions and the pressure distribution of the assembly hydraulic cylinder groups. Therefore, the pressures of 4 groups of reaction hydraulic cylinders (x1~x4), the pressures of 16 groups of assembly hydraulic cylinders (x5~x20), and the pressure difference between the upper and lower propulsion hydraulic cylinder groups (x21) are selected as input variables. In particular, the pressures of the selected 16 groups of assembly hydraulic cylinders should cover the six missing scenarios described in Table 1. Considering that the vertical shield posture can be described by the stroke difference between the upper and lower propulsion hydraulic cylinder groups, it is used as the output target. The statistical characteristics of the selected input variables and output targets are shown in Table 3. In order to prove that there is no strong correlation between all input variables and output targets, Figure 9 A correlation matrix based on Pearson's correlation coefficient is given.

[0091] Table 3 Numerical distribution of selected variables

[0092]

[0093] Note: SD stands for standard deviation.

[0094] (2) Model training and implementation

[0095] Model training for this study was performed on a computer equipped with an Intel i7-12700F CPU, an NVIDIA 3060 GPU, and 32GB of RAM. To ensure the performance of the improved BiLSTM model, a five-fold cross-validation method was used to evaluate the model. The main steps of this method are to randomly divide the data into five subsets, select one subset as the test set, and the other four subsets as the training set. The model is trained and performance is evaluated on the test set. This process is repeated five times, ensuring that each subset is used as the test set once. The final performance evaluation of the model is finally obtained by averaging the performance metrics of the five tests. Since there is no dimensionality difference between the decision variables, no normalization technique is applied.

[0096] To enhance the nonlinear fitting ability of the model, sigmoid and tanh functions are used to activate BiLSTM neurons, while KAN utilizes SiLU activation function. In addition, mean squared error (MSE) is used as the loss function during training, and the model is compiled using Adam optimizer.

[0097] Detailed hyperparameter adjustment is crucial to improving the performance of the improved BiLSTM model. Therefore, a grid search method was used to optimize the model structure and determine the optimal number of hidden layers and neurons in each hidden layer. Specifically, the number of hidden layers and the number of neurons in each hidden layer in the BiLSTM model were set to [1, 2, 3, 4] and [32, 64, 128, 256], respectively. For the KAN network, the number of hidden layers and the number of neurons in each hidden layer were set to [1, 2, 3, 4] and [3, 5, 10, 15], respectively. To improve the computational efficiency of the grid search, the present invention first fixed the hyperparameters of the KAN and then searched for the optimal hidden layer and number of neurons in each hidden layer of the BiLSTM. Due to the low computational efficiency of the KAN, it was initially set to one hidden layer containing five hidden neurons. After determining the optimal hidden layer and number of neurons in each hidden layer for the BiLSTM, a second grid search was performed to find the optimal hyperparameter combination for the KAN. Figure 10 The figure shows the optimization results for the four hyperparameters mentioned above, with a learning rate of 0.001, 1600 iterations, and a grid size of 5. The circles in the figure represent the MSE values corresponding to different hyperparameter combinations, while the red star marks the location with the minimum MSE. It can be observed that the optimal hyperparameter combination consists of two BiLSTM hidden layers and two KAN hidden layers, with the corresponding numbers of hidden neurons being 128, 128, 10, and 10, respectively.

[0098] Figure 11 The figure shows the training loss of the model based on the optimal hyperparameter combination. It can be seen that the loss shows a similar downward trend during each training run, with a rapid decrease during the first 150 iterations and then gradually stabilizing. Taking the fifth training run as an example, the MSE of the training set converged to around 0.68 after 800 iterations, indicating that the selected number of iterations was appropriate. In summary, the model with optimized hyperparameters can effectively capture the nonlinear relationship between feature variables and shield posture, thereby achieving accurate prediction of shield posture.

[0099] After the shield posture prediction model is trained, it is integrated into Figure 5 The execution module is shown in Figure 4. At the same time, the input variables related to the geological conditions and the assembly hydraulic cylinder are loaded into the fixed characteristic parameter module. In order to perform the shield posture optimization task, the key parameters of the MOGWO algorithm are listed in detail in Table 4. Considering that excessive adjustment of the operating parameters of the continuous tunneling shield machine may lead to safety hazards, the adjustable decision variable K is set to p and K i The boundary ranges of are set to [0, 1] and [0, 2] respectively. Since the generated Pareto front provides a set of optimal solutions, the TOPSIS method is applied to rank these Pareto optimal solutions to select the most ideal one.

[0100] Table 4 Parameters of MOGWO algorithm

[0101]

[0102] (3) Results analysis

[0103] In order to verify the reliability of the proposed shield attitude intelligent control method, RMSE, MAE and R 2 The prediction performance of the improved BiLSTM model was comprehensively evaluated. The effectiveness of the shield machine attitude optimization framework was then verified by calculating the percentage improvement in three metrics: settling time (T), overshoot (O), and the integral of the time multiplied by the absolute error (ITAE). The percentage improvement is defined as the relative improvement in a specific performance metric achieved by the adopted shield machine attitude control method compared to the baseline control method. A detailed analysis of the results is provided below.

[0104] (1) Compared with the traditional shield machine, the continuous tunneling shield machine can increase the construction efficiency of a single lining ring by 46.58%. Figure 12 The efficiency comparison between conventional sequential construction and push-and-splice synchronous construction is shown. It can be observed that when the stroke of the propulsion hydraulic cylinder group of the continuous tunneling shield machine extends from the initial 220mm to 482mm, it meets the space requirement for installing the first lining segment B1. At this time, the shield machine switches from the excavation mode to the push-and-splice synchronous construction state. As the cutterhead continues to cut the soil forward, the stroke of the propulsion hydraulic cylinder group gradually increases until it reaches 1377mm. At this time, the excavation of a single lining ring and the installation of six segments have been completed. Therefore, the total time to complete the construction of a single lining ring using a continuous tunneling shield machine is 39 minutes. Figure 12 As shown in the data, using conventional shield construction methods, the total time required to complete a single lining ring is 73 minutes, of which 39 minutes are spent on excavation and 34 minutes on segment installation. Therefore, by applying the simultaneous push-and-splice construction technique, the total construction time can be reduced by 34 minutes. If the construction time of a single lining ring is used as an indicator for evaluating construction efficiency, the efficiency of the continuous tunneling shield machine can be increased by 46.58%.

[0105] (2) The improved BiLSTM model has an R of 0.998 2 The score reliably estimates the shield posture. Table 5 shows the five-fold cross-validation results of the improved BiLSTM model. It can be observed that during the five training iterations, the model always performs slightly better on the training set than on the test set. Specifically, the RMSE, MAE, and R 2 Compared with the training set, the RMSE and MAE on the test set are 0.313mm and 0.145mm higher, respectively, while R 2These results indicate that the model is neither overfitting nor underfitting. To provide a clearer explanation, a more detailed analysis of the test data of the last fold is performed, and the results are as follows Figure 13 and Figure 14 As shown in the figure, the difference between the shield posture predicted by the improved BiLSTM model and the actual shield posture is minimal, with an average prediction error of 0.312 mm. Compared with the deviation range of the actual shield trajectory [-65 mm, 13.8 mm], the model error is acceptable. These results further demonstrate that the proposed model can accurately predict the posture of a continuous tunneling shield machine.

[0106] (3) The pressure of the assembly hydraulic cylinder group, the pressure difference between the upper and lower propulsion hydraulic cylinder groups, and the pressure of the upper reaction hydraulic cylinder group have a significant impact on the model prediction performance. The importance ranking of the decision variables determined using the PFI method is as follows: Figure 15 As shown in Figure 2, it can be found that the first three important factors contributing to the shield posture prediction are the pressure of the assembly hydraulic cylinder group numbered 1, the pressure difference between the upper and lower propulsion hydraulic cylinder groups, and the pressure of the upper reaction hydraulic cylinder group. In order to understand the hidden relationship between specific decision variables and shield posture, Figure 16 The PDP and ICE curves of the first three important decision variables are given. It can be observed that the ICE curves drawn based on different samples are significantly different, which indicates that changes in input variables have different effects on the model prediction results of different samples. This conclusion is supported by the research of Gawde et al. Unlike ICE, PDP reflects the average impact of these features on the prediction of all samples. Figure 16 As shown by the orange curve in , an increase in x5 will reduce the model's prediction. On the contrary, increasing x4 and x21 will improve the model's prediction. To further explore the interaction between the above input variables and their contribution to the model output, Figure 17 An analysis of 2D PDP is given. Figure 17 The contour lines in represent the attitude values of the continuous tunneling shield machine. Figure 17 In (a), the most important variables x5 and x21 are combined. It can be noticed that x5 and x21 have opposite effects on the model’s predictions.

[0107] (4) The proposed intelligent control framework for the attitude of the continuous tunneling shield machine can improve the response performance of the propulsion hydraulic system with an overall improvement percentage of 25.66%. Figure 18A comparison of the response performance of the shield machine system after optimizing PID control parameters using a manual adjustment method and the proposed method is presented. The results demonstrate that this method reliably improves the dynamic and steady-state characteristics of the propulsion hydraulic system. However, the degree of improvement in response performance varies across six different scenarios involving missing assembly hydraulic cylinders. Notably, the overshoot of the continuous tunneling shield machine's attitude after manual and proposed PID controller parameter adjustment is similar, remaining less than 1 mm. Therefore, this paper focuses on analyzing the performance of the proposed method with respect to the T and ITAE metrics. Specifically, in scenario 2, the PID parameters obtained by the proposed method achieve the best average percentage improvement in T and ITAE, reaching 27.43%. In contrast, the method performs slightly worse in scenarios 1, 3, and 6, with average percentage improvements in T and ITAE of 26.74%, 25.76%, and 24.79%, respectively. Further analysis shows that for scenarios 3 and 6, the ITAE improvements achieved by the proposed method exceed the T improvements by 1.22% and 1.91%, respectively. However, for scenario 1, the improvement in T (26.93%) was slightly greater than the improvement in ITAE (26.54%). In scenario 4, the proposed method still reduced T and ITAE by 0.67 s and 214.10 mm·s, respectively, but the optimization was less significant than in other conditions. Nevertheless, the proposed method still achieved an overall improvement of 4.31% compared to the manually adjusted PID parameters. Therefore, the optimal PID parameters obtained using this method can provide effective guidance for operators to optimize the shield posture under different scenarios of missing assembly hydraulic cylinders.

[0108] Table 5 Cross-validation evaluation results of the improved BiLSTM model

[0109]

[0110] Table 6 Comparison of results between manual control and proposed method

[0111]

[0112]

[0113] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVDROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on a carrier medium such as a disk, CD or DVDROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device and its modules of the present invention can be implemented by hardware circuits 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 can be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0114] The above description is only 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 any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent deviation correction control of a continuous tunneling shield machine, characterized in that: The method specifically includes: S1: Conduct push-and-splice synchronous construction experiments to collect data; S2: Integrate BiLSTM and KAN networks to predict the posture of continuous tunneling shield machines; S3: Combined with the MOGWO method, the optimal proportional-integral-derivative (PID) parameters are determined to minimize the posture deviation.

2. The method and system for intelligent deviation correction control of the attitude of a continuous tunneling shield machine according to claim 1, characterized in that: The specific experimental steps of S1 are as follows: (1) Initialize all experimental parameters, system variables and the working status of each component of the continuous tunneling shield machine; (2) Operator adjusts parameters; (3) When the space for installing the first lining ring is prepared, the push-and-spin synchronous construction mode is activated; (4) According to the segment installation sequence, the assembly hydraulic cylinder corresponding to the segment to be installed is retracted. Then, the segment installation machine transports the segment to the designated location, where the installation is completed. This process is repeated until the entire lining ring is assembled. (5) Repeat step 14 until the experiment is completed.

3. The method for intelligent deviation correction control of the attitude of a continuous tunneling shield machine according to claim 1, characterized in that: The input layer of the BiLSTM in step S2 receives raw data for feature extraction and processing, including geological parameters, the pressure difference of the propulsion hydraulic cylinder group, and the pressure of the assembly hydraulic cylinder group. The data then passes through the BiLSTM hidden layer to fully capture and learn the feature information in the data. Afterwards, the output from the BiLSTM hidden layer is transferred to the input layer of KAN; Subsequently, the complex relationships between the variables are further refined and optimized in the hidden layer of KAN to obtain intermediate variables; finally, the output layer of KAN generates the prediction result, which is the shield posture represented by the stroke difference of the propulsion hydraulic cylinder group.

4. The method for intelligent deviation correction control of the attitude of a continuous tunneling shield machine according to claim 1, characterized in that: In S3, a multi-objective optimization framework for intelligent deviation correction of the continuous tunneling shield machine's posture was developed to determine the optimal values of the PID controller parameters. Specifically, this framework combines an improved BiLSTM model, the MOGWO algorithm, and a physical simulation model built in the MATLAB Simulink environment. After using this integrated framework to generate the Pareto front, the TOPSIS method was applied to select the optimal solution, thereby achieving intelligent deviation correction of the continuous tunneling shield machine's posture.

5. An intelligent deviation correction control system for a continuous tunneling shield machine, characterized in that: The system includes: The experimental data collection subsystem is used to perform intelligent deviation correction experiments on the shield machine in the push-and-assemble synchronous construction mode, collecting time series data including the pressure difference of the propulsion hydraulic cylinder group, the pressure of the assembly hydraulic cylinder group, and geological parameters; Bidirectional long short-term memory network module (BiLSTM), used to extract time series features from raw data and construct intermediate state feature vector sequences; Kolmogorov–Arnold network module (KAN), which receives BiLSTM output, models the coupling relationship between nonlinear features, and outputs the shield posture prediction (expressed as the stroke difference of the propulsion hydraulic cylinder group); The multi-objective optimization parameter tuning module performs global search and performance convergence on proportional-integral-derivative (PID) controller parameters based on the multi-objective Grey Wolf Optimization Algorithm (MOGWO) and physical simulation models. The decision integration module selects the optimal control strategy on the Pareto front through the TOPSIS method for attitude closed-loop control.

6. The intelligent control system according to claim 5, characterized in that: The experimental acquisition subsystem supports dynamic sampling mode based on the lining ring installation cycle, including: Initialize the excavation status and various parameter variables; Activate the push-and-splice synchronous construction mode and record the feedback parameters of each hydraulic system during the operation; The system clock is used to drive the sampling mechanism to form a time-aligned data stream for subsequent model training and verification.

7. The intelligent control system according to claim 5, characterized in that: The input layer of the BiLSTM module receives the feature sequence after preprocessing (outlier detection) and encodes the forward and reverse dependencies through a bidirectional gating mechanism, thereby improving the posture prediction model's ability to perceive historical and future trends in the propulsion process. The KAN module uses a learnable activation function mechanism to approximate the objective function, thereby enhancing the interactive expression ability between features.

8. The system according to claim 5, wherein: The MOGWO algorithm integrated in the optimization and parameter adjustment module is built on the basis of the digital twin model in the MATLAB Simulink environment. It simultaneously optimizes three performance indicators: overshoot, adjustment time, and the product integral of time and absolute deviation to form a Pareto optimal solution set. Then, the optimal PID controller parameter combination is selected through the TOPSIS algorithm to realize intelligent correction control of the shield posture.

9. The system according to claim 5, wherein: The posture prediction result output by the KAN module is expressed by the hydraulic cylinder stroke difference, which is used to accurately reflect the posture change trend of the shield machine, such as pitch.

10. The system according to claim 5, wherein: The multiple objectives in the optimization parameter adjustment module are considered with an adjustable weight mechanism. Users can assign different control objective weights according to tunnel working conditions to adapt to the attitude control requirements under different strata and lining ring assembly methods.

Images