A gas separation regulation system based on multi-agent cooperation

Abstract

This invention discloses a gas separation and control system based on multi-agent collaboration, comprising: a data processing module for collecting and preprocessing operational data to form a set of operating condition features; a role allocation module for constructing a multi-agent collaborative structure and assigning roles to generate state feature vectors; a feature processing module for feature decoupling, collaborative encoding, and global attention aggregation to drive policy search; a policy selection module for constructing a role graph and selecting convergent policies; a policy optimization module for constraint optimization and generating parsed instruction sequences; a control execution module for controlling flow rate, pressure, adsorption stage, and cooling capacity allocation and updating feature vectors; and a policy iteration module for evaluating effects, calculating collaborative deviations, and adjusting policies to achieve continuous optimization. This invention achieves high-purity, low-energy-consumption, steady-state precision, and continuous adaptive optimization control of gas separation devices under complex operating conditions through multi-agent collaboration and global attention optimization.

Description

Technical Field

[0001] This invention relates to the field of multi-agent collaborative decision-making technology, and in particular to a gas separation and control system based on multi-agent collaboration. Background Technology

[0002] In the field of gas separation, existing technologies mainly rely on traditional single controllers or distributed control strategies to regulate flow rate, pressure difference, adsorption stage, and cooling capacity distribution in gas separation devices. These methods typically employ fixed-parameter PID control or simple logic control, which struggles to fully utilize the dynamic correlation information between various components. This results in significant fluctuations in gas purity, high energy consumption, and insufficient device stability under multi-disturbance operating conditions. Furthermore, existing control methods lack the ability to comprehensively optimize multiple objectives, failing to achieve optimal energy consumption and steady-state constraint coordination while ensuring gas purity, thus failing to meet the requirements for continuous operation under complex conditions.

[0003] In recent years, artificial intelligence (AI) technology has begun to be applied in gas separation control, including single-agent reinforcement learning control and model prediction-based optimization methods. However, these methods often design agent strategies only for a single control objective or fixed operating conditions, lacking multi-agent collaborative structures and role allocation mechanisms. This results in low strategy generation efficiency, poor online adaptability, and difficulty in achieving global collaboration in prediction, regulation, and energy consumption constraints. Furthermore, existing AI control methods do not fully utilize the temporal characteristics and state dependencies of operating conditions during feature processing and strategy generation, easily leading to slow strategy convergence or local optima, thus limiting their practical application in industrial gas separation devices.

[0004] Therefore, how to provide a gas separation and control system based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a gas separation and control system based on multi-agent collaboration. This invention fully utilizes multi-agent collaborative algorithms, graph neural networks, and global attention mechanisms, and describes in detail the complete control process from operating condition feature acquisition, role allocation, feature decoupling and collaborative encoding, to policy search, constraint optimization, and multi-objective trade-offs, to achieve continuous optimization control of the gas separation device. This invention has the advantages of stable gas purity, low energy consumption, accurate steady-state control, and rapid adaptive iteration of the policy.

[0006] A gas separation and control system based on multi-agent cooperation according to an embodiment of the present invention includes:

[0007] The data processing module is used to collect the operating data of the gas separation unit and preprocess it to form a set of operating condition features.

[0008] The role assignment module is used to construct a multi-agent collaborative structure, assign prediction roles, regulation roles, and energy consumption constraint roles to each agent, and generate a set of feature vectors for each role's corresponding state.

[0009] The feature processing module is used to perform feature decoupling processing, cooperative coding processing and global attention aggregation processing on the feature vector set respectively, driving multiple agents to generate policy search trajectories in the virtual working space;

[0010] The strategy filtering module is used to construct an interactive role graph structure according to role dependencies, and to perform strategy convergence calculation on the strategy search trajectory, and filter strategies with high convergence to form a strategy candidate set.

[0011] The strategy optimization module is used to perform constraint optimization on the set of strategy candidates, generate a set of control instructions, and perform multi-objective trade-off analysis to form an analytical instruction sequence;

[0012] The control and execution module is used to control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and to update the feature vector set through a multi-agent collaborative structure.

[0013] The strategy iteration module is used to evaluate the performance and calculate the cooperation deviation between roles, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

[0014] Optionally, modules can be integrated using the following methods:

[0015] S1. Collect pressure data, flow data, temperature data and component data of the gas separation device, perform noise suppression processing and time-series feature extraction processing to form a set of operating condition features;

[0016] S2. Construct a multi-agent collaborative structure based on the working condition feature set, assign prediction role, regulation role and energy consumption constraint role to each agent, and generate a feature vector set of the corresponding state for each role;

[0017] S3. Perform feature decoupling, co-coding, and global attention aggregation on the feature vector set to drive the multi-agent to generate policy search trajectories in the virtual working space.

[0018] S4. Construct an interactive role graph structure according to role dependencies, calculate the policy convergence of the policy search trajectory, and select policies with high policy convergence to form a policy candidate set.

[0019] S5. Perform constraint optimization on the strategy candidate set, generate a set of control instructions under the conditions of satisfying gas purity constraints, energy consumption constraints and steady state constraints, and perform multi-objective trade-off analytical processing to form an analytical instruction sequence;

[0020] S6. Based on the parsing instruction sequence, control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and update the feature vector set through a multi-agent collaborative structure.

[0021] S7. Perform effect evaluation and inter-role cooperation deviation calculation on the updated feature vector set, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

[0022] Optionally, S2 specifically includes:

[0023] S21. Derive the time window sample set from the working condition feature set, perform standardization and zero mean variance normalization on each time window, calculate the time domain features and frequency domain features, and identify the obtained features according to the sensor source, physical quantity type and sampling rate.

[0024] The physical quantity type and sampling rate are used to group and map homogeneous features during the feature processing stage, so that they can be divided into feature dimensions according to physical attributes in the subsequent encoding process, and maintain time series alignment according to the sampling consistency result.

[0025] S22. Using an unsupervised partitioning method, the identified feature set is divided into several state clusters according to the operating state. A role mapping table is constructed based on the state clusters and the gas separation device. The state clusters related to the prediction task are mapped to the prediction role, the state clusters related to the execution control are mapped to the regulation role, and the state clusters related to the energy consumption constraint are mapped to the energy consumption constraint role.

[0026] S23. Instantiate the multi-agent collaborative structure based on the role mapping table:

[0027] Assign an independent agent identifier, observation domain, and action domain to each mapped role instance;

[0028] The communication topology between agents is set as a directed weighted graph, and the communication weights are determined by the correlation of state clusters;

[0029] Communication latency and bandwidth parameters are registered by the edge computing resource scheduler;

[0030] S24. Construct a set of feature vectors for each instantiated agent corresponding to its state:

[0031] Temporal pruning, sampling uniformity, and missing value imputation are performed on the data features within the observation domain of the state cluster. The preprocessed data features are mapped into fixed-dimensional embedding vectors, and a set of feature vectors is formed in the embedding space according to the time series, with timestamps and confidence indices attached.

[0032] S25. Within the multi-agent collaborative structure, define role-level policy networks for the prediction role, regulation role, and energy consumption constraint role respectively:

[0033] The policy network for predicting roles takes a set of historical feature vectors as input and outputs a predicted vector for the corresponding future state.

[0034] The policy network for controlling the role takes the current set of feature vectors and the prediction vectors from the predicted role as input, and outputs a set of candidate control actions;

[0035] The policy network for the energy-constrained role takes the current energy consumption estimate and the set of control action candidates as input and outputs energy consumption constraint information.

[0036] S26. Establish local collaboration mechanisms within the role level:

[0037] Based on the message passing protocol, the feature vector set, prediction vector, control action candidate set and energy consumption constraint information are cyclically exchanged on the communication topology, and the attention fusion operator is used to weight and aggregate the received messages to update the feature vector set of the local state.

[0038] S27. Perform consistency checks and role dependency corrections on the feature vector set:

[0039] Abnormal vectors are removed based on preset consistency criteria, and interpolation repair is performed on vectors affected by communication delays or packet loss. The correction results are then written back to the role mapping table for initialization of the next round of strategy search trajectories.

[0040] Optionally, S3 specifically includes:

[0041] S31. The feature vector set is segmented according to the time window and local normalization is performed. The time series features are decomposed based on principal component analysis and independent component analysis. The decomposition results are sorted and representative components are retained to form a decoupled sub-feature set.

[0042] S32. Label the decoupled sub-feature sets with associated tags according to the physical quantity type, perform adaptive filtering and short-time Fourier transform on each sub-feature set to construct a frequency domain representation, and encode the time domain representation and the frequency domain representation in parallel;

[0043] S33. Establish a co-encoder structure. The co-encoder consists of several parallel branches. Each branch is responsible for the deep encoding of a sub-feature view. Residual connections and layer normalization are used inside the branches to ensure numerical stability. The branch encoding results are mapped to a unified dimension embedding through a projection layer.

[0044] S34. Perform global attention aggregation processing on the branch encoding results:

[0045] Calculate query, key, and value vectors and construct an attention matrix based on time series;

[0046] A multi-head attention mechanism is used to compute attention weights in parallel.

[0047] The attention weights are standardized using a normalization function, and the value vectors are weighted and summed to generate a global aggregate vector.

[0048] S35. Construct a policy searcher in the virtual working space based on the global aggregation vector, and drive multiple agents to generate policy search trajectories. The policy searcher uses a combination of parallel sampling strategy and heuristic search to perform policy detection. The sampling process records the state-action-reward triplet and writes it into the trajectory buffer. The heuristic search guides the sampling results based on role priority constraints.

[0049] The virtual operating condition space is a high-dimensional representation space constructed based on the operating condition feature set and feature vector set. It is used to simulate the operating status, interactive behavior and strategy search trajectory of the gas separation device under a multi-agent collaborative structure, and realize a simulation environment for strategy generation, evaluation and iteration.

[0050] Optionally, S4 specifically includes:

[0051] S41. A cross-interaction module is built outside the co-encoder to realize message passing between agents through a graph neural network. The graph neural network uses agents as nodes and role dependencies as edges. The edge weights are determined by the correlation of state clusters, and embeddings from different branches are integrated during message passing.

[0052] S42. Perform convergence evaluation on the policy search trajectory within the trajectory buffer:

[0053] The strategy search trajectory is evaluated in segments according to time order, and the evaluation results are sorted to select strategies with high convergence to form a strategy candidate set.

[0054] S43. Perform constraint satisfaction verification on the policy candidate set and feed the verification results back to the co-encoder and graph neural network to update the encoder parameters and message passing weights, completing a closed-loop process from feature decoupling to policy generation.

[0055] Optionally, S5 specifically includes:

[0056] S51. Extract the corresponding action sequence and state response vector for each candidate strategy in the strategy candidate set, perform simulation prediction on the action sequence according to the role priority constraint, and calculate the gas purity, energy consumption and steady-state deviation index of each action sequence under the current working condition.

[0057] S52. Establish a constrained optimization framework, set the gas purity constraint as the minimum threshold condition, the energy consumption constraint as the upper limit condition, and the steady state constraint as the deviation tolerance condition. Perform feasibility screening on the action sequences of candidate strategies, and retain only the strategies that meet the three types of constraints to form a set of feasible strategies.

[0058] S53. Perform multi-objective trade-off analysis on the set of feasible strategies:

[0059] A weighted comprehensive score is calculated based on gas purity, energy consumption, and steady-state deviation indicators. A multi-objective ranking algorithm is used to prioritize the strategies, and the cooperative dependency and resource constraint effects between strategies are considered during the ranking process.

[0060] S54. Generate a set of control instructions based on the multi-objective analysis results:

[0061] The action sequence corresponding to the highest priority strategy is converted into specific control signals, including flow regulation, pressure difference regulation, adsorption stage switching and cooling capacity allocation, along with execution order and time interval information;

[0062] S55. Serialize the set of control commands to form a parsed command sequence:

[0063] Based on action dependencies and role collaboration requirements, instructions are arranged in sequence and synchronized. The role affiliation, execution start and end time and feedback channel of each instruction are recorded so that the multi-agent collaborative structure can update the feature vector set and implement the next round of policy iteration.

[0064] The beneficial effects of this invention are:

[0065] First, this invention achieves global coordinated control of prediction, regulation, and energy consumption constraints through a multi-agent collaborative structure and role allocation mechanism, enabling each agent to search and iterate strategies in a virtual operating space, which can effectively improve the stability of gas purity and reduce fluctuations caused by operating condition disturbances.

[0066] Secondly, by performing feature decoupling, collaborative encoding, and global attention aggregation on the feature vector set, and combining constraint optimization and multi-objective trade-off analysis, this invention can generate efficient, feasible control commands that take into account both energy consumption and steady-state requirements, thereby achieving a synergistic balance between energy consumption optimization and steady-state control and improving the overall operating efficiency of the system.

[0067] Finally, through strategy iteration and closed-loop feedback mechanisms, this invention enables multiple agents to continuously adjust their strategy search trajectories during the continuous optimization process, thereby achieving rapid adaptive control of the gas separation device. This improves the system response speed and strategy convergence, and thus significantly enhances the device's continuous optimization capability and operational reliability while ensuring production safety. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0069] Figure 1 This is a block diagram of a gas separation and control system based on multi-agent collaboration proposed in this invention.

[0070] Figure 2 This is a flowchart of a gas separation and control system based on multi-agent collaboration proposed in this invention.

[0071] Figure 3 This is a flowchart illustrating the strategy selection, optimization, and execution process of a gas separation and control system based on multi-agent collaboration proposed in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figure 1 A gas separation and control system based on multi-agent collaboration, comprising:

[0074] The data processing module is used to collect the operating data of the gas separation unit and preprocess it to form a set of operating condition features.

[0075] The role assignment module is used to construct a multi-agent collaborative structure, assign prediction roles, regulation roles, and energy consumption constraint roles to each agent, and generate a set of feature vectors for each role's corresponding state.

[0076] The feature processing module is used to perform feature decoupling processing, cooperative coding processing and global attention aggregation processing on the feature vector set respectively, driving multiple agents to generate policy search trajectories in the virtual working space;

[0077] The strategy filtering module is used to construct an interactive role graph structure according to role dependencies, and to perform strategy convergence calculation on the strategy search trajectory, and filter strategies with high convergence to form a strategy candidate set.

[0078] The strategy optimization module is used to perform constraint optimization on the set of strategy candidates, generate a set of control instructions, and perform multi-objective trade-off analysis to form an analytical instruction sequence;

[0079] The control and execution module is used to control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and to update the feature vector set through a multi-agent collaborative structure.

[0080] The strategy iteration module is used to evaluate the performance and calculate the cooperation deviation between roles, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

[0081] refer to Figure 2-3 In this embodiment, the modules are interconnected using the following method:

[0082] S1. Collect pressure data, flow data, temperature data and component data of the gas separation device, perform noise suppression processing and time-series feature extraction processing to form a set of operating condition features;

[0083] S2. Construct a multi-agent collaborative structure based on the working condition feature set, assign prediction role, regulation role and energy consumption constraint role to each agent, and generate a feature vector set of the corresponding state for each role;

[0084] S3. Perform feature decoupling, co-coding, and global attention aggregation on the feature vector set to drive the multi-agent to generate policy search trajectories in the virtual working space.

[0085] S4. Construct an interactive role graph structure according to role dependencies, calculate the policy convergence of the policy search trajectory, and select policies with high policy convergence to form a policy candidate set.

[0086] S5. Perform constraint optimization on the strategy candidate set, generate a set of control instructions under the conditions of satisfying gas purity constraints, energy consumption constraints and steady state constraints, and perform multi-objective trade-off analytical processing to form an analytical instruction sequence;

[0087] S6. Based on the parsing instruction sequence, control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and update the feature vector set through a multi-agent collaborative structure.

[0088] S7. Perform effect evaluation and inter-role cooperation deviation calculation on the updated feature vector set, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

[0089] In this embodiment, S2 specifically includes:

[0090] S21. Derive the time window sample set from the working condition feature set, perform standardization and zero mean variance normalization on each time window, calculate the time domain features and frequency domain features, and identify the obtained features according to the sensor source, physical quantity type and sampling rate.

[0091] The physical quantity type and sampling rate are used to group and map homogeneous features during the feature processing stage, so that they can be divided into feature dimensions according to physical attributes in the subsequent encoding process, and maintain time series alignment according to the sampling consistency result.

[0092] S22. Using an unsupervised partitioning method, the identified feature set is divided into several state clusters according to the operating state. A role mapping table is constructed based on the state clusters and the gas separation device. The state clusters related to the prediction task are mapped to the prediction role, the state clusters related to the execution control are mapped to the regulation role, and the state clusters related to the energy consumption constraint are mapped to the energy consumption constraint role.

[0093] S23. Instantiate the multi-agent collaborative structure based on the role mapping table:

[0094] Assign an independent agent identifier, observation domain, and action domain to each mapped role instance;

[0095] The communication topology between agents is set as a directed weighted graph, and the communication weights are determined by the correlation of state clusters;

[0096] Communication latency and bandwidth parameters are registered by the edge computing resource scheduler;

[0097] S24. Construct a set of feature vectors for each instantiated agent corresponding to its state:

[0098] Temporal pruning, sampling uniformity, and missing value imputation are performed on the data features within the observation domain of the state cluster. The preprocessed data features are mapped into fixed-dimensional embedding vectors, and a set of feature vectors is formed in the embedding space according to the time series, with timestamps and confidence indices attached.

[0099] S25. Within the multi-agent collaborative structure, define role-level policy networks for the prediction role, regulation role, and energy consumption constraint role respectively:

[0100] The policy network for predicting roles takes a set of historical feature vectors as input and outputs a predicted vector for the corresponding future state.

[0101] The policy network for controlling the role takes the current set of feature vectors and the prediction vectors from the predicted role as input, and outputs a set of candidate control actions;

[0102] The policy network for the energy-constrained role takes the current energy consumption estimate and the set of control action candidates as input and outputs energy consumption constraint information.

[0103] S26. Establish local collaboration mechanisms within the role level:

[0104] Based on the message passing protocol, the feature vector set, prediction vector, control action candidate set and energy consumption constraint information are cyclically exchanged on the communication topology, and the attention fusion operator is used to weight and aggregate the received messages to update the feature vector set of the local state.

[0105] S27. Perform consistency checks and role dependency corrections on the feature vector set:

[0106] Abnormal vectors are removed based on preset consistency criteria, and interpolation repair is performed on vectors affected by communication delays or packet loss. The correction results are then written back to the role mapping table for initialization of the next round of strategy search trajectories.

[0107] In this embodiment, S3 specifically includes:

[0108] S31. The feature vector set is segmented according to the time window and local normalization is performed. The time series features are decomposed based on principal component analysis and independent component analysis. The decomposition results are sorted and representative components are retained to form a decoupled sub-feature set.

[0109] S32. Label the decoupled sub-feature sets with associated tags according to the physical quantity type, perform adaptive filtering and short-time Fourier transform on each sub-feature set to construct a frequency domain representation, and encode the time domain representation and the frequency domain representation in parallel;

[0110] S33. Establish a co-encoder structure. The co-encoder consists of several parallel branches. Each branch is responsible for the deep encoding of a sub-feature view. Residual connections and layer normalization are used inside the branches to ensure numerical stability. The branch encoding results are mapped to a unified dimension embedding through a projection layer.

[0111] S34. Perform global attention aggregation processing on the branch encoding results:

[0112] Calculate query, key, and value vectors and construct an attention matrix based on time series;

[0113] A multi-head attention mechanism is used to compute attention weights in parallel.

[0114] The attention weights are standardized using a normalization function, and the value vectors are weighted and summed to generate a global aggregate vector.

[0115] S35. Construct a policy searcher in the virtual working space based on the global aggregation vector, and drive multiple agents to generate policy search trajectories. The policy searcher uses a combination of parallel sampling strategy and heuristic search to perform policy detection. The sampling process records the state-action-reward triplet and writes it into the trajectory buffer. The heuristic search guides the sampling results based on role priority constraints.

[0116] The virtual operating condition space is a high-dimensional representation space constructed based on the operating condition feature set and feature vector set. It is used to simulate the operating status, interactive behavior and strategy search trajectory of the gas separation device under a multi-agent collaborative structure, and realize a simulation environment for strategy generation, evaluation and iteration.

[0117] In this embodiment, S4 specifically includes:

[0118] S41. A cross-interaction module is constructed outside the co-encoder to realize message passing between agents through a graph neural network. The graph neural network uses agents as nodes and role dependencies as edges, with edge weights determined by state cluster correlations. Furthermore, embeddings from different branches are fused during message passing. Specifically, this includes:

[0119] Each agent node maintains an embedding vector in the graph neural network, and combines the embedding vector with the weights of the role dependency edges in a weighted combination, generating an aggregated message through the neighbor node aggregation operation;

[0120] Aggregated messages are passed along the graph edges in multiple iterations. In each iteration, linear transformation and nonlinear activation are first performed to update node embeddings, and then the message passing weights are adjusted through a normalization function to ensure that the information passing between different roles conforms to the dependency relationship.

[0121] In each message transmission, the agent node receives feature embeddings from multiple encoding branches and uses an attention mechanism to assign dynamic weights to the embeddings of different branches, while recording the timestamps and confidence indices of messages during the interaction process.

[0122] S42. Perform convergence evaluation on the policy search trajectory within the trajectory buffer:

[0123] The policy search trajectory is evaluated in segments according to time sequence, and the evaluation results are ranked. Policies with high convergence are selected to form a policy candidate set, specifically including:

[0124] Each strategy search trajectory is divided into several continuous segments according to the time series, and the action sequence and state response vector in each segment are normalized to eliminate the difference in dimensions.

[0125] For each sub-segment, the magnitude of change in the action sequence, the degree of fluctuation in the state response, and the stability of the target index are calculated. The cumulative value of the action change rate and the state deviation are statistically analyzed by using a sliding window to quantify the convergence trend of the trajectory in the time dimension.

[0126] The convergence indices of each sub-segment are weighted and combined to form the convergence score of the entire trajectory. The score is then adjusted based on the similarity between trajectories and the multi-agent role dependency relationship so that the score can reflect the overall collaborative effect.

[0127] The trajectories are sorted by convergence score, and strategies with high convergence scores, good motion stability, and satisfying inter-role cooperation constraints are selected to form a strategy candidate set.

[0128] S43. Perform constraint satisfaction verification on the policy candidate set and feed the verification results back to the co-encoder and graph neural network to update the encoder parameters and message passing weights, completing a closed-loop process from feature decoupling to policy generation.

[0129] In this embodiment, S5 specifically includes:

[0130] S51. Extract the corresponding action sequence and state response vector for each candidate strategy in the strategy candidate set, perform simulation prediction on the action sequence according to the role priority constraint, and calculate the gas purity, energy consumption and steady-state deviation index of each action sequence under the current working condition.

[0131] S52. Establish a constrained optimization framework, set the gas purity constraint as the minimum threshold condition, the energy consumption constraint as the upper limit condition, and the steady state constraint as the deviation tolerance condition. Perform feasibility screening on the action sequences of candidate strategies, and retain only the strategies that meet the three types of constraints to form a set of feasible strategies.

[0132] S53. Perform multi-objective trade-off analysis on the set of feasible strategies:

[0133] A weighted comprehensive score is calculated based on gas purity, energy consumption, and steady-state deviation indicators. A multi-objective ranking algorithm is used to prioritize the strategies, and the cooperative dependency and resource constraint effects between strategies are considered during the ranking process.

[0134] S54. Generate a set of control instructions based on the multi-objective analysis results:

[0135] The action sequence corresponding to the highest priority strategy is converted into specific control signals, including flow regulation, pressure difference regulation, adsorption stage switching and cooling capacity allocation, along with execution order and time interval information;

[0136] S55. Serialize the set of control commands to form a parsed command sequence:

[0137] Based on action dependencies and role collaboration requirements, instructions are arranged in sequence and synchronized. The role affiliation, execution start and end time and feedback channel of each instruction are recorded so that the multi-agent collaborative structure can update the feature vector set and implement the next round of policy iteration.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to a large-scale industrial gas separation device. This device primarily handles the separation of multi-component mixed gases, including gas purity control, energy management, and steady-state maintenance. In existing technologies, this device suffers from large fluctuations in gas purity, high energy consumption, and lag in control response under high load and fluctuating gas source conditions, leading to decreased overall production efficiency and increased equipment load. By employing the multi-agent collaborative gas separation control system proposed in this invention, continuous optimized control of the device under complex operating conditions can be achieved.

[0140] In this scenario, the present invention first collects pressure, flow rate, temperature, and gas composition data of the device through a sensor network, and preprocesses the data at the edge to form a set of operating condition features. Then, a multi-agent collaborative structure is constructed, assigning prediction, regulation, and energy consumption constraint roles to each agent, and generating a set of state feature vectors for each role. During feature processing, the feature vector set undergoes feature decoupling to eliminate interference between features of different dimensions; through collaborative coding, the features of each role are mapped to a unified embedding space, achieving cross-agent information fusion; finally, through global attention aggregation, a policy search vector reflecting the global state is generated, enabling the multi-agent to generate efficient policy search trajectories in the virtual operating condition space.

[0141] The generated strategy candidate set undergoes constraint optimization to ensure that gas purity is not lower than a set threshold, energy consumption is within a reasonable range, and system steady-state indicators meet requirements. Then, through multi-objective trade-off analysis, an analytical command sequence is formed, which is controlled by the control execution module to perform flow regulation, pressure difference regulation, adsorption stage switching, and cooling distribution. During continuous iteration, the strategy iteration module evaluates the execution effect, calculates the cooperation deviation between agents, adjusts the strategy search trajectory, and generates the next round of control strategies, thereby ensuring that the entire device maintains optimal operation under complex disturbance conditions.

[0142] In practical applications, the system of this invention continuously monitors and controls gas purity, energy consumption, and steady-state indicators under conditions of high load fluctuations and changes in gas source composition. Through a three-month continuous simulation experiment, data shows that after adopting this invention, the gas purity fluctuation range of the device is reduced to within 0.8%, approximately three times less than existing technologies; energy consumption per unit processed gas volume is reduced by approximately 12%; under high load conditions, pressure and flow response time is reduced from an average of 12 seconds to 4 seconds, and the system steady-state deviation is reduced from ±4% to ±1.2%, significantly improving the device's production efficiency and operational stability.

[0143] To more intuitively demonstrate the effectiveness of this invention, the following table presents continuous monitoring data under typical operating conditions:

[0144] Table 1. Comparison of operational performance of gas separation and control systems based on multi-agent collaboration

[0145] ;

[0146] As can be seen from Table 1, the present invention achieves a comprehensive control effect of high and stable gas purity, low energy consumption, fast response, and small steady-state deviation under different operating conditions, which fully demonstrates the effectiveness and superiority of the multi-agent collaborative strategy in gas separation and regulation.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A gas separation and control method based on multi-agent cooperation, characterized in that, Includes the following steps: S1. Collect pressure data, flow data, temperature data and component data of the gas separation device, perform noise suppression processing and time-series feature extraction processing to form a set of operating condition features; S2. Construct a multi-agent collaborative structure based on the working condition feature set, assign prediction role, regulation role and energy consumption constraint role to each agent, and generate a feature vector set of the corresponding state for each role; S3. Perform feature decoupling, co-coding, and global attention aggregation on the feature vector set to drive the multi-agent to generate policy search trajectories in the virtual working space. S3 specifically includes: S31. The feature vector set is segmented according to the time window and local normalization is performed. The time series features are decomposed based on principal component analysis and independent component analysis. The decomposition results are sorted and representative components are retained to form a decoupled sub-feature set. S32. Label the decoupled sub-feature sets with associated tags according to the physical quantity type, perform adaptive filtering and short-time Fourier transform on each sub-feature set to construct a frequency domain representation, and encode the time domain representation and the frequency domain representation in parallel; S33. Establish a co-encoder structure. The co-encoder consists of several parallel branches. Each branch is responsible for the deep encoding of a sub-feature view. Residual connections and layer normalization are used inside the branches to ensure numerical stability. The branch encoding results are mapped to a unified dimension embedding through a projection layer. S34. Perform global attention aggregation processing on the branch encoding results: Calculate query, key, and value vectors and construct an attention matrix based on time series; A multi-head attention mechanism is used to compute attention weights in parallel. The attention weights are standardized using a normalization function, and the value vectors are weighted and summed to generate a global aggregate vector. S35. Construct a policy searcher in the virtual working space based on the global aggregation vector, and drive multiple agents to generate policy search trajectories. The policy searcher uses a combination of parallel sampling strategy and heuristic search to perform policy detection. The sampling process records the state-action-reward triplet and writes it into the trajectory buffer. The heuristic search guides the sampling results based on role priority constraints. The virtual operating condition space is a high-dimensional representation space constructed based on the operating condition feature set and feature vector set. It is used to simulate the operating status, interactive behavior and strategy search trajectory of the gas separation device under the multi-agent collaborative structure, and realize the simulation environment for strategy generation, evaluation and iteration. S4. Construct an interactive role graph structure according to role dependencies, calculate the policy convergence of the policy search trajectory, and select policies with high policy convergence to form a policy candidate set. S5. Perform constraint optimization on the strategy candidate set, generate a set of control instructions under the conditions of gas purity constraint, energy consumption constraint and steady state constraint, and perform multi-objective trade-off analytical processing to form an analytical instruction sequence; S6. Based on the parsing instruction sequence, control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and update the feature vector set through a multi-agent collaborative structure. S7. Perform effect evaluation and inter-role cooperation deviation calculation on the updated feature vector set, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

2. The gas separation and control method based on multi-agent cooperation according to claim 1, characterized in that, S2 specifically includes: S21. Derive the time window sample set from the working condition feature set, perform standardization and zero mean variance normalization on each time window, calculate the time domain features and frequency domain features, and identify the obtained features according to the sensor source, physical quantity type and sampling rate. The physical quantity type and sampling rate are used to group and map homogeneous features during the feature processing stage, so that they can be divided into feature dimensions according to physical attributes in the subsequent encoding process, and maintain time series alignment according to the sampling consistency result. S22. Using an unsupervised partitioning method, the identified feature set is divided into several state clusters according to the operating state. A role mapping table is constructed based on the state clusters and the gas separation device. The state clusters related to the prediction task are mapped to the prediction role, the state clusters related to the execution control are mapped to the regulation role, and the state clusters related to the energy consumption constraint are mapped to the energy consumption constraint role. S23. Instantiate the multi-agent collaborative structure based on the role mapping table: Assign an independent agent identifier, observation domain, and action domain to each mapped role instance; The communication topology between agents is set as a directed weighted graph, and the communication weights are determined by the correlation of state clusters; Communication latency and bandwidth parameters are registered by the edge computing resource scheduler; S24. Construct a set of feature vectors for each instantiated agent corresponding to its state: Temporal pruning, sampling uniformity, and missing value imputation are performed on the data features within the observation domain of the state cluster. The preprocessed data features are mapped into fixed-dimensional embedding vectors, and a set of feature vectors is formed in the embedding space according to the time series, with timestamps and confidence indices attached. S25. Within the multi-agent collaborative structure, define role-level policy networks for the prediction role, regulation role, and energy consumption constraint role respectively: The policy network for predicting roles takes a set of historical feature vectors as input and outputs a predicted vector for the corresponding future state. The policy network for controlling the role takes the current set of feature vectors and the prediction vectors from the predicted role as input, and outputs a set of candidate control actions; The policy network for the energy-constrained role takes the current energy consumption estimate and the set of control action candidates as input and outputs energy consumption constraint information. S26. Establish local collaboration mechanisms within the role level: Based on the message passing protocol, the feature vector set, prediction vector, control action candidate set and energy consumption constraint information are cyclically exchanged on the communication topology, and the attention fusion operator is used to weight and aggregate the received messages to update the feature vector set of the local state. S27. Perform consistency checks and role dependency corrections on the feature vector set: Abnormal vectors are removed based on preset consistency criteria, and interpolation repair is performed on vectors affected by communication delays or packet loss. The correction results are then written back to the role mapping table for initialization of the next round of strategy search trajectories.

3. The gas separation and control method based on multi-agent cooperation according to claim 1, characterized in that, S4 specifically includes: S41. A cross-interaction module is built outside the co-encoder to realize message passing between agents through a graph neural network. The graph neural network uses agents as nodes and role dependencies as edges. The edge weights are determined by the correlation of state clusters, and embeddings from different branches are integrated during message passing. S42. Perform convergence evaluation on the policy search trajectory within the trajectory buffer: The strategy search trajectory is evaluated in segments according to time order, and the evaluation results are sorted to select strategies with high convergence to form a strategy candidate set. S43. Perform constraint satisfaction verification on the policy candidate set and feed the verification results back to the co-encoder and graph neural network to update the encoder parameters and message passing weights, completing a closed-loop process from feature decoupling to policy generation.

4. The gas separation and control method based on multi-agent cooperation according to claim 1, characterized in that, S5 specifically includes: S51. Extract the corresponding action sequence and state response vector for each candidate strategy in the strategy candidate set, perform simulation prediction on the action sequence according to the role priority constraint, and calculate the gas purity, energy consumption and steady-state deviation index of each action sequence under the current working condition. S52. Establish a constrained optimization framework, set the gas purity constraint as the minimum threshold condition, the energy consumption constraint as the upper limit condition, and the steady state constraint as the deviation tolerance condition. Perform feasibility screening on the action sequences of candidate strategies, and retain only the strategies that meet the three types of constraints to form a set of feasible strategies. S53. Perform multi-objective trade-off analysis on the set of feasible strategies: A weighted comprehensive score is calculated based on gas purity, energy consumption, and steady-state deviation indicators. A multi-objective ranking algorithm is used to prioritize the strategies, and the cooperative dependency and resource constraint effects between strategies are considered during the ranking process. S54. Generate a set of control instructions based on the multi-objective analysis results: The action sequence corresponding to the highest priority strategy is converted into specific control signals, including flow regulation, pressure difference regulation, adsorption stage switching and cooling capacity allocation, along with execution order and time interval information; S55. Serialize the set of control commands to form a parsed command sequence: Based on action dependencies and role collaboration requirements, instructions are sequentially arranged and synchronized. The role affiliation, execution start and end times, and feedback channels of each instruction are recorded so that the multi-agent collaborative structure can update the feature vector set and implement the next round of policy iteration.

5. A gas separation and control system based on multi-agent cooperation, executing the gas separation and control method based on multi-agent cooperation as described in any one of claims 1 to 4, characterized in that, include: The data processing module is used to collect the operating data of the gas separation unit and preprocess it to form a set of operating condition features. The role assignment module is used to construct a multi-agent collaborative structure, assign prediction roles, regulation roles and energy consumption constraint roles to each agent, and generate a set of feature vectors for each role's corresponding state. The feature processing module is used to perform feature decoupling processing, cooperative coding processing and global attention aggregation processing on the feature vector set respectively, driving multiple agents to generate policy search trajectories in the virtual working space; The strategy filtering module is used to construct an interactive role graph structure according to role dependencies, and to perform strategy convergence calculation on the strategy search trajectory, and filter strategies with high convergence to form a strategy candidate set. The strategy optimization module is used to perform constraint optimization on the set of strategy candidates, generate a set of control instructions, and perform multi-objective trade-off analysis to form an analytical instruction sequence; The control and execution module is used to control the gas separation device to perform flow regulation, pressure difference regulation, adsorption stage switching and cooling distribution, and to update the feature vector set through a multi-agent collaborative structure. The strategy iteration module is used to evaluate the performance and calculate the cooperation deviation between roles, adjust the strategy search trajectory and form the next round of control strategy to achieve continuous optimization control of the gas separation device.

Images