Stamping production line self-organizing production system and production method based on twin intelligent agents

Abstract

The invention provides a stamping production line self-organizing production system and method based on a twin intelligent agent, and the twin intelligent agent comprises a data collection layer which is used for collecting the operation state data and production environment parameters of physical equipment; the virtual-real mapping layer is used for constructing a digital twin model of physical equipment and realizing real-time state synchronization and bidirectional control instruction transmission of a physical space and a virtual space; the optimization decision-making layer is used for predicting the performance degradation trend of the equipment based on a deep reinforcement learning algorithm and generating a game parameter adjustment strategy and a preventive maintenance scheme; and the collaborative arbitration layer generates a compromise optimization scheme based on Pareto frontier analysis when the multi-agent strategy conflicts, and the multi-dimensional targets of the production efficiency, the equipment life and the energy consumption are balanced. According to the invention, autonomous task allocation, real-time state monitoring and global resource balance of the stamping production line are realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a self-organizing production system and production method for a stamping production line based on twin intelligent agents. Background Art

[0002] With the deepening implementation of intelligent manufacturing, stamping production lines, as a core component of the manufacturing industry, are facing an increasingly urgent need for automation and intelligent upgrades. However, traditional stamping production lines face challenges such as low production efficiency, slow response times, and high failure rates when faced with complex and ever-changing production tasks. Especially in high-variety, small-batch production models, traditional production lines struggle to quickly adapt to changing production demands, making it difficult to ensure production efficiency and product quality.

[0003] Traditional stamping production lines generally adopt a centralized control architecture, which has problems such as low equipment coordination efficiency, delayed dynamic disturbance response, and insufficient global optimization capabilities. Although existing solutions based on multi-agent systems (MAS) encapsulate equipment and workpieces as intelligent agents, they still have the following shortcomings: 1. Rigid interaction mechanisms: Traditional contract network protocols rely on single-step optimal decision-making, lack a global game perspective, and are prone to falling into local optimality. 2. Insufficient virtual-physical collaboration: The real-time interaction mechanism between physical equipment and digital models is imperfect, making it difficult to support dynamic optimization and predictive maintenance. 3. Unbalanced resource competition: The lack of adaptive negotiation strategies based on game theory between devices leads to uneven load and production bottlenecks. Summary of the Invention

[0004] In response to the above problems, the present invention proposes a self-organizing production system and production method for a stamping production line based on twin intelligent agents, which integrates twin intelligent agents and a multi-agent self-organizing production system of non-cooperative games, and realizes efficient and autonomous operation of the stamping production line through virtual-reality interaction, dynamic game and collaborative optimization mechanism.

[0005] A twin intelligent agent according to the present invention includes: The data collection layer is used to collect the operating status data of physical equipment and production environment parameters; The virtual-reality mapping layer is used to build digital twin models of physical devices, achieving real-time state synchronization and two-way control command transmission between physical and virtual spaces; Optimize the decision-making layer, predict equipment performance degradation trends based on deep reinforcement learning algorithms, and generate game parameter adjustment strategies and preventive maintenance plans; The collaborative arbitration layer generates a compromise optimization solution based on Pareto frontier analysis when multi-agent strategies conflict, balancing the multi-dimensional goals of production efficiency, equipment life, and energy consumption.

[0006] Furthermore, the data acquisition layer includes an OPC UA protocol interface module, an MQTT communication module and an RFID reader / writer; the OPC UA protocol interface module is used to collect pressure sensor data of the stamping machine hydraulic system, temperature data and vibration spectrum data of the servo motor in real time; the MQTT communication module is used to obtain the positioning information of the AGV and the posture data of the end effector of the robotic arm in real time; the RFID reader / writer is used to collect the processing status information bound to the workpiece intelligent body in real time, including the current process completion degree, process parameter deviation and quality inspection results.

[0007] Furthermore, the virtual-reality mapping layer includes a digital thread middleware, a timing database and a three-dimensional kinematic model of the stamping machine; the digital thread middleware is used to map the real-time pressure data of the physical device to the stress cloud map of the virtual model, dynamically displaying the stress distribution in the mold contact area; the timing database is used to store historical data of the equipment to achieve synchronization of virtual and real actions in the stamping cycle; the three-dimensional Unity3D kinematic model of the stamping machine includes the kinematic parameters of the crank-connecting rod mechanism and the deformation simulation module of the mold assembly.

[0008] Furthermore, the optimization decision layer includes a multi-objective optimization model and a PPO algorithm training module; the multi-objective optimization model is used to dynamically adjust the weight coefficient of the game utility function; the PPO algorithm training module is used to train the equipment health prediction model to obtain the predicted value of the remaining service life of the equipment.

[0009] Furthermore, the collaborative arbitration layer includes an NSGA-II algorithm module and a rule engine. When the load balancing difference is greater than a threshold, the NSGA-II algorithm module generates a non-dominated solution set, migrating some processing tasks from high-load stamping machines to low-load equipment to improve production efficiency. When an emergency abnormality occurs, the rule engine freezes the game decision-making process.

[0010] The present invention also proposes a self-organizing production system for a stamping production line based on a twin intelligent agent, comprising: Machine agents, used to detect the status of processing units and execute commands; The workpiece agent uses a state machine model to track the process route, dynamically analyzes process constraints through a semantic description language, and triggers the task bidding process; The scheduling agent initiates bidding and selects the optimal processing unit based on the game strategy; Twin intelligent agents are the twin intelligent agents mentioned above.

[0011] Furthermore, the system realizes dynamic task allocation through a non-cooperative game mechanism. The game mechanism adopts the Nash equilibrium principle. The utility function integrates multi-dimensional indicators such as processing time, load balance and equipment reliability, and the weight coefficient is dynamically adjusted by the twin intelligent agent.

[0012] Furthermore, the multi-agent game decision-making adopts an incomplete information static game model under the non-cooperative game framework. Multiple machine agents make independent decisions based on local information during the bidding stage, and the scheduling agent achieves global optimization through Nash equilibrium solution.

[0013] Furthermore, the incomplete information static game model includes a participant set, a strategy space set and a utility function; The participant set is ,in, represents the i-th machine agent; The strategy space set contains the strategy set of each participant ,in, Indicates submitting a bid that includes processing time and cost parameters. Refusal to participate in the current task bidding; The utility function is constructed based on the equipment's own RUL prediction value, the number of emergency orders, the load condition, and the number of failures this month.

[0014] The present invention also proposes a production method based on a self-organizing production system of a stamping production line based on a twin intelligent agent, comprising the steps of: S1: The workpiece agent analyzes the process matrix, generates a task request package containing machining accuracy and cycle time requirements, and pushes the task requirement data to the scheduling agent through a publish-subscribe model; S2: The scheduling agent calls the process knowledge graph to select a set of machine agents that meet the process capabilities. It performs preliminary screening based on the equipment capability matrix, generates a list of candidate machines, and pushes task information to the candidate machine agents. S3: The candidate machine agent accesses the equipment health prediction interface of the twin agent to obtain the remaining service life prediction value. It then constructs a multidimensional utility function based on its own RUL prediction value, the number of emergency orders, load conditions, and the number of failures this month. Based on the constructed utility function results, it generates a strategy and sends the strategy to the scheduling agent. S4: The scheduling agent constructs a non-cooperative game matrix, where the rows represent the machine agent strategies and the columns represent the task allocation combinations. The Lemke-Howson algorithm is used to solve the mixed strategy Nash equilibrium and obtain the strategy combination that maximizes global utility. The generated initial strategy is pushed to the twin agent in real time through the OPC UC channel. The twin agent performs discrete event simulation on the candidate strategies to verify the feasibility of production capacity. If the simulation result reaches the threshold, the twin agent pushes the simulation result to the scheduling agent, which uses the initial strategy as the final scheduling plan and determines the winning machine agent. If the simulation result does not reach the threshold, the twin agent optimizes the strategy by dynamically adjusting the game parameters until the simulation result reaches the threshold, and generates dynamic parameter recommendations, which are pushed to the scheduling agent in real time through the OPC UC channel. The scheduling agent then generates the final scheduling plan based on the TOPSIS decision algorithm and determines the winning machine agent. S5: The winning machine agent issues processing instructions through the OPC UA protocol and synchronously starts the virtual mirror of the digital twin. The twin agent performs virtual and real data comparison, triggering a three-level abnormal response mechanism: Level 1: microsecond-level data drift → local compensation; Level 2: millisecond-level state deviation → dynamic rescheduling; Level 3: second-level equipment failure → global strategy reconstruction.

[0015] Beneficial effects: Compared with the prior art, the present invention has the following advantages: 1. Improved production efficiency: Through self-organizing production mechanisms and intelligent decision-making optimization, this invention can significantly improve the production efficiency and quality stability of stamping production lines. Interactive negotiation and dynamic adjustment mechanisms between intelligent agents ensure smooth production processes, reducing production interruptions and waste.

[0016] 2. Improved Flexibility: When faced with high-variety, low-volume production tasks, this invention can quickly adjust production parameters and process flows, improving the flexibility and adaptability of the production line. The interactive negotiation mechanism between agents ensures that the production line can maintain efficient operation despite complex and changing production tasks.

[0017] 3. Optimize resource allocation: Through intelligent decision-making and optimization algorithms, this invention can rationally allocate production resources, improve resource utilization and production efficiency. Game-theoretic interaction and negotiation between intelligent agents can ensure that resources are optimally allocated through competition and cooperation.

[0018] 4. Dynamic adaptability: Achieve a balance between resource competition and collaboration through a non-cooperative game mechanism to enhance production line flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 The structure diagram of the twin intelligent body is an embodiment of the present invention.

[0020] Figure 2 This is a diagram of the self-organizing system architecture of a stamping production line based on twin intelligent agents in an embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of the multi-agent collaboration process of an embodiment of the present invention.

[0022] Figure 4 It is a multi-agent game decision-making flowchart of an embodiment of the present invention.

[0023] Figure 5 This is an example diagram of the game tree expansion of an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other.

[0025] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. Example 1

[0026] See also Figures 1 to 5 The present invention proposes a self-organizing production system and method for a stamping production line based on twin intelligent agents. By constructing a collaborative architecture of machine intelligent agents, workpiece intelligent agents, scheduling intelligent agents and twin intelligent agents, and designing an interaction mechanism based on non-cooperative game, combined with virtual and real data fusion and dynamic strategy optimization, the stamping production line can realize autonomous task allocation, real-time status monitoring and global resource balance.

[0027] An embodiment of the present invention provides a twin intelligent agent, which serves as a task allocation center, initiates bidding and selects the optimal processing unit based on game strategy, includes a data acquisition layer, a virtual-reality mapping layer, an optimization decision layer and a collaborative arbitration layer, and supports real-time simulation, abnormal warning and multi-objective strategy recommendation.

[0028] The data acquisition layer collects the operating status data and production environment parameters of physical equipment in real time through the industrial Internet of Things protocol, providing multi-source heterogeneous data support for virtual-reality mapping and optimization decision-making.

[0029] The virtual-reality mapping layer constructs a millimeter-level precision digital twin model of the physical device, realizing real-time state synchronization and two-way control instruction transmission between the physical space and the virtual space.

[0030] The optimization decision layer predicts the equipment performance degradation trend based on the deep reinforcement learning algorithm and generates game parameter adjustment strategies and preventive maintenance plans.

[0031] When multi-agent strategies conflict, the collaborative arbitration layer generates a compromise optimization solution based on Pareto frontier analysis to balance multi-dimensional goals such as production efficiency, equipment life and energy consumption.

[0032] Furthermore, the data acquisition layer includes an OPC UA protocol interface module, an MQTT communication module and an RFID reader / writer; the OPC UA protocol interface module is used to collect pressure sensor data of the stamping machine hydraulic system, temperature data and vibration spectrum data of the servo motor in real time; the MQTT communication module is used to obtain the positioning information of the AGV and the posture data of the end effector of the robotic arm in real time; the RFID reader / writer is used to collect the processing status information bound to the workpiece intelligent body in real time, including the current process completion degree, process parameter deviation and quality inspection results.

[0033] Furthermore, the virtual-reality mapping layer includes a digital thread middleware, a timing database and a three-dimensional kinematic model of the stamping machine; the digital thread middleware is used to map the real-time pressure data of the physical device to the stress cloud map of the virtual model, dynamically displaying the stress distribution in the mold contact area; the timing database is used to store historical data of the equipment to achieve synchronization of virtual and real actions in the stamping cycle; the three-dimensional Unity3D kinematic model of the stamping machine includes the kinematic parameters of the crank-connecting rod mechanism and the deformation simulation module of the mold assembly.

[0034] Furthermore, the optimization decision layer includes a multi-objective optimization model and a PPO algorithm training module; the multi-objective optimization model is used to dynamically adjust the weight coefficient of the game utility function; the PPO algorithm training module is used to train the equipment health prediction model to obtain the predicted value of the remaining service life of the equipment.

[0035] Furthermore, the collaborative arbitration layer includes an NSGA-II algorithm module and a rule engine. When the load balancing difference is greater than a threshold, the NSGA-II algorithm module generates a non-dominated solution set, migrating some processing tasks from high-load stamping machines to low-load equipment to improve production efficiency. When an emergency abnormality occurs, the rule engine freezes the game decision-making process.

[0036] The present invention also proposes a self-organizing production system for a stamping line based on twin agents, comprising a machine agent, a workpiece agent, a scheduling agent, and a twin agent. The system implements dynamic task allocation through a non-cooperative game mechanism and utilizes digital twin technology for virtual-real collaborative optimization.

[0037] The machine agent is bound to physical equipment such as stamping machines, robotic arms, AGVs, etc., and is responsible for equipment status perception (such as pressure, temperature, vibration), action execution and local decision-making.

[0038] The scheduling agent calls the virtual simulation interface of the twin agent during the decision-making process to verify the feasibility of the strategy and optimize the game parameters.

[0039] The workpiece intelligent agent is responsible for managing the process route, processing requirements and real-time status (such as positioning accuracy and processing progress) of the stamping workpiece.

[0040] The twin intelligent agent is deployed on the cloud platform, synchronizes data with physical devices through the OPC UA protocol, and calls the simulation engine for dynamic optimization.

[0041] The game mechanism adopts the Nash equilibrium principle, and the utility function integrates multi-dimensional indicators such as processing time, load balance and equipment reliability. The weight coefficient is dynamically adjusted by the twin intelligent agent.

[0042] The multi-agent game decision-making adopts an incomplete information static game model under the non-cooperative game framework. Each machine agent makes independent decisions based on local information during the bidding phase, and the scheduling agent achieves global optimization by solving the Nash equilibrium. The incomplete information static game model includes three game elements: a participant set, a strategy space set and a utility function.

[0043] Furthermore, the participant set is ,in, represents the i-th machine agent; The strategy space set contains the strategy set of each participant ,in, Indicates submitting a bid that includes processing time and cost parameters. Refusal to participate in the current task bidding; The utility function is constructed based on the equipment's own RUL prediction value, the number of emergency orders, the load condition, and the number of failures this month.

[0044] The present invention also proposes a production method based on a self-organizing production system of a stamping production line based on a twin intelligent agent, comprising the steps of: S1: The workpiece agent analyzes the process matrix, generates a task request package containing machining accuracy and cycle time requirements, and pushes the task requirement data to the scheduling agent through a publish-subscribe model; S2: The scheduling agent calls the process knowledge graph to select a set of machine agents that meet the process capabilities. It performs preliminary screening based on the equipment capability matrix, generates a list of candidate machines, and pushes task information to the candidate machine agents. S3: The candidate machine agent accesses the equipment health prediction interface of the twin agent to obtain the remaining service life prediction value. It then constructs a multidimensional utility function based on its own RUL prediction value, the number of emergency orders, load conditions, and the number of failures this month. Based on the constructed utility function results, it generates a strategy and sends the strategy to the scheduling agent. S4: The scheduling agent constructs a non-cooperative game matrix, where the rows represent the machine agent strategies and the columns represent the task allocation combinations. The Lemke-Howson algorithm is used to solve the mixed strategy Nash equilibrium and obtain the strategy combination that maximizes global utility. The generated initial strategy is pushed to the twin agent in real time through the OPC UC channel. The twin agent performs discrete event simulation on the candidate strategies to verify the feasibility of production capacity. If the simulation result reaches the threshold, the twin agent pushes the simulation result to the scheduling agent, which uses the initial strategy as the final scheduling plan and determines the winning machine agent. If the simulation result does not reach the threshold, the twin agent optimizes the strategy by dynamically adjusting the game parameters until the simulation result reaches the threshold, and generates dynamic parameter recommendations, which are pushed to the scheduling agent in real time through the OPC UC channel. The scheduling agent then generates the final scheduling plan based on the TOPSIS decision algorithm and determines the winning machine agent. S5: The winning machine agent issues processing instructions through the OPC UA protocol and synchronously starts the virtual mirror of the digital twin. The twin agent performs virtual and real data comparison, triggering a three-level abnormal response mechanism: Level 1: microsecond-level data drift → local compensation; Level 2: millisecond-level state deviation → dynamic rescheduling; Level 3: second-level equipment failure → global strategy reconstruction. Example 2

[0045] like Figure 1 As shown, the twin intelligent body includes the following functional modules, namely, data acquisition layer, virtual-reality mapping layer, optimization decision layer, and collaborative arbitration layer.

[0046] like Figure 1 As shown in the figure, the data collection layer deploys an OPC UA protocol interface module, an MQTT communication module, and an RFID reader. The OPC UA protocol interface module collects real-time data from the pressure sensor of the stamping press's hydraulic system, temperature data from the servo motor, and vibration spectrum data. The MQTT communication module obtains real-time positioning information from the AGV and posture data from the robotic arm's end effector. The RFID reader collects real-time processing status information from the workpiece agent, including the completion level of the current process, process parameter deviations, and quality inspection results.

[0047] like Figure 1As shown, the virtual-to-real mapping layer deploys digital thread middleware, a time series database, and a 3D kinematic model of the stamping machine (built on the Unity3D engine). The digital thread middleware maps real-time pressure data from the physical device to a stress contour map of the virtual model, dynamically displaying the stress distribution in the die contact area. The time series database stores historical device data, enabling synchronization of virtual and real-world movements during the stamping cycle. The 3D Unity3D kinematic model of the stamping machine (built on the Unity3D engine) includes the kinematic parameters of the crank-connecting rod mechanism and a deformation simulation module for the die assembly.

[0048] like Figure 1 As shown in Figure 1, the optimization decision layer includes a multi-objective optimization model and a PPO algorithm training module. The multi-objective optimization model dynamically adjusts the weight coefficients of the game utility function. The PPO algorithm training module trains the equipment health prediction model to obtain a predicted value for the equipment's remaining useful life.

[0049] like Figure 1 As shown, the collaborative arbitration layer includes the NSGA-II algorithm module and the rules engine. When load balancing differs significantly, the NSGA-II algorithm module generates a non-dominated solution set, migrating some processing tasks from high-load stamping machines to low-load equipment, improving production efficiency. In the event of an emergency, the rules engine freezes the game decision-making process.

[0050] like Figure 2 As shown, the multi-agent system consists of a machine agent, a workpiece agent, a scheduling agent, and a twin agent. The machine agent is bound to physical equipment such as stamping machines, robotic arms, and AGVs. The perception module acquires real-time equipment status (e.g., idle / busy, queue length). The decision module generates bidding strategies based on local information, and the communication module implements message exchange in accordance with the FIPA-ACL standard. The workpiece agent, acting as the task agent, uses a state machine model to track process routes and dynamically parses process constraints using the semantic description language (OWL-S) to trigger the task bidding process. The scheduling agent, acting as the task allocation hub, initiates bidding and selects the optimal processing unit based on a game-theoretic strategy. The twin agent builds a digital twin model of the physical equipment, integrating sensor data and historical operation and maintenance records to provide virtual-reality mapping, dynamic simulation, and global optimization capabilities.

[0051] like Figure 3 As shown in Figure 1, the multi-agent collaboration process includes five stages of closed-loop control: task triggering, resource matching, game strategy generation, Nash equilibrium decision-making, and execution monitoring.

[0052] like Figure 3As shown in the figure, during the task triggering phase, the workpiece agent first analyzes the process matrix and generates a task request package containing machining accuracy and cycle time requirements. It then pushes the task request data to the scheduling agent through a publish-subscribe model. The data is in JSON format and includes the process sequence and QoS constraints.

[0053] like Figure 3 As shown in the figure, in the resource matching stage, the scheduling agent calls the process knowledge graph to screen the set of machine agents that meet the process capabilities, and then performs preliminary screening based on the equipment capability matrix (stamping tonnage / die compatibility), generates a list of candidate machines, and then pushes the task information to the candidate machine agents.

[0054] like Figure 3 As shown in the figure, in the game strategy generation stage, the candidate machine agent accesses the equipment health prediction interface of the twin agent to obtain the remaining useful life (RUL) prediction value; then it constructs a multidimensional utility function based on its own RUL prediction value, the number of emergency orders, the load situation and the number of failures this month, and then generates a strategy based on the result of the constructed utility function and sends the strategy to the scheduling agent.

[0055] like Figure 3 As shown in the figure, during the Nash equilibrium decision-making phase, the scheduling agent first constructs a non-cooperative game matrix, where rows represent machine agent strategies (bid / no bid) and columns represent task assignment combinations. The Lemke-Howson algorithm is then used to solve the mixed-strategy Nash equilibrium, resulting in a strategy combination that maximizes global utility (the initial strategy). This initial strategy is then pushed to the twin agent in real time via the OPC UC channel. The twin agent first performs discrete event simulation on the candidate strategies to verify capacity feasibility. If the simulation results meet a threshold, the twin agent pushes the results to the scheduling agent, which uses the initial strategy as the final scheduling solution and determines the winning machine agent. If the simulation results fall below the threshold, the twin agent optimizes the strategy by dynamically adjusting the game parameters until the simulation results reach the threshold. It then generates dynamic parameter recommendations and pushes them to the scheduling agent in real time via the OPC UC channel. The scheduling agent then generates a final scheduling solution based on the TOPSIS decision-making algorithm and determines the winning machine agent.

[0056] like Figure 3 As shown in the figure, during the execution and monitoring phase, the winning machine agent issues processing instructions through the OPC UA protocol, synchronously starts the virtual mirror of the digital twin, and the twin agent performs virtual and real data comparison (the deviation threshold is set to 3σ value of sensor accuracy), triggering a three-level abnormal response mechanism: (Level 1: microsecond-level data drift → local compensation; Level 2: millisecond-level state deviation → dynamic rescheduling; Level 3: second-level equipment failure → global strategy reconstruction).

[0057] like Figure 4 As shown in the figure, each machine agent obtains asymmetric information (private information: its own equipment status (remaining service life prediction value, current equipment load rate, etc.); public information: system average load rate, historical failure rate, etc.) through the twin agent, and then substitutes the obtained information as parameters into the utility function to solve it, and then generates a strategy based on the constructed utility function result ( ,in, Indicates submitting a bid that includes processing time and cost parameters. indicates refusal to participate in the current task bidding) and sends the strategy to the scheduling agent.

[0058] like Figure 4 As shown in the figure, the scheduling agent takes the initial state of task allocation as the root node and the strategy selection (S bid or S reject ) is a branch node, and the global utility value corresponding to the final strategy combination is used as the leaf node to construct an extended game tree (such as Figure 5 As shown), the strategy matrix is generated to record all possible strategy combinations. The scheduling agent first constructs the non-cooperative game matrix and finally uses the Lemke-Howson algorithm to solve the mixed strategy Nash equilibrium. The rows represent the machine agent strategy (S bid or S reject ), the columns represent the task allocation combinations, and then the Lemke-Howson algorithm is used to solve the mixed strategy Nash equilibrium to obtain the initial strategy.

[0059] like Figure 4 As shown, the scheduling agent inputs the initial strategy into the discrete event simulation engine of the twin agent to generate the est simulation model .in, Represents a collection of system states (device status, task queues, etc.); Represents a collection of events (task arrival, equipment failure, etc.); represents the state transition function; The twin agent then verifies the simulation results using metrics such as capacity achievement rate, key equipment utilization rate, and the difference between the current equipment load rate and the system average load rate.

[0060] like Figure 4As shown in the figure, if the simulation result reaches the threshold, the twin agent pushes the simulation result to the scheduling agent, which uses the initial strategy as the final scheduling solution and determines the winning machine agent. If the simulation result does not reach the threshold, the twin agent uses the policy gradient algorithm to optimize the game weight coefficients in the game strategy until the simulation result reaches the threshold. After the simulation result reaches the threshold, the twin agent generates dynamic parameter recommendations and pushes the recommendations to the scheduling agent in real time via the OPC UC channel. After receiving the dynamic parameter recommendations from the twin agent, the scheduling agent performs similarity matching using the equipment load rate, order priority, and RUL mean as feature vectors, retrieves the most suitable historical solution from the historical experience library, and then constructs a decision matrix based on the historical solutions, recommended solutions, and equilibrium strategy solutions to select the final scheduling solution using the TOPSIS decision algorithm.

[0061] like Figure 4 As shown, the scheduling agent will generate the final scheduling plan , and then send control instructions to the winning agent through TST. For equipment Start time.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A twin intelligent agent, characterized in that: include: The data collection layer is used to collect the operating status data of physical equipment and production environment parameters; The virtual-reality mapping layer is used to build digital twin models of physical devices, achieving real-time state synchronization and two-way control command transmission between physical and virtual spaces; Optimize the decision-making layer, predict equipment performance degradation trends based on deep reinforcement learning algorithms, and generate game parameter adjustment strategies and preventive maintenance plans; The collaborative arbitration layer generates a compromise optimization solution based on Pareto frontier analysis when multi-agent strategies conflict, balancing the multi-dimensional goals of production efficiency, equipment life, and energy consumption.

2. The twin intelligent agent according to claim 1, characterized in that: The data acquisition layer includes an OPC UA protocol interface module, an MQTT communication module, and an RFID reader / writer; the OPC UA protocol interface module is used to collect pressure sensor data from the stamping machine hydraulic system, temperature data from the servo motor, and vibration spectrum data in real time; the MQTT communication module is used to obtain the positioning information of the AGV and the posture data of the robot arm end effector in real time; the RFID reader / writer is used to collect processing status information bound to the workpiece intelligent body in real time, including the current process completion degree, process parameter deviation, and quality inspection results.

3. The twin intelligent agent according to claim 1, characterized in that: The virtual-reality mapping layer includes digital thread middleware, a timing database and a three-dimensional kinematic model of the stamping machine; the digital thread middleware is used to map the real-time pressure data of the physical device to the stress cloud map of the virtual model, dynamically displaying the stress distribution in the mold contact area; the timing database is used to store historical data of the equipment to achieve synchronization of virtual and real actions in the stamping cycle; the three-dimensional Unity3D kinematic model of the stamping machine includes the kinematic parameters of the crank-connecting rod mechanism and the deformation simulation module of the mold assembly.

4. The twin intelligent agent according to claim 1, characterized in that: The optimization decision layer includes a multi-objective optimization model and a PPO algorithm training module; the multi-objective optimization model is used to dynamically adjust the weight coefficient of the game utility function; the PPO algorithm training module is used to train the equipment health prediction model to obtain the predicted value of the remaining service life of the equipment.

5. The twin intelligent agent according to claim 1, characterized in that: The collaborative arbitration layer includes the NSGA-II algorithm module and a rule engine. When the load balancing difference is greater than a threshold, the NSGA-II algorithm module generates a non-dominated solution set, migrating some processing tasks from high-load stamping machines to low-load equipment to improve production efficiency. When an emergency abnormality occurs, the rule engine freezes the game decision-making process.

6. The self-organizing production system of stamping production line based on twin intelligent agent is characterized by: include: Machine agents, used to detect the status of processing units and execute commands; The workpiece agent uses a state machine model to track the process route, dynamically analyzes process constraints through a semantic description language, and triggers the task bidding process; The scheduling agent initiates bidding and selects the optimal processing unit based on the game strategy; A twin intelligent agent is a twin intelligent agent as described in any one of claims 1 to 5.

7. The self-organizing production system for a stamping production line based on a twin intelligent agent according to claim 6 is characterized in that: The system realizes dynamic task allocation through a non-cooperative game mechanism. The game mechanism adopts the Nash equilibrium principle. The utility function integrates multi-dimensional indicators such as processing time, load balance and equipment reliability, and the weight coefficient is dynamically adjusted by the twin intelligent agent.

8. The self-organizing production system for a stamping production line based on a twin intelligent agent according to claim 7 is characterized in that: The multi-agent game decision-making adopts an incomplete information static game model under the non-cooperative game framework. Multiple machine agents make independent decisions based on local information during the bidding stage, and the scheduling agent achieves global optimization by solving the Nash equilibrium.

9. The self-organizing production system for a stamping production line based on a twin agent according to claim 8 is characterized in that: The incomplete information static game model includes a participant set, a strategy space set and a utility function; The participant set is ,in, represents the i-th machine agent; The strategy space set contains the strategy set of each participant ,in, Indicates submitting a bid that includes processing time and cost parameters. Refusal to participate in the current task bidding; The utility function is constructed based on the equipment's own RUL prediction value, the number of emergency orders, the load condition, and the number of failures this month.

10. The production method of the self-organizing production system of a stamping production line based on a twin intelligent agent according to claim 6 is characterized in that: Including steps: S1: The workpiece agent analyzes the process matrix, generates a task request package containing machining accuracy and cycle time requirements, and pushes the task requirement data to the scheduling agent through a publish-subscribe model; S2: The scheduling agent calls the process knowledge graph to select a set of machine agents that meet the process capabilities. It performs preliminary screening based on the equipment capability matrix, generates a list of candidate machines, and pushes task information to the candidate machine agents. S3: The candidate machine agent accesses the equipment health prediction interface of the twin agent to obtain the remaining service life prediction value. It then constructs a multidimensional utility function based on its own RUL prediction value, the number of emergency orders, load conditions, and the number of failures this month. Based on the constructed utility function results, it generates a strategy and sends the strategy to the scheduling agent. S4: The scheduling agent constructs a non-cooperative game matrix, where the rows represent the machine agent strategies and the columns represent the task allocation combinations. The Lemke-Howson algorithm is used to solve the mixed strategy Nash equilibrium and obtain the strategy combination that maximizes global utility. The generated initial strategy is pushed to the twin agent in real time through the OPC UC channel. The twin agent performs discrete event simulation on the candidate strategies to verify the feasibility of production capacity. If the simulation result reaches the threshold, the twin agent pushes the simulation result to the scheduling agent, which uses the initial strategy as the final scheduling plan and determines the winning machine agent. If the simulation result does not reach the threshold, the twin agent optimizes the strategy by dynamically adjusting the game parameters until the simulation result reaches the threshold, and generates dynamic parameter recommendations, which are pushed to the scheduling agent in real time through the OPC UC channel. The scheduling agent then generates the final scheduling plan based on the TOPSIS decision algorithm and determines the winning machine agent. S5: The winning machine agent issues processing instructions through the OPC UA protocol and synchronously starts the virtual mirror of the digital twin. The twin agent performs virtual and real data comparison, triggering a three-level abnormal response mechanism: Level 1: microsecond-level data drift → local compensation; Level 2: millisecond-level state deviation → dynamic rescheduling; Level 3: second-level equipment failure → global strategy reconstruction.

Images