Progressive hierarchical optimization-based multi-mode pipe jacking machine tunneling posture control construction method

Abstract

The present application relates to the field of tunneling construction, and discloses a multi-mode pipe jacking machine tunneling posture control construction method based on progressive hierarchical optimization, which comprises the following steps: constructing and online updating a generation model, and outputting an updated active inference generation model; using the active inference generation model to generate a set of constrained candidate control actions; performing forward simulation prediction on the constrained candidate control actions to obtain a set of multi-action prediction results; obtaining an action and variational free energy mapping; determining the control level and constraint strategy of the action; based on the equipment process constraint and the control level and constraint strategy of the action, constructing a hierarchical optimization problem with the minimum variational free energy as the target, screening and solving the candidate control actions in the action and variational free energy mapping, and generating the optimal control action of the current control period. The present application realizes dynamic adaptive trade-off between safety and performance.

Description

Technical Field

[0001] This invention relates to the field of tunneling construction, and more specifically, to a construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive layered optimization. Background Technology

[0002] When traversing complex geological formations, multi-mode pipe jacking machines must simultaneously address multiple objectives, including attitude correction, excavation face stability, surface settlement control, and tunneling efficiency. Existing control methods often employ weighted summation to handle these conflicting objectives, such as assigning specific weights to safety and performance indicators for optimization, or switching between different control modes based on preset rules. However, the underground engineering environment is highly unstable. Sudden changes in geological parameters, sensor noise, and equipment wear cause continuous drift in the system's dynamic characteristics. Specific weights or rigid switching rules are difficult to adaptively match real-time operating conditions, easily leading to problems such as prioritizing efficiency under high-risk conditions or being overly conservative under low-risk conditions.

[0003] In recent years, active inference theory has been introduced into the field of complex equipment control. It achieves unified optimization of probabilistic perception and control decisions of hidden states by constructing generative models and minimizing variational free energy. This framework naturally supports the trade-off of prior preferences, observation uncertainties, and control costs within the same probability space, and is theoretically applicable to the multi-objective robust control of multi-mode pipe jacking machines. However, when existing active inference is applied to tunnel boring equipment, it typically designs generative models and free energy functions only for a single control objective, lacking explicit modeling of progressive safety and performance hierarchies, and failing to address the coupling problem between multi-level constraints and free energy optimization. This results in its inability to achieve true hierarchical adaptive control in practical engineering.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] A construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization includes:

[0008] Based on multi-source sensing and construction data during the multi-mode pipe jacking machine's tunneling process, observation vectors are constructed. Combined with active inference theory, a generative model is built to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control action conditions. The generative model is updated online to match the current geological formation and multi-mode pipe jacking machine operating conditions, outputting the updated active inference generative model. Using the active inference generative model, variational inference is performed within the current control cycle to calculate the posterior distribution of state variables. Combined with equipment and process boundary conditions, a constrained set of candidate control actions is generated. Simultaneously, the active inference generative model is used to perform forward simulation of the constrained candidate control actions. The system performs measurements to obtain a set of multi-action prediction results; configures the target prior preference and calculates the variational free energy of candidate control actions in the multi-action prediction result set, forming a mapping between actions and variational free energy; constructs a progressive control hierarchy and threshold rules, and determines the control hierarchy and constraint strategy of actions based on the mapping between actions and variational free energy and the threshold decision quantity corresponding to the posterior distribution of state variables; based on equipment process constraints and the control hierarchy and constraint strategy of actions, it constructs a hierarchical optimization problem with the goal of minimizing variational free energy, filters and solves the candidate control actions in the mapping between actions and variational free energy, and generates the optimal control action for the current control cycle, which is used to control the construction of the multi-mode pipe jacking machine.

[0009] Furthermore, based on the multi-source sensing and construction data during the multi-mode pipe jacking machine's tunneling process, an observation vector is constructed, including: collecting multi-source sensing and construction data in each control cycle, and performing time synchronization and anomaly removal processing on the multi-source sensing and construction data; calculating characteristic quantities to characterize the current working condition on the data after time synchronization and anomaly removal processing, and performing dimensional unification and normalization processing on each characteristic quantity; and splicing the normalized characteristic quantities in a preset order to form the observation vector for the current control cycle.

[0010] Furthermore, combining active inference theory, a generative model is constructed to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control actions. This includes: determining the set of state variables for the generative model based on the attitude control target and working mechanism of the multi-mode pipe jacking machine, including attitude state, guidance deviation state, silo pressure state, slag discharge state, and settlement risk state; establishing an observation function from state variables to observation vectors to characterize the mapping relationship between state variables and observation vectors, and introducing an observation noise term to characterize sensing uncertainty; establishing a state transition function under candidate control actions to characterize the influence of candidate control actions on the evolution of state variables, and introducing a process noise term to characterize formation disturbance and unmodeled dynamics; configuring prior distributions for state variables and generative model parameters, and configuring a target prior preference term for active inference, so that the generative model can express the expected state evolution trend under different control actions; and parameterizing the observation function and state transition function to obtain the computational form for variational inference and forward prediction.

[0011] Furthermore, the generated model is updated online to match the current formation and the multi-mode pipe jacking machine's operating conditions, outputting the updated active inference generated model. This includes: acquiring the most recent observation vector sequence of a preset length and the corresponding control action sequence within the current control cycle to form a data window; performing forward prediction on the data window based on the generated model and calculating the error between the predicted observation and the actual observation to generate a representation of the matching degree between the model and the current formation and operating conditions; determining whether to trigger an online update by comparing the error with a preset threshold; if no online update is triggered, keeping the active inference generated model parameters from the previous control cycle unchanged; if an online update is triggered, determining the update step size and regularization weights, and using recursive optimization to update the observation function parameters and state transition function parameters in the generated model, reducing the free energy of the updated generated model on the data window; performing boundary constraint verification and prediction consistency verification on the updated generated model; if the verification passes, outputting the updated active inference generated model; if the verification fails, reverting to the generated model parameters before the update and freezing the update for this cycle.

[0012] Furthermore, utilizing an active inference-based generative model, variational inference is performed within the current control cycle to calculate the posterior distribution of state variables. Combined with equipment and process boundary conditions, a constrained set of candidate control actions is generated. This includes: prior propagation of the current state distribution under the previous control action condition based on the state transition function to obtain the predicted prior distribution of state variables; constructing a variational distribution and calculating the observation likelihood based on the observation function, iteratively updating it by minimizing the variational free energy to output the posterior distribution of state variables; obtaining the multi-mode pipe jacking machine's capability boundary and process safety boundary to form a constraint-based set; constructing an uncertainty constraint margin based on the posterior and predicted prior distributions of state variables to modify the basic constraint set into a probabilistically feasible constraint set for the current control cycle; generating a set of candidate control actions within the probabilistically feasible constraint set based on the controllable variable dimension and discrete control strategy, and performing feasibility screening on the candidate actions to obtain a constrained set of candidate control actions. Simultaneously, the constrained set of candidate control actions is structured and parameterized.

[0013] Furthermore, using an active inference generative model, forward simulation prediction is performed on constrained candidate control actions to obtain a set of multi-action prediction results. This includes: indexing and temporally expanding the set of constrained candidate control actions to form the action sequence of each candidate action in the prediction time domain; recursively calculating the predicted state distribution based on the state transition function under the condition of the action sequence of each candidate action in the prediction time domain, using the current state posterior distribution as the initial condition; mapping the predicted state distribution to the predicted observation distribution based on the observation function, and incorporating process noise and observation noise into uncertainty propagation; and collecting the predicted state trajectory, predicted observation trajectory, and uncertainty corresponding to each candidate action to form a set of multi-action prediction results.

[0014] Furthermore, the target prior preferences are configured, and the variational free energy of candidate control actions in the multi-action prediction result set is calculated to form a mapping between actions and variational free energy. This includes: obtaining preference indices corresponding to the tunneling target and the safety constraint target; converting the preference indices into a target prior preference distribution, and aligning the prediction distributions of each candidate control action in the multi-action prediction result set with the target prior preference distribution on the same variable space and time index; and calculating the variational free energy of each candidate control action in the prediction time domain.

[0015] ;

[0016] In the formula, Candidate actions The scalar evaluation value of the variational free energy in the prediction time domain. For the first One candidate control action, For the future moment The observed variables, In order to execute candidate actions Under these conditions, the predicted observation distribution obtained from forward simulation by the active inference generation model, For the future moment The target prior preference distribution At the current control cycle time, To predict the length of the time domain, Let KL divergence be the KL divergence. Let cost function be These are weighting coefficients; after truncating, smoothing, or normalizing the variational free energy, a mapping relationship between the action and the corresponding variational free energy is formed.

[0017] Furthermore, a progressive control hierarchy and threshold rules are constructed, and the control hierarchy and constraint strategy for actions are determined based on the mapping between actions and variational free energy and the threshold decision quantity corresponding to the posterior distribution of state variables, including:

[0018] A progressive control hierarchy is constructed, and corresponding constraint strategies are preset for each control level. Based on the posterior distribution of key state variables, threshold rules and threshold decision quantities are constructed. According to the threshold rules, the threshold decision quantities are compared with the equipment process boundary to output the risk level. Free energy evaluation indicators are extracted based on the action-variable free energy mapping. The risk level and free energy evaluation indicators are jointly determined to determine the control level of the current control cycle. According to the determined control level, corresponding constraint strategies are applied to the candidate action set to generate hierarchical action constraint rules.

[0019] Furthermore, the risk level and free energy evaluation index are jointly determined to establish the control level for the current control cycle. This includes: obtaining a recommended control level based on the risk level according to a pre-defined risk level and control level comparison table; extracting the free energy values ​​of all candidate actions in the current cycle from the action-variable free energy mapping results and calculating the free energy determination quantity; obtaining a recommended control level based on free energy according to the pre-defined free energy determination quantity and control level rules; and using the principles of risk priority and conservatism priority, jointly comparing the recommended control level based on the risk level and the recommended control level based on free energy to determine the control level for the current control cycle.

[0020] Furthermore, based on equipment process constraints and the control hierarchy and constraint strategies of actions, a hierarchical optimization problem is constructed with the goal of minimizing variational free energy. Candidate control actions in the mapping between actions and variational free energy are screened and solved to generate the optimal control action for the current control cycle. This includes: obtaining the control hierarchy and corresponding constraint strategies determined for the current control cycle, merging equipment process constraints and constraint strategies to construct a feasible domain constraint set for the current control cycle; constructing a hierarchical objective function for the hierarchical optimization problem by setting hierarchical weights according to the control hierarchy, with the goal of minimizing variational free energy; performing feasibility checks on each candidate control action in the mapping between actions and variational free energy, eliminating candidate control actions that do not satisfy the feasible domain constraint set, and obtaining a subset of feasible candidate actions; calculating hierarchical free energy evaluation values ​​for the subset of feasible candidate actions based on the hierarchical objective function, and comparing and ranking them according to hierarchical priority to obtain a set of candidate actions with the minimum free energy; and performing local continuous optimization in the neighborhood of the candidate action set to obtain the optimal control action that satisfies the feasible domain constraint set.

[0021] The beneficial effects of this invention are as follows:

[0022] 1. This invention achieves a dynamic adaptive trade-off between safety and performance: by jointly determining the control level by risk level and free energy evaluation index, and dynamically adjusting the weight coefficients in the hierarchical optimization objective function accordingly, it automatically converges to the higher-level safety constraint set when there is high risk or low model confidence, and releases the lower-level optimization space when there is low risk and the model is reliable, thus achieving a real-time optimal balance between safety, compliance and economy across the entire operating range.

[0023] 2. Constructing a probabilistic dynamic safety envelope to improve robustness: By utilizing the uncertainty of the posterior distribution of state variables to construct constraint margins, the hard boundaries of equipment processes are modified into a set of probabilistic feasible constraints, ensuring that the probability of candidate control actions violating the real physical boundaries under the current cognitive state is lower than the preset risk threshold; at the same time, forward simulation predicts the complete propagation process and observation noise, so that control decisions are not only based on expected performance, but also quantify the reliability of prediction, effectively suppressing the risk of control divergence under geological changes or sensor anomalies.

[0024] 3. This invention forms a dual guarantee mechanism of closed-loop self-calibration and hierarchical backoff: the generated model is updated online through a sliding window, and the parameter adaptive calibration is triggered by the variational free energy gradient to avoid the prediction failure of the determined model under the condition drift; when the free energy exceeds the limit or the key state index exceeds the limit, it automatically backoffs to a higher priority control level and applies conservative constraints. Combined with boundary verification and prediction consistency check, it forms a redundant safety barrier, which enhances the autonomous control stability and engineering applicability of the multi-mode pipe jacking machine in extreme non-stationary environments. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a multi-mode pipe jacking machine tunneling attitude control construction method based on progressive hierarchical optimization according to an embodiment of the present invention;

[0027] Figure 2 This is a system block diagram according to an embodiment of the present invention;

[0028] Figure 3 This is a graph showing the relationship between the type of pipe jacking machine and the formation particle size according to an embodiment of the present invention;

[0029] Figure 4 This is a diagram showing the relationship between the type of pipe jacking machine and the formation permeability coefficient according to an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of a cutter head according to an embodiment of the present invention. Detailed Implementation

[0031] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0032] According to an embodiment of the present invention, a construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization is provided.

[0033] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization is provided, comprising:

[0034] S1. Based on the multi-source sensing and construction data during the tunneling process of the multi-mode pipe jacking machine, an observation vector is constructed. Combined with active inference theory, a generative model is built to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control action conditions.

[0035] In one embodiment, an observation vector is constructed based on multi-source sensing and construction data during the tunneling process of a multi-mode pipe jacking machine. This includes: collecting multi-source sensing and construction data such as attitude or guidance deviation, propulsion cylinder thrust and distribution, propulsion speed, cutterhead torque and speed, slurry pump flow and differential pressure, screw conveyor speed and torque, excavation chamber or slurry chamber pressure and its fluctuation, and surface settlement or equivalent early warning amount in each control cycle; performing time synchronization and anomaly removal processing on the multi-source sensing and construction data; calculating characteristic quantities to characterize the current working condition from the time synchronization and anomaly removal data, including pressure fluctuation characteristics and attitude deviation characteristics, and performing dimensional unification and normalization processing on each characteristic quantity; and splicing the normalized characteristic quantities in a preset order to form the observation vector for the current control cycle, which is used as the input for the active inference generation model.

[0036] In one embodiment, a generative model is constructed using active inference theory to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control actions. This includes: determining the set of state variables for the generative model based on the attitude control target and working mechanism of a multi-mode pipe jacking machine; state variables including attitude state, guidance deviation state, silo pressure state, slag discharge state, and settlement risk state; establishing an observation function from state variables to observation vectors to characterize the mapping relationship between state variables and observation vectors, and introducing an observation noise term to characterize sensing uncertainty; establishing a state transition function under candidate control actions to characterize the impact of candidate control actions on the evolution of state variables, and introducing a process noise term to characterize formation disturbance and unmodeled dynamics; configuring prior distributions for state variables and generative model parameters, and configuring a target prior preference term for active inference, so that the generative model can express the expected state evolution trend under different control actions; and parameterizing the observation function and state transition function to obtain computational forms for variational inference and forward prediction.

[0037] This invention, based on active inference theory, abstracts the multi-mode pipe jacking machine tunneling system into a probabilistic generative model containing latent states and observable quantities, achieving a unity of perception and prediction. Specifically, it performs spatiotemporal alignment and preprocessing on the collected multi-source heterogeneous sensor data such as thrust, torque, silo pressure, and settlement, extracting normalized feature vectors that characterize the current interaction between the strata and the equipment as observation inputs. Based on the dynamic mechanism of the pipe jacking machine and the construction objectives, it defines latent state variables such as attitude, silo pressure, muck discharge, and settlement risk, and constructs observation functions with noise terms (describing how latent states generate observations) and state transition functions (describing how control actions drive the evolution of latent states). Simultaneously, it embeds target prior preferences to express the desired construction trajectory and safety boundary. By parametrically modeling these functions, a closed-loop generative model is formed that can predict the forward state under candidate control actions and invert the current true state through variational inference, providing a probabilistic environmental representation for subsequent optimization control based on free energy minimization.

[0038] 1. Input Vector: For multi-source sensor data after time synchronization and anomaly removal, pressure fluctuation characteristics are characterized by calculating the standard deviation, peak-to-peak value, and power spectral density of the excavation chamber pressure and slurry circulation pressure within the sliding window to represent pressure stability and pulsation intensity. Attitude deviation characteristics are constructed by building a comprehensive deviation index based on the Euclidean distance and angle change rate between the real-time measured values ​​of the nose tilt angle, yaw angle, and roll angle and their corresponding expected values ​​of the target trajectory. Furthermore, axial position deviation and radial offset are integrated to form a multi-dimensional attitude error vector. All feature quantities are normalized to eliminate dimensional differences and spliced ​​into a unified dimension working condition feature input vector.

[0039] 2. Set of State Variables: The attitude control objective of the multi-mode pipe jacking machine refers to the simultaneous achievement of multiple coordinated requirements in composite strata tunneling, including axis tracking, stable balance of the excavation face, and controllable surface settlement. The working condition mechanism encompasses physical processes such as dynamic response of earth pressure and slurry balance, interaction between the cutterhead and the strata, conservation of slag discharge volume, and dynamic coupling of attitude and guidance. Based on this, the attitude angle and position, which directly reflect the spatial posture of the machine, are set as the attitude state to characterize the core control output. The geometric deviation of the current trajectory from the theoretical axis and its rate of change are set as the guidance deviation state to quantify the correction requirements. The excavation chamber pressure and its fluctuation characteristics are set as the chamber pressure state to characterize the stability of the tunnel face. The ratio and trend of the unit advance slag discharge volume to the theoretical slag discharge volume are set as the slag discharge state to reflect the strata loss and over- and under-excavation risks. The equivalent settlement index, derived from the fusion of chamber pressure, slag discharge, and strata parameters, is set as the settlement risk state to compensate for the time lag defects of measured settlement. This constitutes a set of state variables that both covers the control objective and embeds the working condition mechanism.

[0040] 3. Probabilistic Prediction and Optimization Orientation: Based on the principle of free energy minimization in the active inference framework, domain knowledge is encoded and the uncertainty range of the posterior estimate is constrained by setting conjugate prior distributions (such as Gaussian inverse Wissaud distribution) for state variables and model parameters. At the same time, the target prior preference term is configured as the probabilistic representation of the ideal working condition trajectory of the generative model. This enables the generative model to not only fit the observed data during the Bayesian update process, but also actively align the evolution of the posterior state with the expected trend that conforms to the control objective. Under the conditions of sensor noise and formation disturbance, probabilistic prediction and optimization orientation of the effect of candidate control actions are achieved.

[0041] Furthermore, the abstract mapping relationship between the observation function and the state transition function is transformed into a mathematical form explicitly expressed by finite-dimensional learnable parameters, giving the originally non-parametric or semi-analytical function a definite computational structure. This supports gradient optimization of the posterior distribution in variational inference and numerical propagation in forward prediction, realizing a concrete transformation from theoretical models to trainable and inferable computational instances.

[0042] The observation function is expressed as In the formula, for The observation vector at time t, For a set of state variables, These are the learnable parameters of the observation function (including the kernel hyperparameters of Gaussian process regression and the weights of the heteroscedastic output layer). For Gaussian process regression mapping with heteroscedastic output, The observation noise term is represented by the state transition function, which is expressed as follows: In the formula For candidate control action vectors, This is a physical mechanism model based on the rigid body dynamics and slurry balance equations of a pipe jacking machine. For known or identified parameters in the physical model (such as body mass, cutterhead torque coefficient, formation reaction modulus, etc.). For the residual compensation term of the lightweight multilayer perceptron, These are the weights and bias parameters of the neural network. This is the process noise term. , , The mean is Covariance is The multivariate Gaussian distribution; To observe the noise covariance matrix, Let be the process noise covariance matrix.

[0043] 4. Offline training process of the generative model:

[0044] (1) Collect historical tunneling data of multi-mode pipe jacking machine under typical strata (soft soil, sand layer, rock layer), including multi-source sensor sequence, control action sequence and post-measured settlement / attitude true value; clean the data, time-align it, and construct a supervised training set with the measured true value as the label.

[0045] Equations of rigid body dynamics:

[0046] Mξ¨+C(ξ˙)ξ˙+G(ξ)=τ act +τ ext ;

[0047] In the formula, M is the generalized mass matrix of the machine body, which includes translational mass and rotational inertia, determined by the equipment's factory parameters and assembly status; ξ¨ is the second-order time derivative of the pose, i.e., the generalized acceleration vector; ξ˙ is the first-order time derivative of the pose, i.e., the generalized velocity vector; ξ is the six-degree-of-freedom pose vector of the machine body, representing the three-axis displacement and three-axis Euler angles in the global coordinate system; C(ξ˙) is the Coriolis force and centrifugal force matrix, derived from rigid body kinematics, reflecting the velocity coupling effect; G(ξ) is the equivalent generalized force vector of gravity and buoyancy, which depends on the machine body's attitude, self-weight, and mud-water buoyancy distribution; τ act To actively control the generalized torque vector, it is obtained through geometric mapping from the actuator outputs such as jack thrust, cutter head torque, and articulated cylinder force; τ ext External disturbance generalized moment vectors include uncontrolled loads such as formation reaction force, mud-water pressure resultant force, and segment friction force.

[0048] Slurry balance equation:

[0049] P slurry =P earth +P water +ΔP margin ;

[0050] In the formula, P slurry The actual slurry pressure was measured in the excavation chamber; P earth The static lateral pressure of the soil in front of the excavation face; P water The hydrostatic pressure of groundwater; ΔP margin This represents the safety pressure margin and the positive offset set for the process.

[0051] Physical model parameter identification: Based on rigid body dynamics and slurry balance equations, the least squares or Bayesian identification method is used to estimate physical parameters such as body mass, cutterhead torque coefficient, and formation reaction modulus from historical data, which are used as fixed or adjustable initial values ​​in the state transition function.

[0052] (2) Based on the output of the physical model, train the residual term of the lightweight multilayer perceptron and the Gaussian process regression module in the observation function; the loss function adopts a weighted combination of negative logarithmic marginal likelihood and prediction error, while optimizing the learnable parameters of the observation function, the weights and biases of the neural network, and the noise covariance parameter; the optimizer is AdamW, with a learning rate of 1e-3, a batch size of 64, and 200 training rounds.

[0053] (3) Based on the statistical characteristics of the training set, set the Gaussian prior mean and covariance for the state variables; set the expected distribution parameters of indicators such as trajectory deviation and settlement risk for the target prior preference distribution; store all prior parameters in the controller as the initial belief for online variational inference and update.

[0054] (4) Evaluate the model prediction accuracy and free energy convergence on the reserved test set. Only when the predicted RMSE is lower than the engineering tolerance and the free energy decreases monotonically should the model parameters be fixed.

[0055] This invention achieves probabilistic perception and forward-looking prediction of complex latent states during multi-mode pipe jacking machine tunneling by fusing multi-source sensor data with an active inference and generative model. The introduction of observation and process noise explicitly characterizes the inherent uncertainties of underground engineering, enabling subsequent control decisions to naturally balance risk and performance. The embedding of prior preferences internalizes construction quality and safety requirements as part of the model, preventing a disconnect between optimization objectives and engineering needs. Simultaneously, the set of state variables defined based on the working condition mechanism and the parameterized generative model retain physical interpretability while possessing online learning and adaptive capabilities, providing a decision-making basis for progressive hierarchical optimization.

[0056] S2. Perform online updates on the generated model to match the current formation and multi-mode pipe jacking machine operating conditions, and output the updated active inference generated model.

[0057] In one embodiment, online updates are performed on the generated model to match the current formation and the multi-mode pipe jacking machine's operating conditions, outputting an updated active inference generated model. This includes: acquiring the most recent observation vector sequence of a preset length and the corresponding control action sequence within the current control cycle to form a data window; performing forward prediction on the data window based on the generated model and calculating the error between the predicted observation and the actual observation to generate a representation of the matching degree between the model and the current formation and operating conditions; determining whether to trigger an online update by comparing the error with a preset threshold; if no online update is triggered, keeping the active inference generated model parameters from the previous control cycle unchanged; if an online update is triggered, determining the update step size and regularization weights, and using recursive optimization to update the observation function parameters and state transition function parameters in the generated model, reducing the free energy of the updated generated model on the data window; performing boundary constraint verification and prediction consistency verification on the updated generated model; if the verification passes, outputting the updated active inference generated model; if the verification fails, reverting to the generated model parameters before the update and freezing the update for the current cycle.

[0058] This invention is based on the online variational Bayesian learning mechanism in active inference. The variational free energy gradient corresponding to the prediction error is calculated in real time through a sliding data window and used as a quantitative indicator of model mismatch to trigger condition updates. When the free energy exceeds the threshold, a regularized recursive optimization algorithm is used to incrementally adjust the parameters of the observation and state transition functions. This allows the posterior distribution to quickly adapt to the non-stationary dynamics of the current strata and working conditions while retaining historical prior knowledge. A safe projection operator is constructed through boundary constraints and prediction consistency verification to ensure that the parameter updates always fall within the physical feasible region and the numerical stable region. Otherwise, the system automatically backs down to maintain its robustness.

[0059] 1. Decision Basis: The historical states and control action sequences within the data window are input into the current state transition function and observation function. The theoretical observation value at each moment within the window is calculated recursively according to the time step. The difference between the theoretical observations generated by the model and the actual observations collected by the sensors is calculated moment by moment to obtain the prediction residual sequence. Under the active inference framework, this residual is the prediction error, reflecting the deviation between the internal representation of the model and the external real physical process. The residual sequence is aggregated into a scalar index through the precision-weighted squared error term in the weighted norm or variational free energy. The larger the index, the less accurately the model describes the current stratum characteristics (such as abrupt changes in soil stiffness) or changes in working conditions (such as cutterhead wear, abnormal grouting pressure), and vice versa. This scalar is then used as the decision basis for whether to trigger online updates and the update intensity.

[0060] 2. Threshold Acquisition: Before the pipe jacking machine starts or during the trial excavation stage in typical strata, the generative model is fully trained using historical high-quality datasets, and the distribution of prediction errors of the model under normal matching conditions (such as mean) is statistically analyzed. with standard deviation ), set the initial value of the threshold to ( (Take values ​​2-3). During actual tunneling, a sliding window is used to continuously monitor the statistical characteristics of the error sequence. When a change in stratum type is detected (such as moving from soft soil to rock) or a change in working mode is detected, the error baseline under the current steady state is automatically re-estimated and the threshold is dynamically adjusted. For example, the error variance is tracked by an exponentially weighted moving average, so that the threshold scales with changes in environmental noise levels, avoiding frequent invalid updates due to excessively low thresholds in high-noise hard rock sections, or missed model drift due to excessively high thresholds in low-noise precision control sections. The final value of the threshold also needs to be superimposed with the output of the physical safety verification layer, that is, combined with engineering constraints such as the current attitude deviation tolerance and settlement warning level, the statistical threshold is truncated at the upper and lower limits; when the safety risk increases, the threshold is actively lowered to improve the sensitivity of model updates, and when the system is in a stable and safe zone, the threshold is appropriately relaxed to reduce the computational load.

[0061] In online variational updates, the regularization term uses KL divergence constraints to measure the deviation between the current posterior and the previous posterior. Boundary constraints are validated against the updated parameters. The physical feasible region must be satisfied. Otherwise, project to the nearest feasible point. Prediction consistency verification is used... Regarding the past If the mean square error increases by more than 20% compared to the previous update, the update is deemed invalid.

[0062] The generative model provides adaptive online calibration for multi-mode pipe jacking machines traversing complex and variable strata. Threshold triggering and regularization mechanisms prevent computational burden and parameter oscillations caused by overfitting noise or frequent invalid updates. Dual boundary and consistency checks effectively eliminate the risk of model collapse due to data anomalies or optimization divergence.

[0063] S3. Using the active inference generation model, variational inference is performed within the current control cycle to calculate the posterior distribution of state variables. Combined with equipment and process boundary conditions, a set of constrained candidate control actions is generated. At the same time, the active inference generation model is used to perform forward simulation prediction on the constrained candidate control actions to obtain a set of multi-action prediction results.

[0064] In one embodiment, an active inference generative model is used to perform variational inference within the current control cycle to calculate the posterior distribution of state variables. Combined with equipment and process boundary conditions, a constrained set of candidate control actions is generated. This includes: performing prior propagation of the current state distribution under the previous control action conditions based on the state transition function to obtain the predicted prior distribution of state variables; constructing a variational distribution and calculating the observation likelihood based on the observation function, and iteratively updating it by minimizing the variational free energy to output the posterior distribution of state variables; obtaining the multi-mode pipe jacking machine's capability boundary and process safety boundary to form a constraint set; constraints include upper and lower limits and rate of change constraints for thrust / torque / cutterhead speed / propulsion speed / grouting pressure / bladder pressure; constructing uncertainty constraint margins based on the posterior and predicted prior distributions of state variables to modify the basic constraint set into a probabilistically feasible constraint set for the current control cycle; generating a set of candidate control actions within the probabilistically feasible constraint set based on the controllable variable dimension and discrete control strategy, and performing feasibility screening on the candidate actions to obtain a constrained set of candidate control actions, while structuring and parameterizing the constrained set of candidate control actions.

[0065] In one embodiment, an active inference generative model is used to perform forward simulation prediction of constrained candidate control actions to obtain a set of multi-action prediction results. This includes: indexing and temporally expanding the set of constrained candidate control actions to form the action sequence of each candidate action in the prediction time domain; using the current state posterior distribution as the initial condition, recursively calculating the predicted state distribution based on the state transition function under the condition of the action sequence of each candidate action in the prediction time domain; mapping the predicted state distribution to the predicted observation distribution based on the observation function, and incorporating process noise and observation noise into uncertainty propagation; and collecting the predicted state trajectory, predicted observation trajectory, and uncertainty corresponding to each candidate action to form a set of multi-action prediction results.

[0066] This invention integrates prior dynamic predictions with current observational likelihoods by minimizing variational free energy, obtaining a posterior distribution that accurately characterizes the hidden state and its uncertainties of the system. This posterior uncertainty is mapped to the constraint space of equipment capability and process safety, constructing a probabilistic feasible region to dynamically shrink or expand the allowable range of control actions, and discretizing within this range to generate a set of structured candidate actions. Using the posterior state as initial conditions, a forward simulation considering the uncertainty propagation of the process and observation noise is performed on each candidate action sequence, outputting a set of multi-action results including predicted trajectories, observations, and confidence intervals. This directly embeds the uncertainty of state estimation into the evaluation of future control effects, achieving a closed-loop deduction from current understanding to future expectations.

[0067] 1. Prior Propagation: Using the posterior distribution of the state at the previous moment as initial conditions, this distribution is substituted into the state transition function. Through approximation algorithms such as probability integration or moment matching, the predicted prior distribution of the current state is derived. This distribution represents the theoretical expectation of the generative model regarding the current state of the system, based solely on dynamic laws and historical control actions. It does not incorporate any new observational information, and its uncertainty typically increases compared to the previous moment due to the accumulation of process noise.

[0068] 2. Likelihood Construction and Variational Update: After obtaining the actual observations at the current time, the observation likelihood is calculated based on the observation function, quantifying the probability density of generating the current observation given the hidden state. Since the exact posterior cannot be analytically solved in high-dimensional nonlinear systems, a parameterized variational distribution is introduced as an approximation of the posterior, and the variational free energy is minimized through iterative optimization. In the formula, The Kullback-Leibler divergence; The variational distribution represents the hidden state at the current time step. The parameterized approximation of the true posterior distribution is usually chosen as a Gaussian distribution or other analytically commensurable family of distributions, and its parameters (such as mean and covariance) are variables for iterative optimization. Predict the prior distribution for the state; For variational distribution The mathematical expectation, representing the expectation of the function within the parentheses. Integral averaging is performed over the defined probability space; The objective function is the logarithm of the observation likelihood. The first term constrains the variational distribution to not deviate too far from the dynamic prior, while the second term drives the variational distribution to concentrate in the high-likelihood region. The balance between these two terms yields the optimal variational posterior, which is equivalent to an analytical approximation that best approximates the true posterior in an information geometry sense. The optimal variational posterior simultaneously encodes the point estimates of the state (e.g., mean) and their cognitive uncertainties (e.g., covariance). When the observed information is consistent with the prior, the posterior uncertainty is less than the prior, indicating improved model confidence. When the two conflict, the posterior weighs its respective accuracy, neither blindly trusting potentially noisy sensor data nor adhering to a potentially drifted dynamic model. This posterior serves not only as the state input for the current cycle's control decision but also as the starting point for prior propagation in the next cycle, forming a recursive closed loop.

[0069] 3. Construction of Uncertainty Constraint Margin and Correction of Probabilistic Feasibility Region: The basic constraint set is usually a deterministic hard boundary determined by the equipment's rated parameters or specifications. When there is uncertainty in the state estimation, directly using this hard boundary may lead to the actual physical quantity exceeding the limit. Therefore, the statistical distance between the posterior distribution and the predicted prior (such as KL divergence or covariance trace) is used to quantify the level of cognitive uncertainty at the current moment. The higher the uncertainty, the larger the constraint margin is constructed. By shrinking the basic constraints inward, a probabilistic feasible constraint set is formed. This set ensures that when performing control actions within this range, the probability of the state violating the true physical boundary is lower than a preset risk threshold (such as...). This enables a risk-aware dynamic security envelope.

[0070] 4. Generation and Feasibility Screening of Candidate Control Actions: After obtaining the probabilistically feasible constraint set, sampling or enumeration is performed within the contracted feasible domain, combining controllable variable dimensions (such as independent control channels for thrust, speed, and grouting pressure) and predefined discrete control strategies (such as graded adjustment tables and mode switching logic), to generate an initial set of candidate actions. Simultaneously, rigorous feasibility screening is conducted: on the one hand, it checks whether the action itself satisfies the probabilistically feasible constraints (i.e., whether it is within the dynamic safety envelope); on the other hand, it verifies the coupling compatibility between actions (e.g., whether the combination of high torque and low speed will cause motor overload) and whether the rate of change of the action relative to the previous moment exceeds the limit. Only actions that pass both verifications are retained as constrained candidate control actions, ensuring that each candidate instruction is physically executable and safe under the current cognitive state.

[0071] 5. Structuring and Parameterizing Candidate Action Sets: Original candidate actions are often unordered sets of points or lists, making them difficult to use directly for subsequent forward simulation and optimization evaluation. Therefore, they are structured, for example, by classifying and indexing them according to control modes (such as attitude correction mode, settlement control mode, and uniform speed tunneling mode), or by arranging them according to their temporal sequence within the prediction time domain to form an action sequence matrix. Simultaneously, parameterization is performed, mapping discrete gear positions or switching quantities to continuous numerical vectors, and attaching metadata tags, such as expected energy consumption, risk level, and applicable working condition descriptions.

[0072] 6. Indexing and Temporal Expansion of Action Sequences: The constrained set of candidate control actions represents only single-step or static control options at the current moment. To evaluate their dynamic effects, a unique index needs to be created for each candidate action in the set for parallel processing, based on a preset prediction temporal domain. And control strategies (such as maintaining a constant, linear transition, or following a specific curve) extend single-step actions to lengths of... Timing action sequence This process transforms discrete decision points into continuous control trajectories, enabling forward simulation to capture the cumulative effects and dynamic response characteristics of control actions over time, rather than merely evaluating instantaneous state changes.

[0073] 7. Uncertainty Propagation Recursion Based on Posterior Initial Conditions: The posterior distribution of the state after variational inference correction at the current moment is used as the initial probability condition for the simulation (rather than a single deterministic value), ensuring consistency between the prediction starting point and the actual cognitive state. At each prediction step... Substituting the predicted state distribution from the previous step into the state transition function, and using approximation algorithms such as moment matching, unscented transformation, or particle filtering, the predicted state distribution for the next time step is recursively calculated. During this process, process noise... The covariance is gradually accumulated, causing the uncertainty of the predicted state to naturally diffuse as the prediction time domain increases. At the same time, the state distribution is mapped to the predicted observation distribution through the observation function, incorporating the influence of observation noise. This complete uncertainty propagation mechanism means that the prediction result not only gives the most likely future trajectory, but also quantifies the decay of confidence in the future.

[0074] 8. Structured Collection of Multi-Action Prediction Results: After parallel recursion of all candidate action sequences, the mean and covariance of the predicted state trajectory, the mean and covariance of the predicted observation trajectory, and the cumulative uncertainty indicators (such as prediction entropy or trajectory envelope width) corresponding to each candidate action are packaged into structured data units and organized into a multi-action prediction result set according to the original index. This set is a three-dimensional mapping table of control, expectation, and risk: each row represents a candidate action, and each column represents its expected performance and reliability at various future times. This provides a complete information foundation for the calculation of expected free energy or strategy selection in subsequent active inference. That is, the decision module can not only compare which action brings the best expected performance, but also identify which action's prediction is most reliable and which action has a high risk of divergence in the future, realizing adaptive control decision-making that balances performance, safety, and exploration.

[0075] This invention explicitly transforms state estimation uncertainty into constraint margins, enabling the candidate control action set to adaptively and dynamically adjust with model confidence. When the model has high confidence, it fully utilizes the equipment capability boundary to improve tunneling efficiency, while automatically tightening the feasible region to mitigate safety risks when the model has low confidence. This effectively solves the problem that traditional deterministic constraints in complex geological formations can easily lead to conservatism or exceeding limits. Simultaneously, forward simulation prediction based on posterior distribution not only provides the expected performance indicators of each candidate action but also quantifies the reliability of its prediction results, providing a complete information foundation for subsequent control decisions that balances optimality and robustness. This enhances the adaptability and safety of multi-mode pipe jacking machines in non-stationary conditions for attitude control and settlement management.

[0076] S4. Configure the target prior preference and calculate the variational free energy of candidate control actions in the multi-action prediction result set to form an action-variable free energy mapping; construct the progressive control hierarchy and threshold rules, and determine the control hierarchy and constraint strategy of the action based on the action-variable free energy mapping and the threshold decision quantity corresponding to the posterior distribution of the state variables.

[0077] In one embodiment, a target prior preference is configured, and the variational free energy of candidate control actions in the multi-action prediction result set is calculated to form a mapping between actions and variational free energy. This includes: obtaining preference indices corresponding to the tunneling target and safety constraint target, including trajectory deviation, attitude deviation, surface settlement risk, pressure stability, energy consumption, and construction efficiency; converting the preference indices into a target prior preference distribution; aligning the prediction distributions of each candidate control action in the multi-action prediction result set with the target prior preference distribution on the same variable space and time index; and calculating the variational free energy of each candidate control action in the prediction time domain.

[0078] ;

[0079] In the formula, Candidate actions The scalar evaluation value of the variational free energy in the prediction time domain. For the first One candidate control action, For the future moment The observed variables, In order to execute candidate actions Under these conditions, the predicted observation distribution obtained from forward simulation by the active inference generation model, For the future moment The target prior preference distribution At the current control cycle time, To predict the length of the time domain, Let KL divergence be the KL divergence. Here, the cost function is used to characterize the execution cost of an action. These are weighting coefficients; after truncating, smoothing, or normalizing the variational free energy, a mapping relationship between candidate control actions and corresponding variational free energies is formed.

[0080] In one embodiment, a progressive control hierarchy and threshold rules are constructed, and the control hierarchy and constraint strategy for actions are determined based on the action-variable free energy mapping and the threshold decision quantity corresponding to the posterior distribution of state variables. This includes: constructing a progressive control hierarchy corresponding to attitude correction, settlement disturbance suppression, silo pressure stabilization and surge pressure loss risk suppression, energy consumption wear and continuity optimization, and presetting corresponding constraint strategies for each control hierarchy; constructing threshold rules and threshold decision quantities based on the posterior distribution of key state variables; comparing the threshold decision quantities with the equipment process boundary according to the threshold rules to output the risk level; extracting free energy evaluation indicators based on the action-variable free energy mapping; jointly determining the control hierarchy of the current control cycle by combining the risk level and the free energy evaluation indicators; and applying corresponding constraint strategies to the candidate action set according to the determined control hierarchy to generate hierarchical action constraint rules, including action upper and lower limits, rate of change limits, priority / disabled sets, and safety backoff actions when necessary.

[0081] In one embodiment, the control level for the current control cycle is determined by jointly judging the risk level and the free energy evaluation index. This includes: obtaining a recommended control level based on the risk level according to a pre-set risk level and control level comparison table; extracting the free energy values ​​of all candidate actions in the current cycle from the action-variable free energy mapping results and calculating the free energy determination quantity; obtaining a recommended control level based on free energy according to the pre-set free energy determination quantity and control level rules; and using the principles of risk priority and conservatism priority, jointly comparing the recommended control level based on the risk level and the recommended control level based on free energy to give the control level for the current control cycle.

[0082] This invention transforms multi-objective engineering decision-making into a variational inference problem in a probability space and achieves an adaptive trade-off between safety and performance through a hierarchical threshold mechanism. Heterogeneous preference indicators such as trajectory deviation, settlement risk, and energy consumption are uniformly encoded into a target prior preference distribution, aligning it with the predicted observation distribution output by the generative model in the same variable space. By calculating the KL divergence between the two in the prediction time domain and superimposing the execution cost, a scalarized variational free energy is obtained. The smaller this value, the more the candidate action conforms to the comprehensive target preference and the lower the execution cost. Based on this, threshold rules are triggered according to key statistics of the state posterior distribution (such as mean, variance, or tail probability) to determine the current risk level. Combined with the free energy evaluation index, the control level is dynamically selected according to the principles of risk priority and conservatism priority, from emergency safety suppression to normal efficiency optimization. Corresponding constraint strategies, such as amplitude limiting, prohibition, or backoff, are applied to candidate actions at this level, forming a closed-loop hierarchical decision-making logic.

[0083] 1. Transform the multi-objective engineering decision problem into a variational inference and energy minimization problem in probability space:

[0084] The overall merits of candidate actions are quantified by using a unified target prior preference distribution. This involves encoding heterogeneous and dimensionlessly variable engineering indicators such as trajectory deviation, settlement risk, and energy consumption into a target prior preference distribution defined in the observation variable space. The high-probability region of this distribution represents the ideal state the system aims to achieve. KL divergence is used to measure the deviation between the predicted observation distribution generated by each candidate action and this preference distribution, and this deviation is then applied in the prediction time domain. The deviation and execution cost are summed to obtain the scalarized variational free energy. This is a comprehensive loss function: the KL term measures the degree of achievement and safety of the action towards the future goal; the more overlapping the distributions, the lower the free energy. The cost term penalizes uneconomical behaviors such as high energy consumption or drastic operations. The weighting coefficients... This is used to balance the two. By truncating, smoothing, or normalizing the free energy, the effects of numerical overflow or scale differences are eliminated, forming a stable and comparable mapping relationship between motion and free energy.

[0085] 2. Risk-driven, performance-guided adaptive hierarchical decision-making mechanism:

[0086] The engineering control problem is decoupled into a progressive logical structure with clearly defined priorities. That is, system safety and stability have absolute priority, while efficiency and energy consumption optimization are only activated when the safety margin is sufficient. Specifically, a progressive control hierarchy is predefined, ranging from emergency safety suppression (such as surge pressure loss) to normal performance optimization (such as energy consumption and wear), and each level is bound to a specific constraint strategy, such as mandatory limiting at higher levels and allowing exploration at lower levels. The statistical characteristics of the posterior distribution of the state (such as the probability of exceeding the mean limit and the variance abnormality) are used as threshold quantification, and the risk level is output in real time by comparing it with the equipment process boundary.

[0087] The mechanism extracts free energy evaluation metrics, such as minimum free energy value or free energy distribution entropy, reflecting the quality of the current optimal action from the mapping between actions and free energy. A joint judgment is made using the principles of risk priority and conservatism priority: when the risk level is high, regardless of the free energy evaluation, a forced switch to the corresponding high-priority safety level is implemented; only when the risk is controllable and the free energy metric indicates that the current strategy is effective is the strategy downgraded to the performance optimization level. This mechanism ensures that the candidate action set is always subject to the most appropriate constraints under the current cognitive state, avoiding efficiency losses caused by excessive conservatism under low risk and preventing safety incidents caused by blind optimization under high risk, thus achieving a dynamic adaptive balance between the safety baseline and the performance ceiling.

[0088] Threshold decision quantity:

[0089] (1) For each key state variable For example, pressure, attitude deviation angle, and equivalent settlement index, from their variational posterior distribution Extract the 99th upper quantile As a risk-sensitive indicator.

[0090] 2. Equipment process boundary For dynamic values: ,in For nominal safety limits, The working condition reduction factor is provided in real time by the geological identification module.

[0091] 3. Risk Level Defined as: That is, take the maximum value of the degree of exceeding the limit for each state; when It was determined to be high risk at that time. Medium risk. Low risk.

[0092] 4. For coupled conditions (such as settlement risk), the joint exceedance probability is calculated using posterior samples and incorporated as an independent decision metric in the risk level calculation. For example, the joint exceedance probability is calculated using posterior samples:

[0093] ;

[0094] In the formula, For the state variable of pressure in the excavation chamber or slurry chamber, This is the safe process boundary threshold (dynamic value) for silo pressure. Exceeding this value indicates that the silo pressure is in a risky state. For slag discharge state variables, such as the ratio of unit slag discharge volume to theoretical slag discharge volume, This is the safe process boundary threshold (dynamic value) for slag discharge. Exceeding this value is considered to pose a risk of over-excavation or formation loss.

[0095] 3. Dual-channel redundancy check and priority arbitration mechanism:

[0096] This mechanism ensures the absolute safety and logical completeness of control-level decisions in uncertain environments. It decouples external physical risk from internal cognitive confidence into two independent evaluation dimensions: on the one hand, it directly maps the objective degree of state exceeding limits to a safety catch-all level through a risk level comparison table; on the other hand, it quantifies the cognitive confidence of the current generative model in candidate strategies through free energy determination quantities (such as minimum free energy value, free energy variance, or optimal action proportion), reflecting the convergence quality of the active inference process. In joint determination, the principle of risk priority and conservatism priority is followed; that is, when the recommended levels of the two are inconsistent, the higher priority (more conservative) level is always selected as the final output. The introduction of the free energy channel avoids frequent level jumps caused by relying solely on hard thresholds. When the risk is in a critical ambiguity zone but the free energy indicates a highly credible strategy, a lower level can be appropriately maintained to preserve control continuity, thus achieving a dynamic balance between safety rigidity and optimization flexibility based on the strength of evidence.

[0097] In the progressive control system, the levels are priority sequences ordered according to the urgency of safety and the strength of constraints: the higher levels correspond to emergency states that endanger the safety of the project, such as silo instability and excessive settlement, and have absolute priority and rigid constraints. Once triggered, they will be forcibly taken over to ensure the bottom line safety; the lower levels correspond to performance objectives such as energy consumption optimization and continuous improvement, and are only activated when the risk is controllable and the model confidence is high, allowing for relaxed exploration to pursue comprehensive benefits.

[0098] This invention achieves intelligent decoupling and safety safeguards for multi-objective conflicts under complex working conditions, improving the robustness and adaptability of the autonomous control of the pipe jacking machine. By internalizing engineering experience preferences into a differentiable probability distribution, it avoids the shortcomings of traditional weighted summation methods, such as the difficulty in adjusting weights and the inability to express distributional preferences, making decisions more aligned with actual construction intentions. The progressive control hierarchy and joint decision-making mechanism ensure that the system automatically converges to the safety constraint set under high-risk conditions, preventing the pursuit of optimal performance from ignoring potential instability or equipment damage. At the same time, it fully releases the optimization space under low-risk normal conditions to balance efficiency and energy consumption, thereby achieving a dynamic optimal balance of safety, compliance, and economy across the entire working condition range.

[0099] S5. Based on equipment process constraints and the control hierarchy and constraint strategy of actions, a hierarchical optimization problem with the goal of minimizing variational free energy is constructed. Candidate control actions in the mapping between actions and variational free energy are screened and solved to generate the optimal control action for the current control cycle, which is used to control the construction of the multi-mode pipe jacking machine.

[0100] In one embodiment, based on equipment process constraints and the control hierarchy and constraint strategies of actions, a hierarchical optimization problem is constructed with the goal of minimizing variational free energy. Candidate control actions in the action-variable free energy mapping are screened and solved to generate the optimal control action for the current control cycle. This includes: obtaining the control hierarchy and corresponding constraint strategies determined for the current control cycle, merging the equipment process constraints and constraint strategies to construct a feasible domain constraint set for the current control cycle; constructing a hierarchical objective function for the hierarchical optimization problem with the goal of minimizing variational free energy, setting hierarchical weights according to the control hierarchy; performing feasibility checks on each candidate control action in the action-variable free energy mapping, eliminating candidate control actions that do not satisfy the feasible domain constraint set, and obtaining a subset of feasible candidate actions; calculating hierarchical free energy evaluation values ​​for the subset of feasible candidate actions based on the hierarchical objective function, and comparing and sorting them according to hierarchical priority to obtain a set of candidate actions with the minimum free energy; and performing local continuous optimization in the neighborhood of the candidate action set to obtain the optimal control action that satisfies the feasible domain constraint set.

[0101] This invention transforms control decision-making into a constrained hierarchical variational free energy minimization optimization problem. It integrates the inherent process boundary of the equipment with the current level-specific constraint strategy to form a dynamic feasible region, and strictly eliminates infeasible candidate actions. It constructs an objective function with free energy minimization as the overall goal and weights it according to hierarchical priority. It then performs hierarchical sorting and screening of the remaining feasible solutions to select a low free energy candidate set. It performs local continuous optimization in the neighborhood of the optimal candidate and solves the actual control command that takes into account both safety level and cognitive optimality under the premise of satisfying all hard and soft constraints.

[0102] Hierarchical objective function: ;

[0103] In the formula, These are candidate control actions to be optimized. For the first Each control level corresponds to a variational free energy component. Each level (e.g., safety level, stability level, performance level) has an independent generative model or predictor, therefore its free energy calculation is based on state observations specific to that level. Generate model parameters at the current level This reflects the prediction error of actions at that level; The hierarchical weights are dynamically determined by the control level determined by the current control cycle. When the control level points to a higher level (such as when a safety risk is triggered), the corresponding weights of the higher levels are set to a maximum value (or normalized to close to 1), while the weights of the lower levels are suppressed to close to 0, so that the objective function degenerates into minimizing a single safety free energy. When the control level points to a lower level and the free energy evaluation is reliable, the weights of each level are allocated according to a preset ratio, allowing performance optimization terms to participate in the decision. The weights can also include a conservative priority adjustment factor, that is, when the free energy itself is too high (the model is uncertain), the weights of the higher levels are automatically increased to avoid risks. For the final hierarchical overall objective function, the optimization solution is to find the one that makes the objective function more hierarchical. Minimal action .

[0104] Hierarchical weights Determine as follows:

[0105] 1. Let the set of control levels be... These correspond to attitude correction, settlement suppression, pressure stabilization, and energy consumption optimization, respectively, with priority decreasing in that order.

[0106] 2. Based on the current control level Initialize the basic weight vector; calculate the free energy confidence index. ,in, The baseline value of free energy under historical normal operating conditions is the median of the free energy over the past 50 steps (to resist abnormal disturbances); if , If the free energy anomaly threshold is set, preferably 1.5, then a conservative correction is applied to the base weights. Based on the originally set base weights, differentiated adjustments are applied to different levels:

[0107] For levels strongly correlated with safety, their weight ratios are appropriately increased; for levels focusing on efficiency optimization, their weight ratios are correspondingly decreased; the remaining levels remain largely unchanged or are slightly adjusted. The adjustment magnitude increases linearly with the risk level, meaning the greater the excess of free energy, the more significant the adjustment, but the upper limit is set at 30% to prevent weight imbalance from leading to control rigidity. After adjusting the weights of all levels, they are renormalized according to the principle of a sum of 1 to ensure that the weights of each level still constitute an effective probabilistic allocation coefficient, which can be used for the weighted synthesis of the subsequent objective function.

[0108] In this invention, local continuous optimization employs the projection gradient method combined with an adaptive step size:

[0109] 1. Rank the discrete candidate actions with the best ranking. Define the neighborhood around the center. ,in For Markov normal, The control sub-block is taken from the state posterior covariance matrix. The radius is preferably 0.3 times the operating range of the equipment.

[0110] 2. Objective function The KL divergence term is transformed into a differentiable form using a reparameterization technique, and the gradient is computed by an automatic differentiation framework (such as PyTorch / JAX).

[0111] 3. Execute each iteration: ,in To the feasible region constraint set The projection operator is solved using quadratic programming. Determined by Armijo line search, This is the gradient operator for the candidate control action vector.

[0112] 4. The maximum number of iterations is set to 20. If convergence or projection fails within the time limit, the output will be directly updated. The optimal action for this cycle is marked as optimized and degraded for monitoring by the upper layer; the input is pruned before all gradient calculations.

[0113] This invention employs a two-stage mechanism of discrete screening followed by continuous optimization. This mechanism ensures absolute safety and compliance of actions under high-risk conditions through hierarchical constraints, while minimizing free energy guarantees the optimality of the strategy and the self-consistency of the model under low-risk conditions. This effectively avoids the shortcomings of traditional single-objective optimization, which is prone to getting trapped in local extrema or violating safety boundaries under complex working conditions. At the same time, local continuous optimization compensates for the insufficient resolution of the discrete candidate set, improving the smoothness, robustness, and adaptability of the multi-mode pipe jacking machine's control actions in non-stationary geological environments.

[0114] This invention collects multi-source sensor and construction data in each control cycle, including attitude or guidance deviation, propulsion cylinder thrust and distribution, propulsion speed, cutterhead torque and speed, slurry pump flow rate and differential pressure, screw conveyor speed and torque, excavation bin or slurry bin pressure and its fluctuations, and surface settlement or equivalent early warning data, forming an observation vector. Based on active inference, a generative model of state, observation, and control is established to predict attitude, bin pressure, slag discharge, and settlement risks under different control actions, and the variational free energy is calculated to characterize the prediction error and deviation from prior preferences. A progressive control hierarchy is set: for example, attitude correction, settlement or disturbance suppression, bin pressure stabilization and suppression of gushing or pressure loss risks, energy consumption and continuity optimization, and the hierarchy is switched according to thresholds; when the free energy exceeds the upper limit or the attitude, settlement, or bin pressure indicators exceed the limit, it reverts to the previous or first level, while applying conservative constraints. Minimize the free energy generation control quantities under equipment and process constraints, including cylinder differential or distribution, propulsion parameters, cutterhead parameters, slurry pump and screw compressor parameters, and agitator or modifier injection quantity, to achieve multi-objective progressive optimization and robust control under formation abrupt changes, and execute cyclically.

[0115] In this invention, pipe jacking machines are divided into two types: earth pressure balance pipe jacking machines and slurry balance pipe jacking machines. The earth pressure balance pipe jacking machine balances the water and soil pressure at the excavation face by adjusting the amount of excavated soil accumulated in the soil chamber through changes in the pipe jacking machine's advancing speed and the screw conveyor's slag discharge speed. The slurry balance pipe jacking machine balances the water and soil pressure at the excavation face through the slurry pressure in the pressure regulating chamber. The overall principle for selecting a pipe jacking machine is: focusing on the stability of the excavation face, taking engineering geology and hydrogeology as fundamental points; basing it on stratum particle size, permeability coefficient, and groundwater pressure, and comprehensively considering the specific engineering conditions; ensuring that the selected pipe jacking machine meets the overall goals of being stable (balancing the working face), advancing (cutting the working face), and effectively discharging (discharging excavated soil from the soil chamber).

[0116] Selection based on formation grain size:

[0117] The relationship between the type of earth pressure balance pipe jacking machine and the formation particle size when no soil improvement is performed and when no additives are used in slurry pipe jacking machines, such as... Figure 3 As shown.

[0118] Earth pressure balance (EPB) pipe jacking machines are primarily suitable for construction in silty soils such as silty clay, silty mud, and silty sand, with a particle size range of less than 1.5 mm. Slurry balance (SBP) pipe jacking machines are mainly suitable for construction in sandy, gravelly, and pebble soils, with a particle size range of 0.01–40 mm. Generally, when the total amount of silt and clay particles in the soil or rock reaches more than 40%, an EPB pipe jacking machine is usually chosen; otherwise, a SBP pipe jacking machine is more suitable. However, if the EPB pipe jacking machine is used for soil improvement or if the SBP pipe jacking machine uses appropriate additives, then the applicable ranges for both EPB and SBP pipe jacking machines are the same.

[0119] Selection based on permeability coefficient:

[0120] Selection based on permeability coefficient, such as Figure 4 As shown, when the permeability coefficient of the stratum is less than 10 -7 When the permeability coefficient of the stratum is less than 10 m / s, an earth pressure balance pipe jacking machine should be used; -4 When the permeability coefficient is m / s, a slurry-balanced pipe jacking machine is recommended; when the permeability coefficient is 10... - 7 m / s ~ 10 -4 When the permeability is between m / s, either earth pressure balance pipe jacking machines or slurry balance pipe jacking machines can be used. When selecting a pipe jacking machine based on the permeability coefficient, the specific engineering geological conditions should also be considered. When the clay content in the stratum is less than 10%, it is difficult for a mud film to form, and the excavation face is prone to collapse. In this case, slurry balance pipe jacking machines are not suitable, and earth pressure balance pipe jacking machines should be used instead.

[0121] Model selection based on groundwater pressure:

[0122] Earth pressure balance pipe jacking machines use screw conveyors for muck removal. The pressure in the formation gradually decreases after passing through the soil chamber and screw conveyor, and needs to drop to atmospheric pressure before reaching the muck discharge port of the screw conveyor; otherwise, a blowout will occur. In contrast, slurry balance pipe jacking machines have a mud film at the front to prevent water loss from the formation, and a mud pump at the rear to maintain pressure during muck removal. Therefore, slurry balance pipe jacking machines are suitable for formations with high water pressure and high permeability.

[0123] When selecting a jacking machine based on groundwater pressure, generally speaking, when the groundwater pressure is less than 0.3 MPa, an earth pressure balance jacking machine is preferable; when the groundwater pressure is greater than 0.3 MPa, a slurry balance jacking machine is preferable. However, the specific engineering geological conditions should also be considered during selection: when the groundwater pressure is greater than 0.3 MPa, if an earth pressure balance jacking machine is required due to geological reasons, the length of the screw conveyor needs to be increased, or a two-stage screw conveyor should be used; when the soil improvement effect cannot meet the requirements of the soil plugging effect, an earth pressure balance jacking machine should not be used even when the groundwater pressure is less than 0.3 MPa.

[0124] Table 1 Comparison of Characteristics of Pipe Jacking Machine Types

[0125]

[0126] The advantages of earth pressure balance (EPB) pipe jacking machines are strong muck removal capacity, intuitive soil chamber pressure balance, and relatively simple equipment and operation. The disadvantages are poor adaptability to high water pressure and difficulty in preventing water loss in highly permeable formations. Slurry balance pipe jacking machines have a significant advantage over EPB machines in preventing water loss and can operate in high water pressure and high permeability formations. The disadvantages are that clay formations are prone to mud cake formation, blockage, and difficulty in separation, and sluggish discharge in formations with large amounts of particle size. Furthermore, slurry balance pipe jacking machines are generally of the direct discharge type, making it difficult to maintain slurry chamber pressure in unstable formations, resulting in weaker settlement control compared to EPB machines. Additionally, slurry balance pipe jacking machines have higher procurement costs and site requirements compared to EPB machines.

[0127] The multi-mode pipe jacking machine is equipped with a composite cutterhead, featuring 15.5-inch roller cutters, scrapers, and tearing blades, enabling it to adapt to various geological conditions, including hard rock (uniaxial compressive strength 180MPa), soft soil, composite strata, pebbles, and gravel, truly achieving all-stratum adaptability. The mixing tank has the capability to mix and secondary crush large-diameter soil and rock lumps, and can further crush large-diameter stones from the screw compressor to meet the discharge particle size requirements of the slurry pump, preventing pipe blockage and stagnation. The cutterhead, as... Figure 5 As shown.

[0128] like Figure 2 As shown, according to another embodiment of the present invention, a multi-mode pipe jacking machine tunneling attitude control construction system based on progressive hierarchical optimization is provided, comprising:

[0129] The observation-generated modeling module is used to construct observation vectors based on multi-source sensing and construction data during the tunneling process of the multi-mode pipe jacking machine. Combined with active inference theory, it constructs a generative model to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control action conditions.

[0130] The online model update module is used to perform online updates on the generated model to match the current formation and the working conditions of the multi-mode pipe jacking machine, and output the updated active inference generated model.

[0131] The inference constraint prediction module is used to perform variational inference within the current control cycle using an active inference generation model to calculate the posterior distribution of state variables, and generate a set of constrained candidate control actions by combining equipment and process boundary conditions; at the same time, it uses the active inference generation model to perform forward simulation prediction on the constrained candidate control actions to obtain a set of multi-action prediction results.

[0132] The free energy hierarchical module is used to configure the target prior preference and calculate the variational free energy of candidate control actions in the multi-action prediction result set to form an action-variable free energy mapping; it constructs a progressive control hierarchy and threshold rules, and determines the control hierarchy and constraint strategy of the action based on the action-variable free energy mapping and the threshold decision quantity corresponding to the posterior distribution of the state variables.

[0133] The hierarchical optimization solution module is used to construct a hierarchical optimization problem with the goal of minimizing variational free energy based on equipment process constraints and the control hierarchy and constraint strategies of actions. It filters and solves candidate control actions in the mapping between actions and variational free energy, and generates the optimal control action for the current control cycle, which is used to control the construction of multi-mode pipe jacking machines.

[0134] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization, characterized in that, include: Based on multi-source sensing and construction data during the tunneling process of the multi-mode pipe jacking machine, an observation vector is constructed. Combined with active inference theory, a generative model is built to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control action conditions. The generated model is updated online to match the current formation and the multi-mode pipe jacking machine operating conditions, and the updated active inference generated model is output. Using an active inference generative model, variational inference is performed within the current control cycle to calculate the posterior distribution of state variables. Combined with equipment and process boundary conditions, a set of constrained candidate control actions is generated. At the same time, the active inference generative model is used to perform forward simulation prediction on the constrained candidate control actions to obtain a set of multi-action prediction results. Configure the target prior preference and calculate the variational free energy of the candidate control actions in the multi-action prediction result set to form an action-variable free energy mapping; Construct a progressive control hierarchy and threshold rules, and determine the control hierarchy and constraint strategy of the action based on the action-variable free energy mapping and the threshold decision quantity corresponding to the posterior distribution of the state variable. Based on equipment process constraints and the control hierarchy and constraint strategy of actions, a hierarchical optimization problem with the goal of minimizing variational free energy is constructed. Candidate control actions in the mapping between actions and variational free energy are screened and solved to generate the optimal control action for the current control cycle, which is used to control the construction of multi-mode pipe jacking machine.

2. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The observation vector is constructed based on multi-source sensing and construction data during the multi-mode pipe jacking machine tunneling process, including: Multi-source sensor and construction data are collected in each control cycle, and the multi-source sensor and construction data are synchronized in time and anomaly removal is performed. After time synchronization and anomaly removal, calculate the characteristic quantities used to characterize the current working condition of the data, and perform dimensional unification and normalization on each characteristic quantity. The normalized features are concatenated in a preset order to form the observation vector for the current control cycle.

3. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The aforementioned generative model, constructed by combining active inference theory to characterize the relationship between state variables and observation vectors and to predict state evolution under candidate control actions, includes: Based on the attitude control objectives and working mechanism of the multi-mode pipe jacking machine, the set of state variables of the generated model is determined. The state variables include attitude state, guidance deviation state, silo pressure state, slag discharge state and settlement risk state. An observation function is established to represent the mapping relationship between state variables and observation vectors, and an observation noise term is introduced to characterize sensing uncertainty. A state transition function is established under candidate control action conditions to characterize the influence of candidate control actions on the evolution of state variables, and a process noise term is introduced to characterize formation disturbance and unmodeled dynamics. Prior distributions are configured for state variables and generative model parameters, and target prior preference terms are configured for active inference, so that the generative model can express the expected state evolution trend under different control actions; at the same time, the observation function and state transition function are parameterized to obtain the computational form for variational inference and forward prediction.

4. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The process of performing online updates to the generated model to match the current formation and multi-mode pipe jacking machine operating conditions, and outputting the updated actively inferred generated model, includes: Within the current control cycle, acquire the most recent observation vector sequence of a preset length and the corresponding control action sequence to form a data window; The data window is forward-predicted based on the generative model, and the error between the predicted observation and the actual observation is calculated to generate a representation of the degree of matching between the model and the current strata and working conditions. The system compares the error with a preset threshold to determine whether to trigger an online update. If online updates are not triggered, the parameters of the actively inferred generative model from the previous control cycle remain unchanged; if online updates are triggered, the update step size and regularization weights are determined, and recursive optimization is used to update the observation function parameters and state transition function parameters in the generative model, so that the free energy of the updated generative model on the data window is reduced. The updated generative model is subjected to boundary constraint verification and prediction consistency verification. If the verification passes, the updated active inference generative model is output. If the verification fails, the generative model parameters before the update are rolled back and the update for this period is frozen.

5. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The active inference generation model performs variational inference within the current control cycle to calculate the posterior distribution of state variables, and, combined with equipment and process boundary conditions, generates a constrained set of candidate control actions, including: Based on the state transition function, the state distribution at the current moment is propagated prior to the previous control action to obtain the predicted prior distribution of the state variables. A variational distribution is constructed and the observation likelihood is calculated based on the observation function. The distribution is then iteratively updated by minimizing the variational free energy, and the posterior distribution of the state variable is output. Obtain the capability boundary and process safety boundary of the multi-mode pipe jacking machine to form a constraint set; construct the uncertainty constraint margin according to the posterior distribution and predicted prior distribution of the state variables to modify the basic constraint set into the probabilistic feasible constraint set of the current control cycle. Based on the controllable dimension and discrete control strategy, a set of candidate control actions is generated within the probabilistic feasible constraint set, and the candidate actions are screened for feasibility to obtain a set of constrained candidate control actions. At the same time, the set of constrained candidate control actions is structured and parameterized.

6. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The method utilizes an active inference generation model to perform forward simulation prediction on constrained candidate control actions, obtaining a set of multi-action prediction results, including: The constrained set of candidate control actions is indexed and expanded in the time domain to form the action sequence of each candidate action in the prediction time domain; Using the current state posterior distribution as the initial condition, the predicted state distribution is recursively calculated based on the state transition function under the condition of the action sequence of each candidate action in the prediction time domain. Based on the observation function, the predicted state distribution is mapped to the predicted observation distribution, and process noise and observation noise are incorporated into uncertainty propagation; The predicted state trajectory, predicted observation trajectory, and uncertainty corresponding to each candidate action are collected to form a set of multi-action prediction results.

7. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The configuration target prior preference is set, and the variational free energy of candidate control actions in the multi-action prediction result set is calculated to form an action-variable free energy mapping, including: Obtain the preference indicators corresponding to the tunneling target and the safety constraint target; The preference index is converted into a target prior preference distribution, and the prediction distributions of each candidate control action in the multi-action prediction result set are aligned with the target prior preference distribution on the same variable space and time index. Calculate the variational free energy of each candidate control action in the prediction time domain: ； In the formula, Candidate actions The scalar evaluation value of the variational free energy in the prediction time domain. For the first One candidate control action, For the future moment The observed variables, In order to execute candidate actions Under these conditions, the predicted observation distribution obtained from forward simulation by the active inference generation model, For the future moment The target prior preference distribution At the current control cycle time, To predict the length of the time domain, Let KL divergence be the KL divergence. Let cost function be These are the weighting coefficients; After truncation, smoothing, or normalization of the variational free energy, a mapping relationship between the action and the corresponding variational free energy is formed.

8. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The construction of a progressive control hierarchy and threshold rules, and the determination of the control hierarchy and constraint strategy for actions based on the action-variable free energy mapping and the threshold decision quantity corresponding to the posterior distribution of state variables, include: Construct a progressive control hierarchy and pre-set corresponding constraint strategies for each control hierarchy; Based on the posterior distribution of key state variables, threshold rules and threshold decision quantities are constructed; according to the threshold rules, the threshold decision quantities are compared with the equipment process boundary to output the risk level; and free energy evaluation indicators are extracted based on the action-variable free energy mapping. The risk level and free energy evaluation indicators are used to jointly determine the control level for the current control cycle; Based on the determined control level, corresponding constraint strategies are applied to the candidate action set to generate hierarchical action constraint rules.

9. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 8, characterized in that, The method of jointly determining the risk level and free energy evaluation index to determine the control level of the current control cycle includes: Based on a pre-defined risk level and control level comparison table, a recommended control level is obtained based on the risk level. Extract the free energy values ​​of all candidate actions in the current cycle from the action-variable free energy mapping results, and calculate the free energy determination quantity; Based on the pre-defined free energy determination quantity and control level rules, a recommended control level based on free energy is obtained; Adopting the principles of risk priority and conservatism priority, and jointly comparing the recommended control level based on risk level and the recommended control level based on free energy, the control level for the current control cycle is given.

10. The construction method for controlling the tunneling posture of a multi-mode pipe jacking machine based on progressive hierarchical optimization according to claim 1, characterized in that, The control hierarchy and constraint strategy based on equipment process constraints and actions construct a hierarchical optimization problem with the goal of minimizing variational free energy. Candidate control actions in the action-variable free energy mapping are screened and solved to generate the optimal control action for the current control cycle, including: Obtain the control level and corresponding constraint strategy determined in the current control cycle, and merge the equipment process constraints with the constraint strategy to construct the feasible domain constraint set for the current control cycle; With minimizing variational free energy as the overall objective, hierarchical weights are set according to the control level to construct a hierarchical objective function for the hierarchical optimization problem; The feasibility of each candidate control action in the action-variable free energy mapping is checked, and candidate control actions that do not satisfy the feasible region constraint set are eliminated to obtain a subset of feasible candidate actions; Based on the hierarchical objective function, hierarchical free energy evaluation values ​​are calculated for a subset of feasible candidate actions, and then compared and sorted according to hierarchical priority to obtain a set of candidate actions with the smallest free energy. Local continuous optimization is performed in the neighborhood of the set of alternative actions to obtain the optimal control action that satisfies the feasible region constraint set.

Images