Method and system for automatically correcting posture based on data driving

Abstract

The invention relates to the technical field of data-driven modeling and prediction, and discloses a data-driven attitude automatic correction method and system, which comprises the following steps: by constructing a digital twinborn model of a TBM (Tunnel Boring Machine) and a surrounding rock environment, realizing omnibearing and real-time monitoring of a tunneling process, providing high-precision basic data, and improving the accuracy of the TBM and the surrounding rock environment; and the visualization and prediction of actual tunneling are optimized. The actual position and operation data of the TBM are collected in real time through the attitude and position sensor, the data are synchronized to the digital twinborn model, the high consistency of the model and the actual progress is ensured, the real-time performance and accuracy of the data are greatly improved, and the actual deviation of the track is calculated by comparing the actual track with the design track; after deviation correction each time, required thrust is predicted by combining surrounding rock parameters and the state of the oil cylinder, and continuous optimization of the tunneling process is ensured. The automation degree is improved, the equipment loss is reduced, and the precision, efficiency and safety in the tunneling process are improved by dynamically adjusting the deviation rectifying track.

Description

Technical Field

[0001] The present invention relates to the technical field of data-driven modeling and prediction, and in particular to a method and system for automatic posture correction based on data-driven. Background Art

[0002] With the continuous development of tunnel excavation technology, tunnel boring machines (TBMs) have been widely used in underground projects as an efficient tunnel construction equipment. TBMs have achieved large-scale and efficient tunnel excavation through mechanization and automation, greatly improving construction efficiency and reducing manual operations. However, in actual operations, tunnel boring machines face complex underground environments, such as geological changes, equipment wear, and the accuracy of construction paths. These factors often cause the actual path of the tunnel boring machine to deviate from the designed path. In order to ensure that the tunnel boring machine advances according to the predetermined design trajectory, precise posture control has become a core technology in tunnel construction. In recent years, with the continuous advancement of digitalization and automation technology, the automated control system based on sensor data and digital twin technology has gradually become the main research direction for the posture adjustment and correction of tunnel boring machines. By real-time monitoring of the operating status of the tunnel boring machine and combining it with a data-driven prediction model, the operating posture of the tunnel boring machine can be effectively predicted and adjusted to improve work accuracy.

[0003] Although the existing technology has made some progress in the attitude control and correction of tunnel boring machines, there are still some technical difficulties. Traditional attitude correction methods often rely on simple feedback control algorithms, which are difficult to adapt to complex underground environments and changeable construction conditions, resulting in low efficiency and accuracy of attitude adjustment. In addition, the existing correction system usually uses rule-based algorithms for control, ignoring the possibility of in-depth mining of real-time data and intelligent optimization. Based on these problems, the existing technology has great limitations in the real-time and accuracy of path deviation prediction, pressure application and adjustment, and cannot achieve dynamic adjustment and intelligent optimization. Existing methods often fail to fully utilize the advantages of sensor data and digital twin models, and it is difficult to ensure the efficient operation of tunnel boring machines when dealing with path deviations in complex environments. At the same time, although digital twin technology has been applied in many fields, there is still a lack of mature systems that combine real-time data and prediction models in the precise control of tunnel boring machines. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the path deviation existing in the existing traditional tunnel boring machine attitude control system cannot be corrected in real time and accurately, especially under complex working conditions, manual adjustment and traditional correction methods are difficult to meet the requirements of high precision and high efficiency. By building a digital twin model and a data-driven prediction model, the automatic correction of the tunnel boring machine attitude is realized, overcoming the shortcomings of the traditional method that relies on manual judgment and manual intervention, improving the correction accuracy and automation level, and ensuring the stability and safety of tunnel boring operations.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a data-driven automatic posture correction method, comprising:

[0007] Construct a digital twin model of the TBM and surrounding rock environment;

[0008] The actual position and attitude data of the TBM are collected in real time through attitude and position sensors to determine the actual excavation trajectory of the TBM, and the actual position and attitude data of the TBM are synchronized to the digital twin model;

[0009] Compare the actual trajectory with the designed trajectory and calculate the real deviation of the trajectory;

[0010] When the actual deviation of the trajectory exceeds the correction threshold, a correction trajectory is planned and the expected correction parameters of the actual trajectory and the correction trajectory are calculated;

[0011] The desired correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder are used as input parameters, and the thrust required by the four driving cylinders is predicted through a data-driven model in the digital twin model;

[0012] Apply the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, measure the actual position and posture information of the TBM in the next cycle, determine the actual trajectory, and then calculate the correction parameters for the next cycle;

[0013] The correction parameters, surrounding rock parameters and cylinder thrust are assembled into historical time series data to enhance the training data-driven model.

[0014] As a preferred solution of the data-driven automatic posture correction method described in the present invention, the posture data includes posture data, position data, trajectory deviation, cylinder pressure data, vibration acceleration data, propulsion force data and surrounding rock parameters;

[0015] Remove noise and outliers from real-time operation data, and normalize real-time operation data of different dimensions;

[0016] Synchronizing the posture data to the digital twin model includes transmitting the real-time collected data from the TBM to the digital twin platform via MQTT;

[0017] The real-time operation data is time-synchronized by using a GPS clock, and a timestamp is added after each piece of real-time operation data;

[0018] When MQTT is delayed or interrupted, the cache mechanism ensures that data is not lost and synchronized when MQTT is restored;

[0019] When an abnormality occurs in the cache mechanism, data is backed up through redundant storage to ensure data integrity;

[0020] Synchronize real-time operation data to the digital twin model in real time and update the status of the digital twin model.

[0021] As a preferred solution of the data-driven automatic posture correction method described in the present invention, wherein: calculating the deviation between the correction trajectory and the actual trajectory includes discretizing the correction trajectory and the actual trajectory into data points according to the excavation progress, and calculating the deviation according to the data points;

[0022]

[0023]

[0024] in, It indicates the deviation between the actual trajectory and the designed trajectory. d indicates the deviation correction threshold. The threshold is 0.02 meters in the straight segment and 0.005*R in the turning segment. R indicates the radius of the turning segment. Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the design trajectory at the current time step, Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the design trajectory at the current time step; Indicates that the actual position of the TBM at the current time step is Coordinates on the axis; Indicates the Z-axis coordinate of the design trajectory at the current time step;

[0025] Assume that each excavation step is 0.01 meters long and the time step is 0.5 seconds;

[0026] The TBM position information needs to be updated after each excavation progress.

[0027] The design trajectory is the coordinates of discrete points of the ideal trajectory designed in the previous engineering planning;

[0028] The formula for calculating the expected deviation parameter is expressed as:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] in, represents the component of the expected error in the X direction, X represents the X-axis coordinate in three-dimensional space, Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the current correction track. Indicates the path deviation of the TBM in the Y-axis direction. Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the current correction track; Indicates the path deviation of the TBM in the Z-axis direction. Indicates that the actual position of the TBM at the current time step is The coordinates on the axis, Indicates the Z-axis coordinate of the current correction track; Indicates the actual angle of TBM and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between .

[0036] As a preferred solution of the data-driven automatic posture correction method described in the present invention, the data-driven model includes using the GRU model based on the real-time path deviation parameters to predict the pressure values ​​that need to be applied by the four cylinders. ;

[0037] According to the real-time TBM status and the prediction results of GRU, the Q-learning algorithm is used to select the optimal pressure adjustment scheme, and the result of the adjustment scheme is output as the corresponding actual pressure value of the cylinder.

[0038] As a preferred solution of the data-driven automatic posture correction method described in the present invention, the data-driven model also includes predicting the pressure values ​​that need to be applied by the four cylinders through the GRU model. ;

[0039] The input parameters include the expected correction parameters, the surrounding rock parameters at the current position and the current state of the cylinder;

[0040] The desired correction parameters include , , , , and ;

[0041] Integrate the input parameters at the current moment with the data of the historical time step to form a time series data matrix, and perform preprocessing. The formula is expressed as:

[0042]

[0043] Choose the past The data of time steps is taken as the input sequence, and the historical time series data matrix formula is expressed as:

[0044]

[0045] in, Indicates the current time The surrounding rock parameter vector, Indicates the current time Cylinder status parameters;

[0046] The preprocessing includes using Kalman filtering to remove noise in the time series data, identifying and processing outliers in the time series data, ensuring that all time series data are synchronized in time, and normalizing time series data of different dimensions;

[0047] The input time window size of the GRU model is the time series data of the past 10 time steps;

[0048] Divide the time series data in chronological order into 70% training set, 15% validation set, and 15% test set;

[0049] The hyperparameters of the GRU model include the number of GRU units, time step, learning rate, batch size, and number of training rounds;

[0050] The number of GRU units is 512 GRU layer hidden units;

[0051] The time step is 25 time steps, which represents the length of the entire input sequence;

[0052] The learning rate is 0.0005;

[0053] The batch size is 64 samples used in one training;

[0054] The number of training rounds is 100 rounds;

[0055] Adjust the hyperparameters of the GRU model through grid search;

[0056] The GRU model includes an input layer, a hidden layer, and an output layer;

[0057] The input layer receives the preprocessed time series data;

[0058] The hidden layer includes a plurality of GRU units stacked to extract time series features;

[0059] The formula of the hidden layer is expressed as:

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] in, Indicates the current time Update door of Represents the Sigmoid activation function; Represents the weight matrix input to the update gate, with dimension ,in is the dimension of the hidden state, is the dimension of the input; Indicates the current time The input data vector of Indicates the last moment The hidden state of For the weight matrix of the update gate, the dimension is ; Indicates the last moment The hidden state vector of Represents the bias term of the update gate, with dimension ; Represents the hyperparameter used to adjust the gain of the activation function; Indicates the current time The reset gate controls the reset of the current hidden state; Represents the weight matrix input to the reset gate, with dimension ; Indicates the last moment The hidden state of the reset gate weight matrix has the dimension ; The dimension of the bias term of the reset gate is ; Represents the attenuation factor in the integral function when controlling the reset gate; Indicates the current time The candidate hidden state of the current moment is generated; represents the hyperbolic tangent activation function; Represents the weight matrix input to the candidate hidden state, with dimension ; Indicates the last moment The weight matrix of the hidden state of the candidate hidden state is ; Reset gate Hide the state of the previous moment Weighted operation of Represents element-wise multiplication; Represents the bias term of the candidate hidden state, with dimension ; represents the hyperparameter that controls the weighting of historical information; Indicates the current time The final hidden state of Indicates the current time The final hidden state of Represents the historical time step index; Represents the number of historical time steps; Indicates the last moment The hidden state vector of represents the sum of time steps; represents the time step index, express The hidden state of the moment;

[0066] The output layer includes four output nodes, corresponding to The thrust prediction value is expressed as:

[0067] in, is the weight matrix; is the bias term, which represents the bias of each output; It is the thrust prediction value of the four cylinders, with a dimension of 4, indicating the thrust value of the four cylinders The thrust prediction value.

[0068] As a preferred solution of the data-driven automatic posture correction method described in the present invention, the data-driven model further includes defining the state space, action space and reward function of the Q-learning algorithm;

[0069] The state space includes the current path deviation, the current tunnel boring machine posture data and the current cylinder pressure state;

[0070] The action space includes different pressure adjustment schemes for four cylinders;

[0071] The reward function includes giving a positive reward when the path deviation decreases, and giving a negative reward when the path deviation increases;

[0072] When the cylinder applies pressure and the deviation decreases, a positive reward is given; when the cylinder applies pressure and the deviation increases, a negative reward is given;

[0073] Initialize the Q table to record the Q value of each state-action pair. The initial value of Q is set to 0.

[0074] According to the current state, an ε-greedy strategy is used to select actions to balance exploration and exploitation;

[0075] Apply the selected pressure adjustment scheme to the cylinder to adjust the attitude of the tunnel boring machine;

[0076] Calculate the reward value based on the new path deviation and cylinder pressure status, and update the Q table;

[0077] The formula for updating the Q table is expressed as:

[0078]

[0079] in, Represents the new state after executing the action. represents the best action in the new state, and represents the immediate reward of the current action;

[0080] The iterative process continues until the Q table converges or the preset number of training rounds of 100 is reached, then the training is stopped.

[0081] As a preferred solution of the data-driven automatic posture correction method described in the present invention, the data-driven model further includes optimizing the learning rate, discount factor and exploration rate in the Q-learning algorithm using a genetic algorithm;

[0082] The value ranges of learning rate α, discount factor γ and exploration rate ε are defined as (0,1) respectively;

[0083] The definition range of each parameter (α, γ, ε) is evenly divided into 10 intervals;

[0084] Randomly select a value from the interval of each parameter as a sample point;

[0085] Combine the sample points of each parameter to generate several groups Parameter combination to form the initial population;

[0086] The evaluation criteria of the fitness function include the total cumulative reward and the convergence speed;

[0087] The cumulative reward sum includes: the higher the cumulative reward obtained during training, the higher the fitness;

[0088] The convergence speed includes: The parameter combination makes the Q table converge faster and the fitness is higher;

[0089] Use each set of parameters Train the Q-learning model and record the cumulative rewards and number of convergence rounds during training;

[0090] The formula of the fitness function is expressed as:

[0091] in, represents the fitness function; Represents a combination of parameters, represents the cumulative reward weight, represents the standard deviation, represents the mean, CR(θ) is the cumulative reward; represents the convergence speed weight, represents the convergence speed function, CS) represents the maximum value of convergence speed, represents the time step;

[0092] The value range of F(θ) is from 0 to 1. The closer F(θ) is to 1, the higher the fitness of the parameter combination (α, γ, ε);

[0093] The roulette wheel selection method is used to select the parameter combination (α, γ, ε) with high fitness to enter the crossover and mutation operations;

[0094] Execute the single-point crossover method to generate new individuals;

[0095] Set the mutation probability to 5% to decide whether to mutate each parameter combination (α, γ, ε);

[0096] The selected parameter combination (α, γ, ε) is randomly fine-tuned to generate a new parameter combination;

[0097] Select the parameter combinations with the top 30% fitness from the newly generated parameter combinations and use the hill climbing algorithm to perform local search;

[0098] Local search involves, for each selected parameter combination, generating its fine-tuned parameter combination As a neighborhood solution, the formula is expressed as:

[0099]

[0100] in, represents the optimized parameter combination obtained by local search, Represents the neighborhood solution set of the current parameter combination, argmax represents Find the solution that maximizes the objective function in the neighborhood solution. represents the cumulative reward value of the solution, represents the convergence error of the solution;

[0101] Evaluate the fitness of the neighborhood solution and select the neighborhood solution with the highest fitness as the new solution; combine the better parameters after local search Replace the original parameter combination ;

[0102] Repeat selection, crossover, mutation, and local search to generate a new population; calculate each set of parameter combinations The fitness value of , and retain the individuals with the best performance;

[0103] When the population fitness no longer improves significantly, stop iterating.

[0104] Data-driven automatic posture correction system, including:

[0105] Digital twin model module, which builds a digital twin model of the TBM and the surrounding rock environment;

[0106] A synchronization module collects the actual position and attitude data of the TBM in real time through attitude and position sensors, determines the actual excavation trajectory of the TBM, and synchronizes the actual position and attitude data of the TBM to the digital twin model;

[0107] The comparison module compares the actual trajectory with the designed trajectory and calculates the actual deviation of the trajectory;

[0108] The planning module plans the correction trajectory and calculates the expected correction parameters between the actual trajectory and the correction trajectory when the actual trajectory deviation exceeds the correction threshold.

[0109] A prediction module, which uses the desired correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder as input parameters, and predicts the thrust required by the four driving cylinders through a data-driven model in the digital twin model;

[0110] The calculation module applies the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, and measures the actual position and posture information of the TBM in the next cycle, determines the actual trajectory, and then calculates the correction parameters for the next cycle;

[0111] The enhancement module assembles the correction parameters, surrounding rock parameters and cylinder thrust into historical time series data for enhancing the training data-driven model.

[0112] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0113] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the methods of the present invention.

[0114] The beneficial effects of the present invention are as follows: by constructing a digital twin model of the TBM and the surrounding rock environment, all-round and real-time monitoring of the tunnel excavation process is achieved, high-precision basic data is provided, and the visualization and prediction of the actual excavation are optimized. The actual position and operation data of the TBM are collected in real time by using attitude and position sensors, and these data are synchronized to the digital twin model, ensuring the high consistency between the model and the actual progress, and greatly improving the real-time and accuracy of the data. This synchronization process provides a reliable basis for subsequent trajectory comparison and correction, making it possible to discover and correct trajectory deviations in real time. By comparing the actual trajectory with the designed trajectory, calculating the actual deviation e of the trajectory, and planning the correction trajectory when the deviation exceeds the threshold, the accuracy and timeliness of the correction are further improved. After each deviation correction, the required thrust is predicted in combination with the surrounding rock parameters and the cylinder state, ensuring the continuous optimization of the excavation process. This data-driven closed-loop control mechanism can significantly reduce manual intervention, improve the degree of automation, avoid error accumulation, reduce equipment loss, and improve the accuracy, efficiency and safety of the tunnel excavation process by dynamically adjusting the correction trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0116] Figure 1 An overall flow chart of the data-driven automatic posture correction method provided for the first embodiment of the present invention. DETAILED DESCRIPTION

[0117] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0118] Example 1, reference Figure 1 , is an embodiment of the present invention, and provides a data-driven automatic posture correction method, comprising:

[0119] S1: Construct a digital twin model of the TBM and the surrounding rock environment.

[0120] Collect detailed hardware data of the tunnel boring machine and use CAD to build the geometric model, kinematic model and dynamic model of the TBM.

[0121] The geometric model describes the appearance of the tunnel boring machine and the spatial position relationship of each component. It includes the overall size of the tunnel boring machine, the shape of each component, the kinematic connection and constraint conditions; the geometric characteristics of the tunnel boring machine's cutterhead, support structure, cylinder, hydraulic system, etc., as well as their relative positions and coordination methods.

[0122] The kinematic model describes the relative motion of the various components of the tunnel boring machine, especially the motion trajectory, speed, acceleration, etc. of each cylinder and other moving parts. The model takes into account the motion relationship and coordination of the various components of the tunnel boring machine during operation, and provides the motion state, angle change and constraint conditions of each component during operation.

[0123] The dynamic model describes the mechanical effects and responses of the tunnel boring machine during operation, mainly involving the dynamic behaviors of the tunnel boring machine such as thrust, torque, acceleration, vibration, etc. It includes the external forces, friction, hydraulic pressure on each component during operation, and their effects on the overall motion and posture of the machine. The model can simulate the dynamic response of the tunnel boring machine and provide a theoretical basis for subsequent posture correction control.

[0124] Establish the behavior model of TBM in different working states (such as moving forward, turning, ups and downs, etc.).

[0125] Sensors are installed at key parts of the tunnel boring machine (such as cylinders, power systems, attitude sensors, accelerometers, gyroscopes, etc.) to collect TBM posture data, path deviation data, cylinder pressure data, vibration acceleration data, propulsion force data, and soil and tunnel environment data in real time.

[0126] Furthermore, by establishing a digital twin model of the tunnel boring machine (TBM), combined with a precise description of its geometry, kinematics and dynamics models, a reliable physical and behavioral basis is provided for subsequent posture control and path correction. By installing sensors at key locations of the TBM and collecting various operating data (such as posture, path deviation, cylinder pressure, vibration acceleration, propulsion force, etc.) in real time, dynamic monitoring and precise analysis of the machine status can be achieved. These data provide the necessary input for the subsequent real-time posture correction of the TBM through data-driven control algorithms, thereby achieving more efficient and precise tunnel excavation operations and optimizing its working performance and stability.

[0127] S2: The actual position and posture data of the TBM are collected in real time through the posture and position sensors to determine the actual excavation trajectory of the TBM, and the actual position and posture data of the TBM are synchronized to the digital twin model.

[0128] The posture data includes position data, path deviation data, cylinder pressure data, vibration acceleration data, propulsion force data and soil and tunnel environment data.

[0129] Remove noise and outliers from real-time operation data, and normalize real-time operation data of different dimensions.

[0130] Synchronizing the posture data to the digital twin model includes transmitting the real-time collected data from the tunnel boring machine to the digital twin platform via MQTT.

[0131] The real-time operation data is time-synchronized by a GPS clock, and a timestamp is added after each piece of real-time operation data.

[0132] When MQTT is delayed or interrupted, a cache mechanism is used to ensure that data is not lost and synchronized when MQTT is restored.

[0133] When an exception occurs in the cache mechanism, data is backed up through redundant storage to ensure data integrity.

[0134] Synchronize real-time operation data to the digital twin model in real time and update the status of the digital twin model.

[0135] Furthermore, by collecting various operating data of the tunnel boring machine (such as posture data, path deviation data, cylinder pressure data, etc.) in real time and synchronizing these data to the digital twin model, it is ensured that the model can reflect the actual working status of the tunnel boring machine in real time. The accuracy and consistency of the data are guaranteed by removing noise and outliers and normalizing the real-time data. The MQTT protocol is used for data transmission and time synchronization to ensure the efficiency and timeliness of data transmission. In addition, the stability and integrity of the data transmission process are guaranteed through the cache mechanism and redundant storage technology, and data loss caused by delays or interruptions is prevented, thereby ensuring the accuracy and real-time update of the digital twin model. This process ensures the integrity, real-time and efficiency of the data, and provides a reliable data foundation for subsequent posture correction and control strategy optimization.

[0136] S3: Compare the actual path with the preset design path and calculate the path deviation.

[0137] The correction trajectory and actual trajectory are discretized into data points according to the excavation progress.

[0138]

[0139]

[0140] in, It indicates the deviation between the actual trajectory and the designed trajectory. d indicates the deviation correction threshold. The threshold is 0.02 meters in the straight segment and 0.005*R in the turning segment. R indicates the radius of the turning segment. Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the design trajectory at the current time step, Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the design trajectory at the current time step; Indicates that the actual position of the TBM at the current time step is Coordinates on the axis. Indicates the Z-axis coordinate of the design trajectory at the current time step.

[0141] Assume that each excavation step is 0.01 meters long and the time step is 0.5 seconds.

[0142] The position information of the TBM needs to be updated every time a certain excavation progress is completed.

[0143] Furthermore, by comparing the actual path of the tunnel boring machine with the preset design path, the path deviation is accurately calculated, thereby achieving precise positioning and posture adjustment of the tunnel boring machine. By discretizing the path into a series of points and updating the position of the tunnel boring machine in each time step, the position changes of the tunnel boring machine in real-time operation can be carefully tracked and compared with the preset path. The calculated path deviation (including deviations in the X, Y, and Z directions) provides a quantitative basis for subsequent posture correction. This step ensures that the system can monitor and correct the posture of the tunnel boring machine in real time, avoid the decrease in construction accuracy due to path deviation, and thus improve the safety and efficiency of the tunneling operation.

[0144] S4: When the actual deviation of the trajectory exceeds the correction threshold, a correction trajectory is planned and expected correction parameters between the actual trajectory and the correction trajectory are calculated.

[0145] The designed trajectory is the coordinates of discrete points of the ideal trajectory designed in the early stage of engineering planning.

[0146] The formula for calculating the expected deviation is expressed as:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153] in, represents the component of the expected error in the X direction, X represents the X-axis coordinate in three-dimensional space, Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the current correction track. Indicates the path deviation of the TBM in the Y-axis direction. Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the current correction track; Indicates the path deviation of the TBM in the Z-axis direction. Indicates that the actual position of the TBM at the current time step is The coordinates on the axis, Indicates the Z-axis coordinate of the current correction track; Indicates the actual angle of TBM and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between .

[0154] Furthermore, when the deviation between the actual trajectory of the tunnel boring machine (TBM) and the preset design trajectory exceeds the set correction threshold, the expected correction parameters between the actual trajectory and the correction trajectory are calculated to plan the adjustment path. Specifically, this process includes calculating the path deviation in the X, Y, and Z directions and the angular deviation between the actual posture and the expected posture to ensure that the tunnel boring machine can be adjusted according to the predetermined trajectory. Through the accurate calculation of these deviations, a clear basis can be provided for the application of cylinder thrust, thereby achieving accurate trajectory correction, ensuring that the TBM is smoothly advanced along the ideal trajectory, and avoiding engineering risks and time delays caused by deviations from the design trajectory.

[0155] S5: Using the desired correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder as input parameters, the thrust required for the four driving cylinders is predicted through a data-driven model in the digital twin model.

[0156] The data-driven model includes using the GRU model to predict the pressure values ​​that need to be applied to the four cylinders based on the real-time path deviation. .

[0157] According to the real-time TBM status and the prediction results of GRU, the Q-learning algorithm is used to select the optimal pressure adjustment scheme, and the result of the adjustment scheme is output as the corresponding actual pressure value of the cylinder.

[0158] Use the GRU model to predict the pressure values ​​that need to be applied to the four cylinders .

[0159] The input parameters include the expected correction parameters, the surrounding rock parameters at the current position and the current state of the cylinder.

[0160] The desired correction parameters include , , , , and .

[0161] The input parameters at the current moment are integrated with the data of historical time steps to form time series data, and then preprocessed.

[0162] The preprocessing includes using Kalman filtering to remove noise in the time series data, identifying and processing outliers in the time series data, ensuring that all time series data are synchronized in time, and normalizing time series data of different dimensions.

[0163] Using Kalman filtering to remove noise from time series data includes building a state space model for TBM operation, including position, attitude, and cylinder state variables. Kalman filtering first requires building a state space model for time series data, including the state transition equation: describing how the system transitions from the previous state to the current state within a time step. The formula is expressed as:

[0164]

[0165] in, is the system status at the current moment, is the state transition matrix, is process noise, which indicates the uncertainty of the system during the transition process. Observation equation: describes how to generate observations from the current system state.

[0166] The observation equation formula is expressed as:

[0167]

[0168] in, is the observed value, is the observation matrix, is the observation noise.

[0169] Initial state The first observation of the time series data is often used as an initial estimate.

[0170] Initial error covariance matrix : Indicates the uncertainty of the initial estimate and is usually set to a larger value.

[0171] At each time step, the Kalman filter predicts the state at the current moment by using the state transition equation and updates the estimated covariance matrix.

[0172] Predict current status:

[0173]

[0174] Forecast error covariance matrix:

[0175]

[0176] in, is the process noise covariance matrix. Update step After receiving new observation data, the Kalman filter updates the predicted state estimate through the observation equation and calculates the new error covariance matrix and the Kalman gain , which represents the weight of the current moment to correct the predicted value:

[0177]

[0178] in, is the observation noise covariance matrix.

[0179] Update the estimated status:

[0180]

[0181] in, is the residual between the observed and predicted values. Update the error covariance matrix:

[0182]

[0183] in, is the identity matrix.

[0184] The Kalman filter repeats the prediction and update steps recursively. At each time step, the algorithm optimizes the state estimate based on the previous estimate and the new observation data, and continuously reduces the impact of noise on the data.

[0185] The input time window size of the GRU model is the time series data of the past 10 time steps.

[0186] The time series data is divided into 70% training set, 15% validation set and 15% test set in chronological order.

[0187] The hyperparameters of the GRU model include the number of GRU units, time step, learning rate, batch size, and number of training rounds.

[0188] The number of GRU units is 512 GRU layer hidden units.

[0189] The time step is 25 time steps, representing the length of the entire input sequence.

[0190] The learning rate is 0.0005.

[0191] The batch size is 64 samples used in one training.

[0192] The number of training rounds is 100 rounds.

[0193] The hyperparameters of the GRU model are adjusted by grid search.

[0194] List all hyperparameters to be tuned and their candidate values.

[0195] Learning rate: {0.0001, 0.0005, 0.001}

[0196] Batch size: {32,64,128}

[0197] Number of hidden layer units: {128,256,512}

[0198] Number of training rounds: {50,100,150}

[0199] Generate all possible hyperparameter combinations based on the set candidate values. For the candidate values ​​given above, grid search generates the following combinations:

[0200] (0.0001,32,128,50)

[0201] (0.0001,32,128,100)

[0202] (0.0001,32,128,150)

[0203] (0.0001,64,128,50)

[0204] (0.0001,64,128,100)

[0205] (0.0001,64,128,150)

[0206] (0.0001,128,128,50)

[0207] (0.0001,128,128,100)

[0208] (0.0001,128,128,150) ...

[0210] (all other possible combinations)

[0211] For each set of hyperparameter combinations, the model is trained and the performance of the model is evaluated on the validation set. The mean square error is used as the evaluation metric to select the current hyperparameter combination.

[0212] Train the GRU model on the training set.

[0213] Evaluate the performance of the trained model on the validation set.

[0214] Record the evaluation results. After training and evaluating all hyperparameter combinations, select the hyperparameter combination that performs best on the validation set, that is, select those hyperparameter combinations that can minimize the loss function, maximize the accuracy, or other objectives.

[0215] Using the selected optimal hyperparameter combination, the GRU model is retrained on the entire training set. After training, the model is used to evaluate its generalization ability on the test set to confirm its performance.

[0216] Analyze the evaluation results for all hyperparameter combinations and determine which parameters have the greatest impact on model performance.

[0217] The GRU model consists of an input layer, a hidden layer, and an output layer.

[0218] The input layer receives the preprocessed time series data.

[0219] The hidden layer includes a plurality of GRU units stacked to extract time series features.

[0220] The formula of the hidden layer is expressed as:

[0221]

[0222]

[0223]

[0224]

[0225]

[0226] in, Indicates the current time Update door of Represents the Sigmoid activation function; Represents the weight matrix input to the update gate, with dimension ,in is the dimension of the hidden state, is the dimension of the input; Indicates the current time The input data vector of Indicates the last moment The hidden state of For the weight matrix of the update gate, the dimension is ; Indicates the last moment The hidden state vector of Represents the bias term of the update gate, with dimension ; Represents the hyperparameter used to adjust the gain of the activation function; Indicates the current time The reset gate controls the reset of the current hidden state; Represents the weight matrix input to the reset gate, with dimension ; Indicates the last moment The hidden state of the reset gate weight matrix has the dimension ; The dimension of the bias term of the reset gate is ; Represents the attenuation factor in the integral function when controlling the reset gate; Indicates the current time The candidate hidden state of the current moment is generated; represents the hyperbolic tangent activation function; Represents the weight matrix input to the candidate hidden state, with dimension ; Indicates the last moment The weight matrix of the hidden state of the candidate hidden state is ; Reset gate Hide the state of the previous moment Weighted operation of Represents element-wise multiplication; Represents the bias term of the candidate hidden state, with dimension ; represents the hyperparameter that controls the weighting of historical information; Indicates the current time The final hidden state of Represents the historical time step index; Represents the number of historical time steps; Indicates the last moment The hidden state vector of represents the sum of time steps; represents the time step index, express The hidden state of the moment;

[0227] The output layer includes four output nodes, corresponding to The thrust prediction value is expressed as:

[0228] in, is the weight matrix; is the bias term, which represents the bias of each output; It is the thrust prediction value of the four cylinders, with a dimension of 4, indicating the thrust value of the four cylinders The thrust prediction value.

[0229] Define the state space, action space, and reward function of the Q-learning algorithm.

[0230] The state space includes the current path deviation, the current posture data of the tunnel boring machine and the current cylinder pressure state.

[0231] The action space includes different pressure adjustment schemes for four cylinders.

[0232] The reward function includes giving a positive reward when the path deviation amount decreases, and giving a negative reward when the path deviation amount increases.

[0233] When the cylinder applies pressure and the deviation decreases, a positive reward is given; when the cylinder applies pressure and the deviation increases, a negative reward is given.

[0234] Initialize the Q table and record the Q value of each state-action pair. The initial value of Q is set to 0. According to the current state, the ε-greedy strategy is used to select actions to balance exploration and exploitation.

[0235] Apply the selected pressure adjustment scheme to the cylinder to adjust the TBM posture.

[0236] The reward value is calculated based on the new path deviation and cylinder pressure status, and the Q table is updated.

[0237] The formula for updating the Q table is expressed as:

[0238]

[0239] in, Represents the new state after executing the action. represents the best action in the new state, Indicates the immediate reward of the current action;

[0240] The iterative process continues until the Q table converges or the preset number of training rounds of 100 is reached, then the training is stopped.

[0241] The data-driven model further includes optimizing the learning rate, discount factor and exploration rate in the Q-learning algorithm using a genetic algorithm.

[0242] The value ranges of the learning rate α, discount factor γ and exploration rate ε are defined as (0,1) respectively.

[0243] The definition range of each parameter (α, γ, ε) is evenly divided into 10 intervals.

[0244] A value is randomly selected from the interval of each parameter as a sample point.

[0245] Combine the sample points of each parameter to generate several groups Parameter combination to form the initial population.

[0246] The evaluation criteria of the fitness function include the cumulative reward sum and the convergence speed.

[0247] The cumulative reward sum includes that the higher the cumulative reward obtained during the training process, the higher the fitness.

[0248] The convergence speed includes: The parameter combination makes the Q table converge faster and the fitness is higher.

[0249] Use each set of parameters Train the Q-learning model and record the cumulative rewards and number of convergence rounds during training.

[0250] The formula of the fitness function is expressed as:

[0251] in, represents the fitness function; Represents a combination of parameters, represents the cumulative reward weight, represents the standard deviation, represents the mean; CR(θ) is the cumulative reward; represents the convergence speed weight, represents the convergence speed function, CS) represents the maximum value of convergence speed, Represents the time step. The value range is from 0 to 1. The closer F(θ) is to 1, the higher the fitness of the parameter combination (α, γ, ε).

[0252] The roulette wheel selection method is used to select the parameter combination (α, γ, ε) with high fitness to enter the crossover and mutation operations.

[0253] Calculate the sum of fitness of all candidate parameter combinations . Suppose there are individuals, and the fitness of each individual is ,but .

[0254] For each individual , calculate the probability of being selected :

[0255]

[0256] Individuals with higher fitness will have a higher probability of being selected.

[0257] Construct a cumulative probability distribution based on the selection probability of each individual ,in:

[0258] No. The cumulative probability of an individual is:

[0259]

[0260] in, It represents the sum of fitness of all individuals in the population. Indicates The probability of an individual being selected. Indicates The cumulative probability of an individual. Represents a random number between 0 and 1, used to decide which individual to select. Indicates The fitness value of an individual. Represents the number of individuals in the population.

[0261] Generate a random number between 0 and 1 . Find in the cumulative probability distribution that Individual , and use it as the selected individual. Perform the above steps until the required number of individuals are selected.

[0262] Execute the single-point crossover method to generate new individuals.

[0263] According to the roulette wheel selection method, two individuals with higher fitness are selected as parents. and .

[0264] A random position in the dimension of the parameter vector is chosen as the crossover point. For example, suppose the parameter combination; change the parameters before the crossover point from Copy to child 1, and change the parameters after the intersection from Copy to child 1. Do the reverse operation on the other child 2:

[0265] Offspring 1: ; Offspring 2:

[0266] Through the above operation, two new individuals are obtained as offspring.

[0267] According to the mutation probability, mutation operations are performed on some individuals to randomly change the genes of the individuals.

[0268] The top 30% individuals in terms of fitness are selected from the newly generated individuals for local search.

[0269] Iterate optimization through operations such as selection, crossover, mutation, and local search until the fitness no longer increases significantly.

[0270] The mutation probability is set to 5% to decide whether to mutate each parameter combination (α, γ, ε).

[0271] A set of selected parameter combinations (α, γ, ε) is randomly fine-tuned to generate a new parameter combination.

[0272] The parameter combinations with the top 30% fitness ranking are selected from the newly generated parameter combinations, and the hill climbing algorithm is used for local search.

[0273] Local search involves, for each selected parameter combination, generating its fine-tuned parameter combination As a neighborhood solution, the formula is expressed as:

[0274]

[0275] in, represents the optimized parameter combination obtained by local search, Represents the neighborhood solution set of the current parameter combination, argmax represents Find the solution that maximizes the objective function in the neighborhood solution. represents the cumulative reward value of the solution, represents the convergence error of the solution.

[0276] Evaluate the fitness of the neighborhood solution and select the neighborhood solution with the highest fitness as the new solution; combine the better parameters after local search Replace the original parameter combination .

[0277] Repeat selection, crossover, mutation, and local search to generate a new population; calculate each set of parameter combinations The fitness value of each individual is determined and the individuals with the best performance are retained.

[0278] When the population fitness no longer improves significantly, stop iterating.

[0279] Furthermore, it ensures that the genetic algorithm achieves a balance between efficiency and effectiveness during the optimization process. When the fitness of the population no longer increases significantly, it means that the current parameter combination is close to the optimal solution, and further iteration may not bring substantial performance improvements, and may even lead to a waste of resources. Therefore, by setting the termination condition that the fitness no longer increases significantly, the algorithm can be effectively prevented from over-running, computing resources can be saved, and overfitting problems caused by over-optimization can be avoided. In addition, this step ensures that the algorithm converges to a high-quality parameter combination within a reasonable time, thereby improving the prediction accuracy of the data-driven model and the overall control performance of the system. This design not only improves the efficiency of the optimization process, but also enhances the robustness and reliability of the model in practical applications, ensuring that the tunnel boring machine can move stably and accurately along the designed trajectory.

[0280] S6: Apply the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, measure the actual position and posture information of the TBM in the next cycle, determine the actual trajectory, and then calculate the correction parameters for the next cycle.

[0281] S7: Assemble the correction parameters, surrounding rock parameters and cylinder thrust into historical time series data to enhance the training data-driven model.

[0282] Furthermore, the predicted thrust is applied to the corresponding numbered cylinder to carry out the next cycle of excavation, while the actual position and posture information of the tunnel boring machine (TBM) is measured in real time to obtain the actual trajectory. By comparing the actual trajectory with the preset design trajectory, the new trajectory deviation is calculated to obtain the correction parameters for the next cycle. These correction parameters, surrounding rock parameters, and cylinder thrust are compiled into historical time series data to enhance the performance of the training data-driven model and further optimize the prediction capability. The core of this step is to accurately correct the future thrust prediction based on real-time data by continuously updating and adjusting the model, so that the TBM can efficiently and accurately adjust its posture and trajectory in a complex environment, and achieve more precise automated control.

[0283] Embodiment 2 is an embodiment of the present invention, which provides a data-driven automatic posture correction system, including:

[0284] Digital twin model module, builds a digital twin model of TBM and surrounding rock environment.

[0285] The synchronization module collects the actual position and posture data of the TBM in real time through the posture and position sensors, determines the actual excavation trajectory of the TBM, and synchronizes the actual position and posture data of the TBM to the digital twin model.

[0286] The comparison module compares the actual trajectory with the designed trajectory and calculates the real deviation of the trajectory.

[0287] The planning module plans the correction trajectory and calculates the expected correction parameters between the actual trajectory and the correction trajectory when the actual trajectory deviation exceeds the correction threshold.

[0288] The prediction module uses the expected correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder as input parameters, and predicts the thrust required for the four driving cylinders through a data-driven model in the digital twin model.

[0289] The calculation module applies the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, and measures the actual position and posture information of the TBM in the next cycle, determines the actual trajectory, and then calculates the correction parameters for the next cycle.

[0290] The enhancement module assembles the correction parameters, surrounding rock parameters and cylinder thrust into historical time series data for enhancing the training data-driven model.

[0291] Embodiment 3, an embodiment of the present invention, is different from the first two embodiments in that:

[0292] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0293] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0294] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0295] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0296] Example 4 is an embodiment of the present invention, which provides a data-driven automatic posture correction method and system. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0297] A TBM equipped with high-precision attitude and position sensors was selected, and attitude sensors, position sensors, cylinder pressure sensors, vibration acceleration sensors and propulsion force sensors were installed at key locations to collect attitude data of the TBM in real time. These data are transmitted to the digital twin platform in real time via the MQTT protocol, and time is synchronized via the GPS clock to ensure that each piece of data is accompanied by an accurate timestamp.

[0298] During the test, the TBM began to excavate according to the preset design trajectory, while collecting and transmitting data such as its position, posture, cylinder pressure, vibration acceleration, propulsion force and surrounding rock parameters in real time. By synchronizing and processing real-time data through the digital twin model, the system can instantly determine the actual excavation trajectory of the TBM, compare it with the designed trajectory, and calculate the actual deviation of the trajectory. When the trajectory deviation exceeds the set correction threshold (the straight segment deviation threshold is 0.02 meters, and the turning segment deviation threshold is 0.005 multiplied by the turning radius R), the system automatically starts the correction algorithm. The correction algorithm first plans the correction trajectory and calculates the expected correction parameters between the actual trajectory and the correction trajectory. Then, based on the expected correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder, the data-driven model predicts the thrust required to be applied by the four driving cylinders through the GRU model. The predicted thrust value is then applied to the corresponding numbered cylinder to adjust the TBM posture and proceed to the next cycle of excavation. During this process, the system continuously measures the actual position and posture information of the TBM in the next cycle, determines the new actual trajectory, and recalculates the correction parameters for the next cycle. All correction parameters, surrounding rock parameters and cylinder thrust are assembled into historical time series data, further enhancing the prediction accuracy of the training data-driven model. To ensure data integrity, when MQTT transmission is delayed or interrupted, the system temporarily stores data through a cache mechanism and synchronizes after transmission is restored; if the cache mechanism is abnormal, data is backed up through redundant storage to ensure that data is not lost.

[0299] During the implementation process, the GRU model uses the time series data of historical time steps, combined with the current trajectory deviation and surrounding rock parameters, to accurately predict the thrust required by the cylinder. The Q-learning algorithm selects the optimal pressure adjustment scheme based on the real-time status and the prediction results of the GRU to achieve dynamic and intelligent trajectory correction. The hyperparameters of Q-learning (learning rate, discount factor, exploration rate) are optimized through genetic algorithms to further improve the convergence speed and prediction accuracy of the algorithm. The entire process forms a closed-loop control system to ensure that the TBM can efficiently and accurately excavate along the designed trajectory in a complex surrounding rock environment, significantly reduce trajectory deviation, and improve excavation efficiency and safety.

[0300] In this experiment, 100 excavation cycle tests were conducted, with each cycle step length of 0.01 meters and a time step length of 0.5 seconds. The following are actual data records of some key time steps:

[0301] In the initial stage (steps 1-10), the TBM excavated according to the designed trajectory, with a small trajectory deviation and a maximum deviation of 0.01 meters. As the excavation progressed (step 20), due to the uneven distribution of surrounding rock strength, the TBM had a slight trajectory deviation, with a deviation of 0.015 meters in the X direction, -0.012 meters in the Y direction, and 0.018 meters in the Z direction. When entering step 30, the deviation increased further, reaching 0.025 meters in the X direction, exceeding the correction threshold of 0.02 meters for the straight segment, and the system automatically started the correction algorithm. The GRU model predicted the cylinder thrust values ​​of 16.50MPa, 15.80MPa, 14.90MPa, and 16.20MPa, respectively. After the thrust was applied, the trajectory deviation of the TBM was significantly reduced in step 40, with the deviation in the X direction reduced to 0.005 meters, the deviation in the Y direction reduced to -0.004 meters, and the deviation in the Z direction reduced to 0.007 meters. After 50 steps of excavation, the trajectory deviation was further stabilized, with the deviation in the X direction maintained at 0.003 meters, the deviation in the Y direction at -0.002 meters, and the deviation in the Z direction at 0.004 meters. Throughout the excavation process, the GRU model continuously learned historical data and real-time feedback, gradually optimized the thrust prediction, and achieved accurate trajectory correction.

[0302] Through the analysis of experimental data, it can be clearly seen that the data-driven automatic posture correction method has significant advantages in the TBM excavation process. First, in the initial stage before the correction algorithm is started, the trajectory deviation of the TBM gradually increases with the increase in excavation progress, which shows that the traditional manual adjustment method is difficult to keep the TBM stable along the designed trajectory when facing a complex surrounding rock environment. However, once the trajectory deviation exceeds the preset threshold, the correction algorithm automatically started by the system responds quickly. Through the thrust prediction of the GRU model and the strategy optimization of Q-learning, the thrust adjustment of the four cylinders significantly reduces the trajectory deviation. In this process, the GRU model, with its deep learning ability of time series data, can accurately predict the thrust value required for each cylinder, thereby achieving precise posture adjustment.

[0303] Compared with traditional methods, the automatic posture correction algorithm of the present invention demonstrates its innovation and superiority in many aspects. Traditional methods usually rely on the operator's experience or preset static rules, and it is difficult to cope with real-time changing surrounding rock conditions and dynamic trajectory deviations. The present invention realizes dynamic and real-time trajectory correction through real-time data synchronization of the digital twin model, precise thrust prediction of the GRU model, and intelligent decision-making of the Q-learning algorithm. Experimental data show that after implementing the correction algorithm of the present invention, the trajectory deviation of the TBM is significantly reduced in each excavation cycle step, especially in the key correction step, the trajectory deviation is reduced to below 0.005 meters, which fully proves the efficiency and accuracy of the algorithm.

[0304] In addition, the Q-learning hyperparameters optimized by genetic algorithms make the correction algorithm have faster convergence speed and higher adaptability during the training process. The genetic algorithm improves the performance of the Q-learning algorithm in different excavation environments by optimizing the learning rate, discount factor and exploration rate, making the adjustment of the cylinder thrust more reasonable and efficient. In the experiment, Q-learning optimized by genetic algorithms showed higher cumulative rewards and faster convergence speed after 100 rounds of training, further verifying the application potential and technical advancement of the present invention in complex excavation environments.

[0305] In summary, this embodiment fully demonstrates the innovation and advantages of the data-driven automatic posture correction method in the TBM excavation process through detailed data recording and analysis. The algorithm can not only adjust the posture and trajectory of the TBM in real time and accurately, reducing the deviation during the excavation process, but also improve the excavation efficiency and safety through intelligent thrust prediction and strategy optimization, overcome the shortcomings of traditional methods in dynamic environments, and reflect significant technological progress and application value.

[0306] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A data-driven automatic posture correction method, characterized in that: include: Construct a digital twin model of the TBM and surrounding rock environment; The actual position and attitude data of the TBM are collected in real time through attitude and position sensors to determine the actual excavation trajectory of the TBM, and the actual position and attitude data of the TBM are synchronized to the digital twin model; Compare the actual trajectory with the designed trajectory and calculate the real deviation of the trajectory; When the actual deviation of the trajectory exceeds the correction threshold, a correction trajectory is planned and the expected correction parameters of the actual trajectory and the correction trajectory are calculated; The desired correction parameters, the surrounding rock parameters at the current position, and the current pressure of the oil cylinder are used as input parameters, and the thrust required by the four driving oil cylinders at the next time step is predicted through the data-driven model in the digital twin model; Apply the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, measure the actual position and posture information of the TBM in the next cycle, determine the actual trajectory, and then calculate the correction parameters for the next cycle; The correction parameters, surrounding rock parameters and cylinder thrust are assembled into historical time series data to enhance the training data-driven model.

2. The data-driven automatic posture correction method according to claim 1, characterized in that: The posture data includes posture data, position data, trajectory deviation, cylinder pressure data, vibration acceleration data, propulsion force data and surrounding rock parameters; Remove noise and outliers from real-time operation data, and normalize real-time operation data of different dimensions; Synchronizing the posture data to the digital twin model includes transmitting the real-time collected data from the TBM to the digital twin platform via MQTT; The real-time operation data is time-synchronized by using a GPS clock, and a timestamp is added after each piece of real-time operation data; When MQTT is delayed or interrupted, the cache mechanism ensures that data is not lost and synchronized when MQTT is restored; When an abnormality occurs in the cache mechanism, data is backed up through redundant storage to ensure data integrity; Synchronize real-time operation data to the digital twin model in real time and update the status of the digital twin model.

3. The data-driven automatic posture correction method according to claim 1, characterized in that: Calculating the deviation between the deviation-correcting trajectory and the actual trajectory includes discretizing the deviation-correcting trajectory and the actual trajectory into data points according to the excavation progress, and calculating the deviation according to the data points; ； ； in, It indicates the deviation between the actual trajectory and the designed trajectory. d indicates the deviation correction threshold. The threshold is 0.02 meters in the straight segment and 0.005*R in the turning segment. R indicates the radius of the turning segment. Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the design trajectory at the current time step, Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the design trajectory at the current time step; Indicates that the actual position of the TBM at the current time step is Coordinates on the axis; Indicates the Z-axis coordinate of the design trajectory at the current time step; Assume that each excavation step is 0.01 meters long and the time step is 0.5 seconds; The TBM position information needs to be updated after each excavation progress. The design trajectory is the coordinates of discrete points of the ideal trajectory designed in the previous engineering planning; The formula for calculating the expected deviation parameter is expressed as: ； ； ； ； ； ； in, represents the component of the expected error in the X direction, X represents the X-axis coordinate in three-dimensional space, Indicates the X-axis coordinate of the actual position of the TBM at the current time step, Indicates the X-axis coordinate of the current correction track. Indicates the path deviation of the TBM in the Y-axis direction. Indicates the coordinate of the actual position of the TBM on the Y axis at the current time step; Indicates the Y-axis coordinate of the current correction track; Indicates the path deviation of the TBM in the Z-axis direction. Indicates that the actual position of the TBM at the current time step is The coordinates on the axis, Indicates the Z-axis coordinate of the current correction track; Indicates the actual angle of TBM and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between Indicates that at the current time step, the actual angle and expected angle The deviation between .

4. The data-driven automatic posture correction method according to claim 3, characterized in that: The data-driven model includes using the GRU model to predict the pressure values ​​that need to be applied to the four cylinders based on the real-time path deviation parameters. ; According to the real-time TBM status and the prediction results of GRU, the Q-learning algorithm is used to select the optimal pressure adjustment scheme, and the result of the adjustment scheme is output as the corresponding actual pressure value of the cylinder.

5. The data-driven automatic posture correction method according to claim 4, characterized in that: The data-driven model also includes, through the GRU model, predicting the pressure values ​​that need to be applied to the four cylinders ; The input parameters include the expected correction parameters, the surrounding rock parameters at the current position and the current state of the cylinder; The desired correction parameters include , , , , and ; Integrate the input parameters at the current moment with the data of the historical time step to form a time series data matrix, and perform preprocessing. The formula is expressed as: ； Choose the past The data of time steps is taken as the input sequence, and the historical time series data matrix formula is expressed as: ； in, Indicates the current time The surrounding rock parameter vector, Indicates the current time Cylinder status parameters; The preprocessing includes using Kalman filtering to remove noise in the time series data, identifying and processing outliers in the time series data, ensuring that all time series data are synchronized in time, and normalizing time series data of different dimensions; The input time window size of the GRU model is the time series data of the past 10 time steps; Divide the time series data in chronological order into 70% training set, 15% validation set, and 15% test set; The hyperparameters of the GRU model include the number of GRU units, time step, learning rate, batch size, and number of training rounds; The number of GRU units is 512 GRU layer hidden units; The time step is 25 time steps, which represents the length of the entire input sequence; The learning rate is 0.0005; The batch size is 64 samples used in one training; The number of training rounds is 100 rounds; Adjust the hyperparameters of the GRU model through grid search; The GRU model includes an input layer, a hidden layer, and an output layer; The input layer receives the preprocessed time series data; The hidden layer includes a plurality of GRU units stacked to extract time series features; The formula of the hidden layer is expressed as: ; ； ； ； ； ； in, Indicates the current time Update door of Represents the Sigmoid activation function; Represents the weight matrix input to the update gate, with dimension ,in is the dimension of the hidden state, is the dimension of the input; Indicates the current time The input data vector of Indicates the last moment The hidden state For the weight matrix of the update gate, the dimension is ; Indicates the last moment The hidden state vector of Represents the bias term of the update gate, with dimension ; represents the hyperparameter used to adjust the gain of the activation function; Indicates the current time The reset gate controls the reset of the current hidden state; Represents the weight matrix input to the reset gate, with dimension ; Indicates the last moment The hidden state of the reset gate weight matrix has the dimension ; The dimension of the bias term of the reset gate is ; Represents the attenuation factor in the integral function when controlling the reset gate; Indicates the current time The candidate hidden state of the current moment is generated; represents the hyperbolic tangent activation function; Represents the weight matrix input to the candidate hidden state, with dimension ; Indicates the last moment The weight matrix of the hidden state of the candidate hidden state is ; Reset gate Hide the state of the previous moment Weighted operation of Represents element-wise multiplication; Represents the bias term of the candidate hidden state, with dimension ; represents the hyperparameter that controls the weighting of historical information; Indicates the current time The final hidden state of Represents the historical time step index; Indicates the number of historical time steps; Indicates the last moment The hidden state vector of represents the sum of time steps; represents the time step index, express The hidden state of the moment; The output layer includes four output nodes, corresponding to The thrust prediction value is expressed as: ; in, is the weight matrix; is the bias term, which represents the bias of each output; It is the thrust prediction value of the four cylinders, with a dimension of 4, indicating the thrust value of the four cylinders The thrust prediction value.

6. The data-driven automatic posture correction method according to claim 5, characterized in that: The data-driven model further includes defining a state space, an action space, and a reward function of a Q-learning algorithm; The state space includes the current path deviation, the current tunnel boring machine posture data and the current cylinder pressure state; The action space includes different pressure adjustment schemes for four cylinders; The reward function includes giving a positive reward when the path deviation decreases, and giving a negative reward when the path deviation increases; When the cylinder applies pressure and the deviation decreases, a positive reward is given; when the cylinder applies pressure and the deviation increases, a negative reward is given; Initialize the Q table to record the Q value of each state-action pair. The initial value of Q is set to 0. According to the current state, an ε-greedy strategy is used to select actions to balance exploration and exploitation; Apply the selected pressure adjustment scheme to the cylinder to adjust the attitude of the tunnel boring machine; Calculate the reward value based on the new path deviation and cylinder pressure status, and update the Q table; The formula for updating the Q table is expressed as: ； in, Represents the new state after executing the action. represents the best action in the new state, Indicates the immediate reward of the current action; The iterative process continues until the Q table converges or the preset number of training rounds of 100 is reached, then the training is stopped.

7. The data-driven automatic posture correction method according to claim 6, characterized in that: The data-driven model further includes optimizing the learning rate, discount factor and exploration rate in the Q-learning algorithm using a genetic algorithm; The value ranges of learning rate α, discount factor γ and exploration rate ε are defined as (0,1) respectively; The definition range of each parameter (α, γ, ε) is evenly divided into 10 intervals; Randomly select a value from the interval of each parameter as a sample point; Combine the sample points of each parameter to generate several groups Parameter combination to form the initial population; The evaluation criteria of the fitness function include the total cumulative reward and the convergence speed; The cumulative reward sum includes: the higher the cumulative reward obtained during training, the higher the fitness; The convergence speed includes: The parameter combination makes the Q table converge faster and the fitness is higher; Use each set of parameters Train the Q-learning model and record the cumulative rewards and number of convergence rounds during training; The formula of the fitness function is expressed as: ; in, represents the fitness function; Represents a combination of parameters, represents the cumulative reward weight, represents the standard deviation, represents the mean; CR(θ) is the cumulative reward; represents the convergence speed weight, represents the convergence speed function, CS) represents the maximum value of convergence speed, represents the time step; The value range of F(θ) is from 0 to 1. The closer F(θ) is to 1, the higher the fitness of the parameter combination (α, γ, ε); The roulette wheel selection method is used to select the parameter combination (α, γ, ε) with high fitness to enter the crossover and mutation operations; Execute the single-point crossover method to generate new individuals; Set the mutation probability to 5% to decide whether to mutate each parameter combination (α, γ, ε); The selected parameter combination (α, γ, ε) is randomly fine-tuned to generate a new parameter combination; Select the parameter combinations with the top 30% fitness from the newly generated parameter combinations and use the hill climbing algorithm to perform local search; Local search involves, for each selected parameter combination, generating its fine-tuned parameter combination As a neighborhood solution, the formula is expressed as: ； in, represents the optimized parameter combination obtained by local search, Represents the neighborhood solution set of the current parameter combination, argmax represents Find the solution that maximizes the objective function in the neighborhood solution. represents the cumulative reward value of the solution, represents the convergence error of the solution; Evaluate the fitness of the neighborhood solution and select the neighborhood solution with the highest fitness as the new solution; combine the better parameters after local search Replace the original parameter combination ; Repeat selection, crossover, mutation, and local search to generate a new population; calculate each set of parameter combinations The fitness value of , and retain the individuals with the best performance; When the population fitness no longer improves significantly, stop iterating.

8. A data-driven automatic posture correction system using the method according to any one of claims 1 to 7, characterized in that: Digital twin model module, which builds a digital twin model of the TBM and the surrounding rock environment; A synchronization module collects the actual position and attitude data of the TBM in real time through attitude and position sensors, determines the actual excavation trajectory of the TBM, and synchronizes the actual position and attitude data of the TBM to the digital twin model; The comparison module compares the actual trajectory with the designed trajectory and calculates the actual deviation of the trajectory; The planning module plans the correction trajectory and calculates the expected correction parameters between the actual trajectory and the correction trajectory when the actual trajectory deviation exceeds the correction threshold. A prediction module, which uses the desired correction parameters, the surrounding rock parameters at the current position, and the current state of the cylinder as input parameters, and predicts the thrust required by the four driving cylinders through a data-driven model in the digital twin model; The calculation module applies the predicted thrust to the corresponding numbered cylinder to carry out the next cycle of excavation, and measures the actual position and posture information of the TBM in the next cycle, determines the actual trajectory, and then calculates the correction parameters for the next cycle; The enhancement module assembles the correction parameters, surrounding rock parameters and cylinder thrust into historical time series data for enhancing the training data-driven model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the data-driven automatic posture correction method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the data-driven automatic posture correction method described in any one of claims 1 to 7 are implemented.

Images