Autonomous tunneling method and system for ultra-small turning autonomous tunneling TBM (Tunnel Boring Machine)

Abstract

The invention discloses an ultra-small turning autonomous tunneling TBM (Tunnel Boring Machine) autonomous tunneling method and system. The method comprises the following steps: firstly, collecting multi-sensing data in real time and constructing a state vector; monitoring a safety index and judging whether the safety index exceeds a limit or not, and if yes, switching to a safety control mode; selecting a corresponding basic control strategy based on the geological recognition result, and providing boundary constraints for the control instruction; an Actor-Critic framework is adopted to output the fine adjustment quantity of the control parameters, and the action income is evaluated; dynamically adjusting the maximum allowable speed according to the real-time vibration level; sending the final speed instruction to each thrust cylinder controller, collecting performance data after execution, and calculating a multi-target reward; and state transition experience is stored, batch training is carried out, network parameters of the Actor-Critic architecture are updated, and online optimization of a control strategy is realized. The invention provides a set of complete and feasible technical solution for the autonomous tunneling of the ultra-small turning TBM.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction technology, and in particular to an autonomous tunneling TBM with ultra-small turning radius and an autonomous tunneling method and system. Background Technology

[0002] Large-scale energy projects such as mines and hydropower are important pillars of my country's economic development. Hard rock tunnels are considered the "lifeline" for material transportation, gas drainage, and personnel evacuation in energy extraction. Compared to traditional tunnels, energy tunnels are characterized by multiple interconnected sections and sharp curves, requiring TBMs to be more flexible and capable of rapid relocation. Traditional open-face, shield, and tunnel boring machine (TBM) types are not suitable for these applications. Currently, most construction uses a new type of TBM with parallel "V"-shaped propulsion. Although this type of machine has ultra-small turning capabilities, its attitude control is difficult to operate manually, and the training period for personnel is long. There is an urgent need to develop an autonomous tunneling system for this type of machine. Current autonomous tunneling methods are mainly designed for traditional TBMs or shield tunnels. For example, patent CN202410496837.3 uses a bidirectional long short-term memory recurrent neural network model to predict new tunneling parameters and select tunneling strategies based on vibration data, geological data, and current tunneling parameter data. However, this method ignores the physical model of the actual TBM propulsion system configuration, relies entirely on data for decision-making, and has a relatively simple tunneling strategy, making it difficult to adapt to the tunneling needs of TBMs with ultra-small turns. Patent CN202310186685.2 constructs a deep reinforcement learning neural network model in a virtual simulation environment to train the shield machine attitude control model. However, it focuses on parameters related to cylinder thrust and geological settlement, ignoring the trajectory deviation caused by adverse effects such as vibration impact and off-center loading under hard rock tunneling, and does not consider the impact of attitude changes on the model under ultra-small turn conditions. Furthermore, this method also ignores the physical model of the propulsion system. Patent CN202110748572.8 designs a deep reinforcement learning model for tunnel boring machines (TBMs), outputting the total thrust and torque of the cutterhead and outputting control parameters such as cutterhead rotation speed and propulsion speed through a TBM environment interaction module. This method focuses on evaluating the tunneling efficiency of the TBM but does not consider the TBM's attitude control, and it is not applicable to TBM models. Patent CN202511203614.4 constructs a deep learning model considering cutter wear data to predict tunneling parameters such as cutterhead rotation speed, propulsion speed, and cutterhead torque, but does not address TBM attitude control. Patent CN202510852693.5 also constructs a deep reinforcement learning model to predict TBM tunneling parameters, addressing the problem of low adaptability of automatic control models to uncertainties and external disturbances. However, this method also does not directly address attitude control and ultra-small turning situations.

[0003] In summary, autonomous tunneling control of TBMs / shield tunneling machines is a hot research topic in the industry, with data-driven methods being widely applied. In particular, researchers have increasingly employed deep reinforcement learning to address the uncertainties faced by autonomous TBM tunneling in recent years. However, current research primarily focuses on conventional TBMs and shield tunneling machines, neglecting ultra-small turn conditions. While tunneling parameters and trajectories are mostly similar to or gradually change from the previous moment during non-ultra-small turn tunneling, the TBM's tunneling trajectory changes rapidly under ultra-small turn conditions, making autonomous tunneling control methods more complex and necessitating the development of targeted autonomous tunneling methods. Summary of the Invention

[0004] The purpose of this invention is to provide an autonomous tunneling TBM with ultra-small turns and an autonomous tunneling method and system, which solves the technical problems of large variations in tunneling parameters of ultra-small turns TBM and the inapplicability of existing prediction models. It realizes autonomous tunneling of ultra-small turns TBM under all working conditions such as straight lines and ultra-small turns, taking into account tunneling accuracy, equipment safety and tunneling efficiency.

[0005] To achieve the above objectives, this invention provides an autonomous tunneling TBM with ultra-small turning radius and an autonomous tunneling method, comprising the following steps: S1. Real-time acquisition of multi-sensor data and construction of state vectors; S2. Monitor safety indicators and determine if they exceed limits. If limits are exceeded, switch to safety control mode. S3. Select the corresponding basic control strategy based on the geological identification results to provide boundary constraints for control commands; S4. Employ the Actor-Critic architecture to output fine-tuning values ​​for control parameters and evaluate the benefits of the action; S5. The maximum allowable speed is dynamically adjusted according to the real-time vibration level, and the vibration intensity is inversely proportional to the maximum speed; S6. Send the final speed command to each propulsion cylinder controller, collect the performance data after execution, and calculate the multi-objective reward; S7 stores state transition experience and performs batch training, updates the network parameters of the Actor-Critic architecture, and realizes online optimization of the control strategy.

[0006] Preferably, in S1, geological sensor data, attitude sensor data, vibration sensor data, pressure sensor data, and position sensor data are collected in real time. Constructing state vectors =[Geological characteristics, vibration level, pressure state, trajectory deviation, velocity fluctuation]; The geological characteristic states include four states: turning condition, uniform soft rock, uniform hard rock, and heterogeneous soft-hard rock. The state vector of the geological characteristic state is represented as follows: =[1 / 0,1 / 0,1 / 0,1 / 0]; The values ​​indicate whether the rock is a very small turn, uniform soft rock, uniform hard rock, or uneven rock; where 1 represents yes and 0 represents no. The attitude vector at the current moment is represented as ,in, , , Indicates the spatial coordinates of the current time. , , Represents spatial attitude coordinates; the current attitude vector is guided by the laser to the pose. Visual measurement of pose And the kinematic forward pose based on the propulsion system configuration and cylinder extension position Together, we obtain the following expression: = +f( , ); In the formula, This indicates that the visually measured pose and the orthogonal pose are fused and calculated. The trajectory deviation state represents the amount and trend of the deviation between the current pose and the design axis. The expression for the trajectory deviation vector is as follows: ; The first row represents the deviation between the current pose and the previous spatial pose, and the second row represents the rate of change of the deviation. This represents the rate of change of deviation, with a value range of [-1, 1]. A positive value indicates that the deviation is increasing, while a negative value indicates that the deviation is decreasing. The vibration horizontal state vector is represented as: = ; in, Indicates different vibration measurement points, Indicates vibration intensity; ; in, This indicates the number of measurements or samples taken for vibration values ​​at the same side point. express The maximum value, ; indicates the first The vibration measuring point at the first The magnitude of the vibration level measured in this measurement; The pressure state vector represents the pressure value and pressure balance index of each propulsion cylinder. The expression is: ; in, This indicates the pressure in the large chamber of the hydraulic cylinder. This represents the pressure in the small chamber of the propulsion cylinder. The pressure in both chambers is obtained from a pressure sensor, and the pressure state vector is represented as: ; in, Establish pressure balance indicators for each propulsion cylinder; Speed ​​fluctuations include cutterhead rotation speed fluctuations and feed speed fluctuations. Speed ​​balance indicators are established for each feed cylinder and cutterhead. ; in, This indicates the current cutting head rotation speed and propulsion speed at the moment of propulsion. in, The average speed of the cutterhead and each propulsion cylinder during the current tunneling cycle is expressed as follows: ; in, This indicates the number of historical data points traced back to the beginning.

[0007] Preferably, S2 is as follows: Monitor the pressure of each hydraulic cylinder to ensure it does not exceed the safety threshold. ; Establish pressure balance indicators for each propulsion cylinder: ; The average thrust of each cylinder in the current tunneling cycle is expressed as follows: ; The propulsion pressure command must not exceed the pressure balance index threshold. ; Monitor whether the vibration intensity exceeds the safety threshold. ; Does the monitoring trajectory deviation exceed the safety threshold? ; Monitor whether the speed deviation exceeds the safety threshold. : ; in, Indicates the maximum rotational speed of the cutter head. Indicates the propulsion speed of each hydraulic cylinder; If any safety indicator exceeds the limit, the system will switch to safety control mode, in which low-speed tunneling or manual intervention will be used. The over-limit alarm uses a threshold over-limit alarm, or determines whether to switch to safety control mode based on the number of consecutive alarms or the percentage of over-limit alarms within a set time. Alarm thresholds can be set based on experience or determined online using data distribution analysis.

[0008] Preferably, in S3, the geological identification results are obtained through rock debris image analysis and tool detection methods: For uniform soft rock conditions, a high-speed strategy is adopted; For uniform, weak, and hard rock conditions, a vibration stabilization strategy is adopted; For uneven geological conditions, an equilibrium strategy is adopted; For turning conditions, a precise strategy is adopted; While setting the reference parameters for each operating condition, add attitude deviation constraints. |≤0.2.

[0009] Preferably, in S4, the Actor-Critic architecture includes an Actor Network and a Critic Network: An actor network outputs an action based on the current state. This action follows a Gaussian distribution, including the action mean. and action variance Mean of movement This represents the optimal action, and the action variance. This indicates the network's uncertainty about its current state; The output layer uses the tanh activation function to constrain the mean to the range [-1, 1], and the variance output is ensured to be positive through exp calculation. The final tunneling control parameters are: ; A critic network is used to evaluate the expected total reward of taking an action in a given state. The multi-objective reward function is designed as follows: ; in, , , and Here, are the weighting coefficients; The trajectory accuracy bonus is linearly negatively correlated with the absolute value of the deviation; A piecewise function is used as the reward for vibration control; The incentive for tunneling efficiency is positively correlated with average advance speed and negatively correlated with speed fluctuation. A piecewise function is used for stress safety rewards.

[0010] Preferably, in S5, the maximum allowable speed is dynamically adjusted based on the real-time vibration level, and the vibration intensity is inversely proportional to the maximum speed, as shown in the following formula: ; in, This represents the current pace of progress; The propulsion speed is dynamically adjusted. The velocity scaling factor is based on the real-time vibration intensity. The dynamic speed scaling factor is calculated using the following formula: = / max(P, ); in, This represents the baseline value for vibration intensity, which is a constant.

[0011] Preferably, S7 is as follows: The state transition experience is stored in the replay buffer. After executing the speed command generated by S5, a new round of sensor data is collected to obtain the new state s_{t+1}, and the multi-target reward is calculated. , It is the core indicator for measuring the quality of this action, and is derived by weighted summation of four sub-rewards: trajectory, vibration, efficiency, and safety. Batch training is performed periodically by sampling from the buffer, specifically as follows: fragments of experience The experience segments are stored in the experience replay buffer. A batch of experience segments are randomly selected from the experience replay buffer periodically for training. First, the value prediction accuracy of the critic network is trained. Then, the parameters of the actor network are adjusted according to the value gradient to achieve online optimization of the control strategy.

[0012] The present invention also provides an ultra-small turning autonomous tunneling TBM and an autonomous tunneling system, including a perception layer, a data acquisition layer, a decision-making layer and an execution layer; The sensing layer includes several vibration sensors, a binocular vision measurement system, a visual target, a pressure sensor, a displacement sensor, a laser and vision sensor, and a laser guidance system. The laser guidance system includes a total station and a laser target. The acquisition layer is a distributed data acquisition terminal used to receive data acquisition instructions and synchronously acquire multi-source sensor data from the perception layer according to rules; The decision-making layer is a high-performance edge controller that completes data preprocessing, tunneling strategy generation, autonomous tunneling control parameter calculation, and feeds back to the execution layer. The execution layer includes the TBM main control PLC, the execution system motor controller, and the cylinder / motor control valve, which are used to receive autonomous tunneling control parameters and execute tunneling actions.

[0013] Preferably, the vibration sensor is installed at the TBM cutterhead, main drive, front shield, support shield, and support shoe positions; A binocular vision measurement system and a vision target are installed at both ends of the propulsion cylinder, and pressure or strain sensors are installed on the radial and axial mounting surfaces of the main bearing. A triaxial force sensor is mounted on the tool mounting surface; a pressure sensor is mounted in the large and small oil chambers of the feed cylinder; a displacement sensor is mounted on each feed cylinder. The pressure sensor for the support shoe cylinder is installed on the TBM support shoe cylinder.

[0014] Therefore, the present invention employs the above-mentioned method and system for autonomous tunneling using a TBM with ultra-small turning radius, and the beneficial effects are as follows: (1) This invention applies deep reinforcement learning to TBM multi-objective collaborative control and adopts the Actor-Critic architecture to design the control model, which solves the technical problem that traditional methods cannot simultaneously optimize multiple conflicting objectives such as trajectory accuracy, equipment safety, and tunneling efficiency.

[0015] (2) The present invention designs a four-level hardware architecture of perception, acquisition, decision-making and execution, and with a customized sensor device layout scheme, realizes full-dimensional and high-precision data acquisition for TBM tunneling, providing a data foundation for autonomous tunneling.

[0016] (3) The present invention proposes a geological adaptive strategy switching mechanism, which can automatically select the optimal basic control strategy according to the geological identification results, and adapt to all working conditions such as uniform soft rock, uniform hard rock, soft and hard unevenness, and ultra-small turns.

[0017] (4) The present invention establishes the concept of pressure difference tolerance control, which breaks through the limitations of traditional pressure balance control. Through pressure state monitoring and threshold constraints, it is more in line with the actual construction needs of ultra-small turning TBM.

[0018] (5) The present invention designs a vibration-driven speed limiting mechanism, which is different from the traditional hard fuse safety control. It adopts a soft throttling method to realize a negative feedback closed loop of vibration increase-automatic speed reduction-vibration relief, so as to maximize the tunneling efficiency while ensuring equipment safety.

[0019] (6) This invention realizes the autonomous tunneling of the entire process of ultra-small turning TBM, reduces the intervention of manual operation, reduces the difficulty of attitude control operation, shortens the personnel training cycle, and improves the construction efficiency of energy tunnel hard rock tunnel.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1This is an overall flowchart of an embodiment of the autonomous tunneling TBM with ultra-small turning radius and autonomous tunneling method of the present invention; Figure 2 This is a sensor layout diagram of an embodiment of an ultra-small turning autonomous tunneling TBM and autonomous tunneling system according to the present invention.

[0022] Figure Labels 1. TBM cutterhead; 2. Front shield; 3. Displacement sensor; 4. Visual target; 5. Vibration sensor; 6. Propulsion cylinder; 7. Support shield; 8. Laser target; 9. Binocular vision measurement system; 10. Support shoe; 11. Total station. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0025] like Figure 1 As shown, an autonomous tunneling TBM with ultra-small turning radius and an autonomous tunneling method include the following steps: S1. Multi-source data acquisition and state construction: Real-time acquisition of data from geological sensors, attitude sensors, vibration sensors, pressure sensors, and position sensors; construction of state vectors. =[Geological characteristics, vibration level, pressure state, trajectory deviation, velocity fluctuation].

[0026] 1) Among them, the geological characteristic states include four states: turning condition, uniform soft rock, uniform hard rock, and uneven hardness. The state vector of the geological characteristic state is represented as: =[1 / 0,1 / 0,1 / 0,1 / 0]; The numbers sequentially indicate whether the rock type is a very small turn, uniform soft rock, uniform hard rock, or uneven rock hardness, where 1 represents yes and 0 represents no. For example, when the tunnel is a straight excavation and the rock is uniform soft rock... =[0,1,0,0].

[0027] 2) The attitude vector at the current moment is represented as: ,in, , , Indicates the spatial coordinates of the current time. , , Represents spatial attitude coordinates; the attitude vector at the previous time step is guided by the laser to the pose. Visual measurement of pose And the kinematic forward pose based on the propulsion system configuration and cylinder extension position Together, we obtain the following expression: = +f( , ); In the formula, This refers to the fusion calculation of visually measured pose and forward pose. Generally, Kalman / extended Kalman methods can be used for fusion to eliminate measurement errors and improve the stability and accuracy of pose acquisition. Alternatively, the mean calculation method can be used for fusion.

[0028] The trajectory deviation state represents the amount and trend of the deviation between the current pose and the design axis, such as a decrease or increase in deviation. The deviation trend is determined by fitting the current pose and historical poses using models such as support vector machines, multinomial fitting, or deep learning to predict the deviation value at future moments. This prediction is then compared with the corresponding design axis at future moments to obtain the rate of change of the deviation. The expression for the trajectory deviation vector is as follows: ; The first row represents the deviation between the current pose and the previous spatial pose, and the second row represents the rate of change of the deviation. This represents the rate of change of deviation, with a value range of [-1, 1]. A positive value indicates that the deviation is increasing, while a negative value indicates that the deviation is decreasing.

[0029] 3) The vibration horizontal state vector is represented as: = ; in, Indicates different vibration measurement points, The vibration intensity is represented by the average effective value of vibration or the average peak value of vibration.

[0030] ; in, This indicates the number of measurements or samples taken for vibration values ​​at the same side point. express The maximum value, ; indicates the first The vibration measuring point at the first The magnitude of the vibration level measured in this measurement.

[0031] 4) The pressure state vector represents the pressure value and pressure balance index of each propulsion cylinder. The pressure value of each propulsion cylinder... The expression is: ; in, This indicates the pressure in the large chamber of the hydraulic cylinder. This represents the pressure in the small chamber of the propulsion cylinder. The pressure in both chambers is obtained from a pressure sensor, and the pressure state vector is represented as: ; in, Establish pressure balance indicators for each propulsion cylinder.

[0032] 5) Speed ​​fluctuations include cutterhead rotation speed fluctuations and feed speed fluctuations. Speed ​​balance indicators should be established for each feed cylinder and cutterhead. ; in, This indicates the current cutting head rotation speed and propulsion speed at the moment of propulsion. The average speed of the cutterhead and each propulsion cylinder during the current tunneling cycle is expressed as follows: ; in, This indicates the number of historical data points traced back to the beginning.

[0033] The above states undergo data denoising, filtering, padding, and normalization before being entered into the model.

[0034] S2. Safety Monitoring and Mode Decision: Monitor safety indicators and determine if limits are exceeded. If limits are exceeded, switch to safety control mode. Specifically: Monitor the pressure of each hydraulic cylinder to ensure it does not exceed the safety threshold. .

[0035] To prevent sudden pressure changes, pressure balance indicators are established for each propulsion cylinder: .

[0036] The average thrust of each cylinder in the current tunneling cycle is expressed as follows: .

[0037] The propulsion pressure command must not exceed the pressure balance index threshold. The cylinder pressure here is the cylinder thrust calculated solely based on the pressure sensor and the cylinder's large and small chambers.

[0038] And monitor whether the vibration intensity exceeds the safety threshold. .

[0039] Does the monitoring trajectory deviation exceed the safety threshold? ;

[0040] Monitor whether the speed deviation exceeds the safety threshold. : ; in, Indicates the maximum rotational speed of the cutter head. This indicates the propulsion speed of each hydraulic cylinder, and the speed fluctuation index should be within a reasonable range.

[0041] If any safety indicator exceeds the limit, the system will switch to safety control mode, in which low-speed tunneling or manual intervention will be used.

[0042] The over-limit alarm of this invention can be a threshold over-limit alarm, or it can determine whether to switch to safety control mode based on the number of consecutive alarms or the proportion of over-limit alarms within a set time. The alarm threshold can be set based on experience or determined online using methods such as data distribution analysis.

[0043] S3. Geological Adaptive Strategy Selection: Based on the geological identification results, a corresponding basic control strategy is selected to ensure that the system has a basic, reasonable, and safe operational foundation before complex calculations begin. The geological identification results are obtained through methods such as rock debris image analysis and tool detection. Once the geological conditions are determined, the geological adaptive strategy provides boundary constraints for the control commands output by reinforcement learning.

[0044] For uniform soft rock conditions, a high-speed strategy is adopted, setting a low cutterhead speed, medium thrust, and high penetration. For example, the allowable parameter range is set as follows: cutterhead speed 5±1 r / min, penetration 10±2 mm / r, and total thrust 1500±150 kN. |≤0.3.

[0045] For uniform, weakly hard rock conditions, a vibration stabilization strategy is adopted, which involves setting a high cutterhead rotation speed, high thrust, and small penetration depth. For example, the cutterhead rotation speed is set to 8±1 r / min, the penetration depth to 5±2 mm / r, and the total thrust to 2000±150 kN. |≤0.3.

[0046] To address uneven geological conditions, a balancing strategy is adopted, setting moderate cutterhead speed, moderate thrust, and moderate penetration. For example, the cutterhead speed is set to 3±1 r / min, the penetration to 7±2 mm / r, and the total thrust to 1500±150 kN. |≤0.3.

[0047] For turning conditions, a precise strategy is adopted; while setting the baseline parameters for each condition, attitude deviation constraints are added. |≤0.2; and prepare to dynamically fine-tune the speed of each cylinder according to the real-time attitude difference based on the fine-tuning function in the subsequent steps to prevent "deviation".

[0048] Note: The specific data above are for illustrative purposes only and do not necessarily mean that these are the only possible parameter settings.

[0049] S4. Deep reinforcement learning intelligent control: It adopts an Actor-Critic architecture to output fine-tuning of control parameters and evaluate the benefits of actions; The Actor-Critic architecture includes an Actor Network and a Critic Network: 1) Actor Network, also known as policy network, outputs actions (i.e., fine-tuning of control parameters such as overall propulsion speed and thrust of each cylinder) based on the current state (geological features, posture, vibration, pressure, etc.).

[0050] In the control logic diagram, the actor network is located in the "RL Smart Fine-tuning" module. The output action is not a single velocity value, but a probability distribution, typically a Gaussian distribution, including the action mean. and action variance .

[0051] Action mean This represents the optimal action as perceived by the network, i.e., the baseline adjustment amount for the speed and pressure of each cylinder. For example, [+0.5,-0.2,+0.1,-0.3] represents a proportional adjustment on the base speed. Action variance This indicates the network's uncertainty about its current state.

[0052] The output layer uses the tanh activation function to constrain the mean to the range [-1, 1], facilitating integration with the baseline speed. The variance output is ensured to be positive through an exp operation; therefore, the final tunneling control parameters are: .

[0053] 2) The Critic Network, also known as the Value Network, is used to evaluate the expected total reward of taking an action in a given state, i.e., the state-action value function. In the control logic diagram, the Critic Network is located in the "Policy Network Update" section of the "Learning System".

[0054] The multi-objective reward function is designed as follows: ; in, , , and As weighting coefficients, in this embodiment, several weights can be configured as follows: =0.4, =0.3, =0.2 and =0.1.

[0055] The trajectory accuracy bonus, used to guide the TBM to tunnel strictly along the design axis, reduces efficiency losses and abnormal wear of the shield tail clearance caused by deviation correction, and its expression is: ; The trajectory accuracy reward is linearly negatively correlated with the absolute value of the deviation. The larger the deviation, the greater the penalty. It is a proportionality coefficient, such as =0.1, used to scale millimeter-level deviations to a suitable reward value scale.

[0056] To provide a vibration control incentive and protect critical components such as the cutterhead main bearing and gearbox, preventing fatigue damage or downtime caused by severe vibration, a piecewise incentive function is adopted: When vibration intensity When ≤2g, =+1.0, meaning a fixed positive bonus is given in the low-vibration zone; when the vibration intensity 2g < When ≤5g, =0, meaning in the normal vibration zone, there is neither reward nor penalty; when the vibration intensity... When >5g, =-1.0, meaning a penalty is applied in the dangerous vibration zone. The vibration intensity division range and penalty value are examples and can be adjusted according to the actual situation.

[0057] To incentivize tunneling efficiency, the goal is to maximize tunneling progress per unit time while ensuring safety and accuracy, thereby improving the project's economic benefits. The tunneling efficiency bonus is positively correlated with average advance speed and negatively correlated with speed fluctuations, as follows: ; in, This represents the current average speed. and It is a proportionality coefficient (e.g.) =0.01, =0.3), used to scale speed values ​​and speed fluctuation values ​​to the reward scale.

[0058] To ensure pressure safety and prevent hydraulic system malfunctions, seal damage, or cylinder scoring due to excessive pressure in individual cylinders, and to prevent TBM attitude loss of control due to sudden pressure changes, a piecewise function is adopted: When under pressure [150< ≤250 bar AND When ≤50kN, =0, meaning no intervention in the normal operating range; when pressure state [ >250 bar OR When >50kN, =-1, which means that a penalty is imposed in high-pressure areas and when pressure fluctuations are large.

[0059] S5. Vibration Adaptive Speed ​​Limiting: The maximum allowable speed is dynamically adjusted based on the real-time vibration level, and the vibration intensity is inversely proportional to the maximum speed, as shown in the following formula: ; in, This represents the current pace of progress; The propulsion speed is dynamically adjusted. This is the speed scaling factor. Unlike the S200's rigid "fuse," this is a soft "throttling." The system adjusts based on real-time vibration intensity. The dynamic speed scaling factor is calculated using the following formula: = / max(P, ); in, It represents a vibration intensity benchmark value that can be set manually or assigned by a program, and it is a constant value; When we take 3, we have: =3.0 / max(P,3.0); The greater the vibration, the smaller the coefficient, and the final speed command will be proportionally reduced, thus realizing a negative feedback closed loop of vibration increase -> automatic speed reduction -> vibration mitigation, preventing problems before they occur.

[0060] S6. Control Execution and Performance Evaluation: Send the final speed command to each propulsion cylinder controller, collect the performance data after execution, and calculate the multi-objective reward.

[0061] S7. Online Learning and Policy Update: Store state transition experience and perform batch training to update the network parameters of the Actor-Critic architecture, thereby achieving online optimization of the control policy. Specifically: The state transition experience is stored in the replay buffer. After executing the speed command generated by S5, a new round of sensor data is collected to obtain the new state s_{t+1}, and the multi-target reward is calculated. , It is the core indicator for measuring the quality of this action, and is derived by weighted summation of four sub-rewards: trajectory, vibration, efficiency, and safety.

[0062] Batch training is performed periodically by sampling from the buffer, specifically as follows: A complete decision-making and feedback process As a fragment of experience, it is stored in the experience playback buffer, which is like a "construction log library" that stores a large number of historical cases.

[0063] Update the parameters of the actor network and the critic network: periodically select a batch of experience segments from the experience replay buffer for training. First train the value prediction accuracy of the critic network to make its value prediction Q of state-action pairs more accurate.

[0064] Then, adjust the parameters of the actor network based on the value gradient provided by critics, so that it can propose higher rewards in the future. To achieve better action and optimize the control strategy online.

[0065] like Figure 2 As shown, the present invention also provides an ultra-small turning autonomous tunneling TBM and an autonomous tunneling system for implementing the above method, including a perception layer, a data acquisition layer, a decision-making layer and an execution layer; The sensing layer includes several vibration sensors 5, a binocular vision measurement system 9, a visual target 4, a pressure sensor, a displacement sensor 3, a laser and vision sensor, and a laser guidance system. The laser guidance system includes a total station 11 and a laser target 8.

[0066] Vibration sensors 5 are installed at the locations of the TBM cutterhead 1, main drive, front shield 2, support shield 7, and support shoe 10 to measure vibration data during the tunneling process. A binocular vision measurement system 9 and a visual target 4 are installed at both ends of the propulsion cylinder 6 to identify the position and orientation of the TBM cutterhead 1 relative to the support shoe 10. Combined with a laser guidance system, this enables the measurement of the global position and attitude coordinates of the TBM cutterhead 1.

[0067] A high-frequency pressure sensor is installed in the large and small oil chambers of the propulsion cylinder 6 to acquire millisecond-level pressure data in real time. A high-frequency displacement sensor 3 is installed on each propulsion cylinder 6 to acquire displacement data of each propulsion cylinder 6 in real time. A high-frequency pressure sensor is installed on the TBM support shoe cylinder to obtain the support force; at the same time, data such as the set propulsion speed, set propulsion pressure, and set TBM cutter head 1 rotation speed of each cylinder are acquired through the TBM main control PLC.

[0068] The acquisition layer consists of distributed data acquisition terminals for various sensors. It is used to receive data acquisition instructions such as sampling rate, sampling time, and sampling duration from edge data access terminals, and synchronously acquire multi-source sensing data from each sensor in the perception layer according to corresponding rules.

[0069] The decision-making layer is a high-performance edge controller that performs data preprocessing, tunneling strategy generation, autonomous tunneling control parameter calculation, and feeds back to the execution layer. Specifically: 1) Responsible for managing the data acquisition terminal and building the database. The raw data acquired by multi-source sensors is saved according to timestamp, object, and sampling rules. The data is further normalized, denoised, and aligned using algorithms such as multi-source heterogeneous data alignment and filtering, and then saved to the preprocessing database. 2) Generate a strategy library, run the algorithm model, and generate a tunneling strategy that is adapted to the current geology and whether the tunnel is turning, based on the safety identification results. 3) Calculate the autonomous tunneling control parameters, and fine-tune the generated basic strategy based on conditions such as vibration, cutting tool and main bearing stress, and generate the final control parameters, which are then fed back to the TBM main control PLC.

[0070] The execution layer includes the TBM main control PLC, motor controllers of various execution systems such as the main drive motor / belt conveyor, and cylinder / motor control valves of various components such as the propulsion cylinder 6 and the support shoe cylinder, which are used to receive autonomous tunneling control parameters and execute tunneling actions.

[0071] Therefore, the present invention adopts the above-mentioned autonomous tunneling method and system for ultra-small turning TBMs, which are adapted to the construction requirements of ultra-small turning TBMs. It effectively solves the problems of poor adaptability of existing technical parameters and lack of targeted control strategies, realizes autonomous tunneling under all working conditions, takes into account accuracy, safety and efficiency, and can also iteratively optimize strategies online, which greatly improves the level of intelligence and tunneling efficiency of hard rock tunnel construction.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for autonomous tunneling using a TBM with ultra-small turning radius and autonomous tunneling, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-sensor data and construction of state vectors; S2. Monitor safety indicators and determine if they exceed limits. If limits are exceeded, switch to safety control mode. S3. Select the corresponding basic control strategy based on the geological identification results to provide boundary constraints for control commands; S4. Employ the Actor-Critic architecture to output fine-tuning values ​​for control parameters and evaluate the benefits of the action; S5. The maximum allowable speed is dynamically adjusted according to the real-time vibration level, and the vibration intensity is inversely proportional to the maximum speed; S6. Send the final speed command to each propulsion cylinder controller, collect the performance data after execution, and calculate the multi-objective reward; S7 stores state transition experience and performs batch training, updates the network parameters of the Actor-Critic architecture, and realizes online optimization of the control strategy.

2. The autonomous tunneling method and ultra-small turning TBM according to claim 1, characterized in that, In S1, data from geological sensors, attitude sensors, vibration sensors, pressure sensors, and position sensors are collected in real time. Constructing state vectors =[Geological characteristics, vibration level, pressure state, trajectory deviation, velocity fluctuation]; The geological characteristic states include four states: turning condition, uniform soft rock, uniform hard rock, and heterogeneous soft-hard rock. The state vector of the geological characteristic state is represented as follows: =[1 / 0,1 / 0,1 / 0,1 / 0]； The values ​​indicate whether the rock is a very small turn, uniform soft rock, uniform hard rock, or uneven rock; where 1 represents yes and 0 represents no. The attitude vector at the current moment is represented as ,in, , , Indicates the spatial coordinates of the current time. , , Represents spatial attitude coordinates; the current attitude vector is guided by the laser to the pose. Visual measurement of pose And the kinematic forward pose based on the propulsion system configuration and cylinder extension position Together, we obtain the following expression: = +f( , )； In the formula, This indicates that the visually measured pose and the orthogonal pose are fused and calculated. The trajectory deviation state represents the amount and trend of the deviation between the current pose and the design axis. The expression for the trajectory deviation vector is as follows: ； The first row represents the deviation between the current pose and the previous spatial pose, and the second row represents the rate of change of the deviation. This represents the rate of change of deviation, with a value range of [-1, 1]. A positive value indicates that the deviation is increasing, while a negative value indicates that the deviation is decreasing. The vibration horizontal state vector is represented as: = ； in, Indicates different vibration measurement points, Indicates vibration intensity; ； in, This indicates the number of measurements or samples taken for vibration values ​​at the same side point. express The maximum value, ; indicates the first The vibration measuring point at the first The magnitude of the vibration level measured in this measurement; The pressure state vector represents the pressure value and pressure balance index of each propulsion cylinder. The expression is: ； in, This indicates the pressure in the large chamber of the hydraulic cylinder. This represents the pressure in the small chamber of the propulsion cylinder. The pressure in both chambers is obtained from a pressure sensor, and the pressure state vector is represented as: ； in, Establish pressure balance indicators for each propulsion cylinder; Speed ​​fluctuations include cutterhead rotation speed fluctuations and feed speed fluctuations. Speed ​​balance indicators are established for each feed cylinder and cutterhead. ； in, This indicates the current cutting head rotation speed and propulsion speed at the moment of propulsion. in, The average speed of the cutterhead and each propulsion cylinder during the current tunneling cycle is expressed as follows: ； in, This indicates the number of historical data points traced back to the beginning.

3. The autonomous tunneling method and ultra-small turning TBM according to claim 2, characterized in that, S2 specifically refers to: Monitor the pressure of each hydraulic cylinder to ensure it does not exceed the safety threshold. ; Establish pressure balance indicators for each propulsion cylinder: ； The average thrust of each cylinder in the current tunneling cycle is expressed as follows: ； The propulsion pressure command must not exceed the pressure balance index threshold. ; Monitor whether the vibration intensity exceeds the safety threshold. ; Does the monitoring trajectory deviation exceed the safety threshold? ; Monitor whether the speed deviation exceeds the safety threshold. : ； in, Indicates the maximum rotational speed of the cutter head. Indicates the propulsion speed of each hydraulic cylinder; If any safety indicator exceeds the limit, the system will switch to safety control mode, in which low-speed tunneling or manual intervention will be used. The over-limit alarm uses a threshold over-limit alarm, or determines whether to switch to safety control mode based on the number of consecutive alarms or the percentage of over-limit alarms within a set time. Alarm thresholds can be set based on experience or determined online using data distribution analysis.

4. The autonomous tunneling method for ultra-small turning TBMs according to claim 3, characterized in that, In S3, the geological identification results are obtained through rock debris image analysis and tool detection methods: For uniform soft rock conditions, a high-speed strategy is adopted; For uniform, weak, and hard rock conditions, a vibration stabilization strategy is adopted; For uneven geological conditions, an equilibrium strategy is adopted; For turning conditions, a precise strategy is adopted; While setting the reference parameters for each operating condition, add attitude deviation constraints. |≤0.

2.

5. The autonomous tunneling method for ultra-small turning TBMs according to claim 4, characterized in that, In S4, the Actor-Critic architecture includes an Actor Network and a Critic Network: An actor network outputs an action based on the current state. This action follows a Gaussian distribution, including the action mean. and action variance Mean of movement This represents the optimal action, and the action variance. This indicates the network's uncertainty about its current state; The output layer uses the tanh activation function to constrain the mean to the range [-1, 1], and the variance output is ensured to be positive through exp calculation. The final tunneling control parameters are: ； A critic network is used to evaluate the expected total reward of taking an action in a given state. The multi-objective reward function is designed as follows: ； in, , , and Here, are the weighting coefficients; The trajectory accuracy bonus is linearly negatively correlated with the absolute value of the deviation; A piecewise function is used as the reward for vibration control; The incentive for tunneling efficiency is positively correlated with average advance speed and negatively correlated with speed fluctuation. A piecewise function is used for stress safety rewards.

6. The autonomous tunneling method for ultra-small turning TBMs according to claim 5, characterized in that, In S5, the maximum allowable speed is dynamically adjusted based on the real-time vibration level, and the vibration intensity is inversely proportional to the maximum speed, as shown in the following formula: ； in, This represents the current pace of progress; The propulsion speed is dynamically adjusted. The velocity scaling factor is based on the real-time vibration intensity. The dynamic speed scaling factor is calculated using the following formula: = / max(P, ); in, This represents the baseline value for vibration intensity, which is a constant.

7. The autonomous tunneling method for ultra-small turning TBMs according to claim 6, characterized in that, S7 specifically refers to: The state transition experience is stored in the replay buffer. After executing the speed command generated by S5, a new round of sensor data is collected to obtain the new state s_{t+1}, and the multi-target reward is calculated. , It is the core indicator for measuring the quality of this action, and is derived by weighted summation of four sub-rewards: trajectory, vibration, efficiency, and safety. Batch training is performed periodically by sampling from the buffer, specifically as follows: fragments of experience The experience segments are stored in the experience replay buffer. A batch of experience segments are randomly selected from the experience replay buffer periodically for training. First, the value prediction accuracy of the critic network is trained. Then, the parameters of the actor network are adjusted according to the value gradient to achieve online optimization of the control strategy.

8. A micro-turning autonomous tunneling TBM and an autonomous tunneling system, used to implement the method described in any one of claims 1-7, characterized in that, It includes the perception layer, the data acquisition layer, the decision-making layer, and the execution layer; The sensing layer includes several vibration sensors, a binocular vision measurement system, a visual target, a pressure sensor, a displacement sensor, a laser and vision sensor, and a laser guidance system. The laser guidance system includes a total station and a laser target. The acquisition layer is a distributed data acquisition terminal used to receive data acquisition instructions and synchronously acquire multi-source sensor data from the perception layer according to rules; The decision-making layer is a high-performance edge controller that completes data preprocessing, tunneling strategy generation, autonomous tunneling control parameter calculation, and feeds back to the execution layer. The execution layer includes the TBM main control PLC, the execution system motor controller, and the cylinder / motor control valve, which are used to receive autonomous tunneling control parameters and execute tunneling actions.

9. The ultra-small turning autonomous tunneling TBM and autonomous tunneling system according to claim 8, characterized in that, Vibration sensors are installed at the TBM cutter head, main drive, front shield, support shield, and support shoe positions; A binocular vision measurement system and a vision target are installed at both ends of the propulsion cylinder, and pressure or strain sensors are installed on the radial and axial mounting surfaces of the main bearing. A triaxial force sensor is mounted on the tool mounting surface; a pressure sensor is mounted in the large and small oil chambers of the feed cylinder; a displacement sensor is mounted on each feed cylinder. The pressure sensor for the support shoe cylinder is installed on the TBM support shoe cylinder.

Images