An intelligent management and control method and system for a constant-pressure and reduced-pressure device based on an AI agent

Abstract

The application discloses an AI intelligent agent-based intelligent management and control method and system for a crude oil unit, relates to the field of petroleum chemical industry, and comprises the following steps: constructing a high-fidelity digital twin virtual model by using a process flow diagram, a pipeline instrument flow diagram and historical operation data of the crude oil unit, defining and initializing micro intelligent agents according to a process control loop division principle; performing offline parallel training on the micro intelligent agents in the high-fidelity digital twin virtual model, collecting time series data, using a causal discovery algorithm, constructing a causal correlation graph of the crude oil unit, and calculating the dynamic correlation strength between the behavior data of the micro intelligent agents. Through the decision guidance of the intelligent agent grouping based on the dynamic correlation strength and the causal contribution calculation, the application realizes the deep coupling of the global optimization target and the local control, and overcomes the problem of insufficient long-term self-adaptive ability caused by the lack of causal reasoning in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of petrochemicals, and in particular to an intelligent control method and system for atmospheric and vacuum distillation units based on AI agents. Background Technology

[0002] In advanced process control of atmospheric and vacuum distillation units in the petrochemical industry, the application of multi-agent system frameworks has become an important technical path for achieving distributed optimal control. Existing technologies typically employ rule-based or model predictive control architectures, constructing multiple functional agents corresponding to subsystems such as heaters and distillation columns. Distributed decision-making architectures are used to achieve local optimization, and communication protocols between agents enable coordinated control. By establishing a hybrid model combining data-driven and mechanistic models, the adaptability of the unit to different crude oil ratios is improved to some extent. Existing technologies also employ reinforcement learning algorithms to learn agent strategies online, enabling the system to adjust control parameters according to real-time operating conditions, thus improving the robustness of overall control. These methods reflect the current level of technological development and provide feasible solutions for the intelligent management and control of atmospheric and vacuum distillation units.

[0003] Despite significant advancements in existing technologies, major challenges remain in practical industrial applications. Most current methods rely on direct communication between agents or pre-defined coordination rules, failing to fully consider the dynamic coupling characteristics between the inherent process causality and agent behavior. Due to the complex characteristics of atmospheric and vacuum distillation units, such as strong coupling, nonlinearity, and large time delays, coordination strategies based solely on data correlation often struggle to adapt to the drift in operating conditions caused by time-varying factors such as changes in crude oil properties and equipment fouling. This can lead to a gradual degradation of control performance during long-term operation, requiring frequent manual intervention and parameter tuning. In addressing these challenges, existing methods lack the ability to jointly model the collective behavior of agents and process causality, making it difficult to achieve truly adaptive collaborative optimization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent control method for atmospheric and vacuum distillation equipment based on AI agents, which solves the problem of insufficient long-term adaptive capability caused by the lack of joint modeling of the dynamic coupling between process causality and agent behavior in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent control method for atmospheric and vacuum distillation equipment based on AI intelligent agents, which includes constructing a high-fidelity digital twin virtual model through the process flow diagram, piping and instrumentation diagram and historical operating data of the atmospheric and vacuum distillation equipment, and defining and initializing micro-intelligent agents according to the process control loop division principle;

[0008] Offline parallel training of micro-agents is performed in a high-fidelity digital twin virtual model, and time-series data is collected. Causal discovery algorithm is used to construct a causal correlation map of the atmospheric and vacuum diversion device and calculate the dynamic correlation strength between micro-agent behavioral data.

[0009] Micro-agents are dynamically grouped according to the dynamic correlation strength to form agent groups. The causal correlation graph, agent groups, and trained micro-agents are then deployed to the control unit of the physical atmospheric and vacuum distillation device.

[0010] The global optimization objective is received through the collaborative optimization layer in the control unit. The causal correlation graph and dynamic correlation strength are fused together to calculate the causal contribution of the micro-agent's actions to the global optimization objective and generate decision guidance signals.

[0011] The micro-agent integrates decision-making guidance signals with its own control objectives to generate control commands, which are then sent to the actuators of the physical atmospheric and vacuum distillation unit. Small sample data generated after the physical atmospheric and vacuum distillation unit is in operation are collected, and the agent grouping is updated with a collaborative evolution strategy based on the small sample data.

[0012] As a preferred embodiment of the AI-based intelligent control method for atmospheric and vacuum distillation units according to the present invention, the method includes: constructing a high-fidelity digital twin virtual model using the process flow diagram, piping and instrumentation diagram, and historical operating data of the atmospheric and vacuum distillation unit; defining and initializing micro-intelligent agents according to the process control loop partitioning principle; and including the following steps:

[0013] The topology, equipment connections, and fundamental differential-algebraic equations of a high-fidelity digital twin virtual model are defined using process flow diagrams and piping and instrumentation diagrams; a long short-term memory network compensator is trained using historical operating data; and a high-fidelity digital twin virtual model is constructed using the hybrid modeling method of the long short-term memory network compensator.

[0014] The physical laws derived from the process flow diagram and piping and instrumentation diagram are transformed into mathematical constraints. The square of the degree of constraint violation is multiplied by the penalty coefficient and added to the overall loss function to construct the loss function of the physical constraint penalty term.

[0015] The high-fidelity digital twin virtual model is trained and optimized using a loss function with a physical constraint penalty term.

[0016] A global sensitivity analysis is performed on a high-fidelity digital twin virtual model to quantify the influence of manipulated variables on controlled variables. A global sensitivity matrix is ​​generated based on the process control loop partitioning principle. A spectral clustering algorithm is used to partition the strongly coupled manipulated and controlled variables into the same cluster. The cluster boundary defines the micro-agent. The micro-agent is initialized using a proximal policy optimization algorithm framework to obtain the policy network of the micro-agent.

[0017] As a preferred embodiment of the intelligent control method for an atmospheric and vacuum distillation unit based on AI agents described in this invention, the method includes: offline parallel training of micro-agents in a high-fidelity digital twin virtual model, collecting time-series data and using a causal discovery algorithm to construct a causal correlation graph of the atmospheric and vacuum distillation unit, and calculating the dynamic correlation strength between the behavioral data of the micro-agents, comprising the following steps:

[0018] In a high-fidelity digital twin virtual model, a multi-agent proximal policy optimization algorithm is used to set up an adversarial course from easy to difficult.

[0019] A performance threshold is set based on a specific percentage of the maximum reward value obtained by a micro-agent in a high-fidelity digital twin virtual model during the adversarial course phase.

[0020] After the micro-agent reaches the performance threshold in each course stage, it is then moved to a more difficult course to complete the offline parallel training of the micro-agent.

[0021] During offline parallel training, time-series data of micro-agent action sequences and high-fidelity digital twin virtual model states are collected. Using the time-series data and injecting process knowledge constraints, a causal discovery algorithm is executed to construct a causal relationship map of the atmospheric and vacuum distillation unit.

[0022] Based on the causal correlation graph of the atmospheric and vacuum distillation unit, the linear non-Gaussian model method is used to quantify the causal effect strength of each edge in the causal correlation graph of the atmospheric and vacuum distillation unit.

[0023] Based on the action sequences of micro-agents, the minimum warped path distance between any two action sequences of micro-agents is calculated using the dynamic time warping algorithm, thereby obtaining the dynamic correlation strength between micro-agent behavior data.

[0024] As a preferred embodiment of the intelligent control method for an atmospheric and vacuum distillation device based on AI agents according to the present invention, the method includes the following steps: dynamically grouping micro-agents according to dynamic correlation strength to form agent groups, and deploying the causal correlation graph, agent groups, and trained micro-agents to the control unit of the physical atmospheric and vacuum distillation device:

[0025] Based on the dynamic correlation strength, the spectral clustering algorithm is used to dynamically group the micro-agents to obtain the preliminary agent grouping structure.

[0026] The preliminary agent grouping structure and the causal relationship map of the atmospheric and vacuum distillation device are input into the Louvain community discovery algorithm. The agent grouping is formed by iterative calculation and optimization by maximizing the modularity Q value as the objective function.

[0027] The intelligent agents are grouped, the trained micro-agent policy network parameters are used, and the causal relationship graph of the atmospheric and vacuum distillation unit is deployed to the control unit of the physical atmospheric and vacuum distillation unit through a secure communication protocol.

[0028] As a preferred embodiment of the intelligent control method for atmospheric and vacuum distillation devices based on AI agents described in this invention, the method includes the following steps: receiving a global optimization objective through a collaborative optimization layer in the control unit of the atmospheric and vacuum distillation device, fusing a causal correlation graph and dynamic correlation strength, calculating the causal contribution of the micro-agent's actions to the global optimization objective, and generating a decision guidance signal.

[0029] The control unit of the atmospheric and vacuum distillation device consists of a lower-level executive intelligent agent layer, a middle-level collaborative optimization layer, and an upper-level cognitive decision-making layer.

[0030] The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit receives the global optimization target from the upper layer. The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit integrates the dynamic correlation strength between the causal correlation graph of the atmospheric and vacuum distillation unit and the behavioral data of micro-intelligent agents to construct a joint influence network.

[0031] Based on the joint influence network, counterfactual reasoning is used to calculate the causal contribution of each micro-agent's candidate actions to the global optimization objective;

[0032] The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit adjusts the causal contribution to a range that matches the dimensions of the micro-agent's own reward function through linear scaling, and transforms the scaled value into a decision guidance signal in the form of additional reward items.

[0033] As a preferred embodiment of the intelligent control method for atmospheric and vacuum distillation devices based on AI agents described in this invention, the micro-agent integrates decision guidance signals with its own control objectives to generate control commands, and sends the control commands to the actuators of the physical atmospheric and vacuum distillation devices, including the following steps:

[0034] The reward function for the micro-agent's own control objective is obtained by weighted summing of the squared term of the tracking error of the setpoint of the micro-agent's control objective and the squared term of the change in the manipulated variable.

[0035] The local reward value is derived based on the reward function of the micro-agent's own control objective;

[0036] The total reward function of the micro-agent is obtained by adding the fused decision guidance signal to the local reward value.

[0037] The policy network of the micro-agent generates the original action probability distribution based on the current observation state, selects an action index through the Boltzmann exploration policy, and converts it into the original control command value through the linear mapping function;

[0038] The original control command value is input to a first-order hysteresis filter for smoothing, and the final execution command is generated and sent to the actuator of the physical atmospheric and vacuum distillation unit through the industrial communication protocol.

[0039] As a preferred embodiment of the intelligent control method for an atmospheric and vacuum distillation unit based on AI agents according to the present invention, the method includes: collecting small sample data generated after the physical atmospheric and vacuum distillation unit has been running, and updating the collaborative evolution strategy of the agents based on the small sample data, comprising the following steps:

[0040] A mini-batch of data is extracted proportionally from the small sample data generated after the operation of the physical atmospheric and vacuum distillation unit.

[0041] The local fitness of each micro-agent is calculated using the extracted mini-batch data. Within each agent group, an elite group of micro-agents is selected based on the local fitness value, and simulated binary crossover and polynomial mutation operations are performed to generate offspring policies.

[0042] The knowledge distillation loss function is used to mimic the output distribution of the elite group's policy in the offspring policy, resulting in an optimized offspring policy.

[0043] Replace the micro-agent policy with the lowest local fitness in the agent group with the optimized offspring policy.

[0044] Secondly, the present invention provides an intelligent control system for atmospheric and vacuum distillation equipment based on AI agents, including a model building module, which constructs a high-fidelity digital twin virtual model through the process flow diagram, piping and instrumentation flow diagram and historical operating data of the atmospheric and vacuum distillation equipment, and defines and initializes micro agents according to the process control loop division principle;

[0045] The offline parallel training module performs offline parallel training on micro-agents in a high-fidelity digital twin virtual model, collects time-series data, uses a causal discovery algorithm to construct a causal correlation graph of the atmospheric and vacuum distillation device, and calculates the dynamic correlation strength between the behavioral data of micro-agents.

[0046] The grouping module dynamically groups micro-agents according to the dynamic correlation strength to form agent groups, and then deploys the causal correlation graph, agent groups, and trained micro-agents to the control unit of the physical atmospheric and vacuum distillation device.

[0047] The decision guidance module receives the global optimization objective through the collaborative optimization layer in the control unit, integrates the causal correlation graph and the dynamic correlation strength, calculates the causal contribution of the micro-agent's actions to the global optimization objective, and generates a decision guidance signal.

[0048] The adaptive evolution module integrates decision guidance signals with its own control objectives to generate control commands, which are then sent to the actuators of the physical atmospheric and vacuum distillation unit. It collects small sample data generated after the physical atmospheric and vacuum distillation unit is in operation and updates the collaborative evolution strategy of the agents based on the small sample data.

[0049] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent control method for a normal and low pressure device based on an AI agent as described in the first aspect of the present invention.

[0050] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent control method for a normal and low pressure ventilation device based on an AI agent as described in the first aspect of the present invention.

[0051] The beneficial effects of this invention are as follows: By constructing a high-fidelity digital twin virtual model and initializing micro-agents, accurate mapping of the physicochemical processes of atmospheric and vacuum distillation units and the establishment of a distributed control architecture are achieved; furthermore, by conducting offline parallel training and causal discovery in a virtual environment, combined with a dynamic time warping algorithm, a dynamic knowledge system integrating process causal relationships and agent behavior correlations is constructed, solving the problem of separation between mechanism models and data-driven approaches in multi-agent collaborative control; through decision guidance based on agent grouping based on dynamic correlation strength and causal contribution calculation, deep coupling between global optimization objectives and local control is achieved, overcoming the problem of insufficient long-term adaptive capability caused by the lack of causal reasoning in traditional methods. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of an intelligent control method for atmospheric and vacuum distillation equipment based on AI agents.

[0054] Figure 2 This is a schematic diagram of an intelligent control system for atmospheric and vacuum distillation equipment based on AI agents. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figures 1-2 As one embodiment of the present invention, this embodiment provides an intelligent control method for a normal and low pressure ventilation device based on an AI agent, comprising the following steps:

[0059] S1. Construct a high-fidelity digital twin virtual model using the process flow diagram, piping and instrumentation diagram, and historical operating data of the atmospheric and vacuum distillation unit. Define and initialize micro-intelligent agents according to the process control loop division principle.

[0060] S1.1 Define the topology, equipment connections, and fundamental differential-algebraic equations of a high-fidelity digital twin virtual model using process flow diagrams and piping / instrumentation diagrams; train a long short-term memory network compensator using historical operating data; and construct a high-fidelity digital twin virtual model using a hybrid modeling method that integrates the long short-term memory network compensator.

[0061] Furthermore, in constructing a high-fidelity digital twin virtual model, the process flow diagram is used to determine the spatial layout and material flow logic of the core equipment such as the heater, distillation column, and heat exchanger in the atmospheric and vacuum distillation unit. The piping and instrumentation diagram clarifies the connection methods of the pipelines between the equipment, the valve positions, and the distribution of instrument measurement points. Based on these topological and connection relationships, material balance equations are established according to the law of conservation of mass, heat balance equations according to the law of conservation of energy, and pressure balance equations according to the law of conservation of momentum. Together, these form the basic differential-algebraic equations of the high-fidelity digital twin virtual model. At the same time, a long short-term memory network compensator is trained using historical operating data. The historical operating data includes time-series data of temperature, pressure, and flow rate collected by the DCS. The input of the long short-term memory network compensator is the process parameter sequence in the historical operating data, and the output is the dynamic compensation value of the time-varying parameters in the basic differential-algebraic equations. Finally, by weightedly fusing the output of the mechanism model with the output of the long short-term memory network compensator, a high-fidelity digital twin virtual model is formed.

[0062] S1.2 The physical laws derived from the process flow diagram and piping and instrumentation diagram are transformed into mathematical constraints. The square of the degree of constraint violation is multiplied by the penalty coefficient and added to the overall loss function to construct the loss function of the physical constraint penalty term.

[0063] Furthermore, when the physical laws derived from the process flow diagram and piping instrumentation diagram are transformed into mathematical constraints, the material balance law is transformed into an equality constraint that the total feed flow rate equals the sum of the output flow rates of each side line; the energy balance law is transformed into an equality constraint that the heating furnace supply equals the sum of the heat absorbed by the material flow and the heat loss; and the equipment safe operation law is transformed into an inequality constraint that the temperature gradient of the tray decreases from top to bottom. The loss function of the physical constraint penalty term is constructed by multiplying the square value of the violation degree of the mathematical constraint by the penalty coefficient and then adding it to the overall loss function. The overall loss function also includes the mean square error term between the output value of the high-fidelity digital twin virtual model and the historical operating data. The overall loss function is optimized by the gradient descent algorithm.

[0064] S1.3. The high-fidelity digital twin virtual model is trained and optimized using a loss function with physical constraint penalty terms.

[0065] Furthermore, when training and optimizing the high-fidelity digital twin virtual model using a loss function with a physical constraint penalty term, a composite loss function is constructed that includes a data fitting term and a physical constraint term. The data fitting term calculates the mean square error between the output value of the high-fidelity digital twin virtual model and the true value in the historical operating data. The physical constraint term forces the output of the high-fidelity digital twin virtual model to meet the basic physical laws derived from the process principles, such as the temperature of each tray in the tower equipment must be maintained to increase gradually from top to bottom, and the material flow rate must meet the constraint that the cumulative inlet amount equals the cumulative outlet amount. The physical constraint term is added to the loss function in the form of a penalty term. When the output of the high-fidelity digital twin virtual model violates the physical laws, the loss value is increased. The gradient descent algorithm is used to simultaneously optimize the parameters of the mechanistic model and the parameters of the long short-term memory network compensator, so that the high-fidelity digital twin virtual model can accurately fit the historical operating data while strictly following the physical laws.

[0066] S1.4 Perform global sensitivity analysis on the high-fidelity digital twin virtual model to quantify the influence of the manipulated variable on the controlled variable. Generate a global sensitivity matrix according to the process control loop partitioning principle. Use spectral clustering algorithm to partition the strongly coupled manipulated variable and controlled variable into the same cluster. The cluster boundary defines the micro-agent. Use the near-end policy optimization algorithm framework to initialize the micro-agent and obtain the policy network of the micro-agent.

[0067] Furthermore, when performing global sensitivity analysis on a high-fidelity digital twin virtual model, the Sobol index method is used to calculate the first-order sensitivity index and the overall sensitivity index of each manipulated variable on all controlled variables, quantifying the intensity of the main effect and interaction effect of the manipulated variables on the controlled variables. Based on the process control loop partitioning principle, the calculated sensitivity indices between all manipulated and controlled variables are organized into a global sensitivity matrix. The rows of this matrix correspond to the manipulated variables, the columns to the controlled variables, and the element values ​​are the sensitivity indices. Using the global sensitivity matrix as feature input, a spectral clustering algorithm is applied to jointly cluster the manipulated and controlled variables. The spectral clustering algorithm first calculates... The Laplacian matrix of the global sensitivity matrix is ​​obtained and its eigenvalues ​​are decomposed. Then, k-means clustering is performed on the eigenvectors to group the strongly coupled manipulated and controlled variables into the same cluster. The boundary of each cluster clearly defines the control scope of a micro-agent, which includes all controlled variables within the cluster as the micro-agent's control target and all manipulated variables within the cluster as the micro-agent's action space. A proximal policy optimization algorithm framework is used to initialize each micro-agent. The policy network of the proximal policy optimization algorithm framework includes an observation state encoding layer and an action probability output layer. The policy network parameters are initialized using the Xavier method to obtain the policy network of the micro-agent.

[0068] It should be noted that global sensitivity analysis of high-fidelity digital twin virtual models can accurately quantify the nonlinear influence relationship between manipulated and controlled variables, avoiding the subjectivity of traditional empirical division; the global sensitivity matrix generated based on the process control loop division principle objectively reflects the coupling strength between variables, providing a data-driven basis for agent grouping; the spectral clustering algorithm is used to group strongly coupled manipulated and controlled variables into the same cluster, ensuring high cohesion of variables within the control scope of each micro-agent, significantly reducing the coordination complexity across agents; the cluster boundaries clearly define the scope of responsibility of micro-agents, avoiding control objective conflicts; and the micro-agent policy network is initialized using a proximal policy optimization algorithm framework.

[0069] S2. Perform offline parallel training on micro-agents in a high-fidelity digital twin virtual model, collect time-series data, use causal discovery algorithms to construct and update the causal correlation map of the atmospheric and vacuum distillation device, and calculate the dynamic correlation strength between the behavioral data of micro-agents.

[0070] S2.1 In a high-fidelity digital twin virtual model, a multi-agent proximal policy optimization algorithm is adopted to set up an adversarial course from easy to difficult.

[0071] Furthermore, when training micro-agents using the multi-agent proximal policy optimization algorithm in a high-fidelity digital twin virtual model, an adversarial course is first set up, progressing from easy to difficult. This course increases in difficulty by gradually increasing the perturbation range of the input parameters of the high-fidelity digital twin virtual model. For example, in the initial stage, the perturbation range of crude oil density is ±2% of the baseline value, expanding to ±10% in the final stage. Each course stage has a corresponding performance threshold, set based on a specific percentage of the theoretical maximum reward that the micro-agent group can obtain in the high-fidelity digital twin virtual model at the current course stage. The micro-agent group is trained using the multi-agent proximal policy optimization algorithm in each course stage. This algorithm employs a centralized training and distributed execution architecture. The critics network receives observation information from all micro-agents and calculates the advantage function, while the actors network updates the policy parameters based on the advantage function. When the average reward of the micro-agent group in the current course stage continuously exceeds the performance threshold, it automatically enters the next more difficult course stage for training, until the training objectives of all course stages are completed.

[0072] S2.2. Set a performance threshold based on a specific percentage of the maximum reward value obtained by the micro-agent in a high-fidelity digital twin virtual model during the adversarial course phase.

[0073] Furthermore, when setting the performance threshold based on a specific percentage of the maximum reward value obtained by micro-agents in a high-fidelity digital twin virtual model during the adversarial course phase, the ideal maximum reward value that the micro-agent group may obtain in each course phase is first determined through theoretical derivation and simulation testing. The ideal maximum reward value represents the upper limit of reward that can be achieved by a perfectly controlled strategy under the current course difficulty parameters. The performance threshold is set as a fixed percentage of this ideal maximum reward value, for example, taking 90% of the ideal maximum reward value as the advancement threshold. The percentage value remains constant and is applied to all course phases to ensure the uniformity of the course advancement standards. The specific value of the performance threshold is calculated before the course begins and embedded in the training process, serving as the sole quantitative standard for judging whether the micro-agent group has reached the current course training objective.

[0074] S2.3 After the micro-agent reaches the performance threshold in each course stage, it advances to a more difficult course to complete the offline parallel training of the micro-agent.

[0075] Furthermore, the micro-agents are trained using a multi-agent proximal policy optimization algorithm at each course stage. When the average reward value of the micro-agent group in the current course stage continuously exceeds the performance threshold set for that course stage, the training process automatically adjusts the perturbation range of the input parameters of the high-fidelity digital twin virtual model to a higher difficulty level preset for the next course stage. The micro-agent group then immediately continues training in the new difficulty course stage, and completes the entire offline parallel training process after going through all preset course stages in sequence. The advancement judgment of each course stage is strictly based on the comparison result between the performance threshold and the measured average reward value, ensuring that the micro-agent group has fully mastered the control strategy of the current difficulty before entering a higher difficulty course.

[0076] S2.4 During offline parallel training, time-series data of micro-agent action sequences and high-fidelity digital twin virtual model states are collected. The time-series data is used and process knowledge constraints are injected to execute the causal discovery algorithm to construct the causal relationship map of the atmospheric and vacuum distillation unit.

[0077] Furthermore, during offline parallel training, time-series data of micro-agent action sequences and high-fidelity digital twin virtual model states are continuously collected. This time-series data includes the action values ​​executed by the micro-agent at each time step and the corresponding state response values ​​of the high-fidelity digital twin virtual model. When using the time-series data to execute the causal discovery algorithm, process knowledge constraints extracted from the process flow diagram and piping instrumentation diagram are first injected. These constraints clearly define the causal directionality between variables, stipulating that the material flow direction can only be from upstream equipment to downstream equipment, and that changes in tray temperature can only affect the lower tray from the upper tray. Guided by the process knowledge constraints, the PC algorithm is used to perform conditional independence checks on the time-series data, gradually eliminating edges that do not meet the constraints, and constructing a causal relationship graph of the atmospheric and vacuum distillation unit.

[0078] It should be noted that during offline parallel training, the time-series data of the collected micro-agent action sequences and the state of the high-fidelity digital twin virtual model are continuously expanded into the historical operation database. When the accumulated data reaches the level that detects operating condition drift, the causal discovery algorithm is re-executed, merging the newly added time-series data with the original historical operation data as input, and injecting process knowledge constraints extracted from the process flow diagram and piping instrumentation diagram again. Conditional independence testing and causal orientation are performed through the PC algorithm to generate an updated causal relationship map of the atmospheric and vacuum distillation unit. The updated causal relationship map of the atmospheric and vacuum distillation unit will replace the old version.

[0079] S2.5. Based on the causal correlation graph of the atmospheric and vacuum distillation unit, the causal effect intensity of each edge in the causal correlation graph of the atmospheric and vacuum distillation unit is quantified by using a linear non-Gaussian model method.

[0080] Furthermore, based on the causal correlation graph of the atmospheric and vacuum distillation unit, when quantifying the causal effect strength of each edge in the causal correlation graph using the linear non-Gaussian model method, each directed edge in the causal correlation graph is first transformed into a regression problem in a structural equation model, with the parent node variable as the independent variable and the child node variable as the dependent variable. The coefficient matrix in the linear non-Gaussian model is estimated using the independent component analysis algorithm, and the exact causal direction is identified and the causal effect strength is calculated using the statistical characteristics of non-Gaussian noise. The causal effect strength value of each edge is the regression coefficient of the corresponding independent variable in the linear non-Gaussian model regression equation, representing the expected value of the change in the child node variable caused by a unit change in the parent node variable, thus obtaining the causal correlation graph of the atmospheric and vacuum distillation unit with weighted values.

[0081] S2.6. Based on the action sequences of micro-agents, the minimum warped path distance between any two action sequences of micro-agents is calculated using the dynamic time warping algorithm to obtain the dynamic correlation strength between micro-agent behavior data.

[0082] The expression for the minimum regularized path distance is:

[0083] ;

[0084] in, For the first sequence at the first micro-agent time The action value and the second sequence at the second micro-agent time The Euclidean distance between action values, For the first micro-intelligent agent Time index of action sequence, For the second micro-agent Time index of action sequence, For the second micro-agent The specific numerical values ​​of the action, For the first micro-intelligent agent The specific numerical values ​​of the action, For the first micro-agent index, This is the index for the second micro-agent.

[0085] Furthermore, based on the action sequences of micro-agents, when using the dynamic time warping algorithm to calculate the minimum warping path distance between any two action sequences of micro-agents, the first step is to extract the action value sequences of the two micro-agents within the same time range recorded during offline parallel training. The dynamic time warping algorithm finds the optimal warping path between two action sequences by constructing a cumulative cost matrix. Each element of the cumulative cost matrix calculates the Euclidean distance between corresponding point pairs and accumulates the previous minimum cumulative cost. The minimum warping path distance is the value of the lower right element of the cumulative cost matrix. The smaller this distance value, the more similar the action patterns of the two micro-agents are in the time dimension, thus obtaining the dynamic correlation strength between the micro-agent behavior data.

[0086] S3. Dynamically group the micro-agents according to the dynamic correlation strength to form agent groups, and deploy the causal correlation map, agent groups and trained micro-agents to the control unit of the physical atmospheric and vacuum distillation device.

[0087] S3.1. Based on the dynamic association strength, the spectral clustering algorithm is used to dynamically group the micro-agents to obtain the preliminary agent grouping structure.

[0088] Furthermore, based on the dynamic correlation strength matrix among the behavioral data of micro-agents, when using the spectral clustering algorithm to dynamically group micro-agents, the dynamic correlation strength matrix is ​​first input into the spectral clustering algorithm as a similarity matrix. The spectral clustering algorithm calculates the normalized Laplacian matrix and performs eigenvalue decomposition, selecting the top K eigenvectors to form a new feature matrix. K-means clustering analysis is performed on the row vectors of the feature matrix to group micro-agents with similar behavioral patterns into the same cluster. Each cluster contains a group of micro-agents with high dynamic correlation strength, forming a preliminary agent grouping structure.

[0089] S3.2 Input the preliminary agent grouping structure and the causal relationship map of the atmospheric and vacuum distillation device into the Louvain community discovery algorithm, and perform iterative calculation and optimization to form agent groups by maximizing the modularity Q value as the objective function.

[0090] The expression for the modularity Q-value is:

[0091] ;

[0092] in, For modularity, For the first intelligent agent With the second intelligent agent Connection strength in causal relationship graphs For the first intelligent agent Aggregation affects weights. For the second intelligent agent Aggregation affects weights. For the first intelligent agent Identify the source node, As a second intelligent agent Identify the target node. This represents the total weight of all edges in the network.

[0093] Furthermore, when the initial agent grouping structure and the causal relationship graph of the atmospheric and vacuum distillation unit are input into the Louvain community detection algorithm, the causal relationship graph of the atmospheric and vacuum distillation unit is first converted into a weighted network structure, where nodes represent process variables controlled by micro-agents and edge weights represent the strength of causal effects. The Louvain community detection algorithm performs multiple rounds of iterative calculation with the objective function of maximizing the modularity Q value. Each round of iteration includes two stages: local movement and community aggregation. In the local movement stage, each node is adjusted to a community that maximizes the modularity Q value. In the community aggregation stage, nodes belonging to the same community are merged into supernodes and the network structure is updated. After multiple iterations, the modularity Q value no longer increases, and the optimized agent grouping structure is output.

[0094] S3.3 Deploy the grouped agents, the trained micro-agent policy network parameters, and the causal relationship graph of the atmospheric and vacuum distillation unit to the control unit of the physical atmospheric and vacuum distillation unit through a secure communication protocol.

[0095] Furthermore, when deploying the agent grouping structure, trained micro-agent policy network parameters, and causal relationship graph of the atmospheric and vacuum distillation unit to the control unit of the physical atmospheric and vacuum distillation unit via a secure communication protocol, a secure transmission channel is first established using the TLS encryption protocol. The agent grouping structure data, micro-agent policy network parameter data, and causal relationship graph data of the atmospheric and vacuum distillation unit to be transmitted are digitally signed and encrypted. The encrypted data packets are transmitted to the control unit of the physical atmospheric and vacuum distillation unit via industrial Ethernet. After receiving the data, the control unit verifies the digital signature and decrypts the data. The decrypted agent grouping structure data is used to configure the collaborative optimization logic in the control unit, the micro-agent policy network parameter data is loaded into the execution processors of each control loop, and the causal relationship graph data of the atmospheric and vacuum distillation unit is stored in a real-time database for online optimization. After all data loading is completed, the control unit sends a deployment ready signal to the monitoring center.

[0096] It should be noted that the introduction of agent grouping is based on a deep analysis of process coupling and behavioral synergy. By grouping strongly correlated manipulated variables and controlled variables into the same group, the coordination complexity among multiple agents is significantly reduced, avoiding the target conflict problem commonly found in traditional distributed control. The grouping structure concentrates communication and computing resources among agents within the group, greatly reducing unnecessary global interaction overhead and improving response efficiency. The grouping design provides a natural boundary for subsequent collaborative evolution updates, allowing intra-group policy optimization to focus more on the in-depth exploration of local coupling relationships. This ensures the targeting of optimization while avoiding the risk of chain oscillations in policy updates through inter-group isolation, thus achieving a synergistic improvement in control accuracy and operational efficiency.

[0097] S4. Receive the global optimization objective through the collaborative optimization layer in the atmospheric and vacuum distillation unit control unit, fuse the causal correlation graph and dynamic correlation strength, calculate the causal contribution of the micro-agent's actions to the global optimization objective, and generate decision guidance signals.

[0098] S4.1 The control unit of the atmospheric and vacuum distillation device consists of a lower-level executive intelligent agent layer, a middle-level collaborative optimization layer, and an upper-level cognitive decision-making layer.

[0099] Furthermore, the atmospheric and vacuum distillation unit control unit adopts a three-layer architecture. The upper cognitive decision-making layer is deployed on a real-time database server, communicating with the production management system via the OPC UA protocol to receive global optimization goals in the form of key performance indicators. The middle collaborative optimization layer is deployed on an advanced control server, connecting with the upper cognitive decision-making layer and the lower execution intelligent agent layer via industrial Ethernet. It is responsible for receiving global optimization goals, integrating multi-source knowledge, calculating causal contribution, and generating decision guidance signals. The lower execution intelligent agent layer is deployed in the edge controllers of each control loop, connecting with the actuators via fieldbus. It is responsible for receiving decision guidance signals, integrating local rewards, generating control commands, and issuing them for execution. The three layers exchange data through a standard communication protocol. The upper cognitive decision-making layer sends the parsed optimization goals downwards, the middle collaborative optimization layer feeds back the optimization status upwards and sends guidance signals downwards, and the lower execution intelligent agent layer feeds back the execution status upwards, forming a closed-loop control system.

[0100] S4.2 The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit receives the global optimization target issued by the upper layer. The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit integrates the dynamic correlation strength between the causal correlation map of the atmospheric and vacuum distillation unit and the behavioral data of micro-intelligent agents to construct a joint influence network.

[0101] Furthermore, after receiving the global optimization target instruction from the upper-level production management system, the collaborative optimization layer in the atmospheric and vacuum distillation unit control unit immediately initiates the joint influence network construction process. The collaborative optimization layer reads the causal relationship graph of the atmospheric and vacuum distillation unit from the real-time database, and at the same time obtains the latest dynamic correlation strength matrix between micro-agent behavior data from memory. The causal relationship graph of the atmospheric and vacuum distillation unit is used as the basic network topology, and the correlation relationships in the dynamic correlation strength matrix between micro-agent behavior data that are higher than a set threshold are added to the basic network as supplementary edges. The weight of the supplementary edges is assigned using the reciprocal of the normalized distance value in the dynamic correlation strength matrix between micro-agent behavior data, forming a joint influence network that simultaneously includes process causal relationships and behavioral synergistic relationships.

[0102] S4.3 Based on the joint influence network, counterfactual reasoning is used to calculate the causal contribution of each micro-agent's candidate actions to the global optimization objective.

[0103] The expression for causal contribution is:

[0104] ;

[0105] in, For causal contribution, For the observation condition set, To optimize the target variable globally, For the expectation value operator, The standard deviation is denoted as .

[0106] Furthermore, based on the joint influence network, when calculating the causal contribution of each micro-agent's candidate action to the global optimization objective using counterfactual reasoning, an intervention operation is first performed on each candidate action of each micro-agent on the joint influence network, fixing the value of the corresponding node to the candidate action value; the expected impact of the intervention operation on the global optimization objective variable is calculated through the causal propagation mechanism on the joint influence network, with causal propagation proceeding along the directed path in the joint influence network, comprehensively considering direct and indirect effects; the expected value of the global optimization objective after intervention is subtracted from the baseline expected value before intervention to obtain the absolute causal effect of the candidate action; finally, the absolute causal effect is divided by the historical standard deviation of the global optimization objective variable to obtain the standardized causal contribution value.

[0107] S4.4 The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit adjusts the causal contribution to a range that matches the dimensions of the micro-agent's own reward function through linear scaling, and transforms the scaled value into a decision guidance signal in the form of additional reward items.

[0108] Furthermore, the collaborative optimization layer in the atmospheric and vacuum distillation unit control unit performs dimensional matching processing on the calculated causal contribution value. The historical standard deviation of the micro-agent's own reward function output value and the historical standard deviation of the causal contribution value are used as the ratio of the two standard deviations as a linear scaling factor. This scaling factor is then multiplied by the causal contribution value to obtain an adjusted value consistent with the dimensions of the micro-agent's own reward function. The adjusted value is then directly used as the amplitude value of the additional reward item to construct a decision guidance signal.

[0109] S5. The micro-intelligent agent integrates the decision guidance signal with its own control target to generate control commands, and sends the control commands to the actuator of the physical atmospheric and vacuum distillation device.

[0110] S5.1, by weighted summing the squared term of the tracking error of the setpoint of the micro-agent's control objective and the squared term of the change in the manipulated variable, the reward function of the micro-agent's own control objective is obtained.

[0111] Furthermore, the reward function for the micro-agent's own control objective is constructed through a weighted summation. The squared term of the tracking error of the controlled variable reflects the degree of deviation between the actual value and the set value of the controlled variable; the squared term of the change in the manipulated variable reflects the intensity of the change in the magnitude of the manipulated variable's action at adjacent time points; the squared term of the tracking error of the set value and the squared term of the change in the manipulated variable are linearly combined according to preset weight coefficients, which are allocated according to the importance of the control objective. The resulting weighted sum is used as the output value of the reward function for the micro-agent's own control objective.

[0112] S5.2. The local reward value is obtained based on the reward function of the micro-agent's own control objective.

[0113] Furthermore, when deriving the local reward value based on the reward function of the micro-agent's own control objective, the micro-agent derives the control objective deviation, the tracking error of the controlled variable setpoint, and the change in the manipulated variable based on the current observation state. The reward function of the micro-agent's own control objective is constructed in the form of a negative sum of squares of deviations, combining the square of the temperature setpoint tracking error with the square of the flow rate change according to a certain weight; the reward function calculates the reward value in real time under the current state.

[0114] S5.3 Add the fused decision guidance signal of the micro-agent to the local reward value to obtain the total reward function of the micro-agent.

[0115] Furthermore, the decision guidance signal received by the micro-agent is algebraically added to the local reward value calculated based on the reward function of the micro-agent's own control objective. The decision guidance signal is directly superimposed on the local reward value as an additional reward item. The result of the addition constitutes the output of the micro-agent's total reward function, which reflects both the achievement of the local control objective and the degree of contribution to global optimization. The output of the micro-agent's total reward function serves as the target value for policy network updates, guiding the optimization of the micro-agent's subsequent decision-making direction.

[0116] S5.4 The policy network of the micro-agent generates the original action probability distribution based on the current observation state, selects an action index through the Boltzmann exploration policy, and converts it into the original control command value through the linear mapping function.

[0117] Furthermore, the policy network of the micro-agent receives the current observation state as input, calculates the activation values ​​of all nodes in the output layer through forward propagation, and generates the original action probability distribution after normalization by the softmax function. Based on the original action probability distribution, a Boltzmann exploration strategy is used for random sampling. The Boltzmann exploration strategy controls the exploration degree through a temperature parameter and extracts an action index number from the original action probability distribution. The sampled action index number is input into a preset linear mapping function, which converts the discrete action index number into a continuous original control command value.

[0118] S5.5 The original control command value is input to a first-order hysteresis filter for smoothing, generating the final execution command, which is then sent to the actuator of the physical atmospheric and vacuum distillation unit via an industrial communication protocol.

[0119] Furthermore, when the original control command value is input to the first-order hysteresis filter for smoothing, the first-order hysteresis filter uses a recursive filtering algorithm to calculate the final execution command. The final execution command at the current moment is equal to the weighted sum of the final execution command at the previous moment and the current original control command value. The weighting coefficient determines the filtering strength. After generating the final execution command, it is encapsulated into a standard data message through the OPC UA industrial communication protocol. The data message contains the target actuator address identifier and the command value. The encapsulated data message is transmitted to the corresponding actuator of the physical atmospheric and vacuum distillation unit via industrial Ethernet. The actuator receives and parses the data message and then executes the final execution command.

[0120] S6. Collect small sample data generated after the operation of the physical atmospheric and vacuum distillation device, and update the cooperative evolution strategy of the intelligent agents based on the small sample data.

[0121] S6.1 Extract a Mini-batch of data proportionally from the small sample data generated after the operation of the atmospheric and vacuum distillation unit.

[0122] Furthermore, when extracting a Mini-batch of data proportionally from the small sample data generated after the operation of the atmospheric and vacuum distillation unit, the absolute value of the temporal difference error of each experience in the small sample data is first calculated as a priority; sampling is carried out proportionally according to the priority value, and the experience with the larger the absolute value of the temporal difference error is selected into the Mini-batch data; the sampling process adopts a roulette wheel selection algorithm, and the probability of each experience being selected is proportional to its absolute value of the temporal difference error; the final Mini-batch data is then extracted.

[0123] S6.2. Calculate the local fitness of each micro-agent using the extracted Mini-batch data. Within each agent group, select the elite group of micro-agents based on the local fitness value, and perform simulated binary crossover and polynomial mutation operations to generate offspring policies.

[0124] Furthermore, the local fitness value of each micro-agent is obtained using the extracted mini-batch data. The local fitness value is the sum of the discounted cumulative rewards of all experience data in the mini-batch data. The discounted cumulative rewards are obtained using a time difference method. Within each agent group, micro-agents are ranked according to their local fitness values, and the top 50% of micro-agents are selected as the elite group. The policy network parameters of the elite group are subjected to simulated binary crossover, which generates new parameter combinations by simulating the chromosome crossover process in nature. The parameters after crossover are subjected to polynomial mutation, which increases policy diversity by adding random perturbations. Finally, offspring policies containing new parameter combinations are generated.

[0125] S6.3. Use the knowledge distillation loss function to mimic the output distribution of the elite group policy in the offspring policy to obtain the optimized offspring policy.

[0126] Furthermore, when using the knowledge distillation loss function to mimic the output distribution of the elite group policy in the offspring policy, the mini-batch data is first input into both the offspring policy and the elite group policy simultaneously, and the action probability distribution outputs of the two policies are calculated separately. The knowledge distillation loss function includes a Kullback-Leibler divergence term and a mean squared error term. The Kullback-Leibler divergence term measures the difference between the offspring policy and the elite group policy in terms of action probability distribution, while the mean squared error term measures the difference between the two in terms of state value function output. The knowledge distillation loss function is minimized using the gradient descent algorithm to optimize the network parameters of the offspring policy, so that the output distribution of the offspring policy gradually approaches the output distribution of the elite group policy, ultimately obtaining an optimized offspring policy that maintains diversity while inheriting the knowledge of the elite policy.

[0127] S6.4 Replace the micro-agent policy with the lowest local fitness in the agent group with the optimized offspring policy.

[0128] Furthermore, when replacing the policy of the micro-agent with the lowest local fitness in the agent group with the optimized offspring policy, firstly, all micro-agents in the agent group are sorted in ascending order according to the previously calculated local fitness values; then, several micro-agents with the lowest ranking are selected as the replacement targets, and the number of replacements is consistent with the number of generated offspring policies; the network parameters of the optimized offspring policy are directly assigned to the replaced micro-agents, overwriting their original policy network parameters; after the replacement is completed, the agent group maintains its original size.

[0129] This embodiment also provides an intelligent control system for atmospheric and vacuum distillation units based on AI agents, including: a model building module, which constructs a high-fidelity digital twin virtual model through the process flow diagram, piping and instrumentation diagram and historical operating data of the atmospheric and vacuum distillation unit, and defines and initializes micro agents according to the process control loop division principle.

[0130] The offline parallel training module performs offline parallel training on micro-agents in a high-fidelity digital twin virtual model, collects time-series data, uses a causal discovery algorithm to construct and update the causal relationship map of the atmospheric and vacuum distillation device, and calculates the dynamic correlation strength between the behavioral data of micro-agents.

[0131] The grouping module dynamically groups micro-agents according to the dynamic correlation strength to form agent groups. The causal correlation graph, agent groups, and trained micro-agents are then deployed to the control unit of the physical atmospheric and vacuum distillation device.

[0132] The decision guidance module receives the global optimization objective through the collaborative optimization layer in the atmospheric and vacuum distillation unit control unit, integrates the causal correlation graph and dynamic correlation strength, calculates the causal contribution of the micro-agent's actions to the global optimization objective, and generates a decision guidance signal.

[0133] The adaptive evolution module integrates decision guidance signals with its own control objectives to generate control commands, which are then sent to the actuators of the physical atmospheric and vacuum distillation unit. It collects small sample data generated after the physical atmospheric and vacuum distillation unit is in operation and updates the collaborative evolution strategy of the agents based on the small sample data.

[0134] This embodiment also provides a computer device applicable to the intelligent control method of atmospheric and vacuum distillation apparatus based on AI intelligent agents, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent control method of atmospheric and vacuum distillation apparatus based on AI intelligent agents as proposed in the above embodiment.

[0135] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0136] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent control method for an AI-based atmospheric and vacuum depressurization device as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0137] In summary, this invention achieves precise mapping of the physicochemical processes of an atmospheric and vacuum distillation unit and establishes a distributed control architecture by constructing a high-fidelity digital twin virtual model and initializing micro-agents. Furthermore, through offline parallel training and causal discovery in the virtual environment, combined with a dynamic time warping algorithm, a dynamic knowledge system integrating process causal relationships and agent behavior correlations is constructed, solving the problem of separating the mechanism model from data-driven approaches in multi-agent collaborative control. Through decision guidance based on agent grouping based on dynamic correlation strength and causal contribution calculation, deep coupling between global optimization objectives and local control is achieved, overcoming the problem of insufficient long-term adaptive capability caused by the lack of causal reasoning in traditional methods.

[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent control of an atmospheric and vacuum distillation device based on an AI agent, characterized in that: This includes constructing a high-fidelity digital twin virtual model using the process flow diagram, piping and instrumentation diagram, and historical operating data of the atmospheric and vacuum distillation unit, and defining and initializing micro-intelligent agents according to the process control loop partitioning principle; Offline parallel training of micro-agents is performed in a high-fidelity digital twin virtual model, and time-series data is collected. Causal discovery algorithms are used to construct a causal correlation graph of the atmospheric and vacuum diversion device, and the dynamic correlation strength between the behavioral data of micro-agents is calculated. The process includes the following steps: In a high-fidelity digital twin virtual model, a multi-agent proximal policy optimization algorithm is used to set up an adversarial course from easy to difficult. A performance threshold is set based on a specific percentage of the maximum reward value obtained by a micro-agent in a high-fidelity digital twin virtual model during the adversarial course phase. After the micro-agent reaches the performance threshold in each course stage, it is then moved to a more difficult course to complete the offline parallel training of the micro-agent. During offline parallel training, time-series data of micro-agent action sequences and high-fidelity digital twin virtual model states are collected. Using the time-series data and injecting process knowledge constraints, a causal discovery algorithm is executed to construct a causal relationship graph of the atmospheric and vacuum distillation unit. Based on the causal correlation graph of the atmospheric and vacuum distillation unit, the linear non-Gaussian model method is used to quantify the causal effect strength of each edge in the causal correlation graph of the atmospheric and vacuum distillation unit. Based on the action sequences of micro-agents, the minimum warped path distance between any two action sequences of micro-agents is calculated using the dynamic time warping algorithm to obtain the dynamic correlation strength between micro-agent behavior data. Micro-agents are dynamically grouped according to the dynamic correlation strength to form agent groups. The causal correlation graph, agent groups, and trained micro-agents are then deployed to the control unit of the physical atmospheric and vacuum distillation device. The global optimization objective is received through the collaborative optimization layer in the control unit. The causal correlation graph and dynamic correlation strength are fused to calculate the causal contribution of the micro-agent's actions to the global optimization objective, generating a decision guidance signal. This process includes the following steps: The control unit of the atmospheric and vacuum distillation device consists of a lower-level executive intelligent agent layer, a middle-level collaborative optimization layer, and an upper-level cognitive decision-making layer. The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit receives the global optimization target from the upper layer. The collaborative optimization layer in the atmospheric and vacuum distillation unit control unit integrates the dynamic correlation strength between the causal correlation map of the atmospheric and vacuum distillation unit and the behavioral data of micro-intelligent agents to construct a joint influence network. Based on the joint influence network, counterfactual reasoning is used to calculate the causal contribution of each micro-agent's candidate actions to the global optimization objective; The collaborative optimization layer in the control unit of the atmospheric and vacuum distillation unit adjusts the causal contribution to a range that matches the dimensions of the micro-agent's own reward function through linear scaling, and transforms the scaled value into a decision guidance signal in the form of additional reward items; The micro-agent integrates decision-making guidance signals with its own control objectives to generate control commands, which are then sent to the actuators of the physical atmospheric and vacuum distillation unit. Small sample data generated after the physical atmospheric and vacuum distillation unit is in operation are collected, and the agent grouping is updated with a collaborative evolution strategy based on the small sample data.

2. The intelligent control method for atmospheric and vacuum distillation devices based on AI agents as described in claim 1, characterized in that: Using the process flow diagram, piping and instrumentation diagram, and historical operating data of the atmospheric and vacuum distillation unit, a high-fidelity digital twin virtual model is constructed. Based on the process control loop partitioning principles, micro-agents are defined and initialized, including the following steps: The topology, equipment connections, and fundamental differential-algebraic equations of a high-fidelity digital twin virtual model are defined using process flow diagrams and piping and instrumentation diagrams; a long short-term memory network compensator is trained using historical operating data; and a high-fidelity digital twin virtual model is constructed using the hybrid modeling method of the long short-term memory network compensator. The physical laws derived from the process flow diagram and piping and instrumentation diagram are transformed into mathematical constraints. The square of the degree of constraint violation is multiplied by the penalty coefficient and added to the overall loss function to construct the loss function of the physical constraint penalty term. The high-fidelity digital twin virtual model is trained and optimized using a loss function with a physical constraint penalty term. A global sensitivity analysis is performed on a high-fidelity digital twin virtual model to quantify the influence of manipulated variables on controlled variables. A global sensitivity matrix is ​​generated based on the process control loop partitioning principle. A spectral clustering algorithm is used to partition the strongly coupled manipulated and controlled variables into the same cluster. The cluster boundary defines the micro-agent. The micro-agent is initialized using a proximal policy optimization algorithm framework to obtain the policy network of the micro-agent.

3. The intelligent control method for a normal and low pressure ventilation device based on an AI agent as described in claim 1, characterized in that: The micro-agents are dynamically grouped according to the dynamic correlation strength to form agent groups. The causal correlation graph, agent groups, and trained micro-agents are then deployed to the control unit of the physical atmospheric and vacuum diversion device, including the following steps: Based on the dynamic correlation strength, the spectral clustering algorithm is used to dynamically group the micro-agents to obtain the preliminary agent grouping structure. The preliminary agent grouping structure and the causal relationship graph of the atmospheric and vacuum distillation device are input into the Louvain community discovery algorithm. The agent grouping is formed by iterative calculation and optimization with the objective function of maximizing the modularity Q value. The intelligent agents are grouped, the trained micro-agent policy network parameters are used, and the causal relationship graph of the atmospheric and vacuum distillation unit is deployed to the control unit of the physical atmospheric and vacuum distillation unit through a secure communication protocol.

4. The intelligent control method for atmospheric and vacuum distillation devices based on AI agents as described in claim 3, characterized in that: The micro-agent integrates decision-making guidance signals with its own control objectives to generate control commands, and then sends these control commands to the actuators of the physical atmospheric and vacuum distillation unit, including the following steps: The reward function for the micro-agent's own control objective is obtained by weighted summing of the squared term of the tracking error of the setpoint of the micro-agent's control objective and the squared term of the change in the manipulated variable. The local reward value is derived based on the reward function of the micro-agent's own control objective; The total reward function of the micro-agent is obtained by adding the fused decision guidance signal to the local reward value. The policy network of the micro-agent generates the original action probability distribution based on the current observation state, selects an action index through the Boltzmann exploration policy, and converts it into the original control command value through the linear mapping function; The original control command value is input to a first-order hysteresis filter for smoothing, and the final execution command is generated and sent to the actuator of the physical atmospheric and vacuum distillation unit through the industrial communication protocol.

5. The intelligent control method for a normal and low pressure vacuum device based on an AI agent as described in claim 4, characterized in that: Collect small sample data generated after the operation of the physical atmospheric and vacuum distillation unit, and update the cooperative evolution strategy of the intelligent agents based on the small sample data, including the following steps: A mini-batch of data is extracted proportionally from the small sample data generated after the operation of the physical atmospheric and vacuum distillation unit. The local fitness of each micro-agent is calculated using the extracted mini-batch data. Within each agent group, an elite group of micro-agents is selected based on the local fitness value, and simulated binary crossover and polynomial mutation operations are performed to generate offspring policies. The knowledge distillation loss function is used to mimic the output distribution of the elite group's policy in the offspring policy, resulting in an optimized offspring policy. Replace the micro-agent policy with the lowest local fitness in the agent group with the optimized offspring policy.

6. An intelligent control system for an atmospheric and vacuum distillation apparatus based on an AI agent, comprising the intelligent control method for an atmospheric and vacuum distillation apparatus based on an AI agent as described in any one of claims 1 to 5, characterized in that: This includes a model building module, which constructs a high-fidelity digital twin virtual model using the process flow diagram, piping and instrumentation diagram, and historical operating data of the atmospheric and vacuum distillation unit, and defines and initializes micro-intelligent agents according to the process control loop division principle; The offline parallel training module performs offline parallel training on micro-agents in a high-fidelity digital twin virtual model, collects time-series data, uses a causal discovery algorithm to construct a causal correlation graph of the atmospheric and vacuum distillation device, and calculates the dynamic correlation strength between the behavioral data of micro-agents. The grouping module dynamically groups micro-agents according to the dynamic correlation strength to form agent groups, and then deploys the causal correlation graph, agent groups, and trained micro-agents to the control unit of the physical atmospheric and vacuum distillation device. The decision guidance module receives the global optimization objective through the collaborative optimization layer in the control unit, integrates the causal correlation graph and the dynamic correlation strength, calculates the causal contribution of the micro-agent's actions to the global optimization objective, and generates a decision guidance signal. The adaptive evolution module integrates decision guidance signals with its own control objectives to generate control commands, which are then sent to the actuators of the physical atmospheric and vacuum distillation unit. It collects small sample data generated after the physical atmospheric and vacuum distillation unit is in operation and updates the collaborative evolution strategy of the agents based on the small sample data.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent control method for atmospheric and vacuum distillation devices based on AI agents as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent control method for atmospheric and vacuum distillation devices based on AI agents as described in any one of claims 1 to 5.

Images