Digital twin workshop production scheduling decision system and method fusing virtual and real simulation
By integrating virtual and real simulation into a digital twin workshop production scheduling decision system, and utilizing data acquisition and fusion, scheduling engine, simulation sandbox, and dynamic optimization modules, the system solves the problems of slow response and poor robustness in existing technologies, achieves rapid response to production disturbances and optimization of scheduling strategies, and improves the level of intelligence in production scheduling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-05-22
- Publication Date
- 2026-06-19
AI Technical Summary
Existing reinforcement learning-based digital twin shop floor scheduling methods are slow to respond and have poor robustness when facing highly dynamic and randomly disturbed production environments, and modeling errors between the model and the physical entity lead to unstable decision-making.
The digital twin workshop production scheduling decision system, which integrates virtual and real simulation, updates the digital twin model in real time through the data acquisition and fusion module. Combined with the scheduling engine, simulation sandbox module, and dynamic optimization module, it enables rapid iteration and closed-loop adjustment to optimize the scheduling strategy.
It enables rapid response to production disturbances, improves the agility and robustness of scheduling decisions, reduces the risk of scheduling conflicts, and has continuous learning capabilities, thereby enhancing the level of intelligence in production scheduling.
Smart Images
Figure CN122239657A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent industrial manufacturing, and in particular to a digital twin workshop production scheduling decision-making system and method that integrates virtual and real simulation. Background Technology
[0002] In the field of workshop production scheduling, digital twin technology, as the core of achieving deep integration of cyber-physical systems, provides a new solution for virtual-physical mapping, real-time interaction, and iterative optimization. By constructing a digital model synchronized with the physical workshop, scheduling schemes can be simulated and verified in virtual space. The performance of workshop scheduling optimization is directly related to indicators such as production efficiency and machine utilization.
[0003] To overcome the drawbacks of traditional static scheduling methods that are detached from real-time production status, digital twins, due to their powerful simulation and information synchronization capabilities, have been widely applied in shop floor scheduling frameworks to seek better scheduling strategies. For example, Chinese Patent CN115373353A discloses an improved digital twin shop floor scheduling method based on reinforcement learning. This method proposes using reinforcement learning algorithms to optimize scheduling decisions. The core of this type of method lies in: decomposing the physical shop floor into multiple subsystems in the digital twin virtual environment, building corresponding digital twin models, and generating a large amount of training data by repeatedly running the digital twin models to train a reinforcement learning policy model, thereby obtaining a near-optimal scheduling strategy. Compared with traditional heuristic rules, this method shows potential advantages in handling complex constraints and finding global optimal solutions, and is an important technical path to achieve intelligent scheduling.
[0004] However, in-depth research has revealed that existing technical solutions combining reinforcement learning and digital twins still have inherent limitations when facing highly dynamic and randomly disturbed real-world production environments. The core issues are slow response to disturbances and the potential instability caused by direct coupling between simulation and execution.
[0005] Regarding the issue of slow disturbance response, existing solutions typically train reinforcement learning agents in a pre-defined virtual environment. After training, the policy's response to disturbances still requires complete retraining. When unexpected disturbances occur in the workshop, such as equipment downtime, material shortages, or emergency orders, the slow iteration and response fail to meet the real-time response requirements of production scheduling, causing a deviation between the real and training environments and a sharp decline in the decision-making performance of the trained agent. Essentially, this stems from the lack of an online learning and adjustment mechanism capable of rapidly iterating and dynamically evolving in sync with changes in the physical environment.
[0006] The risk of decision-making under model mismatch arises because there are inherent modeling errors between the digital twin model and the physical entity, and these errors are amplified under dynamic perturbations. Applying a policy trained on an ideal model to a physical entity may result in suboptimal or even infeasible scheduling instructions, such as assigning a task to a device that is about to fail or whose efficiency has been significantly reduced.
[0007] In summary, while existing reinforcement learning-based digital twin scheduling methods have improved the intelligence level of scheduling optimization, their reliance on offline training makes them exhibit poor robustness and insufficient adaptability when facing continuous dynamic disturbances in real workshops. Summary of the Invention
[0008] To address the issues of information lag in response to disturbances, trial-and-error costs, and instability that may result from direct coupling between simulation and execution in existing technologies, this invention discloses a digital twin workshop production scheduling decision system that integrates virtual and real simulations. This system can sense disturbances, iterate rapidly, and adjust scheduling strategies in a closed loop, thereby achieving forward-looking decision-making, agile response, and closed-loop optimization in production scheduling.
[0009] A digital twin workshop production scheduling decision-making system integrating virtual and real simulation includes a physical workshop and a multi-dimensional digital twin model mapped to the physical workshop entity. The system also includes the following modules:
[0010] The data acquisition and fusion module collects production data from the physical workshop in real time through the Internet of Things sensor network, and processes (cleaning, denoising and semantic processing) and fuses the data to drive the state update of the digital twin model and maintain the dynamic consistency between the virtual space and the physical workshop.
[0011] The scheduling engine responds to changes in production plans or disturbance events by generating an initial scheduling scheme based on a rule model and real-time data.
[0012] The simulation sandbox module, as an isolated running instance of the digital twin model, is used to receive the initial scheduling scheme generated by the scheduling engine, perform parallel simulation verification in the virtual space, and output simulation evaluation results based on the comprehensive objective function.
[0013] The dynamic optimization module triggers an iterative optimization process when verification is unsatisfactory or when there are sudden disturbances. This process includes a fast screening layer and a fine optimization layer. The fast screening layer uses a lightweight model to pre-screen the initial schemes, while the fine optimization layer accurately evaluates the candidate schemes based on high-fidelity simulation and dynamically adjusts the optimization parameters through a feedback mechanism until a feasible scheduling scheme is obtained.
[0014] The instruction verification module is used to parse the optimization scheme into an atomic control instruction sequence, perform resource conflict and logic error verification in the simulation sandbox, and generate the final scheduling strategy after the verification is passed.
[0015] The distribution and monitoring module distributes the final scheduling strategy to the execution unit of the physical workshop in the form of an instruction sequence, and monitors the execution status of the digital twin model and the physical workshop to form a closed-loop feedback control.
[0016] Furthermore, the data acquisition and fusion module specifically includes:
[0017] The heterogeneous data acquisition unit integrates IoT sensors, equipment PLC controllers, RFID readers and writers and manufacturing execution system API interfaces to collect equipment operating status parameters, work-in-process and material identity and location information, production process progress data and order execution status data in real time or near real time.
[0018] The data preprocessing subunit cleans, denoises, detects outliers, and standardizes the format of the raw data. It also assigns a unified semantic identifier to the processed data, enabling it to establish a mapping relationship with the entity objects and attributes in the digital twin model.
[0019] The data synchronization subunit updates the status, attributes, and location information of the corresponding entity objects in the digital twin model according to a preset cycle or event triggering method.
[0020] Furthermore, the simulation sandbox module supports parallel simulation verification of multiple scheduling schemes, and the evaluation of the schemes is based on a comprehensive objective function:
[0021]
[0022] In the formula, T is the maximum completion time, L is the equipment load balancing index, C is the penalty term for violating the constraint, and ω1, ω2, ω3 are adjustable weight coefficients.
[0023] Furthermore, the feedback mechanism is as follows: when the fine optimization layer finds that the evaluation results of all candidate solutions are not ideal, it feeds back this result and bottleneck analysis to the fast screening layer. The fast screening layer triggers a dynamic adjustment strategy, including adaptive adjustment of rule weights, relaxation of constraint tightness, or search space guidance, so as to guide the next iteration to generate a more promising set of candidate solutions.
[0024] Furthermore, in the feedback mechanism, the evaluation results are all unsatisfactory when the average comprehensive evaluation value of the candidate solution set output by the fine optimization layer is... Better than historical benchmark When the amplitude is less than the preset threshold δ, the following condition is met: If the current candidate solution set is not of ideal quality, then it is determined that the quality of the current candidate solution set is not ideal.
[0025] Furthermore, dynamic adjustment strategies include:
[0026] The rule weights are adaptively adjusted. The fast filtering layer maintains a rule base R = {R1, R2, ..., Rn} and a corresponding weight vector W = {W1, W2, ..., Wn}, with initial weights being equal. The feedback mechanism dynamically adjusts the parameters of W based on the bottleneck analysis report provided by the fine-tuning layer. The weight adjustment formula is as follows:
[0027]
[0028] In the formula: W k (new) represents the updated rule weight, W k (old) represents the rule weight before the update. Let R be the correlation index between rule R and the bottleneck in this round (given by the bottleneck report), and α be the learning rate. After adjustment, W is normalized.
[0029] The constraint tightness is relaxed. The fast filtering layer maintains a set of constraints C = {C1, C2, ..., Cm} and corresponding relaxation variables S = {S1, S2, ..., Sm}. When the quality of the solution set remains unsatisfactory, non-core constraints are relaxed to expand the search space. The initial value is 0, and it is updated according to the feedback as follows:
[0030]
[0031] In the formula: β is the relaxation step size, and Q is the quality score of the current round's scheme set;
[0032] Guided by the search space and based on the identification results of bottleneck devices, the fast filtering layer will prioritize allocating a wider set of available devices for processes involving bottleneck devices or reserving a longer buffer time when generating new solutions, thereby guiding the scheduling engine to explore new solutions that can effectively alleviate bottlenecks.
[0033] Furthermore, when the dynamic adjustment strategy is triggered, it prioritizes matching feasible scheduling schemes similar to the current disturbance scenario from the historical strategy pool as the initial point of iteration. Disturbance events include sudden equipment shutdown, material shortage, or emergency order insertion.
[0034] Furthermore, the system also includes a self-learning module, which records the events, schemes, simulation results, and actual execution effects in the dynamic adjustment strategy as decision cases; based on the accumulated case data, it periodically optimizes the constraints in the rule model or the optimization parameters of the scheduling engine to achieve continuous learning and performance evolution of the system.
[0035] This invention also discloses a digital twin workshop production scheduling decision support method that integrates virtual and real simulation.
[0036] Step 1: Construct a digital twin model that maps to the physical workshop entity at a preset ratio;
[0037] Step 2, Workshop data acquisition: Real-time acquisition of production data from the physical workshop through the Internet of Things sensor network, and processing and fusion of the data to drive the digital twin model to update its state and maintain dynamic consistency between the virtual space and the physical workshop.
[0038] Step 3: Initial scheduling scheme generation. In response to production plan changes or disturbance events, the scheduling engine generates an initial scheduling scheme and imports it into a simulation sandbox composed of the digital twin model. Based on the rule model and real-time data, the execution process of the scheme is simulated in the virtual space, and simulation results including estimated completion time, process sequencing and equipment load balancing are output to evaluate its feasibility.
[0039] Step 4: Dynamic adjustment of scheduling strategy. When the simulation results fail to meet the preset target or a sudden disturbance occurs in the physical workshop, the dynamic optimization process is triggered: the optimization strategy is adjusted to generate alternative solutions. In the simulation sandbox, the alternative solutions are subjected to rapid iterative simulation and comparative evaluation. At this time, the scheduling engine and the simulation sandbox work together to form a two-layer simulation optimization engine, including a rapid screening layer and a fine optimization layer. Through the two-layer screening, a feasible scheduling solution that meets the target is obtained and stored in the strategy pool.
[0040] Step 5, instruction conflict verification: Select the target scheduling scheme from the fine optimization layer in the policy pool, parse it into an atomic control instruction sequence, and perform conflict verification on the instruction sequence in the simulation sandbox. After the verification is successful, the final scheduling policy is generated.
[0041] Step 6: Issue scheduling instructions, sending the final scheduling strategy in the form of an instruction sequence to the corresponding execution unit in the physical workshop, and monitoring the execution status of the digital twin model and the physical workshop to form a closed-loop feedback.
[0042] Furthermore, in step 2, IoT sensors, equipment PLC controllers, RFID readers, and manufacturing execution system API interfaces deployed on the workshop site are integrated to collect multi-source heterogeneous data in real time or near real time. The multi-source heterogeneous data includes at least equipment operating status parameters, work-in-process and material identity and location information, production process progress data, and order execution status data.
[0043] It also includes a data preprocessing sub-step, which receives raw data from the heterogeneous data acquisition unit and performs data cleaning, noise reduction, outlier detection and format standardization on the data; at the same time, it assigns a unified semantic identifier to the processed data so that it establishes a mapping relationship with specific entity objects and attributes in the digital twin model.
[0044] It also includes a data synchronization sub-step, which, based on the preprocessed and semantically encoded data, updates the state, attributes, and location information of the corresponding entity objects in the digital twin model through the data interface provided by the digital twin model management module, according to a preset period or event triggering method, thereby driving the virtual model to keep synchronized with the physical workshop.
[0045] Furthermore, in step 3, the simulation sandbox is an isolated running instance of the digital twin model, supporting parallel simulation verification of multiple scheduling schemes. The evaluation of the schemes is based on a comprehensive objective function:
[0046]
[0047] In the formula: T is the maximum completion time, L is the equipment load balancing index, C is the penalty term for violating the constraint, and ω1, ω2, ω3 are adjustable weight coefficients.
[0048] Furthermore, step 4 specifically includes:
[0049] Step 4.1: When a disturbance is triggered, select feasible scheduling schemes similar to the current disturbance scenario from the historical strategy pool as the initial point for iteration. The sudden situation includes sudden equipment shutdown, material shortage or emergency order insertion.
[0050] Step 4.2: The fast screening layer uses a lightweight simulation model to quickly pre-screen a large number of initial solutions generated by the scheduling engine. The fast pre-screening uses the satisfaction of the primary constraint as the judgment criterion, and eliminates obviously infeasible or low-performance solutions within seconds, outputting a reduced set of candidate solutions. The fine optimization layer receives the set of candidate solutions from the fast screening layer and uses the high-fidelity simulation sandbox to perform full-process, high-precision simulation verification on each candidate solution in the set. This layer accurately evaluates and ranks the solutions based on the comprehensive objective function, and serves as the basis for subsequent dynamic adjustment of the scheduling strategy.
[0051] Step 4.3: When the fine optimization layer finds that the evaluation results of all candidate solutions are not ideal, it will feed back this result and bottleneck analysis to the fast screening.
[0052] Step 4.4: When the fine optimization layer outputs the average comprehensive evaluation value of the candidate solution set... Better than historical benchmark When the amplitude is less than the preset threshold δ, it satisfies...
[0053]
[0054] If the current set of candidate solutions is deemed unsatisfactory, the following dynamic adjustment process will be triggered:
[0055] The rule weights are adaptively adjusted. The fast filtering layer maintains a rule base R = {R1, R2, ..., Rn} and a corresponding weight vector W = {W1, W2, ..., Wn}, with initial weights being equal. The feedback mechanism dynamically adjusts the parameters of W based on the bottleneck analysis report provided by the fine-tuning layer. The weight adjustment formula is as follows:
[0056]
[0057] In the formula: W k (new) represents the updated rule weight, W k (old) represents the rule weight before the update. Let α be the correlation index between rule R and the bottleneck in this round, and let α be the learning rate. After adjustment, W is normalized.
[0058] The constraint tightness is relaxed. The fast filtering layer maintains a set of constraints C = {C1, C2, ..., Cm} and corresponding relaxation variables S = {S1, S2, ..., Sm}. When the quality of the solution set remains unsatisfactory, non-core constraints are relaxed to expand the search space. The initial value is 0, and it is updated according to the feedback as follows:
[0059]
[0060] In the formula: β is the relaxation step size, and Q is the quality score of the current round's scheme set;
[0061] Guided by the search space and based on the identification results of bottleneck devices, the fast filtering layer will prioritize allocating a wider set of available devices for processes involving bottleneck devices or reserving a longer buffer time when generating new solutions, thereby guiding the scheduling engine to explore new solutions that can effectively alleviate bottlenecks.
[0062] Step 4.5: Record the events, schemes, simulation results and actual execution effects in the scheduling decision-making process as decision cases; based on the accumulated case data, periodically optimize the constraints in the rule model or the optimization parameters of the scheduling engine to achieve continuous learning and performance evolution of the system.
[0063] Furthermore, in step 5, if a resource conflict or logic error is detected during the final conflict verification of the instruction sequence in the simulation sandbox, the issuance process is interrupted and an alarm is issued. At the same time, the dynamic optimization process is automatically triggered to start a new round of dynamic iterative optimization.
[0064] Beneficial effects:
[0065] (1) The digital twin workshop production scheduling decision system disclosed in this invention has agile dynamic response and strong anti-interference ability: the dynamic iterative optimization mechanism based on event triggering enables the system to quickly generate and verify new solutions after disturbance occurs, significantly shortening the scheduling decision time and improving the flexibility and robustness of workshop operation.
[0066] (2) The digital twin workshop production scheduling decision system disclosed in this invention is safe and controllable in operation, and the risks are effectively isolated: through the isolation verification of the simulation sandbox and the instruction-level security issuance mechanism, a "double insurance" for scheduling instructions is constructed, which effectively avoids the on-site operation risks caused by scheduling conflicts.
[0067] (3) The digital twin workshop production scheduling decision system disclosed in this invention has intelligent evolution and continuous decision optimization: Through knowledge accumulation and self-learning modules, the system can continuously learn from historical decisions, automatically optimize models and parameters, realize the continuous improvement of scheduling capabilities over time, and has long-term adaptability. Attached Figure Description
[0068] Figure 1 This is an overall flowchart of a digital twin workshop production scheduling decision-making system that integrates virtual and real simulation, according to an embodiment of the present invention.
[0069] Figure 2 This is a detailed flowchart of workshop data acquisition according to an embodiment of the present invention;
[0070] Figure 3 A detailed flowchart of the initial scheduling scheme generated according to an embodiment of the present invention;
[0071] Figure 4 A detailed flowchart illustrating the dynamic adjustment of the scheduling strategy according to an embodiment of the present invention;
[0072] Figure 5 This is a detailed flowchart of instruction conflict verification according to an embodiment of the present invention. Detailed Implementation
[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.
[0074] Example 1
[0075] This embodiment discloses a digital twin workshop production scheduling decision system integrating virtual and real simulation. The system applies rapidly iterative digital twin workshop scheduling and includes a physical workshop and a multi-dimensional digital twin model. The multi-dimensional digital twin model is constructed based on the geometry, physical attributes, and production behavior logic of the physical workshop, and is mapped 1:1 to the physical workshop entity. The system also includes the following modules:
[0076] The data acquisition and fusion module collects production data from the physical workshop in real time through an IoT sensor network, processes and fuses the data to drive the state update of the digital twin model and maintain dynamic consistency between the virtual space and the physical workshop. This data acquisition and fusion module further includes several sub-modules:
[0077] The heterogeneous data acquisition unit integrates IoT sensors, equipment PLC controllers, RFID readers, and manufacturing execution system API interfaces to acquire multi-source heterogeneous data in real-time or near real-time. This multi-source heterogeneous data includes at least: equipment operating status parameters, tool life, work-in-process and material identification and location information, production process progress data, and order execution status data.
[0078] The data preprocessing subunit cleans, denoises, detects outliers, and standardizes the format of the raw data. It also assigns a unified semantic identifier to the processed data, enabling it to establish a mapping relationship with the entity objects and attributes in the digital twin model.
[0079] The data synchronization subunit, based on preprocessed and semantically encoded data, updates the status, attributes, and location information of the corresponding entity objects in the digital twin model through the data interface provided by the digital twin model management module, according to a preset cycle or event triggering method, thereby driving the virtual model to keep synchronized with the physical workshop.
[0080] The scheduling engine, in response to changes in the production plan or disruptive events (such as the urgent insertion of orders), generates an initial scheduling plan based on a rule model and real-time data. The rule model includes constraints (such as a process must start after the preceding process is completed) and optimization objectives (such as minimizing completion time), while real-time data comes from the data acquisition module (such as the current utilization rate of equipment).
[0081] The simulation sandbox module, acting as an isolated running instance of the digital twin model, receives the initial scheduling scheme generated by the scheduling engine and performs parallel simulation verification in a virtual space, outputting simulation evaluation results based on a comprehensive objective function. The simulation sandbox supports parallel simulation verification of multiple scheduling schemes, with the scheme evaluation based on a comprehensive objective function:
[0082]
[0083] In the formula: T is the maximum completion time, L is the equipment load balancing index, C is the penalty term for violating the constraint, and ω1, ω2, ω3 are adjustable weight coefficients.
[0084] The dynamic optimization module, responding to changes in production plans or sudden disturbances, triggers an iterative optimization process. It includes a rapid screening layer and a fine-grained optimization layer. The rapid screening layer uses a lightweight model to pre-screen initial solutions, while the fine-grained optimization layer uses high-fidelity simulation to accurately evaluate candidate solutions and dynamically adjusts optimization parameters through a feedback mechanism until a feasible scheduling solution is obtained. Specifically, the dynamic adjustment of optimization parameters through the feedback mechanism involves…
[0085] When a disturbance is triggered, a feasible scheduling scheme similar to the current disturbance scenario is selected from the historical strategy pool as the initial point for iteration. The sudden situation includes sudden equipment shutdown, material shortage or emergency order insertion.
[0086] The fast screening layer employs a lightweight simulation model to rapidly pre-screen a large number of initial solutions generated by the scheduling engine. Fast pre-screening uses primary constraint satisfaction as the criterion, eliminating obviously infeasible or low-performance solutions within seconds, and outputting a reduced set of candidate solutions. The fine-tuning layer receives the candidate solution set from the fast screening layer and uses a high-fidelity simulation sandbox to perform full-process, high-precision simulation verification on each candidate solution in the set. This layer accurately evaluates and ranks the solutions based on the comprehensive objective function, serving as the basis for subsequent dynamic adjustments to the scheduling strategy.
[0087] When the fine optimization layer finds that the evaluation results of all candidate solutions are unsatisfactory, it will feed this result and bottleneck analysis back to the fast screening.
[0088] When the fine optimization layer outputs the average comprehensive evaluation value of the candidate solution set Better than historical benchmark When the amplitude is less than the preset threshold δ, the following condition is met:
[0089]
[0090] If the current set of candidate solutions is deemed unsatisfactory, the following dynamic adjustment process will be triggered:
[0091] The rule weights are adaptively adjusted. The fast filtering layer maintains a rule base R = {R1, R2, ..., Rn} and a corresponding weight vector W = {W1, W2, ..., Wn}, with initial weights being equal. The feedback mechanism dynamically adjusts the parameters of W based on the bottleneck analysis report provided by the fine-tuning layer. The weight adjustment formula is as follows:
[0092]
[0093] In the formula: Let α be the correlation index between rule R and the bottleneck in this round, and let α be the learning rate. After adjustment, W is normalized.
[0094] The constraint tightness is relaxed. The fast filtering layer maintains a set of constraints C = {C1, C2, ..., Cm} and corresponding relaxation variables S = {S1, S2, ..., Sm}. When the quality of the solution set remains unsatisfactory, non-core constraints are relaxed to expand the search space. The initial value is 0, and it is updated according to the feedback as follows:
[0095]
[0096] In the formula: β is the relaxation step size, and Q is the quality score of the current round's scheme set;
[0097] Guided by the search space and based on the identification results of bottleneck devices, the fast filtering layer will prioritize allocating a wider set of available devices for processes involving bottleneck devices or reserving a longer buffer time when generating new solutions, thereby guiding the scheduling engine to explore new solutions that can effectively alleviate bottlenecks.
[0098] The instruction verification module is used to parse the optimization scheme into an atomic control instruction sequence, verify resource conflicts and logic errors in the simulation sandbox, and generate the final scheduling strategy after the verification is passed.
[0099] The distribution and monitoring module distributes the final scheduling strategy to the execution unit of the physical workshop in the form of an instruction sequence, and monitors the execution status of the digital twin model and the physical workshop to form a closed-loop feedback control.
[0100] In this embodiment, the system also includes a self-learning module, which records events, schemes, simulation results and actual execution effects in the scheduling decision-making process as decision cases; based on the accumulated case data, the system periodically automatically optimizes the constraints in the rule model or the optimization parameters of the scheduling engine to achieve continuous learning and performance evolution of the system.
[0101] In this embodiment, the 3D modeling and rendering engine used for digital twin modeling is Unity3D, which is used to construct and display the geometric model of the workshop. The data platform uses the Industrial Internet of Things platform (ThingsBoard), which is responsible for the collection, cleaning, and storage of multi-source heterogeneous data. The simulation kernel uses the engine of discrete event simulation software, embedded in the system as the computing core of the simulation sandbox. The iterative optimization algorithm layer uses integrated open-source optimization libraries (such as OR-Tools), genetic algorithms (GA), particle swarm optimization (PSO), etc.
[0102] Example 2
[0103] This embodiment discloses a digital twin workshop production scheduling decision-making method that integrates virtual and real simulation, such as... Figure 1 As shown, it includes the following steps:
[0104] The entire scheduling decision-making process begins with proactive plan changes (such as new order arrivals or plan adjustments) or passive sudden disturbances (such as equipment failures, material shortages, or emergency order insertions). The system's response mechanism is activated by taking dynamic changes from the physical workshop or management demands from the planning layer as input.
[0105] Step 1: Establish a digital twin model. By collecting real-time data on the physical workshop's status and static processes through a sensor network, a dynamic, high-fidelity model synchronized with the physical workshop is constructed in virtual space. For example... Figure 2 The diagram illustrates the detailed process of workshop data acquisition and digital twin model status updates. The input raw data originates from heterogeneous acquisition devices such as IoT sensors and PLC controllers. The data preprocessing sub-steps are executed sequentially, including data cleaning, data denoising, outlier detection, format standardization, and data semanticization. Then, the data synchronization sub-step determines the triggering conditions: if an event trigger is met or a preset cycle is reached, the twin model is updated immediately; otherwise, the model is updated periodically. Finally, the model status update is completed, achieving dynamic consistency between the virtual and physical workshops.
[0106] Step 2: The scheduling engine generates an initial scheduling scheme. Based on the current workshop status, production targets (such as delivery time and efficiency), and preset rules, the system automatically calculates and generates one or more preliminary production operation plans (scheduling schemes). The initial scheduling scheme is then imported into a simulation sandbox composed of a digital twin model for execution.
[0107] Step 3: Simulation Sandbox Virtual-Real Simulation Verification. Based on the rule model and real-time data, the execution process of the scheduling scheme is simulated in a virtual space, and simulation results including estimated completion time, process sequencing, and equipment load balancing are output to evaluate its feasibility. The detailed process of initial scheduling scheme generation and feasibility verification is as follows: Figure 3 As shown, this embodiment employs parallel verification, meaning multiple independent sandboxes (Instance 1, Instance 2, etc.) can be created simultaneously to perform synchronous simulations and deductions on different initial schemes or different parameters of the same scheme. The feasibility of the scheme is judged based on feasibility evaluation indicators, including maximum completion time, equipment load balancing index, and detection of constraint violations. In this embodiment, the scheme evaluation is based on a comprehensive objective function:
[0108]
[0109] In the formula: T is the maximum completion time, L is the equipment load balancing index, C is the penalty term for violating the constraint, and ω1, ω2, ω3 are adjustable weight coefficients.
[0110] If the assessment is feasible, the simulation results data will be output and the solution will be stored in the policy pool, and then proceed to the instruction conflict verification stage; if it is not feasible, then dynamic adjustment will be initiated.
[0111] Step 4: Dynamically adjust the scheduling strategy. Adjust and optimize the strategy to generate alternative solutions. Perform rapid iterative simulation and comparative evaluation of these alternative solutions in the simulation sandbox. At this point, the scheduling engine and the simulation sandbox work together to form a two-layer simulation optimization engine, including a rapid screening layer and a fine-grained optimization layer. Through this two-layer screening, a feasible scheduling solution that meets the objective is obtained and stored in the strategy pool. Figure 4 As shown, when a sudden situation occurs in the physical workshop or the simulation results fail to meet the standards, a dynamic optimization process is triggered. First, feasible scheduling schemes for similar scenarios are matched from the strategy pool as the initial iteration point. Then, the scheduling engine generates a large number of candidate schemes, which are input into a two-layer simulation optimization engine for screening and optimization: The rapid screening layer uses a lightweight simulation model, prioritizing primary constraint satisfaction, to quickly screen within seconds and output a set of candidate schemes; the fine-tuning layer uses a high-fidelity simulation sandbox to perform full-process high-precision simulation verification of the candidate schemes, accurately evaluating them based on the objective function. Finally, it determines whether the evaluation result is ideal. If ideal, the feasible scheme is stored in the strategy pool; if not, a bottleneck analysis report is output, and a dynamic adjustment mechanism is initiated. This includes adaptive adjustment of rule weights, relaxation of constraint tightness, and search space guidance to allocate more resources to the bottleneck process. After adjustment, the scheme re-enters the two-layer simulation optimization engine. The process executes in parallel, recording decision cases and optimizing the rule model and engine parameters based on the case data, ultimately achieving continuous performance evolution of the system.
[0112] Specifically,
[0113] Step 4.1: When a disturbance is triggered, select feasible scheduling schemes similar to the current disturbance scenario from the historical strategy pool as the initial point for iteration. The sudden situation includes sudden equipment shutdown, material shortage or emergency order insertion.
[0114] Step 4.2: The fast screening layer uses a lightweight simulation model to quickly pre-screen a large number of initial solutions generated by the scheduling engine. The fast pre-screening uses the satisfaction of the primary constraint as the judgment criterion, and eliminates obviously infeasible or low-performance solutions within seconds, outputting a reduced set of candidate solutions. The fine optimization layer receives the set of candidate solutions from the fast screening layer and uses the high-fidelity simulation sandbox to perform full-process, high-precision simulation verification on each candidate solution in the set. This layer accurately evaluates and ranks the solutions based on the comprehensive objective function, and serves as the basis for subsequent dynamic adjustment of the scheduling strategy.
[0115] Step 4.3: When the fine optimization layer finds that the evaluation results of all candidate solutions are not ideal, it will feed back this result and bottleneck analysis to the fast screening.
[0116] Step 4.4, when the fine optimization layer outputs the average comprehensive evaluation value of the candidate solution set. Better than historical benchmark When the amplitude is less than the preset threshold δ, the following condition is met:
[0117]
[0118] If the current set of candidate solutions is deemed unsatisfactory, the following dynamic adjustment process will be triggered:
[0119] The rule weights are adaptively adjusted. The fast filtering layer maintains a rule base R = {R1, R2, ..., Rn} and a corresponding weight vector W = {W1, W2, ..., Wn}, with initial weights being equal. The feedback mechanism dynamically adjusts the parameters of W based on the bottleneck analysis report provided by the fine-tuning layer. The weight adjustment formula is:
[0120]
[0121] In the formula: Let R be the correlation index between rule R and the bottleneck in this round, and α be the learning rate. W is then normalized after adjustment.
[0122] The constraint tightness is relaxed. The fast filtering layer maintains a set of constraints C = {C1, C2, ..., Cm} and corresponding relaxation variables S = {S1, S2, ..., Sm}. When the quality of the solution set remains unsatisfactory, non-core constraints are relaxed to expand the search space. The initial value is 0, and it is updated according to the feedback as follows:
[0123]
[0124] In the formula: β is the relaxation step size, and Q is the quality score of the current round of schemes (i.e., the average comprehensive evaluation value of all feasible scheduling schemes in this round of iteration). The quality score obtained after normalization.
[0125] Guided by the search space and based on the identification results of bottleneck devices, the fast filtering layer will prioritize allocating a wider set of available devices for processes involving bottleneck devices or reserving a longer buffer time when generating new solutions, thereby guiding the scheduling engine to explore new solutions that can effectively alleviate bottlenecks.
[0126] Step 4.5: Record the events, schemes, simulation results and actual execution effects in the scheduling decision-making process as decision cases; based on the accumulated case data, periodically optimize the constraints in the rule model or the optimization parameters of the scheduling engine to achieve continuous learning and performance evolution of the system.
[0127] Step 5, Instruction Conflict Verification: Select the target scheduling scheme from the fine-tuning layer in the policy pool, parse it into an atomic control instruction sequence, and perform conflict verification on the instruction sequence in the simulation sandbox. Conflict verification includes resource conflict detection and logical sequence verification. After successful verification, the final scheduling policy is generated.
[0128] The detailed process for instruction conflict verification is as follows: Figure 5 As shown, a target scheduling scheme is selected from the fine-grained optimization layer of the strategy pool. The scheduling parser decomposes the scheduling scheme into three types of atomic instructions and generates an atomic instruction sequence by sorting them according to execution time. Subsequently, the instruction verifier performs relaxation constraint checks in the simulation sandbox: it verifies the time window, process duration, material availability, and equipment capacity for each instruction; and performs core constraint checks on the entire instruction sequence, including equipment exclusivity and process sequence. The verifier summarizes all violation records and generates a verification report: if all core constraints pass, the final scheduling strategy is generated and the scheduling instructions are issued; if there is a violation of the core constraints, the issuance is interrupted and a detailed conflict log is output, while the dynamic optimization process is automatically triggered and returned. Figure 4 The dynamic adjustment process will initiate a new round of optimization.
[0129] This invention, through the aforementioned system and method, creatively integrates digital twins and simulation sandboxes to construct a dynamic decision support solution that systematically addresses the problems of information lag and high trial-and-error costs in static scheduling during workshop production scheduling. Compared with existing technologies, this invention enables simulation verification of the feasibility of scheduling schemes, solving the problem that traditional static scheduling cannot predict execution consequences; through a dynamic iterative optimization engine, it significantly improves the rapid response and recovery capabilities to sudden disturbances such as equipment downtime and material shortages; and through an instruction-level security verification mechanism, it achieves a secure closed-loop control from virtual decision-making to physical execution, eliminating the risk of scheduling conflicts. In summary, this invention is beneficial for improving the intelligence level, decision-making foresight, and system robustness of workshop production scheduling, effectively replacing the rigid and high-risk traditional scheduling model.
Claims
1. A digital twin workshop production scheduling decision-making system integrating virtual and real simulation, comprising a physical workshop and a multi-dimensional digital twin model mapped to the physical workshop entity, characterized in that, Includes the following modules: The data acquisition and fusion module collects production data from the physical workshop in real time through the Internet of Things sensor network, processes and fuses the data to drive the state update of the digital twin model and maintain the dynamic consistency between the virtual space and the physical workshop. The scheduling engine responds to changes in production plans or disturbance events by generating an initial scheduling scheme based on a rule model and real-time data. The simulation sandbox module, as an isolated running instance of the digital twin model, is used to receive the initial scheduling scheme generated by the scheduling engine, perform parallel simulation verification in the virtual space, and output simulation evaluation results based on the comprehensive objective function. The dynamic optimization module triggers an iterative optimization process when verification is unsatisfactory or when there are sudden disturbances. This process includes a fast screening layer and a fine optimization layer. The fast screening layer uses a lightweight model to pre-screen the initial schemes, while the fine optimization layer accurately evaluates the candidate schemes based on high-fidelity simulation and dynamically adjusts the optimization parameters through a feedback mechanism until a feasible scheduling scheme is obtained. The instruction verification module is used to parse the optimization scheme into an atomic control instruction sequence, perform resource conflict and logic error verification in the simulation sandbox, and generate the final scheduling strategy after the verification is passed. The distribution and monitoring module distributes the final scheduling strategy to the execution unit of the physical workshop in the form of an instruction sequence, and monitors the execution status of the digital twin model and the physical workshop to form a closed-loop feedback control.
2. The digital twin workshop production scheduling decision system according to claim 1, characterized in that, The data acquisition and fusion module specifically includes: The heterogeneous data acquisition unit integrates IoT sensors, equipment PLC controllers, RFID readers and writers and manufacturing execution system API interfaces to collect equipment operating status parameters, work-in-process and material identity and location information, production process progress data and order execution status data in real time or near real time. The data preprocessing subunit cleans, denoises, detects outliers, and standardizes the format of the raw data. It also assigns a unified semantic identifier to the processed data, enabling it to establish a mapping relationship with the entity objects and attributes in the digital twin model. The data synchronization subunit updates the status, attributes, and location information of the corresponding entity objects in the digital twin model according to a preset cycle or event triggering method.
3. The digital twin workshop production scheduling decision system according to claim 1, characterized in that, The simulation sandbox module supports parallel simulation verification of multiple scheduling schemes, and the evaluation of the schemes is based on a comprehensive objective function: ， In the formula, T is the maximum completion time, L is the equipment load balancing index, C is the penalty term for violating the constraint, and ω1, ω2, ω3 are adjustable weight coefficients.
4. The digital twin workshop production scheduling decision system according to claim 1, characterized in that, In the dynamic optimization module, the feedback mechanism is as follows: when the fine optimization layer finds that the evaluation results of all candidate solutions are not ideal, it feeds back this result and bottleneck analysis to the fast screening layer. The fast screening layer triggers a dynamic adjustment strategy, including adaptive adjustment of rule weights, relaxation of constraint tightness, or search space guidance, so as to guide the next iteration to generate a more promising set of candidate solutions.
5. The digital twin workshop production scheduling decision system according to claim 4, characterized in that, In the feedback mechanism, the evaluation results are not ideal when the average comprehensive evaluation value of the candidate solution set output by the fine optimization layer is not satisfactory. Better than historical benchmark When the amplitude is less than the preset threshold δ, the following condition is met: If the current candidate solution set is not of ideal quality, then it is determined that the quality of the current candidate solution set is not ideal.
6. The digital twin workshop production scheduling decision system according to claim 4 or 5, characterized in that, The dynamic adjustment strategy includes: The rule weights are adaptively adjusted. The fast filtering layer maintains a rule base R = {R1, R2, ..., Rn} and a corresponding weight vector W = {W1, W2, ..., Wn}, with initial weights being equal. The feedback mechanism dynamically adjusts the parameters of W based on the bottleneck analysis report provided by the fine-tuning layer. The weight adjustment formula is as follows: ， In the formula: W k (new) represents the updated rule weight, W k (old) represents the rule weight before the update. Let R be the correlation index between rule R and the bottleneck in this round, and α be the learning rate; after adjustment, W is normalized. The constraint tightness is relaxed. The fast filtering layer maintains a constraint set C = {C1, C2, ..., Cm} and corresponding relaxation variables S = {S1, S2, ..., Sm}. When the quality of the solution set remains unsatisfactory, non-core constraints are relaxed to expand the search space. The initial value is 0, and it is updated according to the feedback as follows: ， In the formula: β is the relaxation step size, and Q is the average comprehensive evaluation value of all feasible scheduling schemes in this iteration. The quality score obtained after normalization; Guided by the search space and based on the identification results of bottleneck devices, the fast filtering layer will prioritize allocating a wider set of available devices for processes involving bottleneck devices or reserving a longer buffer time when generating new solutions, thereby guiding the scheduling engine to explore new solutions that can effectively alleviate bottlenecks.
7. The digital twin workshop production scheduling decision system according to claim 4, characterized in that, When the dynamic adjustment strategy is triggered, a feasible scheduling scheme similar to the current disturbance scenario is matched from the historical strategy pool as the initial point for iteration. The disturbance events include sudden equipment shutdown, material shortage, or emergency order insertion.
8. The digital twin workshop production scheduling decision system according to claim 4, characterized in that, The system also includes a self-learning module, which records events, schemes, simulation results, and actual execution effects in the dynamic adjustment strategy as decision cases; based on the accumulated case data, it automatically optimizes the constraints in the rule model or the optimization parameters of the scheduling engine on a regular basis to achieve continuous learning and performance evolution of the system.
9. A digital twin workshop production scheduling decision-making method integrating virtual and real simulation, characterized in that, Includes the following steps: Step 1: Construct a digital twin model that maps to the physical workshop entity at a preset ratio; Step 2: Real-time collection of production data from the physical workshop via IoT sensor network, processing and fusion of the data, and driving the digital twin model to update its status, maintaining dynamic consistency between the virtual space and the physical workshop. Step 3: In response to changes in the production plan or disturbance events, the scheduling engine generates an initial scheduling plan and imports the initial plan into the simulation sandbox for execution. Virtual simulation is performed based on the rule model and real-time data, and the simulation results are output to evaluate the feasibility. Step 4: When the simulation results fail to meet the preset target or a sudden disturbance occurs, the dynamic iterative optimization process is triggered: alternative solutions are quickly generated and verified in the simulation sandbox, and then screened and evaluated through a two-layer simulation optimization engine until a feasible scheduling solution is obtained. Step 5: Select the target scheme from the optimization schemes, parse it into an atomic control instruction sequence, perform conflict verification in the simulation sandbox, and generate the final scheduling strategy after the verification is passed. Step 6: The final scheduling strategy is issued to the physical workshop execution unit in the form of a sequence of instructions, and the execution status is monitored to form a closed-loop feedback.
10. The digital twin workshop production scheduling decision-making method according to claim 9, characterized in that, In step 4, the dynamic iterative optimization process includes: prioritizing matching similar scenario solutions from the historical strategy pool as the initial point of iteration; performing lightweight pre-screening through a fast screening layer to output a candidate solution set; performing high-precision simulation evaluation through a fine optimization layer, and triggering feedback adjustment when the solution set quality is not ideal to dynamically adjust the rule weights or constraint tightness.