In-situ preparation method of high water content soft soil curing agent

Abstract

The application provides a field preparation method of a high-water-content soft soil in-situ curing agent, and belongs to the technical field of field preparation of the high-water-content soft soil in-situ curing agent.The application constructs a soil body geological feature vector by collecting a field soft soil sample, inputs the layered reinforcement learning driven reconfigurable control network to output a target formula parameter vector, and then selects an optimal formula in a Pareto frontier through a multi-objective evolution-gradient hybrid optimization algorithm based on hydration kinetics constraints, and prepares a uniform powder mixture through a three-stage series mixing process according to the formula, and prepares a curing slurry through a multivariable adaptive model predictive control and a double feedback closed loop of an online rheological sensor vibration probe, and drives a grouting pump to inject the slurry into a soft soil layer according to an adaptive pressure-flow collaborative grouting scheduling algorithm based on nonlinear seepage mechanics, so that the technical problem of lacking real-time closed loop feedback in the grouting process control is solved.

Description

Technical Field

[0001] This invention belongs to the field technology of in-situ preparation of in-situ solidifying agents for soft soil with high moisture content. Specifically, it relates to an in-situ preparation method of an in-situ solidifying agent for soft soil with high moisture content. Background Technology

[0002] Treatment of soft soil foundations with high water content is an important topic in geotechnical engineering, and in-situ grouting with curing agents is currently the mainstream method for improving the bearing capacity of soft soil. Existing technologies typically pre-determine curing agent formulations based on laboratory tests, inject grout by manually adjusting the grouting pump pressure and flow rate, and estimate the grout diffusion range using empirical formulas. However, existing technologies have significant shortcomings under varying field conditions. The water content, mineral composition, and temperature of soft soil in the field often exhibit high heterogeneity along both the depth and planar directions, and pre-determined formulations cannot cover these dynamic changes. During grouting, the grout viscosity continuously evolves with the hydration reaction, making it difficult for fixed-parameter controllers to track the non-stationary changes in seepage status in real time. Furthermore, the permeability of the stratum decays exponentially during the curing process, and traditional scheduling models based on linear Darcy's law deviate significantly from actual seepage behavior, easily leading to localized overpressure or uneven curing. In existing technologies, due to the spatial heterogeneity of soft soil geological characteristics and the time-varying nonlinearity of the grout hydration process, there is a lack of real-time information exchange mechanism between the two decision-making levels of hardener formulation optimization and grouting process control. This leads to a disconnect between formulation decision-making and process execution, making it impossible to simultaneously achieve adaptive formulation optimization and closed-loop control of the grouting process under dynamic operating conditions. In other words, existing technologies suffer from the technical problem of lacking real-time closed-loop feedback in hardener control during the grouting process. Summary of the Invention

[0003] In view of this, the present invention provides a method for on-site preparation of an in-situ solidifying agent for soft soil with high moisture content, which can solve the technical problems in the prior art where the solidifying agent formulation cannot be dynamically adapted according to the geological characteristics of soft soil and the grouting process control lacks real-time closed-loop feedback.

[0004] This invention is achieved as follows: This invention provides a method for the on-site preparation of an in-situ solidification agent for soft soil with high water content, comprising the following steps:

[0005] Collect soft soil samples from the site, determine the moisture content, mineral composition and temperature, construct a soil geological feature vector, input it into a hierarchical reinforcement learning-driven reconfigurable control network, and output the target formula parameter vector.

[0006] Based on the target formulation parameter vector, a multi-objective evolutionary-gradient mixing optimization algorithm based on hydration kinetic constraints is run to select the optimal formulation that satisfies the field constraints from the Pareto front formulation set and determine the mass ratio of each component.

[0007] Weigh each curing agent component powder according to the optimal formula, and prepare a uniform powder mixture through a three-stage series mixing process;

[0008] A solidified slurry is prepared by mixing powder mixture with water. The water-cement ratio is adjusted by multivariate adaptive model predictive control. The viscosity of the slurry is monitored in real time by an online rheological sensor with a tuning fork vibration probe and temperature compensation is performed to form a dual feedback closed-loop control.

[0009] Based on the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics, the joint scheduling sequence of grouting pressure and flow is solved in a rolling manner to drive the grouting pump to inject solidified grout into the soft soil layer;

[0010] After grouting is completed, multiple parallel tests are conducted to record the unconfined compressive strength, peak hydration exothermic value and diffusion radius. The test data are then fed back to a hierarchical reinforcement learning-driven reconfigurable control network and a multi-objective evolutionary-gradient hybrid optimization algorithm to achieve iterative model updates.

[0011] The soil geological feature vector is specifically a state description vector composed of water content, estimated mineral composition, and on-site temperature as components, used to characterize the physicochemical properties of soft soil.

[0012] Specifically, the hierarchical reinforcement learning-driven reconfigurable control network is a two-time-scale hierarchical partially observable Markov decision process composed of a high-level policy network and a low-level policy network. The high-level policy network takes the soil geological feature vector as the state and outputs the target formula parameter vector. The low-level policy network takes the real-time sensing signal of the equipment as the state and the high-level sub-target as the condition, and outputs pump speed and valve opening control actions.

[0013] In the hierarchical reinforcement learning-driven reconfigurable control network, the actor network uses a super network to dynamically generate the weight parameters of the low-level policy network. When the statistical process control module detects a change in operating conditions, the high-level network outputs a new network parameter vector to the super network, realizing real-time reconstruction of the policy network. The reconstruction response time is lower than the reconstruction response time threshold.

[0014] The learning rate of the low-level policy network of the hierarchical reinforcement learning-driven reconfigurable control network is dynamically adjusted by an adaptive adjustment function. The adaptive adjustment function takes the mean of the current batch time difference error, the standard deviation of the current batch sensor signal, and the number of updates completed as inputs, outputs a learning rate adjustment coefficient, and selects the corresponding learning rate in segments according to the adjustment coefficient interval.

[0015] Specifically, the multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints employs a third-generation non-dominated sorting genetic algorithm for global population search in the outer layer, and an inner layer... , , , The four-phase hydration rate equation is solved using the fourth-order Runge-Kutta integral to verify its thermodynamic feasibility; elite individuals additionally execute the finite-memory-Broyden-Fletcher-Goldfarb-Shanno-constrained algorithm for local gradient refinement.

[0016] Specifically, the three-stage series mixing process involves the first stage using Venturi jet premixing to wet the particles, the second stage using high-speed shearing to break up agglomerates, and the third stage using low-speed circulating stirring for homogenization. The combination of blade spacing and rotation speed is determined through computational fluid dynamics simulation optimization. The coefficient of variation for mixing uniformity must be reduced to below the coefficient of variation threshold.

[0017] Specifically, the multivariable adaptive model predictive control uses an online rheometer to measure viscosity and an online densitometer to measure density to form a dual feedback loop, establishing a data-driven dynamic model of water-cement ratio, viscosity, temperature, and density. The closed-loop adjustment period does not exceed the adjustment period threshold and is equipped with an anti-integral saturation mechanism.

[0018] Among them, the temperature compensation algorithm of the online rheological sensor of the tuning fork vibration probe corrects the viscosity measurement value in real time based on the viscosity-temperature relationship calibration curve; the slurry thixotropic recovery kinetics proxy model predicts the peak pressure of pipeline start-up based on the online identification of Houška model parameters and adjusts the pump speed ramp control strategy in advance.

[0019] Specifically, the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics uses the measured sequence of grouting pressure and flow as observations, employs an extended Kalman filter to identify the seepage state vector online, drives a simplified seepage finite difference model, and aims to minimize the solidification uniformity index. Within the model prediction and control framework, it continuously solves the joint scheduling sequence of grouting pressure and flow.

[0020] The solidification uniformity index is specifically the variance of the diffusion radius at each observation point during the grouting solidification process; the constraints include the maximum rated pressure of the pump, the safety margin of the formation fracturing critical pressure estimated online based on the Vesić formula, and the upper limit of the daily grouting volume.

[0021] The training of the hierarchical reinforcement learning-driven reconfigurable control network includes a pre-training stage, a transfer fine-tuning stage, and an online update stage. The transfer fine-tuning stage uses a hierarchical Bayesian transfer learning framework to transfer indoor pre-training parameters to field conditions. The online update stage uses a priority experience replay strategy to trigger incremental training of the low-level policy network.

[0022] The critic network is embedded with a Lagrange constraint layer to ensure that the upper limits of equipment pressure and temperature are not violated; the Lagrange multipliers are adaptively updated during training through a dual gradient ascent algorithm.

[0023] The distribution of the diffusion radius is determined by a combination of core sampling after grouting and on-site resistivity tomography; the number of parallel tests is not less than the threshold for the number of parallel tests, and each set of tests covers different water-cement ratios and mix proportion gradients.

[0024] The reconstruction response time threshold is 100ms; the coefficient of variation threshold is 5%; the adjustment period threshold is 50ms; the number of parallel trials threshold is 5 groups; and the linear velocity of the second-stage high-speed shear is not less than 15. The online rheological sensor of the tuning fork vibration probe has a sampling rate of not less than 10. The water-ash ratio ranges from 0.6 to 1.2; the population size is 200; and the number of neurons in the hidden layer of the supernetwork is 512 and 256, respectively.

[0025] This invention constructs a hierarchical reinforcement learning-driven reconfigurable control network, which models formulation optimization and process control as a hierarchical decision framework with two time scales, realizing real-time information interaction between the formulation decision layer and the process execution layer.

[0026] In traditional technologies, formulation design and grouting control are independent processes, lacking dynamic coupling. This invention's high-level strategy network takes soil geological characteristics as input and dynamically outputs a target formulation parameter vector, which is then passed to the low-level strategy network as sub-objectives. Under the constraints of the high-level sub-objectives, the low-level strategy network, combined with real-time sensor signals, adjusts grouting pressure and flow rate in a closed loop, thus achieving continuous coupling between formulation decision-making and process control at the information level. The mechanism of the supernetwork dynamically generating low-level weights allows the system to reconstruct strategies without offline retraining during operating condition changes, overcoming the fundamental deficiency of fixed-parameter controllers in their insufficient generalization ability under multiple operating conditions. In summary, this invention solves the technical problem of lacking real-time closed-loop feedback in the control of curing agents during the grouting process. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method of the present invention.

[0028] Figure 2 The graph shows the changes in diffusion radius and solidification uniformity index at each observation point during the grouting process as a function of the number of scheduling steps.

[0029] Figure 3 The distribution of unconfined compressive strength and peak temperature of hydration exothermic reaction over 28 days under different water-cement ratios is shown in the figure.

[0030] Figure 4 The graph shows the evolution of the correlation between the output of the high-level policy network and the measured strength after multiple iterations of a reconfigurable control network driven by hierarchical reinforcement learning. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0032] like Figure 1 The diagram shown is a flowchart of a method for on-site preparation of an in-situ solidification agent for soft soil with high moisture content, provided by this invention. This method includes the following steps:

[0033] S01. Collect soft soil samples from the site, determine the moisture content, mineral composition and temperature, construct the soil geological feature vector, input the hierarchical reinforcement learning driven reconfigurable control network, and output the target formula parameter vector.

[0034] S02. Based on the target formulation parameter vector, run the multi-objective evolutionary-gradient mixing optimization algorithm based on hydration kinetic constraints, select the optimal formulation that meets the field constraints from the Pareto front formulation set, and determine the mass ratio of each component.

[0035] S03. Weigh each curing agent component powder according to the optimal formula, and prepare a uniform powder mixture through a three-stage series mixing process. The coefficient of variation of the mixing uniformity must be reduced to below 5%.

[0036] S04. Mix the powder mixture with water to prepare a solidified slurry. Use a multivariate adaptive model to predict and control the water-cement ratio. Use a tuning fork vibration probe online rheological sensor to monitor the slurry viscosity in real time and perform temperature compensation to form a dual feedback closed-loop control.

[0037] S05. Based on the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics, the joint scheduling sequence of grouting pressure and flow is solved in a rolling manner, and the grouting pump is driven to inject the solidified grout into the soft soil layer.

[0038] S06. After grouting is completed, multiple parallel tests are conducted to record the 28-day unconfined compressive strength, hydration exothermic peak value and diffusion radius. The test data are fed back to a hierarchical reinforcement learning-driven reconfigurable control network and a multi-objective evolutionary-gradient hybrid optimization algorithm to achieve iterative model updates.

[0039] The soil geological feature vector refers to a state description vector composed of water content, estimated mineral composition, and on-site temperature as components, used to characterize the physicochemical properties of soft soil under different working conditions.

[0040] The basic formula is as follows: ultrafine composite mineral admixture (slag powder and fly ash in a mass ratio of 6:4) accounts for 40%–50% of the total mass of the curing agent; ordinary silicate cement accounts for 30%–40% of the total mass of the curing agent; high-performance marine special cementitious materials account for 10%–20% of the total mass of the curing agent; activator (sodium sulfate and gypsum in a mass ratio of 7:3) accounts for 3%–8% of the total mass of the curing agent; the water-cement ratio (the ratio of water to curing agent powder mass) is 0.6–1.2, dynamically adjusted by a hierarchical reinforcement learning-driven reconfigurable control network based on the measured moisture content. The above proportion range was determined through indoor orthogonal experiments (moisture content gradient 40%–120%, mineral composition gradient covering three typical soft soil types: kaolinite-dominated, montmorillonite-dominated, and illite-dominated) combined with variance analysis to determine the initial range, which was then iteratively corrected through multiple sets of field experiments, and finally refined in the Pareto front by a multi-objective evolutionary-gradient hybrid optimization algorithm.

[0041] The three-stage cascade mixing process employs Venturi jet premixing in the first stage to achieve particle wetting, and in the second stage, a linear velocity of not less than 15 is used. The high-speed shearing breaks up the agglomerates, and the third stage uses low-speed circulation stirring to homogenize them; the combination of blade spacing and rotation speed is determined by computational fluid dynamics simulation optimization, with simulation mesh accuracy of not less than 1 mm, and simulation conditions covering the entire formulation range.

[0042] The coefficient of variation of the uniformity of mixing refers to the ratio of the standard deviation to the mean of the concentration distribution of each component of the curing agent in the mixture. It is used to quantitatively describe the degree of uniformity of micro-dispersion of the powder. The target value is less than 5%, which is determined based on the correlation test between the spatial distribution uniformity of hydration products and 28-day strength.

[0043] The multivariable adaptive model predictive control uses an online rheometer to measure viscosity and an online densitometer to measure density to form a dual feedback loop, establishing a data-driven dynamic model of water-cement ratio, viscosity, temperature, and density. The closed-loop adjustment cycle does not exceed 50ms and is equipped with an anti-integral saturation mechanism. The control model parameters are determined by fitting laboratory calibration experiments with 10 initial sets of field experimental data, and are subsequently updated online by a reconfigurable control network driven by hierarchical reinforcement learning.

[0044] The online rheological sensor of the tuning fork vibration probe has a sampling rate of not less than 10. The viscosity measurement value is corrected in real time through a temperature compensation algorithm. The temperature compensation algorithm is based on the viscosity-temperature relationship calibration curve, which is determined by isothermal viscosity measurement experiments at 5℃ intervals within the range of 20℃ to 60℃. The slurry thixotropic recovery kinetic proxy model is based on online identification of Houška model parameters to predict the peak pressure of pipeline startup and adjust the pump speed ramp control strategy in advance to suppress the malfunction of safety valve. The initial values ​​of Houška model parameters are obtained by fitting static recovery experiments (static time gradient from 0.5min to 5min, step size 0.5min).

[0045] The diffusion radius refers to the effective solidification boundary radius of the solidified grout penetrating outward from the grouting point in the soft soil layer. The diffusion radius distribution is determined by a combination of core sampling after grouting and on-site resistivity tomography.

[0046] The number of parallel trials is no less than 5, with each trial covering different water-cement ratios and mixing gradients, used to support online posterior updates of reconfigurable control networks driven by hierarchical reinforcement learning and Pareto front correction of multi-objective evolutionary-gradient hybrid optimization algorithms.

[0047] The specific structure of the hierarchical reinforcement learning-driven reconfigurable control network (HRL-RCN) is as follows: This network models the formulation optimization and process control of the on-site preparation of the curing agent as a hierarchical partially observable Markov decision process with two time scales, consisting of a high-level policy network and a low-level policy network. The high-level policy network (formulation optimization layer, decision cycle approximately 60s) uses soil geological feature vectors (moisture content, mineral composition estimation, temperature) as its state and outputs a target formulation parameter vector as a sub-objective. It is trained using a proximal policy optimization algorithm with intrinsic incentives. The intrinsic incentives are based on latent space novelty calculations to encourage exploration of uncovered formulation spaces. The intrinsic incentive weight coefficients are determined iteratively through pre-training experiments under five typical soft soil conditions. The low-level policy network (process control layer, decision cycle approximately 50ms) uses real-time sensor signals from the equipment (pump pressure, flow rate, viscosity, temperature) as its state and the high-level sub-objectives as its conditions. It outputs specific control actions (pump speed, valve opening) and is trained using a soft actor critic algorithm to ensure control robustness. The temperature hyperparameter of the soft actor critic algorithm is automatically adjusted by an adaptive adjustment function. In the reconfigurable mechanism, the actor network uses a supernetwork to dynamically generate the weight parameters of the low-level policy network; the input dimension of the supernetwork corresponds to the total number of parameters of the low-level policy network; the hidden layer adopts a two-layer fully connected structure with 512 and 256 neurons respectively, and the activation function is a rectified linear unit; the weight matrix between neurons is stored in blocks, each block corresponding to a specific set of sensor input channels to reduce redundant computation across neurons; the gradient flow between layers is stabilized by spectral normalization, and the spectral normalization coefficient is determined by Lipschitz constraint theory; the temporal batches of data loops are allocated in a circular buffer according to a priority empirical playback strategy, and the priority weights are updated in real time based on the temporal difference error; the memory allocation adopts static graph pre-compilation, allocating the high-level policy network and the low-level policy network to independent CUDA streams, with the high-level network allocated to CUDA stream 0 and the low-level network allocated to CUDA stream 1, avoiding inter-stream blocking; memory allocation adopts pinned memory (Pinned memory). Memory mapping of sensor input tensors reduces data transmission latency between the host and devices. In CUDA's different kernel function allocation strategies, matrix multiplication kernel functions are allocated to computationally intensive stream processor clusters, while normalization kernel functions are allocated to memory-bandwidth intensive clusters, achieving kernel-level parallelism. In hierarchical allocation, the value estimation head of the critic network and the action sampling head of the actor network are placed in different computation graph branches, supporting asynchronous backpropagation. When the statistical process control module detects a change in operating conditions (based on exceeding the limits of cumulative sum and statistics), the higher-level network outputs a new network parameter vector to the supernetwork, enabling real-time reconstruction of the policy network without retraining, with a reconstruction response time of less than 100ms. The cumulative sum and statistics thresholds are determined by analyzing the subject operating characteristic curves of historical operating condition switching data.The commentator network embeds a Lagrange constraint layer to ensure that safety constraints (upper limit of equipment pressure, upper limit of temperature) are not violated; the initial value of the Lagrange multiplier is set to 0 and is adaptively updated during training through the dual gradient ascent algorithm; the upper limit constraint value is jointly determined by the equipment nameplate rating and the on-site safety specifications.

[0048] The steps for establishing the training dataset for the hierarchical reinforcement learning-driven reconfigurable control network specifically include: collecting indoor simulated grouting test data for five typical soft soil types (kaolinite-dominated, montmorillonite-dominated, illite-dominated, mixed type, and organic matter-enriched type) with a moisture content range of 40% to 120%, with no less than 200 sets of sensor time-series records collected for each type of working condition; collecting historical grouting operation data from the field, covering no less than three engineering projects, with continuous recording for no less than 50 minutes for each project; performing time alignment and normalization on the indoor and field data, constructing a state-action-reward triplet sequence, with the reward function consisting of a weighted average of 28-day intensity, peak hydration heat release, and material cost, with the weight coefficients initially set by expert experience and then iteratively adjusted through Bayesian optimization; and using a hierarchical Bayesian transfer learning framework to perform distribution alignment on small sample field data, eliminating the distribution offset between indoor and field data to support posterior updates.

[0049] The specific steps of training the hierarchical reinforcement learning-driven reconfigurable control network include: a pre-training phase, in which the high-level policy network and the low-level policy network are trained offline using indoor data, with the high-level network trained for 300 rounds using the proximal policy optimization algorithm and the low-level network trained for 500 rounds using the soft actor critic algorithm; a transfer fine-tuning phase, in which the indoor pre-trained parameters are transferred to the field conditions using a hierarchical Bayesian transfer learning framework, and the low-level field condition-specific parameters are quickly updated posteriorly using a small amount of field data (no less than 5 sets), and the global super-prior parameters are initialized using a class-based learning approach; and an online update phase, in which after each set of field grouting operations is completed, the priority experience replay buffer is updated with newly collected data, triggering incremental training of the low-level policy network, with an incremental training batch size of 32 and a learning rate dynamically adjusted by an adaptive adjustment function.

[0050] The adaptive adjustment function is used to adjust the learning rate of the lower-level policy network in the hierarchical reinforcement learning-driven reconfigurable control network, and the formula is expressed as follows:

[0051] ;

[0052] in, This represents the average time difference error (ms) for the current batch. The initial time difference error baseline value (ms) is used. The standard deviation of the sensor signal for the current batch. The initial batch standard deviation of the sensor signal. This indicates the number of updates completed. Maximum number of updates and Determined by the statistical mean of the last 50 batches in the pre-training phase. The total number of grouting operations is estimated on-site and then rounded down; when At that time, the learning rate was adopted (Initial maximum learning rate) to accelerate parameter convergence; when At that time, the learning rate was adopted To balance convergence speed and stability; when At that time, the learning rate was adopted (Minimum protection learning rate) to prevent overfitting and The learning rate sensitivity scan experiment during the pre-training phase (learning rate range) to (Determined by logarithmic uniform sampling of 20 points).

[0053] The hierarchical reinforcement learning-driven reconfigurable control network decouples formulation optimization decisions and process control actions into two time scales, ensuring that high-level formulation decisions are not affected by millisecond-level control disturbances, while low-level controllers respond quickly to sensor changes under high-level sub-objective constraints. The mechanism of dynamically generating weights in the supernetwork allows the model to complete policy reconstruction without offline retraining when switching operating conditions, overcoming the problem of insufficient generalization ability of traditional fixed-parameter controllers under multiple operating conditions. The Lagrange constraint layer enforces the satisfaction of equipment safety boundaries from the optimization structure, eliminating the risk of constraint violation due to improper weight setting of soft constraint penalty terms. The hierarchical Bayesian transfer learning framework effectively suppresses overfitting under small sample field data conditions through global super-prior sharing, enabling the model to quickly achieve reliable prediction accuracy under new operating conditions, thereby improving the overall stability and adaptability of the curing agent field preparation process.

[0054] The principle and specific implementation of the multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints are as follows: This algorithm addresses a high-dimensional multi-objective nonlinear programming problem for optimizing multi-component formulations of curing agents. The optimization objectives are to maximize 28-day strength, minimize material cost, and minimize the peak temperature of hydration exothermic reaction. These three objectives constitute the objective function vector for multi-objective optimization. The initial values ​​of the weight coefficients for the three objectives are set by expert experience and then iteratively adjusted through Pareto front interactive preference learning. The outer layer of the algorithm uses an improved third-generation non-dominated sorting genetic algorithm for global population search, with a population size of 200. Diversity is maintained through a Pareto front guided by reference points. Each candidate formulation solution is fed into the inner layer validator, which uses the Arrhenius parameterized hydration kinetic equations (including...) , , , The four-phase hydration rate equation was solved using a fourth-order Runge-Kutta integral to verify thermodynamic feasibility. Infeasible solutions were mapped to constraint surfaces using gradient projection operators, with the gradient projection direction determined by the constraint surface normal vector. Elite individuals (top 10%) underwent additional local gradient refinement using a finite-memory-Broyden-Fletcher-Goldfarb-Shanno constraint algorithm, utilizing automatic differentiation to calculate accurate gradients and rapidly converge to a Pareto local optimum. Evolutionary operations were performed in a normalized recipe space, with normalization based on the mass fraction of each component divided by its upper limit of allowable range. Crossover probability and mutation magnitude were adaptively adjusted by the evolutionary progress (the ratio of the current generation to the maximum generation), with the adjustment curve determined by joint optimization of convergence speed and diversity indices in pre-experiments. Finally, a Pareto front recipe set was output for engineers to select based on their preferences according to actual constraints. This algorithm ensures that every formulation solution entering local refinement satisfies hydration kinetic feasibility by embedding hydration thermodynamic constraints into the inner layer of the evolutionary search, thereby avoiding thermodynamic failure of the optimization results in engineering implementation. The multi-objective framework preserves the trade-off between strength, cost, and exothermic peak, providing structured support for formulation decisions under different engineering constraints. Local gradient refinement significantly improves the resolution of elite solutions on the Pareto surface, making up for the insufficient efficiency of pure evolutionary algorithms in local search in high-dimensional space. Overall, it improves the engineering adaptability and thermal stability of curing agent formulations.

[0055] The principle and specific implementation of the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics are as follows: The algorithm treats the soft soil layer as a saturated heterogeneous porous medium, where the grout seepage satisfies the nonlinear Darcy-Fochheimer equation, with permeability decreasing exponentially with solidification degree, and grout viscosity varying with hydration time and temperature. Using the measured sequence of grouting pressure and flow rate as observations, an extended Kalman filter is used to identify the state vector containing initial permeability, decay coefficient, initial viscosity, and activation energy online. The identified parameters drive a simplified 5×5 grid seepage finite difference model to calculate the current solidification diffusion radius distribution. The solidification uniformity index (diffusion...) is then used to... With the goal of minimizing the radius variance, the joint scheduling sequence of grouting pressure and flow rate with a prediction step size of 10 steps is solved in a rolling manner within the model predictive control framework. Constraints include the maximum rated pressure of the pump, the safety margin of the formation fracturing critical pressure estimated online based on the Vesić formula, and the upper limit of the daily grouting volume. The extended Kalman filter state is updated in real time every control step (30s) and the model predictive control problem is solved again to achieve online adaptive closed-loop scheduling. The spacing of the 5×5 grid nodes is determined by the experiment of balancing the grouting influence radius and the calculation accuracy. The initial value of the noise covariance matrix of the extended Kalman filter process is obtained by the maximum likelihood estimation of the indoor seepage calibration experimental data. This algorithm embeds a nonlinear seepage mechanics model into an online identification and rolling optimization framework, ensuring that grouting pressure and flow scheduling decisions are always based on the actual permeability of the current formation rather than fixed empirical parameters. Extended Kalman filtering tracks the dynamic changes in permeability and viscosity in real time, eliminating the impact of formation parameter non-stationarity on scheduling accuracy during the solidification process. Using the solidification uniformity index as the optimization objective, it drives the uniform diffusion of grout in the formation from a scheduling structure perspective, reducing the risk of formation fracturing or uneven solidification due to local overpressure, and comprehensively improving the spatial uniformity and engineering safety of grouting solidification.

[0056] The solidification uniformity index refers to the variance of the diffusion radius at each observation point during the grouting solidification process. It is used to quantitatively characterize the uniformity of the spatial distribution of the solidified grout in the formation. The smaller the variance, the more uniform the solidification distribution.

[0057] The intrinsic incentive refers to the additional reward signal given to the agent for accessing novel regions in the latent state space during reinforcement learning training. Novelty is measured by the statistical distance between the current state and the distribution of historical states in the latent space, and is used to drive the high-level policy network to explore the unseen recipe space.

[0058] The supernetwork refers to a neural network structure that takes a condition vector as input and dynamically outputs the weight parameters of another network. In this scheme, the high-level policy network outputs a condition description vector to the supernetwork, and the supernetwork generates the weight matrix of the low-level policy network accordingly, thereby realizing the dynamic reconstruction of the policy.

[0059] The cumulative sum statistic refers to a statistic used to detect shifts in the mean of a sequence. By accumulating the difference between the observed value and the reference mean, a change in operating condition is determined when the cumulative sum exceeds a preset threshold. The threshold is determined by analyzing the subject operating characteristic curves of historical operating condition change data.

[0060] The spectral normalization refers to normalizing the weight matrix of the neural network according to its maximum singular value, which is used to constrain the Lipschitz constant of the network, thereby stabilizing the gradient magnitude during the training process.

[0061] The Lagrange constraint layer refers to a structural layer that embeds inequality safety constraints in the form of Lagrange multipliers into the network optimization objective. The multipliers are adaptively adjusted during training using the dual gradient ascent algorithm, so that the network output structurally satisfies the upper pressure and upper temperature constraints.

[0062] The priority experience replay strategy refers to assigning priority weights to historical samples in the experience replay buffer according to the absolute value of the time difference error, prioritizing the sampling of samples with larger errors, so that the network training pays more attention to the state region with higher current fitting error.

[0063] The hierarchical Bayesian transfer learning framework refers to a Bayesian model structure that divides multi-condition data learning into a global hyperprior layer and a condition-specific parameter layer. The global hyperprior parameters are learned jointly from historical data of multiple conditions, while the condition-specific parameters are quickly updated posteriorly through a small amount of field data. Its class-based learning approach originates from the parameter initialization idea of ​​model-independent meta-learning.

[0064] The Vesić formula is a theoretical formula used to estimate the ultimate lateral resistance and end resistance of soil. In this scheme, it is used to calculate the critical splitting pressure of the formation online. The formula parameters (internal friction angle, cohesion) are obtained by interpolation of the formation state vector identified by extended Kalman filter and the initial calibration values ​​of the indoor triaxial test.

[0065] The Darcy-Fochheimer equation is a nonlinear equation describing fluid seepage in porous media. At low flow rates, it degenerates into a linear Darcy law, while at high flow rates, an inertial effect term needs to be taken into account. The equation parameters (permeability, inertial drag coefficient) are identified online by extended Kalman filtering.

[0066] The fourth-order Runge-Kutta integral refers to a numerically accurate explicit method for solving systems of ordinary differential equations, which is used in this scheme for time-step solution. , , , The integration time step of the four-phase hydration rate equation is determined by numerical stability analysis (Courant condition), which is obtained by pre-integration stability tests on the hydration rate equations of different formulations.

[0067] The third-generation non-dominated sorting genetic algorithm is a multi-objective evolutionary algorithm that maintains population diversity through a Pareto front decomposition strategy guided by reference points. Compared with the second-generation algorithm, it has a more uniform front distribution in high-dimensional multi-objective problems with three or more objectives. The number and distribution of reference points are jointly determined by the number of objectives and the population size.

[0068] The specific implementation of step S01 is as follows: Before grouting, technicians collect soil samples at the soft soil site according to a grid layout. At each sampling point, the moisture content is determined using the ring cutter method, the mineral composition ratio is estimated using X-ray diffraction analysis, and the on-site temperature is recorded using a thermocouple probe. These parameters are organized into a soil geological feature vector consisting of a moisture content component, an estimated mineral composition component, and a temperature component. This vector serves as the state input to a hierarchical reinforcement learning-driven reconfigurable control network. After inference by the near-end policy optimization algorithm of the high-level policy network, a target formulation parameter vector is output, containing the target range of mass ratios for each component and the target water-cement ratio, used for initialization of subsequent formulation optimization steps.

[0069] The specific implementation of step S02 is as follows: Using the target formulation parameter vector output in step S01 as the initialization condition, a multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints is initiated. The outer layer of the algorithm performs a global search in the formulation space using a third-generation non-dominated sorting genetic algorithm with a population size of 200, guided by reference points to maintain Pareto front diversity. The optimization objectives are to maximize the 28-day unconfined compressive strength, minimize material cost, and minimize the peak temperature of hydration exothermic reaction. Each candidate formulation solution enters the inner validator... , , , The four-phase Arrhenius parameterized hydration rate equation is solved stepwise using the fourth-order Runge-Kutta integral to verify thermodynamic feasibility; infeasible solutions are mapped to constraint surfaces using gradient projection operators. The top 10% of elite individuals undergo additional local gradient refinement using a finite-memory—Broyden—Fletcher—Goldfarb—Shanno—constraint algorithm, utilizing automatic differentiation to calculate the accurate gradient and rapidly converge to a Pareto local optimum. The final output is a Pareto front formulation set, and technicians select the optimal formulation based on site constraints to determine the mass ratio of each component. The basic proportion range is: ultrafine composite mineral admixture (slag powder and fly ash in a 6:4 ratio) 40%–50%, ordinary Portland cement 30%–40%, high-performance marine special cementitious materials 10%–20%, activator (sodium sulfate and gypsum in a 7:3 ratio) 3%–8%, and water-cement ratio 0.6–1.2.

[0070] The specific implementation of step S03 is as follows: Weigh the powder of each curing agent component according to the optimal formula, and sequentially enter the three-stage series mixing process. The first stage uses Venturi jet premixing, utilizing the negative pressure generated by the Bernoulli effect to entrain the powder into a high-speed airflow, achieving particle surface wetting and reducing subsequent shear energy consumption. The second stage uses a linear velocity of not less than 15... The high-speed shear rotor uses a strong shear field to break up agglomerates formed by electrostatic adsorption or van der Waals forces, ensuring thorough dispersion of all component particles. The third stage employs a low-speed circulating impeller for homogenization, eliminating local concentration gradients. The blade spacing and rotational speed combination are determined through computational fluid dynamics simulation optimization, with a simulation mesh accuracy of at least 1 mm, covering all formulation conditions. After mixing, a multi-point sampling method is used to detect the concentration distribution of each component, and the coefficient of variation is calculated; it must be below 5% before proceeding to the next step.

[0071] The specific implementation of step S04 is as follows: The powder mixture obtained in step S03 is added to a mixing tank, water is added according to the target water-cement ratio, and a multivariate adaptive model predictive control system is activated. The control system uses online rheometer measurements of slurry viscosity and online densitometer measurements of slurry density as dual feedback signals. Based on the data of water-cement ratio, viscosity, temperature, and density, a dynamic model is driven to adjust the water addition in a closed loop. The adjustment cycle does not exceed 50ms, and an anti-integral saturation mechanism is provided to prevent saturation of the actuator. The tuning fork vibration probe online rheological sensor has a value of not less than 10... The viscosity signal is continuously acquired at a high sampling rate, and the measured value is corrected in real time using a temperature compensation algorithm based on a calibration curve from 20℃ to 60℃. The slurry thixotropic recovery kinetics proxy model is based on the online identification of Houška model parameters, predicts the pressure peak at pipeline startup, and outputs pump speed ramp control commands in advance, thereby suppressing safety valve malfunctions and forming a dual feedback closed-loop control of viscosity and density.

[0072] The specific implementation of step S05 is as follows: An adaptive pressure-flow rate coordinated grouting scheduling algorithm based on nonlinear seepage mechanics is initiated. The soft soil stratum is modeled as a saturated heterogeneous porous medium, where the grout seepage satisfies the nonlinear Darcy-Fochheimer equation, and the permeability decreases exponentially with the degree of solidification. Using the measured sequence of grouting pressure and flow rate as observations, an extended Kalman filter is used to identify the state vector containing initial permeability, attenuation coefficient, initial viscosity, and activation energy online, driving a 5×5 grid simplified seepage finite difference model to calculate the current solidification diffusion radius distribution. With the objective of minimizing the solidification uniformity index (i.e., the variance of the diffusion radius at each observation point), a joint scheduling sequence of grouting pressure and flow rate with a prediction step size of 10 steps is solved within the model predictive control framework. Constraints include the maximum rated pump pressure, the safety margin of the formation fracturing critical pressure estimated online based on the Vesić formula, and the upper limit of the daily grouting volume. The extended Kalman filter state is updated and the model predictive control problem is solved again every control step (30s) to achieve online adaptive closed-loop scheduling.

[0073] The specific implementation of step S06 is as follows: After grouting is completed, at least five parallel tests are arranged within the grouting influence area, each covering different water-cement ratios and mix proportion gradients. The unconfined compressive strength, peak hydration exothermicity, and diffusion radius are recorded after 28 days. The diffusion radius is determined jointly by core sampling and on-site resistivity tomography. The above test data are organized into state-action-reward triplets, the priority experience replay buffer is updated, incremental training of the low-level policy network is triggered, and the Pareto front formulation set is corrected simultaneously, achieving collaborative iterative updates of the hierarchical reinforcement learning-driven reconfigurable control network and the multi-objective evolutionary-gradient hybrid optimization algorithm.

[0074] It should be noted that the key technologies of this invention include: a hierarchical reinforcement learning-driven reconfigurable control network decouples formulation optimization and process control into two time scales, ensuring that high-level formulation decisions are not affected by millisecond-level control disturbances; a hypernetwork dynamic weight generation mechanism enables the system to complete strategy reconstruction without offline retraining during operating condition switching, overcoming the fundamental defect of insufficient generalization ability of traditional fixed-parameter controllers; a multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints embeds thermodynamic feasibility verification into the inner layer of evolutionary search, and combined with local gradient refinement, significantly improves the Pareto front resolution, making the formulation optimization results have engineering physical reliability; and an adaptive pressure-flow collaborative grouting scheduling algorithm based on nonlinear seepage mechanics tracks the dynamic changes of permeability and viscosity in real time through extended Kalman filtering, ensuring that scheduling decisions are always based on the actual formation seepage state. The synergistic effect of these three key technologies enables formulation design, grout preparation, and grouting scheduling to form a complete closed loop at the information level, comprehensively ensuring the adaptability and spatial uniformity of the on-site preparation of the curing agent.

[0075] It should be noted that in the case of organic-rich soft soil with high water content, the organic acids dissolved in the pore water of the soil will undergo complexation reactions with cement hydration products, leading to systematic deviations in the reaction rate constants of each phase in the hydration rate equation from the standard values. Simultaneously, the adsorption of water film by organic matter causes the slurry viscosity to exhibit a strongly nonlinear response to temperature, significantly increasing the error of traditional rheological control models based on linear viscosity-temperature relationships. The above technical problems arise because organic ligands such as humic acid in organic-rich soft soil... Chelation inhibited Hydration products The crystal growth of organic matter and clay minerals causes the hydration exothermic peak to occur earlier and the peak temperature to rise. Simultaneously, the complex colloidal structure formed by organic matter and clay minerals complicates the pore geometry, leading to a deviation of the inertial drag coefficient in nonlinear seepage from the standard porous media model in terms of its decay with solidification degree. Furthermore, if the process noise covariance matrix of the extended Kalman filter is initialized with standard soft soil data, identification bias will accumulate. A common solution to these technical problems is to add an organic passivating agent (such as activated magnesium oxide) to the formulation and manually calibrate the rheological control parameters offline. However, manual calibration relies on numerous field tests and cannot respond in real-time to spatial variations in formation organic matter content. Improperly set passivating agent dosage can negatively impact the hydration synergy efficiency of the activator and cementitious material, resulting in a decrease in 28-day strength. This invention effectively solves this technical problem. The hierarchical Bayesian transfer learning framework incorporates organic-rich soft soil as an independent working condition category into the training dataset. Its working condition-specific parameter layer is rapidly updated a posteriori using a small amount of field data, enabling the high-level policy network to output a target formulation parameter vector containing the combined adjustment of passivating agent dosage and water-cement ratio when an organic-rich working condition is identified (triggered by the organic matter indicator component in the mineral composition estimate). The viscosity-temperature dynamic model of the low-level policy network is continuously corrected using measured data from an online rheological sensor with a tuning fork vibration probe, ensuring that the temperature compensation algorithm maintains accuracy in the nonlinear response range. The process noise covariance matrix of the extended Kalman filter is automatically updated with priority experience playback data during the incremental training phase, allowing the identification of the inertial drag coefficient to gradually converge to the true value of the organic-rich strata, thereby ensuring the spatial uniformity of grouting scheduling under this special working condition.

[0076] Specifically, the principle of this invention is:

[0077] The fundamental reason why this invention can solve the above-mentioned technical problems is that it unifies the two traditionally independent decision-making links of formula optimization and process control in terms of mathematical structure through a hierarchical Markov decision process framework.

[0078] In traditional solutions, formulation design relies on indoor test data, and grouting control relies on fixed parameter regulators; there is no runtime information channel between the two. The high-level strategy network of this invention periodically maps soil geological feature vectors to target formulation parameter vectors. This vector is simultaneously passed as a sub-objective to the low-level strategy network. Within each control cycle, the low-level strategy network uses real-time sensor signals as the state and the sub-objectives as constraints to output specific pump speeds and valve actions. This two-level information link allows changes in formulation decisions to be transmitted to the execution layer within millisecond-level control cycles, structurally eliminating the disconnect between formulation and control.

[0079] The introduction of the supernetwork mechanism further solves the generalization problem under multiple operating conditions. When the cumulative sum and statistics detect a change in operating condition, the higher-level network outputs a new condition vector to the supernetwork. Based on this, the supernetwork dynamically generates all the weights of the lower-level policy network, enabling the control policy to be reconstructed without retraining. The rationale for this mechanism lies in the fact that the supernetwork essentially parameterizes the mapping relationship between the operating condition description vector and the control policy, allowing navigation in the policy space to be achieved through continuous changes in the condition vector, rather than relying on discrete model switching.

[0080] The Lagrange constraint layer enforces the satisfaction of equipment safety boundaries from an optimized structural perspective, eliminating the risk of constraint violations due to improper weighting of soft constraint penalty terms. This is crucial for the safe operation of on-site grouting equipment. Based on the nonlinear Darcy-Fosieheimer equations, a seepage model uses extended Kalman filtering to identify dynamic changes in permeability and viscosity online. This ensures that grouting scheduling decisions are always based on the actual seepage state of the current formation, guaranteeing the achievability of the solidification uniformity optimization objective from a physical model perspective. A multi-objective evolutionary-gradient hybrid optimization algorithm embeds hydration kinetic constraints into the evolutionary search layer, ensuring that each formulation solution entering local refinement satisfies thermodynamic feasibility, thus providing physical reliability for the formulation optimization results in engineering implementation. These technical elements work together to form a complete closed-loop technical chain covering formulation design, grout preparation, grouting scheduling, and model iteration, theoretically guaranteeing the system's adaptability and control accuracy under dynamic operating conditions.

[0081] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0082] The specific implementation of step S01 is to collect soft soil samples from the site and determine their moisture content. (Unit: %), mineral composition and temperature (Unit: °C), construct the soil geological feature vector:

[0083] ;

[0084] In the formula, This is the normalized baseline value for moisture content (unit: %). , , The mass fractions of kaolinite, montmorillonite, and illite are respectively. , , To correspond to the normalized baseline value, This is a temperature normalized baseline value (unit: °C), all of which are the average values ​​of the operating conditions used in the indoor pre-training phase. This vector is input to the high-level policy network of the hierarchical reinforcement learning-driven reconfigurable control network, and outputs the target formula parameter vector:

[0085] ;

[0086] In the formula, The mass fraction of slag powder. This refers to the mass fraction of fly ash. This represents the cement mass fraction. This represents the mass fraction of marine cementitious materials. The activator mass fraction, The water-cement ratio is given, and all components are within the allowable range.

[0087] The specific implementation of step S02 is as follows: based on the target formula parameter vector A multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints is run, with the three objective function vectors as follows:

[0088] ;

[0089] In the formula, Predicted 28-day unconfined compressive strength (unit: ), The intensity normalization reference value (unit: ), Material cost (unit: yuan) t), The normalized cost benchmark (unit: yuan) t), The peak temperature of hydration heat release (unit: °C). This is the normalized reference value for the peak exothermic temperature (unit: °C). Let be the formula decision vector, where , , , , These are the mass fractions of slag powder, fly ash, cement, marine cementitious materials, and activator. The water-cement ratio is given. The inner verifier solves the hydration kinetic equations for each candidate solution, and the reaction rate equations for each hydration phase are as follows:

[0090] ;

[0091] In the formula, For the first Mutually( ) Hydration degree (dimensionless) For time (unit: ), For the first Phase frequency factor (unit: ), The normalized reference value for the frequency factor (unit: ), For the first Phase activation energy (unit: ), The gas constant (take) ), Current temperature (unit: ), This is the normalized temperature reference value (unit: ), For the first The reaction order (dimensionless) was determined by fitting the reaction order of each phase using an isothermal hydration heat experiment in the laboratory. The normalized baseline value for hydration degree (dimensionless, taken as 1) is used. The above system of equations is integrated stepwise using the fourth-order Runge-Kutta method, with the integration step size... From the Courant stability condition Confirmed, among which Maximum reaction rate for each phase (unit: Peak temperature of hydration exothermic reaction. (Unit: °C) Estimated from the heat release during hydration of each phase:

[0092] ;

[0093] In the formula, The initial temperature of the slurry (unit: °C). Slurry density (unit: ), Normalized reference value for slurry density (unit: ), For the first The mass fraction of the phase in the curing agent, For the first Heat released during complete phase hydration (unit: ), For the first Final degree of hydration of the phase (dimensionless). Volumetric heat capacity of slurry (unit: ), Normalized reference value for volumetric heat capacity (unit: Infeasible solutions are mapped to the constraint surface using the gradient projection operator:

[0094] ;

[0095] In the formula, constraint function about Jacobian matrix, The residual vector represents the equality constraint. Elite individuals (top 10%) undergo additional local refinement using a limited-memory—Broyden—Fletcher—Goldfarb—Sharno—constraint algorithm, calculating gradients using automatic differentiation. ,in For the first Each target component.

[0096] The specific implementation of step S03 is as follows: weigh the powder components according to the optimal formula, and prepare a homogeneous powder mixture through a three-stage series mixing process. The coefficient of variation of the mixing homogeneity is:

[0097] ;

[0098] In the formula, Standard deviation of concentration at each sampling point (unit: ), Average concentration (unit: ), The content must be below 5%. The first-stage Venturi jet premixing achieves particle wetting, and the second-stage high-speed shear velocity is not less than 15. The third stage involves low-speed circulating mixing and homogenization. The combination of blade spacing and rotational speed is simulated using computational fluid dynamics (mesh accuracy not less than 1). ) Optimization and determination.

[0099] The specific implementation of step S04 is as follows: A solidified slurry is prepared by mixing the powder mixture with water. A multivariate adaptive model predictive control is used to establish a data-driven dynamic model of water-cement ratio, viscosity, temperature, and density. The state equation is:

[0100] ;

[0101] In the formula, For the first Step state vector (dimensionless). For the first Step viscosity (unit: ), Normalized viscosity reference value (unit: ), For the first Step density (unit: ), Density normalized reference value (unit: ), For the first Step temperature (unit: °C). For the first Step water-cement ratio, This is the normalized baseline value for the water-cement ratio. and The system matrix is ​​a data-driven identification system. For the first Step control input vector (including normalized values ​​of control quantities such as pump speed and valve opening). For the first Step-by-step modeling of the error vector, closed-loop adjustment period not exceeding 50. The online rheological sensor for the tuning fork vibration probe has a sampling rate of no less than 10. The viscosity correction value after temperature compensation is:

[0102] ;

[0103] In the formula, For the first Step-by-step viscosity measurement (unit: ), For calibration curve interpolation function (unit: ), This is a reference temperature (taken as 20℃). Viscosity after temperature compensation (unit: The thixotropic recovery kinetics surrogate model for slurry is based on online parameter identification of the Hoshka model, and the structural parameter dynamic equations are as follows:

[0104] ;

[0105] The shear stress equation is:

[0106] ;

[0107] In the formula, These are the slurry structure parameters (dimensionless). The structural recovery rate coefficient (dimensionless). Time-normalized reference value (unit: ), Structural failure rate coefficient (unit: ), Shear rate (unit: ), Shear stress (unit: ), Infinite shear viscosity (unit: ), Zero shear viscosity (unit: ), It is a structural index (dimensionless). Yield stress (unit: The initial parameter values ​​were obtained from the static recovery experiment (static time 0.5). Up to 5 Step size 0.5 (This was obtained through fitting.)

[0108] The specific implementation of step S05 is that the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics treats the formation as a saturated heterogeneous porous medium, and the grout seepage satisfies the nonlinear Darcy-Foceiheimer equation:

[0109] ;

[0110] In the formula, Pore ​​pressure (unit: ), For the first Directional spatial coordinates (unit: ), For the first Directional seepage velocity components (unit: ), Normalized reference value for seepage velocity (unit: ), To vary with the degree of curing Equivalent penetration rate with exponential decay (unit: ), Initial permeability (unit: ), The permeability decay coefficient (dimensionless). Inertial drag coefficient (unit: ), The current local solidification degree (dimensionless, ranging from 0 to 1) is estimated by integrating the state vector identified by the extended Kalman filter with the hydration kinetic equation. The extended Kalman filter state vector is:

[0111] ;

[0112] In the formula, The initial permeability normalized baseline value (unit: ), This is the normalized baseline value (dimensionless) for the permeability decay coefficient. Initial viscosity of the slurry (unit: ), Normalized reference value for initial viscosity of slurry (unit: ), Activation energy for slurry viscosity (unit: ), Normalized reference value for activation energy of slurry viscosity (unit: The state prediction equation is:

[0113] ;

[0114] The observation equation is:

[0115] ;

[0116] In the formula, For the first The prior state estimation vector is given. For the first The posterior state estimation vector is given by step. This is the process noise vector. For the first Step observation vector (including normalized values ​​of measured grouting pressure and flow rate). For nonlinear observation functions, The observed noise vector and the process noise covariance matrix. The curing uniformity index was obtained from the maximum likelihood estimation of indoor seepage calibration test data.

[0117] ;

[0118] In the formula, This represents the total number of observation points (25 nodes in a 5×5 grid). For the first Radius of diffusion at each observation point (unit: ), The mean diffusion radius at each observation point (unit: ), Normalized baseline value for diffusion radius (unit: The model predictive control rolling optimization problem is as follows:

[0119] ;

[0120] Constraints include , The maximum daily grouting volume is given by the formula. Grouting pressure (unit: ), Grouting flow rate (unit: ), The maximum rated pressure of the pump (unit: ), The first estimate for the Vese formula Critical pressure for fracturing of the formation (unit: ), This is the safety margin factor (empirical value is 0.1). For the first Predicted curing uniformity index for each control step (30 steps) The extended Kalman filter state is updated in real time and the optimization problem is solved again.

[0121] The specific implementation of step S06 is as follows: After grouting is completed, at least five parallel tests are conducted, recording the 28-day unconfined compressive strength, hydration exothermic peak value, and diffusion radius. The test data are then fed back to a hierarchical reinforcement learning-driven reconfigurable control network and a multi-objective evolutionary-gradient hybrid optimization algorithm. The high-level policy network of the hierarchical reinforcement learning-driven reconfigurable control network employs a proximal policy optimization algorithm with intrinsic incentives, and the objective function is:

[0122] ;

[0123] In the formula, These are trainable parameters for a high-level policy network. For the probability distribution of high-level strategies, This is the probability distribution of the high-level strategy before the update. For the first Step-up high-level actions (i.e., target recipe parameter vector). For the first Step soil geological feature vector, For the first Step dominance function estimate (dimensionless). This is the cutoff factor (empirical value is 0.2). Indicates will Cut off to interval Inside, The intrinsic incentive weighting coefficient (dimensionless). For the first Intrinsic incentive rewards (dimensionless). The intrinsic motivation normalization benchmark value (dimensionless). For time step The expected value operator. Intrinsic incentive reward is measured by the novelty of the latent space:

[0124] ;

[0125] In the formula, Current state The latent space vector mapped by the encoder, Let be the mean vector of the historical latent space state distribution. It is the Euclidean norm. The normalized baseline value for latent space distance (and) (With the same units of measurement), determined by the mean distance between historical latent space sample pairs during the pre-training phase. The low-level policy network employs the soft actor-critic algorithm, with the objective function being:

[0126] ;

[0127] In the formula, These are the trainable parameters for the low-level policy network. For the first Step-lower layer sensor state vector (dimensionless). For the first Pump pressure (unit: ), Pump pressure normalized reference value (unit: ), For the first Step-by-step measured flow rate (unit: ), Normalized baseline value for flow (unit: ), For the first Viscosity after temperature compensation (unit: ), For the first Step-lower level control action vector (including normalized values ​​of pump speed and valve opening). For the Critics Network (with) (for trainable parameters) The trainable parameters of the critic network are updated by gradient descent using the soft actor-critic algorithm. This is a temperature hyperparameter (dimensionless). This is the normalized baseline value for the temperature hyperparameter (dimensionless). This represents the probability distribution of low-level policies. The hypernetwork outputs a task description vector at higher levels. As input, dynamically generate all weight matrices for the low-level policy network:

[0128] ;

[0129] In the formula, This is the weight matrix of the low-level policy network. These are the trainable parameters of the supernetwork. It is a two-layer fully connected structure (with 512 and 256 neurons in the hidden layers, and linear rectified units as the activation function). Spectral normalization affects the weight matrix. By maximum singular value Normalization process:

[0130] ;

[0131] In the formula, The normalized weight matrix, Weight matrix The maximum singular value is used to constrain the Lipschitz constant of the network. The optimization objective of the commentator network's Lagrangian constraint layer is:

[0132] ;

[0133] In the formula, For the first The Lagrange multipliers corresponding to each constraint (initialized to 0). The normalized reference values ​​for Lagrange multipliers are... For the first A constraint function value (such as pump pressure or temperature, in units of...). Consistent) For the first Each constraint has an upper bound, and the multipliers are updated using the dual gradient ascent algorithm. Cumulative sum statistic. The update formula is:

[0134] ;

[0135] In the formula, For the first Step cumulative sum statistic (dimensionless). The reference sensor signal mean vector is used as the reference operating condition. The reference value for normalized signal deviation (and) (same dimensions) The allowable deviation coefficient (empirical value is 0.5) is used when... Exceeding the threshold Timely determination of operating condition switching The study was determined by analyzing the subject operating characteristic curves from historical operational switching data. The sample size in the priority experience playback strategy was [not specified]. The sampling priority is:

[0136] ;

[0137] In the formula, For the sample absolute value of time difference error (unit: ), To prevent small positive numbers with a priority of zero (empirical value is 0.01). The normalization constant is the sum of the priorities of all samples. The adaptive adjustment formula for the learning rate of the low-level policy network is:

[0138] ;

[0139] In the formula, The mean time difference error of the current batch (unit: ), Initial time difference error reference value (unit: ), The standard deviation of the sensor signal for the current batch. The initial batch standard deviation of the sensor signal. This indicates the number of updates completed. This is the maximum number of updates. and Determined by the statistical mean of the last 50 batches during the pre-training phase. At that time, the learning rate is taken ;when At that time, the learning rate is taken ;when At that time, the learning rate is taken , and These represent the maximum learning rate and the minimum protected learning rate, as determined by the learning rate sensitivity scan experiment during the pre-training phase (learning rate range). to The incremental training batch size is 32, determined by (logarithmically uniform sampling of 20 points).

[0140] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians selected a coastal reclamation soft soil foundation grouting and solidification project as a test scenario. Through the implementation of the entire process of on-site sampling, formula optimization, grout preparation and grouting scheduling, the actual operation effect of the method of the invention was recorded and analyzed.

[0141] The soft soil at the site was a typical high-water-content montmorillonite-dominated marine sedimentary soft soil, with a measured average water content of 87%, an estimated montmorillonite mass fraction of 52%, an estimated illite mass fraction of 28%, and an estimated kaolinite mass fraction of 20%, and a site temperature of 24℃. Technicians collected soil samples from five grid sampling points, measured the above parameters, constructed a soil geological feature vector, and input it into a high-level policy network of a hierarchical reinforcement learning-driven reconfigurable control network for inference, outputting a target formulation parameter vector.

[0142] Based on the target formulation parameter vector, a multi-objective evolutionary-gradient hybrid optimization algorithm constrained by hydration kinetics was initiated. After 200 generations of evolution and refinement using local gradients of elite individuals, the optimal formulation satisfying the on-site construction constraints was selected from the Pareto front formulation set. The optimal formulations determined after optimization are shown in Table 1.

[0143] Table 1. Optimal formulation component mass ratio table

[0144]

[0145] Weigh the powder components according to the formula in Table 1, and sequentially introduce them into a three-stage series mixing process. The linear velocity of the first-stage Venturi jet premixed gas flow is set to 22. The particle wetting is completed; the linear velocity of the second-stage high-speed shear rotor is set to 18. Break up agglomerates; the third-stage low-speed circulating impeller moves at 45 degrees. Homogenize for 10 minutes. After mixing, the concentration distribution of each component at 5 sampling points is measured, and the coefficient of variation is calculated. The result is 3.2%, which is lower than the target threshold of 5%, meeting the requirements for proceeding to the next step.

[0146] The slurry preparation stage then commences. A multivariate adaptive model predictive control system, with a target water-cement ratio of 0.82 as the initial setpoint, uses a closed-loop control to regulate the water addition. An online rheological sensor with a tuning fork vibrating probe operates at a 12... The viscosity signal was continuously acquired by sampling rate, corrected by temperature compensation algorithm, and then fed back to the control system. The measured slurry viscosity remained stable within the target range, with an adjustment cycle of 38ms, meeting the 50ms threshold requirement. After online identification of Houška model parameters, the peak pressure of pipeline start-up was predicted, and the pump speed ramp control command was output in advance. The safety valve did not malfunction during the grouting start-up process.

[0147] During the grouting scheduling phase, an adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics is initiated, using an extended Kalman filter with an initial permeability identification value. m^2$, with an attenuation coefficient identification value of 0.034, drives a 5×5 mesh finite difference model to calculate the solidification diffusion radius distribution. For example... Figure 2 As shown in the figure, the dynamic change curves of the diffusion radius of each grouting point during the scheduling process indicate that the solidification uniformity index (variance of diffusion radius at each observation point) continuously decreases with the increase of the number of scheduling steps, indicating that the model predictive control framework effectively drives the uniform diffusion of grout in the formation. The joint scheduling sequence of pressure and flow rate at each grouting point during the grouting process is shown in Table 2, and the constraints were not violated.

[0148] Table 2. Grouting Process Pressure-Flow Co-dispatch Record Table

[0149]

[0150] After grouting was completed, technicians arranged six sets of parallel tests, covering a water-cement ratio gradient of 0.75 to 0.90, and recorded the 28-day unconfined compressive strength and the peak value of hydration exothermic reaction. Figure 3 As shown in the figure, the distribution curve of the 28-day unconfined compressive strength as a function of water-cement ratio indicates that the strength reaches its peak value near the optimal water-cement ratio of 0.82, the peak temperature of hydration exothermic reaction remains within a reasonable range, and the risk of thermodynamic failure is effectively controlled. The main test results of each group of parallel tests are shown in Table 3.

[0151] Table 3. Record of peak intensity and hydration heat release during 28 days of parallel experiments.

[0152]

[0153] The parallel experimental data mentioned above are fed back to the hierarchical reinforcement learning-driven reconfigurable control network and multi-objective evolutionary-gradient hybrid optimization algorithm, triggering incremental training of the low-level policy network and correcting the Pareto front formula set, thus completing this round of model iteration update. Figure 4 As shown, after multiple iterations, the correlation between the target formula parameter vector output by the high-level policy network and the final measured intensity of the hierarchical reinforcement learning-driven reconfigurable control network continues to improve, indicating that the adaptive prediction capability of the model under dynamic working conditions gradually increases with the accumulation of field data.

[0154] Compared to traditional methods, this invention achieves the following technological advancements: Traditional formulation design relies on static indoor test results, failing to perceive spatial changes in on-site geological characteristics. In contrast, this invention uses a hierarchical reinforcement learning-driven reconfigurable control network to map geological feature vectors to formulation decisions in real time, eliminating the disconnect between formulation and on-site conditions from the information link structure. Traditional grouting control uses fixed parameter regulators to adjust pressure and flow, failing to track the non-stationary evolution of formation permeability during the solidification process. In contrast, this invention uses extended Kalman filtering to identify seepage states online, ensuring that scheduling decisions are always based on real formation physical parameters. Traditional multi-condition switching relies on manual intervention or offline model switching. In contrast, the mechanism of dynamically generating weights in this invention's hypernetwork enables strategy reconstruction to be completed within 100ms after condition switching, fundamentally solving the structural defect of insufficient generalization ability of control strategies under multiple conditions.

[0155] It should be noted that the variables involved in this invention are explained in detail in Tables 4, 5, and 6.

[0156] Table 4. Variable Explanation Table (Part 1)

[0157]

[0158] Table 5. Variable Explanation Table (Part Two)

[0159]

[0160] Table 6. Variable Explanation Table (Part 3)

[0161]

[0162] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for on-site preparation of an in-situ solidification agent for soft soil with high moisture content, characterized in that, Includes the following steps: Collect soft soil samples from the site, determine the moisture content, mineral composition and temperature, construct a soil geological feature vector, input it into a hierarchical reinforcement learning-driven reconfigurable control network, and output the target formula parameter vector. Based on the target formulation parameter vector, a multi-objective evolutionary-gradient mixing optimization algorithm based on hydration kinetic constraints is run to select the optimal formulation that satisfies the field constraints from the Pareto front formulation set and determine the mass ratio of each component. Weigh each curing agent component powder according to the optimal formula, and prepare a uniform powder mixture through a three-stage series mixing process; A solidified slurry is prepared by mixing powder mixture with water. The water-cement ratio is adjusted by multivariate adaptive model predictive control. The viscosity of the slurry is monitored in real time by an online rheological sensor with a tuning fork vibration probe and temperature compensation is performed to form a dual feedback closed-loop control. Based on the adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics, the joint scheduling sequence of grouting pressure and flow is solved in a rolling manner to drive the grouting pump to inject solidified grout into the soft soil layer; After grouting is completed, multiple parallel tests are conducted to record the unconfined compressive strength, peak hydration exothermic value and diffusion radius. The test data are then fed back to a hierarchical reinforcement learning-driven reconfigurable control network and a multi-objective evolutionary-gradient hybrid optimization algorithm to achieve iterative model updates.

2. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 1, characterized in that, The soil geological feature vector is specifically a state description vector composed of water content, estimated mineral composition, and on-site temperature as components, used to characterize the physicochemical properties of soft soil.

3. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 2, characterized in that, The hierarchical reinforcement learning-driven reconfigurable control network is specifically a two-timescale hierarchical partially observable Markov decision process consisting of a high-level policy network and a low-level policy network. The high-level policy network takes the soil geological feature vector as the state and outputs the target formula parameter vector. The low-level policy network takes the real-time sensing signal of the equipment as the state and the high-level sub-objective as the condition, and outputs pump speed and valve opening control actions.

4. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 3, characterized in that, In the hierarchical reinforcement learning-driven reconfigurable control network, the actor network uses a super network to dynamically generate the weight parameters of the low-level policy network; when the statistical process control module detects a change in operating conditions, the high-level network outputs a new network parameter vector to the super network, realizing real-time reconstruction of the policy network, with the reconstruction response time being lower than the reconstruction response time threshold.

5. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 4, characterized in that, The learning rate of the low-level policy network of the hierarchical reinforcement learning-driven reconfigurable control network is dynamically adjusted by an adaptive adjustment function. The adaptive adjustment function takes the mean of the current batch time difference error, the standard deviation of the current batch sensor signal, and the number of updates completed as inputs, outputs a learning rate adjustment coefficient, and selects the corresponding learning rate segmented according to the adjustment coefficient interval.

6. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 5, characterized in that, The multi-objective evolutionary-gradient hybrid optimization algorithm based on hydration kinetic constraints specifically employs a third-generation non-dominated sorting genetic algorithm for global population search in the outer layer, and an inner layer... , , , The four-phase hydration rate equation is solved using the fourth-order Runge-Kutta integral to verify its thermodynamic feasibility; elite individuals additionally execute the finite-memory-Broyden-Fletcher-Goldfarb-Shanno-constrained algorithm for local gradient refinement.

7. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 6, characterized in that, The three-stage series mixing process specifically involves the first stage using Venturi jet premixing to wet particles, the second stage using high-speed shearing to break up agglomerates, and the third stage using low-speed circulating stirring for homogenization. The combination of blade spacing and rotation speed is determined through computational fluid dynamics simulation optimization. The coefficient of variation for mixing uniformity must be reduced to below the coefficient of variation threshold.

8. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 7, characterized in that, The multivariable adaptive model predictive control specifically uses an online rheometer to measure viscosity and an online densitometer to measure density to form a dual feedback loop, establishing a data-driven dynamic model of water-cement ratio, viscosity, temperature, and density. The closed-loop adjustment period does not exceed the adjustment period threshold and is equipped with an anti-integral saturation mechanism.

9. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 8, characterized in that, The temperature compensation algorithm of the online rheological sensor of the tuning fork vibration probe corrects the viscosity measurement value in real time based on the viscosity-temperature relationship calibration curve. The thixotropic recovery kinetics proxy model for slurry is based on online parameter identification of the Houška model, predicts the peak pressure at pipeline start-up, and adjusts the pump speed ramp control strategy in advance.

10. The method for on-site preparation of the in-situ solidification agent for high moisture content soft soil according to claim 9, characterized in that, The adaptive pressure-flow coordinated grouting scheduling algorithm based on nonlinear seepage mechanics specifically uses the measured sequence of grouting pressure and flow as observations, employs extended Kalman filtering to identify the seepage state vector online, drives a simplified seepage finite difference model, and aims to minimize the solidification uniformity index. Within the model prediction and control framework, it continuously solves the joint scheduling sequence of grouting pressure and flow.

Images