Shaft driving equipment posture feedforward compensation method based on formation-electromechanical multi-source characteristics

Abstract

The application discloses a shaft driving equipment posture feedforward compensation method based on stratum-mechanical and electrical multi-source features, the method constructs a space-time coupling matrix by fusing non-uniform stratum spatial distribution and high-frequency mechanical and electrical time sequence response, uses a physical information neural network with embedded dynamics mechanism to decouple and extract deep hidden features of posture evolution, then inversely calculates a globally optimal feedforward compensation vector based on a model predictive control framework, and finally dynamically corrects a microsecond-level space vector with the aid of a deterministic network and a bottom-layer multi-actuator, so that active feedforward and accurate management and control of a complex deep well driving trajectory are realized before macroscopic posture instability of the equipment. The application breaks through the hysteresis bottleneck of traditional driving equipment relying on single mechanical parameter passive correction, realizes a leap from hysteresis feedback to advanced prediction and active feedforward through space-time coupling of stratum-mechanical and electrical multi-source features and deep feature decoupling under physical mechanism constraints.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring of the operational safety of tunnel engineering structures, and relates to a method for attitude feedforward compensation of shaft excavation equipment, specifically a method for attitude feedforward compensation of shaft excavation equipment based on the multi-source characteristics of strata and electromechanical systems. Background Technology

[0002] Existing shaft tunneling equipment or similar shield tunneling equipment primarily relies on passive feedback control mechanisms (such as PID control and its variants) for attitude control. The conventional operating mode involves sensing the equipment's current geometric attitude deviation through inclinometers, displacement sensors, etc. When the deviation exceeds a set threshold, the control system sends a correction command to the propulsion system (such as zoned hydraulic cylinders or multi-layer support shoes) to adjust the tunneling trajectory. However, when facing long-distance shaft tunneling, existing technologies have the following significant drawbacks:

[0003] (1) The control mechanism has serious lag: Traditional feedback control is a passive adjustment based on the error that has already occurred. Due to the large inertia, large time delay and nonlinear characteristics of deep well tunneling equipment, and the response delay of the fluid transmission system, when the system calculates the correction amount and executes the electromechanical-hydraulic action, the equipment has often already deviated further, which can easily cause repeated oscillations or overshoots in attitude (i.e., "dragon-drawing" trajectory), which seriously increases the abnormal wear of the cutter and the tunnel segment.

[0004] (2) Lack of forward-looking coupling with the geological environment: Existing control systems are mostly limited to the mechanical parameters (thrust, torque, etc.) and apparent geometry of the equipment itself, which severs the strong coupling and mutual feedback relationship between the equipment and the strata to be traversed. When facing interfaces with uneven hardness or abrupt geological changes, the equipment cannot predict the trend of uneven load on the strata because it fails to perceive the non-uniform distribution characteristics of the stratum impedance in advance. It often only responds passively after severe deviation or even jamming occurs.

[0005] (3) Difficulty in capturing the deep temporal patterns of attitude degradation: The electromechanical high-frequency sensing data (such as vibration and stress mutation) during the tunneling process contains rich equipment-rock interaction information. Traditional control methods only perform simple threshold judgments on these multi-source data, failing to deeply explore the dynamic evolution characteristics and implicit mapping relationships of the sensing data in the spatiotemporal dimension, and thus failing to achieve advanced prediction of the attitude degradation trend of the equipment.

[0006] In summary, there is an urgent need for a control method that can integrate the characteristics of the geological formations and the electromechanical sensing data of the equipment, extract the deep evolution characteristics of the attitude, and realize active feedforward compensation and dynamic correction, so as to solve the engineering problems of attitude instability and lag in correction during deep well excavation in complex formations. Summary of the Invention

[0007] To address the engineering challenges of attitude instability and delayed correction during deep well excavation in complex geological formations, this invention provides an attitude feedforward compensation method for vertical shaft excavation equipment based on the multi-source characteristics of formation and electromechanical systems.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics.

[0010] Step 1: Spatiotemporal mapping and coupling matrix construction of heterogeneous data:

[0011] Using the real-time mileage of the vertical shaft excavation and the three-dimensional coordinates of the cutterhead space as a dynamic benchmark, the Kriging spatial interpolation algorithm is used to reconstruct the non-uniform geological distribution surface of the advanced detection in a grid. The dynamic time warping algorithm is used to eliminate the sampling rate lag between the high-frequency electromechanical sensing data of the equipment and the spatial sequence of the strata. Furthermore, through coordinate transformation, the time series parameters are mapped to the spatial grid features to construct a high-dimensional spatiotemporal tensor matrix containing multi-channel features of geological impedance and electromechanical response.

[0012] Step 2: Decoupling deep pose features under physical constraints:

[0013] The high-dimensional spatiotemporal tensor matrix is ​​input into the physical information neural network with embedded tunneling dynamics partial differential equations for joint training; a mutual information minimization constraint mechanism is introduced into the network feature extraction layer to orthogonally decouple the strongly coupled surrounding rock geological interference features and equipment active control response features, and extract the deep latent feature vector that uniquely maps the attitude degradation caused by the multi-module force bias.

[0014] Step 3: Calculate the feedforward compensation amount based on model prediction:

[0015] Using deep latent eigenvectors as input state variables, a nonlinear dynamics proxy model of equipment based on data-physics dual-drive is established to predict the partial derivatives of attitude deviation for the next advance. On this basis, within the model predictive control framework, with the dual objectives of maximizing the smoothness of the attitude correction trajectory and minimizing the energy consumption of the multi-layer support shoe and combined propulsion system, a multi-objective optimization cost function is constructed. The inverse kinematic constraints of the Jacobian matrix are solved using a multi-objective heuristic algorithm, and the optimal feedforward compensation vector solution set of the hydraulic cylinder thrust and support shoe pressure in each zone is calculated in reverse.

[0016] Step 4: Dynamic correction of space vectors for multiple actuators:

[0017] Based on a deterministic scheduling mechanism using time-sensitive networks, the optimal feedforward compensation solution set calculated in reverse is parsed into underlying synchronous control signaling. A spatial torque cross-decoupling control law is constructed. Before the equipment's macroscopic posture exceeds the limit, the variable frequency motor, combined propulsion module, and multi-layer support shoe are synchronously controlled via fieldbus to complete the vectorized adaptive redistribution of torque output, zone differential pressure feed, and stepping circumferential support force, thereby achieving smooth feedforward correction of the kilometer-level vertical shaft excavation posture in complex strata.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] 1. This invention breaks through the bottleneck of the lag in traditional tunneling equipment that relies on passive correction of a single mechanical parameter. By decoupling the spatiotemporal coupling of the multi-source features of the stratum and electromechanical systems with the deep features under the constraints of physical mechanisms, it realizes the leap from lag feedback to advance prediction and active feedforward in attitude control.

[0020] 2. This invention effectively eliminates trajectory oscillations and system overshoot caused by abrupt changes in complex strata, and completes dynamic correction of spatial vectors before the equipment's macroscopic pose becomes unstable, thereby improving the trajectory accuracy and well completion quality of kilometer-level vertical shaft excavation. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for attitude feedforward compensation of shaft tunneling equipment based on the multi-source characteristics of formation and electromechanical systems.

[0022] Figure 2 A comparative graph showing the change of attitude angle deviation (yaw angle) over time in an implementation example.

[0023] Figure 3 A comparison diagram of the response timing of the left and right partition propulsion cylinder thrust difference control in the implementation example. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0025] This invention provides a posture feedforward compensation method for shaft boring equipment based on formation-electromechanical multi-source characteristics. The method constructs a spatiotemporal coupling matrix integrating non-uniform formation spatial distribution and high-frequency electromechanical timing response. It utilizes a physical information neural network with embedded dynamic mechanisms to decouple and extract deep implicit features of posture evolution. Then, relying on a model predictive control framework, it back-calculates the globally optimal feedforward compensation vector. Finally, it uses a deterministic network to coordinate with multiple underlying actuators for microsecond-level dynamic correction of the spatial vector, thereby achieving proactive feedforward and precise control of complex deep shaft boring trajectories before macroscopic posture instability occurs. Figure 1As shown, the specific steps are as follows:

[0026] Step 1: Spatiotemporal mapping and coupling matrix construction of heterogeneous data:

[0027] Using the real-time mileage of the vertical shaft excavation and the three-dimensional coordinates of the cutterhead space as a dynamic benchmark, the Kriging spatial interpolation algorithm is employed to reconstruct the non-uniform geological distribution surface of the advanced exploration in a gridded manner. A dynamic time warping algorithm is used to eliminate the sampling rate hysteresis between the high-frequency electromechanical sensing data of the equipment and the spatial sequence of the strata. Furthermore, through coordinate transformation, the time series parameters are mapped onto the spatial grid features to construct a high-dimensional spatiotemporal tensor matrix containing multi-channel features of geological impedance and electromechanical response, which serves as the prior physical boundary for subsequent pose evolution inference. The specific steps are as follows:

[0028] Step 11: Dynamic benchmark acquisition and geological spatial reconstruction:

[0029] Using the real-time mileage of the shaft excavation and the three-dimensional coordinates of the cutterhead as a dynamic reference, the Kriging spatial interpolation algorithm is used to perform gridded calculations on the discrete data from the advance detection, reconstructing a non-uniform geoimpedance distribution surface:

[0030]

[0031] In the formula, These are the estimated geological impedance values ​​for spatial grid points. These are the optimal unbiased interpolation weight coefficients obtained based on the spatial variogram. These are the measured values ​​at known detection points. The total number of valid known probe points participating in the interpolation calculation of this grid point; The number of the detection point, ranging from 1 to... .

[0032] Step 12: Elimination of timing hysteresis in heterogeneous data:

[0033] The minimum cumulative distance warping path between high-frequency electromechanical sensing data sequences and stratigraphic spatial sequences is solved using a dynamic time warping algorithm, thereby eliminating the spatiotemporal lag caused by heterogeneous sampling rates.

[0034]

[0035] In the formula, This represents the cumulative distance matrix element of the currently aligned node; For the first Individual electromechanical sensing data sequence points With the Geological spatial sequence points Local Euclidean distance between them; The sequence number of the electromechanical sensing data sequence, with a value ranging from 1 to... , This is the total length of the time series; This refers to the sequence number of the geological spatial sequence, with a value ranging from 1 to... , This is the total length of the spatial sequence.

[0036] Step 13: Construction of the high-dimensional spatiotemporal tensor matrix:

[0037] By transforming and mapping the coordinate space, the time-aligned electromechanical parameters are embedded into the three-dimensional geological mesh, thus constructing a high-dimensional spatiotemporal tensor matrix that integrates the dual-source characteristics of geological impedance and electromechanical response. As the prior physical boundary for pose derivation, in the formula, To construct the high-dimensional spatiotemporal tensor matrix, , , Characterizing the three-dimensional spatial grid in front of the shaft excavation, Characterizes multi-channel feature dimensions that include geological and electromechanical parameters.

[0038] Step 2: Decoupling deep pose features under physical constraints:

[0039] A high-dimensional spatiotemporal tensor matrix is ​​input into a physical information neural network embedded with partial differential equations of tunneling dynamics for joint training. A mutual information minimization constraint mechanism is introduced into the network feature extraction layer to orthogonally decouple the strongly coupled surrounding rock geological disturbance characteristics and equipment active control response characteristics, extracting a unique deep latent feature vector that maps the attitude degradation caused by multi-module force imbalance. The specific steps are as follows:

[0040] Step 21: Construction of the physical information neural network and embedding of dynamic equations:

[0041] A high-dimensional spatiotemporal tensor matrix is ​​input into a Physical Information Neural Network (PINNs). Joint training is performed under the dual constraints of physical laws and data distribution by embedding the residuals of the partial differential equations of shaft excavation dynamics into the loss function. The total loss function for joint training is:

[0042]

[0043] In the formula, The total loss function for joint training of the network; The mean square error data-driven loss is calculated based on multi-source measured data; Physical residual loss calculated for the physical partial differential equations of force and attitude evolution of multiple modules of an embedded tunneling machine; This is the set of weights and bias parameters for the neural network. A dynamic adaptive penalty weighting coefficient to balance data loss and physical loss.

[0044] Step 22: Minimize mutual information of strongly coupled features and orthogonally decouple:

[0045] A mutual information minimization constraint mechanism is introduced into the feature extraction layer of the network to forcibly penalize and separate redundant correlations in the feature space, thereby achieving orthogonal decoupling of the strongly coupled surrounding rock geological disturbance features and equipment active control response features:

[0046]

[0047] In the formula, The penalty term is minimized for mutual information, and its minimization is pursued during training; The latent space feature vector extracted by the network represents the geological disturbance of the surrounding rock. The latent space feature vector extracted by the network represents the active control response of the equipment. This represents the joint probability distribution of the two types of features; The marginal probability distribution of the corresponding feature; For expectation operator, This represents all feature pairs sampled from the joint distribution of the hidden layer feature space. Find the expected value of its log-likelihood ratio.

[0048] Step 23: Extraction of deep latent feature vectors for posture degradation:

[0049] Based on the pure feature space after decoupling from physical constraints and orthogonality, a deep latent feature vector that uniquely maps the force eccentricity of multiple modules and drives the attitude degradation of the equipment is extracted through a nonlinear mapping network. As the input state variable for the subsequent proxy model, in the formula, This is the final extracted deep latent feature vector that maps the multi-module force offset and drives the equipment attitude degradation. This is an operator for concatenating and fusing orthogonally decoupled feature vectors; and These are the weight matrix and bias vector of the nonlinear mapping network at this layer, respectively; It is a non-linear activation function.

[0050] Step 3: Calculate the feedforward compensation amount based on model prediction:

[0051] Using deep latent eigenvectors as input state variables, a nonlinear dynamics surrogate model of the equipment based on data-physics dual-drive is established to predict the partial derivatives of attitude deviation for the next advance. Based on this, a multi-objective optimization cost function is constructed within the model predictive control framework (with the dual objectives of maximizing the smoothness of the attitude correction trajectory and minimizing the energy consumption of the multi-layer support shoe and combined propulsion system). A multi-objective heuristic algorithm is used to solve the inverse kinematic constraints of the Jacobian matrix, and the optimal feedforward compensation vector solution set for the cylinder thrust and support shoe pressure in each zone is calculated in reverse. The specific steps are as follows:

[0052] Step 31: Construction of Nonlinear Dynamics Surrogate Model and State Deduction:

[0053] Using the extracted deep latent feature vectors as input state variables, a nonlinear dynamics proxy model of the equipment based on data-physics dual-drive is established to predict the evolution trend of the attitude deviation of the shaft boring machine in the next prediction time domain:

[0054]

[0055] In the formula, Based on the current number The first step of the state advance simulation Step attitude deviation prediction vector; For data-physics dual-driven nonlinear dynamics surrogate model functions; For the current number The deep latent feature vectors extracted in the first step; To control the sequence of input commands in the time domain, i.e. ; To predict the number of time-domain steps; To control the number of time-domain steps.

[0056] Step 32: Construction of the multi-objective prediction control cost function:

[0057] Within the model predictive control framework, a multi-objective optimization cost function considering the dynamic evolution of the system is constructed with the dual objectives of maximizing the smoothness of the attitude correction trajectory (i.e., minimizing attitude deviation) and minimizing the energy consumption of the multi-layer support shoe and combined propulsion system:

[0058]

[0059] In the formula, Optimize the cost function for multiple objectives; The desired ideal attitude reference vector (usually with zero deviation); To control the input increment, the action energy consumption and wear of the actuator are characterized; Q is the penalty weight matrix for attitude tracking error, used to ensure trajectory smoothness; R is the penalty weight matrix for controlling energy consumption, used to suppress the violent oscillation of the actuator. The norm of a matrix is ​​represented.

[0060] Step 33: Optimal compensation vector solution based on inverse kinematic constraints:

[0061] The cost function is solved using a multi-objective heuristic algorithm. Under the inverse kinematic physical constraints represented by the Jacobian matrix, the optimal feedforward compensation control vector solution set, which includes the thrust of each hydraulic cylinder and the pressure of the support shoe, is calculated in reverse.

[0062]

[0063] In the formula, The optimal feedforward compensation control vector solution set is finally calculated in reverse (i.e., the actual control command for the next advance). This indicates that the cost function Minimize the value of the variable; st represents the constraint condition; The inverse kinematic constraint equations are given, where Let be the Jacobian partial derivative matrix of the system attitude with respect to the actuator input; and These represent the lower and upper limits of the physical output of the hydraulic cylinder thrust and the multi-layer support shoe pressure in each zone, respectively.

[0064] Step 4: Dynamic correction of space vectors for multiple actuators:

[0065] Based on a deterministic scheduling mechanism using time-sensitive networks, the optimal feedforward compensation solution set calculated in reverse is parsed into underlying synchronous control signaling. A spatial torque cross-decoupling control law is constructed. Before the equipment's macroscopic posture exceeds the limit, the variable frequency motor, combined propulsion module, and multi-layer support shoes are synchronously controlled via fieldbus to complete the vectorized adaptive redistribution of torque output, zoned differential pressure feed, and stepping circumferential support force, achieving smooth feedforward correction of the kilometer-level vertical shaft excavation posture in complex strata. The specific steps are as follows:

[0066] Step 41: Deterministic analysis of underlying synchronization control signaling:

[0067] Based on a time-sensitive network-based deterministic scheduling mechanism, the current step control command in the optimal feedforward compensation vector solution set calculated by backpropagation is parsed into low-level synchronous control signaling with high-precision timestamps after compensating for network transmission and mechanical response delays.

[0068]

[0069] In the formula, This is the continuous-time synchronization control signaling that is parsed and then sent to the fieldbus; This is a deterministic scheduling and delay compensation mapping function for time-sensitive networks; For the optimal feedforward compensation sequence in the current kth ... The actual instructions issued for each step; This is the total system delay constant, which includes bus communication delay and dead zone of hydraulic / electric actuators.

[0070] Step 42, Spatial moment cross-decoupling and vector adaptive redistribution:

[0071] A spatial torque cross-decoupling control law is constructed to eliminate the parasitic coupling effect between the rotation of the variable frequency motor, the propulsion of the hydraulic cylinder, and the radial support of the support shoe. Based on the spatial physical topology of the shaft boring machine, the bus compensation signaling is vectorized and redistributed as the reference driving force for each actuator.

[0072]

[0073] In the formula, , , These are the torque of the cutterhead variable frequency motor after redistribution, the pressure difference of the combined propulsion cylinders in each zone, and the circumferential support force vector of the multi-layer support shoe; The cross-decoupling gain matrix is ​​constructed based on the actual spatial geometry and topology of the equipment; The spatial three-dimensional attitude compensation torque / force vector is derived from the synchronous control signaling. , , To maintain the steady-state reference working vector for the current foundation tunneling operation.

[0074] Step 43: Feedforward offsetting and smoothing correction before pose exceeds limits:

[0075] Before the equipment's macroscopic posture exceeds the safety warning boundary, multiple actuators are driven by the field industrial bus to perform high-frequency synchronous actions strictly according to the redistributed vector. This physically offsets the off-center load disturbance caused by the non-uniform impedance of the formation in advance, achieving stable feedforward correction of the kilometer-level vertical shaft excavation posture.

[0076]

[0077] In the formula, The actual three-dimensional pose state of the equipment at the next moment after the action is performed; This describes the state transition process of a shaft boring machine in a real physical environment. Configure the dynamic input matrix for each actuator; The maximum safe over-limit threshold for the macroscopic posture of the vertical shaft excavation allowed by the project; For weighted The weighted state norm emphasizes that the feedforward hedging action must ensure that the pose at the next moment is strictly constrained within the safety envelope.

[0078] Example:

[0079] This embodiment takes a kilometer-level vertical shaft excavation project as the background and constructs a mechanical-hydraulic dynamics simulation model of a full-section vertical shaft excavation equipment (excavation diameter Φ=10m, total weight 1200t).

[0080] (1) Geological abrupt boundary: When the tunneling reaches a depth of 800m underground, there is a typical half-hard rock and half-soft rock abrupt boundary 2m ahead (the uniaxial compressive strength UCS on the left side is 120MPa, and the soft rock UCS on the right side is 30MPa).

[0081] (2) Initial tunneling parameters: the initial penetration depth is set to 6 mm / r, the rated speed of the cutterhead is 4.5 rpm, and the total thrust is 24,000 kN.

[0082] Based on simulation calculations, the performance comparison between the method of this invention and the traditional method is shown in Table 1.

[0083] Table 1 Comparison of attitude control performance under abrupt changes in uneven geological formations.

[0084]

[0085] contrast Figure 2 As shown by the two curves, the existing technology suffers from huge transient deviations and long oscillation convergence processes due to control lag; while the present invention reduces the maximum attitude deviation by 85.8% through advanced sensing and feedforward action, and basically eliminates the trajectory oscillation phenomenon.

[0086] Figure 3 This invention visually demonstrates the core mechanism of feedforward compensation. Existing technologies correct deviations upon observation, but the actions always lag behind sudden load changes; this invention anticipates and adjusts, controlling actions to be ahead of or synchronized with sudden load changes, thereby ensuring stable attitude.

[0087] Numerical simulations fully demonstrate that this invention overcomes the lag limitations of traditional passive feedback control by employing a spatiotemporal coupling of formation-electromechanical multi-source characteristics and an active feedforward compensation mechanism. Under extreme formation abrupt changes, it achieves response and attitude maintenance within a minimal error range, providing an effective technical means for the precise and safe construction of kilometer-level vertical shafts.

Claims

1. A method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics, characterized in that... The method includes the following steps: Step 1: Spatiotemporal mapping and coupling matrix construction of heterogeneous data: Using the real-time mileage of the vertical shaft excavation and the three-dimensional coordinates of the cutterhead space as a dynamic benchmark, the Kriging spatial interpolation algorithm is used to reconstruct the non-uniform geological distribution surface of the advanced detection in a grid. The dynamic time warping algorithm is used to eliminate the sampling rate lag between the high-frequency electromechanical sensing data of the equipment and the spatial sequence of the strata. Furthermore, through coordinate transformation, the time series parameters are mapped to the spatial grid features to construct a high-dimensional spatiotemporal tensor matrix containing multi-channel features of geological impedance and electromechanical response. Step 2: Decoupling deep pose features under physical constraints: The high-dimensional spatiotemporal tensor matrix is ​​input into the physical information neural network with embedded tunneling dynamics partial differential equations for joint training; a mutual information minimization constraint mechanism is introduced into the network feature extraction layer to orthogonally decouple the strongly coupled surrounding rock geological interference features and equipment active control response features, and extract the deep latent feature vector that uniquely maps the attitude degradation caused by the multi-module force bias. Step 3: Calculate the feedforward compensation amount based on model prediction: Using deep latent eigenvectors as input state variables, a nonlinear dynamics proxy model of equipment based on data-physics dual-drive is established to predict the partial derivatives of attitude deviation for the next advance. On this basis, within the model predictive control framework, with the dual objectives of maximizing the smoothness of the attitude correction trajectory and minimizing the energy consumption of the multi-layer support shoe and combined propulsion system, a multi-objective optimization cost function is constructed. The inverse kinematic constraints of the Jacobian matrix are solved using a multi-objective heuristic algorithm, and the optimal feedforward compensation vector solution set of the hydraulic cylinder thrust and support shoe pressure in each zone is calculated in reverse. Step 4: Dynamic correction of space vectors for multiple actuators: Based on a deterministic scheduling mechanism using time-sensitive networks, the optimal feedforward compensation solution set calculated in reverse is parsed into underlying synchronous control signaling. A spatial torque cross-decoupling control law is constructed. Before the equipment's macroscopic posture exceeds the limit, the variable frequency motor, combined propulsion module, and multi-layer support shoe are synchronously controlled via fieldbus to complete the vectorized adaptive redistribution of torque output, zone differential pressure feed, and stepping circumferential support force, thereby achieving smooth feedforward correction of the kilometer-level vertical shaft excavation posture in complex strata.

2. The method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics according to claim 1, characterized in that... The specific steps of step 1 are as follows: Step 11: Dynamic benchmark acquisition and geological spatial reconstruction: Using the real-time mileage of the shaft excavation and the three-dimensional coordinates of the cutterhead as a dynamic reference, the Kriging spatial interpolation algorithm is used to perform gridded calculations on the discrete data from the advance detection, reconstructing a non-uniform geoimpedance distribution surface: In the formula, These are the estimated geological impedance values ​​for spatial grid points. These are the optimal unbiased interpolation weight coefficients obtained based on the spatial variogram. These are the measured values ​​at known detection points. The total number of valid known probe points participating in the interpolation calculation of this grid point; The serial number of the detection point; Step 12: Elimination of timing hysteresis in heterogeneous data: The minimum cumulative distance warping path between high-frequency electromechanical sensing data sequences and stratigraphic spatial sequences is solved using a dynamic time warping algorithm, thereby eliminating the spatiotemporal lag caused by heterogeneous sampling rates. In the formula, This represents the cumulative distance matrix element of the currently aligned node; For the first Individual electromechanical sensing data sequence points With the Geological spatial sequence points Local Euclidean distance between them; This refers to the sequence number of the electromechanical sensing data sequence; This refers to the sequence number of the geological spatial sequence; Step 13: Construction of the high-dimensional spatiotemporal tensor matrix: By transforming and mapping the coordinate space, the time-aligned electromechanical parameters are embedded into the three-dimensional geological mesh, thus constructing a high-dimensional spatiotemporal tensor matrix that integrates the dual-source characteristics of geological impedance and electromechanical response. As the prior physical boundary for pose derivation, in the formula, To construct the high-dimensional spatiotemporal tensor matrix, , , Characterizing the three-dimensional spatial grid in front of the shaft excavation, Characterizes multi-channel feature dimensions that include geological and electromechanical parameters.

3. The method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics according to claim 1, characterized in that... The specific steps of step 2 are as follows: Step 21: Construction of the physical information neural network and embedding of dynamic equations: By inputting the high-dimensional spatiotemporal tensor matrix into the physical information neural network, and embedding the residual of the partial differential equation of shaft excavation dynamics into the loss function, joint training is carried out under the dual constraints of physical laws and data distribution. Step 22: Minimize mutual information of strongly coupled features and orthogonally decouple: A mutual information minimization constraint mechanism is introduced into the feature extraction layer of the network to forcibly penalize and separate redundant correlations in the feature space, thereby achieving orthogonal decoupling of the strongly coupled surrounding rock geological disturbance features and equipment active control response features: In the formula, The penalty term is minimized for mutual information, and its minimization is pursued during training; The latent space feature vector extracted by the network represents the geological disturbance of the surrounding rock. The latent space feature vector extracted by the network represents the active control response of the equipment. This represents the joint probability distribution of the two types of features; The marginal probability distribution of the corresponding feature; For expectation operator, This represents all feature pairs sampled from the joint distribution of the hidden layer feature space. Find the expected value of its log-likelihood ratio; Step 23: Extraction of deep latent feature vectors for posture degradation: Based on the pure feature space after decoupling from physical constraints and orthogonality, a deep latent feature vector that uniquely maps the force eccentricity of multiple modules and drives the attitude degradation of the equipment is extracted through a nonlinear mapping network. As the input state variable for the subsequent proxy model, in the formula, This is the final extracted deep latent feature vector that maps the multi-module force offset and drives the equipment attitude degradation. This is an operator for concatenating and fusing orthogonally decoupled feature vectors; and These are the weight matrix and bias vector of the nonlinear mapping network at this layer, respectively; It is a non-linear activation function.

4. The method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics according to claim 3, characterized in that... The total loss function for the joint training is: In the formula, The total loss function for joint training of the network; The mean square error data-driven loss is calculated based on multi-source measured data; Physical residual loss calculated for the physical partial differential equations of force and attitude evolution of multiple modules of an embedded tunneling machine; This is the set of weights and bias parameters for the neural network. A dynamic adaptive penalty weighting coefficient to balance data loss and physical loss.

5. The method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics according to claim 1, characterized in that... The specific steps of step 3 are as follows: Step 31: Construction of Nonlinear Dynamics Surrogate Model and State Deduction: Using the extracted deep latent feature vectors as input state variables, a nonlinear dynamics proxy model of the equipment based on data-physics dual-drive is established to predict the evolution trend of the attitude deviation of the shaft boring machine in the next prediction time domain: In the formula, Based on the current number The first step of the state advance simulation Step attitude deviation prediction vector; For data-physics dual-driven nonlinear dynamics surrogate model functions; For the current number The deep latent feature vectors extracted in the first step; To control the sequence of input commands in the time domain, i.e. ; To predict the number of time-domain steps; To control the number of time-domain steps; Step 32: Construction of the multi-objective prediction control cost function: Within the model predictive control framework, a multi-objective optimization cost function considering the dynamic evolution of the system is constructed, with the dual objectives of maximizing the smoothness of the attitude correction trajectory and minimizing the energy consumption of the multi-layer support shoe and combined propulsion system. In the formula, Optimize the cost function for multiple objectives; The desired ideal attitude reference vector; To control the input increment, the action energy consumption and wear of the actuator are characterized; Q is the penalty weight matrix for attitude tracking error, used to ensure trajectory smoothness; R is the penalty weight matrix for controlling energy consumption, used to suppress the violent oscillation of the actuator. Represents the norm of a matrix; Step 33: Optimal compensation vector solution based on inverse kinematic constraints: The cost function is solved using a multi-objective heuristic algorithm. Under the inverse kinematic physical constraints represented by the Jacobian matrix, the optimal feedforward compensation control vector solution set, which includes the thrust of each hydraulic cylinder and the pressure of the support shoe, is calculated in reverse. In the formula, This is the final optimal feedforward compensation control vector solution set obtained through reverse calculation; This indicates that the cost function Minimize the value of the variable; st represents the constraint condition; The inverse kinematic constraint equations are given, where Let be the Jacobian partial derivative matrix of the system attitude with respect to the actuator input; and These represent the lower and upper limits of the physical output of the hydraulic cylinder thrust and the multi-layer support shoe pressure in each zone, respectively.

6. The method for attitude feedforward compensation of shaft tunneling equipment based on formation-electromechanical multi-source characteristics according to claim 1, characterized in that... The specific steps of step 4 are as follows: Step 41: Deterministic analysis of underlying synchronization control signaling: Based on a time-sensitive network-based deterministic scheduling mechanism, the current step control command in the optimal feedforward compensation vector solution set calculated by backpropagation is parsed into low-level synchronous control signaling with high-precision timestamps after compensating for network transmission and mechanical response delays. In the formula, This is the continuous-time synchronization control signaling that is parsed and then sent to the fieldbus; This is a deterministic scheduling and delay compensation mapping function for time-sensitive networks; For the optimal feedforward compensation sequence in the current kth ... The actual instructions issued for each step; This is the total system delay constant, which includes bus communication delay and dead zone of hydraulic / electric actuators; Step 42, Spatial moment cross-decoupling and vector adaptive redistribution: A spatial torque cross-decoupling control law is constructed to eliminate the parasitic coupling effect between the rotation of the variable frequency motor, the propulsion of the hydraulic cylinder, and the radial support of the support shoe. Based on the spatial physical topology of the shaft boring machine, the bus compensation signaling is vectorized and redistributed as the reference driving force for each actuator. In the formula, , , These are the torque of the cutterhead variable frequency motor after redistribution, the pressure difference of the combined propulsion cylinders in each zone, and the circumferential support force vector of the multi-layer support shoe; The cross-decoupling gain matrix is ​​constructed based on the actual spatial geometry and topology of the equipment; The spatial three-dimensional attitude compensation torque / force vector is derived from the synchronous control signaling. , , To maintain the steady-state reference working vector for the current foundation tunneling operation; Step 43: Feedforward offsetting and smoothing correction before pose exceeds limits: Before the equipment's macroscopic posture exceeds the safety warning boundary, multiple actuators are driven by the field industrial bus to perform high-frequency synchronous actions strictly according to the redistributed vector. This physically offsets the off-center load disturbance caused by the non-uniform impedance of the formation in advance, achieving stable feedforward correction of the kilometer-level vertical shaft excavation posture. In the formula, The actual three-dimensional pose state of the equipment at the next moment after the action is performed; This describes the state transition process of a shaft boring machine in a real physical environment. Configure the dynamic input matrix for each actuator; The maximum safe over-limit threshold for the macroscopic posture of the vertical shaft excavation allowed by the project; For weighted The weighted state norm emphasizes that the feedforward hedging action must ensure that the pose at the next moment is strictly constrained within the safety envelope.

Images