Method for predicting and rectifying deviation of large-diameter shield tunneling axis of subsea tunnel

Abstract

The invention relates to the technical field of tunnel shield tunneling, and particularly discloses a subsea tunnel large-diameter shield tunneling axis deviation prediction and correction method, which comprises the following steps: constructing a geology-fluid coupling matrix to obtain a geology-fluid coupling tensor; extracting a stratum heterogeneity feature tensor; performing axis deviation prediction according to a short-time, medium-time and long-time multi-scale sliding window mode to obtain a prediction deviation value and a deviation type; a preliminary deviation correction result is obtained; a closed-loop correction deviation is obtained; obtaining a final axis correction deviation; according to the method, through deep fusion of multi-source heterogeneous data and coupling of a physical information neural network, the seabed environment, geological characteristics and shield mechanics are cooperatively input into a unified prediction framework, and accurate description of tunneling axis deviation essential driving factors is achieved. A coupling matrix of a tidal pressure gradient and salinity ratio is introduced, the recessive influence of a seabed fluid field on the shield attitude is effectively compensated, and the model can maintain prediction stability under complex tide and salinity disturbance.

Description

Technical Field

[0001] This invention belongs to the field of tunnel shield tunneling technology, specifically relating to a method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels. Background Technology

[0002] In the construction of subsea tunnels, large-diameter shield tunneling technology has become a key means of traversing complex seabed geological conditions due to its advantages such as high efficiency and safety. However, due to the extremely complex seabed environment, controlling the axis deviation during shield tunneling faces many severe challenges, and existing technologies are clearly insufficient in addressing these challenges.

[0003] The seabed environment possesses unique geological and fluid characteristics. On one hand, the seabed strata exhibit significant heterogeneity, with vastly different geological structures and soil properties across different regions, including a complex interweaving of soft soil, rock layers, faults, and other geological formations. This heterogeneity results in uneven distribution of ground resistance experienced by the tunnel boring machine (TBM) during excavation, easily leading to TBM attitude deviation and consequently, axial deviation. On the other hand, the seabed fluid environment is complex and variable. Tidal forces cause periodic fluctuations in seawater pressure, while seawater salinity varies depending on region and season. Tidal pressure fluctuations alter the water-soil pressure balance around the TBM, while changes in seawater salinity affect the permeability of pore water in the strata. The combined effect of these two factors exerts a complex influence on the TBM's excavation attitude.

[0004] Existing technologies for predicting axis deviation have several limitations. Most prediction methods rely on only a single or a few data sources, such as mechanical parameters or geological exploration data during tunnel boring machine (TBM) excavation, failing to fully integrate multi-source heterogeneous data from complex seabed environments. For example, they neglect the implicit influence of tidal pressure fluctuations and seawater salinity changes on TBM attitude, leading to inaccurate characterization of the geological-fluid coupling effect and an inability to accurately capture the essential driving factors causing axis deviation. In terms of prediction model construction, existing technologies mostly employ traditional data-driven models or simple physical models, lacking a mechanism for deeply integrating physical information with data-driven approaches. Traditional data-driven models are prone to large prediction errors in complex seabed environments due to data sparsity or overfitting; while simple physical models struggle to accurately reflect complex characteristics such as stratigraphic heterogeneity and cannot adapt to dynamic changes in the seabed environment. Furthermore, most existing prediction methods employ single-timescale prediction strategies, failing to simultaneously address the needs of instantaneous fine-tuning and long-term trend prediction, thus failing to provide comprehensive and accurate information support for deviation correction control.

[0005] In response, this application proposes a method for predicting and correcting the axial deviation of large-diameter shield tunneling in submarine tunnels, in order to solve the above-mentioned problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting and correcting the axial deviation of large-diameter shield tunneling in submarine tunnels, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Methods for predicting and correcting axis deviation during large-diameter shield tunneling of submarine tunnels include: S1: Collect sonar terrain scanning data, tidal pressure fluctuation data and seawater salinity data to construct the geological-fluid coupling matrix G1 and obtain the geological-fluid coupling tensor; S2: Input the geological-fluid coupling tensor and the shield cutterhead vibration spectrum into the three-dimensional convolutional network 3D-GTN, and combine it with the shield machine dynamics equations to extract the formation heterogeneity feature tensor G2; S3: The formation heterogeneity characteristic tensor and historical axis deviation sequence are used to predict the axis deviation in a multi-scale sliding window manner with short-time, medium-time and long-time, so as to obtain the predicted deviation value δpred and the deviation type. S4: Based on the predicted deviation value δpred and the deviation type, use reinforcement learning decision-making to generate hydraulic thrust ΔF, grouting volume Qgrout and segment rotation angle φs, which are used to perform multimodal collaborative correction and obtain preliminary correction results; S5: In the digital twin simulation environment, the preliminary correction results are compared with real-time sensing data to train the error model online and update the prediction network parameters to obtain the closed-loop correction deviation. S6: Feedback the closed-loop correction deviation to the grouting system and cutterhead torque distribution strategy to optimize subsequent construction control and obtain the final axis correction deviation.

[0008] Preferably, the formula for calculating the geological-fluid coupling matrix in S1 is: ; Wherein, α: tidal gradient weight coefficient (dimensionless), to balance the tidal influence; Tidal pressure gradient (Pa / m) is obtained by differentiating continuous time series data; β: Salinity sensitivity coefficient (dimensionless), which modulates the salinity response; Ssal: Real-time seawater salinity (‰); Sref: Reference salinity (‰), regional average.

[0009] By fusing tidal pressure and salinity variations, the system accurately characterizes the seabed fluid environment, providing environmental coupling input for subsequent feature extraction. In the complex geological environment 20m ahead of the tunnel boring machine, under the combined influence of tidal and salinity changes, this matrix fuses two key factors, outputting a "coupling tensor" G1 as the foundational input for subsequent heterogeneous feature extraction. This precisely depicts the coupling effect of seabed tides and salinity, avoiding misjudgments due to overlooked changes in osmotic pressure. It also provides richer input dimensions for the 3D-GTN network, improving the environmental adaptability of the prediction model.

[0010] Preferably, the shield tunneling machine dynamics equations in S2 are: ; Wherein, τbend is the bending moment (N·m) generated during the tunnel boring process. J is the equivalent inertia of the cutterhead and transmission system (kg·m²), reflecting the inertial resistance of the shield tunnel's diagonal acceleration; θ is the shield deflection angle (rad); C is the damping coefficient (N·m·s / rad), representing the viscous damping and mechanical friction of the soil layer; K is the equivalent stiffness coefficient of the stratum (N·m / rad), which reflects the restoring force of the heterogeneous stratum on the tunnel boring machine.

[0011] Before extracting the stratigraphic heterogeneity feature tensor, the above equation is incorporated to constrain the temporal feature learning, so that the network can take into account the physical and dynamic laws when capturing historical deviation trends; the pure data-driven temporal model is coupled with mechanical dynamics to enhance the robustness of response to sudden stratigraphic changes (such as soft soil to karst zone); and the prediction error caused by model overfitting or data sparsity is reduced.

[0012] Preferably, the formula for reinforcement learning decision-making in S3 is: ; Where δ: real-time axis deviation (m); ΔC: Segment misalignment (m); Evib: Vibrational energy (J); ω1, ω2, and ω3 are three weights (dimensionless) that can be adjusted to control the focus.

[0013] The system takes into account multiple objectives, including deviation, assembly accuracy, and vibration, and generates the optimal correction action through strategy optimization to achieve multimodal collaborative control.

[0014] Preferably, the formula for calculating the hydraulic thrust in S4 is: ; Where, ΔFij: the thrust adjustment value of the j-th hydraulic cylinder in the i-th direction; δpred: Predicted attitude deviation (e.g., pitch angle deviation); δ·pred: The first derivative of the deviation (i.e., the rate of change of the deviation); Kp and Kd are the proportional-differential control weights that converge during reinforcement learning training, respectively. Applicable scenarios: Real-time fine-tuning of the shield tunneling front end, especially dynamic feedback control within short-term sliding windows.

[0015] Preferably, the formula for calculating the grouting volume in S4 is:

[0016] Wherein, Q0: the basic value of conventional grouting volume; γ: The increment coefficient determined by reinforcement learning; The integral value of the prediction bias represents the trend of the cumulative bias. η·Ppore: Compensation term for formation pore water pressure Ppore, used to enhance support strength; T: is the analysis period window (e.g., 10 rings).

[0017] Applicable scenarios: When the medium- to long-term prediction results are continuously skewed, it is used to stabilize the tail structure of the tunnel boring machine and the strata.

[0018] Preferably, the formula for calculating the segment rotation angle in S4 is: ; Where: φs: suggested segment rotation angle (around the vertical axis); δpred,y: Horizontal axis deviation value; R: Inner diameter / radius of the shield tunnel, used to convert linear offset into angular compensation.

[0019] Applicable scenarios: Mid-term structural correction measures, corresponding to mid-term sliding windows, to ensure that the assembly direction compensates for deviation trends.

[0020] Preferably, the multi-scale sliding window in S3 includes a short-time sliding window, a medium-time sliding window, and a long-time sliding window: The short-time sliding window is a 30-second short-time sliding window used to generate hydraulic fine-tuning commands; The aforementioned time window is a 10-ring time window, lasting 5-10 minutes, used to optimize segment selection; The long-term sliding window is for long-term trend analysis, ranging from 30 to 60 minutes, and is used for risk warning and pre-adjustment of grouting parameters.

[0021] Preferably, the closed-loop correction deviation described in S5 is used to adjust the grouting pump pressure and cutterhead torque distribution to ensure that the final axis deviation does not exceed ±3mm.

[0022] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention integrates the deep fusion of multi-source heterogeneous data with the coupling of physical information neural networks, and inputs the seabed environment, geological characteristics and shield mechanics into a unified prediction framework, thereby achieving an accurate characterization of the essential driving factors of tunnel axis deviation. First, the coupling matrix of tidal pressure gradient and salinity ratio is introduced to effectively compensate for the implicit influence of the seabed fluid field on the shield attitude, enabling the model to maintain prediction stability under complex tidal and salinity disturbances.

[0023] (2) The multi-scale prediction mechanism in this invention is carried out simultaneously in the three time and ring number dimensions of short, medium and long time, realizing the seamless connection from instantaneous fine-tuning to trend-level decision-making; while the multi-modal collaborative correction strategy organically combines the three means of hydraulic, grouting and segment rotation, and automatically weighs various control objectives through reinforcement learning, which significantly enhances the accuracy and execution efficiency of the correction action.

[0024] (3) Based on the closed-loop verification and online incremental learning mechanism of digital twin, this invention ensures the continuous adaptation and dynamic correction of the model throughout the tunneling process, avoids the accumulation of prediction deviation and sudden increase in error, and theoretically pushes the shield axis control accuracy to the millimeter level and the response speed to the second level, providing an innovative solution for large-diameter submarine tunnel excavation that combines system adaptability and engineering operability. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method for predicting and correcting axis deviation in large-diameter shield tunneling of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: Please refer to Figure 1 As shown, the method for predicting and correcting axis deviation during large-diameter shield tunneling of submarine tunnels includes: S1: Collect sonar terrain scanning data, tidal pressure fluctuation data and seawater salinity data to construct the geological-fluid coupling matrix and obtain the geological-fluid coupling tensor; The formula for calculating the geological-fluid coupling matrix is ​​as follows: ; Wherein, α: tidal gradient weight coefficient (dimensionless), to balance the tidal influence; Tidal pressure gradient (Pa / m) is obtained by differentiating continuous time series data; β: Salinity sensitivity coefficient (dimensionless), which modulates the salinity response; Ssal: Real-time seawater salinity (‰); Sref: Reference salinity (‰), regional average.

[0028] By fusing tidal pressure and salinity variations, the system accurately characterizes the seabed fluid environment, providing environmental coupling input for subsequent feature extraction. In the complex geological environment 20m ahead of the tunnel boring machine, under the combined influence of tidal and salinity changes, this matrix fuses two key factors, outputting a "coupling tensor" G1 as the foundational input for subsequent heterogeneous feature extraction. This precisely depicts the coupling effect of seabed tides and salinity, avoiding misjudgments due to overlooked changes in osmotic pressure. It also provides richer input dimensions for the 3D-GTN network, improving the environmental adaptability of the prediction model.

[0029] S2: Input the geological-fluid coupling tensor and the shield cutterhead vibration spectrum into the three-dimensional convolutional network 3D-GTN, and combine it with the shield machine dynamics equations to extract the formation heterogeneity feature tensor; The dynamic equations of the tunnel boring machine are: ; Wherein, τbend is the bending moment (N·m) generated during the tunnel boring process. J is the equivalent inertia of the cutterhead and transmission system (kg·m²), reflecting the inertial resistance of the shield tunnel's diagonal acceleration; θ is the shield deflection angle (rad); C is the damping coefficient (N·m·s / rad), representing the viscous damping and mechanical friction of the soil layer; K is the equivalent stiffness coefficient of the stratum (N·m / rad), which reflects the restoring force of the heterogeneous stratum on the tunnel boring machine.

[0030] Before extracting the stratigraphic heterogeneity feature tensor, the above equation is incorporated to constrain the temporal feature learning, so that the network can take into account the physical and dynamic laws when capturing historical deviation trends; the pure data-driven temporal model is coupled with mechanical dynamics to enhance the robustness of response to sudden stratigraphic changes (such as soft soil to karst zone); and the prediction error caused by model overfitting or data sparsity is reduced.

[0031] S3: The formation heterogeneity characteristic tensor and historical axis deviation sequence are used to predict the axis deviation in a multi-scale sliding window manner with short-time, medium-time and long-time, so as to obtain the predicted deviation value δpred and the deviation type. The multi-scale sliding window includes short-time sliding windows, medium-time sliding windows, and long-time sliding windows: The short-time sliding window is a 30-second short-time sliding window used to generate hydraulic fine-tuning commands; The aforementioned time window is a 10-ring time window, lasting 5-10 minutes, used to optimize segment selection; The long-term sliding window is for long-term trend analysis, ranging from 30 minutes to 1 hour, and is used for risk warning and pre-adjustment of grouting parameters.

[0032] S4: Based on the predicted deviation value δpred and the deviation type, use reinforcement learning decision-making to generate hydraulic thrust ΔF, grouting volume Qgrout and segment rotation angle φs, which are used to perform multimodal collaborative correction and obtain preliminary correction results; The formula for the reinforcement learning decision is: ; Where δ: real-time axis deviation, in meters; ΔC: Segment misalignment, in meters (m); Evib: Vibrational energy, unit J; ω1, ω2, ω3: Three weights that can be adjusted to control the focus; The formula for calculating the hydraulic thrust is:

[0033] Where, ΔFij: the thrust adjustment value of the j-th hydraulic cylinder in the i-th direction; δpred: Predicted attitude deviation (e.g., pitch angle deviation); δ·pred: The first derivative of the deviation (i.e., the rate of change of the deviation); Kp,Kd: The proportional-differential control weights converged during reinforcement learning training; The formula for calculating the grouting volume is: ; Wherein, Q0: conventional grouting volume (basic value); γ: The increment coefficient determined by reinforcement learning; The integral value of the prediction bias represents the trend of the cumulative bias. η·Ppore: Compensation term for formation pore water pressure Ppore, used to enhance support strength; T: is the analysis period window (e.g., 10 rings); The formula for calculating the segment rotation angle is: ; Where φs: the suggested segment rotation angle (around the vertical axis); δpred,y: Horizontal axis deviation value; R: Inner diameter / radius of the shield tunnel, used to convert linear offset into angular compensation.

[0034] S5: In the digital twin simulation environment, the preliminary correction results are compared with real-time sensing data to train the error model online and update the prediction network parameters to obtain the closed-loop correction deviation. The closed-loop correction deviation is used to adjust the grouting pump pressure and cutterhead torque distribution to ensure that the final axis deviation does not exceed ±3mm. S6: Feedback the closed-loop correction deviation to the grouting system and cutterhead torque distribution strategy to optimize subsequent construction control and obtain the final axis correction deviation.

[0035] As shown above, by deeply fusing multi-source heterogeneous data and coupling it with a physical information neural network, the seabed environment, geological characteristics, and shield tunneling mechanics are collaboratively input into a unified prediction framework, achieving an accurate characterization of the essential driving factors of tunneling axis deviation. Firstly, the introduction of a coupling matrix between tidal pressure gradient and salinity ratio effectively compensates for the implicit influence of the seabed fluid field on the shield tunneling attitude, enabling the model to maintain prediction stability under complex tidal and salinity disturbances.

[0036] Example 2: A method for predicting and correcting axis deviation during large-diameter shield tunneling of submarine tunnels, including the following steps: Environment and Equipment: Shield cutterhead diameter: 14.5m; equivalent moment of inertia J = 1.2 × 10⁶ kg·m 2 ; Sensor: Tidal pressure sensor (TP-X200, ±0.2 kPa, 0.5 Hz); Salinity sensor (SS-300, ±0.1‰, 0.2Hz); Sonar topographic scanner (HY-S700, ±0.05m, 1Hz); FBG vibration sensing network (1kHz) + IMU (200Hz); S1: Acquire sonar topographic scan data, tidal pressure fluctuation data, and seawater salinity data to construct the geological-fluid coupling matrix and obtain the geological-fluid coupling tensor. Data acquisition: Tidal pressure and salinity data were continuously collected for 10 minutes in an area 20m ahead of the 120th ring.

[0037] Calculation process: Tidal pressure gradient is obtained by difference. =0.01 kPa / m.

[0038] The average salinity Ssal = 34.7‰ and the reference salinity Sref = 35.0‰ are taken. Assigning coefficients α=0.8 and β=1.1, substituting them yields: =0.8×0.01+1.1×ln(34.7 / 35.0)=0.008-0.0100=-0.0020; The combined effects of seabed tides and salinity on the tunnel boring machine's attitude are quantified by using a coupling matrix, providing a stable input for subsequent spatial-temporal feature extraction.

[0039] S2: The geological-fluid coupling tensor and the shield cutterhead vibration spectrum are input into a three-dimensional convolutional network 3D-GTN, and combined with the shield machine dynamics equations to extract the formation heterogeneity feature tensor: Input: Ground penetrating radar slice: 80×80×40 voxels; Cutter head vibration spectrum: 0-600Hz, 512-point FFT; Network and formula integration: 3D-GTN extracts the spatial tensor G2∈R16×16×8.

[0040] Incorporating the kinetic equations into Bi-STGU: ; (Take C = 4.5 × 10) 4 K = 2.0 × 10 5 This allows the timing update gate to take mechanical response into account.

[0041] We obtained a high-dimensional feature sequence that integrates stratigraphic heterogeneity and mechanical dynamics.

[0042] S3: Using the formation heterogeneity characteristic tensor and historical axis deviation sequence, the axis deviation is predicted using a multi-scale sliding window method (short-time, medium-time, and long-time) to obtain the predicted deviation value and deviation type. Mechanism: Short time (30s) → hydraulic fine adjustment; medium time (10 rings ≈ 120s) → segment selection; long time (500s) → trend warning.

[0043] Physical constraints are introduced into the loss function to ensure that the acceleration output by the network is consistent with the kinematics of the shield tunneling theory.

[0044] Prediction result: δpred = +0.012 m.

[0045] S4: Based on the predicted deviation value and deviation type, use reinforcement learning decision-making to generate hydraulic thrust, grouting volume and segment rotation angle, which are used to perform multimodal collaborative correction and obtain preliminary correction results; Input: δpred and deviation type (horizontal + torsion); Reinforcement learning reward function: ; The weights ω1:ω2:ω3 = 5:2:1.

[0046] Output action: The hydraulic thrust increment ΔF = +1.3 × 10 5 N; Grouting volume Qgrout = 0.9 m 3 ; The segment rotation angle φs = 2.0°; S5: In the digital twin simulation environment, the preliminary correction results are compared with real-time sensing data to train the error model online and update the prediction network parameters to obtain the closed-loop correction deviation. Comparison: Measured deviation δmeas = +0.014 m; Condition: MAE(0.012,0.014)=0.002 m<εth(0.005 m), online learning is not triggered.

[0047] S6: Feedback the closed-loop correction deviation to the grouting system and cutterhead torque distribution strategy to optimize subsequent construction control and obtain the final axis correction deviation.

[0048] Application: Maintain the above correction parameters; Result: The final deviation converged to +0.002 m.

[0049] Example 3: A method for predicting and correcting axis deviation during large-diameter shield tunneling in submarine tunnels, including the following steps: Environment and Equipment: Cutter head diameter: 13.8m; J = 1.0 × 10⁶ kg·m 2 ; Other sensor configurations are the same as in Example 2; S1: Collect sonar terrain scanning data, tidal pressure fluctuation data and seawater salinity data to construct the geological-fluid coupling matrix and obtain the geological-fluid coupling tensor; =0.012 kPa / m, Ssal=34.9 ‰; Coefficients α = 0.85, β = 1.0. G1=0.85×0.012+1.0×ln(34.9 / 35.0)=0.0102-0.0029=0.0073; S2: Input the geological-fluid coupling tensor and the shield cutterhead vibration spectrum into the three-dimensional convolutional network 3D-GTN, and combine it with the shield machine dynamics equations to extract the formation heterogeneity feature tensor; 3D-GTN outputs the formation heterogeneity characteristic tensor G2; Bi-STGU is incorporated into τbend (C=4.0×10 4 K = 2.2 × 10 5 ).

[0050] S3: The formation heterogeneity characteristic tensor and historical axis deviation sequence are used to predict axis deviation in a multi-scale sliding window manner with short-time, medium-time and long-time, so as to obtain the predicted deviation value and deviation type. Synchronization window configuration; The loss function is constrained to λ=0.1; δpred=−0.009 m.

[0051] S4: Based on the predicted deviation value and deviation type, use reinforcement learning decision-making to generate hydraulic thrust, grouting volume and segment rotation angle, which are used to perform multimodal collaborative correction and obtain preliminary correction results; Vertical deviation; RL reward weights 5:2:1; Output: ΔF = −1.1 × 10 5 N, Qgrout = 1.0 m 3 φs=0°.

[0052] S5: In the digital twin simulation environment, the preliminary correction results are compared with real-time sensing data to train the error model online and update the prediction network parameters to obtain the closed-loop correction deviation. Measured δmeas = −0.011 m, MAE = 0.002 m < 0.005 m → No triggering; S6: Feedback the closed-loop correction deviation to the grouting system and cutterhead torque distribution strategy to optimize subsequent construction control and obtain the final axis correction deviation.

[0053] Final deviation –0.001m.

[0054] Comparative example: Traditional error correction process: Bias prediction: Method: Based on a unidirectional LSTM model, deviation prediction is performed using historical axis sensing data; Output: Prediction bias value δ trad ;

[0055] Hydraulic thrust adjustment (ΔF) trad ) Strategy: Use a fixed PID controller and perform closed-loop control based on the predicted deviation. ; Features: Parameters , , Parameters were adjusted based on on-site experience. It cannot adaptively adjust dynamically based on geological changes or vibration characteristics; The adjustment range and response time are both fixed, typically with a delay of 20–30 seconds.

[0056] Grouting volume determined by (Q_grout_trad) Strategy: Preset constants + experience-based adjustments, no real-time trend compensation. ; Note: The basic grouting volume Q0trad is set based on the average permeability of the formation; It is issued manually by the site engineer only after a clear cavity or collapse is measured, usually with a lag of 5–10 minutes.

[0057] Segment rotation angle (Ф_segment_trad) Strategy: Generally, segment rotation compensation is not performed, or manual rotation adjustment at a fixed angle (such as ±3°) is only adopted after the cumulative deviation exceeds 0.05m; Note: Primarily relies on the engineer's experience and judgment; Structural correction is delayed and can only be implemented after the large ring assembly is completed, with a response cycle ranging from tens of minutes to several hours. The data from Example 2, Example 3, and the comparative example are compared in Table 1 below: Table 1

[0058] As can be seen from the above, the multi-scale prediction mechanism unfolds simultaneously in the short, medium and long time and ring number dimensions, realizing a seamless connection from instantaneous fine-tuning to trend-level decision-making; while the multi-modal collaborative correction strategy organically combines hydraulic, grouting and segment rotation methods, and automatically weighs various control objectives through reinforcement learning, significantly enhancing the accuracy and execution efficiency of correction actions. The closed-loop verification and online incremental learning mechanism based on digital twins ensures the continuous adaptation and dynamic correction of the model throughout the tunneling process, avoiding the accumulation of prediction deviations and sudden increases in errors. Theoretically, it pushes the shield axis control accuracy to the millimeter level and the response speed to the second level, providing an innovative solution for large-diameter subsea tunnel excavation that combines system adaptability and engineering operability.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and correcting axis deviation during large-diameter shield tunneling of submarine tunnels, characterized in that... Includes the following steps: S1: Collect sonar terrain scanning data, tidal pressure fluctuation data and seawater salinity data to construct the geological-fluid coupling matrix and obtain the geological-fluid coupling tensor; S2: Input the geological-fluid coupling tensor and the shield cutterhead vibration spectrum into a three-dimensional convolutional network 3D-GTN, and combine it with the shield machine dynamics equations to extract the formation heterogeneity feature tensor; S3: The formation heterogeneity characteristic tensor and historical axis deviation sequence are used to predict axis deviation in a multi-scale sliding window manner with short-time, medium-time and long-time, so as to obtain the predicted deviation value and deviation type. S4: Based on the predicted deviation value and deviation type, use reinforcement learning decision-making to generate hydraulic thrust, grouting volume and segment rotation angle, which are used to perform multimodal collaborative correction and obtain preliminary correction results; S5: In the digital twin simulation environment, the preliminary correction results are compared with real-time sensing data to train the error model online and update the prediction network parameters to obtain the closed-loop correction deviation. S6: Feedback the closed-loop correction deviation to the grouting system and cutterhead torque distribution strategy to optimize subsequent construction control and obtain the final axis correction deviation.

2. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The formula for calculating the geological-fluid coupling matrix mentioned in S1 is: ； Wherein, α: tidal gradient weighting coefficient, used to balance the effects of tides; Tidal pressure gradient, in Pa / m, obtained by differential analysis of continuous time series data; β: Salinity sensitivity coefficient, which regulates the salinity response; Ssal: Real-time seawater salinity, in ‰; Sref: Reference salinity, in ‰, regional average.

3. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The shield tunneling machine dynamics equations described in S2 are: ； Wherein, τbend is the bending moment generated during shield tunneling, in N·m; J is the equivalent inertia of the cutterhead and transmission system, in kg·m², reflecting the inertial resistance of the shield tunnel's diagonal acceleration. θ is the shield deflection angle, in rad; C is the damping coefficient, with units of N·m·s / rad, representing the viscous damping and mechanical friction of the soil layer; K is the equivalent stiffness coefficient of the stratum, with units of N·m / rad, which reflects the restoring force of the heterogeneous stratum on the tunnel boring machine.

4. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The formula for reinforcement learning decision-making described in S3 is: ； Where δ: real-time axis deviation, in meters; ΔC: Segment misalignment, in meters (m); Evib: Vibrational energy, unit J; ω1, ω2, and ω3 are three weights that can be adjusted to control the focus.

5. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The formula for calculating the hydraulic thrust mentioned in S4 is: ； Where, ΔFij: the thrust adjustment value of the j-th hydraulic cylinder in the i-th direction; δpred: Predicted attitude deviation value; δ·pred: the first derivative of the deviation; Kp and Kd are the proportional-differential control weights that converge during reinforcement learning training.

6. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The formula for calculating the grouting volume mentioned in S4 is as follows: ； Wherein, Q0: the basic value of conventional grouting volume; γ: The increment coefficient determined by reinforcement learning; The integral value of the prediction bias represents the trend of the cumulative bias. η·Ppore: Compensation term for formation pore water pressure Ppore, used to enhance support strength; T: The analysis period window.

7. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The formula for calculating the segment rotation angle mentioned in S4 is: ； Where: φs: suggested segment rotation angle; δpred,y: Horizontal axis deviation value; R: Inner diameter / radius of the shield tunnel, used to convert linear offset into angular compensation.

8. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The multi-scale sliding window described in S3 includes short-time sliding windows, medium-time sliding windows, and long-time sliding windows: The short-time sliding window is a 30-second short-time sliding window used to generate hydraulic fine-tuning commands; The aforementioned time window is a 10-ring time window, lasting 5-10 minutes, used to optimize segment selection; The long-term sliding window is for long-term trend analysis, ranging from 30 minutes to 1 hour, and is used for risk warning and pre-adjustment of grouting parameters.

9. The method for predicting and correcting axis deviation in large-diameter shield tunneling for submarine tunnels according to claim 1, characterized in that, The closed-loop correction deviation described in S5 is used to adjust the grouting pump pressure and cutterhead torque distribution to ensure that the final axis deviation does not exceed ±3mm.

Images