Factory interaction method and system based on digital twinning

Abstract

The invention relates to the technical field of smart factories, in particular to a digital twinning-based factory interaction method and system, and the method comprises a data fusion step, a construction step, an intelligent optimization step, an interaction verification step and an execution step. The system comprises a processor and a memory in communication connection with the processor. According to the technical scheme, an operator can intuitively evaluate the comprehensive influence of different decisions, so that the scheme is screened and perfected before actual execution.

Description

Technical Field

[0001] This application relates to the technical field of smart factories, and in particular to a factory interaction method and system based on digital twins. Background Technology

[0002] Digital twins, as a key enabling technology for achieving two-way mapping and dynamic interaction between the physical world and virtual information space, are profoundly transforming the traditional factory operation and management model. By constructing a high-precision model of a physical entity in virtual space and using real-time data to drive its operation, digital twin technology provides unprecedented possibilities for transparent monitoring, process simulation, and performance optimization of factories.

[0003] Existing digital twin-based factory management systems typically follow a basic paradigm of "data acquisition - model building - visualization." Specifically, they first collect operational data such as equipment status through sensor networks, and then combine this data with a 3D CAD model of the factory to build a static or low-frequency updated digital twin scenario. Based on this, the system provides production data visualization functions based on 2D dashboards or 3D scenarios. Some advanced systems integrate discrete event simulation modules for offline verification and bottleneck analysis of production plans.

[0004] However, existing technologies suffer from the following shortcomings: they lack a closed-loop interactive mechanism capable of real-time perception, concurrent analysis, collaborative optimization, and multimodal intervention of the factory's operational status across all dimensions (such as production efficiency, equipment health, energy consumption, and quality control). This results in optimization decisions often being localized and delayed, and it is difficult to verify their comprehensive effects before actual implementation. Furthermore, it cannot address the multi-objective, strongly coupled dynamic optimization needs of complex production systems, thus limiting the development of digital twin technology towards enabling deeper intelligent decision-making. Summary of the Invention

[0005] This application provides a factory interaction method and system based on digital twins, which can at least partially solve the above-mentioned technical problems.

[0006] Firstly, this application provides a factory interaction method based on digital twins, which adopts the following technical solution: A factory interaction method based on digital twins includes the following steps: Data fusion: Real-time collection of equipment operating parameters and environmental sensing data through the Industrial Internet of Things, and integration of factory CAD drawings, production process data and quality inspection data, followed by cleaning, alignment and fusion, to build a twin data pool with a unified spatiotemporal benchmark; Construction: Based on the aforementioned digital twin data pool, a multi-scale modeling method integrating geometric, physical, and behavioral rules is used to construct a digital twin of the factory; the digital twin is then driven to be updated synchronously through real-time data streams. Intelligent optimization: On the factory digital twin, production process simulation and equipment health status prediction simulation are executed concurrently; the simulation results and real-time data are input into a multi-objective optimization algorithm, with production efficiency, equipment reliability and resource utilization as optimization objectives, to generate a collaborative optimization scheme, which includes production scheduling strategy, predictive maintenance plan and process parameter adjustment suggestions; Interactive verification: The twin, optimization scheme and key indicators are visualized in three dimensions through a VR / AR interactive interface; it supports operators to simulate operation and parameter adjustment of the optimization scheme in a natural interactive way, and to deduce the multi-dimensional performance changes after adjustment in real time in the twin; Execution: Based on the validated optimization scheme, control instructions or decision guidance are generated and issued to the physical plant for execution.

[0007] By adopting the above technical solution and driven by real-time data streams, the twin can serve as a dynamic mirror of the physical factory, providing a high-fidelity environment for online analysis. Based on this, concurrent execution of production process and equipment health simulations allows for simultaneous assessment of efficiency and reliability. Combined with collaborative optimization schemes generated by multi-objective optimization algorithms, multiple conflicting objectives can be coordinated, providing comprehensive improvement strategies. Interactive verification steps allow personnel to deduce and adjust solutions in a virtual environment, intuitively evaluating the multi-dimensional impact of different decisions, thereby reducing the risk of directly executing infeasible solutions. Finally, the execution steps complete the closed loop from virtual decision-making to physical execution.

[0008] Optionally, it also includes historical deviation accumulation modeling: in the twin data pool, the deviation data between the actual execution results of each optimization scheme and the simulation prediction value is continuously recorded, and a cumulative deviation knowledge graph is constructed and updated according to the dimensions of equipment, process section and product batch. This graph represents the systematic error trend and uncertainty distribution of the simulation model under specific production conditions. Based on the cumulative deviation knowledge graph, the parameters of the corresponding equipment or process model in the digital twin are adaptively calibrated and labeled with confidence level.

[0009] By adopting the above technical solutions, the system can quantify and visualize the prediction error trend and uncertainty distribution of the model under different production conditions, transforming one-time simulation comparisons into systematic knowledge that can be accumulated and queried. Based on this map, the parameters of the digital twin model are adaptively calibrated and confidence level labeled, enabling the model to self-correct using historical experience and gradually improve its prediction accuracy in specific scenarios. At the same time, the confidence level labeling provides a basis for downstream steps to identify reliable areas and weak links in the model.

[0010] Optionally, in the intelligent optimization step, a cumulative deviation constraint term is introduced into the optimization objective function of the multi-objective optimization algorithm. This constraint term is weighted according to the historical uncertainty of the corresponding decision path in the cumulative deviation knowledge graph to penalize highly volatile solutions.

[0011] By adopting the above technical solution, the optimization algorithm no longer merely pursues the theoretical optimal solution under an ideal model. Instead, it proactively avoids decision paths that have historically exhibited high volatility—that is, where actual results deviate significantly from predictions—by introducing a weighted cumulative deviation constraint term. This allows the generated collaborative optimization scheme to consider not only performance indicators but also the stability and predictability of the scheme, thereby significantly improving the success rate and robustness of the optimization scheme in actual production environments and achieving a transformation from "ideal optimization" to "robust optimization."

[0012] Optionally, it also includes: Fluctuation monitoring: Systemic high-deviation-risk locations identified in the cumulative deviation knowledge graph are defined as key areas of concern; the physical equipment operating parameters or process indicators in these areas are monitored in real time, their short-term fluctuation values ​​are calculated and compared with a preset dynamic fluctuation threshold; when the short-term fluctuation value exceeds the dynamic fluctuation threshold three times in a row, a real-time warning is generated.

[0013] The above technical solution enables continuous and quantitative monitoring of key weak links. By comparing short-term fluctuation values ​​with dynamic thresholds, the system can sensitively detect abnormal signs in the physical entity's operating status and issue timely alarms when fluctuations continuously exceed the threshold. This is equivalent to adding a layer of "microscope"-like focused monitoring for known risk points on top of the global monitoring of the digital twin, transforming offline deviation analysis capabilities into online proactive risk perception capabilities, and gaining a time window for early intervention.

[0014] Optionally, it also includes: Causal tracing: When a high-fluctuation concern point A continuously triggers an early warning, its fluctuation time series is analyzed; if the linear fitting slope of its fluctuation mean is consistently positive or negative and passes the significance test, it is determined that there is a directional drift trend; then, in the cumulative deviation knowledge graph, along the reverse transmission path of material flow, energy flow, or information flow, the direct upstream node B that has a leading-lag relationship with the fluctuation trend of point A in time series and whose own volatility increases in the same period is identified as the suspected root cause node causing the trend fluctuation of point A.

[0015] By adopting the above technical solutions, the system has achieved an intelligent diagnostic leap from "monitoring abnormal phenomena" to "preliminarily locating potential causes"; it no longer simply reports a problem at a certain point, but can automatically infer the direct upstream root cause that may lead to the trend fluctuation along the production logic chain, providing clear intervention targets for subsequent accurate backup.

[0016] Optionally, in the causal tracing step, the fluctuation pattern of the identified suspected root cause node B is further analyzed: if the fluctuation of B is characterized by random high fluctuation without direction, it is marked as a first-level correction target; if the fluctuation of B also shows a directional drift trend in the same direction as A, the tracing continues upstream until a node with a fluctuation pattern of random high fluctuation is found and it is marked as a first-level correction target.

[0017] By adopting the above technical solutions, the intelligence and accuracy of diagnosis are increased. By distinguishing whether the fluctuation of a node is "undirected random high fluctuation" or "directional drift," the system can decide whether the tracing should stop. If it is random high fluctuation, it is determined that the point itself may have a stability problem and should be the final correction target. If it is directional drift, it indicates that the point may also be the result of influence from upstream sources and needs to be traced further. This discrimination mechanism ensures that the tracing process can lock onto the real, original stability defect node and avoid erroneous intervention in intermediate transmission links.

[0018] Optionally, for nodes marked as primary correction targets, a directional deterministic compensation is initiated in the operational parameter settings of their digital twin models: If the target node is a device, then adjust the set value of its key control parameters by a fixed percentage in the opposite direction of its current fluctuation, or adjust it to the median value of its historical stable operating range. If the target node is a process parameter, then revise the standard operating procedure value of that parameter to the historical average value used during the most recent period of stable output of that node.

[0019] By adopting the above technical solutions, clear and automated primary intervention methods are provided. For equipment, rules are used to reverse the adjustment or reset to historical stable values; for process parameters, the settings are rolled back to the most recent successful settings. These rules are derived from engineering experience and are characterized by clear logic and rapid response. They can directly and quickly correct the identified root cause of instability, attempting to suppress the generation of fluctuations at the source.

[0020] Optionally, it also includes: Buffer: While taking corrective measures for the suspected root cause node B, the following minor adjustments are made to the initial high volatility concern point A and its direct downstream node C, which are affected by the volatility trend: In the simulation logic of the digital twin, the acceptable range for the output quality of point A is temporarily relaxed; Slightly reduce the production cycle time or feed rate of downstream node C to allow for buffer time to handle potential unconventional inputs from point A.

[0021] The adoption of the above technical solutions embodies a systematic intervention strategy. While attempting to "address the root cause" (correct the underlying problem), it also addresses the "symptoms" (buffer the impact). By temporarily relaxing the quality assessment range and reserving buffer time for downstream processes, this step aims to absorb or attenuate the immediate impact of fluctuations on the production chain during their propagation, protecting downstream links and maintaining the overall continuity of the production flow. This reduces the risk that local fluctuations may evolve into systemic paralysis before the root cause is fully corrected.

[0022] Optionally, it also includes: Priority reallocation: During a preset observation period after the execution of the directional deterministic compensation and preventive buffer adjustment, the fluctuation data of node B, the output quality data of node A, and the production smoothness data of node C in the physical plant are collected simultaneously. Based on the volatility data of node B, its short-term statistical volatility is calculated and compared with the benchmark volatility before execution to generate the root cause correction effectiveness coefficient. Based on the output quality data of node A and the relaxed judgment range, its pass rate is calculated and compared with the historical pass rate before the warning is triggered to generate a quality buffer adaptation coefficient. The product of the root cause correction effectiveness coefficient and the quality buffer adaptation coefficient is defined as the comprehensive handling efficiency index for this chain of fluctuation events. Based on the magnitude of the comprehensive handling efficiency index, the risk priority labels of nodes A, B, and C involved in this event in the cumulative deviation knowledge graph are dynamically recalibrated: if the efficiency index is higher than the high threshold, its priority is reduced; if the efficiency index is lower than the low threshold, its priority is increased. This recalibration result is fed back to the dynamic fluctuation monitoring and threshold early warning step to adjust the focus of subsequent monitoring resource allocation and the sensitivity of the early warning threshold.

[0023] By adopting the above technical solutions, a closed-loop learning and adaptive management mechanism based on practical feedback is formed. The system can quantitatively evaluate the actual effect of each intervention action and feed this effect back to the knowledge base to adjust the allocation of future monitoring resources (such as reducing the monitoring intensity of effective treatment points and increasing the attention to ineffective treatment points) and the sensitivity of the early warning threshold. This enables the entire system to have the ability to self-optimize and self-evolve, and to allocate resources and define risks more and more accurately over time.

[0024] Secondly, this application provides a factory interaction system based on digital twins, which adopts the following technical solution: A factory interaction system based on digital twins includes: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes the computer program stored on the computer-readable storage medium, it implements a factory interaction method based on digital twins.

[0025] In summary, this application includes at least one of the following beneficial technical effects: 1. A coherent workflow has been constructed, encompassing real-time fusion of multi-source data, high-fidelity digital modeling, concurrent simulation and multi-objective optimization, immersive interactive verification, and finally, instruction issuance and execution. This enables factory optimization decisions to move away from offline, fragmented analysis and instead allow for global simulation and verification in a virtual environment synchronized with the physical world. Operators can intuitively assess the combined impact of different decisions, thereby selecting and refining solutions before actual execution. This reduces over-reliance on human experience and minimizes quality risks and unplanned downtime caused by inappropriate decisions directly impacting the production line, making the factory's operational optimization process more systematic, controllable, and reliable. 2. By continuously accumulating historical data and constructing a deviation knowledge graph, the system can automatically identify systematic errors and uncertainties in model predictions. This knowledge is not only used to dynamically calibrate the digital twin model itself and improve its prediction fidelity under specific working conditions, but is also deeply applied to guide future optimization decisions and resource allocation. Based on historical performance feedback, the system can automatically adjust the monitoring intensity and early warning thresholds for different production links and optimize its diagnostic logic. The accuracy of its model and the adaptability of its decisions continuously improve over time.

[0026] 3. Simultaneously implement a combined strategy of "source correction" and "impact buffering": On the one hand, make targeted adjustments to the root cause parameters, and on the other hand, buffer the parameters of related processes. This systematic intervention aims to quickly suppress the source of disturbance, absorb the energy of fluctuations, and prevent local problems from spreading along the production line. After long-term operation, the system can accumulate experience in dealing with various fluctuations, thereby improving the inherent resilience of the production system to resist internal and external interference and maintain stable and efficient operation. Attached Figure Description

[0027] Figure 1 This is a flowchart of the factory interaction method in Embodiment 1 of this application; Figure 2 This is a flowchart of the factory interaction method in Embodiment 2 of this application; Figure 3 This is a flowchart of the factory interaction method in Embodiment 3 of this application; Figure 4 This is a flowchart of the factory interaction method in Embodiment 4 of this application. Detailed Implementation

[0028] The following combination Figures 1 to 4 This application will be described in further detail.

[0029] This embodiment discloses a factory interaction method based on digital twins.

[0030] Example 1: Refer to Figure 1 A factory interaction method based on digital twins includes the following steps: Data fusion: Real-time collection of equipment operating parameters and environmental sensing data through the Industrial Internet of Things, and integration of factory CAD drawings, production process data and quality inspection data, followed by cleaning, alignment and fusion, to build a twin data pool with a unified spatiotemporal benchmark; Specifically, industrial IoT sensor networks deployed on the production site collect equipment operating parameters in real time, including but not limited to spindle speed. (Unit: rpm), operating temperature (Unit: °C) Vibration amplitude (Unit: mm / s) and current (Unit: A). Environmental sensing data is collected by temperature and humidity sensors and particulate matter sensors distributed throughout the workshop, and is denoted as environmental vector. Simultaneously, the CAD 3D model files of the equipment are imported from the factory's PLM (Product Lifecycle Management) system, and the currently executing production process data is read from the MES (Manufacturing Execution System). (such as feed rate F, depth of cut D) and quality inspection data obtained from the online inspection station. (such as dimensional error) Surface roughness The data cleaning process includes removing abnormal jump values ​​from sensors (using the Raida criterion, for values ​​exceeding...). Data points within a certain range are considered outliers and removed, and transient missing values ​​caused by network outages are filled in using linear interpolation. The alignment process involves adding a uniform timestamp to all data. (Source: Factory-grade NTP time server) and spatial tags (Equipment or workstation number) is used for data fusion. The fusion process will combine the processed data according to timestamps. and space tags For indexing, data is integrated and stored in a time-series database, forming a twin data pool with a unified spatiotemporal benchmark. .

[0031] Construction: Based on the aforementioned digital twin data pool, a multi-scale modeling method integrating geometric, physical, and behavioral rules is used to construct a digital twin of the factory; the digital twin is then driven to be updated synchronously through real-time data streams. Specifically, based on the twin data pool DP, a multi-scale modeling method is used to construct the factory digital twin DT; geometric modeling utilizes CAD files and generates a 3D mesh model suitable for real-time rendering through lightweight processing; physical modeling assigns physical properties to key equipment, for example, defining mass m, stiffness k, and damping c parameters for the machine tool spindle, whose motion equations can be simplified to... Where F(t) is determined according to the processing technology Calculated load capacity; behavioral rule modeling defines state transition rules by parsing the PLC control logic code and production process flow card of the equipment, such as "when..." At that time, the status changes from 'running' to 'alarm'; the final generated digital twin DT is a computable model. This is achieved by subscribing to real-time data streams in the twin data pool DP (such as...). , This drives the parameter updates of the corresponding model in DT, achieving synchronization with the physical factory, with the synchronization delay controlled within... Within.

[0032] Intelligent optimization: On the factory digital twin, production process simulation and equipment health status prediction simulation are executed concurrently; the simulation results and real-time data are input into a multi-objective optimization algorithm, with production efficiency, equipment reliability and resource utilization as optimization objectives, to generate a collaborative optimization scheme, which includes production scheduling strategy, predictive maintenance plan and process parameter adjustment suggestions; the optimization objective function of the multi-objective optimization algorithm introduces a cumulative deviation constraint term, which is weighted according to the historical uncertainty of the corresponding decision path in the cumulative deviation knowledge graph, and penalizes high volatility schemes.

[0033] Specifically, two simulation threads are executed concurrently on the digital twin (DT). The production process simulation thread uses a discrete event simulation engine to simulate orders. At workstation The flow between, based on the total completion time The scheduling analysis is performed with minimization as the objective. The equipment health status prediction simulation thread uses a prediction model based on LSTM (Long Short-Term Memory) network. The vibration sequence of the input device at the most recent N=512 time points Output the predicted remaining useful life (RUL) and health index (HI) after M=24 future time points. Define the vector objective function of the multi-objective optimization problem as follows: ,in: , represents production efficiency (the negative sign indicates minimizing negative throughput, i.e. maximizing throughput). To simulate the output per unit time.

[0034] This indicates overall equipment unreliability. Let be the simulated health index of the i-th device.

[0035] This represents the cost of resource utilization.

[0036] The decision variable x includes the production scheduling sequence, equipment maintenance trigger points, and process parameter setpoints. To improve the robustness of the solution, a cumulative deviation constraint term is introduced. Definition Where J represents the historical scheme set, Let x be the Euclidean distance between the current solution x and the historical solution j on the key decision variables. This is the normalized weight of the deviation ratio (|actual value - predicted value| / predicted value) between the actual performance and the simulated predicted performance of historical scheme j. The final optimization objective is to satisfy... ( Under the constraint of 0.3 (an empirical value, usually taken as 0.3), the Pareto optimal front (PF) is solved. The NSGA-II algorithm is used for the solution, with a population size of 100 and 200 iterations. Finally, an equilibrium solution is selected from the PF as the output of the co-optimization scheme, which includes a specific scheduling Gantt chart, maintenance time window suggestions, and process parameter setting tables.

[0037] Interactive verification: The twin, optimization scheme and key indicators are visualized in three dimensions through a VR / AR interactive interface; it supports operators to simulate operation and parameter adjustment of the optimization scheme in a natural interactive way, and to deduce the multi-dimensional performance changes after adjustment in real time in the twin; Specifically, the digital twin (DT) and collaborative optimization scheme are imported into the VR interaction system. The system generates a 3D immersive scene of the workshop. Operators wearing VR headsets can use controllers to grasp virtual devices and view their real-time parameter panels (display). , (e.g., HI). For the scheduling plan in the optimization scheme, the operator can use a handle to drag and adjust the position of the order on the timeline in the virtual scene. The system calculates and displays the adjusted position in real time. Changes. For maintenance plans, operators can click on virtual devices to trigger simulated disassembly and view suspected worn parts indicated by the system. All interactive operations trigger corresponding simulation state changes in DT and are updated in real time on the performance dashboard on one side of the interface. , , The estimated value.

[0038] Execution: Based on the validated optimization scheme, control instructions or decision guidance are generated and issued to the physical plant for execution.

[0039] Specifically, the final solution, confirmed through interactive verification, is automatically converted by the system into two types of instructions: 1) Control instructions: such as the optimized process parameters. , 1) Encapsulate the commands into standard G-code or Modbus TCP protocol packets; 2) Decision guidance: such as generating PDF documents containing maintenance work order numbers, equipment numbers, time windows, and recommended actions. Control commands are directly sent to the controllers of physical equipment via the OPC UA protocol; decision guidance is pushed to the mobile terminals of maintenance personnel or the MES work order system.

[0040] Example of overall application scenario: Using a CNC machine tool processing unit to produce two types of parts ( , For example, the data fusion step involves real-time data acquisition. machine tools , , ,as well as parts , And other data, and integrate them. CAD models and middle The records are used to build a data pool. The construction steps include establishing a data pool containing... Geometric model, physical model of thermal deformation of the main shaft and " A digital twin of the "timely alarm" behavior rules. Intelligent optimization step simulation identifies the current production schedule. , }of It is 360 minutes, and the prediction is... The HI will be below 0.8 after 8 hours. The optimization algorithm, under cumulative bias constraints (historically similar loads), The scheme with a larger prediction deviation has a higher weight ω), and a new scheme is obtained: the adjustment order is { , }, and reduce F to 180mm / min, HI can be maintained above 0.85; in the interactive verification, the engineer confirmed in the VR scene that the new schedule is feasible and the equipment status is better; the execution step sends the new F=180mm / min parameter to The controller then sends a work order to the maintenance team to "check the spindle bearing in 8 hours".

[0041] This embodiment provides a complete digital twin factory interaction method from data acquisition, model building, optimization simulation, immersive verification to command issuance; through multi-source data fusion and high-fidelity modeling, a precise virtual mapping of the physical factory is achieved; by introducing multi-objective concurrent optimization with historical deviation constraints, a collaborative solution that balances efficiency, reliability, and resources can be generated in the virtual environment; through VR / AR interactive verification, operators are allowed to intuitively evaluate and adjust the solution before execution, thereby reducing the risk of physical resource waste or production interruption due to improper decision-making, and realizing a closed loop from perception, analysis to decision-making and verification.

[0042] Example 2: Refer to Figure 2 The difference between this embodiment and embodiment 1 is that, after the data fusion step and before the construction step, it also includes historical deviation accumulation modeling: in the twin data pool, the deviation data between the actual execution results of each optimization scheme and the simulation prediction value is continuously recorded, and the cumulative deviation knowledge graph is constructed and updated according to the dimensions of equipment, process section and product batch. This graph represents the systematic error trend and uncertainty distribution of the simulation model under specific production conditions. Based on the cumulative deviation knowledge graph, the parameters of the corresponding equipment or process model in the digital twin are adaptively calibrated and labeled with confidence level.

[0043] Specifically, the system establishes a deviation recording cycle (e.g., after each production shift or order batch is completed). At the end of each cycle, all optimization schemes executed during that cycle are extracted from the twin data pool DP. The simulation predictions and corresponding values ​​are obtained, along with the actual execution results from the MES and SCADA systems. For each scheme k, calculate its deviation in three dimensions: equipment i, process segment s, and product batch b. Taking equipment as an example, the deviation is calculated as the difference between the predicted and actual values ​​of the key performance indicators (KPIs): Common KPIs include Output Per Unit Time (UPP), Overall Equipment Effectiveness (OEE), and Defect Rate (DR). Subsequently, a cumulative deviation knowledge graph (KG) is constructed, with equipment, process segment, and product batch as entity nodes, and relationships such as "belongs to," "inflows," and "configurations" between them as edges.

[0044] Each entity node stores its historical deviation sequence. And calculate two key attributes: 1) Systematic error trend 1) The arithmetic mean of the sequence; 2) Uncertainty That is, the standard deviation of the sequence. For example, the device Producing " under the "precision milling" process section The historical OEE deviation of the batch is the mean. (Prediction slightly lower), standard deviation is .

[0045] Model calibration based on KG: For digital twin DT In precision milling OEE prediction model at time ,according to After compensation, the calibrated prediction is At the same time, a confidence level is assigned to the model output. , The larger the value, the lower the confidence level C. The confidence level C will be visually represented (e.g., color depth, transparency) in the interactive verification interface.

[0046] Overall application scenario example (continued from Example 1): Assuming the system recorded the following over the past 10 shifts: machine tools in , Processing When calculating the OEE of a component, the simulation prediction averaged 85%, but the actual average was only 83%, with a deviation of the mean. Standard deviation This data is recorded in the KG. "Precision milling" of nodes "Under the attribute. Before this new build step, the system reads this..." , for DT The OEE prediction logic was calibrated, causing its baseline prediction value to be lowered by 2%. Meanwhile, because... The calculated confidence level C≈0.87. In the subsequent VR interface, An "87%" confidence level label will be displayed next to the OEE prediction value. During the intelligent optimization step, the cumulative deviation constraint term g(x) will calculate the difference between the new solution and historical values. The algorithm is guided to avoid similar high-uncertainty decision paths by using a larger (i.e., high-uncertainty) "distance" between similar working condition solutions and applying a higher penalty weight ω.

[0047] This embodiment endows the digital twin system with the ability to continuously learn and self-calibrate through historical deviation accumulation modeling; the structured storage of the cumulative deviation knowledge graph (KG) enables the analysis of model errors to move from fragmented to systematic; based on The adaptive calibration directly improves the prediction accuracy of the digital twin (DT) under specific operating conditions. The confidence level label C provides users with a quantitative reference for the reliability of the model's predictions; simultaneously, As a measure of uncertainty, the algorithm is fed back to the optimization algorithm, enabling the generated solution to proactively avoid paths with high historical volatility and enhance the robustness of the decision-making. This allows the system to move from "accurate one-time modeling" to "continuous evolution during operation".

[0048] Example 3: Reference Figure 3 The difference between this embodiment and Embodiment 2 is that, after the intelligent optimization step, it further includes: Fluctuation monitoring: Systemic high-deviation-risk locations identified in the cumulative deviation knowledge graph are defined as key areas of concern; the physical equipment operating parameters or process indicators in these areas are monitored in real time, their short-term fluctuation values ​​are calculated and compared with a preset dynamic fluctuation threshold; when the short-term fluctuation value exceeds the dynamic fluctuation threshold three times in a row, a real-time warning is generated.

[0049] Specifically, the system filters out uncertainties from the cumulative deviation knowledge graph KG. Greater than the threshold (like Corresponding confidence level The entity nodes of ) are defined as key areas of interest. For each The system subscribes in real time to its corresponding physical sensor data stream (such as...) vibration ). Calculate the data stream within the sliding time window. Short-term fluctuation value within (e.g., W=30 minutes) . The calculation uses the coefficient of variation of the data within the window: ,in and These are the standard deviation and mean of the data sequence within window W, respectively. The preset dynamic fluctuation threshold. It is not a fixed value, but rather depends on this. The nodes are recorded in KG as having historically stable operation ( (small) period The statistic is determined, for example, by taking the period. The 95th quantile of the distribution, Q95, is used as When calculated in real time The requirement is met for three consecutive monitoring cycles (e.g., one cycle every 5 minutes). At that time, generate information about Real-time early warning.

[0050] After generating real-time alerts, the following is also included: Causal tracing: When a high-fluctuation concern point A continuously triggers an early warning, its fluctuation time series is analyzed; if the linear fitting slope of its fluctuation mean is consistently positive or negative and passes the significance test, it is determined that there is a directional drift trend; then, in the cumulative deviation knowledge graph, along the reverse transmission path of material flow, energy flow, or information flow, the direct upstream node B that has a leading-lag relationship with the fluctuation trend of point A in time series and whose own volatility increases in the same period is identified as the suspected root cause node causing the trend fluctuation of point A.

[0051] In the causal tracing step, the fluctuation pattern of the identified suspected root cause node B is further analyzed: if the fluctuation of B is characterized by random high fluctuation without direction, it is marked as a first-level correction target; if the fluctuation of B also shows a directional drift trend in the same direction as A, the tracing continues upstream until a node with a fluctuation pattern of random high fluctuation is found and marked as a first-level correction target.

[0052] For nodes marked as primary correction targets, initiate a directional deterministic compensation in the operational parameter settings of their digital twin models: If the target node is a device, then adjust the set value of its key control parameters by a fixed percentage in the opposite direction of its current fluctuation, or adjust it to the median value of its historical stable operating range. If the target node is a process parameter, then revise the standard operating procedure value of that parameter to the historical average value used during the most recent period of stable output of that node.

[0053] Specifically, when targeting If an alert (denoted as node A) continues to be triggered, the system will trace the fluctuation time series of A over the past L=100 sampling points (e.g., approximately 50 minutes). .right Perform linear fitting If the p-value of the hypothesis test for the slope β ,and ( If the minimum significant drift rate is 0.01 units / sampling interval, then A is determined to have a directional drift trend. Subsequently, all direct upstream nodes U of A are searched backwards along the "material / energy / information flow" edge in KG. For each U, its contemporaneous... Perform a Granger causality test. If... Statistically significant Granger induced Changes (test p-value) ), and U itself If it also increases by more than 20% during the same period, then U is marked as a suspected root cause node B. Analyze the fluctuation pattern of B: calculate... Skewness and kurtosis. If Furthermore, if the Kurtosis value is close to 3, the fluctuations are considered random; if... If the fluctuation is large and exhibits a clear monotonic trend, it is considered directional. If B is a random high-volatility pattern, it is marked as a first-level correction target. If B is drifting in the same direction, then continue to use B as the new A and repeat the above tracing process.

[0054] right Initiate targeted deterministic compensation: If For equipment (such as) Its key control parameter is currently valued at (like The median value of the historical stable operating range is The adjustment amount is Where γ is a fixed percentage (e.g., γ=5%), and its direction is opposite to the current fluctuation direction. That is, the new set value. .

[0055] like If it is a process parameter (such as F), then directly query the most recent time for that node from KG. The historical average value corresponding to the point where the value is below the stability threshold And revise the value in the standard operating procedure (SOP) to .

[0056] Overall application scenario example (continued from Example 2): KG display In precision milling hour higher ( Therefore, it was listed as Real-time system monitoring vibration , discovered its The value was 0.25 within the 10:00-10:30 window, exceeding its historical stable period Q95=0.20 for three consecutive cycles, triggering an alert. Analysis ( The sequence showed a significant upward trend (β>0). In KG, One of its upstream suppliers is the coolant. The inspection revealed... Export pressure sequence Granger caused change( ),and of The value increased from 0.15 to 0.19 during the same period. (Analysis) Sequence, its , It was determined to be random with high fluctuations, therefore... Marked as Query Median of historical stable interval Current value .calculate The direction is negative (because the current value is higher than the median value). Therefore, a deterministic correction command is generated: the outlet pressure setting of pump_P1 is reduced from 5.3MPa to 5.285MPa.

[0057] This embodiment achieves a leap from passively recording deviations to proactive early warning, diagnosis, and rapid intervention. Dynamic fluctuation monitoring based on a knowledge graph (KG) intelligently focuses on weak links with high model uncertainty. The causal tracing step, combined with statistical testing and graph relationships, automates and quantitatively locates the root causes of fluctuations, reducing the time and reliance on experience in manual diagnosis. Deterministic correction rules, based on historical stable data, provide rapid and interpretable primary intervention methods. This enables the system to implement precise source control before potential problems lead to significant production losses or quality defects, shifting the role of digital twins from "post-event analysis and optimization" to "proactive in-process control."

[0058] Example 4: Reference Figure 4 The difference between this embodiment and embodiment 3 is that, before the interactive verification step, it also includes: Buffer: While taking corrective measures for the suspected root cause node B, the following minor adjustments are made to the initial high volatility concern point A and its direct downstream node C, which are affected by the volatility trend: In the simulation logic of the digital twin, the acceptable range for the output quality of point A is temporarily relaxed; Slightly reduce the production cycle time or feed rate of downstream node C to allow for buffer time to handle potential unconventional inputs from point A.

[0059] Specifically, when initiating the investigation of suspected root cause node B (as in the previous example) While correcting the issue, the system also performs adjustments to the affected node A. ) and its direct downstream node C (such as the one responsible for Preventive buffer adjustment of the inspection station I1 in the next process.

[0060] For node A ( ): In the digital twin DT, temporarily modify its output quality pass / fail judgment logic. Assume the original... The critical dimensional tolerance of the part is The buffer has been adjusted to be temporarily relaxed to . As a buffer amplitude, its value is determined based on the current fluctuation at point A. Exceeding the threshold proportion Calculation, for example The basic tolerance zone (e.g., 0.005 mm).

[0061] For node C(I1): In the DT production process simulation logic, its production cycle time (or material feed rate) is determined. Make a slight downward adjustment. New rate. ρ is the downward adjustment coefficient, based on the process buffer time from A to C. And A's estimated recovery time Estimate, ρ = min(0.05, The maximum reduction will not exceed 5%.

[0062] Following the execution steps, the following is also included: Priority reallocation: During a preset observation period after the execution of the directional deterministic compensation and preventive buffer adjustment, the fluctuation data of node B, the output quality data of node A, and the production smoothness data of node C in the physical plant are collected simultaneously. Based on the volatility data of node B, its short-term statistical volatility is calculated and compared with the benchmark volatility before execution to generate the root cause correction effectiveness coefficient. Based on the output quality data of node A and the relaxed judgment range, its pass rate is calculated and compared with the historical pass rate before the warning is triggered to generate a quality buffer adaptation coefficient. The product of the root cause correction effectiveness coefficient and the quality buffer adaptation coefficient is defined as the comprehensive handling efficiency index for this chain of fluctuation events. Based on the magnitude of the comprehensive handling efficiency index, the risk priority labels of nodes A, B, and C involved in this event in the cumulative deviation knowledge graph are dynamically recalibrated: if the efficiency index is higher than the high threshold, its priority is reduced; if the efficiency index is lower than the low threshold, its priority is increased. This recalibration result is fed back to the dynamic fluctuation monitoring and threshold early warning step to adjust the focus of subsequent monitoring resource allocation and the sensitivity of the early warning threshold.

[0063] Specifically, after the correction and buffering instructions are issued and executed, a preset observation period is initiated. (e.g., 4 hours). During the period, synchronously collect data from the physical factory: B ( Pressure fluctuation data, A ( ) produced The quality inspection data of the parts, and the work-in-process queue length data of station C(I1).

[0064] Calculate the root cause correction effectiveness coefficient : Calculate the observation period separately Same duration as before the correction Standard deviation of pressure data at point B and . .like ,but Set it to 0.

[0065] Calculate the mass buffer fit coefficient Statistical observation period Within the relaxed decision range, point A's output falls within this range. The pass rate within And the historical pass rate within the same production duration prior to the warning trigger. . .

[0066] Calculate the comprehensive disposal efficiency index .

[0067] according to The size of the risk priority label Pri (divided into high, medium, and low levels) of nodes A, B, and C in KG is dynamically recalibrated: like (like If the assessment is efficient, the Pri values ​​of all three will be reduced by one level (e.g., high -> medium, medium -> low).

[0068] like (like The decision was deemed inefficient, and all three were upgraded by one level.

[0069] like If Pri remains unchanged, then Pri will remain unchanged.

[0070] This recalibration result will be fed back to the dynamic fluctuation monitoring step: for nodes with rising Pri, the system will shorten their monitoring window W (e.g., from 30 minutes to 15 minutes) and may lower their dynamic threshold. (For example, replacing Q95 with Q90) to improve monitoring sensitivity; for nodes with reduced Pri, W may be extended or a more lenient approach may be adopted. To save computing resources.

[0071] Overall application scenario example (following Example 3): During the execution of downscaling Under pressure, the system Production Bore diameter tolerance of parts Temporarily relax restrictions, take , Therefore, the temporary allowance is For the downstream inspection station I1, the estimated ρ=0.02, so its theoretical cycle time is slightly adjusted from 60 pieces / hour to 58.8 pieces / hour.

[0072] go through observe: Standard deviation of pressure fluctuation is Down to , . Output The pass rate for parts under temporary tolerances was 98.5%, and the historical pass rate was... , . .

[0073] because The system determined that the overall processing efficiency was low. Despite good quality buffer adaptation ( However, the effectiveness of root cause correction is questionable. Very low. Therefore, in KG, will , The Pri labels for both I1 and I2 have been upgraded from "Medium" to "High". Following this feedback, subsequent monitoring of these three nodes will employ a shorter window and more sensitive thresholds, and may indicate the need for further adjustments. Impact of pressure fluctuations We need to re-examine the causal relationship of "vibration".

[0074] This embodiment improves the closed loop of intelligent intervention by adding buffering and priority reallocation steps, and introduces a system self-optimization mechanism based on performance evaluation. The buffering step, by pre-setting tolerance intervals and adjusting downstream rhythms in virtual space, buys time for corrective measures in the real world to take effect, absorbs uncertainties, protects the continuity of production flow, and reflects system-level collaborative control thinking. The priority reallocation step, by quantitatively evaluating the combined practical effects of "correction" and "buffering" and feeding the results back to the knowledge graph and monitoring module, enables the system to "review" and "adjust strategies" of its own intervention behavior. This allows the system to identify which models, diagnoses, or intervention rules need further optimization, thereby dynamically adjusting monitoring resources and focus, and driving the entire digital twin system to continuously evolve autonomously towards greater precision and efficiency.

[0075] This application also discloses a factory interaction system based on digital twins.

[0076] A factory interaction system based on digital twins is characterized by comprising: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes the computer program stored on the computer-readable storage medium, it implements a factory interaction method based on digital twins.

[0077] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A factory interaction method based on digital twins, characterized in that: Includes the following steps: Data fusion: Real-time collection of equipment operating parameters and environmental sensing data through the Industrial Internet of Things, and integration of factory CAD drawings, production process data and quality inspection data, followed by cleaning, alignment and fusion, to build a twin data pool with a unified spatiotemporal benchmark; Construction: Based on the aforementioned digital twin data pool, a multi-scale modeling method integrating geometric, physical, and behavioral rules is used to construct a digital twin of the factory; the digital twin is then driven to be updated synchronously through real-time data streams. Intelligent optimization: On the factory digital twin, production process simulation and equipment health status prediction simulation are executed concurrently; the simulation results and real-time data are input into a multi-objective optimization algorithm, with production efficiency, equipment reliability and resource utilization as optimization objectives, to generate a collaborative optimization scheme, which includes production scheduling strategy, predictive maintenance plan and process parameter adjustment suggestions; Interactive verification: The twin, optimization scheme and key indicators are visualized in three dimensions through a VR / AR interactive interface; it supports operators to simulate operation and parameter adjustment of the optimization scheme in a natural interactive way, and to deduce the multi-dimensional performance changes after adjustment in real time in the twin; Execution: Based on the validated optimization scheme, control instructions or decision guidance are generated and issued to the physical plant for execution.

2. The factory interaction method based on digital twins according to claim 1, characterized in that: It also includes a historical deviation accumulation modeling step: In the twin data pool, the deviation data between the actual execution results and the simulation prediction values ​​of each optimization scheme are continuously recorded, and a cumulative deviation knowledge graph is constructed and updated according to the dimensions of equipment, process section and product batch. This graph represents the systematic error trend and uncertainty distribution of the simulation model under specific production conditions. Based on the cumulative deviation knowledge graph, the parameters of the corresponding equipment or process model in the digital twin are adaptively calibrated and labeled with confidence level.

3. The factory interaction method based on digital twins according to claim 2, characterized in that: In the intelligent optimization step, the objective function of the multi-objective optimization algorithm introduces a cumulative deviation constraint term. This constraint term is weighted according to the historical uncertainty of the corresponding decision path in the cumulative deviation knowledge graph, and penalizes highly volatile solutions.

4. The factory interaction method based on digital twins according to claim 2, characterized in that: Also includes: Fluctuation monitoring: Systemic high-risk locations identified in the cumulative deviation knowledge graph are defined as key areas of focus; The system monitors the operating parameters or process indicators of physical equipment in the area in real time, calculates their short-term fluctuation values ​​and compares them with preset dynamic fluctuation thresholds; when the short-term fluctuation value exceeds the dynamic fluctuation threshold three times in a row, a real-time warning is generated.

5. The factory interaction method based on digital twins according to claim 4, characterized in that: Also includes: Causal tracing: When a high-volatility concern point A continuously triggers warnings, analyze its volatility time series; If the slope of the linear fit of its mean fluctuation is consistently positive or negative and passes the significance test, it is determined that it has a directional drift trend. Subsequently, in the cumulative deviation knowledge graph, along the reverse transmission path of material flow, energy flow, or information flow, we search for the direct upstream node B that has a leading-lag relationship with the fluctuation trend of point A in time sequence and whose own volatility increases in the same period. We identify B as the suspected root cause node causing the trend fluctuation of point A.

6. The factory interaction method based on digital twins according to claim 5, characterized in that: In the causal tracing step, the fluctuation pattern of the identified suspected root cause node B is further analyzed: if the fluctuation of B is characterized by random high fluctuation without direction, it is marked as a first-level correction target; if the fluctuation of B also shows a directional drift trend in the same direction as A, the tracing continues upstream until a node with a fluctuation pattern of random high fluctuation is found and marked as a first-level correction target.

7. The factory interaction method based on digital twins according to claim 6, characterized in that: For nodes marked as primary correction targets, initiate a directional deterministic compensation in the operational parameter settings of their digital twin models: If the target node is a device, then adjust the set value of its key control parameters by a fixed percentage in the opposite direction of its current fluctuation, or adjust it to the median value of its historical stable operating range. If the target node is a process parameter, then revise the standard operating procedure value of that parameter to the historical average value used during the most recent period of stable output of that node.

8. The factory interaction method based on digital twins according to claim 7, characterized in that: Also includes: Buffer: While taking corrective measures for the suspected root cause node B, the following minor adjustments are made to the initial high volatility concern point A and its direct downstream node C, which are affected by the volatility trend: In the simulation logic of the digital twin, the acceptable range for the output quality of point A is temporarily relaxed; Slightly reduce the production cycle time or feed rate of downstream node C to allow for buffer time to handle potential unconventional inputs from point A.

9. The factory interaction method based on digital twins according to claim 8, characterized in that: Also includes: Priority reallocation: During a preset observation period after the execution of the directional deterministic compensation and preventive buffer adjustment, the fluctuation data of node B, the output quality data of node A, and the production smoothness data of node C in the physical plant are collected simultaneously. Based on the volatility data of node B, its short-term statistical volatility is calculated and compared with the benchmark volatility before execution to generate the root cause correction effectiveness coefficient. Based on the output quality data of node A and the relaxed judgment range, its pass rate is calculated and compared with the historical pass rate before the warning is triggered to generate a quality buffer adaptation coefficient. The product of the root cause correction effectiveness coefficient and the quality buffer adaptation coefficient is defined as the comprehensive handling efficiency index for this chain of fluctuation events. Based on the magnitude of the comprehensive handling efficiency index, the risk priority labels of nodes A, B, and C involved in this event in the cumulative deviation knowledge graph are dynamically recalibrated: if the efficiency index is higher than the high threshold, its priority is reduced; if the efficiency index is lower than the low threshold, its priority is increased. This recalibration result is fed back to the dynamic fluctuation monitoring and threshold early warning step to adjust the focus of subsequent monitoring resource allocation and the sensitivity of the early warning threshold.

10. A factory interaction system based on digital twins, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in any one of claims 1-9.

Images