Mass concrete curing system and method based on AI intelligent agent

Abstract

The invention discloses a mass concrete curing system and method based on an AI intelligent agent, and the core innovation of the method lies in that the AI intelligent agent with the deep reinforcement learning (DRL) capability is introduced as a decision-making core, the curing system is upgraded from automation to intelligence, and the fundamental transformation from passive response to active prediction and collaborative optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and intelligent construction technology in civil engineering, and in particular to a large-volume concrete curing system and method based on AI intelligent agents. Background Technology

[0002] Constructing large-volume concrete in high-altitude, cold regions presents severe challenges due to the extreme environment. These areas experience significant diurnal temperature variations, dry air, high wind speeds, and unpredictable wind directions. After concrete pouring, the heat released during cement hydration interacts with the complex external environment, easily leading to excessive temperature differences between the inside and outside of the structure and rapid surface moisture evaporation. This can trigger temperature stress cracks and shrinkage cracks, seriously threatening the durability and safety of the structure, a fact well-recognized by the engineering community. Traditional curing methods to address this global challenge rely heavily on manual experience, employing techniques such as regular watering, covering with insulation materials (like straw mats and blankets), or erecting heated sheds in the concrete pouring area, using integrated heat storage methods or heated sheds to maintain curing temperatures. However, these methods suffer from inherent drawbacks such as delayed response, inefficient control, and resource waste, making it difficult to achieve precise and uniform temperature and humidity control, and especially unable to effectively suppress the significant temperature gradient generated within the concrete due to the heat of hydration.

[0003] In recent years, with the development of sensing and automation technologies, intelligent curing technology has begun to be applied to large-volume concrete projects. This type of technology typically involves deploying distributed temperature sensors and constructing cooling water circulation systems (such as burying serpentine cooling water pipes) and intelligent spraying subsystems. By monitoring the internal temperature of the concrete in real time, it automatically adjusts the cooling water flow rate or spraying strategy to achieve temperature control and crack prevention. Some systems also attempt to introduce heat recovery devices to improve energy efficiency. However, most of these systems still fall under the category of "automation," and their control logic often relies on preset fixed thresholds or rules (e.g., starting cooling water circulation when the temperature at a monitoring point exceeds a certain set value). They lack a deep understanding and dynamic extrapolation ability regarding the physical laws of concrete hydration thermodynamics and heat transfer, and cannot perform forward-looking predictions and global collaborative optimization like human experts. Therefore, they are difficult to effectively cope with the complex working conditions of rapidly changing and uneven heat dissipation in high-altitude areas, especially when sensor data is missing or the working conditions exceed preset ranges, the control effect will be greatly reduced.

[0004] As artificial intelligence (AI) technology penetrates the engineering field, AI has been attempted for use in the temperature and humidity curing decisions of large-volume concrete. These systems, through intelligent agents, output control commands to actuators such as cooling, spraying, and insulation, hoping to replace human experience and achieve more precise and energy-efficient curing. However, in actual deployment, such AI agents have exposed a significant "hallucination" problem. In an engineering context, "hallucination" refers to the agent generating seemingly reasonable decision sequences that violate physical laws, engineering specifications, or actual conditions. The root cause lies in the fact that the core of these AI agents is usually a complex probabilistic model that learns the mapping relationship between states and actions through trial and error, aiming to maximize cumulative rewards. However, it does not truly "understand" the physical laws of concrete hydration thermodynamics and heat transfer; its decisions are entirely based on statistical correlations in training data (or simulation experience). When encountering extreme conditions outside the training data distribution (such as specific wind speeds, sunshine, and low-temperature combinations unique to high-altitude areas and never fully appearing in the training data), the model tends to make "probabilistic guesses" based on existing patterns rather than reasoning based on physical principles, resulting in various decisions that violate facts and common sense in engineering. For example, there may be a fact-conflicting illusion: when the internal temperature of the concrete is already high, the agent may still decide to reduce cooling and strengthen insulation, causing the temperature difference between the inside and outside of the concrete (ΔT1(t)) to far exceed the safe threshold of 25°C; there may also be an illusion of fabrication: in areas lacking effective humidity sensor data, the agent "imagines" that the humidity is sufficient, thus deciding to stop spraying, causing shrinkage cracks on the concrete surface; there are also illusions of instruction misunderstanding and logical error: failing to correctly understand the composite goal of "prioritizing energy saving under the constraint of temperature difference," getting stuck in local optima, such as shutting down all equipment for a long time in pursuit of extremely low energy consumption, completely deviating from the fundamental purpose of curing. These illusion problems seriously restrict the reliability and effectiveness of AI technology in concrete curing and are key technical defects that urgently need to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a large-volume concrete curing system and method based on AI intelligent agents to address the above-mentioned shortcomings, thereby solving the problem that existing quasi-static analysis methods are insufficient to accurately reflect the time-varying effects throughout the construction process of railway-related rotating bridges.

[0006] This invention is achieved through the following scheme: A large-volume concrete curing system based on an AI agent includes a sensing system, an AI agent decision-making system, and a collaborative execution system. The sensing system is deployed inside the concrete structure, on its surface, and in the environment. The AI ​​agent decision-making system is used to collect data from the sensing system and issue control commands according to corresponding strategies. The collaborative execution system is used to execute the control commands issued by the AI ​​agent decision-making system.

[0007] The collaborative execution system includes a zoned intelligent sprinkler system, an adjustable insulation / heating system, and a central intelligent water cooling system; the sprinkler heads in the zoned intelligent sprinkler system are independently controlled by the AI ​​intelligent agent decision-making system; the sprinkler network is arranged in zones according to the BIM model.

[0008] The adjustable insulation / heating system uses independently controlled intelligent air-fin membranes as the insulation layer; the AI ​​intelligent decision-making system dynamically adjusts the inflation pressure of each air-fin membrane and the opening of local ventilation openings according to the real-time heat dissipation of each facade; the internal intelligent water cooling system: the cooling water pipeline is arranged in layers inside the concrete.

[0009] This solution also discloses a method for curing large-volume concrete based on an AI agent, using the system described in any one of claims 1 to 3, characterized by comprising the following steps: Step S1: Construct a perception system, an AI intelligent agent decision-making system, and a collaborative execution system; Step S2: Optimize and train the AI ​​agent decision-making system; Step S3: The optimized and trained AI agent decision-making system is used to perform intelligent curing of the concrete to be cured.

[0010] In step S2, the AI ​​agent decision-making system is optimized and trained, which includes the following steps: Step S21: Establish simulation training, knowledge-based simulation pre-training—laying a factual foundation; Step S22, reinforcement reward learning, deep reinforcement learning based on reinforcement rewards—constrained optimization, by establishing reinforcement reward function Rt, basic reward R, uncertainty penalty term Ut, the agent shifts from imitation to autonomous optimization, but is strictly constrained within the boundary of facts and knowledge through reinforcement reward function; Step S23, online alignment of experts in the loop—achieving human-machine collaborative evolution.

[0011] Step S21 specifically includes the following steps: Step S211, Constructing the simulation environment: In addition to integrating the traditional thermodynamic model, a deep learning-based concrete hydration heat-temperature field-stress field coupled prediction model is introduced. Long short-term memory networks are used to learn complex historical data mapping relationships to more accurately simulate concrete behavior. Step S212, construct the knowledge base: It is a structured, queryable database containing: ① Physical laws and constraints: Fourier's law of heat conduction, the formula for heat of hydration of concrete, and core engineering specifications are encoded into executable logical rules or inequality constraints; ②Expert Case Study Library: A large number of successful historical maintenance cases are abstracted into "state patterns, action patterns" pairs; ③ Material property database: Stores thermal parameters and hydration heat curves of concrete with different mix proportions; ④ Behavior cloning: Supervised pre-training of the agent's policy network using expert cases from the knowledge base.

[0012] The formula for establishing the enhanced reward function Rt in step S22 is as follows:

[0013] In the formula: R—base reward value; —Penalty value; The formula for establishing the basic reward R is:

[0014] In the formula: The maximum temperature difference between the core and the surface; Surface temperature; Ambient temperature; To set the humidity; This represents the actual humidity. Energy consumption; α, β, γ, λ are weighting coefficients; in, and The temperatures need to be controlled below 25℃ and 20℃ respectively. The surface humidity needs to be maintained above 95% for actual measurement; the goal of the agent is to learn the strategy π that maximizes long-term cumulative rewards.

[0015] The uncertainty penalty term function is established as follows:

[0016] ω: An adaptive weight coefficient greater than 0; it is not a fixed value, but is dynamically adjusted according to the training phase, the uncertainty level of the current state, or expert feedback; : This is the "factual deviation function"; it is a scalar function used to quantify the actions selected by the agent. Relative to the current state The degree of deviation from the domain knowledge base K; its value range is designed as [0, 1], where 0 indicates complete conformity with knowledge and facts, and 1 indicates complete deviation or high uncertainty; --action; --state; K – Self-built knowledge base;

[0017] The core is the "comprehensive confidence function," which integrates information from multiple sources to evaluate actions. The rationality of the action is determined by its range [0, 1]; the higher the C value, the more credible the action and the greater the penalty. The lower the value, the better; the calculation of the C function is... The key to its effectiveness:

[0018] In the formula: , , The normalized weighting coefficients, i.e. These represent the reliability weights of different knowledge sources, and the weights of each parameter can be adjusted according to the application scenario. — Rule validation confidence —Case matching confidence —Confidence level of model prediction consistency The hard physical constraints and engineering specifications (such as ΔT1(t) ≤ 25℃, cooling rate ≤ 2℃ / d) in the source knowledge base K are calculated as a continuous satisfaction function; Source knowledge base K contains a historical expert success case study database; by searching the case study database for cases related to the current state... The N most similar historical states { }; Calculate actions Expert actions corresponding to these historical states { The average similarity of} is the average similarity. ; Based on the concrete hydration heat-temperature field prediction model embedded or associated in the knowledge base K; the actions are simulated by using the prediction model. The trajectory of state evolution within a short time window in the future; assess the "smoothness" and "reasonableness" of this trajectory.

[0019] Among them, the uncertainty penalty term function Usage and workflow: During training and deployment, The function is integrated into the enhanced reward function according to the following process. middle: Step S221, Real-time Evaluation: At each decision time step t, the AI ​​agent proposes candidate actions based on the policy network. Then, confidence level fusion is performed, and the system executes three evaluations in parallel: Call the rules engine to calculate ; Search the case library and calculate ; Run the fast prediction model and calculate ; The weights are dynamically adjusted based on the current stage. , , The weighted sum is used to obtain the overall confidence level C; Step S222, calculate the penalty: based on Calculate the deviation; at the same time, adaptively adjust the weight ω: when the system detects that it is in an unfamiliar environment or that experts have intervened frequently recently, it automatically increases ω to strengthen the penalty for uncertainty, prompting the agent to become more conservative or actively seek human confirmation. Step S223, Reward Generation and Strategy Update, will generate the calculated reward... Substitution The agent learns through the DRL algorithm that the actions that can simultaneously obtain high basic rewards R and high confidence C are the optimal choices in the long run. Step S224, expert feedback closed loop: Domain experts can see not only the AI's decisions on the monitoring interface, but also the "decision confidence" C value and its decomposition given by the system; experts provide feedback based on this, and this feedback data will be directly used to update the case library and adjust the weight coefficients, so as to realize the system's continuous online learning and human-machine alignment.

[0020] Step S23 specifically includes the following steps: Step S231, Expert Feedback Interface: Through an intuitive monitoring interface, the following is displayed to domain experts: ① Real-time sensor data and prediction curves; ② Decision-making instructions suggested by the AI ​​agent; ③ Basis for decision-making; Step S232, Feedback Mechanism and Data Processing: Real-time intervention: Experts can "approve," "question," or "reject" decisions; if rejected, experts can directly input correction instructions; the system will execute the expert instructions and update the system accordingly. (Expert actions) serve as a high-quality new sample; Post-hoc labeling: After maintenance is completed, experts review the AI's historical decision-making sequence and label the decisions that were later proven to be suboptimal or wrong. Step S233, Continue learning: Knowledge Base K Update: After cleaning and generalizing, the correction instructions and new samples provided by experts are added to the case library of Knowledge Base K as new successful cases, so that it is constantly enriched and closer to reality. Step S234, Strategy fine-tuning: Periodically use newly added expert feedback data to make incremental fine-tuning of the strategy.

[0021] In summary, due to the adoption of the above technical solution, the beneficial effects of this solution are: 1. This invention creatively introduces deep reinforcement learning AI agents into the field of engineering maintenance, constructing an intelligent maintenance system with autonomous perception, predictive judgment, collaborative decision-making, and dynamic execution capabilities. It fundamentally solves the concrete maintenance problems caused by extreme environments and uneven heat dissipation in high-altitude and cold regions, achieving a leap from "experience-driven" to "data and intelligence-driven," and from "passive response" to "proactive optimization."

[0022] 2. This method not only significantly improves the construction quality and long-term durability of concrete structures, but also effectively saves water and energy resources, which is in line with the development direction of green building and provides a reliable and advanced intelligent solution for the construction of major infrastructure in similar extreme environments.

[0023] 3. This solution can significantly reduce the rate of illusions: By using process constraints of knowledge-enhanced rewards and result correction through expert feedback, the probability of AI making serious factual errors in complex and unfamiliar situations can be reduced by more than 70%. The agent's decisions are more interpretable because its actions can often be justified in the knowledge base.

[0024] 4. This solution can improve maintenance quality and safety: When pursuing goals such as energy conservation, intelligent agents will be... The system is forced to "pull back" to a track that conforms to the laws of physics, thereby ensuring that core quality indicators (temperature difference and humidity) are always under control, and fundamentally eliminating maintenance accidents caused by AI illusions.

[0025] 5. This solution enables dynamic human-machine integration: the system does not completely replace experts, but rather constructs a highly efficient collaborative model where "AI dominates daily operations, while experts supervise and correct deviations." Experts' experience is digitally accumulated (into knowledge base K) and continuously feeds back into the evolution of AI, forming a virtuous cycle where the more it is used, the smarter it becomes.

[0026] 6. This solution possesses strong generalization ability and engineering practicality: The method framework does not rely on the absolute accuracy of a specific simulation model, but rather adapts to real-world discrepancies through an online alignment mechanism. Therefore, it can be extended to high-altitude concrete curing scenarios in different projects and under varying climatic conditions, demonstrating broad engineering application prospects. Attached Figure Description

[0027] Figure 1 This is a diagram of the architecture of the high-altitude, cold-region large-volume concrete system based on AI agents according to the present invention. Detailed Implementation

[0028] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0029] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0030] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a predetermined orientation, or be constructed and operated in a predetermined orientation. Therefore, they should not be construed as limitations on this invention.

[0031] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0032] Example 1 like Figure 1 As shown: This invention provides a technical solution: A large-volume concrete curing system based on an AI agent includes a sensing system, an AI agent decision-making system, and a collaborative execution system. The sensing system is deployed inside the concrete structure, on its surface, and in the environment. The AI ​​agent decision-making system is used to collect data from the sensing system and issue control commands according to corresponding strategies. The collaborative execution system is used to execute the control commands issued by the AI ​​agent decision-making system.

[0033] Based on the above structure, this solution introduces an AI agent with deep reinforcement learning (DRL) capabilities as the decision-making core, upgrading the maintenance system from "automation" to "intelligence," achieving a fundamental shift from passive response to proactive prediction and collaborative optimization. Specifically, the system is a closed-loop control system integrating IoT sensing, AI decision-making, and electromechanical collaborative execution. This invention creatively introduces a deep reinforcement learning AI agent into the field of engineering maintenance, constructing an intelligent maintenance system with autonomous perception, predictive analysis, collaborative decision-making, and dynamic execution capabilities. It fundamentally solves the concrete maintenance challenges caused by extreme environments and uneven heat dissipation in high-altitude, cold regions, achieving a leap from "experience-driven" to "data and intelligence-driven," and from "passive response" to "proactive optimization."

[0034] As an example, the sensing system comprehensively collects data such as temperature, humidity, wind speed, and solar radiation through a multi-source sensor network deployed inside the concrete structure, in various surface zones, and in the environment. It also uses an infrared thermal imager to scan the surface temperature field, thus constructing a real-time "digital twin" of the concrete structure.

[0035] The AI-powered decision-making system receives data from the perception layer, performs state fusion and trend prediction (e.g., using an LSTM model), and finally generates optimal control commands through a policy network trained using deep reinforcement learning. Its decision objective is to minimize overall energy consumption while satisfying all maintenance constraints (e.g., temperature difference ≤ 25℃).

[0036] As an example, the collaborative execution system may include a zoned intelligent sprinkler system, an adjustable heat preservation / heating system, and a zoned intelligent water cooling system; the sprinkler heads in the zoned intelligent sprinkler system are independently controlled by an AI agent; the sprinkler network is arranged in zones according to the BIM model; it can not only be turned on and off, but also adjust the atomization level and water volume, and regulate the surface micro-area temperature by using water evaporation while moisturizing; The adjustable insulation / heating system uses independently controlled intelligent air-fin membranes as the main insulation layer. The AI-powered system can dynamically adjust the inflation pressure (i.e., insulation thickness) of each air-fin membrane based on the real-time heat dissipation of each facade, and even open ventilation openings in certain areas to achieve dynamic insulation. In extreme low temperatures, it can be linked with built-in heaters or steam pipes for auxiliary heating.

[0037] Internal intelligent water cooling system: The cooling water pipe network is arranged in layers inside the concrete. The AI ​​intelligent body dynamically adjusts the water flow and temperature based on the core temperature prediction, and works in conjunction with the external insulation measures to achieve the best combination of "internal cooling and external insulation", accurately controlling the maximum internal temperature (usually required to be ≤70℃) and the cooling rate (preferably ≤2℃ / d).

[0038] Example 2 This invention provides a technical solution: A method for curing large-volume concrete based on AI intelligent agents includes the following steps: Step S1: Construct a perception system, an AI intelligent agent decision-making system, and a collaborative execution system. Step S2: Optimize and train the AI ​​agent decision-making system; Step S3: The optimized and trained AI agent decision-making system is used to perform intelligent curing of the concrete to be cured.

[0039] In step S2, the AI ​​agent decision-making system is optimized and trained, which includes the following steps: Step S21, establish simulation training, knowledge-based simulation pre-training—laying a factual foundation. The goal of this stage is to inject basic engineering common sense and physical laws into the intelligent agent, avoiding dangerous and inefficient random exploration from scratch. Step S22, reinforcement reward learning, deep reinforcement learning based on reinforcement rewards—constrained optimization. In this stage, by establishing the reinforcement reward function Rt, the basic reward R, and the uncertainty penalty term Ut, the agent is transformed from imitation to autonomous optimization, but is strictly constrained within the boundary of facts and knowledge through the reinforcement reward function. Step S23, online alignment of experts in the loop—achieving human-machine collaborative evolution.

[0040] The core point of step S21 is to build a high-fidelity simulation environment and knowledge base. Step S21 specifically includes the following steps; Step S211, Constructing the simulation environment: In addition to integrating the traditional thermodynamic model, a deep learning-based concrete hydration heat-temperature field-stress field coupled prediction model is introduced. For example, a long short-term memory network (LSTM) is used to learn complex historical data mapping relationships to more accurately simulate concrete behavior. Step S212, Constructing a knowledge base: This is a key innovative component of the present invention. It is a structured, queryable database containing: ① Physical laws and constraints: Fourier's law of heat conduction, the heat of hydration formula of concrete, and core engineering specifications (such as "the temperature difference between the inside and the surface ΔT1(t) should not be greater than 25℃" and "the cooling rate should not be greater than 2℃ / d") are encoded into executable logical rules or inequality constraints.

[0041] ② Expert Case Study Library: A large number of historical successful maintenance cases are abstracted into pairs of "(state mode, action mode)". For example, when "internal temperature > 60℃ and surface temperature < 35℃ and ambient wind speed > 5m / s", historical experts usually take the action combination of "increasing the cooling water flow rate by 10% and strengthening the insulation of the leeward side".

[0042] ③ Material property database: Stores thermal parameters (such as specific heat and thermal conductivity) and hydration heat curves of concrete with different mix proportions.

[0043] ④ Behavior Cloning: Supervised pre-training of the agent's policy network is performed using expert cases from the knowledge base. This is equivalent to having the AI ​​"imitate" the expert's actions in various states, quickly obtaining an initial policy with basic factual basis and without serious low-level errors.

[0044] The formula for establishing the enhanced reward function Rt in step S22 is as follows:

[0045] In the formula: R—base reward value; —Penalty value; The formula for establishing the basic reward R is:

[0046] In the formula: The maximum temperature difference between the core and the surface; Surface temperature; Ambient temperature; To set the humidity; This represents the actual humidity. Energy consumption; α, β, γ, λ are weighting coefficients.

[0047] in, and The temperatures need to be controlled below 25℃ and 20℃ respectively. The surface humidity needs to be maintained above 95% for actual measurement. The goal of the agent is to learn a policy π that maximizes long-term cumulative rewards. The uncertainty penalty term function is established as follows:

[0048] ω: An adaptive weight coefficient greater than 0. It is not a fixed value, but can be dynamically adjusted based on the training phase, the uncertainty level of the current state, or expert feedback.

[0049] : This is the "factual deviation function"; it is a scalar function used to quantify the actions selected by the agent. Relative to the current state The degree of deviation from the domain knowledge base K. Its value range is designed as [0, 1], where 0 represents complete conformity with knowledge and facts, and 1 represents complete deviation or high uncertainty.

[0050] --action; --state; K – Self-built knowledge base;

[0051] The core is the "comprehensive confidence function," which integrates information from multiple sources to evaluate actions. The rationality of the action is determined by its range [0, 1]. A higher C value indicates a more credible action and a lower penalty. The lower the value, the better. The calculation of the C function is... The key to its effectiveness. This invention proposes a multi-layered, computable fusion framework:

[0052] In the formula: , , The normalized weighting coefficients, i.e. These represent the reliability weights of different knowledge sources, and the weights of each parameter can be adjusted according to the application scenario.

[0053] — Rule validation confidence; —Case matching confidence level; —Confidence level of model prediction consistency; The source is the hard physical constraints and engineering specifications in knowledge base K (e.g., ΔT1(t) ≤ 25℃, cooling rate ≤ 2℃ / d). The calculation is a continuous satisfaction function; for example, if the action... If the predicted temperature difference at the next moment instantaneously exceeds 25°C, then... = 0; if all constraints are fully satisfied, then = 1; If in a critical state, it can be designed as a linear or sigmoid function that satisfies the degree of satisfaction. Its function is to ensure that the decision never violates the safety red line, corresponding to the resolution of "fact-conflict type illusion".

[0054] Source knowledge base K contains a historical expert success case library (state-action pairs). This is achieved by retrieving cases from the case library that match the current state. The N most similar historical states { (e.g., based on Euclidean distance or cosine similarity of state vectors). Calculate actions. Expert actions corresponding to these historical states { The average similarity of} is the average similarity. Its function is to guide intelligent agents to make decisions within the scope of experience, avoiding "out-of-the-box" guesswork.

[0055] Based on the concrete hydration heat-temperature field prediction model embedded or associated in the knowledge base K, actions are simulated using the prediction model. The trajectory of state evolution within a short future time window (e.g., the next hour). Evaluate the "smoothness" and "reasonableness" of this trajectory, for example: calculate the variance of the trajectory, whether there are non-differentiable cusps, or deviations from the ideal maintenance path. Normalize this deviation to a confidence level of [0, 1]; the smaller the deviation, the higher the confidence level. The higher the level, the better. Its function is to utilize the forward predictive capabilities of physical models to conduct rapid "thought experiments" on the long-term consequences of decisions, and to identify actions that may lead to adverse trends in advance.

[0056] Among them, the uncertainty penalty term function Usage and workflow: During training and deployment, The function is integrated into the enhanced reward function according to the following process. middle: Step S221, Real-time Evaluation: At each decision time step t, the AI ​​agent proposes candidate actions based on the policy network. Then, confidence level fusion is performed, and the system executes three evaluations in parallel: Call the rules engine to calculate ; Search the case library and calculate ; Run the fast prediction model and calculate ; The weights are dynamically adjusted based on the current stage (e.g., more trust in the model during the warming period, more trust in the case during the stable period). , , The weighted sum is used to obtain the overall confidence level C.

[0057] Step S222, calculate the penalty: based on Calculate the deviation. At the same time, adaptively adjust the weight ω: when the system detects that it is in an unfamiliar environment (such as a generally low case matching degree) or that experts have intervened frequently recently, it automatically increases ω to strengthen the penalty for uncertainty, prompting the agent to become more conservative or actively seek human confirmation.

[0058] Step S223, Reward Generation and Strategy Update, will generate the calculated reward... Substitution The agent learns through DRL algorithms (such as PPO) that those agents can simultaneously obtain high basic rewards R (good quality, low energy consumption) and have high confidence C (i.e., low... The action of [doing something] is the optimal choice in the long run.

[0059] Step S224, Expert Feedback Closed Loop: Domain experts see not only the AI's decisions on the monitoring interface, but also the "decision confidence" C value and its decomposition (e.g., "rule validation passed, but case matching degree is low"). Experts provide feedback based on this (approval, questioning, rejection), and this feedback data will be directly used to update the case library (adding new cases) and adjust the weight coefficients (λ series and ω), enabling the system to continuously learn online and achieve human-machine alignment. Step S23 specifically includes the following steps: Step S231, Expert Feedback Interface: Through an intuitive monitoring interface, the following is displayed to domain experts (such as senior maintenance engineers): ① Real-time sensor data and prediction curves; ② Decision-making instructions suggested by the AI ​​agent (e.g., "Increase the frequency of cooling pump No. 1 to 45Hz, and inflate the insulation film on the sunny side to 80%"). ③ Decision-making basis (e.g., the system prompts that this decision mainly references a similar case in the knowledge base); Step S232, Feedback Mechanism and Data Processing: Real-time intervention: Experts can "approve," "question," or "reject" decisions. If rejected, experts can directly input correction instructions. The system will execute the expert's instructions and update the system accordingly. (Expert action) as a high-quality new sample.

[0060] Post-hoc annotation: After maintenance is completed, experts can review the AI's historical decision-making sequence and label the decisions that were later proven to be suboptimal or wrong. Step S233, Continue learning: Knowledge Base (K) Update: After cleaning and generalization, the correction instructions and new samples provided by experts are added to the case library of Knowledge Base K as new successful cases, so as to continuously enrich it and make it closer to reality; Step S234, Strategy Fine-tuning: Periodically (e.g., weekly) use newly added expert feedback data to incrementally fine-tune the strategy. This is equivalent to allowing the AI ​​to continue "learning" from experts in the real world, gradually correcting biases or erroneous associations formed in the simulation environment, and achieving continuous optimization of the strategy and human-machine alignment.

[0061] Example 2 This solution provides a more specific implementation plan, addressing the challenges of traditional methods: manual maintenance struggles to monitor the temperature difference of over 15°C between the sunny and shady sides of the dam in real time; and the insulation blanket application is often indiscriminate, resulting in insufficient heat dissipation on the sunny side and inadequate insulation on the shady side, easily leading to cracks. This solution addresses these challenges: 1. Initial Heating Phase (Days 1-3): The sensing system indicates that the internal temperature is rising too rapidly. The AI ​​prediction module determines that afternoon sunlight will exacerbate the temperature rise on the sun-facing side. Intelligent Agent Decisions: Instruction ① Appropriately increase the cooling water flow rate on the shaded side; Instruction ② Instruct the intelligent air-fin membrane on the sun-facing side to open 30% of its ventilation area at midday to assist in heat dissipation, while simultaneously activating low-frequency atomizing spray on the sun-facing side; Instruction ③ Predicting strong winds and temperature drops at night, instruct all insulation membranes to be completely sealed and pressurized before sunset.

[0062] 2. Mid-term cooling phase: The system continuously monitors to ensure that the temperature difference between inside and outside is ≤20℃ and the cooling rate is stable. When the infrared thermal image of a certain area shows that the surface humidity is below the 95% threshold, the intelligent agent immediately dispatches the nearest spray unit for precise humidification.

[0063] 3. Benefits Summary: Through the aforementioned predictive and collaborative curing methods, the maximum internal and external temperature difference was successfully controlled within 18℃ throughout the curing period, the surface humidity of the concrete remained above 95%, and no harmful cracks were generated. Furthermore, compared to traditional constant temperature and flow cooling methods, overall energy savings were approximately 25%.

[0064] This invention creatively introduces deep reinforcement learning AI agents into the field of engineering maintenance, constructing an intelligent maintenance system with autonomous perception, predictive judgment, collaborative decision-making, and dynamic execution capabilities. It fundamentally solves the concrete maintenance challenges caused by extreme environments and uneven heat dissipation in high-altitude, cold regions, achieving a leap from "experience-driven" to "data and intelligence-driven," and from "passive response" to "proactive optimization."

[0065] This method not only significantly improves the construction quality and long-term durability of concrete structures, but also effectively saves water and energy resources, which is in line with the development direction of green building and provides a reliable and advanced intelligent solution for the construction of major infrastructure in similar extreme environments.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-volume concrete curing system based on AI intelligent agents, characterized in that, It includes a perception system, an AI agent decision-making system, and a collaborative execution system. The perception system is deployed inside the concrete structure, on its surface, and in the environment. The AI ​​agent decision-making system is used to collect data from the perception system and issue control commands according to corresponding strategies. The collaborative execution system is used to execute the control commands issued by the AI ​​agent decision-making system.

2. The large-volume concrete curing system based on AI intelligent agents according to claim 1, characterized in that: The collaborative execution system includes a zoned intelligent sprinkler system, an adjustable insulation / heating system, and a central intelligent water cooling system; the sprinkler heads in the zoned intelligent sprinkler system are independently controlled by the AI ​​intelligent agent decision-making system; the sprinkler network is arranged in zones according to the BIM model.

3. The large-volume concrete curing system based on AI intelligent agents according to claim 2, characterized in that: The adjustable insulation / heating system uses independently controlled intelligent air-fin membranes as the insulation layer; the AI ​​intelligent decision-making system dynamically adjusts the inflation pressure of each air-fin membrane and the opening of local ventilation openings according to the real-time heat dissipation of each facade; the internal intelligent water cooling system: the cooling water pipeline is arranged in layers inside the concrete.

4. A method for curing large-volume concrete based on an AI agent, employing the system described in any one of claims 1 to 3, characterized in that: Includes the following steps: Step S1: Construct a perception system, an AI intelligent agent decision-making system, and a collaborative execution system; Step S2: Optimize and train the AI ​​agent decision-making system; Step S3: The optimized and trained AI agent decision-making system is used to perform intelligent curing of the concrete to be cured.

5. A method for curing large-volume concrete based on an AI agent according to claim 4, characterized in that: In step S2, the AI ​​agent decision-making system is optimized and trained, which includes the following steps: Step S21: Establish simulation training, knowledge-based simulation pre-training—laying a factual foundation; Step S22, reinforcement reward learning, deep reinforcement learning based on reinforcement rewards—constrained optimization, by establishing reinforcement reward function Rt, basic reward R, uncertainty penalty term Ut, the agent shifts from imitation to autonomous optimization, but is strictly constrained within the boundary of facts and knowledge through reinforcement reward function; Step S23, online alignment of experts in the loop—achieving human-machine collaborative evolution.

6. A method for curing large-volume concrete based on an AI agent according to claim 5, characterized in that: Step S21 specifically includes the following steps: Step S211, Constructing the simulation environment: In addition to integrating the traditional thermodynamic model, a deep learning-based concrete hydration heat-temperature field-stress field coupled prediction model is introduced. Long short-term memory networks are used to learn complex historical data mapping relationships to more accurately simulate concrete behavior. Step S212, construct the knowledge base: It is a structured, queryable database containing: ① Physical laws and constraints: Fourier's law of heat conduction, the formula for heat of hydration of concrete, and core engineering specifications are encoded into executable logical rules or inequality constraints; ②Expert Case Study Library: A large number of successful historical maintenance cases are abstracted into "state patterns, action patterns" pairs; ③ Material property database: Stores thermal parameters and hydration heat curves of concrete with different mix proportions; ④ Behavior cloning: Supervised pre-training of the agent's policy network using expert cases from the knowledge base.

7. The method for curing large-volume concrete based on AI intelligent agents according to claim 6, characterized in that: The formula for establishing the enhanced reward function Rt in step S22 is as follows: In the formula: R—base reward value; —Penalty value; The formula for establishing the basic reward R is: In the formula: The maximum temperature difference between the core and the surface; Surface temperature; Ambient temperature; To set the humidity; This represents the actual humidity. Energy consumption; α, β, γ, λ are weighting coefficients; in, and The temperatures need to be controlled below 25℃ and 20℃ respectively. The surface humidity needs to be maintained above 95% for actual measurement; the goal of the agent is to learn the strategy π that maximizes long-term cumulative rewards.

8. A method for curing large-volume concrete based on an AI agent according to claim 7, characterized in that: The uncertainty penalty term function is established as follows: ω: An adaptive weight coefficient greater than 0; it is not a fixed value, but is dynamically adjusted according to the training phase, the uncertainty level of the current state, or expert feedback; : This is the "factual deviation function"; it is a scalar function used to quantify the actions selected by the agent. Relative to the current state The degree of deviation from the domain knowledge base K; Its value range is designed as [0, 1], where 0 represents complete conformity with knowledge and facts, and 1 represents complete deviation or high uncertainty; --action; --state; K – Self-built knowledge base; The core is the "comprehensive confidence function," which integrates information from multiple sources to evaluate actions. The rationality of this is that its range is [0, 1]; A higher C value indicates a more credible action and a higher penalty. The lower the value, the better; the calculation of the C function is... The key to its effectiveness: In the formula: , , The normalized weighting coefficients, i.e. These represent the reliability weights of different knowledge sources, and the weights of each parameter can be adjusted according to the application scenario. — Rule validation confidence —Case matching confidence —Confidence level of model prediction consistency The hard physical constraints and engineering specifications (such as ΔT1(t) ≤ 25℃, cooling rate ≤ 2℃ / d) in the source knowledge base K are calculated as a continuous satisfaction function; Source knowledge base K contains a historical expert success case study database; by searching the case study database for cases related to the current state... The N most similar historical states { }; Calculate actions Expert actions corresponding to these historical states { The average similarity of} is the average similarity. ; Based on the concrete hydration heat-temperature field prediction model embedded or associated in the knowledge base K; the actions are simulated by using the prediction model. The trajectory of state evolution within a short time window in the future; assess the "smoothness" and "reasonableness" of this trajectory.

9. A method for curing large-volume concrete based on an AI agent according to claim 8, characterized in that: Among them, the uncertainty penalty term function Usage and workflow: During training and deployment, The function is integrated into the enhanced reward function according to the following process. middle: Step S221, Real-time Evaluation: At each decision time step t, the AI ​​agent proposes candidate actions based on the policy network. Then, confidence level fusion is performed, and the system executes three evaluations in parallel: Call the rules engine to calculate ; Search the case library and calculate ; Run the fast prediction model and calculate ; The weights are dynamically adjusted based on the current stage. , , The weighted sum is used to obtain the overall confidence level C; Step S222, calculate the penalty: based on Calculate the deviation; at the same time, adaptively adjust the weight ω: when the system detects that it is in an unfamiliar environment or that experts have intervened frequently recently, it automatically increases ω to strengthen the penalty for uncertainty, prompting the agent to become more conservative or actively seek human confirmation. Step S223, Reward Generation and Strategy Update, will generate the calculated reward... Substitution The agent learns through the DRL algorithm that the actions that can simultaneously obtain high basic rewards R and high confidence C are the optimal choices in the long run. Step S224, expert feedback closed loop: Domain experts can see not only the AI's decisions on the monitoring interface, but also the "decision confidence" C value and its decomposition given by the system; experts provide feedback based on this, and this feedback data will be directly used to update the case library and adjust the weight coefficients, so as to realize the system's continuous online learning and human-machine alignment.

10. A method for curing large-volume concrete based on an AI agent according to claim 9, characterized in that: Step S23 specifically includes the following steps: Step S231, Expert Feedback Interface: Through an intuitive monitoring interface, the following is displayed to domain experts: ① Real-time sensor data and prediction curves; ② Decision-making instructions suggested by the AI ​​agent; ③ Basis for decision-making; Step S232, Feedback Mechanism and Data Processing: Real-time intervention: Experts can "approve," "question," or "reject" decisions; if rejected, experts can directly input correction instructions; the system will execute the expert instructions and record the changes. (Expert actions) serve as a high-quality new sample; Post-hoc labeling: After maintenance is completed, experts review the AI's historical decision-making sequence and label the decisions that were later proven to be suboptimal or wrong. Step S233, Continue learning: Knowledge Base K Update: After cleaning and generalizing, the correction instructions and new samples provided by experts are added to the case library of Knowledge Base K as new successful cases, so that it is constantly enriched and closer to reality. Step S234, Strategy fine-tuning: Periodically use newly added expert feedback data to make incremental fine-tuning of the strategy.

Images