An automated adjustment method and system for TBM tunneling process

Abstract

This invention relates to the field of intelligent control technology for tunnel boring equipment, and more specifically, to an automated adjustment method and system for the TBM shield tunneling process. It addresses the shortcomings of existing technologies, which cannot construct a dynamic causal knowledge graph rooted in geological conditions, cannot integrate Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game mechanisms, and struggle to achieve interpretable modeling and adaptive decision-making for the "geology-equipment-control" coupling relationship. This invention constructs a dynamic causal knowledge graph rooted in geological conditions, integrates Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game mechanisms, and achieves interpretable modeling and adaptive decision-making for the "geology-equipment-control" coupling relationship. It can capture causal changes caused by geological evolution online, support rapid strategy transfer and hierarchical action generation, and output precise control commands that conform to engineering priorities.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for tunnel boring equipment, and more specifically, to an automated adjustment method and system for the TBM shield tunneling process. Background Technology

[0002] In traditional TBM construction, tunneling parameters are often set manually. The surrounding rock condition, control parameters, and support schemes during the tunneling process mostly rely on engineers' analysis and judgment. Due to the lack of experience of some personnel, problems such as machine jamming and local instability can easily occur.

[0003] Referring to patent application CN115217484A, an automatic control method and system for tunnel boring machines (TBMs) is disclosed to address the problem that existing TBMs rely excessively on the experience and work attitude of the main operator, leading to difficulties in controlling tunnel formation quality. The control system of this invention includes an expert system connected to a PLC controller. The expert system comprises a knowledge base, an inference engine, an interpreter, and a human-machine interface (HMI). The knowledge base and HMI are both connected to the inference engine, and both are connected to the interpreter. The interpreter is connected to the HMI. The inference engine is connected to the PLC controller via a host computer, and the knowledge base is connected to the host computer. This invention fully utilizes expert knowledge and rules in the TBM tunneling process to provide an automatic control scheme for the entire TBM construction process. This invention has a high degree of automation and intelligence and is easy to implement, improving the intelligence level of existing tunneling control and attitude control, ensuring equipment and personnel safety, and significantly improving TBM construction efficiency.

[0004] However, while the aforementioned reference patents achieve intelligent automatic control covering the entire process of startup, ascent, stabilization, temporary shutdown, and shutdown by integrating expert knowledge of tunnel boring machines (TBMs) with multi-factor collaborative control, significantly improving tunneling efficiency and reducing the load on the main operator while ensuring construction safety and accuracy, they cannot construct a dynamic causal knowledge graph based on geological conditions. They also cannot integrate Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game mechanisms, making it difficult to achieve interpretable modeling and adaptive decision-making for the coupling relationship between geology, equipment, and control. Furthermore, they cannot integrate the main shaft attitude, stratum curvature, and circumferential support stress field to construct a coupled control model, and cannot adopt a strategy-execution dual-layer architecture to achieve high-precision trajectory planning and dynamic support coordination, thus reducing the controllability of the TBM trajectory, the timeliness of support, and the robustness of the system in complex curvature strata.

[0005] To address these issues, we propose an automated adjustment method and system for the TBM (Tunnel Boring Machine) tunneling process. Summary of the Invention

[0006] The purpose of this invention is to provide an automated adjustment method and system for the TBM (Tunnel Boring Machine) process. This addresses the problems of existing technologies, which cannot construct a dynamic causal knowledge graph based on geological conditions, cannot integrate Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game mechanisms, and struggle to achieve interpretable modeling and adaptive decision-making for the coupling relationship between geology, equipment, and control. Furthermore, these technologies cannot integrate the main shaft attitude, stratum curvature, and circumferential support stress field to construct a coupled control model, and cannot employ a strategy-execution dual-layer architecture to achieve high-precision trajectory planning and dynamic support coordination, thus reducing the controllability of the TBM trajectory, the timeliness of support, and the robustness of the system in complex curvature strata.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] An automated adjustment method for TBM (Tunnel Boring Machine) tunneling process includes the following steps:

[0009] Step 1: Receive the raw parameters during the shield tunneling process, preprocess the raw parameters during the shield tunneling process, and output the time series of the preprocessed raw parameters;

[0010] Step 2: Based on the preprocessed original parameter time series, construct a causal knowledge graph with geological state representation as the root variable. Dynamically adjust the structure of the causal knowledge graph based on Bayesian structure learning, and generate adaptive decision-making strategies by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory.

[0011] Step 3: Integrate the main shaft attitude, stratum curvature and circumferential support stress field to construct a coupled control model. Adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching.

[0012] Step 4: Based on the fuzzy PID control structure, the feed speed, cutter head speed and screw conveyor speed are adjusted in real time in a closed loop, responding to the multivariable control commands output by the decision strategy;

[0013] Step 5: Based on continuous alignment of original parameters, control commands and execution deviations, monitor the system operation status in real time, identify data anomalies, command conflicts or tracking instability, and trigger fault tolerance mechanisms through multi-source redundant information.

[0014] In a preferred embodiment of the present invention, the process of constructing a causal knowledge graph with geological state characterization as the root variable based on the preprocessed original parameter time series in step two includes:

[0015] Obtain a preprocessed raw parameter time series, calculate the corresponding geological state variable time series, label the stratum strength, stratum stability, tunneling resistance level, and geological hardness index as exogenous variables, and label the soil chamber pressure, total thrust, cutterhead torque, shield pitch angle, thrust speed, cutterhead motor current, and screw conveyor discharge flow rate as endogenous variables, and merge the time series of exogenous and endogenous variables into the observation dataset of the current segment;

[0016] For any two variables in the observed dataset, traverse all possible subsets of conditional variables and perform conditional independence tests. If any two variables are determined to be conditionally independent under a certain subset of conditional variables, remove the connection edge between the two variables in the variable graph. After completing all conditional independence tests, an undirected graph is formed.

[0017] Traverse all triplet structures in the undirected graph. If two variables in a triplet are connected to a third variable but there is no connecting edge between the two variables, then orient the two edges to the third variable. Repeat the following two operations based on the existing directed edges until no more directed edges are added to the variable graph.

[0018] Output the local causal graph corresponding to the current time segment. Repeat all the above operations to process each time segment in the preprocessed original parameter time series in turn, merge all local causal graphs, and generate a complete causal knowledge graph.

[0019] In a preferred embodiment of the present invention, step two, which involves dynamically adjusting the structure of the causal knowledge graph based on Bayesian structural learning and combining model-independent meta-learning mechanisms, hierarchical reinforcement learning, and multi-objective game theory to generate adaptive decision-making strategies, includes:

[0020] Obtain the geological state characterization of the current time step, obtain the construction control parameter observation values ​​of the current time step, obtain the observation data of the most recent fixed number of time steps, form the current window data, and attempt to perturb the structure based on the current causal graph. For the perturbed causal graph structure, calculate its matching score with the current window data. If the score of the perturbed causal graph is higher than that of the current causal graph, then use it as the new causal graph; otherwise, retain the current causal graph.

[0021] Load the initial policy parameters that match the current causal graph from the policy library. Based on the current geological task identifier, construct the policy evaluation context using the variable dependencies depicted by the latest causal graph. Calculate the loss value of the current policy under this task. Perform a gradient update on the initial policy parameters based on this loss value to obtain the optimized adaptation parameters.

[0022] The geological state characterization is input into the high-level strategy network, which outputs the control mode identifier. The control mode identifier and the current state are input into the low-level strategy network, which outputs the original control action. Four target values ​​are calculated: tunneling distance per unit time, equipment mechanical and electrical load, energy consumption per unit advance, and shield attitude deviation. Among all candidate strategies, the strategy that meets the following conditions is selected as the final strategy. Six instructions are extracted from the final strategy: total thrust setting value, cutterhead torque setting value, thrust speed setting value, motor current upper limit value, cutterhead speed setting value, and screw conveyor speed setting value.

[0023] In a preferred embodiment of the present invention, the process of constructing a coupled control model by integrating the principal axis attitude, formation curvature, and circumferential support stress field in step three includes:

[0024] Acquire shield machine attitude sensor data, extract the main shaft attitude angle from the attitude sensor data, acquire circumferential stress sensor data within the tunnel segment, extract the circumferential support stress component associated with the main shaft attitude from the stress sensor data, acquire the vertical displacement sequence and corresponding axial coordinate sequence of the tunneling trajectory, calculate the second derivative of the vertical displacement with respect to the axial coordinate, output the ground curvature, acquire the current control input, and send the main shaft attitude angle, circumferential support stress component, ground curvature, and current control input to the coupled evolution calculation unit. In the coupled evolution calculation unit, calculate the state at the next moment, output the predicted value of the main shaft attitude angle at the next moment, and output the predicted value of the circumferential support stress component at the next moment.

[0025] In a preferred embodiment of the present invention, step three employs a strategy-execution two-layer architecture, combining fuzzy prediction, adaptive loop closure, and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching.

[0026] The system obtains the spindle attitude angle and the stratum curvature. Based on the spindle attitude angle and the stratum curvature, it calculates the initial tunneling trajectory, obtains the circumferential support stress, the advance speed, and the cutterhead torque. Based on the circumferential support stress, the advance speed, and the cutterhead torque, it generates operational safety constraints, obtains the soil chamber pressure, and obtains the spatial distribution of stratum strength. Based on the spatial distribution of stratum strength, it queries a predefined fuzzy rule table, matches the control action corresponding to the current stratum strength from the fuzzy rule table, and outputs the matched control action as an auxiliary instruction.

[0027] Obtain the immediate reward, obtain the current state vector, obtain the current control action, invoke the reinforcement learning update process, calculate the state-action value function correction during the reinforcement learning update process, and generate the final reference control sequence based on the action output by reinforcement learning, combined with the direction guidance of the initial tunneling trajectory, the boundary constraints of the running safety constraints, and the auxiliary instruction correction provided by fuzzy rules.

[0028] The system acquires the actual feedback values ​​of each actuator, calculates the deviation between the reference control sequence and the actual feedback values, calculates the proportional component, integral component, and differential component of the deviation, and adjusts the proportional gain, integral gain, and differential gain online based on the current formation strength and the main shaft attitude deviation and their changing trends. The proportional component, integral component, and differential component are multiplied by their respective gains and then superimposed to generate the final drive signal. The final drive signal is then output to the propulsion system, cutterhead drive system, and grouting system.

[0029] Obtain the current circumferential support stress value, determine whether the current circumferential support stress value exceeds the safety threshold. If the current circumferential support stress value exceeds the safety threshold, switch to the next level of support mode in the preset support mode sequence. If the current support mode is already the highest level, maintain the current mode and output an alarm signal.

[0030] In a preferred embodiment of the present invention, step four, which involves real-time closed-loop adjustment of the propulsion speed, cutterhead speed, and screw conveyor speed based on a fuzzy PID control structure, and responding to the multivariable control commands output by the decision strategy, includes:

[0031] The propulsion speed setpoint and actual value are obtained, the current deviation of the propulsion speed is calculated, the propulsion speed deviation of the previous cycle is obtained, the change of propulsion speed deviation is calculated, fuzzy PID parameter self-tuning is performed based on the current deviation and the change of deviation, and the propulsion speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change of deviation.

[0032] Obtain the setpoint and actual value of the cutter head speed, calculate the current deviation of the cutter head speed, obtain the deviation of the cutter head speed in the previous cycle, calculate the change in the deviation of the cutter head speed, perform fuzzy PID parameter self-tuning based on the current deviation and the change in deviation, and combine the self-tuned parameters, current deviation, deviation integral and deviation change to calculate the cutter head speed control output and output the cutter head speed control command.

[0033] The setpoint and actual values ​​of the screw conveyor speed are obtained, the current deviation of the screw conveyor speed is calculated, the deviation of the screw conveyor speed in the previous cycle is obtained, the change in the screw conveyor speed deviation is calculated, fuzzy PID parameter self-tuning is performed based on the current deviation and the change in deviation, and the screw conveyor speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change in deviation, and the screw conveyor speed control command is output.

[0034] In a preferred embodiment of the present invention, step five, which involves continuously aligning the original parameters, control commands, and execution deviations, monitoring the system's operating status in real time, identifying data anomalies, command conflicts, or tracking instability, and triggering a fault-tolerance mechanism through multi-source redundant information, includes:

[0035] Obtain the status observation value of the main channel, determine whether the status observation value of the main channel is valid, if it is invalid, obtain the status observation value of the redundant channel, determine whether the status observation value of the redundant channel is valid, if the redundancy is valid, adopt the redundant value, otherwise adopt the valid status of the previous cycle.

[0036] Acquire three independent state perception outputs, determine whether there is at least two outputs with a difference less than the preset tolerance. If so, take the average value as the valid state; otherwise, determine that the state perception has failed. Further verify whether the valid state is time-synchronized and within a reasonable range. If the verification fails, output a zero-value control command and trigger a fault report.

[0037] The system acquires decision control commands and monitoring control commands, calculates the difference between them, and if the difference exceeds the allowable range, it is determined to be a command conflict. The execution of the current command is suspended, and a preset safe and conservative command is adopted instead, while triggering a conflict alarm.

[0038] Obtain the current execution deviation and historical execution deviation sequence, determine whether the deviation shows a continuous upward trend, if the trend duration reaches the upper limit, determine that the tracking is unstable, if the instability is determined, load the backup control parameters, and start the model calibration process;

[0039] If the identified data is abnormal, the redundant channel is activated or the previous valid state is maintained, and the range of control command changes is limited. If the identified commands conflict, the system switches to a safe and conservative control mode and suspends high-risk actions. If the identification tracking becomes unstable, the tunneling parameters are downgraded.

[0040] As a preferred embodiment of the present invention, an automated adjustment system for the TBM (tunnel boring machine) process includes:

[0041] TBM Multi-Source Sensing Processing Module: Used to receive raw parameters during the shield tunneling process, preprocess the raw parameters during the shield tunneling process, and output the preprocessed raw parameter time series;

[0042] TBM Adaptive Parameter Tuning Module: Based on the preprocessed original parameter time series, a causal knowledge graph with geological state representation as the root variable is constructed. The structure of the causal knowledge graph is dynamically adjusted based on Bayesian structure learning. Adaptive decision-making strategies are generated by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory.

[0043] TBM attitude support coordination module: used to integrate the main shaft attitude, stratum curvature and circumferential support stress field, construct a coupled control model, adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching.

[0044] TBM parameter execution module: Based on the fuzzy PID control structure, it performs real-time closed-loop adjustment of the feed speed, cutter head speed and screw conveyor speed, and responds to the multivariable control commands output by the decision strategy;

[0045] TBM safety fault tolerance module: Based on continuous alignment of original parameters, control commands and execution deviations, it monitors the system operating status in real time, identifies data anomalies, command conflicts or tracking instability, and triggers fault tolerance mechanisms through multi-source redundant information.

[0046] Compared with the prior art, the advantages of this invention are:

[0047] In this invention, by constructing a dynamic causal knowledge graph based on geological conditions, and integrating Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game mechanism, interpretable modeling and adaptive decision-making of the coupling relationship between "geology-equipment-control" are achieved. It can capture causal changes caused by geological evolution online, support rapid policy transfer and hierarchical action generation, and find Pareto optimal solutions among safety, efficiency, energy consumption, and attitude objectives. It outputs precise control commands that meet engineering priorities, significantly improving the system's autonomy, robustness, and construction safety in complex strata.

[0048] In this invention, a coupled control model is constructed by integrating the main shaft attitude, stratum curvature, and circumferential support stress field. A strategy-execution dual-layer architecture is adopted to achieve high-precision trajectory planning and dynamic support coordination. The shield attitude and support response are predicted synchronously through nonlinear evolution mapping, which strengthens the modeling of soil-machine-structure coupling. Combined with fuzzy rule guidance, reinforcement learning optimization, and adaptive PID closed-loop control, multi-mode support switching is triggered based on real-time stress state. While ensuring structural safety, tunneling efficiency and energy consumption are also taken into account, significantly improving the controllability of the shield trajectory, the timeliness of support, and the robustness of the system in complex curvature strata. Attached Figure Description

[0049] Figure 1 This is a flowchart of the automated adjustment method for the TBM tunneling process in this invention.

[0050] Figure 2 This is a system block diagram of Embodiment 3 of the present invention. Detailed Implementation

[0051] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0052] Example 1: As Figure 1 As shown, the present invention proposes an automated adjustment method for the TBM (Tunnel Boring Machine) process, which includes the following steps:

[0053] Step 1: Receive the raw parameters during the shield tunneling process, including soil chamber pressure, total thrust, cutterhead torque, shield pitch angle, thrust speed, cutterhead motor current, and auger conveyor discharge flow rate. Preprocess the raw parameters during the shield tunneling process. The preprocessing operations include time synchronization, range verification, outlier detection and correction, noise filtering, and missing value interpolation. Output the preprocessed raw parameter time series.

[0054] Step one involves preprocessing the raw parameters of the tunnel boring machine (TBM) (such as soil chamber pressure, total thrust, cutterhead torque, etc.) through time synchronization, range verification, outlier detection and correction, noise filtering, and missing value interpolation. This significantly improves the quality, consistency, and reliability of the data. It not only effectively eliminates sensor errors and interference noise, ensuring accurate alignment of multi-source parameters under a unified time reference, but also provides stable, continuous, and clean high-quality time series inputs for subsequent status identification, risk warning, intelligent control, and high-precision modeling. This enhances the system's robustness, supports automated decision-making, and lays a solid foundation for building a historical database and promoting the transformation of TBM construction towards data-driven approaches.

[0055] Step 2: Based on the preprocessed original parameter time series, construct a causal knowledge graph with geological state representation as the root variable. Dynamically adjust the structure of the causal knowledge graph based on Bayesian structure learning, and generate adaptive decision-making strategies by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory.

[0056] Step two involves constructing a causal knowledge graph with geological state characteristics as the root variable based on the preprocessed original parameter time series.

[0057] Obtain a preprocessed raw parameter time series, which includes earth chamber pressure, total thrust, cutterhead torque, shield pitch angle, thrust speed, cutterhead motor current, and screw conveyor discharge flow rate. Calculate the corresponding geological state variable time series based on the engineering model or expert rules, including formation strength (continuous numerical value), formation stability (binary discrete category, valued as "stable" or "unstable"), tunneling resistance level (ordered discrete variable, valued as level one, two, or three), and geological hardness index (continuous numerical value). Combine these four geological parameters... State variables are labeled as exogenous variables (i.e., potential root causes), while soil pressure, total thrust, cutterhead torque, shield pitch angle, thrust speed, cutterhead motor current, and screw conveyor discharge flow rate are labeled as endogenous variables. To capture temporal causal dependencies, an extended variable set containing the current time t and the previous time t-1 is constructed, that is, the values ​​of all variables at times t and t-1 are included in the observation. For each time segment (e.g., a sliding window of length L with a step size of S), the values ​​of exogenous and endogenous variables at times t and t-1 are merged to form the observation dataset for the current segment.

[0058] For any two variables in the observed dataset, traverse all possible subsets of conditional variables and perform conditional independence tests. If any two variables are determined to be conditionally independent under a certain subset of conditional variables, remove the connection edge between the two variables in the variable graph. After completing all conditional independence tests, an undirected graph is formed.

[0059] Traverse all triplet structures in the undirected graph. If two variables in a triplet are each connected to a third variable, but there is no connecting edge between these two variables, then orient the two edges to the third variable. Based on the existing directed edges, repeat the following two operations until no more directed edges are added to the variable graph:

[0060] If there exists a directed edge from variable A to variable B, an undirected edge between variable B and variable C, and no connection between variable A and variable C, then the edge between variable B and variable C is directed from variable B to variable C.

[0061] If there exists a directed path from variable A to variable B, and then from variable B to variable C, and there is an undirected edge between variable A and variable C, then the edge between variable A and variable C will be directed from variable A to variable C.

[0062] Output the local causal graph corresponding to the current time segment. Repeat all the above operations to process each time segment in the preprocessed original parameter time series in turn. Merge all local causal graphs to generate a complete causal knowledge graph. The complete causal knowledge graph is a directed acyclic graph. In the complete causal knowledge graph, all exogenous variables have no parent nodes (which meets the root cause hypothesis). Endogenous variables may have zero or more parent nodes. However, if an endogenous variable has no parent node after merging, it is marked as "no significant causal driving was observed" and spurious edges are not forcibly added.

[0063] Step two involves dynamically adjusting the structure of the causal knowledge graph based on Bayesian structure learning. This process, which combines model-independent meta-learning mechanisms, hierarchical reinforcement learning, and multi-objective game theory to generate adaptive decision-making strategies, includes:

[0064] Obtain the geological state characterization at the current time step, the observed values ​​of construction control parameters at the current time step, and the observation data from the most recent fixed number of time steps to form the current window data. Based on the current causal graph, attempt structural perturbation: attempt to add a directed edge; if this action causes a loop, abandon the addition; if adding an edge cannot improve the causal graph, attempt to delete an existing directed edge; if the graph structure does not change after deletion, abandon the deletion; if deletion is also not feasible, attempt to reverse the direction of an existing directed edge; if reversal causes a loop, abandon the reversal. For the perturbed causal graph structure, calculate its matching score with the current window data, mathematically expressed as:

[0065] ,in The causal graph structure at time t is defined as the set of directed edges between variables. This represents all observation data from the start of tunneling to time t. This represents the observation data from the most recent W time steps. This indicates that, given a cause-effect graph structure Below, the likelihood values ​​of the window data. Represents the structural transition probability, when and The value is 1 if the difference between the sets of edges is no more than one, and 0 otherwise. This represents the matching score between the perturbed causal graph and the current window data under all historical data conditions. The updated causal graph is used as the input for the next time step. If the score of the perturbed causal graph is higher than that of the current causal graph, it is used as the new causal graph. Otherwise, the current causal graph is retained, and the updated causal graph is used as the input for the next time step.

[0066] Based on the current geological task identifier, a policy evaluation context is constructed using the variable dependencies depicted by the latest causal graph. The loss value of the current policy under this task is calculated. Based on this loss value, a gradient update is performed on the initial parameters of the policy to obtain the optimized adaptive parameters. The mathematical expression of the policy optimization objective is:

[0067] ,in This represents the initial parameters of the strategy, and N represents the number of geological tasks used during the training phase. This represents the i-th geological task. Indicates by parameters Defined strategy, Representing the strategy in the task The loss value on, This represents the learning rate of the inner loop. This represents the gradient of the loss function with respect to the policy parameters. This indicates the updated parameters. This indicates the updated strategy. Indicates the task Expectation calculation;

[0068] The geological state characterization is input into the high-level strategy network, which outputs the control mode identifier. The control mode identifier and the current state are input into the low-level strategy network, which outputs the original control actions. Four target values ​​are calculated: tunneling distance per unit time (the larger the better), equipment mechanical and electrical load (the smaller the better), energy consumption per unit advance (the smaller the better), and shield attitude deviation (the smaller the better). Among all candidate strategies, the strategy whose target vector is located at the Pareto front and is ranked first according to the preset priority (such as safety > efficiency > energy consumption) is selected as the final strategy. Six instructions are extracted from the final strategy: total thrust setting value, cutterhead torque setting value, thrust speed setting value, motor current upper limit value, cutterhead speed setting value, and screw conveyor speed setting value. The specific outputs are as follows: the total thrust setting value is sent to the propulsion system actuator, the cutterhead torque setting value is sent to the cutterhead drive actuator, the thrust speed setting value is sent to the propulsion system actuator, the motor current upper limit value is sent to the electrical control system actuator, the cutterhead speed setting value is sent to the cutterhead drive actuator, and the screw conveyor speed setting value is sent to the soil removal system actuator.

[0069] Step two constructs a dynamic causal knowledge graph based on geological conditions and integrates Bayesian structural learning, model-independent meta-learning, hierarchical reinforcement learning, and multi-objective game theory to achieve interpretable modeling and adaptive decision-making for the coupling relationship between "geology-equipment-control" during tunnel boring machine (TBM) excavation. This enables the TBM to capture causal evolution caused by geological changes online, ensuring that the model is synchronized with the working conditions. It also enables the TBM to quickly transfer strategies, generate control actions in layers, and find Pareto optimal solutions among multiple objectives such as safety, efficiency, energy consumption, and attitude. Finally, it outputs precise control commands that meet the engineering priorities, significantly improving the system's autonomy, robustness, and construction safety in complex geological formations.

[0070] Step 3: Integrate the main shaft attitude, stratum curvature and circumferential support stress field to construct a coupled control model. Adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching.

[0071] Step three, which integrates the principal axis attitude, formation curvature, and circumferential support stress field to construct a coupled control model, includes:

[0072] Acquire shield machine attitude sensor data, extract the main shaft attitude angle from the attitude sensor data, acquire circumferential stress sensor data within the tunnel segments, extract the circumferential support stress components associated with the main shaft attitude from the stress sensor data, acquire the vertical displacement sequence and corresponding axial coordinate sequence of the tunneling trajectory, calculate the second derivative of the vertical displacement with respect to the axial coordinates, output the ground curvature, acquire the current control input, and send the main shaft attitude angle, circumferential support stress components, ground curvature, and current control input to the coupled evolution calculation unit. In the coupled evolution calculation unit, calculate the state at the next moment, mathematically expressed as:

[0073] ,in The principal axis attitude angle at the current moment. This represents the circumferential support stress component associated with the current spindle attitude. The current geological curvature. For the control input at the current moment, This is the predicted value of the principal axis attitude angle at the next moment. The predicted values ​​of the circumferential support stress components for the next moment. This represents a nonlinear evolution mapping composed of shield dynamics, soil-machine interaction, and support structure response.

[0074] Output the next moment prediction value of the spindle attitude angle and the next moment prediction value of the circumferential support stress components;

[0075] Step three employs a two-layer strategy-execution architecture, combining fuzzy prediction, adaptive loop closure, and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching. The process includes:

[0076] The system obtains the spindle attitude angle and the stratum curvature. Based on the spindle attitude angle and the stratum curvature, it calculates the initial tunneling trajectory, obtains the circumferential support stress, the advance speed, and the cutterhead torque. Based on the circumferential support stress, the advance speed, and the cutterhead torque, it generates operational safety constraints, obtains the soil chamber pressure, and obtains the spatial distribution of stratum strength. Based on the spatial distribution of stratum strength, it queries a predefined fuzzy rule table, matches the control action corresponding to the current stratum strength from the fuzzy rule table, and outputs the matched control action as an auxiliary instruction.

[0077] Obtain the immediate reward, obtain the current state vector, obtain the current control action, and invoke the reinforcement learning update process. During the reinforcement learning update process, calculate the state-action value function correction, with the update rule as follows:

[0078] ,in This is a state vector, with elements including principal axis attitude, formation curvature, circumferential support stress, propulsion speed, and cutterhead torque. To control the action, the value is taken from an element in a predefined set of actions. For immediate rewards, the scores are calculated by weighting the tunneling efficiency score, safety score, and energy consumption score according to their respective weighting coefficients. For learning rate, Using the discount factor as a basis, the action output by reinforcement learning is combined with the directional guidance of the initial tunneling trajectory, the boundary constraints of operational safety constraints, and the auxiliary instruction correction provided by fuzzy rules to generate the final reference control sequence.

[0079] The system acquires the actual feedback values ​​of each actuator, calculates the deviation between the reference control sequence and the actual feedback values, calculates the proportional component, integral component, and differential component of the deviation, and adjusts the proportional gain, integral gain, and differential gain online based on the current formation strength and the main shaft attitude deviation and their changing trends. The proportional component, integral component, and differential component are multiplied by their respective gains and then superimposed to generate the final drive signal. The final drive signal is then output to the propulsion system, cutterhead drive system, and grouting system.

[0080] Obtain the current circumferential support stress value, determine whether the current circumferential support stress value exceeds the safety threshold, if the current circumferential support stress value exceeds the safety threshold, switch to the next level of support mode in the preset support mode sequence, if the current support mode is already the highest level, maintain the current mode and output an alarm signal;

[0081] Step 3 involves constructing a coupled control model by integrating the main shaft attitude, stratum curvature, and circumferential support stress field. A two-tiered strategy-execution architecture is employed to achieve high-precision attitude trajectory planning and dynamic support coordination. On one hand, based on nonlinear evolution mapping, the shield attitude and support response are predicted synchronously, enhancing the modeling capability of soil-machine-structure coupling. On the other hand, fuzzy rules provide prior knowledge guidance, reinforcement learning enables online strategy optimization, adaptive PID closed-loop precise tracking, and multi-mode support switching is triggered based on real-time stress state, ensuring structural safety while balancing tunneling efficiency and energy consumption. This design combines physical mechanism support with intelligent decision-making capabilities, significantly improving the controllability of the shield trajectory, the timeliness of support, and the overall robustness of the system in complex curvature strata.

[0082] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that:

[0083] like Figure 1 As shown, step four: Based on the fuzzy PID control structure, the feed speed, cutter head speed and screw conveyor speed are adjusted in real time in a closed loop, responding to the multivariable control commands output by the decision strategy;

[0084] Step four involves real-time closed-loop regulation of the feed speed, cutterhead speed, and screw conveyor speed based on a fuzzy PID control structure, responding to the multivariable control commands output by the decision strategy. This process includes:

[0085] The propulsion speed setpoint and actual value are obtained, the current deviation of the propulsion speed is calculated, the propulsion speed deviation of the previous cycle is obtained, the change of propulsion speed deviation is calculated, fuzzy PID parameter self-tuning is performed based on the current deviation and the change of deviation, and the propulsion speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change of deviation.

[0086] Obtain the setpoint and actual value of the cutter head speed, calculate the current deviation of the cutter head speed, obtain the deviation of the cutter head speed in the previous cycle, calculate the change in the deviation of the cutter head speed, perform fuzzy PID parameter self-tuning based on the current deviation and the change in deviation, and combine the self-tuned parameters, current deviation, deviation integral and deviation change to calculate the cutter head speed control output and output the cutter head speed control command.

[0087] The setpoint and actual value of the screw conveyor speed are obtained, the current deviation of the screw conveyor speed is calculated, the deviation of the screw conveyor speed in the previous cycle is obtained, the change of the screw conveyor speed deviation is calculated, the fuzzy PID parameter self-tuning is performed based on the current deviation and the change of deviation, and the screw conveyor speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change of deviation, and the screw conveyor speed control command is output.

[0088] Step four employs a fuzzy PID control structure to perform real-time closed-loop adjustment of the propulsion speed, cutterhead speed, and screw conveyor speed: the PID parameters are dynamically tuned by the deviation and its rate of change to adaptively cope with nonlinear, time-varying loads and strong coupling interference during the tunneling process; while retaining the practicality of traditional PID engineering, it integrates the adaptability of fuzzy logic to complex working conditions, significantly improving control accuracy, response speed and stability, and ensuring high-fidelity execution of upper-level multivariable decision commands.

[0089] Step 5: Based on continuous alignment of original parameters, control commands and execution deviations, monitor the system operation status in real time, identify data anomalies, command conflicts or tracking instability, and trigger fault tolerance mechanisms (including correcting control commands, switching support modes or implementing tunneling parameter degradation) through multi-source redundant information.

[0090] Step five involves continuously aligning the original parameters, control commands, and execution deviations, monitoring the system's operating status in real time, identifying data anomalies, command conflicts, or tracking instability, and triggering a fault-tolerance mechanism through multi-source redundant information. This process includes:

[0091] Obtain the status observation value of the main channel, determine whether the status observation value of the main channel is valid, if it is invalid, obtain the status observation value of the redundant channel, determine whether the status observation value of the redundant channel is valid, if the redundancy is valid, adopt the redundant value, otherwise adopt the valid status of the previous cycle.

[0092] Acquire three independent state perception outputs, determine whether there is at least two outputs with a difference less than the preset tolerance. If so, take the average value as the valid state; otherwise, determine that the state perception has failed. Further verify whether the valid state is time-synchronized and within a reasonable range. If the verification fails, output a zero-value control command and trigger a fault report.

[0093] The system acquires decision control commands and monitoring control commands, calculates the difference between them, and if the difference exceeds the allowable range, it is determined to be a command conflict. The execution of the current command is suspended, and a preset safe and conservative command is adopted instead, while triggering a conflict alarm.

[0094] Obtain the current execution deviation and historical execution deviation sequence, determine whether the deviation shows a continuous upward trend (such as monotonically increasing for multiple consecutive periods or the fitting slope is positive), if the trend duration reaches the upper limit, then determine that the tracking is unstable. If instability is determined, load the backup control parameters and start the model calibration process.

[0095] If the identified data is abnormal, the redundant channel is activated or the previous valid state is maintained, and the range of control command changes is limited. If the identified commands conflict, the system switches to a safe and conservative control mode and suspends high-risk actions. If the identification tracking becomes unstable, the tunneling parameters are downgraded, and an emergency shutdown is triggered if necessary.

[0096] Step five establishes a real-time status monitoring and proactive fault-tolerance mechanism based on multi-source redundant information by continuously aligning the original parameters, control commands, and execution deviations. This mechanism can not only promptly identify risks such as data anomalies, command conflicts, or tracking instability, but also dynamically enable redundant data, switch to a safe and conservative control mode, downgrade tunneling parameters, or trigger emergency shutdowns based on consistency verification, trend analysis, and channel effectiveness assessment. This mechanism significantly improves the system's reliability, safety, and anti-interference capabilities, ensuring stable shield tunneling operation even under complex conditions such as sensor failure, communication delays, or control mismatches, providing crucial fault resilience for the intelligent tunneling system.

[0097] Example 3: The technical solution of this embodiment of the invention differs from that of Example 1 and Example 2 in that:

[0098] like Figure 2 As shown, this automated adjustment system for the TBM tunneling process includes:

[0099] TBM Multi-Source Sensing Processing Module: Used to receive raw parameters during the shield tunneling process, preprocess the raw parameters during the shield tunneling process, and output the preprocessed raw parameter time series;

[0100] TBM Adaptive Parameter Tuning Module: Based on the preprocessed original parameter time series, a causal knowledge graph with geological state representation as the root variable is constructed. The structure of the causal knowledge graph is dynamically adjusted based on Bayesian structure learning. Adaptive decision-making strategies are generated by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory.

[0101] TBM attitude support coordination module: used to integrate the main shaft attitude, stratum curvature and circumferential support stress field, construct a coupled control model, adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching.

[0102] TBM parameter execution module: Based on the fuzzy PID control structure, it performs real-time closed-loop adjustment of the feed speed, cutter head speed and screw conveyor speed, and responds to the multivariable control commands output by the decision strategy;

[0103] TBM Safety Fault Tolerance Module: Based on continuous alignment of original parameters, control commands and execution deviations, it monitors the system operating status in real time, identifies data anomalies, command conflicts or tracking instability, and triggers fault tolerance mechanisms through multi-source redundant information;

[0104] This TBM (Tunnel Boring Machine) automated adjustment system constructs an integrated intelligent closed loop encompassing perception, decision-making, execution, and safety through the coordinated operation of five modules: multi-source perception provides high-quality data; adaptive parameter tuning integrates causal graphs, Bayesian learning, and meta-reinforcement learning to generate geological driving strategies that balance safety, efficiency, and energy consumption; attitude support coordination enables precise trajectory planning and dynamic support linkage; the fuzzy PID execution module ensures high-fidelity tracking of key parameters; and the safety fault-tolerant module identifies anomalies in real time based on multi-source redundancy and triggers command correction, mode switching, or parameter degradation. The system possesses interpretability, robustness, and fault resilience under complex geological conditions, significantly improving tunneling safety and intelligence.

[0105] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. An automated adjustment method for the TBM (Tunnel Boring Machine) shield tunneling process, characterized in that, Includes the following steps: Step 1: Receive the raw parameters during the shield tunneling process, preprocess the raw parameters during the shield tunneling process, and output the time series of the preprocessed raw parameters; Step 2: Based on the preprocessed original parameter time series, construct a causal knowledge graph with geological state representation as the root variable. Dynamically adjust the structure of the causal knowledge graph based on Bayesian structure learning, and generate adaptive decision-making strategies by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory. The process of constructing a causal knowledge graph with geological state characteristics as the root variable based on the preprocessed original parameter time series in step two includes: Obtain a preprocessed raw parameter time series, calculate the corresponding geological state variable time series, label the stratum strength, stratum stability, tunneling resistance level, and geological hardness index as exogenous variables, and label the soil chamber pressure, total thrust, cutterhead torque, shield pitch angle, thrust speed, cutterhead motor current, and screw conveyor discharge flow rate as endogenous variables, and merge the time series of exogenous and endogenous variables into the observation dataset of the current segment; For any two variables in the observed dataset, traverse all possible subsets of conditional variables and perform conditional independence tests. If any two variables are determined to be conditionally independent under a certain subset of conditional variables, remove the connection edge between the two variables in the variable graph. After completing all conditional independence tests, an undirected graph is formed. Traverse all triplet structures in the undirected graph. If two variables in a triplet are connected to a third variable but there is no connecting edge between the two variables, then orient the two edges to the third variable. Repeat the following two operations based on the existing directed edges until no more directed edges are added to the variable graph. Output the local causal graph corresponding to the current time segment. Repeat all the above operations to process each time segment in the preprocessed original parameter time series in turn. Merge all local causal graphs to generate a complete causal knowledge graph. Step 3: Integrate the main shaft attitude, stratum curvature and circumferential support stress field to construct a coupled control model. Adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching. The process of integrating the principal axis attitude, formation curvature, and circumferential support stress field to construct the coupled control model in step three includes: Acquire shield machine attitude sensor data, extract the main shaft attitude angle from the attitude sensor data, acquire circumferential stress sensor data inside the tunnel segment, extract the circumferential support stress component associated with the main shaft attitude from the stress sensor data, acquire the vertical displacement sequence and corresponding axial coordinate sequence of the tunneling trajectory, calculate the second derivative of the vertical displacement with respect to the axial coordinate, output the stratum curvature, acquire the current control input, and send the main shaft attitude angle, circumferential support stress component, stratum curvature and the current control input to the coupled evolution calculation unit. In the coupled evolution calculation unit, calculate the state at the next moment, output the predicted value of the main shaft attitude angle at the next moment, and output the predicted value of the circumferential support stress component at the next moment. Step 4: Based on the fuzzy PID control structure, the feed speed, cutter head speed and screw conveyor speed are adjusted in real time in a closed loop, responding to the multivariable control commands output by the decision strategy; Step 5: Based on continuous alignment of original parameters, control commands and execution deviations, monitor the system operation status in real time, identify data anomalies, command conflicts or tracking instability, and trigger fault tolerance mechanisms through multi-source redundant information.

2. The automated adjustment method for the TBM (tunnel boring machine) process according to claim 1, characterized in that, Step two, which involves dynamically adjusting the structure of the causal knowledge graph based on Bayesian structural learning and combining model-independent meta-learning mechanisms, hierarchical reinforcement learning, and multi-objective game theory to generate adaptive decision-making strategies, includes the following: Obtain the geological state characterization of the current time step, obtain the construction control parameter observation values ​​of the current time step, obtain the observation data of the most recent fixed number of time steps, form the current window data, and attempt to perturb the structure based on the current causal graph. For the perturbed causal graph structure, calculate its matching score with the current window data. If the score of the perturbed causal graph is higher than that of the current causal graph, then use it as the new causal graph; otherwise, retain the current causal graph. Load the initial policy parameters that match the current causal graph from the policy library. Based on the current geological task identifier, construct the policy evaluation context using the variable dependencies depicted by the latest causal graph. Calculate the loss value of the current policy under this task. Perform a gradient update on the initial policy parameters based on this loss value to obtain the optimized adaptation parameters. The geological state characterization is input into the high-level strategy network, which outputs the control mode identifier. The control mode identifier and the current state are input into the low-level strategy network, which outputs the original control action. Four target values ​​are calculated: tunneling distance per unit time, equipment mechanical and electrical load, energy consumption per unit advance, and shield attitude deviation. Among all candidate strategies, the strategy that meets the following conditions is selected as the final strategy. Six instructions are extracted from the final strategy: total thrust setting value, cutterhead torque setting value, thrust speed setting value, motor current upper limit value, cutterhead speed setting value, and screw conveyor speed setting value.

3. The automated adjustment method for TBM tunneling process according to claim 1, characterized in that, Step three employs a two-layer strategy-execution architecture, combining fuzzy prediction, adaptive loop closure, and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching. This process includes: The system obtains the spindle attitude angle and the stratum curvature. Based on the spindle attitude angle and the stratum curvature, it calculates the initial tunneling trajectory, obtains the circumferential support stress, the advance speed, and the cutterhead torque. Based on the circumferential support stress, the advance speed, and the cutterhead torque, it generates operational safety constraints, obtains the soil chamber pressure, and obtains the spatial distribution of stratum strength. Based on the spatial distribution of stratum strength, it queries a predefined fuzzy rule table, matches the control action corresponding to the current stratum strength from the fuzzy rule table, and outputs the matched control action as an auxiliary instruction. Obtain the immediate reward, obtain the current state vector, obtain the current control action, invoke the reinforcement learning update process, calculate the state-action value function correction during the reinforcement learning update process, and generate the final reference control sequence based on the action output by reinforcement learning, combined with the direction guidance of the initial tunneling trajectory, the boundary constraints of the running safety constraints, and the auxiliary instruction correction provided by fuzzy rules. The system acquires the actual feedback values ​​of each actuator, calculates the deviation between the reference control sequence and the actual feedback values, calculates the proportional component, integral component, and differential component of the deviation, and adjusts the proportional gain, integral gain, and differential gain online based on the current formation strength and the main shaft attitude deviation and their changing trends. The proportional component, integral component, and differential component are multiplied by their respective gains and then superimposed to generate the final drive signal. The final drive signal is then output to the propulsion system, cutterhead drive system, and grouting system. Obtain the current circumferential support stress value, determine whether the current circumferential support stress value exceeds the safety threshold. If the current circumferential support stress value exceeds the safety threshold, switch to the next level of support mode in the preset support mode sequence. If the current support mode is already the highest level, maintain the current mode and output an alarm signal.

4. The automated adjustment method for the TBM (tunnel boring machine) process according to claim 1, characterized in that, Step four, based on a fuzzy PID control structure, involves real-time closed-loop adjustment of the propulsion speed, cutterhead speed, and screw conveyor speed, responding to the multivariable control commands output by the decision strategy. This process includes: The propulsion speed setpoint and actual value are obtained, the current deviation of the propulsion speed is calculated, the propulsion speed deviation of the previous cycle is obtained, the change of propulsion speed deviation is calculated, fuzzy PID parameter self-tuning is performed based on the current deviation and the change of deviation, and the propulsion speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change of deviation. Obtain the setpoint and actual value of the cutter head speed, calculate the current deviation of the cutter head speed, obtain the deviation of the cutter head speed in the previous cycle, calculate the change in the deviation of the cutter head speed, perform fuzzy PID parameter self-tuning based on the current deviation and the change in deviation, and combine the self-tuned parameters, current deviation, deviation integral and deviation change to calculate the cutter head speed control output and output the cutter head speed control command. The setpoint and actual values ​​of the screw conveyor speed are obtained, the current deviation of the screw conveyor speed is calculated, the deviation of the screw conveyor speed in the previous cycle is obtained, the change in the screw conveyor speed deviation is calculated, fuzzy PID parameter self-tuning is performed based on the current deviation and the change in deviation, and the screw conveyor speed control output is calculated by combining the self-tuned parameters, the current deviation, the deviation integral and the change in deviation, and the screw conveyor speed control command is output.

5. The automated adjustment method for TBM tunneling process according to claim 1, characterized in that, Step five, based on continuous alignment of original parameters, control commands, and execution deviations, monitors the system's operating status in real time, identifies data anomalies, command conflicts, or tracking instability, and triggers a fault-tolerant mechanism through multi-source redundant information, includes the following: Obtain the status observation value of the main channel, determine whether the status observation value of the main channel is valid, if it is invalid, obtain the status observation value of the redundant channel, determine whether the status observation value of the redundant channel is valid, if the redundancy is valid, adopt the redundant value, otherwise adopt the valid status of the previous cycle. Acquire three independent state perception outputs, determine whether there is at least two outputs with a difference less than the preset tolerance. If so, take the average value as the valid state; otherwise, determine that the state perception has failed. Further verify whether the valid state is time-synchronized and within a reasonable range. If the verification fails, output a zero-value control command and trigger a fault report. The system acquires decision control commands and monitoring control commands, calculates the difference between them, and if the difference exceeds the allowable range, it is determined to be a command conflict. The execution of the current command is suspended, and a preset safe and conservative command is adopted instead, while triggering a conflict alarm. Obtain the current execution deviation and historical execution deviation sequence, determine whether the deviation shows a continuous upward trend, if the trend duration reaches the upper limit, determine that the tracking is unstable, if the instability is determined, load the backup control parameters, and start the model calibration process; If the identified data is abnormal, the redundant channel is activated or the previous valid state is maintained, and the range of control command changes is limited. If the identified commands conflict, the system switches to a safe and conservative control mode and suspends high-risk actions. If the identification tracking becomes unstable, the tunneling parameters are downgraded.

6. An automated adjustment system for the TBM (Tunnel Boring Machine) process, characterized in that, An automated adjustment method for a TBM (tunnel boring machine) process as described in claim 1, comprising: TBM Multi-Source Sensing Processing Module: Used to receive raw parameters during the shield tunneling process, preprocess the raw parameters during the shield tunneling process, and output the preprocessed raw parameter time series; TBM Adaptive Parameter Tuning Module: Based on the preprocessed original parameter time series, a causal knowledge graph with geological state representation as the root variable is constructed. The structure of the causal knowledge graph is dynamically adjusted based on Bayesian structure learning. Adaptive decision-making strategies are generated by combining model-independent meta-learning mechanism, hierarchical reinforcement learning and multi-objective game theory. TBM attitude support coordination module: used to integrate the main shaft attitude, stratum curvature and circumferential support stress field, construct a coupled control model, adopt a strategy-execution two-layer architecture, and combine fuzzy prediction, adaptive closed loop and reinforcement learning to perform attitude trajectory planning and execute multi-mode support switching. TBM parameter execution module: Based on the fuzzy PID control structure, it performs real-time closed-loop adjustment of the feed speed, cutter head speed and screw conveyor speed, and responds to the multivariable control commands output by the decision strategy; TBM safety fault tolerance module: Based on continuous alignment of original parameters, control commands and execution deviations, it monitors the system operating status in real time, identifies data anomalies, command conflicts or tracking instability, and triggers fault tolerance mechanisms through multi-source redundant information.

Images