Heading machine attitude control method based on multi-source data fusion and dynamic solution

Abstract

The invention provides a heading machine attitude control method based on multi-source data fusion and dynamic solution. The method comprises the following steps: firstly, extracting geometric deformation characteristics and defect characteristics of a well wall in real time; establishing a suspender-well wall-heading machine coupling dynamic model fused with the well wall constraint effect, taking heading machine state data, well wall geometric deformation characteristics, defect characteristics and suspender dynamic response data as observation input, and resolving the current attitude state of the heading machine in real time through an adaptive state estimation algorithm; and generating a control strategy for adjusting the attitude of the heading machine and realizing well wall obstacle avoidance by using a model prediction control algorithm containing an environmental risk penalty term. According to the method, state calculation is carried out through the built suspender-well wall-heading machine coupling dynamic model, the assumed defect that a suspender is regarded as an ideal rigid rod in a traditional geometric method is fundamentally overcome, the attitude calculation precision is improved to the centimeter level or even the millimeter level from the decimeter level, and the reliability of a guiding system is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of tunneling machine attitude control, and in particular to a tunneling machine attitude control method based on multi-source data fusion and dynamic calculation. Background Technology

[0002] As underground space development moves towards deeper and larger diameters, the geological environment faced by shaft excavation projects is becoming increasingly complex, placing higher demands on the precise guidance and safety control of the excavation process. Accurately and in real-time acquiring the three-dimensional attitude (including position and rotation angle) of the tunneling machine (such as a shield tunneling machine or a raise boring machine) in a deep shaft, and achieving stable excavation along the design axis, is crucial for ensuring the verticality of the project, preventing shaft deviation, and avoiding equipment jamming or collisions with the shaft wall.

[0003] Currently, the mainstream guidance method for shaft boring machines (TBMs) mainly relies on indirect measurements using sensors mounted on the boom (or stabilizer bar). A typical technical approach involves installing tilt sensors, accelerometers, or fiber optic gyroscopes at the top of the boom or specific points. By measuring the boom's tilt angle and combining this with geometric parameters such as boom length and wire rope lowering depth, the spatial position of the TBM's bottom is calculated using simple trigonometric relationships. More advanced systems incorporate optical measurement methods such as total stations and laser plumb bobs for periodic verification and calibration.

[0004] However, the aforementioned existing technologies have revealed the following significant defects and limitations in engineering practice: 1. Large model simplification errors, failing to reflect true dynamic states: Existing technologies generally simplify the boom as an "ideal rigid rod" or only consider a "flexible beam" with static bending, and their solution models are based on static or quasi-static geometric relationships. In reality, during tunneling, the boom is subjected to multiple sources of excitation, such as tunneling machine vibration, cutterhead disturbance, underground turbulence, and wire rope oscillation, resulting in complex dynamic bending, torsion, and higher-order vibrations. Ignoring these dynamic effects and directly treating the inclination angle at the top of the boom as a direct mapping of the bottom attitude introduces non-negligible dynamic errors. Especially when the boom length is large and the excitation is significant, the calculated results may deviate from the actual position and attitude by decimeters, failing to meet the requirements for high-precision guidance.

[0005] 2. Sensitive to environmental disturbances and poor robustness: The internal environment of the shaft is harsh, with various disturbances such as water flow impact, rockfall, and equipment noise. The sensors used in existing systems (such as inclinometers) are easily affected by these transient disturbances, resulting in data jumps or drifts. More critically, existing technologies lack the ability to directly perceive the state of the shaft wall. When the tunneling machine or boom approaches an irregular shaft wall, encounters local protrusions or convergence sections, traditional geometric calculation models are completely unable to perceive this "contact" or "near-contact" constraint. This leads to a significant discrepancy between the calculated "theoretical position" and the actual reachable "physical position," easily inducing guidance decision errors and causing equipment scraping or even shield jamming accidents.

[0006] 3. Lagging and blind control strategies, lacking foresight and safety: Current control systems mostly employ classic PID feedback control based on the current attitude deviation. This control method is "post-event correction," unable to predict the relative motion trend between the tunnel boring machine and the well wall in the next few steps. Because it lacks integrated real-time environmental perception (especially well wall distance information), the control system is "unaware" of impending collision risks and can only respond passively after a collision or anomaly occurs, resulting in low safety margins. Furthermore, control parameters are usually fixed and cannot be adaptively adjusted according to different geological conditions (such as soft rock, hard rock, or the presence of fracture zones) and well wall conditions.

[0007] 4. The system is isolated and rigid, lacking self-learning and evolution capabilities: Existing guidance systems typically operate as a functionally fixed "black box." Once its internal model parameters are set, they are difficult to adjust, making it impossible to utilize the massive amounts of data continuously generated during tunneling (such as boom response under different geological formations and changes in wellbore morphology) for self-optimization. As the tunneling distance increases and geological conditions change, system performance may gradually degrade. Simultaneously, the system lacks the ability to monitor and diagnose the health status of the sensors themselves, making it difficult to guarantee measurement reliability during long-term service. Summary of the Invention

[0008] In view of the above situation, the main objective of this invention is to propose a tunneling machine attitude control method based on multi-source data fusion and dynamic calculation to solve the above-mentioned technical problems.

[0009] This invention proposes a method for attitude control of a tunneling machine based on multi-source data fusion and dynamic calculation. The method includes the following steps: Step 1: Deploy at least two smart sensor nodes on the boom. During the tunneling operation, the smart sensor nodes will synchronously collect dynamic response data of the boom itself, multi-directional distance data from the shaft wall, and shaft wall environment data. Step 2: Based on the dynamic response data of the boom itself and the distance data of the well wall in multiple directions, establish a coupled dynamic model of boom-well wall-tunneling machine that integrates the constraint effect of the well wall. Based on the observation state and the state data of the tunneling machine, the coupled dynamic model of boom-well wall-tunneling machine is solved in real time through an adaptive state estimation algorithm to obtain the current attitude state of the tunneling machine. Step 3: Generate environmental risk characteristics using multi-directional wellbore distance data and wellbore thermal imaging data. Based on the current attitude state of the tunneling machine and the environmental risk characteristics, generate the control strategy for the tunneling machine using model predictive control algorithms. Step 4: Based on the multi-directional wellbore distance data accumulated during the tunneling process, construct and update the wellbore geometric model online; Step 5: Use the shaft geometry model as a dynamic spatial constraint for the future trajectory of the tunnel boring machine in the model predictive control algorithm to optimize the control strategy, obtain the final control strategy, and use the final control strategy to adjust the attitude and parameters of the tunnel boring machine in real time.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Traditional methods rely solely on the tilt angle of the boom top or a single geometric relationship to calculate the tunneling machine's attitude, failing to detect the boom's dynamic deformation and the well wall constraints. Errors increase dramatically when the boom swings, vibrates, or approaches the well wall. This invention, by deploying multi-mode sensing nodes on the boom and introducing a laser ranging module to directly sense the boom-well wall distance, can acquire real-time data on the boom's dynamic bending, torsion, and interaction with the well wall. By establishing a coupled dynamic model of the boom-well wall-tunneling machine for state calculation, it fundamentally overcomes the traditional geometric method's assumption of treating the boom as an "ideal rigid rod," improving attitude calculation accuracy from decimeters to centimeters or even millimeters, significantly enhancing the reliability of the guidance system.

[0011] 2. In complex operating conditions such as irregular well walls, water seepage, strong water flow impact, or rock collisions, sensor data is easily contaminated, and the guidance system is prone to failure or issuing erroneous commands. This invention, through multi-source data fusion and feature extraction, enables the system to proactively identify environmental interference (e.g., identifying water seepage through infrared thermal imaging and identifying well wall protrusions through a ranging array). In the state resolution stage, the adaptive extended Kalman filter dynamically adjusts the noise matrix based on the real-time perceived minimum distance, "recognizing" the increased uncertainty when approaching the well wall; the deep learning-assisted compensation module learns and corrects system errors caused by well wall irregularities. This allows the system to maintain a stable and reliable operating state even in harsh, non-ideal well environments.

[0012] 3. Existing control strategies are mostly based on PID regulation of the current deviation, which is "post-mortem correction" and cannot predict future states, and the response to well wall collision risks is delayed. This invention adopts model predictive control based on environmental perception, introducing an environmental risk penalty term into the objective function, enabling the control system to automatically avoid high-risk trajectories close to the well wall during optimization. By constructing a well wall accessibility map and calculating the optimal passage direction, active obstacle avoidance control is achieved. At the same time, the system can dynamically adjust control constraints based on identified well wall defects (such as protrusions and water seepage), for example, reducing the maximum correction force in the direction of protrusions to avoid "head-on collisions". This forward-looking and intelligent control strategy can effectively prevent accidents such as shield jamming and well wall scraping, ensure construction safety, reduce downtime caused by obstacle avoidance, and improve pure tunneling efficiency.

[0013] 4. Existing technologies typically only output the tunneling machine's position and orientation, lacking in-depth perception and presentation of the process, environment, and the equipment's own status. This invention not only outputs high-precision guidance results but also provides real-time multi-dimensional information such as wellbore ellipticity convergence curves, defect distribution heatmaps, boom vibration spectra, and environmental risk warnings. This comprehensive status visualization enables operators to transition from "blind operation" to "transparent management," allowing for early detection of geological anomalies, assessment of wellbore quality, and prediction of potential risks. This leads to more scientific and timely construction decisions, enhancing the overall intelligence level of project management.

[0014] 5. This invention updates the wellbore geometry model through online learning, allowing the system to "memorize" and predict the shape of the well wall ahead, providing forward-looking information for control. For harmful vibrations such as boom swaying, the system can actively suppress them by finding the optimal combination of tunneling parameters through reinforcement learning. Furthermore, the self-diagnosis and automatic calibration functions of the sensors ensure the accuracy of long-term measurements. This enables the entire system to continuously optimize its performance as the tunneling progresses, adapting to the needs of long-distance, geologically variable tunneling, extending the system's effective service life, and reducing maintenance costs.

[0015] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0016] Figure 1 The flowchart shows the tunneling machine attitude control method based on multi-source data fusion and dynamic calculation proposed in this invention. Figure 2 This is a schematic diagram of the tunneling process of the tunneling machine attitude control method based on multi-source data fusion and dynamic calculation proposed in this invention; In the diagram, 1. Tunneling machine, 2. boom, 3. steel wire, 4. electric winch, 5. ground sealing plate, 6. station changing device, 7. after-sales equipment, 8. total station, 9. intelligent sensor node, 10. ring reflective strip. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0019] The intelligent guidance and control system for shaft boring machines based on multi-source data fusion provided in this invention mainly includes the following hardware components: an array of intelligent sensor nodes deployed on the boom, a central processing and main control unit at the wellhead or fixed platform, and actuators on the boring machine body (such as correction cylinders and cutterhead drive systems). The system transmits data and power between the boom nodes and the ground unit via a slip-ring type integrated power supply and communication cable. The software components include a data synchronization acquisition module, a multi-source feature extraction module, a coupled dynamic calculation module, a model prediction control module, and a self-learning management module running on the central processing unit.

[0020] Please see Figure 1 This embodiment provides a method for attitude control of a tunneling machine based on multi-source data fusion and dynamic calculation. The method includes the following steps: Step 1: Deploy at least two smart sensor nodes on the boom. During the tunneling operation, the smart sensor nodes will synchronously collect dynamic response data of the boom itself, multi-directional distance data from the shaft wall, and shaft wall environment data. like Figure 2 As shown in the schematic diagram, the main components include a tunneling machine (1), a boom (2), a steel wire (3), an electric winch (4), a ground sealing plate (5), a station-changing fixing device (6), a rear-mounted accessory (7), and a total station (8). The electric winch is connected to the boom via a steel wire, which in turn is connected to the tunneling machine via a steel wire. The rear-mounted accessory is positioned above the tail of the tunneling machine and secured by a station-changing fixing pile. The ground sealing plate covers the wellhead. The total station is positioned directly above the wellhead. Intelligent sensor nodes (9) and a ring-shaped reflective strip (10) are mounted on the boom.

[0021] Taking a boom with a length of 1500mm and a diameter of 200mm as an example, two smart sensor nodes are deployed on each boom.

[0022] The upper node S1 is installed 300mm from the top of the boom; the lower node S2 is installed 300mm from the bottom of the boom (near the tunneling machine connection).

[0023] Each smart sensing node integrates the following sensors: Triaxial MEMS accelerometer: range ±8g, bandwidth 500Hz, used to measure the vibration acceleration of a boom in three directions.

[0024] Dual-axis tilt sensor: accuracy ±0.01°, used to measure the tilt angle of the boom in two mutually perpendicular directions.

[0025] Four-way laser ranging module: Four ranging heads are arranged at 90° intervals along the circumference, with a measurement range of 0.1-5m and an accuracy of ±2mm. It is used to directly measure the distance between the boom and the surrounding well walls at the node, providing multi-directional well wall distance data.

[0026] Miniature infrared thermal imager: Used to capture infrared images of the well wall, monitor temperature anomalies, and collect thermal image data of the well wall.

[0027] Temperature and humidity sensor: Monitors the environment around the node to collect ambient temperature and humidity data.

[0028] Wellbore environmental data includes wellbore thermal imaging data and ambient temperature and humidity data.

[0029] The dynamic response data of the boom itself includes vibration acceleration and tilt angle.

[0030] The tunneling machine status data includes the advance speed, the thrust of the correction cylinder, and the cutterhead status parameters, which are acquired by the tunneling machine control system.

[0031] The surface of the boom features enhanced design: laser etching technology is used to create a three-dimensional coded pattern (e.g., a QR code matrix composed of micro-grooves with a depth of 0.1-0.5 mm) on the boom surface. This pattern can serve as feature points for high-precision pose registration during LiDAR scanning, and can also encode and store the boom's unique identification ID. Furthermore, a 20 mm wide annular reflective strip is wrapped around each installation node to further enhance the reflection intensity of the LiDAR signal.

[0032] All sensor nodes have built-in clock chips supporting the IEEE 1588 PTP protocol, and are uniformly synchronized by a high-precision master clock deployed at the wellhead, achieving microsecond-level time synchronization of all data streams. Spatial synchronization is achieved through initial calibration and online registration: when the boom is stationary, a total station is used to accurately measure the coordinates of the upper node S1 and the lower node S2 in the global coordinate system to establish a reference; during operation, the lidar identifies the three-dimensional coded pattern and reflective strip on the boom surface, calculates the boom's pose relative to the radar coordinate system in real time, and, combined with known installation relationships, unifies all sensor data into the global coordinate system.

[0033] Data acquisition is performed at a fixed period of T=0.5s. The data collected in each period includes: data from two smart sensor nodes and tunneling machine status data.

[0034] Step 2: Based on the dynamic response data of the boom itself and the distance data of the well wall in multiple directions, establish a coupled dynamic model of boom-well wall-tunneling machine that integrates the constraint effect of the well wall. Based on the observation state and the state data of the tunneling machine, the coupled dynamic model of boom-well wall-tunneling machine is solved in real time through an adaptive state estimation algorithm to obtain the current attitude state of the tunneling machine. Coupled dynamic modeling: Treating the boom as an Euler-Bernoulli beam subjected to tension T, distributed damping c, external excitation, and wellbore forces, we establish its partial differential equation for lateral vibration in an asymmetric wellbore: ; in, This indicates the linear density of the boom. This represents the equivalent viscous damping coefficient of the suspension system. This indicates the bending stiffness of the suspension rod. This indicates the axial tension applied by the wire rope to the boom. This represents the lateral displacement of the boom at axial position z and time t. This represents the lateral acceleration of the boom at position z and time t. This represents the lateral velocity of the boom at position z and time t. This represents the spatial derivative related to bending deformation. The spatial derivative related to axial tension is obtained by approximating the rod curvature, which is calculated by differentiating the tilt angle and combining it with the rod length. This refers to the external excitation force acting on the boom, such as tunneling machine vibration or water flow impact. This represents the contact or near-field force exerted by the wellbore on the boom, the magnitude of which is determined by the boom displacement. and multi-directional laser ranging data Jointly determined. Wellbore forces. Using real-time multi-directional laser ranging data The inversion uses a simplified contact spring-damping model. When the distance is less than a set threshold, the force term increases significantly to simulate wellbore constraints.

[0035] Adaptive Extended Kalman Filter Solution: State vector design: It has 14 dimensions in total, including the tunneling machine's pose, velocity, and angular velocity. Among them, This indicates the three-dimensional position of the tunnel boring machine's tail section in the global coordinate system. This indicates the roll angle, pitch angle, and yaw angle of the tunneling machine. This indicates the linear velocity of the tunneling machine in three directions. This represents the angular velocity of the tunneling machine about its three axes. This indicates the transpose operation.

[0036] Observation status: It includes data collected by all intelligent sensing nodes, namely the dynamic response data of the boom itself, multi-directional wellbore distance data, and wellbore environment data.

[0037] Based on the coupled dynamics model, an adaptive extended Kalman filter is used to drive state prediction with the current state data of the tunnel boring machine, and the predicted values ​​are updated with the observed states. The state vector is recursively estimated to obtain the estimated value of the state vector. The corresponding process has the following relationship: ; in, This represents the estimated value of the state vector calculated using the extended Kalman filter based on the coupled dynamics model of the boom-shaft-tunneling machine. Indicates the observation status. Indicates extended Kalman filtering, This represents the noise covariance matrix of the adaptive process. This indicates the status data of the tunneling machine. Represents the observation noise covariance matrix. The nonlinear state transition function of the coupled dynamic model of boom-bore-tunneling machine based on well wall constraints is obtained by spatial and temporal discretization of the coupled dynamic model of boom-bore-tunneling machine based on well wall constraints. The calculation process of the noise covariance matrix in the adaptive process follows the following relationship: ; in, Represents the reference noise matrix. Indicates the safe distance threshold. It represents a very small positive number; when the real-time minimum distance between the boom and the well wall is less than the safety threshold (e.g., 200mm), the process noise covariance corresponding to the position (x, y, z) component in the state estimate is automatically increased to characterize the increased model uncertainty due to possible contact with the well wall, thereby improving the robustness of the filter under dangerous conditions and realizing the adaptive process.

[0038] Deep learning-assisted dynamic compensation: The dynamic response data of the boom itself includes vibration acceleration and tilt angle. A sliding window FFT analysis is performed on the vibration acceleration to obtain the vibration spectrum of the boom. A pre-trained ConvLSTM hybrid neural network is run in parallel. The pre-trained ConvLSTM hybrid neural network takes as input an 8-axis distance sequence (8×10 matrix) over the past 10 cycles (5 seconds) and a boom vibration spectrum, and outputs an attitude correction. This correction is primarily used to compensate for EKF estimation residuals caused by extreme wellbore irregularities and nonlinear factors not fully covered by the model.

[0039] Final Fusion: ,in, Indicates adaptive fusion weights, This represents the correction amount output by the deep learning model (ConvLSTM is used in this embodiment), which mainly compensates for nonlinear factors such as extreme irregularities of the wellbore. This represents the current optimal state estimate, i.e., the current attitude state of the tunneling machine. The adaptive fusion weights are adaptively adjusted based on the real-time calculated wellbore roughness (standard deviation of the distance sequence). When the roughness is high, the adaptive fusion weights are increased, relying more on data-driven compensation. The weights are dynamically adjusted between 0 and 1 based on confidence indicators such as wellbore roughness.

[0040] Step 3: Generate environmental risk characteristics using multi-directional wellbore distance data and wellbore thermal imaging data. Based on the current attitude state of the tunneling machine and the environmental risk characteristics, generate the control strategy for the tunneling machine using model predictive control algorithms. Optimize problem formulation: Prediction Model: The coupled dynamic model of boom-wellbore-tunneling machine under wellbore constraint is used as the internal prediction model.

[0041] Objective function: ; in, This represents the prediction of the system state at time t+k using the current optimal state estimate at the current time t. This indicates the sequence of future control strategies to be optimized, including the steering cylinder thrust, cutter head speed adjustment, and propulsion speed to be optimized. The reference trajectory represents the geometric state of the tunneling machine moving along the design axis; This represents the total number of future steps predicted by the MPC algorithm. Indicates the step index in the prediction time domain. Indicates the current moment. This represents the weighting coefficients of the tracking accuracy requirements for different components in the state vector within the optimization objective. The weighting coefficients represent the energy consumption of the correction cylinders in different directions in the optimization objective. This represents the weighting coefficient of environmental risk in the optimization objective. Indicates the use of Weighted square norm, Indicates the use of Weighted square norm, This indicates environmental risk penalties; among them, , This indicates the initial state of MPC; The calculation process for environmental risk penalties follows the following formula: ; in, The normalized value of the seepage risk index is obtained by: performing temperature field analysis on the thermal imaging data of the well wall, identifying low temperature or high temperature anomaly areas where the average temperature difference with the surrounding well wall exceeds 3℃, marking them as potential seepage points, obtaining the seepage risk index, and normalizing the seepage risk index to obtain the normalized value of the seepage risk index. The normalized value representing the characteristics of well wall defects is obtained by comparing adjacent points of the well wall distance data in multiple directions. If the distance difference between consecutive measuring points exceeds 50mm, the area is marked as having defects, thus obtaining the characteristics of well wall defects. The characteristics of well wall defects are then normalized to obtain the normalized value of the characteristics of well wall defects. The seepage risk index and the characteristics of well wall defects together constitute the environmental risk characteristics. This represents the actual environmental risk value calculated based on the measured real-time data at the current time t. This represents a given absolute safety distance threshold. These represent the influence factors of humidity and wellbore defects, respectively, adjusting their weights in risk assessment. At the same distance from the wellbore, a high-humidity environment (meaning a more slippery, unstable, or leaky wellbore) will trigger a stronger obstacle avoidance control strategy. Furthermore, even at a greater distance, as long as a defect area is identified ahead (defect features are marked in the wellbore geometry model), the control strategy will be adjusted in advance to adjust the attitude, achieving proactive, predictive obstacle avoidance.

[0042] Dynamic adjustment of constraints: When the feature extraction module identifies a well wall protrusion in a specific direction, the upper limit of the correction force in that direction is instantly reduced by 30%; when a seepage area is identified, constraints are added to the optimization problem to prevent the control strategy from causing the tunneling machine to stay in that area for a long time.

[0043] The basic control quantity is obtained by iteratively optimizing the objective function. .

[0044] Wellbore adaptive obstacle avoidance: Based on 8 ranging points, the well wall distance distribution of 128 points in the circumferential direction is generated by interpolation, a accessibility map is constructed, and areas on the accessibility map with a point distance of less than 250mm are marked as "high-risk areas".

[0045] Calculate the optimal travel direction based on the accessibility map. To keep it as far away from high-risk areas as possible, i.e., to solve ,in, Indicates the optimal travel direction. Indicates a circumferential angle; This represents the weighting function, used to penalize directions that are closer to high-risk areas; This represents the continuous distance distribution function with respect to the circumferential angle θ, fitted using an interpolation algorithm.

[0046] Based on the optimal travel direction and the current orientation angle of the tunneling machine, the obstacle avoidance feedforward term is calculated: ; in, This represents the feedforward compensation amount used for real-time obstacle avoidance. Indicates obstacle avoidance control gain, and Inversely proportional, the closer the distance, the stronger the obstacle avoidance intervention; Indicates the optimal travel direction. This indicates the current orientation angle of the tunneling machine.

[0047] The basic control quantity calculated by MPC By superimposing an obstacle avoidance feedforward term, the control strategy of the tunnel boring machine is obtained: ,in, It indicates the control strategy of the tunneling machine and sends it to the correction cylinder and cutterhead drive system, etc.

[0048] Abnormal operating condition response: Real-time system monitoring Implement a three-level early warning system: A yellow warning light illuminates on the interface when the distance is less than 300mm. Automatically reduce propulsion speed by 50% when the distance is less than 200mm; When the distance is less than 150mm, an emergency pause is triggered and manual confirmation is requested.

[0049] When a "large protrusion" (three consecutive measuring points < 200mm) is detected, the human-machine interface will display a prompt suggesting that the cutterhead speed be reduced by 20% in this area. When a seepage point (temperature difference > 5℃) is detected, the location will be automatically marked in the geological model and a prompt suggesting grouting treatment will be displayed.

[0050] Step 4: Based on the multi-directional wellbore distance data accumulated during the tunneling process, construct and update the wellbore geometric model online.

[0051] The construction and online updating of the wellbore geometric model are detailed below: A parameterized shaft geometry model is constructed, with the tunneling distance L as the independent variable. The shaft geometry model is derived from the shaft centerline function. and the equivalent radius function of the wellbore Composition; where the shaft centerline function describes the position of the ideal shaft axis in three-dimensional space, and the coordinates of the shaft center point at the current position of the tunneling machine are represented as: The equivalent radius function of the wellbore is used to describe the change in the size of the wellbore at each cross-section along the centerline. The size of the wellbore at the centerline cross-section at the current moment is expressed as... .

[0052] Wellbore geometry model online update: The least squares method is used to fit the elliptical model of the well wall of the current tunneling section in real time for the multi-directional well wall distance data to obtain the major axis a, minor axis b and the offset of the ellipse center; and the ellipticity is calculated based on the major axis a, minor axis b and the offset of the ellipse center; local polynomial fitting is performed on the multi-directional well wall distance data to calculate the local curvature, and the ellipticity and local curvature constitute the geometric deformation characteristics of the well wall; After tunneling a set distance, the system automatically summarizes all multi-directional laser ranging data collected within the current distance, the synchronously calculated high-precision tunneling machine trajectory points and their corresponding tunneling distance L, as well as the well wall geometric deformation features corresponding to each trajectory point, forming a structured data package for model updates. PCA analysis is performed on the projection of all tunneling machine trajectory points in the current structured data package onto the horizontal plane to extract the direction of the first principal component; the direction of the first principal component is determined as the main trend of the current well section centerline on the horizontal plane; In the vertical direction, based on the characteristic that tunneling operations are basically vertically downward, the vertical coordinates of the centerline are directly determined by linear correlation with the tunneling mileage L or by taking the average value of the measured vertical coordinates of that segment. Project all trajectory points onto the direction of the first principal component and calculate the average coordinates of the projected points as an estimated point of the current segment's mileage centerline; Set a sliding window with a fixed length and step size (10 meters long and 5 meters step size), perform cubic spline interpolation on the centerline estimation points of multiple consecutive mileages along the tunneling direction, and finally generate a globally continuous and smooth wellbore centerline function. For each tunneling machine trajectory point in the structured data packet, the normal distance from each sampling point on the well wall on the current cross section to the current centerline is calculated by combining the pose of the intelligent sensor node on the boom and its corresponding multi-directional laser ranging value when the trajectory point is collected. The median of all normal distances on the same cross section is taken as the instantaneous equivalent radius at the current tunneling mileage L; Collect all mileage-instantaneous equivalent radius data pairs within the current structured data package, fit them using a local weighted regression scatter smoothing method to obtain the radius variation trend of the current well section, and update the global wellbore equivalent radius function using the radius variation trend of the current well section. Simultaneously, all mileage-instantaneous ellipticity data pairs and mileage-instantaneous curvature data pairs within the current structured data package are collected, and the local weighted regression scatter smoothing method is used to fit them respectively to obtain the ellipticity change trend function and curvature change trend function of the current well section. A shaft geometric model is constructed based on the shaft centerline function, shaft equivalent radius function, ellipticity variation trend function, and curvature variation trend function. The shaft centerline function, shaft equivalent radius function, ellipticity variation trend function, and curvature variation trend function are updated by increasing the tunneling mileage to update the shaft geometric model.

[0053] The updated shaft centerline function, shaft equivalent radius function, shaft ellipticity function, and shaft curvature function are used together to predict the possible shape of the shaft wall in the unexplored area ahead. In the model predictive control algorithm, this predictive information is transformed into dynamic spatial constraints on the future trajectory of the tunnel boring machine and is reflected in the forward-looking environmental risk penalty term.

[0054] Specifically, based on the wellbore centerline function and the wellbore equivalent radius function, the foundation boundary of the well wall at the future tunneling mileage can be constructed, i.e., the radius is... A circular pipe. For safety, the controller employs a recessed safety boundary. As a dynamic space constraint, where... Indicates safety margin. Tunneling machine predicted position. The distance from this safety boundary directly affects the forward-looking risk penalty items.

[0055] Simultaneously, based on the wellbore ellipticity function and wellbore curvature function, the degree of geometric irregularity of the well wall at future mileages can be assessed. Greater ellipticity or curvature indicates a more irregular well wall at that location, and a higher tunneling risk. This information is also incorporated into the forward-looking risk penalty, guiding the controller to adopt a more conservative control strategy when traversing irregular areas.

[0056] By using a wellbore geometric model, the MPC algorithm is provided with the geometric constraints and risk characteristics of the future wellbore, enabling the controller to simultaneously consider basic spatial constraints and local irregularity risks when optimizing the objective function, thereby generating an optimal control strategy that balances safety and efficiency.

[0057] The specific optimization of the tunneling strategy is as follows: By performing time-frequency analysis on the minimum distance in the distance sequence, the periodic fluctuations caused by the boom swing are identified, and their dominant swing frequency is extracted. .

[0058] When monitoring the dominant oscillation frequency When the frequency is >0.5Hz and the oscillation amplitude is large, the system enters reinforcement learning mode.

[0059] In reinforcement learning mode, the system initiates a reinforcement learning process based on the proximal policy optimization algorithm, optimizing the tunneling parameters online through reinforcement learning. Reinforcement learning includes a state space, an action space, and a reward function.

[0060] The state space describes the system conditions that the learning algorithm needs to observe at each decision moment. The state variables defined in this embodiment include: the boom's swing frequency, swing amplitude, well wall ellipticity, minimum distance between the boom and the well wall, current tunneling mileage, and tunneling machine state data, including propulsion speed, hydraulic cylinder thrust, and cutterhead state parameters. To ensure that variables with different physical dimensions are treated equally in the algorithm, the system standardizes all state variables before input, unifying their values ​​to between 0 and 1.

[0061] The action space represents the adjustment operations that the learning algorithm can perform. In this embodiment, an action is defined as a relative adjustment to two key tunneling parameters (i.e., the exploration boundary of the reinforcement learning algorithm's action space): one is the percentage adjustment of the advance speed relative to its rated value, with an allowable adjustment range of [-20%, +20%]; the other is the percentage adjustment of the cutterhead rotation speed relative to its rated value, with an allowable adjustment range of [-15%, +15%]. Both of these adjustment amounts are continuously changing values, allowing for fine-tuning.

[0062] The reward function is the core of reinforcement learning, quantifying the merits of each action. In this embodiment, the reward function aims to simultaneously achieve the goals of "suppressing oscillations" and "maintaining efficiency." The reward value consists of the following four parts, the sum of which constitutes the total reward at that moment: Swing amplitude penalty: This item is negatively correlated with the current swing amplitude. That is, the larger the swing amplitude, the greater the penalty and the lower the reward value.

[0063] Propulsion speed adjustment penalty: This item is negatively correlated with the absolute value of the propulsion speed adjustment. The larger the adjustment range, the greater the penalty, in order to avoid excessively frequent or drastic speed changes.

[0064] Cutter head speed adjustment penalty: This item is negatively correlated with the absolute value of the cutter head speed adjustment. Its function is similar to that of the speed adjustment penalty, aiming to maintain the stability of parameter adjustment.

[0065] Tunneling efficiency bonus: This bonus is positively correlated with the tunneling progress per unit time. The greater the progress, the higher the bonus, to ensure that tunneling efficiency is not excessively sacrificed.

[0066] The system balances the priority between sway suppression and tunneling efficiency by adjusting the weight coefficients of these four components. For example, when safety requirements are high, the weight of the sway amplitude penalty term can be increased; when efficiency requirements are high, the weight of the tunneling efficiency reward term can be increased.

[0067] This embodiment employs a proximal policy optimization algorithm as the core framework of reinforcement learning. This algorithm comprises two neural networks: a policy network and a value network. The policy network is responsible for evaluating and selecting the optimal action based on the current state; the value network evaluates the value of the current state, guiding the policy network's update direction. Both networks are designed as fully connected neural networks with two hidden layers, where the first hidden layer contains 64 neurons and the second hidden layer contains 32 neurons, using linear rectified units as activation functions. The algorithm trains the network by interacting with a simulation environment built based on an established shaft wall constraint-shaft wall-tunneling machine coupled dynamics model, or by replaying recorded historical operation data for offline learning, thereby continuously optimizing the network parameters. Finally, long-term optimal policy parameters are generated through reinforcement learning, including recommended propulsion speed and recommended cutterhead rotation speed.

[0068] By using the predicted wellbore morphology of the unexplored area as a forward-looking spatial safety constraint for the objective function, and using the long-term optimal strategy parameters as a dynamic reference part of the objective function, the objective function is adjusted to obtain the adjusted objective function. The corresponding process has the following relationship: ; in, This represents the coordinates of the wellbore center point at the predicted future time t+k. This represents the equivalent radius of the wellbore corresponding to the predicted future time t+k. This indicates a forward-looking environmental risk penalty item; Indicates from The location portion extracted is the predicted location of the tunneling machine; This represents the dynamic reference trajectory, which embodies the desired state that integrates a predetermined geometric path with the real-time optimization process objective. The calculation process for forward-looking environmental risk penalties follows the following formula: ; in, express and and The minimum distance between them This represents the ellipticity at future mileage predicted from the wellbore geometry model. This represents the curvature value predicted from the wellbore geometry model at future mileage. These represent the weights corresponding to the ellipticity and curvature values ​​at future mileage, respectively. In this embodiment, the current values ​​of humidity and defects are approximated as future values.

[0069] The calculation process of the dynamic reference trajectory is governed by the following relationship: ; in, This represents the geometric state at a future moment, as determined by the design axis. This represents the parameters of the long-term optimal strategy.

[0070] Using the adjusted objective function, a basic control variable is generated, and then combined with the feedforward compensation variable to generate the final control strategy.

[0071] Sensor health management: Laser ranging module self-diagnosis: During daily maintenance, the boom is raised to the calibrated position at a known distance from the wellhead, and automatic zero-position calibration is performed. The system periodically checks the consistency of ranging values ​​in four directions at the same node in a uniform well section. If the deviation of a certain direction from the average value of other directions is consistently greater than 5%, the channel is marked as "abnormal," and its weight is reduced or an alarm is triggered during data fusion.

[0072] Infrared thermal imager calibration: Every 24 hours, using the stable ambient temperature at the wellhead as a reference, the thermal imager is automatically calibrated at one point to correct any possible drift.

[0073] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for attitude control of a tunneling machine based on multi-source data fusion and dynamic calculation, characterized in that, The method includes the following steps: Step 1: Deploy at least two smart sensor nodes on the boom. During the tunneling operation, the smart sensor nodes will synchronously collect dynamic response data of the boom itself, multi-directional distance data from the shaft wall, and shaft wall environment data. Step 2: Based on the dynamic response data of the boom itself and the distance data of the well wall in multiple directions, establish a coupled dynamic model of boom-well wall-tunneling machine that integrates the constraint effect of the well wall. Based on the observation state and the state data of the tunneling machine, the coupled dynamic model of boom-well wall-tunneling machine is solved in real time through an adaptive state estimation algorithm to obtain the current attitude state of the tunneling machine. Step 3: Generate environmental risk characteristics using multi-directional wellbore distance data and wellbore thermal imaging data. Based on the current attitude state of the tunneling machine and the environmental risk characteristics, generate the control strategy for the tunneling machine using model predictive control algorithms. Step 4: Based on the multi-directional wellbore distance data accumulated during the tunneling process, construct and update the wellbore geometric model online; Step 5: Use the shaft geometry model as a dynamic spatial constraint for the future trajectory of the tunnel boring machine in the model predictive control algorithm to optimize the control strategy, obtain the final control strategy, and use the final control strategy to adjust the attitude and parameters of the tunnel boring machine in real time.

2. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 1, characterized in that, In step 1, the intelligent sensing node integrates a triaxial accelerometer, a biaxial inclinometer, and at least one laser ranging module for measuring the distance between the boom and the well wall, a miniature infrared thermal imager, and a temperature and humidity sensor; it is used to collect dynamic response data of the boom itself, multi-directional well wall distance data, well wall infrared images, and tunneling machine status data. The dynamic response data of the boom itself includes vibration acceleration and tilt angle; Tunneling machine status data includes propulsion speed and cutterhead rotation speed parameters; The surface of the boom is etched with a three-dimensional coded pattern for LiDAR pose registration and boom identification; the boom is also equipped with a ring-shaped reflective strip.

3. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 2, characterized in that, In step 2, a coupled dynamic model of the boom-shaft wall-tunneling machine, incorporating the constraints of the shaft wall, is established based on the boom's own dynamic response data and multi-directional shaft wall distance data. The corresponding process has the following relationship: ； in, This indicates the linear density of the boom. This represents the equivalent viscous damping coefficient of the suspension system. This indicates the bending stiffness of the suspension rod. This indicates the axial tension applied by the wire rope to the boom. This represents the lateral displacement of the boom at axial position z and time t. This represents the lateral acceleration of the boom at position z and time t. This represents the lateral velocity of the boom at position z and time t. Represents the spatial derivative related to bending deformation; The spatial derivative related to axial tension is obtained by approximating the rod curvature, which is calculated by differentiating the tilt angle and combining it with the rod length. This represents the external excitation force acting on the boom. This represents the contact or near-field force exerted by the wellbore on the boom, the magnitude of which is determined by the boom displacement. and multi-directional laser ranging data A joint decision.

4. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 3, characterized in that, In step 2, based on the observed state and the tunneling machine's state data, an adaptive state estimation algorithm is used to perform real-time calculations on the coupled dynamic model of the boom-shaft-tunneling machine to obtain the current attitude state of the tunneling machine. Specifically, this includes the following steps: Define the state vector as the tunneling machine's pose, velocity, and angular velocity; The observation status is defined as the dynamic response data of the boom itself, the multi-directional wellbore distance data, and the wellbore environment data; A coupled dynamic model of the boom-bore-tunneling machine (TBM) based on wellbore constraints is employed. An adaptive extended Kalman filter is used, with the current TBM state data driving state prediction. The predicted values ​​are updated using observed states, and the state vector is recursively estimated to obtain the estimated state vector values. The corresponding process follows the following relationship: ； in, This represents the estimated value of the state vector calculated using the extended Kalman filter based on the coupled dynamics model of the boom-shaft-tunneling machine. Indicates the observation status. Indicates extended Kalman filtering, This represents the noise covariance matrix of the adaptive process. This indicates the status data of the tunneling machine. Represents the observation noise covariance matrix. The nonlinear state transition function of the coupled dynamics model of the boom-bore-tunneling machine based on wellbore constraints is obtained by spatial and temporal discretization of the coupled dynamics model of the boom-bore-tunneling machine based on wellbore constraints; the calculation process of the adaptive process noise covariance matrix follows the following relationship: ； in, Represents the reference noise matrix. Indicates the safe distance threshold. Represents extremely small positive numbers. This indicates the real-time minimum distance between the boom and the well wall; The dynamic response data of the boom itself includes vibration acceleration and tilt angle. A sliding window FFT analysis is performed on the vibration acceleration to obtain the vibration spectrum of the boom. A pre-trained deep learning model is constructed, using multi-directional wellbore distance data and boom vibration spectrum as inputs, to predict the attitude correction amount; The estimated state vector is fused with the attitude correction value to obtain the current optimal state estimate. The corresponding process has the following relationship: ； in, Indicates adaptive fusion weights, This represents the correction amount output by the deep learning model; This represents the current optimal state estimate, i.e., the current attitude state of the tunneling machine.

5. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 4, characterized in that, In step 3, environmental risk characteristics are generated using multi-directional wellbore distance data and wellbore thermal imaging data. Based on the current attitude state of the tunneling machine and the environmental risk characteristics, a control strategy for the tunneling machine is generated using a model predictive control algorithm. Specifically, this includes the following steps: A coupled dynamic model of the boom-bore-tunneling machine constrained by the wellbore is adopted as the internal prediction model, and an objective function is constructed. The corresponding process has the following relationship: ； in, This represents the prediction of the system state at time t+k using the current optimal state estimate at the current time t. This indicates the sequence of future control strategies to be optimized, including the steering cylinder thrust, cutter head speed adjustment, and propulsion speed to be optimized. The reference trajectory represents the geometric state of the tunneling machine moving along the design axis; This represents the total number of future steps predicted by the MPC algorithm. Indicates the step index in the prediction time domain. Indicates the current moment. This represents the weighting coefficients of the tracking accuracy requirements for different components in the state vector within the optimization objective. The weighting coefficients represent the energy consumption of the correction cylinders in different directions in the optimization objective. This represents the weighting coefficient of environmental risk in the optimization objective. Indicates the use of Weighted square norm, Indicates the use of Weighted square norm, This indicates environmental risk penalties; among them, , This indicates the initial state of MPC; The calculation process for environmental risk penalties follows the following formula: ； in, This represents the normalized value of the seepage risk index; Normalized values ​​representing the characteristics of wellbore defects; This represents the actual environmental risk value calculated based on the measured real-time data at the current time t. This represents a given absolute safety distance threshold. These represent the influence factors of humidity and wellbore defects, respectively, and the weights of humidity and wellbore defects in the risk assessment are adjusted accordingly. The basic control quantity is obtained by iteratively optimizing the objective function. ; Interpolate the wellbore distance data from multiple directions to generate a wellbore distance distribution of 128 points in the circumferential direction, construct a accessibility map, and mark the areas on the accessibility map where the point distance is less than the preset distance as "high-risk areas"; Based on the accessibility map, the optimal travel direction is calculated, and the corresponding process follows the following formula: ； in, Indicates the optimal travel direction. Indicates a circumferential angle; This represents the weighting function, used to penalize directions that are closer to high-risk areas; This represents the continuous distance distribution function with respect to the circumferential angle θ, fitted using an interpolation algorithm. Based on the optimal travel direction and the current orientation angle of the tunneling machine, the obstacle avoidance feedforward term is calculated, and the corresponding process has the following relationship: ； in, This represents the feedforward compensation amount used for real-time obstacle avoidance. Indicates obstacle avoidance control gain, and Inversely proportional, the closer the distance, the stronger the obstacle avoidance intervention; Indicates the optimal travel direction. This indicates the current orientation angle of the tunneling machine; Basic control quantity By superimposing an obstacle avoidance feedforward term, the control strategy of the tunneling machine is obtained, and the corresponding process has the following relationship: ； in, This refers to the control strategy of the tunneling machine.

6. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 5, characterized in that, In step 4, based on the multi-directional wellbore distance data accumulated during the tunneling process, a wellbore geometric model is constructed and updated online, specifically including the following steps: The least squares method is used to fit the elliptical model of the well wall of the current tunneling section in real time for the multi-directional well wall distance data to obtain the major axis a, minor axis b and the offset of the ellipse center; and the ellipticity is calculated based on the major axis a, minor axis b and the offset of the ellipse center; local polynomial fitting is performed on the multi-directional well wall distance data to calculate the local curvature, and the ellipticity and local curvature constitute the geometric deformation characteristics of the well wall; After tunneling a set distance, the system automatically summarizes all multi-directional laser ranging data collected within the current distance, the synchronously calculated high-precision tunneling machine trajectory points and their corresponding tunneling distance L, as well as the well wall geometric deformation features corresponding to each trajectory point, forming a structured data package for model updates. PCA analysis is performed on the projection of all tunneling machine trajectory points in the current structured data package onto the horizontal plane to extract the direction of the first principal component; the direction of the first principal component is determined as the main trend of the current well section centerline on the horizontal plane; The vertical coordinates of the centerline are determined by linear correlation with the tunneling mileage L; Project all trajectory points onto the direction of the first principal component and calculate the average coordinates of the projected points as an estimated point of the current segment's mileage centerline; Set a sliding window with a fixed length and step size, perform cubic spline interpolation on the centerline estimation points of multiple consecutive mileages along the tunneling direction, and finally generate a globally continuous and smooth wellbore centerline function. For each tunneling machine trajectory point in the structured data packet, the normal distance from each sampling point on the well wall on the current cross section to the current centerline is calculated by combining the pose of the intelligent sensor node on the boom and its corresponding multi-directional laser ranging value when the trajectory point is collected. The median of all normal distances on the same cross section is taken as the instantaneous equivalent radius at the current tunneling mileage L; Collect all mileage-instantaneous equivalent radius data pairs in the current structured data package, use local weighted regression scatter smoothing method to fit the data, obtain the radius change trend of the current well section, and use the radius change trend of the current well section to update the global wellbore equivalent radius function. Simultaneously, all mileage-instantaneous ellipticity data pairs and mileage-instantaneous curvature data pairs within the current structured data package are collected, and the local weighted regression scatter smoothing method is used to fit them respectively to obtain the ellipticity change trend function and curvature change trend function of the current well section. A shaft geometric model is constructed based on the shaft centerline function, shaft equivalent radius function, ellipticity variation trend function, and curvature variation trend function. The shaft centerline function, shaft equivalent radius function, ellipticity variation trend function, and curvature variation trend function are updated by increasing the tunneling mileage to update the shaft geometric model.

7. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 6, characterized in that, In step 5, the wellbore geometric model is used as a dynamic spatial constraint for the future trajectory of the tunnel boring machine in the model predictive control algorithm to optimize the control strategy. This specifically includes the following steps: By using the updated wellbore centerline function and wellbore equivalent radius function, the possible shape of the well wall in the unexplored area ahead can be predicted; By using the predicted wellbore morphology of the unexplored area as a forward-looking spatial safety constraint for the objective function, and using the long-term optimal strategy parameters as a dynamic reference part of the objective function, the objective function is adjusted to obtain the adjusted objective function. The corresponding process has the following relationship: ； in, This represents the coordinates of the wellbore center point at the predicted future time t+k. This represents the equivalent radius of the wellbore corresponding to the predicted future time t+k. Indicates safety margin, This indicates a forward-looking environmental risk penalty item; Indicates from The location portion extracted is the predicted location of the tunneling machine; This represents the dynamic reference trajectory, which embodies the desired state that integrates a predetermined geometric path with the real-time optimization process objective. The calculation process for forward-looking environmental risk penalties follows the following formula: ； in, express and and The minimum distance between them This represents the ellipticity at future mileage predicted from the wellbore geometry model. This represents the curvature value predicted from the wellbore geometry model at future mileage. These represent the weights corresponding to the ellipticity and curvature values ​​at future mileages, respectively. The calculation process of the dynamic reference trajectory is governed by the following relationship: ； in, This represents the geometric state at a future moment, as determined by the design axis. Indicates the parameters of the long-term optimal strategy; Using the adjusted objective function, a basic control variable is generated, and then combined with the feedforward compensation variable to generate the final control strategy.

8. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 7, characterized in that, The long-term optimal policy parameters are obtained using reinforcement learning, and the specific steps are as follows: When time-frequency analysis of the minimum distance in the distance sequence identifies that the dominant swing frequency of the boom is greater than 0.5Hz and the swing amplitude exceeds the safety threshold, the reinforcement learning mode is automatically activated. In reinforcement learning mode, the system initiates a reinforcement learning process based on the proximal policy optimization algorithm, the framework of which is defined as follows: State space: including boom swing frequency, swing amplitude, well wall ellipticity, minimum distance between boom and well wall, current tunneling mileage, as well as tunneling machine advance speed, cutterhead speed and cutterhead torque; All state variables are standardized before input; Action space: defined as the continuous adjustment amount of two key tunneling parameters, namely: the percentage adjustment of the advance speed relative to its rated value, with an adjustment range of [-20%, +20%]; and the percentage adjustment of the cutterhead rotation speed relative to its rated value, with an adjustment range of [-15%, +15%]. The reward function is a weighted sum of the following four components: a swing amplitude penalty term that is negatively correlated with the swing amplitude, a stability penalty term that is negatively correlated with the absolute value of the advance speed adjustment, a stability penalty term that is negatively correlated with the absolute value of the cutterhead speed adjustment, and an efficiency reward term that is positively correlated with the tunneling footage per unit time. The reinforcement learning process is trained by interacting with a simulation environment constructed based on the coupling dynamics model of boom-bore-tunneling machine under wellbore constraints. The final output of the long-term optimal strategy parameters includes the recommended propulsion speed and the recommended cutterhead speed.

9. The tunneling machine attitude control method based on multi-source data fusion and dynamic calculation according to claim 8, characterized in that, At least two intelligent sensing nodes are fixedly installed at intervals along the axial direction of the main body of the boom; when there are two intelligent sensing nodes, the two intelligent sensing nodes are respectively installed at the bottom and top of the boom.

Images