Virtual-real interaction digital twin full closed loop control method and system

Abstract

This application relates to the field of intelligent control technology and discloses a virtual-physical interactive digital twin full-closed-loop control method and system. The method includes: collecting multimodal time-series sensor data from manufacturing equipment to establish an execution-layer semantic benchmark library and a cognitive-layer confidence parameter set; selecting corresponding commands from three levels of control commands based on the confidence level and issuing them to the actuator; registering a semantic blind zone hypothesis when the confidence level is below a threshold and calculating the optimal detection action to superimpose and issue it to the actuator; and continuously issuing corrective control commands after updating the execution-layer semantic benchmark library using the detection residual vector. This application solves the problems of existing digital twin control systems being unable to recognize semantic mapping failures under operating condition drift and the virtual model being unable to actively apply directional detection to the physical actuator to eliminate cognitive blind zones.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to a virtual-real interactive digital twin full closed-loop control method and system. Background Technology

[0002] Digital twin technology, by constructing a digital mapping model of a physical entity in virtual space, enables real-time monitoring and control decisions of the physical entity's operational status. It has been widely applied in fields such as intelligent manufacturing, collaborative scheduling of industrial robots, and drone swarm control. Existing digital twin control systems typically employ a single-layer semantic mapping architecture, where physical entity sensor data is directly converted into virtual model drive commands via fixed mapping rules. The virtual model then issues control commands to the physical entity based on the mapping results, achieving data interaction between the physical and virtual sides. In this architecture, the semantic mapping rules are established based on known operating condition data during system deployment and remain static. The virtual model continuously receives sensor data from the physical side and outputs control commands accordingly, forming a basic control closed loop where physical entity status data flows to the virtual model, and virtual model control commands flow to the physical actuators.

[0003] However, the aforementioned single-layer static semantic mapping architecture has inherent defects during the long-term operation of manufacturing equipment. The actual operating conditions of manufacturing equipment continuously drift with production task switching, equipment wear and tear, and environmental changes. When the real-time operating conditions deviate from the known operating condition range on which the semantic mapping rules were established, the behavioral intent judgment output by the fixed semantic mapping rules will be biased. The control commands generated by the virtual model accordingly will also become inaccurate. Furthermore, since the existing system does not have the ability to assess the reliability of the semantic mapping, the system cannot detect the above deviations when they occur, resulting in erroneous control commands being continuously issued to the manufacturing equipment actuators, forming an unstable state where the more control is applied, the more skewed the deviation becomes.

[0004] If a semantic credibility assessment mechanism is introduced into the system to detect mapping deviations, the assessment mechanism itself also relies on a baseline model built from known operating condition data. When the operating condition drifts beyond the known range, the judgment ability of the assessment mechanism also fails. That is, the system not only does not know whether the control command is correct, but also does not know whether its ability to judge whether the control command is correct is reliable, forming a cognitive blind spot. To eliminate the cognitive blind spot, new data under the current drifting operating condition needs to be obtained. However, the data acquisition of the existing digital twin system relies entirely on the sensor data naturally generated by the physical entity during normal production. It is impossible to actively apply targeted detection to the physical entity to obtain targeted verification data. As a result, the cognitive blind spot cannot be eliminated for a long time under normal production cycle. The online correction of semantic mapping rules also lacks effective data support and cannot be realized. The semantic baseline library of the execution layer is solidified for a long time, and the control accuracy continues to degrade with the accumulation of running time. Summary of the Invention

[0005] This application provides a virtual-real interactive digital twin full closed-loop control method and system, which solves the problems of existing digital twin control systems being unable to recognize semantic mapping failure under operating condition drift and the virtual model being unable to actively apply directional detection to the physical actuator to eliminate cognitive blind spots. It realizes online collaborative updating of the execution layer semantic reference library and the cognitive layer confidence parameter set, so that the manufacturing equipment actuator can still receive corrective control commands based on the updated semantic reference under continuous operating condition drift.

[0006] Firstly, this application provides a fully closed-loop control method for virtual-real interactive digital twins, the fully closed-loop control method for virtual-real interactive digital twins comprising: Step S1: Collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set that include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. Step S2: Match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer, output the current confidence level from the confidence level parameter set of the cognition layer, and select the corresponding control command from the three levels of weighted fusion command, conservative subset command, and direct drive command according to the confidence level to issue to the execution mechanism of the manufacturing equipment. When the confidence level is lower than the preset threshold, register the semantic blind zone hypothesis. Step S3: Based on the semantic blind zone hypothesis, with an amplitude constraint of not exceeding the preset proportion of the conservative subset instruction norm, calculate the optimal detection action, superimpose the optimal detection action onto the conservative subset instruction, and issue it to the manufacturing equipment actuator. Record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. Step S4: Perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library based on the probe residual vector, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

[0007] Secondly, this application provides a fully closed-loop control system for virtual-real interactive digital twins, the fully closed-loop control system comprising: The acquisition module is used to collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and to establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set, which include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. The registration module is used to match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer. The current confidence level is output by the confidence level parameter set of the cognition layer. Based on the confidence level, the corresponding control command is selected from three levels: weighted fusion command, conservative subset command, and direct drive command and issued to the execution mechanism of the manufacturing equipment. When the confidence level is lower than a preset threshold, the semantic blind zone hypothesis is registered. The recording module is used to calculate the optimal detection action based on the semantic blind zone hypothesis, with an amplitude constraint of not exceeding the preset proportion of the conservative subset instruction norm, and to issue the optimal detection action to the manufacturing equipment actuator after superimposing the conservative subset instruction, and to record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. The generation module is used to perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library based on the probe residual vector, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

[0008] Thirdly, a virtual-real interactive digital twin fully closed-loop control device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the virtual-real interactive digital twin fully closed-loop control device to execute the aforementioned virtual-real interactive digital twin fully closed-loop control method.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned virtual-real interactive digital twin full-closed-loop control method.

[0010] The technical solution provided in this application fundamentally changes the existing digital twin system's reliance on single-layer static mapping rules to issue control commands to manufacturing equipment actuators by establishing a two-layer semantic structure based on multimodal temporal sensing data under known operating conditions. The execution-layer semantic benchmark library stores the semantic knowledge of known operating conditions in a three-element organization of operating condition feature vectors, behavioral intent labels, and operating condition boundary domains. The cognitive-layer confidence parameter set dynamically summarizes the credibility of the current semantic judgment based on the historical prediction error sequence of the execution-layer semantics. The two layers are dynamically correlated through the error sequence, enabling the system to quantitatively assess the credibility of the current semantic mapping before issuing control commands to the manufacturing equipment actuators. This overcomes the fundamental deficiency in existing technologies where the system cannot perceive the reliability of its own semantic judgment capabilities. Based on this, the corresponding control command is dynamically selected from three levels—weighted fusion command, conservative subset command, and direct drive command—and issued to the manufacturing equipment actuator according to the current confidence level output by the cognitive layer confidence parameter set. The three-level switching mechanism enables the generation strategy of control commands to be adaptively adjusted according to the change of semantic confidence level. When the confidence level is sufficient, the predictive ability of the virtual model is maximized. When the confidence level decreases, the command coverage range is narrowed to the high deterministic region. When the confidence level is severely insufficient, it switches to pure physical measurement direct drive, ensuring that the manufacturing equipment actuator can receive control commands that match the current semantic confidence level under any confidence level condition.

[0011] When the confidence level is below a preset threshold, a semantic blind zone hypothesis is registered. The optimal detection action is calculated with an amplitude constraint not exceeding a preset proportion of the conservative subset instruction norm. This optimal detection action is then superimposed onto the conservative subset instructions and actively issued to the manufacturing equipment actuator. The core contribution of this mechanism lies in transforming the virtual model's role from a passive mirror receiving physical sensor data to an active detector capable of issuing directional detection commands to the manufacturing equipment actuator. The calculation of the detection action is based on the working condition positioning information and error statistics recorded in the semantic blind zone hypothesis, with the optimization objective of maximizing information gain. Within the amplitude constraint range, the detection action that contributes most to eliminating the cognitive blind zone is selected and injected into the normal control commands, ensuring that the detection process does not interfere with the normal production cycle of the manufacturing equipment actuator. The detection residual vector recorded after detection serves as the core data basis for updating the execution layer semantic benchmark library, driving the expansion of the working condition boundary domain or the generation of new semantic entries. The updated execution layer semantic benchmark library and the cognitive layer confidence parameter set re-participate in the working condition boundary domain matching and confidence output, continuously issuing corrective control commands to the manufacturing equipment actuators, forming a truly two-way active interactive closed loop between the physical and virtual sides, enabling the manufacturing equipment actuators to receive effective control commands based on the online updated semantic benchmarks for a long time even under conditions of continuous working condition drift. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of the virtual-real interactive digital twin full closed-loop control method in this application. Figure 2 This is a schematic diagram comparing the actual response of the actuator with the predicted response of the virtual model in an embodiment of this application; Figure 3 This is a schematic diagram illustrating how the normalized amplitude of the detection residual changes with the detection action number in an embodiment of this application. Detailed Implementation

[0014] This application provides a virtual-real interactive digital twin fully closed-loop control method and system. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the virtual-real interactive digital twin full closed-loop control method in this application includes: Step S1: Collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set that include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. Specifically, the execution layer semantic benchmark library is a structured database formed by causal modeling of displacement, velocity, torque, and temperature data collected from manufacturing equipment under multiple known operating conditions. Each record in the library contains three data items: an operating condition feature vector, a behavioral intent label, and an operating condition boundary domain. The operating condition feature vector is aggregated from the aforementioned four-dimensional sensor data. The behavioral intent label describes the semantic category of the manufacturing equipment's actions under that operating condition. The operating condition boundary domain is a Mahalanobis distance hyperellipsoid constructed with the operating condition feature vector as the center and the covariance matrix as the shape parameter. The Mahalanobis distance threshold is set to 3 to cover the effective range of sensor data distribution under that operating condition. The cognitive layer confidence parameter set is obtained by fitting historical prediction error sequences using Gaussian process regression. Its output value range is a continuous quantity between 0 and 1, reflecting the credibility of the execution layer semantics under the current operating condition. This parameter set is continuously updated with the control cycle and is not a fixed static parameter.

[0016] Step S2: Match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer. Output the current confidence level from the confidence level parameter set of the cognition layer. Select the corresponding control command from the three levels of weighted fusion command, conservative subset command, and direct drive command according to the confidence level and issue it to the execution mechanism of the manufacturing equipment. Register the semantic blind zone hypothesis when the confidence level is lower than the preset threshold. Specifically, the selection boundary for the three levels of control commands is determined based on the comparison between the current confidence level output by the cognitive layer confidence parameter set and two preset thresholds. When the confidence level is not lower than the upper threshold, the virtual model prediction driving weight and the physical measurement direct driving weight are linearly weighted and superimposed according to the confidence level value to form a weighted fusion command issued to the actuator. When the confidence level is between the upper and lower thresholds, only the driving components corresponding to the records with the smallest Mahalanobis distance in the execution layer semantic benchmark library are extracted, and their intersection is taken to form a conservative subset of commands, ensuring that the command range is narrowed to the high confidence region. When the confidence level is lower than the lower threshold, the control quantity is directly calculated based entirely on real-time physical measurements to form a direct drive command, without inference from the virtual model. The semantic blind zone hypothesis is registered when the confidence level reaches the lower threshold, recording the real-time operating condition feature vector at the trigger time, the mean and variance of the prediction error within the current time window. These three data together describe the operating condition location and degree of semantic failure.

[0017] Step S3: Based on the semantic blind zone hypothesis, with the amplitude constraint not exceeding the preset proportion of the conservative subset instruction norm, calculate the optimal detection action, superimpose the optimal detection action onto the conservative subset instruction and issue it to the manufacturing equipment actuator, and record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. Specifically, the calculation of the optimal detection action is based on the real-time operating condition feature vector recorded in the semantic blind zone hypothesis. The uncovered intervals between the boundary domains of each operating condition and this vector are identified in the execution layer semantic benchmark library, forming a semantic coverage gap vector. In the candidate detection action set, the reduction in cognitive layer information entropy after applying the action is calculated for each candidate action, and the action with the largest reduction is selected as the optimal detection action. Amplitude constraints limit the Euclidean norm of the candidate actions to within 5% of the norm of the conservative subset of instructions. This percentage is set based on the premise of generating observable response differences without triggering the overtravel protection of the manufacturing equipment actuator. If the percentage is too small, the response differences will be submerged in sensor noise; if the percentage is too large, it will interfere with normal process execution. The detection residual vector is obtained by subtracting the measured four-dimensional response of the manufacturing equipment actuator in the next sampling period after the detection action is executed, and the predicted four-dimensional response of the virtual model for the same action, component by component. The normalized amplitude of this vector, after being normalized by the standard deviation of each channel, is used to determine the update method of the execution layer semantic benchmark library.

[0018] Step S4: Based on the detection residual vector, perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

[0019] Specifically, based on the comparison between the normalized magnitude of the probe residual vector and a preset magnitude threshold, two types of updates are performed on the execution layer semantic benchmark library. When the normalized magnitude does not exceed the preset magnitude threshold, the current operating condition is determined to be a slight extension of the existing operating condition boundary domain. A column vector is constructed by the difference between the real-time operating condition feature vector and the nearest operating condition boundary domain center vector. A rank-one update is performed on the covariance matrix of this boundary domain, expanding the boundary domain towards the current operating condition. The learning rate is set to 0.1 to control the magnitude of each expansion. When the normalized magnitude exceeds the preset magnitude threshold, the current operating condition is determined to fall outside the existing semantic coverage. The probe residual vector is substituted into the structural causal model in the execution layer semantic benchmark library to calculate the posterior probability of each behavioral intent label. The label with the highest probability is combined with the real-time operating condition feature vector and the initial covariance matrix to write a new semantic entry. The initial covariance matrix is ​​set to a small multiple of the identity matrix to limit the initial coverage of the new entry, which will be expanded further after subsequent probe data accumulation. After the update is completed, the execution layer semantic benchmark library re-participates in the working condition boundary domain matching, and the cognitive layer confidence parameter set outputs the updated confidence based on this, driving the manufacturing equipment actuator to receive the corrected control command.

[0020] In one specific embodiment, step S1 includes: Multimodal time-series sensing data is constructed from displacement, velocity, torque, and temperature data collected from manufacturing equipment under known operating conditions. Causal inference is performed on the multimodal time-series sensing data based on a structural causal model to obtain the causal strength coefficient between operating condition variables and behavioral intention variables. Based on the causality strength coefficient, the condition variables under each known working condition are aggregated to obtain the working condition feature vector and behavioral intention label corresponding to each known working condition. Based on the covariance matrix of the characteristic vectors of the working conditions, the coverage of each known working condition is modeled using a Mahalanobis distance hyperellipsoid, thus obtaining the boundary domain of each known working condition. The operating condition feature vector, behavioral intention label, and operating condition boundary domain are written into the execution layer semantic benchmark library. Based on the error sequence between the predicted and actual behavioral intention values ​​in the historical control cycle, Gaussian process regression is used to fit the error sequence to obtain the cognitive layer confidence parameter set.

[0021] Specifically, when manufacturing equipment operates continuously under known operating conditions, four-dimensional sensing quantities—displacement, velocity, torque, and temperature—are combined to form multimodal time-series sensing data with a sampling period Δt. The structural causal model treats each dimension of the four-dimensional sensing quantities as a node in a causal directed graph, with directed edges between nodes representing causal influence paths. The causal strength coefficient α is calculated using the do-calculus framework, specifically representing the conditional probability of the behavioral intention variable taking a corresponding value after an intervention operation is applied to the operating condition variable. The closer the α value is to 1, the stronger the determinant effect of the operating condition variable on the behavioral intention. The process of aggregating operating condition variables into an operating condition feature vector involves selecting sensing components with a causal strength coefficient of not less than 0.5 and concatenating them according to their mean within a time window to form the operating condition feature vector corresponding to the known operating condition. Simultaneously, the semantic category of the equipment action under this operating condition is recorded as a behavioral intention label, which is pre-labeled by process engineers based on the actual production process. In Mahalanobis distance hyperellipsoid modeling, the covariance matrix is ​​obtained by statistically analyzing the feature vectors of all sampling points under the known working condition. The Mahalanobis distance threshold is set to 3, which corresponds to a sample coverage rate of about 99.7% under the multivariate normal distribution. If the coverage rate is too low, the boundary domain will be too narrow, leading to frequent triggering of the semantic blind zone hypothesis. If the coverage rate is too high, the boundary domain will be too wide, leading to aliasing between different working conditions.

[0022] In the Gaussian process regression fitting of the error sequence, the time steps of the historical control cycle are used as input variables, and the absolute value of the difference between the intended predicted value and the actual observed value at each time step is used as the observed value. The Matérn 5 / 2 kernel function is selected to establish the Gaussian process model. This kernel function achieves a balance between smoothness and local variation capability, making it suitable for describing the non-stationary characteristics of the error sequence during the drift of manufacturing equipment operating conditions. The length scale parameter l and the signal variance parameter of the kernel function are determined by maximizing the logarithmic marginal likelihood function. The partial derivative of the logarithmic marginal likelihood function with respect to the kernel parameters is taken and set to zero, and the optimal parameter combination is obtained through iterative solution. After fitting, the Gaussian process regression model outputs the predicted mean confidence level and 95% confidence interval at the current time. The predicted mean confidence level is written into the cognitive layer confidence parameter set. When the confidence interval width exceeds 0.2, the switching threshold of the three control commands is increased by 0.1 to switch to a more conservative control level in advance under high uncertainty conditions.

[0023] In one specific embodiment, step S2 includes: The displacement, velocity, torque, and temperature values ​​of the manufacturing equipment during the current sampling period are input into the semantic benchmark library of the execution layer. The Mahalanobis distance is calculated on the feature vectors of each working condition boundary domain to obtain the matching distance value between the current real-time working condition feature vector and each working condition boundary domain. Based on the matching distance value, the current confidence level is extracted from the confidence level parameter set of the cognitive layer. Based on the relationship between the current confidence level and the preset threshold, the corresponding control command is determined from the three levels of weighted fusion command, conservative subset command, and direct drive command, and the control command is issued to the actuator of the manufacturing equipment. Based on the judgment result that the current confidence level is lower than the preset threshold, the real-time working condition feature vector, the mean error and the error variance within the current time window are associated and written to obtain the semantic blind zone hypothesis carrying the working condition positioning information and error statistics.

[0024] Specifically, the displacement, velocity, torque, and temperature output by the manufacturing equipment within the current sampling period are filtered according to the causal strength coefficient and concatenated into a real-time operating condition feature vector. This vector is then used to calculate the Mahalanobis distance between each record's operating condition feature vector in the execution layer semantic benchmark library. The record with the smallest distance is taken as the nearest semantic entry for the current operating condition. The error sequence corresponding to this entry is input into a Gaussian process regression model in the cognitive layer confidence parameter set, and the current confidence level is output. A current confidence level of 0.7 or higher is considered the first level. The virtual model-predicted driving quantity and the directly calculated driving quantity from physical measurements are linearly weighted and superimposed using the current confidence level and its complement. The higher the confidence level, the greater the proportion of the virtual model-predicted driving quantity. The sum of the two constitutes a weighted fusion command issued to the manufacturing equipment's execution mechanism. When the current confidence level is between 0.3 and 0.7, it is classified as the second level. The intersection of the driving components corresponding to several records with the smallest Mahalanobis distance is extracted from the execution layer semantic benchmark library to form a conservative subset of instructions with narrowed coverage, which is then issued to the actuator. This level design ensures that when the semantic confidence level is moderately low, the control instructions only cover the high-determinism region. When the current confidence level is below 0.3, it is classified as the third level. The control quantities are directly calculated based entirely on real-time physical measurements to form direct drive instructions, which are then issued to the actuator without going through the virtual model inference stage.

[0025] A judgment result with a confidence level below 0.3 triggers the registration of the semantic blind zone hypothesis. The time window takes the prediction error sequence within the previous 100 sampling periods. The mean and variance are calculated for the absolute values ​​of the differences between the predicted and actual observed values ​​at each moment in the sequence. The mean reflects the average degree of the current semantic deviation, and the variance reflects the severity of the deviation fluctuation; together, they constitute the error statistics. The real-time operating condition feature vector, the mean error, and the variance error are correlated and written into the same record. The real-time operating condition feature vector serves as the operating condition positioning information, marking the location of the semantic failure. The mean error and variance error serve as error statistics, describing the degree of failure. These three data points together constitute the semantic blind zone hypothesis, carrying operating condition positioning information and error statistics. This hypothesis records the current operating condition's positioning coordinates and a quantitative description of the failure state in the semantic benchmark database, driving the calculation of subsequent optimal detection actions. The time window length is set to 100 sampling periods. This value strikes a balance between the stationarity of the error sequence and the response speed to changes in operating conditions. If the window is too short, the statistics will be greatly affected by noise; if the window is too long, the perception of rapid shifts in operating conditions will lag.

[0026] In one specific embodiment, step S3, based on the semantic blind zone assumption and with a magnitude constraint not exceeding a preset proportion of the conservative subset instruction norm, calculates the optimal detection action, including: Input the real-time operating condition feature vector and error statistics in the semantic blind zone hypothesis into the execution layer semantic benchmark library, locate the coverage gap between each operating condition boundary domain and the real-time operating condition feature vector, and obtain the semantic coverage gap vector. Based on the semantic coverage gap vector, the reduction in cognitive layer information entropy after each candidate detection action is applied in the candidate detection action set is calculated to obtain the information gain value corresponding to each candidate detection action. The Euclidean norm of each candidate detection action is compared with a preset ratio of the norm of the conservative subset instruction. Candidate detection actions whose Euclidean norm exceeds the preset ratio are eliminated to obtain a subset of candidate detection actions that satisfy the amplitude constraint. The optimal detection action is obtained by extracting the candidate detection action with the largest information gain value from the subset of candidate detection actions.

[0027] Specifically, after the real-time operating condition feature vector in the semantic blind zone hypothesis is input into the semantic benchmark library of the execution layer, it is compared with the operating condition boundary domain of each record in the library to determine whether the vector falls within the coverage range of the Mahalanobis distance hyperellipsoid of each operating condition boundary domain. For all operating condition boundary domains that fail to cover the vector, the difference vector between the operating condition feature vector at the center of each boundary domain and the real-time operating condition feature vector is calculated. These difference vectors are then weighted and summed according to the mean error in the error statistics from largest to smallest to obtain the semantic coverage gap vector. Each component of the semantic coverage gap vector reflects the direction and degree of deviation between the current operating condition and the known semantic coverage area in the corresponding sensing dimension. The operating condition boundary domain with a larger mean error weight contributes more to the gap vector, indicating that the direction with a more severe semantic failure dominates in gap localization. The candidate detection action set is enumerated by several discrete value combinations of displacement increment, velocity increment, torque increment, and temperature increment. Each candidate detection action is a specific value combination of the above four-dimensional increments, and the value range is generated by uniform discrete sampling based on the upper limit of the physical stroke of each dimension of the manufacturing equipment actuator.

[0028] For each candidate action in the candidate detection action set, it is superimposed onto the current conservative subset of instructions and input into the virtual model to predict the response state of the manufacturing equipment actuator under that action. The predicted response state is then substituted into the Gaussian process regression model of the cognitive layer confidence parameter set to calculate the reduction in cognitive layer prediction uncertainty after applying the candidate action. This reduction is the information gain value of the candidate action; a larger reduction indicates a stronger contribution to eliminating semantic blind spots. Amplitude constraints limit the Euclidean norm of the candidate detection actions to within 5% of the norm of the conservative subset of instructions. This percentage is set based on the upper limit of acceptable disturbances during normal process execution of the manufacturing equipment actuator. If the percentage is below 2%, the detection action amplitude is submerged in sensor noise, resulting in unobservable response differences. If the percentage is above 10%, the detection action interferes with the normal process cycle. 5% represents an empirically balanced point between these two extremes. After filtering out candidate detection actions whose Euclidean norm exceeds the specified proportion, the action with the largest information gain value is selected from the remaining candidate detection action subset as the optimal detection action. This action has the strongest effect on eliminating semantic blind spots under the premise of satisfying the amplitude constraint. It is then superimposed on the conservative subset instructions and issued to the manufacturing equipment execution mechanism for execution.

[0029] In one specific embodiment, step S3, which involves superimposing the optimal detection action onto the conservative subset instruction and then issuing it to the manufacturing equipment actuator, includes: The displacement increment, velocity increment, torque increment, and temperature increment in the optimal detection action are superimposed component by component with the corresponding displacement component, velocity component, torque component, and temperature component in the conservative subset instruction to obtain the detection drive instruction. The speed deviation value is obtained by calculating the difference between the current speed component of the manufacturing equipment actuator and the current process reference speed. The speed deviation value is compared with the speed tolerance. If the speed deviation value does not exceed the speed tolerance and the current process node is a non-critical process node, the detection drive command is issued to the manufacturing equipment actuator.

[0030] Specifically, the optimal detection action consists of four components: displacement increment, velocity increment, torque increment, and temperature increment. The conservative subset instruction consists of the corresponding displacement component, velocity component, torque component, and temperature component. The specific operation of component-by-component superposition is as follows: the displacement increment of the optimal detection action is added to the displacement component of the conservative subset instruction to obtain the displacement component of the detection drive instruction; the velocity increment is added to the velocity component to obtain the velocity component of the detection drive instruction; the torque increment is added to the torque component to obtain the torque component of the detection drive instruction; and the temperature increment is added to the temperature component to obtain the temperature component of the detection drive instruction. The four superimposed results together constitute the detection drive instruction. After the detection drive instruction is generated, the real-time velocity component of the manufacturing equipment actuator within the current sampling period is read and subtracted from the current process reference speed issued by the process scheduling system. The absolute value of the difference is the speed deviation value. The speed tolerance is set to two percent of the current process reference speed. This value is based on the engineering allowable range of speed fluctuation of the manufacturing equipment actuator during normal processing. If it is less than one percent, it is too stringent, resulting in very few detection action injection windows. If it is more than five percent, the speed deviation has affected the processing accuracy and is not suitable for superimposing detection disturbances.

[0031] Critical process nodes are pre-marked by the process scheduling system according to the process specifications. Specifically, nodes that directly affect machining dimensional accuracy or equipment safety, such as clamping, tool setting, tool changing, workpiece handover, and the final stage of finishing, are marked as critical process nodes; other nodes are marked as non-critical process nodes. A detection drive command is issued to the manufacturing equipment actuator when both conditions are met: the speed deviation does not exceed the speed tolerance and the current process node is a non-critical process node. If either condition is not met, the detection drive command is temporarily suspended, the current control cycle continues to execute a conservative subset of commands, and the semantic blind zone assumption remains registered. The two conditions are re-evaluated in the next sampling cycle until both conditions are met before the detection drive command is injected again. This waiting mechanism ensures that detection actions are only injected when the manufacturing equipment actuator is operating stably and the process safety allows, without interfering with the normal machining cycle.

[0032] In one specific embodiment, step S3, recording the probe residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model, includes: In the next sampling period after the detection drive command is issued to the manufacturing equipment actuator, the displacement, velocity, torque and temperature of the manufacturing equipment actuator are collected to obtain the actual response vector of the actuator; The optimal detection action is input into the virtual model, and the displacement, velocity, torque and temperature after the detection drive command is executed are predicted to obtain the virtual model prediction response vector. The detection residual vector is obtained by subtracting the actual response vector of the actuator from the predicted response vector of the virtual model on a component-by-component basis.

[0033] Specifically, after the detection drive command is issued to the manufacturing equipment actuator, a complete sampling period Δt is waited for. At the end of this sampling period, the displacement, velocity, torque, and temperature of the manufacturing equipment actuator are synchronously read. The four-dimensional measured values ​​are arranged in the same order as the characteristic vector of the working condition, forming the actual response vector of the actuator. The synchronization between the acquisition time and the completion of the detection drive command is guaranteed by a hardware interrupt signal. The interrupt signal is triggered by the actuator controller after the current command is executed. The acquisition operation is executed immediately after the interrupt response, ensuring that the measured values ​​reflect the steady-state response after the completion of the detection drive command, rather than the transient values ​​during the transition process. The input for the virtual model to predict the response vector consists of two parts: one part is the state vector of the manufacturing equipment actuator at the moment before the detection drive command is issued, which serves as the initial state; the other part is the four-dimensional increment of the optimal detection action. The virtual model starts from the initial state and uses the optimal detection action as the driving quantity. It extrapolates one sampling period forward according to the kinematics and dynamics model of the equipment, outputting the predicted displacement, velocity, torque, and temperature. The four-dimensional predicted quantities are arranged in the same order to form the virtual model's predicted response vector.

[0034] The specific operation of component-by-component subtraction involves subtracting the displacement of the virtual model's predicted response vector from the actual displacement of the actuator's response vector to obtain the displacement residual; the difference in velocity quantities yields the velocity residual; the difference in torque quantities yields the torque residual; and the difference in temperature quantities yields the temperature residual. These four residual values ​​are arranged in sequence to form the probe residual vector. The dimensions of each component of the probe residual vector are in units of length, velocity, torque, and temperature, respectively. Since these four dimensions are different, direct comparison is meaningless. Therefore, the probe residual vector must be normalized before being passed into the execution layer semantic benchmark library update stage. The normalization method is to divide each component by the standard deviation of the corresponding sensor channel under known operating conditions. After eliminating the dimensional differences, the components are comparable. The square root of the sum of the squares of the normalized four-dimensional residual values ​​is the probe residual normalization amplitude. This amplitude value is used to determine the update type of the execution layer semantic benchmark library, whether it is an extension of the execution boundary or the generation of new semantic entries.

[0035] Figure 2This is a schematic diagram comparing the actual response of the actuator with the predicted response of the virtual model in an embodiment of this application. The solid line in the figure represents the measured displacement response value collected by the actuator of the manufacturing equipment after the detection drive command is executed. The dashed line represents the predicted response value of the virtual model to the same detection drive command before the execution layer semantic benchmark library is updated. The dotted line represents the predicted response value of the virtual model after the execution layer semantic benchmark library is updated. The filled area on the right axis corresponds to the detection residuals before and after the update, respectively. As can be seen from the figure, the residual amplitude between the predicted response and the actual response of the virtual model after the update is significantly narrowed, reflecting the correction effect of the execution layer semantic benchmark library on the boundary domain of the working condition after the detection residual vector feedback.

[0036] In one specific embodiment, step S4 includes: The normalized amplitude of the detection residual is obtained by taking the square root of the sum of the squares of the sum of the squares of each component of the detection residual vector divided by the standard deviation of the corresponding sensor channel under known operating conditions. Based on the relationship between the normalized magnitude of the probe residual and the preset magnitude threshold, the execution layer semantic benchmark library is updated in one of the following two ways: When the normalized magnitude of the probe residual does not exceed the preset magnitude threshold, the difference between the real-time working condition feature vector and the working condition feature vector of the nearest working condition boundary region in the execution layer semantic benchmark library is used to form a column vector. The covariance matrix of the nearest working condition boundary region is updated to rank one to obtain the expanded working condition boundary region. When the normalized magnitude of the probe residual exceeds the preset magnitude threshold, the probe residual vector is input into the structural causal model in the execution layer semantic benchmark library. The posterior probability of each behavioral intent label is calculated. The behavioral intent label with the highest posterior probability is combined with the real-time working condition feature vector and the initial covariance matrix to obtain a new semantic entry and write it into the execution layer semantic benchmark library. The updated execution layer semantic benchmark library and the cognitive layer confidence parameter set are re-executed with Mahalanobis distance matching and confidence output to obtain the current corrected control command; The modified control command is issued to the actuator of the manufacturing equipment.

[0037] Specifically, the normalization magnitude of the detection residual is obtained by taking the square root of the sum of the squares of each component of the detection residual vector divided by the standard deviation of the corresponding sensor channel. The preset magnitude threshold is set to 1.0. This value corresponds to the situation where the joint deviation of the normalized residuals of each sensor channel is within one standard deviation. A value below 1.0 is considered a slight extension of the existing working condition boundary domain, and a value above 1.0 is considered a new working condition region falling outside the existing semantic coverage. When the normalization magnitude does not exceed 1.0, the difference between the real-time working condition feature vector and the working condition feature vector of the center of the working condition boundary domain with the nearest Mahalanobis distance is taken to form a column vector. This column vector is multiplied by its own transpose and then multiplied by the learning rate. It is then superimposed on the original covariance matrix of the working condition boundary domain to obtain the expanded covariance matrix, which is the expanded working condition boundary domain. The learning rate is set to 0.1. This value controls the correction magnitude of the boundary domain by a single detection data. If the learning rate is too large, the boundary domain will be over-expanded due to a single noise disturbance. If the learning rate is too small, the boundary domain update speed will not be able to keep up with the slow drift of the working condition. When the normalization magnitude exceeds 1.0, the probe residual vector is used as new observation evidence and input into the structural causal model in the execution layer semantic benchmark library. The structural causal model calculates the posterior probability of each existing behavioral intent label under the observation evidence one by one, and takes the behavioral intent label with the highest posterior probability as the behavioral intent label of the new semantic entry. It is combined with the real-time working condition feature vector, and the initial covariance matrix is ​​set to 0.01 times the identity matrix. This value limits the initial coverage of the new semantic entry to a very small range, so as to avoid the new entry from over-covering the boundary domain of adjacent existing working conditions when the data accumulation is insufficient. After subsequent probe data is written, it is gradually expanded through rank-one update.

[0038] After the update, the Mahalanobis distance is recalculated for each of the real-time operating condition feature vectors and all operating condition boundary domains in the updated execution layer semantic benchmark library. The error sequence corresponding to the record with the smallest distance is then re-input into the Gaussian process regression model of the cognitive layer confidence parameter set, and the updated current confidence level is output. Based on the comparison between this confidence level value and the three-level switching threshold, the current gear is re-determined, and a corrected control command is generated according to the command calculation rules for the corresponding gear. Compared with the control command before the detection action injection, the corrected control command incorporates the new operating condition information carried by the detection residual vector into its gear selection basis. Boundary expansion or the writing of new semantic entries makes the execution layer semantic benchmark library more accurate in covering the current operating condition. The confidence level output by the cognitive layer confidence parameter set is closer to the true semantic credibility of the current operating condition. After the corrected control command is issued to the manufacturing equipment actuator, the actuator executes according to the corrected displacement component, velocity component, torque component, and temperature component, completing the command output of this round of virtual-real interactive closed-loop control.

[0039] Figure 3This is a schematic diagram illustrating how the normalization amplitude of the probe residual changes with the probe action number in an embodiment of this application. Circular markers indicate probe actions where the normalization amplitude does not exceed a preset amplitude threshold of 1.0, corresponding to the execution layer semantic benchmark library performing boundary domain expansion updates. Square markers indicate probe actions where the normalization amplitude exceeds the preset amplitude threshold of 1.0, corresponding to the execution layer semantic benchmark library performing new semantic entry generation. The horizontal dashed line is the preset amplitude threshold judgment baseline. The filled area in the diagram distinguishes the amplitude range corresponding to the two types of updates.

[0040] The virtual-real interactive digital twin full closed-loop control method in the embodiments of this application has been described above. The virtual-real interactive digital twin full closed-loop control system in the embodiments of this application is described below. One embodiment of the virtual-real interactive digital twin full closed-loop control system in the embodiments of this application includes: The acquisition module is used to collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and to establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set, which include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. The registration module is used to match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer. The current confidence level is output by the confidence level parameter set of the cognition layer. Based on the confidence level, the corresponding control command is selected from three levels: weighted fusion command, conservative subset command, and direct drive command and issued to the execution mechanism of the manufacturing equipment. When the confidence level is lower than a preset threshold, the semantic blind zone hypothesis is registered. The recording module is used to calculate the optimal detection action based on the semantic blind zone hypothesis, with an amplitude constraint of not exceeding the preset proportion of the conservative subset instruction norm, and to issue the optimal detection action to the manufacturing equipment actuator after superimposing the conservative subset instruction, and to record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. The generation module is used to perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library based on the probe residual vector, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

[0041] This invention also provides a virtual-real interactive digital twin fully closed-loop control device, which can be a server. The virtual-real interactive digital twin fully closed-loop control device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the virtual-real interactive digital twin fully closed-loop control device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the virtual-real interactive digital twin fully closed-loop control device stores the data corresponding to this embodiment. The network interface of the virtual-real interactive digital twin fully closed-loop control device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0042] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the virtual-real interactive digital twin full closed-loop control method.

[0043] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0044] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a virtual-physical interactive digital twin fully closed-loop control device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0045] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A virtual-real interactive digital twin full closed-loop control method, characterized in that, The method includes: Step S1: Collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set that include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. Step S2: Match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer, output the current confidence level from the confidence level parameter set of the cognitive layer, and select the corresponding control command from the three levels of weighted fusion command, conservative subset command, and direct drive command according to the confidence level to issue to the execution mechanism of the manufacturing equipment. When the confidence level is lower than the preset threshold, register the semantic blind zone hypothesis. Step S3: Based on the semantic blind zone hypothesis, with an amplitude constraint of not exceeding the preset proportion of the conservative subset instruction norm, calculate the optimal detection action, superimpose the optimal detection action onto the conservative subset instruction, and issue it to the manufacturing equipment actuator. Record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. Step S4: Perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library based on the probe residual vector, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

2. The virtual-real interaction digital twin full closed-loop control method according to claim 1, characterized in that, Step S1 includes: The displacement, velocity, torque and temperature data collected from the manufacturing equipment under known operating conditions constitute multimodal time-series sensing data. Based on the structural causal model, causal inference is performed on the multimodal time-series sensing data to obtain the causal strength coefficient between the operating condition variables and the behavioral intention variables. Based on the causal strength coefficient, the condition variables under each known working condition are aggregated to obtain the working condition feature vector and behavioral intention label corresponding to each known working condition. Based on the covariance matrix of the feature vectors of the working conditions, the coverage of each known working condition is modeled using a Mahalanobis distance hyperellipsoid to obtain the boundary domain of each known working condition. The working condition feature vector, the behavioral intention label, and the working condition boundary domain are written into the execution layer semantic benchmark library. Based on the error sequence between the predicted behavioral intention value and the actual observed value in the historical control cycle, the error sequence is fitted by Gaussian process regression to obtain the cognitive layer confidence parameter set.

3. The virtual-real interactive digital twin full closed-loop control method according to claim 1, characterized in that, Step S2 includes: The displacement, velocity, torque, and temperature of the manufacturing equipment during the current sampling period are input into the semantic reference library of the execution layer. The Mahalanobis distance is calculated on the feature vectors of each working condition boundary domain to obtain the matching distance value between the current real-time working condition feature vector and each working condition boundary domain. Based on the matching distance value, the current confidence level is extracted from the confidence level parameter set of the cognitive layer. Based on the relationship between the current confidence level and the preset threshold, the control command of the corresponding level is determined from the three levels of weighted fusion command, conservative subset command, and direct drive command, and the control command is issued to the manufacturing equipment actuator. Based on the determination result that the current confidence level is lower than the preset threshold, the real-time working condition feature vector, the mean error and the error variance within the current time window are associated and written to obtain the semantic blind zone hypothesis carrying working condition positioning information and error statistics.

4. The virtual-real interactive digital twin full closed-loop control method according to claim 3, characterized in that, In step S3, based on the semantic blind zone hypothesis, and with a magnitude constraint not exceeding a preset proportion of the conservative subset instruction norm, the optimal detection action is calculated, including: The real-time operating condition feature vector and the error statistic in the semantic blind zone hypothesis are input into the execution layer semantic benchmark library to locate the coverage gap between each operating condition boundary domain and the real-time operating condition feature vector, and obtain the semantic coverage gap vector. Based on the semantic coverage gap vector, the reduction in cognitive layer information entropy after applying each candidate detection action in the candidate detection action set is calculated to obtain the information gain value corresponding to each candidate detection action. The Euclidean norm of each candidate detection action is compared with a preset ratio of the norm of the conservative subset instruction, and candidate detection actions whose Euclidean norm exceeds the preset ratio are filtered out to obtain a subset of candidate detection actions that satisfy the amplitude constraint. The optimal detection action is obtained by extracting the candidate detection action with the largest information gain value from the subset of candidate detection actions.

5. The virtual-real interactive digital twin full closed-loop control method according to claim 4, characterized in that, In step S3, the optimal detection action is superimposed on the conservative subset instruction and then issued to the manufacturing equipment execution mechanism, including: The displacement increment, velocity increment, torque increment, and temperature increment in the optimal detection action are superimposed component by component with the corresponding displacement component, velocity component, torque component, and temperature component in the conservative subset instruction to obtain the detection drive instruction. The speed deviation value is obtained by calculating the difference between the current speed component of the manufacturing equipment actuator and the current process reference speed. The speed deviation value is compared with the speed tolerance. When the speed deviation value does not exceed the speed tolerance and the current process node is a non-critical process node, the detection drive command is issued to the manufacturing equipment actuator.

6. The virtual-real interactive digital twin full closed-loop control method according to claim 5, characterized in that, The step S3, which records the probe residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model, includes: In the next sampling period after the detection drive command is issued to the manufacturing equipment actuator, the displacement, velocity, torque and temperature of the manufacturing equipment actuator are collected to obtain the actual response vector of the actuator; The optimal detection action is input into the virtual model, and the displacement, velocity, torque and temperature after the detection drive command is executed are predicted to obtain the virtual model prediction response vector. The actual response vector of the actuator is subtracted from the predicted response vector of the virtual model on a component-by-component basis to obtain the detection residual vector.

7. The virtual-real interactive digital twin full closed-loop control method according to claim 1, characterized in that, Step S4 includes: The square root of the sum of the squares of the components in the detection residual vector is obtained by dividing each component by the standard deviation of the corresponding sensor channel under known operating conditions. Based on the relationship between the normalized magnitude of the probe residual and the preset magnitude threshold, one of the following two types of updates is performed on the execution layer semantic benchmark library: When the normalized magnitude of the probe residual does not exceed the preset magnitude threshold, the difference between the real-time operating condition feature vector and the operating condition feature vector of the nearest operating condition boundary region in the execution layer semantic benchmark library is taken to form a column vector, and the covariance matrix of the nearest operating condition boundary region is updated to rank one to obtain the expanded operating condition boundary region; When the normalized magnitude of the probe residual exceeds the preset magnitude threshold, the probe residual vector is input into the structural causal model in the execution layer semantic benchmark library, the posterior probability of each behavioral intent label is calculated, and the behavioral intent label with the highest posterior probability is combined with the real-time operating condition feature vector and the initial covariance matrix to obtain a new semantic entry and write it into the execution layer semantic benchmark library; The updated execution layer semantic benchmark library and the cognitive layer confidence parameter set are re-executed with Mahalanobis distance matching and confidence output to obtain the current corrected control instruction; The correction control command is issued to the manufacturing equipment actuator.

8. A fully closed-loop control system for virtual-real interactive digital twins, characterized in that, For implementing the virtual-real interactive digital twin full closed-loop control method as described in any one of claims 1-7, the virtual-real interactive digital twin full closed-loop control system comprises: The acquisition module is used to collect multimodal time-series sensing data of manufacturing equipment under known operating conditions, and to establish an execution layer semantic benchmark library and a cognitive layer confidence parameter set, which include operating condition feature vectors, behavioral intention labels and operating condition boundary domains. The registration module is used to match the real-time operating condition feature vector of the manufacturing equipment with the boundary domains of each operating condition in the semantic benchmark library of the execution layer. The current confidence level is output by the confidence level parameter set of the cognition layer. Based on the confidence level, the corresponding control command is selected from three levels: weighted fusion command, conservative subset command, and direct drive command and issued to the execution mechanism of the manufacturing equipment. When the confidence level is lower than a preset threshold, the semantic blind zone hypothesis is registered. The recording module is used to calculate the optimal detection action based on the semantic blind zone hypothesis, with an amplitude constraint of not exceeding the preset proportion of the conservative subset instruction norm, and to issue the optimal detection action to the manufacturing equipment actuator after superimposing the conservative subset instruction, and to record the detection residual vector between the actual response of the manufacturing equipment actuator and the predicted response of the virtual model. The generation module is used to perform boundary expansion or generate new semantic entries for the execution layer semantic benchmark library based on the probe residual vector, and continuously issue correction control commands to the manufacturing equipment actuator based on the updated execution layer semantic benchmark library and the cognitive layer confidence parameter set.

9. A virtual-real interactive digital twin fully closed-loop control device, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the virtual-real interactive digital twin full closed-loop control method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the virtual-real interactive digital twin full closed-loop control method as described in any one of claims 1 to 7.

Images