Intelligent real-time production line regulation and control system and method based on digital twinning

Abstract

The invention relates to the technical field of intelligent manufacturing and industrial internet of things, in particular to an intelligent real-time production line regulation and control system and method based on digital twinning, and the system comprises a state sensing and collection module which is used for obtaining the real-time operation parameters of production equipment, material flow state information and environment disturbance characteristics of a production line; the digital twin modeling module is used for fusing multi-source data, constructing accurate mapping of a production line physical entity and a virtual model through a twin model driven by dynamic data, and predicting a performance degradation trend and a disturbance propagation path of production line equipment in real time; the intelligent decision-making module is used for generating a production line real-time scheduling strategy through a distributed collaborative reinforcement learning algorithm and outputting a production line regulation and control instruction set; and the real-time regulation and control feedback module executes production line regulation and control actions through a multi-target dynamic balance algorithm and automatically generates a production line production optimization report with equipment health degree compensation. Therefore, the problems of static model construction, lagging of a decision mechanism, lack of regulation and control feedback and the like in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial Internet of Things (IoT) technology, specifically to an intelligent real-time control system and method for production lines based on digital twins. Background Technology

[0002] In traditional manufacturing systems, production line scheduling and control largely rely on pre-set fixed strategies or real-time intervention based on human experience. During operation, production equipment typically schedules and executes production based on pre-set production cycles, process sequences, and equipment capacity, or is manually adjusted by on-site operators according to real-time conditions. These methods can maintain basic operation in static or predictable production environments: on the one hand, fixed strategies simplify system complexity and reduce the burden of real-time computation; on the other hand, human intervention can flexibly respond to unexpected situations, ensuring the basic operational goals of uninterrupted production and preventing backlogs. This is a crucial technical means supporting production activities in discrete manufacturing, process industries, and other fields.

[0003] In related technologies, CN119717747B discloses a method and system for intelligent scheduling of production lines. This system relies solely on historical production data to construct a static scheduling model, lacking a dynamic data-driven digital twin mapping mechanism. It cannot synchronize the physical state of the production line with the virtual simulation process in real time, making it difficult to accurately capture equipment performance degradation trends and environmental disturbance propagation paths. It generates single-dimensional production scheduling instructions only through a centralized rule engine, without introducing distributed collaborative reinforcement learning algorithms to optimize multi-device collaboration strategies, nor establishing a multi-objective trade-off mechanism for efficiency, quality, and cost. When faced with concurrent disturbances at multiple workstations, the decision-making response delay exceeds 30 seconds, making it difficult to dynamically optimize the allocation of production line resources. Most importantly, it lacks an equipment health compensation mechanism and a closed-loop control feedback module, only outputting basic scheduling instructions without real-time monitoring data and production optimization reports. It cannot assess changes in equipment performance and product quality fluctuations after control execution, leading to a disconnect between the scheduling strategy and actual production line optimization needs, making it difficult to effectively cope with complex situations such as equipment health degradation and dynamic disturbances in the production line.

[0004] While existing production line scheduling and control systems have been widely applied to improve production efficiency and ensure production continuity, the current technological system still has significant limitations: Existing data collection is mostly focused on basic operating parameters such as equipment start / stop status and output counts, lacking real-time equipment health status, accurate material flow tracking, and synchronous perception of environmental disturbances. This makes it difficult to construct a dynamic profile of all elements of the production line, and the data dimensions are too limited to comprehensively assess the synergistic relationship between equipment efficiency, material availability, and production cycle time. Model construction is static, relying heavily on historical data fitting or offline simulation models, lacking a digital twin-driven virtual-real mapping and real-time correction mechanism, making it impossible to accurately predict the propagation effect of disturbances and the effectiveness of dynamic adjustment strategies. The decision-making mechanism is lagging, primarily relying on centralized scheduling or rule engines, lacking distributed collaborative perception and autonomous decision-making capabilities, and failing to incorporate dynamic optimization algorithms for real-time rescheduling. This results in high decision-making delays and poor adaptability of control strategies in the face of multiple concurrent disturbances. Control feedback is lacking; during execution, there is a lack of real-time monitoring and compensation for factors such as equipment performance degradation, material delivery delays, and quality fluctuations, making it difficult to form a closed loop of perception-decision-execution-evaluation, resulting in insufficient guarantees for the overall efficiency and robustness of the production line. These deficiencies restrict the production line's response speed, resource utilization, and long-term stable operation in complex and ever-changing environments, necessitating the introduction of innovative technological solutions to overcome existing bottlenecks. Summary of the Invention

[0005] This application provides a digital twin-based intelligent real-time control system and method for production lines to solve problems such as static model construction, lagging decision-making mechanisms, and lack of control feedback in existing technologies.

[0006] The first aspect of this application provides an intelligent real-time control system for a production line based on digital twins, comprising: a state perception and acquisition module, a digital twin modeling module, an intelligent decision-making module, and a real-time control feedback module; wherein, the state perception and acquisition module is used to acquire real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics of the production line; the digital twin modeling module is used to fuse multi-source data, construct an accurate mapping between the physical entity and virtual model of the production line through a dynamic data-driven twin model, and predict the performance degradation trend and disturbance propagation path of the production line equipment in real time; the intelligent decision-making module is used to generate a real-time scheduling strategy for the production line through a distributed collaborative reinforcement learning algorithm, and output a production line control instruction set by combining multi-objective optimization reasoning with knowledge graphs; the real-time control feedback module, based on the production line control instruction set, executes production line control actions through a multi-objective dynamic trade-off algorithm, monitors the equipment efficiency and production quality of the production line in real time, and automatically generates a production line optimization report with equipment health compensation.

[0007] Preferably, the status perception and acquisition module includes an equipment operation parameter acquisition unit, a material flow status tracking unit, and an environmental disturbance perception unit. The equipment operation parameter acquisition unit is used to acquire real-time operating parameters of the production equipment through an IoT interface and sensor network, recording key indicators such as equipment vibration, temperature, power consumption, and operating cycle. The material flow status tracking unit is used to track the location, quantity, and status information of materials in the production line in real time. The environmental disturbance perception unit is used to monitor environmental disturbance characteristics through temperature, humidity, vibration, and electromagnetic interference sensors, identify external interference sources, and quantify their impact on production line stability.

[0008] Preferably, the digital twin modeling module includes a multi-source data fusion unit, a dynamic mapping construction unit, and a performance prediction unit. The multi-source data fusion unit cleans, aligns, and fuses the data using spatiotemporal alignment and feature extraction algorithms to eliminate data heterogeneity and noise interference. The dynamic mapping construction unit, based on the fused data, constructs a high-fidelity dynamic mapping between the physical entity of the production line and the virtual model, synchronizing equipment behavior, material flow, and environmental disturbances between the virtual and real worlds. The performance prediction unit analyzes the relationship between equipment performance indicators and disturbance propagation in the dynamic mapping to predict equipment performance degradation trends and the impact of disturbance events on production cycle time and quality.

[0009] Preferably, the intelligent decision-making module includes a distributed collaborative reinforcement learning unit and a multi-objective optimization reasoning unit. The distributed collaborative reinforcement learning unit trains and optimizes multi-agent policies through a distributed collaborative reinforcement learning algorithm to generate a real-time distributed scheduling strategy. The multi-objective optimization reasoning unit performs multi-objective trade-off analysis through a knowledge graph multi-objective optimization reasoning algorithm and, in conjunction with the distributed scheduling strategy, outputs a set of control instructions that satisfy multiple objectives such as efficiency, quality, and cost.

[0010] Preferably, the distributed collaborative reinforcement learning algorithm formula is as follows: ; in, For intelligent agents The optimal action value function; State; For joint action; For intelligent agents The reward received immediately after performing the action; Discount factor; The total number of agents; Indexing for intelligent agents; For collaborative weighting coefficients; For intelligent agents optimal value; The next state; For the next joint action.

[0011] Preferably, the real-time control feedback module includes a dynamic trade-off execution unit and an optimization report generation unit. The dynamic trade-off execution unit uses a multi-objective dynamic trade-off algorithm to perform real-time trade-offs and decision-making on key indicators of production efficiency, equipment load, and product quality during the control process. The optimization report generation unit collects and stores production line status data by real-time monitoring of equipment comprehensive efficiency and production quality indicators, and combines the equipment health compensation coefficient to perform correlation analysis on the monitoring data and control results, automatically generating an optimization report containing suggestions for adjusting production parameters.

[0012] The second aspect of this application provides a method for intelligent real-time control of a production line based on digital twins, comprising: acquiring real-time operating parameters of the production equipment, material flow status information, and environmental disturbance characteristics of the production line; fusing the real-time operating parameters of the production equipment, material flow status information, and environmental disturbance characteristics of the production line, constructing a precise mapping between the physical entity and the virtual model of the production line through a dynamic data-driven twin model, predicting the performance degradation trend of the production line equipment and the disturbance propagation path in real time, and obtaining prediction results; based on the prediction results, generating a real-time scheduling strategy for the production line through a distributed collaborative reinforcement learning algorithm, combining multi-objective optimization reasoning with a knowledge graph, and outputting a production line control instruction set; based on the production line control instruction set, executing production line control actions through a multi-objective dynamic trade-off algorithm, monitoring the equipment efficiency and production quality of the production line in real time, collecting and storing production line status data, and automatically generating a production line production optimization report with equipment health compensation.

[0013] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the intelligent real-time control method for a production line based on digital twins as described in the above embodiments.

[0014] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the intelligent real-time control method for a production line based on digital twins as described in the above embodiments.

[0015] The fifth aspect of this application provides a computer program product, including a computer program or instructions, for implementing the intelligent real-time control method for a production line based on digital twins as described in the above embodiments.

[0016] Therefore, this application has the following beneficial effects:

[0017] This application embodiment acquires real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics through a state perception and acquisition module. This comprehensively perceives the production line's operating status and external disturbance factors, providing full-element data support for subsequent analysis and decision-making. Combined with a digital twin modeling module, the dynamic data-driven model integrates multi-source data to construct a precise mapping between the physical entity and virtual model of the production line. This allows for real-time prediction of equipment performance degradation trends and disturbance propagation paths, enhancing the depth and foresight of production line state perception. The intelligent decision-making module generates real-time scheduling strategies using a distributed collaborative reinforcement learning algorithm. Combined with multi-objective optimization reasoning based on knowledge graphs, it outputs a set of control instructions that meet multiple objective requirements, enhancing the synergy and adaptability of scheduling decisions under complex disturbance environments. The real-time control feedback module executes control actions through a multi-objective dynamic trade-off algorithm, monitors equipment efficiency and production quality in real time, automatically generates production optimization reports with equipment health compensation, and continuously optimizes production line performance and dynamically maintains equipment health, improving the system's robustness and overall efficiency in dynamic environments. This solves the problems of static model construction, lagging decision-making mechanisms, and lack of control feedback in existing technologies.

[0018] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0020] Figure 1 This is a schematic diagram of the structure of a digital twin-based intelligent real-time control system for production lines provided in accordance with an embodiment of this application;

[0021] Figure 2 This is a schematic diagram of a state-aware acquisition module according to an embodiment of this application;

[0022] Figure 3 This is a schematic diagram of a digital twin modeling module provided according to an embodiment of this application;

[0023] Figure 4 This is a schematic diagram of an intelligent decision-making module provided according to an embodiment of this application;

[0024] Figure 5 This is a schematic diagram of a real-time control feedback module provided according to an embodiment of this application;

[0025] Figure 6 A flowchart of a digital twin-based intelligent real-time control method for production lines according to an embodiment of this application;

[0026] Figure 7This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] The following description, with reference to the accompanying drawings, describes an embodiment of the intelligent real-time control system and method for production lines based on digital twins. To address the high detection error rate mentioned in the background technology, this application provides a digital twin-based intelligent real-time control system for production lines. In this system, a state perception and acquisition module acquires real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics, comprehensively perceiving the production line's operating status and external disturbance factors, providing comprehensive data support for subsequent analysis and decision-making. Combined with a dynamic data-driven model from a digital twin modeling module, multi-source data is integrated to construct a precise mapping between the physical entity and virtual model of the production line, predicting equipment performance degradation trends and disturbance propagation paths in real time, thus enhancing the depth and foresight of production line state perception. A distributed collaborative reinforcement learning algorithm from an intelligent decision-making module generates real-time scheduling strategies, combined with multi-objective optimization reasoning using knowledge graphs, outputting a set of control instructions that meet multiple objective requirements, enhancing the synergy and adaptability of scheduling decisions under complex disturbance environments. A real-time control feedback module executes control actions through a multi-objective dynamic trade-off algorithm, monitoring equipment efficiency and production quality in real time, automatically generating production optimization reports with equipment health compensation, continuously optimizing production line performance and dynamically maintaining equipment health, thereby improving the system's robustness and overall efficiency in dynamic environments. This solves the problems of static model building, lagging decision-making mechanisms, and lack of regulatory feedback in existing technologies.

[0029] Figure 1 A schematic diagram of the structure of a digital twin-based intelligent real-time control system for production lines provided in this application embodiment.

[0030] This application provides an intelligent real-time control system for a production line based on digital twins. The system 10 includes:

[0031] The system includes a state perception and acquisition module 100, a digital twin modeling module 200, an intelligent decision-making module 300, and a real-time control and feedback module 400.

[0032] Among them, the state perception and acquisition module 100 is used to acquire real-time operating parameters of production equipment, material flow status information and environmental disturbance characteristics of the production line; the digital twin modeling module 200 is used to integrate multi-source data, construct a precise mapping between the physical entity and virtual model of the production line through a dynamic data-driven twin model, and predict the performance degradation trend and disturbance propagation path of the production line equipment in real time; the intelligent decision-making module 300 is used to generate real-time scheduling strategies for the production line through a distributed collaborative reinforcement learning algorithm, and output the production line control instruction set by combining multi-objective optimization reasoning with knowledge graph; the real-time control feedback module 400 executes production line control actions based on the production line control instruction set through a multi-objective dynamic trade-off algorithm, monitors the equipment efficiency and production quality of the production line in real time, and automatically generates a production line production optimization report with equipment health compensation.

[0033] It is understood that in this embodiment, the state perception acquisition module acquires real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics to comprehensively perceive the production line's operating status and external disturbance factors, providing full-element data support for subsequent analysis and decision-making. Combined with the dynamic data-driven model of the digital twin modeling module, multi-source data is integrated to construct a precise mapping between the physical entity and virtual model of the production line, predicting equipment performance degradation trends and disturbance propagation paths in real time, thus enhancing the depth and foresight of production line state perception. The distributed collaborative reinforcement learning algorithm of the intelligent decision-making module generates real-time scheduling strategies, combined with multi-objective optimization reasoning using knowledge graphs, outputting a set of control instructions that meet multi-objective needs, enhancing the synergy and adaptability of scheduling decisions under complex disturbance environments. The real-time control feedback module executes control actions through a multi-objective dynamic trade-off algorithm, monitors equipment efficiency and production quality in real time, automatically generates production optimization reports with equipment health compensation, continuously optimizes production line performance, and dynamically maintains equipment health, improving the system's robustness and overall efficiency in dynamic environments. This solves the problems of static model construction, lagging decision-making mechanisms, and lack of control feedback in existing technologies.

[0034] In this embodiment of the application, the state awareness acquisition module 100 includes: as follows Figure 2 As shown, there are equipment operation parameter acquisition unit, material flow status tracking unit, and environmental disturbance sensing unit.

[0035] The equipment operation parameter acquisition unit is used to acquire real-time operating parameters of production equipment through the equipment IoT interface and sensor network, and record key indicators such as equipment vibration, temperature, power consumption and operating cycle; the material flow status tracking unit is used to track the position, quantity and status information of materials in the production line in real time; the environmental disturbance sensing unit is used to monitor the disturbance characteristics of the production environment through temperature, humidity, vibration and electromagnetic interference sensors, identify external interference sources and quantify their impact on the stability of the production line.

[0036] It is understood that the embodiments of this application acquire real-time operating parameters of production equipment through the equipment operation parameter acquisition unit via the equipment IoT interface and sensor network, recording key indicators such as equipment vibration, temperature, power consumption, and operating cycle, effectively capturing core status information during equipment operation, and providing real-time data support for equipment health diagnosis and fault early warning; the material flow status tracking unit tracks the location, quantity, and status information of materials in the production line in real time, clearly presenting the dynamics of the entire material flow process, avoiding material stagnation, loss, or mismatch, and ensuring the continuity and efficiency of the production process; the environmental disturbance perception unit relies on temperature, humidity, vibration, and electromagnetic interference sensors to monitor the disturbance characteristics of the production environment, identify external interference sources and quantify their impact on production line stability, promptly discover the potential impact of environmental factors on production, and provide decision-making basis for environmental optimization and production line anti-interference capability improvement; improve the real-time performance of equipment status monitoring, the accuracy of material flow tracking, and the scientific nature of environmental interference assessment, and provide a comprehensive and reliable raw data foundation for subsequent production process optimization, production line stability assurance, and anomaly troubleshooting through multi-dimensional data acquisition, reducing the probability of missed equipment anomaly detection, chaotic material management, and untimely handling of environmental interference.

[0037] In this embodiment of the application, the digital twin modeling module 200 includes: as follows Figure 3 As shown, there are a multi-source data fusion unit, a dynamic mapping construction unit, and a performance prediction unit.

[0038] The multi-source data fusion unit cleans, aligns, and fuses data using spatiotemporal alignment and feature extraction algorithms to eliminate data heterogeneity and noise interference. The dynamic mapping construction unit builds a high-fidelity dynamic mapping between the physical entity and the virtual model of the production line based on the fused data, and synchronizes the virtual and real data of equipment behavior, material flow, and environmental disturbances. The performance prediction unit predicts the trend of equipment performance degradation and the impact of disturbance events on production cycle time and quality by analyzing the relationship between equipment performance indicators and disturbance propagation in the dynamic mapping.

[0039] It is understood that the multi-source data fusion unit in this application uses spatiotemporal alignment and feature extraction algorithms to clean, align, and fuse data, effectively eliminating data heterogeneity and noise interference to ensure data validity. The dynamic mapping construction unit relies on the fused data to construct a high-fidelity dynamic mapping between the physical entity and the virtual model of the production line, synchronizing equipment behavior, material flow, and environmental disturbances in the virtual and real worlds to provide accurate mapping support for the collaborative management and control of the virtual and real worlds of the production line. The performance prediction unit analyzes the relationship between equipment performance indicators and disturbance propagation in the dynamic mapping to predict the trend of equipment performance degradation and the impact of disturbance events on production cycle and quality, improving the accuracy of equipment performance prediction, the comprehensiveness of disturbance impact assessment, and the foresight of production management and control. Through the output of high-fidelity dynamic mapping and performance prediction results, it provides a reliable analytical basis for production line equipment management, production cycle control, and quality risk prevention and control, reducing data heterogeneity interference, virtual-real synchronization deviation, and the probability of performance prediction errors. These three units effectively solve the problem of static model construction.

[0040] For example, in the final assembly workshop of an automobile manufacturing company, its digital twin modeling module uses a multi-source data fusion unit to collect real-time data from 32 types of sensors (frequency up to 100Hz) including equipment vibration, temperature, and pressure, as well as order information from the MES system. It employs a spatiotemporal alignment algorithm to eliminate data heterogeneity and combines feature extraction technology to filter out 99.7% of noise interference. The dynamic mapping construction unit, based on the Unity engine and ANSYS TwinBuilder, builds a 1:1 virtual production line model. Through the OPCUA protocol, it synchronizes with physical equipment in real time, achieving millisecond-level virtual-real mapping of welding robot posture, AGV material delivery paths, and environmental temperature and humidity, with a positioning accuracy of ±2mm. The performance prediction unit uses an LSTM neural network to analyze the relationship between equipment OEE indicators and disturbance propagation in the dynamic mapping, predicting the probability of welding robot motor failure 72 hours in advance (accuracy 89.3%), and simulating the impact of material shortages on production cycle time, automatically generating alternative delivery plans, reducing downtime losses by 42%. After the system went online, the changeover time in the final assembly workshop was reduced by 40%, the overall equipment efficiency (OEE) was improved by 28%, and unplanned downtime was reduced by 320 hours per year.

[0041] In this embodiment of the application, the intelligent decision-making module 300 includes: Figure 4 As shown, there are distributed collaborative reinforcement learning units and multi-objective optimization inference units.

[0042] Among them, the distributed cooperative reinforcement learning unit trains and optimizes multi-agent policies through distributed cooperative reinforcement learning algorithms to generate real-time distributed scheduling policies; the multi-objective optimization reasoning unit performs multi-objective trade-off analysis through knowledge graph multi-objective optimization reasoning algorithms, and outputs a set of control instructions that meet multiple objectives such as efficiency, quality, and cost in combination with the distributed scheduling policy.

[0043] It is understood that the embodiments of this application employ a distributed collaborative reinforcement learning unit to conduct distributed training and optimization of multi-agent policies using a distributed collaborative reinforcement learning algorithm, effectively performing iterative upgrades and precise optimization of multi-agent policies to generate real-time distributed scheduling strategies. The multi-objective optimization reasoning unit relies on a knowledge graph multi-objective optimization reasoning algorithm to conduct multi-objective trade-off analysis, and integrates multi-objective trade-off logic with scheduling strategy information in conjunction with the distributed scheduling strategy to provide decision-making basis for the output of the control instruction set, outputting a control instruction set that satisfies multiple objectives such as efficiency, quality, and cost. Through distributed training optimization of multi-agent policies and collaborative analysis of multi-objective trade-offs, the system improves the distributed collaboration of policy training, the comprehensiveness of multi-objective analysis, and the adaptability of the output of the control instruction set. The structured output of the control instruction set provides reliable control support for relevant scenarios, reducing policy optimization deviations and the probability of misjudgment of control instructions. These two units effectively address the problem of lagging decision-making mechanisms.

[0044] For example, taking the intelligent scheduling and process control scenario of an automotive engine block machining production line as an example: In the intelligent decision-making module, the distributed collaborative reinforcement learning unit trains a multi-device collaboration strategy of "machine tool-AGV-conveyor belt" based on historical machining cycle time, real-time load of 12 CNC machine tools, material delivery trajectory of 6 AGVs, and process connection data of 3 conveyor belts through a distributed collaborative reinforcement learning algorithm. This strategy includes dynamically optimizing the cycle time matching of roughing and finishing processes (avoiding machine tool waiting for materials or conveyor belt congestion) and adjusting the material delivery priority according to the real-time position of the AGVs (such as prioritizing the delivery of cylinder block blanks to machine tools with lower loads). This generates a real-time production scheduling strategy: when the load of a certain finishing machine tool exceeds 85%, the subsequent 3 batches of orders are automatically diverted to adjacent machine tools of the same model, and the AGVs are coordinated to deliver the corresponding tools 10 minutes in advance. Subsequently, the multi-objective optimization reasoning unit calls upon the production line knowledge graph (including related information such as tool wear curves, cylinder hole accuracy tolerance standards, customer order delivery penalty coefficients, and equipment failure history) to weigh the three objectives of "shortening the production cycle (efficiency), ensuring cylinder machining accuracy (quality), and reducing tool wear costs (cost)". If it is detected that the defect rate of a certain process rises to 1.2% due to tool wear (exceeding the target threshold of 0.5%), the system will combine the information in the knowledge graph that "the remaining life of the tool is only 2 hours" and "there is a tool of the same specification in the spare tool library and the AGV is idle" to adjust the strategy: suspend the current machine tool operation, dispatch the AGV to retrieve the spare tool for replacement, and postpone the subsequent orders of the machine tool by 15 minutes to prioritize the operation of other processes at the original cycle time. Finally, it outputs a set of control instructions including "machine tool replacement instructions", "AGV material delivery adjustment list" and "temporary correction parameters for process cycle time". Through this process, the average production cycle of the production line was shortened by 28%, the defect rate of cylinder block processing was reduced from 0.9% to 0.4%, the tool wear cost was reduced by 18%, and the scheduling response time when dealing with emergency order insertions was reduced from 12 minutes to 3 minutes, achieving flexible and efficient production under multiple objective constraints.

[0045] In this embodiment of the application, the formula for the distributed cooperative reinforcement learning algorithm is as follows: ; in, For intelligent agents The optimal action value function; State; For joint action; For intelligent agents The reward received immediately after performing the action; Discount factor; The total number of agents; Indexing for intelligent agents; For collaborative weighting coefficients; For intelligent agents optimal value; The next state; For the next joint action.

[0046] It is understood that the embodiments of this application protect the data privacy of each distributed node and avoid the risks of centralization by leveraging the characteristics of distributed learning through distributed collaborative reinforcement learning algorithms. They also improve training efficiency by relying on collaborative knowledge sharing to solve the problem of data sparsity or heterogeneity in some nodes and combining it with reinforcement learning mechanisms for dynamic policy optimization. At the same time, they ensure the generalization ability of the policy to the differences of multiple nodes through collaborative aggregation and adaptation, and efficiently complete the collaborative training and optimization of distributed policies to generate a distributed scheduling policy that is "globally adaptable, locally usable and dynamically adjustable". This policy not only provides high-quality dynamic input for multi-objective optimization reasoning of knowledge graphs, but also lays a key foundation for accurately outputting production line control instruction sets by combining it with temporal causal analysis.

[0047] For example, in a scenario of intelligent control of an automotive parts production line based on digital twins, assuming the total number of intelligent agents N=3 (corresponding to the control agents for stamping, welding, and painting equipment respectively), let's take the stamping equipment agent (i.e., i=1) as an example, its optimal action value function... This is used to measure the long-term optimal expected return under a specific state and action; at this time, the production line state s is "load rate 60%, queue length of parts to be processed 50", and the joint action a is "stamping equipment speed up 20%, welding equipment maintains speed, and painting equipment enters inspection mode". The immediate reward of this agent is... =8 Future Reward Discount Factor =0.9. During the collaborative process, the referenced agent indices j are 2 (welding agent) and 3 (painting agent), respectively, with corresponding collaborative weights. =0.4、 =0.3, and the two agents in subsequent states ("Load rate 70%, length of waiting parts queue 40") and subsequent joint actions The optimal Q values ​​under the conditions of "stamping equipment maintaining speed, welding equipment increasing speed by 15%, and painting equipment maintaining speed" are respectively =12、 =9, substituting these values ​​into the formula, we can finally obtain the stamping agent's value in this scenario. =14.75, fully demonstrating the practical application logic of multi-agent collaborative reinforcement learning in the dynamic control of production lines.

[0048] In this embodiment of the application, the real-time control feedback module 400 includes, as follows: Figure 5 As shown, there are dynamic trade-off execution unit and optimization report generation unit.

[0049] The dynamic trade-off execution unit uses a multi-objective dynamic trade-off algorithm to weigh and make decisions on key indicators of production efficiency, equipment load, and product quality in real time during the control process. The optimization report generation unit collects and stores production line status data by monitoring the comprehensive efficiency and production quality indicators of equipment in real time, and combines the equipment health compensation coefficient to perform correlation analysis on the monitoring data and control results, and automatically generates an optimization report containing suggestions for adjusting production parameters.

[0050] It should be noted that the equipment health compensation coefficient is a dynamically calculated normalized parameter, typically ranging from [0.8, 1.5], used to quantify the impact of equipment health status on production scheduling and maintenance strategies. When the coefficient is <1.0, it indicates good equipment health, and the maintenance cycle can be appropriately extended or the load increased; when the coefficient is >1.0, it indicates potential risks to the equipment, requiring earlier maintenance or reduced load operation.

[0051] It is understood that the embodiments of this application employ a multi-objective dynamic trade-off algorithm through a dynamic trade-off execution unit to conduct real-time trade-offs and decision-making execution on key indicators of production efficiency, equipment load, and product quality during the control process. This effectively achieves dynamic balance and decision implementation for each key control indicator, ensuring the scientific and efficient nature of the production control process. The optimization report generation unit relies on real-time monitoring of equipment comprehensive efficiency and production quality indicators, collects and stores production line status data, and combines the equipment health compensation coefficient to conduct correlation analysis on monitoring data and control results. It automatically generates an optimization report containing suggestions for production parameter adjustments, providing targeted directions for production line parameter optimization, while also achieving effective retention and in-depth analysis of production line operation data. This improves the real-time nature of control decisions, the accuracy of data correlation analysis, and the practicality of optimization suggestions. Through the output of structured optimization reports, it provides a reliable reference for research on continuous improvement of the production process and equipment efficiency enhancement, reducing the probability of indicator trade-off bias, data application disconnect, and lack of optimization guidance. These two units effectively solve the problem of missing control feedback.

[0052] For example, in the welding workshop of an automobile manufacturing company, the real-time control and feedback module has achieved intelligent upgrades to the production process through a dynamic balance execution unit and an optimization report generation unit. When the production line is producing new energy vehicles in mixed-flow mode, the dynamic balance execution unit detects that the welding robot load rate has reached 92% (exceeding the threshold of 85%) and the weld defect rate has risen to 1.2% (exceeding the standard of 0.8%). It immediately triggers a multi-objective dynamic balance algorithm: within 0.3 seconds, the welding speed is reduced by 18% to alleviate equipment pressure, while the current parameter is increased by 5% and the electrode cap contact angle is adjusted, so that the defect rate drops back to 0.6%, the equipment load is stabilized at 82%, and a dynamic balance between production efficiency and product quality is achieved. Meanwhile, the optimization report generation unit collects 28 indicators in real time, including equipment vibration spectrum (10kHz sampling rate) and OEE (Overall Equipment Effectiveness). Combined with an equipment health model trained on historical data (containing 12-dimensional features such as vibration and temperature), it generates "Health Assessment and Parameter Optimization Recommendations for Robot at Workstation No. 32." The recommendations suggest shortening the robot's maintenance cycle from 300 hours to 200 hours and adjusting the acceleration curve in the welding path (from 0.8g to 0.6g). After optimization, the OEE of this workstation increased from 78% to 89.5%, unplanned downtime decreased by 40%, and the number of welding defects per vehicle body decreased by 35%. This example verifies the core value of the real-time control feedback module in achieving precise decision-making through data closed-loop in complex manufacturing scenarios.

[0053] The intelligent real-time control system for production lines based on digital twins proposed in this application acquires real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics through a state perception and acquisition module. This comprehensively perceives the production line's operating status and external disturbance factors, providing full-element data support for subsequent analysis and decision-making. Combined with a dynamic data-driven model from a digital twin modeling module, it integrates multi-source data to construct a precise mapping between the physical entity and virtual model of the production line. This allows for real-time prediction of equipment performance degradation trends and disturbance propagation paths, enhancing the depth and foresight of production line state perception. The intelligent decision-making module generates real-time scheduling strategies using a distributed collaborative reinforcement learning algorithm. Combined with multi-objective optimization reasoning based on knowledge graphs, it outputs a set of control instructions that meet multiple objective requirements, enhancing the synergy and adaptability of scheduling decisions under complex disturbance environments. The real-time control feedback module executes control actions through a multi-objective dynamic trade-off algorithm, monitors equipment efficiency and production quality in real time, automatically generates production optimization reports with equipment health compensation, and continuously optimizes production line performance and dynamically maintains equipment health, improving the system's robustness and overall efficiency in dynamic environments. This solves the problems of static model construction, lagging decision-making mechanisms, and lack of control feedback in existing technologies.

[0054] The following will illustrate a specific embodiment of a digital twin-based intelligent real-time control system for production lines, including:

[0055] In a leading new energy vehicle group's smart manufacturing benchmark factory in the Yangtze River Delta, the factory focuses on the flexible mixed-line production of pure electric sedans and high-end SUVs, with a planned annual production capacity of 300,000 vehicles. Its core production chain covers four core processes: stamping, welding, painting, and final assembly, as well as 12 auxiliary workstations. The intelligent real-time control system for the production line based on digital twins has achieved closed-loop management of the entire chain, from physical equipment perception and virtual model mapping to decision optimization and feedback execution. This has become a key technical support for the factory to overcome the bottleneck of mixed-line production of multiple models and improve dynamic response capabilities.

[0056] Among them, the state perception and acquisition module serves as the system's data entry point, constructing a perception network that is "fully covered, high-frequency and accurate, and multi-dimensional collaborative." The factory has deployed over 3,800 sets of high-precision sensing equipment across its four main processes. These devices cover a wide range of types, including: piezoelectric vibration sensors (sampling rate up to 1kHz, capable of capturing micron-level vibration displacement of equipment with an accuracy of ±0.001mm) and PT100 platinum resistance temperature sensors (real-time monitoring of stamping die temperature, measurement range -200℃~600℃, error ≤±0.1℃) in the stamping workshop; and 5-megapixel industrial CMOS cameras (capturing weld point images every 0.08 seconds, coupled with a deep learning-based defect detection algorithm to identify 12 common problems such as weld point misalignment and cold welds) and laser contour sensors (scanning frequency 200Hz, generating three-dimensional weld point data) in the welding workshop. The cloud model, with a deviation controlled within ±0.15mm compared to the preset process standard; the painting workshop uses an infrared spectrometer (to detect paint film thickness and orange peel in real time, with a paint film thickness measurement range of 0~200μm and an accuracy of ±0.3μm) and a temperature and humidity sensor (sampling interval of 10 seconds to ensure that the humidity of the spraying environment is stable at 50%~60% and the temperature is 23℃±2℃); the final assembly workshop uses an AGV (automatic guided vehicle) positioning sensor (based on UWB technology, with a positioning accuracy of ±10cm, and real-time collection of 28 operating parameters such as AGV driving speed, load weight, and battery power) and a torque sensor (to monitor bolt tightening torque, with a range of 0~500N・m and an error of ≤±1%). Furthermore, the module interacts in real-time with the factory's MES (Manufacturing Execution System) and ERP (Enterprise Resource Planning) systems via API interfaces to acquire management data such as order production progress, material inventory levels, and supplier delivery cycles. It collects and cleans over 1.2TB of multi-source data daily, building a dynamic database containing 32,000 data traceability points, providing comprehensive data support for subsequent digital twin modeling and decision analysis. For example, at a side panel welding station for a certain vehicle model in the welding workshop, the system, through the collaborative acquisition of data from 8 industrial cameras and 2 laser sensors, can complete a comprehensive inspection of 24 welding points at that station within 0.3 seconds. If a welding point is found to have an offset of 0.5mm (exceeding the 0.3mm process threshold), an anomaly marker is immediately triggered and synchronized to the digital twin model, providing data for subsequent adjustments.

[0057] The digital twin modeling module, relying on the "full-element modeling + multi-physics simulation" capabilities of the Fuli Code Cloud platform, constructs a virtual production line that maps 1:1 to the physical factory, achieving hierarchical modeling and dynamic collaboration from the equipment level, workstation level to the workshop level. The module first standardizes multi-source heterogeneous data using ETL (Extract-Transform-Load) tools: it converts CAD-designed 3D equipment models (STEP format), real-time control commands from PLC controllers (Modbus protocol), production task data from the MES system (JSON format), and physical parameters collected by sensors (CSV format) into a unified twin data format supported by the platform. Furthermore, it eliminates data redundancy and noise through data fusion algorithms (e.g., using Kalman filtering to smooth high-frequency fluctuation data from vibration sensors, resulting in a 40% improvement in the signal-to-noise ratio after filtering). Based on this, the module constructs a three-level twin model: the equipment-level model (such as the joint motion model of the welding robot and the slider dynamics model of the stamping machine) can simulate the operating state of a single piece of equipment with an accuracy of 98%, and can calculate the stress and wear of key components of the equipment (such as the robot reducer and the crankshaft of the stamping machine) in real time; the workstation-level model (such as the welding side panel workstation and the painting and spraying workstation) integrates the interaction relationship between all equipment and materials in the workstation, and can simulate the material flow efficiency (such as the waiting time for AGV to deliver materials to the workstation) and the matching degree of process parameters (such as the adaptability of welding current and steel plate thickness); the workshop-level model (such as the entire welding workshop and the final assembly workshop) realizes the global collaborative simulation of all equipment, materials and personnel in the workshop, and supports the optimization of production cycle and the identification of bottleneck workstations. Meanwhile, the module constructs an equipment performance degradation prediction model based on LSTM (Long Short-Term Memory) neural network. By inputting the equipment's operating parameters (such as vibration frequency, temperature, and energy consumption) and fault records (more than 1,200 fault samples) over the past 6 months, the model can provide early warning of equipment failure risks up to 72 hours in advance. The accuracy rate for predicting joint wear of welding robots reaches 91%, and the accuracy rate for predicting block jamming of stamping machines reaches 89%. In addition, the module also constructs a disturbance propagation model through graph theory algorithms, which can simulate the propagation path and impact range of various disturbances (such as material shortages caused by AGV failures and paint film defects caused by parameter drift of painting equipment) in the production line. For example, when an AGV battery fails and cannot deliver materials, the model can calculate within 10 seconds that the failure will cause the No. 3 welding station to stop and further affect the vehicle assembly sequence in the final assembly workshop, providing the decision-making module with a predictive basis for disturbance response.

[0058] The intelligent decision-making module adopts a "distributed collaboration + multi-algorithm fusion" architecture to achieve millisecond-level generation and multi-objective optimization of production scheduling strategies. The module deploys five edge computing nodes (using NVIDIA Jetson AGXXavier hardware with a computing power of 32 TOPS) in the factory, responsible for local data processing and preliminary decision-making in the four major workshops of stamping, welding, painting, and final assembly, as well as auxiliary workstations. Simultaneously, two high-performance servers (Intel Xeon Platinum 8375C CPU and NVIDIA A100 GPU) are deployed in the factory cloud to handle the collaborative optimization of global production resources. The edge nodes and the cloud interact in real time via a 5G private network (latency ≤10ms, bandwidth ≥1Gbps). At the algorithmic level, the module employs a distributed collaborative reinforcement learning algorithm to decompose the production scheduling problem into three levels: equipment-level scheduling (such as optimizing the work sequence of welding robots and path planning for AGVs) is implemented using the DQN (Deep Q-Network) algorithm, aiming to minimize the idle time of individual equipment; production line-level scheduling (such as the allocation of workstations for a certain car model in the welding workshop and the adjustment of the painting sequence in the painting workshop) is implemented using the PPO (Proximity Policy Optimization) algorithm, aiming to balance the production load of each workstation; and factory-level scheduling (such as prioritizing the production of multiple car models and cross-workshop material allocation) is implemented using a hybrid algorithm of genetic algorithm (GA) and particle swarm optimization (PSO), aiming to maximize overall capacity and order delivery rate. To ensure the rationality of the decision, the algorithm's reward function design incorporates multiple dimensions of objectives: capacity achievement rate (weight 40%), equipment load balance (weight 30%), and product quality pass rate (weight 30%). For example, when the equipment load rate of a certain process exceeds 95%, the reward function will trigger a negative penalty, guiding the algorithm to adjust the task allocation. Meanwhile, the module constructs a knowledge graph covering "equipment-process-quality-management". The graph contains ledger information for 1200+ sets of equipment (such as equipment model, rated parameters, and maintenance cycle), process specifications for 200+ processes (such as welding current range and spraying pressure standards), 5 million+ historical optimization cases (such as scheduling plans for dealing with order fluctuations and equipment failures in the past 3 years), and 1000+ fault tree analysis (FTA) nodes (such as the correlation between weld defects and welding parameters). Multi-objective optimization reasoning is achieved through GNN (Graph Neural Network).For example, when a factory receives a sudden 30% increase in orders for a certain SUV model (from an original daily production capacity of 120 vehicles to 156 vehicles), the intelligent decision-making module processes the following: First, edge nodes collect real-time equipment load data from the welding and final assembly workshops, discovering that the welding side panel station and the final assembly chassis station are already at full capacity; then, the cloud server adjusts the production plan through a hybrid algorithm: optimizing the mold switching sequence in the stamping workshop (the original mold switching time for a certain sedan was 2 hours, which was reduced to 1.5 hours through preheating the mold and automated mold changing), and simultaneously coordinating with upstream suppliers to provide the SUV model... The frequency of stamped parts delivery was increased from 3 times to 4 times per day. Then, the knowledge graph used GNN inference to match historical optimization solutions under the scenario of "high order volume + full equipment load", and suggested diverting some auxiliary processes in the final assembly workshop (such as seat installation) to adjacent idle workstations. Finally, the module generated a scheduling strategy containing 15 control instructions within 200 milliseconds, covering dimensions such as equipment parameter adjustment (such as increasing the welding current of the welding robot from 180A to 190A), material delivery optimization (such as adjusting the AGV delivery path), and workstation task allocation (such as adding 2 workers to the No. 3 workstation in the final assembly).

[0059] The real-time control feedback module, serving as the system's execution exit, achieves a closed-loop control system through "precise command issuance + dynamic performance monitoring + intelligent report generation." The module first sends the control commands generated by the intelligent decision-making module to the PLC controllers (such as the Siemens S7-1500 series) at each process stage via the OPCUA protocol (an industrial communication standard). The command transmission delay is controlled within 5 milliseconds to ensure rapid response from the actuators. For example, regarding the predicted deviation in paint film thickness at the painting station in the coating workshop (the virtual model predicts that the paint film thickness in a certain area will drop to 75μm, below the process standard of 80μm), the module issues control commands: adjusting the spraying pressure of the painting robot from 0.3MPa to 0.32MPa, reducing the spray gun movement speed from 50mm / s to 45mm / s, and extending the spraying time in that area by 0.5 seconds. After receiving the commands, the PLC controller drives the painting robot to complete the parameter adjustments within 1 second, ensuring that the paint film thickness returns to the standard range. During the execution of control actions, the module monitors key production line indicators in real time through edge computing nodes, including equipment efficiency indicators (such as equipment OEE, downtime, and energy consumption), production quality indicators (such as weld pass rate, paint film thickness compliance rate, and bolt tightening torque compliance rate), and material flow indicators (such as AGV delivery on-time rate and material inventory turnover rate). The monitoring frequency is 10 times per second. If an indicator is found to exceed a threshold (such as the welding robot OEE dropping to 80%, below the target value of 85%), secondary control will be triggered immediately. In addition, the module constructs an equipment health compensation model based on a particle filter algorithm, which can dynamically correct the health assessment results according to the real-time operating parameters of the equipment. For example, when the crankshaft temperature of the stamping press is detected to be 5°C higher than normal, the model adjusts the health compensation coefficient from 1.0 to 1.2, corrects the remaining life prediction of the crankshaft from 1200 hours to 950 hours, and adds a maintenance suggestion to "shorten the crankshaft lubrication cycle from 7 days to 5 days" to the control instructions. After each day's production is completed, the module automatically generates a production line optimization report containing "control effects + problem analysis + optimization suggestions": The first part of the report presents the day's control execution status (e.g., a total of 32 control commands were executed, with a 100% success rate, resolving 28 production anomalies); the second part analyzes changes in key indicators (e.g., equipment OEE reached 86.2%, an increase of 1.8 percentage points from yesterday; weld point qualification rate was 99.6%, exceeding the target value by 0.1 percentage points; order delivery on time rate was 98.5%, an increase of 2.3 percentage points from last week); the third part proposes targeted optimization suggestions (e.g., based on the health compensation data of the stamping press, it is recommended to conduct crankshaft wear testing on 3 stamping presses next week; based on the energy consumption data of the painting workshop, it is recommended to adjust the peak and off-peak electricity usage ratio of the spraying equipment to reduce energy costs).

[0060] In summary, this application's embodiments utilize state perception and 1:1 twin modeling to achieve visualized monitoring of all elements of the production line, predicting equipment degradation and disturbance propagation in advance, reducing emergency handling time for equipment failures from 1 hour to within 15 minutes; relying on the intelligent decision-making module, it responds to order fluctuations and mixed-line production scenarios at millisecond level, generating optimal scheduling strategies, shortening capacity ramp-up time by 20%, and increasing overall equipment efficiency (OEE) from 78% to 86.2%, significantly improving production flexibility; through closed-loop control feedback, while protecting equipment health, it increases the weld point qualification rate from 92% to 99.5% and reduces the product defect rate to below 0.5%, stabilizing quality; it can also optimize maintenance through health compensation and automatically generate reports, saving over 12 million yuan in annual maintenance costs, and the virtual twin can shorten the training cycle for new employees by 50%, while providing data support for process iteration and capacity planning, helping enterprises transform from traditional production to intelligent and lean production.

[0061] Next, referring to the accompanying drawings, a method for intelligent real-time control of a production line based on digital twins, according to an embodiment of this application, is described.

[0062] like Figure 6 As shown, this intelligent real-time control method for production lines based on digital twins includes the following steps:

[0063] In step S101, the real-time operating parameters of the production equipment, material flow status information, and environmental disturbance characteristics of the production line are obtained.

[0064] It is understandable that this application, by acquiring real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics of the production line, forms the fundamental premise for constructing a precise digital twin mapping and realizing intelligent control. Real-time operating parameters of production equipment directly reflect the equipment's health status and performance fluctuations, providing crucial evidence for predicting equipment degradation trends. Material flow status information depicts the dynamic connection of the production process, supporting the identification of bottlenecks and breakpoint risks. Environmental disturbance characteristics quantify the impact of external non-steady-state factors on the production system, avoiding prediction biases from a single equipment perspective. The integration of these three elements comprehensively depicts the real-time status of the physical entity of the production line from multiple dimensions—equipment, materials, and environment—providing high-fidelity input for a dynamic data-driven twin model. This ensures precise synchronization between the virtual model and the physical factory, laying a data foundation for subsequent performance degradation prediction, disturbance propagation analysis, and the generation of collaborative control strategies. Ultimately, it enables closed-loop optimization of the entire chain from status perception to intelligent decision-making, improving the production line's adaptability and operational reliability in the face of complex operating conditions.

[0065] In step S102, the real-time operating parameters of the production equipment, material flow status information and environmental disturbance characteristics of the production line are fused together. A precise mapping between the physical entity and the virtual model of the production line is constructed through a dynamic data-driven twin model. The performance degradation trend of the production line equipment and the disturbance propagation path are predicted in real time to obtain the prediction results.

[0066] Dynamic data-driven refers to an operational mode that uses real-time or continuously updated dynamic data as the core driving force to promote system operation, process optimization, or decision-making, and enables it to adjust in real time as the data changes.

[0067] It is understood that the embodiments of this application use multi-source data fusion to dynamically map physical entities and virtual models. By leveraging the continuous updating characteristics of real-time data, the state of the twin model is calibrated, improving the timeliness and accuracy of the model's prediction of equipment performance degradation trends and disturbance propagation paths. This enhances the ability to predict abnormal responses in production lines, reduces the risk of control failures caused by model lag, and reduces reliance on offline static models to improve system adaptability. It also provides high-fidelity dynamic data support for distributed collaborative reinforcement learning scheduling strategy generation and knowledge graph optimization reasoning.

[0068] For example, in an engine assembly line at an automobile manufacturing plant, a dynamic data-driven twin model is deployed with 128 vibration sensors (sampling frequency 10kHz), RFID tags on the material conveyor line (99.9% recognition rate), and environmental temperature and humidity sensors (accuracy ±0.5℃ / ±3%RH). It synchronously collects equipment operating parameters, material flow status, and environmental disturbance data at 50ms intervals, achieving precise mapping between the physical entity and the virtual model. The system uses intelligent algorithms to predict the wear trend of the spindle motor bearings in real time (error ≤3%) and provides a 4-hour advance warning. When material delivery delays are detected, simulation analysis identifies the cascading effects on 27 workstations, automatically generating a dynamic scheduling plan. A multi-objective trade-off algorithm adjusts the robot's motion cycle (e.g., reducing bolt tightening time from 8 seconds to 7.2 seconds) and simultaneously optimizes the AGV path (reducing it by 23 meters). This solution reduces unplanned downtime by 60%, increases material inventory turnover by 25%, and generates daily production optimization reports with equipment health compensation, creating a millisecond-level closed loop from data acquisition to decision execution.

[0069] In step S103, based on the prediction results, a real-time scheduling strategy for the production line is generated through a distributed collaborative reinforcement learning algorithm, and combined with multi-objective optimization reasoning based on knowledge graphs, a set of production line control instructions is output.

[0070] Among them, distributed cooperative reinforcement learning algorithms refer to reinforcement learning algorithms that use multiple agents or computing nodes to be deployed in a distributed manner, cooperate with each other in trial and error interaction with the environment, and jointly optimize the global goal, improve learning efficiency and obtain the optimal joint strategy.

[0071] Distributed cooperative reinforcement learning algorithm: ; in, For strategy parameters, For the next time step; For the whole; The number of distributed intelligent agents in the production line; This is the starting index for the summation; For the current moment Local policy parameters of each agent; This is the global regularization coefficient; For global policy parameters Find the partial derivative; The loss function; This is the current global value function; The objective is the global value function.

[0072] It is understood that the embodiments of this application utilize a distributed collaborative reinforcement learning algorithm, based on a multi-agent collaborative decision-making mechanism, to efficiently generate real-time production line scheduling strategies and perform real-time collaborative knowledge graph multi-objective optimization reasoning. Even in production line scenarios with multiple coupled disturbances, it can still accurately output a highly adaptable set of production line control instructions. Its distributed collaborative optimization capability can ensure the globality and real-time performance of the scheduling strategy, providing synergy and effectiveness for the precise execution of production line control actions and the dynamic monitoring of equipment performance. At the same time, it supports the automatic generation of production line optimization reports with equipment health compensation. It overcomes the limitations of traditional centralized scheduling algorithms, improves the flexibility and anti-disturbance of production line control, and ensures that control actions can accurately adapt to changes in equipment performance degradation and disturbance propagation, providing a timely and reliable decision-making basis for production line equipment performance evaluation and production quality optimization.

[0073] For example, in the real-time scheduling scenario of an automobile assembly line, the distributed collaborative reinforcement learning algorithm can be deployed as follows: The production line is divided into three core workstations: chassis assembly, powertrain integration, and interior installation. Each workstation deploys an independent intelligent agent (a total of three agents). Each agent is equipped with a local observation module (collecting the operating status of the equipment at its workstation, the AGV material delivery delay rate, and the amount of work-in-process inventory) and a communication interface, sharing global information (such as order delivery countdown, the overall OEE target of the production line, and a historical disturbance event database). The algorithm employs a centralized training-distributed execution (CTDE) framework: During the training phase, a digital twin model simulates 100,000 disturbance scenarios (such as sudden robot shutdown at a workstation or fluctuations in material batch quality). Each agent learns local actions (such as adjusting assembly rhythm, switching AGV priorities, and calling spare tooling) based on proximal policy optimization (PPO). Simultaneously, the collaborative strategy is corrected through a global reward function (combining production cycle reduction rate, equipment energy consumption reduction, and order on-time delivery rate). During the execution phase, when the twin model predicts in real time that "the chassis workstation will experience a 15-second delay due to tool wear," the chassis workstation agent immediately reduces the speed of non-critical processes (local action) and informs the power system workstation agent that "semi-finished products need to be output 10 seconds earlier." The power system workstation agent synchronously adjusts the robot's material picking rhythm and coordinates upstream AGVs to accelerate delivery. Ultimately, through multi-agent collaboration, the impact of the disturbance is limited to a local area, reducing the overall production line efficiency loss from 8% under traditional rules to 2%. This algorithm adaptively optimizes production line scheduling strategies under complex disturbances through dynamic collaboration and continuous learning among agents.

[0074] In step S104, based on the production line control instruction set, the production line control action is executed through a multi-objective dynamic trade-off algorithm, the equipment efficiency and production quality of the production line are monitored in real time, production line status data is collected and stored, and a production line production optimization report with equipment health compensation is automatically generated.

[0075] Among them, equipment health compensation refers to the measures taken by the equipment provider or maintenance entity to restore or improve the health level of the equipment through technical debugging, resource replenishment, component maintenance, etc., when the health status of the equipment fails to meet the preset standard, or when performance loss or functional degradation occurs, so as to ensure the normal operation of the equipment.

[0076] It is understood that the embodiments of this application align the health features of production line status data and match multi-source information by using a unified equipment health assessment benchmark. This enables the dynamic data-driven twin model to have an accurate reference for predicting equipment performance degradation. By constraining prediction deviations with features such as the real-time change trend of equipment health and historical degradation patterns, the prediction accuracy of equipment performance degradation trends and disturbance propagation paths is improved. This enhances the adaptability of real-time scheduling strategies and the effectiveness of control actions, reduces unplanned equipment downtime or production quality fluctuations, and reduces reliance on high-frequency manual inspections or redundant maintenance resources to control costs. This provides data support for subsequent intelligent control such as equipment lifecycle management and continuous optimization of production processes.

[0077] For example, during the operation of a chip mounter on an automotive electronics production line, data fusion analysis of vibration sensors and current monitoring revealed that the health of the nozzle module had dropped to 75% (out of 100%) due to long-term high-frequency use. It was predicted that the probability of component cold solder joints due to nozzle wear would exceed 60% after 24 hours. Based on a digital twin model, the system simulated different compensation strategies and automatically generated the following adjustment plan: reducing the mounting speed by 15% to reduce mechanical stress, triggering a predictive maintenance work order, and scheduling the robot to replace the nozzle with a spare in advance; simultaneously increasing the sampling rate in the quality inspection stage by 30%, and mapping the adjusted production cycle and quality fluctuation trend in real time in the digital twin model. Through a multi-objective dynamic trade-off algorithm, after equipment health compensation, the cold solder joint defect rate at this station decreased from 42% to 9%. Although daily output decreased by 8%, the overall quality loss cost decreased by 67%, and the risk of entire line downtime due to sudden failures was avoided. The production optimization report details the specific parameters of health compensation (such as nozzle wear rate and health threshold of 70%), the dynamic response time of the control action (2 minutes and 15 seconds), and the quantitative impact on capacity, quality, and maintenance costs, providing data closed-loop support for subsequent production line iterations.

[0078] The intelligent real-time control method for production lines based on digital twins proposed in this application acquires real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics through a state perception and acquisition module. This comprehensively perceives the production line's operating status and external disturbance factors, providing full-element data support for subsequent analysis and decision-making. Combined with the dynamic data-driven model of the digital twin modeling module, multi-source data is integrated to construct a precise mapping between the physical entity and virtual model of the production line. This allows for real-time prediction of equipment performance degradation trends and disturbance propagation paths, enhancing the depth and foresight of production line state perception. The distributed collaborative reinforcement learning algorithm of the intelligent decision-making module generates real-time scheduling strategies. Combined with multi-objective optimization reasoning using knowledge graphs, a set of control instructions that meets multiple objective requirements is output, enhancing the synergy and adaptability of scheduling decisions under complex disturbance environments. The real-time control feedback module executes control actions through a multi-objective dynamic trade-off algorithm, monitors equipment efficiency and production quality in real time, automatically generates production optimization reports with equipment health compensation, continuously optimizes production line performance, and dynamically maintains equipment health, improving the system's robustness and overall efficiency in dynamic environments. This solves the problems of static model construction, lagging decision-making mechanisms, and lack of control feedback in existing technologies.

[0079] The following will illustrate a specific embodiment of a method for intelligent real-time control of a production line based on digital twins, including:

[0080] At 7:55 AM, the day shift is about to begin. The data acquisition gateway, located on the factory edge server, starts operating at high speed. It acquires real-time operating parameters of the production line equipment through thousands of sensors deployed on each device: the current temperature of the spindle in machining center 1 is 28.5℃, the vibration spectrum of the X-axis lead screw is within the normal range, but the current ripple of its drive motor shows a slight fluctuation of 0.5%; the cutting fluid pressure in machining center 2 is stable at 0.6MPa, but the pH sensor shows its acidity / alkalinity is slowly approaching the corrosive threshold. Simultaneously, the system also acquires material flow status information: the AGV scheduling system reports that a specific type of bearing for production order 3 is en route and is expected to arrive at the lineside warehouse at 8:15 AM; the vision inspection station reports that the reference surface roughness Ra value of the previous batch of workpieces has a statistical upward trend of 0.01 micrometers. In addition, the system also detected environmental disturbances: the factory's energy management system pushed a message that due to partial maintenance of the city's power grid, the factory area may face voltage fluctuations of ±5% from 2 pm to 4 pm today; the central air conditioning system also reported that due to the sudden rise in outdoor temperature today, the environmental temperature control accuracy of the cleanroom will face challenges, with an expected fluctuation of ±0.3℃.

[0081] All this massive, heterogeneous, real-time (millisecond-level) data is aggregated into a data lake in the cloud. Next, the dynamic, data-driven digital twin model begins to play a crucial role. This model is not a static 3D visualization model, but a "living" model that deeply integrates physical mechanisms (such as machine tool dynamics, thermal deformation theory, and cutting mechanics) with real-time data (continuously calibrated through deep learning algorithms). It uses the aforementioned real-time data as input to construct a precise mapping between the physical entity of the production line and the virtual model. In the virtual space, the "digital twin" of machining center No. 1 begins to predict the performance degradation trend of the production line equipment in real time based on spindle temperature and current ripple data. By analyzing historical data and failure physics models, the model calculates that the remaining service life (RUL) of the spindle bearing decreases from the estimated 1500 hours to 1420 hours, and predicts that after completing the current batch of 50 workpieces, its thermal deformation will cause an 85% probability that the machining error of the next critical hole diameter will exceed the tolerance band by 1.5 micrometers. Meanwhile, the model predicts the disturbance propagation path: it simulates the impact of afternoon voltage fluctuations on the servo system of machining center No. 2, deducing that this will lead to an increase in the roundness error of its precision boring holes; furthermore, fluctuations in ambient temperature will have a coupling effect with the thermal deformation of machine tool No. 1, further exacerbating the deterioration of accuracy. This prediction result, which includes equipment health warnings and disturbance effects, is clearly presented in the digital twin cockpit of the factory's central control room.

[0082] Faced with these predicted risks, the system does not wait for failures to occur but proactively intervenes. Based on the prediction results, a distributed collaborative reinforcement learning algorithm is triggered. In this algorithm architecture, each machining center, robot, and AGV is an agent. They share a common goal—maximizing the OEE and product quality of the entire production line—but also need to negotiate resources (such as tools, time, and materials). The algorithm performs millions of simulations and strategic games at millisecond speeds in a "simulation sandbox" provided by the digital twin model. For example, for the predicted performance degradation of machine tool No. 1, the algorithm may generate several alternative solutions: Solution A, immediately dispatch a backup machine tool to take over, but this will cause a 30-minute production interruption; Solution B, dynamically adjust the cutting parameters for the next workpiece of machine tool No. 1 (reduce the speed, increase the feed), sacrificing 15 seconds of cycle time for accuracy, but this will accelerate tool wear; Solution C, adjust the production sequence, prioritizing the allocation of workpieces with slightly lower accuracy requirements to machine tool No. 1, and processing critical parts after it cools down. These agents learn through "competition-cooperation" and ultimately generate a real-time production line scheduling strategy. Subsequently, the knowledge graph multi-objective optimization reasoning module intervenes. This knowledge graph stores information such as domain experts' experience rules, process standards, customer order priorities, and cost matrices. It performs a multi-objective (quality, efficiency, cost, equipment lifespan) trade-off evaluation of the scheduling strategy. For example, it determines that although Option B maintains the current cycle time, it violates the expert rule of "prohibiting proactive process degradation for critical dimensions" and is therefore rejected. Option C is feasible, but the knowledge graph retrieves information indicating that the customer has an urgent need for order number 3, which cannot be delayed. Ultimately, under the coordination of the inference engine, a fusion solution was formed: outputting a production line control instruction set—Instruction 1: Immediately swap the 5th high-precision workpiece in the current task sequence of machine tool 1 with the 8th medium-precision workpiece in the sequence of machine tool 2; Instruction 2: Issue a new machining code to machine tool 1, and fine-tune and compensate the cutting parameters of its 5th workpiece (originally the task of machine tool 2); Instruction 3: Notify the AGV system to upgrade the delivery priority of the specific tool T-234 required by machine tool 1; Instruction 4: Schedule a maintenance engineer to perform preventive inspection and status monitoring of the spindle of machine tool 1 during the evening shift.

[0083] Based on this production line control instruction set, the physical world begins to respond. Instructions are distributed to various equipment controllers via the Industrial Internet of Things (IIoT) platform. The multi-objective dynamic trade-off algorithm, when executing production line control actions, does not rigidly execute them but continuously monitors the equipment efficiency and production quality of the production line in real time. When the swapped workpiece begins processing on machine tool No. 1, the online measurement probe feeds back the aperture data to the system in real time. The system detects that despite sequential adjustments, the error is still hovering at the critical point due to the accumulation of thermal deformation. At this point, the dynamic trade-off algorithm immediately activates the compensation mechanism: based on real-time feedback data, it performs rapid iterative calculations in the digital twin model, dynamically fine-tuning the machine tool's tool compensation value, successfully bringing the machining accuracy back to the tolerance zone center. Throughout the process, every control action, every change in equipment status, and every workpiece's quality data is collected and stored by the system as production line status data. This data, in turn, is used to train and optimize the twin model and reinforcement learning algorithm, forming a closed loop of "perception-decision-execution-learning."

[0084] When production ends for the day, the system no longer simply summarizes output. It automatically generates a production line optimization report with equipment health compensation. This report not only includes traditional KPIs such as 105 qualified spindle boxes produced today and OEE reaching 86.5%, but more importantly, it includes in-depth analysis based on equipment health status: the report points out that proactive intervention on machine tool No. 1 successfully avoided the generation of 3 potential defective products, equivalent to saving XX yuan in value loss; based on its spindle RUL prediction, it recommends preventative maintenance during the evening shift 72 hours later, and includes a maintenance checklist; the report also analyzes production data during afternoon voltage fluctuations, verifying the accuracy of the twin model prediction, and recommends installing a voltage regulator on machine tool No. 2 to improve long-term stability. This report provides managers with a new perspective for decision-making, moving from "post-event statistics" to "pre-event prediction and in-event optimization," truly realizing intelligent, real-time, and autonomous control of the production line.

[0085] In summary, the embodiments of this application accurately predict equipment performance degradation and disturbance propagation through digital twin technology, enabling the system to proactively intervene before potential faults and quality deviations occur. Combined with distributed collaborative reinforcement learning and knowledge graph reasoning, a multi-objective optimization scheduling strategy that simultaneously balances efficiency, quality, cost, and equipment lifespan is generated, dynamically executed, and compensated for, ultimately forming a closed loop of "perception-decision-execution-learning". This significantly improves production efficiency and product quality, greatly reduces downtime and scrap costs, and extends equipment lifespan, driving the production system towards intelligence and autonomy.

[0086] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0087] The memory 701, the processor 702, and the computer program stored on the memory 701 and capable of running on the processor 702.

[0088] When the processor 702 executes the program, it implements the intelligent real-time control method for production lines based on digital twins provided in the above embodiments.

[0089] Furthermore, electronic devices also include:

[0090] Communication interface 703 is used for communication between memory 701 and processor 702.

[0091] The memory 701 is used to store computer programs that can run on the processor 702.

[0092] The memory 701 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0093] If the memory 701, processor 702, and communication interface 703 are implemented independently, then the communication interface 703, memory 701, and processor 702 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0094] Optionally, in a specific implementation, if the memory 701, processor 702, and communication interface 703 are integrated on a single chip, then the memory 701, processor 702, and communication interface 703 can communicate with each other through an internal interface.

[0095] The processor 702 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0096] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described intelligent real-time control method for production lines based on digital twins.

[0097] Furthermore, this application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described intelligent real-time control method for production lines based on digital twins.

[0098] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0099] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0100] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0101] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0102] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0103] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A production line intelligent real-time control system based on digital twins, characterized in that, include: The system comprises a state perception and acquisition module, a digital twin modeling module, an intelligent decision-making module, and a real-time control and feedback module; among which, The status perception and acquisition module is used to acquire real-time operating parameters of production equipment, material flow status information and environmental disturbance characteristics of the production line. The digital twin modeling module is used to integrate multi-source data and construct a precise mapping between the physical entity and virtual model of the production line through a dynamic data-driven twin model, so as to predict the performance degradation trend and disturbance propagation path of the production line equipment in real time. The intelligent decision-making module is used to generate real-time scheduling strategies for the production line through a distributed collaborative reinforcement learning algorithm, and output a set of production line control instructions by combining multi-objective optimization reasoning based on knowledge graphs. The real-time control feedback module, based on the production line control instruction set, executes production line control actions through a multi-objective dynamic trade-off algorithm, monitors the equipment efficiency and production quality of the production line in real time, and automatically generates a production line optimization report with equipment health compensation.

2. The intelligent real-time control system for production lines based on digital twins according to claim 1, characterized in that, The status perception and acquisition module includes an equipment operation parameter acquisition unit, a material flow status tracking unit, and an environmental disturbance perception unit. The equipment operation parameter acquisition unit is used to acquire real-time operating parameters of the production equipment through the equipment IoT interface and sensor network, and record key indicators such as equipment vibration, temperature, power consumption, and operating cycle. The material flow status tracking unit is used to track the position, quantity, and status information of materials in the production line in real time. The environmental disturbance perception unit is used to monitor the disturbance characteristics of the production environment through temperature, humidity, vibration, and electromagnetic interference sensors, identify external interference sources, and quantify their impact on the stability of the production line.

3. The intelligent real-time control system for production lines based on digital twins according to claim 1, characterized in that, The digital twin modeling module includes a multi-source data fusion unit, a dynamic mapping construction unit, and a performance prediction unit. The multi-source data fusion unit cleans, aligns, and fuses data using spatiotemporal alignment and feature extraction algorithms to eliminate data heterogeneity and noise interference. The dynamic mapping construction unit, based on the fused data, constructs a high-fidelity dynamic mapping between the physical entity of the production line and the virtual model, synchronizing equipment behavior, material flow, and environmental disturbances between the virtual and physical worlds. The performance prediction unit analyzes the relationship between equipment performance indicators and disturbance propagation in the dynamic mapping to predict equipment performance degradation trends and the impact of disturbance events on production cycle time and quality.

4. The intelligent real-time control system for production lines based on digital twins according to claim 1, characterized in that, The intelligent decision-making module includes a distributed collaborative reinforcement learning unit and a multi-objective optimization reasoning unit. The distributed collaborative reinforcement learning unit trains and optimizes multi-agent policies through a distributed collaborative reinforcement learning algorithm to generate a real-time distributed scheduling strategy. The multi-objective optimization reasoning unit performs multi-objective trade-off analysis through a knowledge graph multi-objective optimization reasoning algorithm, and outputs a set of control instructions that meet the multiple objectives of efficiency, quality, and cost in conjunction with the distributed scheduling strategy.

5. The intelligent real-time control system for production lines based on digital twins according to claim 1, characterized in that, The formula for the distributed collaborative reinforcement learning algorithm is as follows: in, For intelligent agents The optimal action value function; State; For joint action; For intelligent agents The reward received immediately after performing the action; Discount factor; The total number of agents; Indexing for intelligent agents; For collaborative weighting coefficients; For intelligent agents optimal value; The next state; For the next joint action.

6. The intelligent real-time control system for production lines based on digital twins according to claim 1, characterized in that, The real-time control and feedback module includes a dynamic trade-off execution unit and an optimization report generation unit. The dynamic trade-off execution unit uses a multi-objective dynamic trade-off algorithm to perform real-time trade-offs and decision-making on key indicators such as production efficiency, equipment load, and product quality during the control process. The optimization report generation unit monitors the comprehensive efficiency and production quality indicators of equipment in real time, collects and stores production line status data, and combines the equipment health compensation coefficient to perform correlation analysis on the monitoring data and control results, automatically generating an optimization report containing suggestions for adjusting production parameters.

7. A method for intelligent real-time control of a production line based on digital twins, applicable to any one of claims 1-6, characterized in that, include: S101: Acquire real-time operating parameters of production equipment, material flow status information, and environmental disturbance characteristics of the production line; S102: The real-time operating parameters of the production equipment, material flow status information and environmental disturbance characteristics of the production line are integrated. A precise mapping between the physical entity and the virtual model of the production line is constructed through a dynamic data-driven twin model. The performance degradation trend of the production line equipment and the disturbance propagation path are predicted in real time to obtain the prediction results. S103: Based on the prediction results, a real-time scheduling strategy for the production line is generated through a distributed collaborative reinforcement learning algorithm, and a set of production line control instructions is output by combining multi-objective optimization reasoning based on knowledge graph. S104: Based on the production line control instruction set, execute production line control actions through a multi-objective dynamic trade-off algorithm, monitor the equipment efficiency and production quality of the production line in real time, collect and store production line status data, and automatically generate a production line optimization report with equipment health compensation.

8. An electronic device, characterized in that, include: The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent real-time control method for a production line based on digital twins as described in claim 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the intelligent real-time control method for production lines based on digital twins as described in claim 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the intelligent real-time control method for production lines based on digital twins as described in claim 7.

Images