Underneath passing existing station horizontal freezing construction effect data prediction and intelligent control method

Abstract

The invention relates to the technical field of civil engineering construction, and discloses an underneath pass existing station horizontal freezing construction effect data prediction and intelligent control method, which comprises the following steps: constructing a three-dimensional physical model, and collecting multi-source monitoring data in real time; based on simulation and real-time data, an AI agent model capable of predicting the construction effect in real time is generated and synchronously calibrated, and a digital twinning environment is formed; generating a control instruction set for performing differential adjustment on each freezing pipe through a hierarchical reinforcement learning method; quantizing the strategy uncertainty during decision making of the high-level intelligent agent, and generating an active sensing instruction when the strategy uncertainty is greater than a preset threshold value; and executing a control and sensing instruction, performing closed-loop control on physical construction, and optimizing a data acquisition process. According to the technical scheme, the AI agent model is adopted to deduce the construction effect in real time, the reinforcement learning agent generates the prospective control instruction, decision making can be conducted based on prediction of the future state, and active intervention on the freezing construction process is achieved.

Description

Technical Field

[0001] This invention relates to the field of civil engineering construction technology, specifically to a method for predicting and intelligently controlling the effects of horizontal freezing construction on existing railway stations. Background Technology

[0002] In modern urban development, underground engineering projects such as subway tunnels often need to pass under existing railway stations, highways, or subway stations. To ensure the safety of the existing structures above, the horizontal artificial ground freezing method is widely used. This technology freezes loose soil into a frozen soil wall with temporary support capacity by circulating a low-temperature medium in the soil. However, the volume expansion of water during the freezing process inevitably causes ground heave, posing a potential threat to the structural safety of existing stations. Therefore, accurate prediction and refined control of the frost heave effect throughout the construction process are core technical requirements for ensuring the safety of such projects and mitigating operational risks.

[0003] Existing technologies have developed a standardized implementation process to address such engineering challenges. Before construction begins, engineers typically use numerical simulation methods such as the finite element method to establish a physical model based on geological survey data. This model is used to pre-analyze the temperature and stress fields during the freezing process, providing an important theoretical basis for the preliminary design of the freezing scheme. During construction, a dense network of sensors is deployed at key locations on the existing station structure and strata to monitor key indicators such as temperature and displacement in real time. This monitoring data provides on-site engineers with direct and real physical quantity evidence to judge the freezing effect, make manual interventions, and activate emergency plans.

[0004] However, existing technologies still have inherent technical shortcomings in practice. First, there is a temporal disconnect between prediction and control. Pre-construction numerical simulations are static and cannot reflect unexpected changes during construction, while on-site monitoring is a post-construction observation. By the time significant anomalies appear in the monitoring data, adverse physical effects have already occurred, and control adjustments at this point can only serve as remedial measures, lacking foresight. Second, the global optimality of control decisions is difficult to guarantee. On-site control relies on engineers' experience and usually adjusts for local problems. This approach struggles to coordinate and balance multiple interdependent objectives such as structural safety, freezing effect, and system energy consumption, often resulting in suboptimal control behavior at the global level. Finally, there is a lack of an effective closed-loop correction mechanism between the prediction model and physical reality. There will inevitably be deviations between the initial physical model and actual geological conditions, and existing data acquisition is unidirectional and passive. The system cannot proactively identify which areas of its prediction model have the highest distortion and use this information to guide data acquisition in a targeted manner, leading to a decrease in model accuracy over time. Therefore, those skilled in the art have proposed a data prediction and intelligent control method for the horizontal freezing construction effects of tunneling under existing stations to address these problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting and intelligently controlling the effects of horizontal freezing construction under existing stations, which solves the problems of prediction and control lag, difficulty in decision optimization, and model accuracy decay in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for predicting and intelligently controlling the effects of horizontal freezing construction under existing stations. This method constructs a digital twin environment synchronized with the physical world and uses a hierarchical reinforcement learning architecture in this environment to perform forward-looking strategy optimization. At the same time, it combines an active perception mechanism driven by decision uncertainty to form a closed-loop technical solution that coordinates prediction, decision-making, control and perception.

[0008] The method specifically includes the following steps:

[0009] First, a three-dimensional physical model coupled with the temperature field, seepage field, and stress field is constructed to describe the physical and mechanical behavior of the soil during the freezing construction process.

[0010] Meanwhile, a multi-source monitoring network containing temperature, displacement, and stress sensors is deployed at the physical construction site, and multi-source monitoring data reflecting the frozen construction state in the physical world is collected in real time through data transmission technology.

[0011] Secondly, an AI agent model capable of predicting construction effects in real time is generated and simultaneously calibrated.

[0012] Specifically, the freezing construction system is abstracted into a graph structure containing sensor monitoring points and freezing pipe control units, and a graph neural network is constructed as an AI agent model. The AI ​​agent model is then trained offline using simulation data generated from the aforementioned three-dimensional physical model.

[0013] During construction, real-time multi-source monitoring data is continuously compared with the predictions of the AI ​​proxy model, and the model is calibrated online based on the deviation between the two. This calibrated AI proxy model constitutes a digital twin environment capable of rapidly extrapolating future construction effects.

[0014] Subsequently, within the digital twin environment, a set of control instructions is generated using a hierarchical reinforcement learning method. This method comprises a high-level agent and one or more low-level agents.

[0015] The high-level intelligent agent is responsible for global and macro-level decision-making. Based on the global construction effect state predicted by the AI ​​agent model, such as the predicted displacement of key structural points of existing stations and the global instability risk index, it generates a macro-level control strategy that dynamically divides all frozen pipes into functional zones, such as a strong attack zone, a stability maintenance zone, and an active pressure relief zone.

[0016] The lower-level agent is responsible for local and executive decision-making. Based on the macro-control strategy issued by the higher-level agent and the local state within its assigned zone, such as the real-time temperature of the freezing tubes within the zone, it generates specific and differentiated control parameters for each freezing tube in that zone. All the control parameters generated by the lower-level agents together constitute a control instruction set for future time periods.

[0017] Next, the policy uncertainty of the high-level agent in the hierarchical reinforcement learning method is quantified.

[0018] Specifically, the Monte Carlo Dropout method can be used to perform multiple random forward propagations on the policy network of high-level agents, and the variance of the resulting decision action distribution can be used to obtain the quantitative value of policy uncertainty.

[0019] When the quantification value is greater than a preset uncertainty threshold, the system determines that the AI ​​agent model has a high risk of data deviation in the physical area corresponding to the current state, and generates an active perception instruction accordingly.

[0020] The active sensing commands may include alarm commands that suggest manual reconnaissance of areas with high uncertainty, and parameter adjustment commands that automatically increase the sampling frequency of fixed sensor data in that area.

[0021] Finally, the instructions generated in the preceding steps are executed. The control instruction set is sent to the field programmable logic controller to automatically adjust the operating conditions of each freezing tube;

[0022] Simultaneously, the active sensing command is executed to optimize the acquisition process of the multi-source monitoring data. The entire method is executed cyclically according to a preset time period. After execution, the changes in the physical world state are re-acquired through multi-source monitoring data and used for the calibration of the AI ​​agent model and the generation of control decisions in the next round, thereby forming a continuously self-optimizing closed-loop control.

[0023] A second aspect of the present invention provides a data prediction and intelligent control system for the horizontal freezing construction effect of tunneling under existing stations. This system is used to implement the aforementioned method and is characterized by comprising:

[0024] The data and model fusion module is used to integrate the three-dimensional physical model containing temperature field, seepage field and stress field with the multi-source monitoring dataset collected in real time from the field sensor network, and output the fused system status information.

[0025] The real-time prediction agent module, which is connected to the data and model fusion module, is used to receive the fused system status information and generate a digital twin environment that can quickly predict future construction effects through an online calibrated AI agent model.

[0026] A hierarchical decision generation module, connected to the real-time prediction agent module, is used in the digital twin environment to perform policy inference through a hierarchical reinforcement learning architecture containing high-level and low-level agents, generate a set of control instructions for differentially adjusting each frozen tube, and generate active perception instructions based on the decision uncertainty of the high-level agent.

[0027] The closed-loop control execution module, which is connected to the hierarchical decision generation module, is used to receive the control instruction set and the active sensing instruction to drive the programmable logic controller and data acquisition system on site, and feeds back the changes in the physical world state after execution as new monitoring data to the data and model fusion module to form a closed loop.

[0028] This invention provides a method for predicting and intelligently controlling the effects of horizontal freezing construction when tunneling under existing stations.

[0029] It has the following beneficial effects:

[0030] 1. The present invention employs an AI agent model to simulate construction effects in real time and generates forward-looking control commands by a reinforcement learning agent. This enables decision-making based on predictions of future states and achieves proactive intervention in the frozen construction process. Compared with existing technologies that rely on manual adjustments after monitoring data exceeds limits, this invention solves the problem of difficulty in avoiding structural safety risks in advance due to delayed control response.

[0031] 2. This invention utilizes a hierarchical reinforcement learning architecture, in which a high-level agent performs global risk assessment and macro-level partitioning, while a low-level agent executes refined and differentiated local control. Through a global and local collaborative decision-making mechanism, it achieves comprehensive optimization of multiple objectives such as structural safety and system energy consumption. This differs from existing technologies that rely on experience for unified or local adjustments, and solves the problems of difficulty in multi-objective collaborative optimization and unnecessary energy waste.

[0032] 3. This invention quantifies the policy uncertainty in the control decision-making process and generates proactive perception instructions based on this, thereby optimizing the data acquisition process in reverse. It establishes a feedback loop from control decision-making to data perception, which can proactively identify and compensate for the shortcomings of its own model. Unlike the fixed mode in the prior art where data acquisition and control models are independent of each other, this invention solves the technical defect of its prediction model, which suffers from decreased fidelity due to its inability to dynamically align with physical reality. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the system functional modules of the present invention;

[0034] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0035] Figure 3 This is a schematic diagram of the functional architecture of the hierarchical reinforcement learning method of the present invention;

[0036] Figure 4 This is a schematic diagram of the active sensing instruction generation process of the present invention.

[0037] Among them, 10 is the data and model fusion module; 20 is the real-time prediction agent module; 30 is the hierarchical decision generation module; and 40 is the closed-loop control execution module. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] See attached document Figure 1 , Figure 1 This is a schematic diagram of the functional modules of a data prediction and intelligent control system for the horizontal freezing construction effect of an existing station underpass, according to an embodiment of the present invention. The system can be deployed on one or more server computing devices. The computing device includes a processor, a memory, and a communication interface. The memory stores a computer program, which, when executed by the processor, implements the following functional modules and methods.

[0040] This invention provides a data prediction and intelligent control system for the horizontal freezing construction effect of tunneling under existing stations. The system includes: a data and model fusion module 10, a real-time prediction agent module 20, a hierarchical decision generation module 30, and a closed-loop control execution module 40.

[0041] The data and model fusion module 10 integrates the initial input 3D physical model parameters with multi-source monitoring data collected in real time from the construction site sensor network. The 3D physical model is a pre-constructed coupled model containing temperature, seepage, and stress fields. The multi-source monitoring data includes temperature data collected from a distributed fiber optic temperature measurement system, displacement data collected from a hydrostatic level, and stress data collected from a stress gauge. The data and model fusion module 10 fuses the model parameters with the pre-processed monitoring data to output system state information representing the current physical world during the frozen construction phase.

[0042] A real-time prediction agent module 20 has its input connected to the output of the data and model fusion module 10. This module receives system state information and, through an internally integrated AI agent model, generates a digital twin environment capable of rapidly predicting future construction effects. In one embodiment, the AI ​​agent model is a graph neural network. The module uses the received system state information to perform online calibration of the graph neural network to continuously reduce the deviation between its predicted values ​​and the actual values ​​in the physical world. After calibration, the module outputs prediction results for the construction effects at one or more future time steps, and these prediction results constitute the future state of the digital twin environment.

[0043] A hierarchical decision generation module 30, whose input is connected to the output of the real-time prediction agent module 20, is used to generate control strategies in a digital twin environment. This module deploys a hierarchical reinforcement learning architecture comprising high-level and low-level agents. The high-level agent receives global features from the future state of the digital twin environment and generates a macro-level control strategy that functionally partitions all frozen tubes. The low-level agent receives the macro-level control strategy and the local state features within its assigned partition, generating differentiated control parameters for each frozen tube within that partition. The control parameters generated by all low-level agents together constitute a control instruction set.

[0044] Furthermore, the hierarchical decision generation module 30 is also used to quantify the uncertainty of the policy network output when the high-level agent makes decisions. When the calculated uncertainty value exceeds a preset threshold, the module generates an additional active perception instruction. The hierarchical decision generation module 30 outputs its generated control instruction set and active perception instruction together.

[0045] The closed-loop control execution module 40 has its input connected to the output of the hierarchical decision generation module 30. This module receives control instruction sets and active sensing instructions, and transmits them to the corresponding execution units in the physical world. Specifically, the control instruction sets are transmitted to the field programmable logic controller (PLC) controlling the operation of the freeze tube. The active sensing instructions are transmitted to the field management platform or data acquisition system to guide field personnel or adjust the operating parameters of the sensors.

[0046] After the closed-loop control execution module 40 executes the command, the state of the physical world changes. This change is collected as new multi-source monitoring data via the sensor network at the construction site and transmitted to the data and model fusion module 10 as input for the next calculation cycle, thus forming a complete and continuously operating closed loop.

[0047] See attached document Figure 2 , Figure 2 This is a flowchart illustrating a method for predicting and intelligently controlling the effects of horizontal freezing construction on existing stations under a tunnel, according to an embodiment of the present invention.

[0048] The method may include:

[0049] S100 constructs a three-dimensional physical model that includes temperature field, seepage field and stress field, and collects multi-source monitoring data in real time that reflects the frozen construction state of the physical world.

[0050] S200, based on simulation data generated from a 3D physical model and real-time multi-source monitoring data, generates and synchronously calibrates an AI agent model that can predict construction effects in real time, thus forming a digital twin environment.

[0051] In a digital twin environment, the S300 generates a set of control instructions for future time periods, which includes differentiated adjustments to each freeze tube, through a hierarchical reinforcement learning method that includes high-level and low-level agents.

[0052] S400 quantifies the policy uncertainty of high-level agents in hierarchical reinforcement learning methods when making decisions, and generates an active perception instruction when the policy uncertainty exceeds a preset threshold.

[0053] The S500 executes a set of control instructions to control the frozen construction of the physical world in a closed loop, and executes active sensing instructions to optimize the acquisition process of multi-source monitoring data, thereby realizing the iterative process of the method.

[0054] The following will provide a detailed explanation of each of the above steps.

[0055] In S100, a three-dimensional physical model incorporating temperature, seepage, and stress fields is constructed, and multi-source monitoring data reflecting the frozen construction state in the physical world are acquired in real time. This step provides the initial model and real-time data input for subsequent prediction and control.

[0056] Specifically, the construction process of the three-dimensional physical model begins with integrating the input data. This data includes: physical and mechanical parameters such as soil layer distribution, porosity, density, specific heat capacity, thermal conductivity, and permeability obtained through engineering geological surveys; information such as station structural dimensions, material properties, and foundation types obtained through existing station building information models (BIM) or design drawings; and design parameters such as the layout, diameter, and design temperature of the freezing pipes for this horizontal freezing project.

[0057] Based on the above input data, a three-dimensional finite element or finite difference model coupled with a thermo-hydraulic-mechanical (THM) multiphysics field is established in a numerical simulation software environment. This model is defined by a set of partial differential governing equations.

[0058] The temperature field of the model follows the law of conservation of energy, and its governing equations are:

[0059]

[0060] In the formula: T is temperature; t is time Q T The internal heat source per unit volume refers to the cooling capacity provided by the freezing pipe; C is the equivalent volumetric heat capacity; λ is the equivalent thermal conductivity.

[0061] Considering the water-ice phase transition process, C and λ are defined as follows:

[0062]

[0063] In the formula: ρ s ρ w ρ i These are the densities of the soil skeleton, water, and ice, respectively; c s c w c i λ represents the specific heat capacity of the soil skeleton, water, and ice, respectively; s , λ w , λ i λ represents the thermal conductivity of the soil skeleton, water, and ice, respectively; n represents the soil porosity; I m is the volume fraction of ice per unit pore volume; L is the latent heat of phase change of water; w u The unfrozen water content is a function of temperature T.

[0064] The seepage field of the model follows the law of conservation of mass and Darcy's law, and its governing equations are:

[0065]

[0066] In the formula: v w Q is the seepage velocity vector of water; M is the source and sink term per unit volume; k is the soil permeability tensor, whose value is a function of temperature and ice content; μw p is the dynamic viscosity of water. w ρ is the pore water pressure; g is the gravitational acceleration vector.

[0067] The stress field of the model obeys the law of conservation of momentum, and its governing equations are:

[0068]

[0069] Its constitutive relation is:

[0070] dσ=D ep (dε-dε th );

[0071] In the formula: σ is the total stress tensor; F v dσ is the volume force vector; dε is the total stress increment tensor; dε is the total strain increment tensor; th This is the thermal strain increment tensor, which includes the frost heave effect caused by water freezing into ice; D ep This is an elastoplastic matrix, which reflects the mechanical properties of soil under different temperature and stress states.

[0072] While constructing the three-dimensional physical model, a multi-source sensor network was deployed at the physical construction site. This network deployment followed the principle of key point coverage. Specifically, a distributed fiber optic temperature measurement system was deployed along the direction of the freezing pipe to obtain the continuous temperature field distribution of the frozen wall. Static levels and inclinometers were deployed at key structural locations such as the existing station's foundation slab and load-bearing columns to monitor the vertical and horizontal displacements of the structure. Stress gauges and pore water pressure gauges were deployed in the soil between the frozen wall and the existing station structure to monitor stress and water pressure changes in key areas.

[0073] Finally, using IoT technology, all monitoring data collected by the sensor network is transmitted in real time to a central data server via wireless or wired means. The server performs preprocessing operations such as data cleaning, noise filtering, and time series alignment on the received raw data to form a standardized multi-source monitoring dataset that can be directly used in subsequent steps.

[0074] In S200, based on simulation data generated from the 3D physical model constructed in S100, and multi-source monitoring data collected in real time from S100, an AI proxy model capable of predicting construction effects in real time is generated and simultaneously calibrated to form a digital twin environment. This step aims to address the problem of time-consuming calculations for 3D physical models, providing a rapid prediction tool that can be used for real-time decision-making.

[0075] This step begins with the offline generation of a training dataset. Using a 3D physical model built in S100, extensive simulation calculations are performed by setting different boundary conditions, combinations of soil parameters, and freezing pipe control sequences. Each simulation recreates a complete freezing construction process, generating data on the evolution of the system state over time, including temperature, displacement, and stress at each node. This data is recorded to form a simulation dataset encompassing multiple working conditions.

[0076] Secondly, the network structure of the AI ​​agent model is constructed. In one embodiment, the AI ​​agent model adopts a graph neural network (GNN) structure. The construction process includes abstracting the construction system into a graph structure G = (V, E). The node set V includes: all sensor monitoring points deployed in S100, several key structural points of focus on the existing station, and each independently controllable freezing pipe unit. The edge set E is established based on the physical influence relationships and spatial proximity relationships between nodes; for example, connections are established between two spatially adjacent sensor nodes, and connections are established between a sensor node and its nearest structural point.

[0077] Then, the graph neural network is trained offline. The goal of this training is to improve the graph neural network model f. GNN Learning a mapping from the current system state and control inputs to the future system state can be represented as:

[0078]

[0079] In the formula: t is the current time; k is the prediction time step; X t Let A be the node feature matrix at time t. Each row of this matrix corresponds to a node in the graph and contains the node's current monitored value (e.g., temperature, displacement) and static attributes (e.g., node type, spatial coordinates). t This is the control action matrix applied to each node at time t, and its main content is the power or flow regulation value applied to each freeze tube control unit. The nodal state matrix predicted by the model at future time t+k includes predicted values ​​for physical quantities such as temperature and displacement of each node; Θ GNN This represents all trainable parameters included in the graph neural network model.

[0080] The training process uses the aforementioned generated simulation dataset, and the parameter Θ is adjusted through an optimization algorithm. GNN To minimize the predicted value The error between the actual value and the value in the simulation dataset.

[0081] Finally, after physical construction begins, the AI ​​agent model is calibrated online synchronously. At each time period, the system uses multi-source monitoring data collected in real-time from S100 as the real physical world state Y. t Compare the actual state with the predictions made by the AI ​​agent model at the corresponding time point. The two are compared, and the deviation or residual between them is calculated. This deviation is then fed into an online learning or calibration algorithm, such as an incremental learning algorithm, to fine-tune the parameters Θ of the graph neural network model. GNN This online calibration mechanism operates continuously, ensuring that the digital twin environment created by the AI ​​agent model remains synchronized with the physical construction site.

[0082] See attached document Figure 3 , Figure 3 This is a schematic diagram of the functional architecture of a hierarchical reinforcement learning method according to an embodiment of the present invention.

[0083] In step S300, within the digital twin environment generated by step S200, a hierarchical reinforcement learning method comprising high-level and low-level agents is used to generate a set of control instructions for future time periods, which includes differentiated adjustments to each frozen tube. This step is the core of intelligent decision-making in this invention.

[0084] This hierarchical reinforcement learning method decomposes the complex global control problem into two levels: macro-policy planning and micro-execution control, which are handled by high-level agents and low-level agents, respectively.

[0085] The high-level intelligent agent performs global macro-level policy planning. Its state space S h The input, i.e., state s h This is a feature vector extracted from the global state predicted by the AI ​​agent model. The vector includes: the maximum predicted displacement value among all key structural points of the existing station, the average temperature and temperature field uniformity index characterizing the overall morphology of the frozen wall, and the global instability risk index calculated based on the predicted stress field.

[0086] The output of the action space G of the high-level intelligent agent, i.e., action a h , is a macro-level goal or sub-task g. In one embodiment, the action a h It dynamically divides all the freezing tubes in the physical world into several functional zones, such as a strong attack zone, a stable maintenance zone, and an active pressure relief zone, and sets an overall control target for each zone.

[0087] The learning objective of a high-level intelligent agent is to maximize a pre-defined long-term cumulative reward. Its reward function R at a single time step... h Defined as:

[0088] R h=-w disp ·f disp (max(d pred ))-w ener ·E total -w risk ·S risk ;

[0089] In the formula: d pred f represents the displacement vectors of all key structural points in the station predicted by the AI ​​agent model. disp It is a superlinear penalty function for the maximum displacement; the larger the displacement, the higher the penalty value. total S represents the predicted total energy consumption of the system in the next time period. risk This refers to the aforementioned global instability risk index; w disp w ener and w risk These are the weighting coefficients corresponding to the three objective items: structural deformation, system energy consumption, and global risk, respectively.

[0090] Lower-level agents are responsible for executing fine-grained local control. Each functional area divided by higher-level agents corresponds to a lower-level agent. The state space S of the lower-level agents... l The input, i.e., state s l It consists of two parts: one part is the local state s within the partition it is responsible for. local The data includes the real-time temperature of each freezing pipe in the area and the stress state of the adjacent soil; another part is the macroscopic target g received from the high-level intelligent agent.

[0091] Action space A of the low-level intelligent agent l The output, i.e., action a l This refers to the specific control parameter value for each freezing pipe within its assigned zone. In one embodiment, this parameter value is the adjustment value for the percentage of cooling power or refrigerant flow rate of each freezing pipe.

[0092] The learning objective of a low-level agent is to maximize its local reward while following instructions from higher levels. Its reward function R at a single time step... l Defined as:

[0093] R l =w goal ·f goal (s local ,g)-w eloc ·E local -w unif ·U local ;

[0094] In the formula: f goal It is a measure of the current local state s localE is a function of the degree to which the macroscopic goal g is achieved. local This represents the local energy consumption of this functional zone; U local This represents the variance of the temperature of each freezing pipe within the zone, used to characterize the local uniformity of the temperature field; w goal w eloc and w unif These are the corresponding weighting coefficients.

[0095] In each decision cycle, the high-level agent first generates a macro-level control strategy based on the global state. Then, each low-level agent generates refined control parameters in parallel based on the macro-level strategy and their respective local states. The control parameters output by all low-level agents are aggregated to form a complete and differentiated set of freeze control instructions for a future time period.

[0096] See attached document Figure 4 , Figure 4 This is a schematic diagram of the active sensing instruction generation process according to an embodiment of the present invention.

[0097] In S400, the policy uncertainty of the high-level agent in the hierarchical reinforcement learning method of S300 is quantified when making decisions, and an active perception instruction is generated when the policy uncertainty exceeds a preset threshold. This step aims to evaluate the reliability of the control decision and, if the reliability is insufficient, actively correct model bias by optimizing data perception.

[0098] This step is performed on a high-level agent in a hierarchical reinforcement learning method, whose policy network integrates a Dropout layer during training. During the decision-making phase, this Dropout layer remains active.

[0099] Specifically, the process of quantifying policy uncertainty employs the Monte Carlo Dropout method. This is achieved by acquiring the current high-level state s. h Through the policy network π of high-level intelligent agents h Perform N random forward propagations. Due to the presence of the Dropout layer, each forward propagation randomly deactivates some neurons in the network, resulting in a slightly different network structure and outputting a different macroscopic control policy action. These N forward propagations ultimately yield a distribution set of decision actions {a}. h,1 ,a h,2 ,...,a h,N}

[0100] Strategy uncertainty The variance, defined as the set of decision action distributions, is calculated as follows:

[0101]

[0102] Where: πh (s h |Θ j ) represents the state s in the j-th forward propagation. h and the effective network parameter subset Θ generated by this random Dropout j The policy network outputs the decision action; Var[·] is the function for calculating the variance; N is the total number of random forward propagations. A large variance value indicates the policy network's response to the current state s. h There is considerable disagreement regarding the optimal decision under these circumstances.

[0103] Subsequently, the calculated strategy uncertainty With a preset uncertainty threshold θ U The comparison is performed. This threshold can be preset based on the project's safety level and historical data.

[0104] When the condition is met At that time, the system determines that the AI ​​agent model is in relation to the current state s h The relevant physical area has a high risk of data bias or model distortion, and an active sensing command is automatically triggered. In one embodiment, the proactive sensing instruction comprises two parts: one part is an alarm instruction sent to the field management platform to suggest that field personnel conduct manual surveys of the physical area pointed to by the source of uncertainty or deploy mobile sensing devices; the other part is a parameter adjustment instruction sent to the data acquisition system to automatically increase the data sampling frequency and upload frequency of fixed sensors in the physical area.

[0105] In step S500, the control instruction set generated in step S300 and the active sensing instruction generated in step S400 are executed to achieve closed-loop control of the frozen construction in the physical world and optimize the acquisition process of multi-source monitoring data, thus realizing the iterative iteration of the method. This step is the link connecting the virtual decision-making world and the physical execution world and forms the basis of the closed-loop characteristic of the method of this invention.

[0106] In one embodiment, this step first includes an instruction review process. The control instruction set generated in S300 and the active sensing instructions generated in S400 are submitted to a human-machine interface before being sent to the physical execution unit. On-site engineers or technical managers conduct a final review and confirmation of the instructions' rationality and safety on this interface. Only after the review is passed is the instruction allowed to be sent.

[0107] After the command is issued, the control command set is transmitted to the programmable logic controller (PLC) or equivalent automation control unit deployed at the construction site. Based on the specific parameter values ​​set for each freezing pipe in the received command set, the PLC automatically adjusts the operating conditions of the refrigeration unit connected to it, such as adjusting the valve opening of the refrigerant or the power of the circulation pump, thereby achieving differentiated and precise control of each freezing pipe.

[0108] Simultaneously, if an active sensing command is generated, it is sent to the corresponding execution terminal. Alarm commands are sent to the field management platform, prompting managers, either textually or graphically, to pay attention to or manually verify specific physical areas. Parameter adjustment commands are sent to the data acquisition system deployed in the S100, which adjusts the parameters of the specific sensors it manages, such as increasing their data sampling frequency or data upload frequency.

[0109] The closed-loop characteristic of this method is achieved through iterative iteration. After the command in S500 is executed, the state of the physical world (including soil temperature field, stress field, station structure displacement, etc.) will change accordingly. These changes are captured by the multi-source sensor network deployed in S100 and collected as real-time monitoring data for the new cycle. Subsequently, the system repeats all steps from S100 to S500 at a preset time period (e.g., 15 minutes or 30 minutes). In the new cycle, the system uses the latest monitoring data containing the control effects of the previous cycle to calibrate the AI ​​agent model (S200) and generate new control and sensing commands that are more adapted to the current state (S300, S400), thus forming a closed-loop control system that continuously optimizes and evolves.

[0110] To further illustrate the collaborative working process of the technical solution of this invention, a specific working scenario example will be used below.

[0111] The engineering scenario in this example involves a subway tunnel horizontally passing under an operating shallow-foundation three-story frame-structure railway station in a city. The geological conditions of the site consist of an upper saturated soft soil layer and a lower layer of sandy soil with high bearing capacity. The key technical challenge of this project lies in the need to strictly control the heave deformation of the station structure caused by the freezing construction, with clear limits on both the heave rate and the cumulative heave value.

[0112] During the construction preparation phase, according to S100, a three-dimensional thermo-hydraulic-mechanical coupled physical model corresponding to the scenario was first established based on the geological survey report and the station's structural design drawings. Simultaneously, static levels were installed at locations such as the station's base slab and key load-bearing columns C5 and C6, and distributed temperature-measuring optical fibers were laid around the freezing pipes. Subsequently, according to the offline training section of S200, multi-condition simulations were performed using this physical model to generate a simulation dataset for training the AI ​​agent model, and offline pre-training of the graph neural network agent model was completed.

[0113] On the 15th day after the frozen construction entered a stable development phase, the intelligent control system of this invention operated in a preset 30-minute cycle. At the beginning of a specific cycle:

[0114] S100 was executed, and the system collected the latest multi-source monitoring data from the field sensor network. The data showed that the status of the other monitoring points was stable, but the hydrostatic level located below the load-bearing column C5 detected a slight, continuous increasing trend in its bulging rate.

[0115] S200 was executed, and the AI ​​agent model received the latest monitoring data, including the change in the bulging rate of load-bearing column C5, and completed an online calibration based on this data. Subsequently, the model quickly extrapolated the construction effects over the next 24 hours. The extrapolation results showed that if the current freezing pipe control strategy continued, the cumulative bulging value of load-bearing column C5 would exceed its warning threshold within the next 18 hours.

[0116] S300 is executed, and the hierarchical decision generation module receives the aforementioned prediction results. The input state of the high-level agent includes the risk of C5 column exceeding the warning level. Based on this, the high-level agent generates a macro-control strategy, designating the area directly affecting the three freezing tubes (numbered F10, F11, F12) below C5 column as the "active pressure relief zone," and setting the surrounding freezing tube area as the "stability maintenance zone."

[0117] Subsequently, the lower-level agent responsible for the "active pressure relief zone" receives the macro-level command and real-time temperature data of the three freezing tubes F10, F11, and F12 within the zone. To achieve the pressure relief target, the lower-level agent generates specific control parameters: reducing the cooling power of freezing tube F11, which is directly below column C5, by 15%, while increasing the cooling power of freezing tubes F10 and F12 on either side of it by 5% to ensure the overall continuity of the frozen wall. The lower-level agents responsible for other areas maintain their original control parameters. These parameters together constitute the control command set for this cycle.

[0118] S400 was executed, and simultaneously with S300, the system quantified the uncertainty of the decision-making process of the high-level agent. The calculation results showed that the uncertainty of the high-level agent's decision regarding the "active pressure relief zone" was below the threshold. However, when making decisions about another area B, which is far from the main station structure, the calculated uncertainty exceeded the preset threshold due to the complex geological conditions and some noise in the sensor data. Therefore, the system generated an additional active sensing instruction: suggesting manual inspection of area B and automatically increasing the data sampling frequency of temperature sensor T31 in that area from once every 30 minutes to once every 10 minutes.

[0119] S500 is executed, and the generated control instruction set and active sensing instructions are submitted to the engineer's review interface. After the engineer confirms that everything is correct, the instructions are issued. The field PLC receives the control instructions and adjusts the cooling power of the freezing tubes F10, F11, and F12 accordingly. At the same time, the data acquisition system receives the sensing instructions and automatically adjusts the sampling frequency of sensor T31.

[0120] This completes a full decision-making and control cycle. In the next cycle, the system will collect higher-quality monitoring data after control adjustment and perception optimization, thereby initiating a new round of more accurate prediction and control, keeping the entire construction process under continuous closed-loop optimization management.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and intelligently controlling the effects of horizontal freezing construction under existing stations, characterized in that: Includes the following steps: A three-dimensional physical model containing temperature field, seepage field and stress field is constructed, and multi-source monitoring data reflecting the frozen construction state of the physical world are collected in real time. Based on the simulation data generated by the three-dimensional physical model and the real-time collected multi-source monitoring data, an AI agent model that can predict construction effects in real time is generated and synchronously calibrated, and the AI ​​agent model constitutes a digital twin environment. In the digital twin environment, a hierarchical reinforcement learning method that includes high-level and low-level agents is used to generate a set of control instructions for future time periods, which includes differentiated adjustments to each freezing tube. The policy uncertainty of the high-level agent in the hierarchical reinforcement learning method is quantified, and when the policy uncertainty is greater than a preset threshold, an active perception instruction is generated. The control instruction set is executed to control the frozen construction of the physical world in a closed loop, and the active sensing instruction is executed to optimize the acquisition process of the multi-source monitoring data.

2. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 1, characterized in that, The construction of a three-dimensional physical model including temperature field, seepage field, and stress field, and the real-time acquisition of multi-source monitoring data reflecting the frozen construction state in the physical world, includes: The three-dimensional physical model was established by integrating engineering geological parameters, existing station structural information, and frozen pipeline design. A sensor network, including a distributed fiber optic temperature measurement system, a hydrostatic level, and a stress gauge, is deployed at the construction site. The temperature, displacement, and stress data collected by the sensor network are aggregated in real time to the data center using Internet of Things (IoT) technology, serving as the multi-source monitoring data.

3. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 1, characterized in that, The generated and synchronously calibrated AI proxy model capable of predicting construction effects in real time includes: The construction system is abstracted as a graph structure, where nodes are sensor monitoring points and freezing pipe control units, and edges are the physical influence relationships between nodes. The graph structure is used to construct a graph neural network as the AI ​​agent model; The graph neural network is trained offline using simulation data generated from the three-dimensional physical model. During construction, the graph neural network is calibrated online by using the deviation between the real-time collected multi-source monitoring data and the predicted values ​​of the AI ​​agent model.

4. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 1, characterized in that, The generation of a control instruction set oriented towards future time periods, which includes differentiated adjustments to each freezing tube, includes: Based on the global construction effect state predicted by the AI ​​agent model, the high-level intelligent agent generates a macro-control strategy that dynamically divides all frozen pipes into functional zones. The lower-level agent generates specific control parameters for each freezing tube in the partition based on the macro-control strategy issued by the higher-level agent and the local state within its assigned partition, thus forming the control instruction set.

5. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 4, characterized in that, The macro-control strategy generated by the high-level intelligent agent is as follows: The input state of the high-level intelligent agent is the maximum displacement of the station's key structural points, the global frozen wall uniformity index, and the global instability risk index predicted by the AI ​​agent model. The output action of the high-level intelligent agent is a partitioning command to divide the frozen pipeline network into a strong attack zone, a stable maintenance zone, and an active pressure relief zone. The reward function of the high-level intelligent agent is set as a weighted combination of maximizing structural security, minimizing total energy consumption, and minimizing global risk.

6. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 4, characterized in that, The specific control parameters for the generation of the low-level intelligent agent are as follows: The input states of the lower-level intelligent agent are the real-time temperature of the frozen pipe in the area it is responsible for, the local soil stress state, and the zoning instructions issued by the higher-level intelligent agent. The output action of the lower-level intelligent agent is to adjust the cooling power or flow rate of each freezing pipe in the area it is responsible for. The reward function of the low-level agent is set as a weighted combination of maximizing macroscopic instruction fulfillment, minimizing local energy consumption, and minimizing local temperature non-uniformity.

7. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 1, characterized in that, The generated active sensing instructions include: The Monte Carlo Dropout method is used to perform multiple random forward propagations on the policy network of the high-level agent to obtain a decision action distribution. Calculate the variance of the distribution of the decision actions to quantify the uncertainty of the strategy; The calculated uncertainty of the strategy is compared with a preset uncertainty threshold. When the uncertainty exceeds the threshold, the active perception instruction is generated.

8. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 7, characterized in that, The active sensing instructions include: It is recommended to conduct manual surveys or deploy alarm commands from mobile sensing devices for the physical areas corresponding to the uncertainties in the aforementioned strategy. Automatically increase the parameter adjustment instructions for the data sampling frequency and upload frequency of fixed sensors within the physical area corresponding to the uncertainty of the strategy.

9. The method for predicting and intelligently controlling the horizontal freezing construction effect of tunneling under existing stations according to claim 1, characterized in that, The execution of the control instruction set and the execution of the active sensing instruction include: The control instruction set is sent to the programmable logic controller on site to automatically adjust the operating conditions of each freezing tube; The active sensing command is sent to the field management platform to guide encrypted observation or adjustment of the sensor's sampling frequency; The entire process is executed cyclically according to a preset time period, forming a closed-loop control system that continuously optimizes itself.

10. A data prediction and intelligent control system for the horizontal freezing construction effect of tunneling under existing stations, applied to the data prediction and intelligent control method for the horizontal freezing construction effect of tunneling under existing stations as described in any one of claims 1-9, characterized in that, include: The data and model fusion module is used to integrate a three-dimensional physical model containing temperature field, seepage field and stress field with a multi-source monitoring dataset collected in real time from the field sensor network, and output the fused system status information. The real-time prediction agent module is used to receive the fused system status information and generate a digital twin environment that can quickly predict future construction effects through an online calibrated AI agent model. The hierarchical decision generation module is used in the digital twin environment to perform policy inference through a hierarchical reinforcement learning architecture containing high-level and low-level agents, generate a set of control instructions for differentially adjusting each freezing tube, and generate active perception instructions based on the decision uncertainty of the high-level agents. The closed-loop control execution module is used to receive the control instruction set and the active sensing instruction, drive the programmable logic controller and data acquisition system on site, and feed back the changes in the physical world state after execution as new monitoring data to the data and model fusion module to form a closed loop.

Images