Real-time monitoring method, device and equipment for intelligent manufacturing production line and medium

Abstract

The invention relates to an intelligent manufacturing production line real-time monitoring method and device, equipment and a medium. The method comprises the following steps: acquiring observation data and simulation output data in real time, performing time synchronization alignment processing to obtain a physical-virtual sensor synchronization data stream, and aggregating all comprehensive mismatch degrees according to the comprehensive mismatch degrees of each physical sensor-virtual sensor pair to obtain a global mismatch index; when the global mismatch index exceeds a preset threshold value, identifying a mismatch salient region based on each comprehensive mismatch degree, and performing sensitivity analysis on key physical parameters of the mismatch salient region to obtain a suspicious parameter subset; constructing a loss function according to the observation data and the simulation output data, and minimizing the loss function to solve the suspicious parameter subset to obtain an optimal parameter correction value; and updating the model parameters based on the optimal parameter correction value to obtain a corrected digital twinborn model. By adopting the method, the digital twin model can keep high-fidelity virtual mapping capability in the full life cycle of the production line.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, and in particular relates to a method, device, equipment and medium for real-time monitoring of intelligent manufacturing production lines. Background Technology

[0002] With the development of intelligent manufacturing and Industry 4.0 technologies, digital twin technology has emerged. By constructing a high-fidelity virtual mapping of the physical production line and realizing real-time interaction and collaborative simulation between the virtual and real spaces, it provides core technical support for the intelligent operation of the production line.

[0003] In traditional technologies, addressing the mismatch between the state of digital twins and physical entities mainly relies on periodic manual inspections and offline calibration. This involves technicians manually adjusting the internal parameters of the digital twin model using production line data collected at specific times to compensate for the discrepancy between changes in the physical entity and the static settings of the model.

[0004] However, the above methods suffer from significant problems such as low efficiency and slow response. They cannot capture sudden or gradual minor anomalies during production line operation, nor can they achieve continuous and accurate mapping of the physical production line throughout its entire life cycle. The few studies that attempt to introduce adaptive filtering or parameter estimation methods are mostly focused on adjusting single equipment or local parameters. They lack a mismatch monitoring and correction system for the entire production line, multiple dimensions, and closed-loop automation. If the mismatch between the state of the digital twin and the physical entity cannot be effectively resolved, it will significantly reduce the credibility of real-time monitoring, fault diagnosis, capacity prediction, and process optimization based on digital twins, and may even lead to serious consequences such as false alarms and misscheduling. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for real-time monitoring of intelligent manufacturing production lines that can quantify the degree of mismatch in real time, automatically locate the root cause of mismatch, and complete model parameter identification and correction online, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for real-time monitoring of an intelligent manufacturing production line, including:

[0007] The system collects observation data from multiple physical sensors on the physical production line in real time and simulation output data from multiple virtual sensors deployed in the corresponding physical sensors in the digital twin model. It then performs time synchronization and alignment processing on the observation data and simulation output data to obtain a physical-virtual sensor synchronized data stream.

[0008] Based on a preset sliding time window, the comprehensive mismatch degree between the observation data and simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream is calculated, and the comprehensive mismatch degree of all physical sensor-virtual sensor pairs is aggregated to obtain the global mismatch index.

[0009] When the global mismatch index exceeds a preset threshold, the significant mismatch region is identified based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair, and the key physical parameters of the corresponding significant mismatch region in the digital twin model are subjected to sensitivity analysis to obtain a subset of suspicious parameters that lead to mismatch.

[0010] Based on the differences between the observation data and simulation output data of each physical sensor-virtual sensor pair in the significant mismatch region, a loss function for the suspicious parameter subset is constructed. With the goal of minimizing the loss function, an online optimization algorithm is used to solve the suspicious parameter subset and obtain the optimal parameter correction value.

[0011] Based on the optimal parameter correction value, the model parameters of the corresponding suspicious parameter subset in the digital twin model are incrementally updated to obtain the corrected digital twin model.

[0012] In one embodiment, based on a preset sliding time window, the comprehensive mismatch between the observation data and simulation output data of each physical sensor-virtual sensor pair within the physical-virtual sensor synchronous data stream is calculated, including:

[0013] Based on a preset sliding time window, the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair are extracted from the physical-virtual sensor synchronous data stream;

[0014] The mean, standard deviation, skewness, and kurtosis of the observed data sequence and the simulated output data sequence of each physical sensor-virtual sensor pair are calculated respectively to obtain the statistical characteristics of each physical sensor-virtual sensor pair.

[0015] Based on statistical characteristics, the static feature mismatch degree of each physical sensor-virtual sensor pair is calculated using weighted Euclidean distance; the expression for the static feature mismatch degree is: ,in, , , and These are the mean, standard deviation, skewness, and kurtosis of the observed data series, respectively. , , and These represent the mean, standard deviation, skewness, and kurtosis of the simulation output data sequence, respectively. , , and These are weighting coefficients preset according to the sensor type;

[0016] Short-time Fourier transforms are performed on the observation data sequences and simulation output data sequences of each physical sensor-virtual sensor pair to obtain the physical time spectrum and virtual time frequency of each physical sensor-virtual sensor pair.

[0017] For each time frame, the physical time spectrum and the virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution of each physical sensor-virtual sensor pair.

[0018] Based on the probability distribution of the physical sequence and the probability distribution of the virtual sequence, calculate the KL divergence on each time frame, and calculate the average value of the KL divergence of all time frames within the entire sliding time window to obtain the dynamic characteristic mismatch of each physical sensor-virtual sensor pair.

[0019] After normalizing the static feature mismatch and dynamic characteristic mismatch of each physical sensor-virtual sensor pair, the normalized static feature mismatch and dynamic characteristic mismatch are weighted and fused to obtain the comprehensive mismatch of each physical sensor-virtual sensor pair.

[0020] In one embodiment, for each time frame, the physical time spectrum and the virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution for each physical sensor-virtual sensor pair, including:

[0021] For each time frame, the physical energy values ​​of the physical time spectrum at each frequency are summed to obtain the total physical energy. Based on the physical energy values ​​and the total physical energy, the physical probability values ​​of the observed data sequence at each frequency are calculated to obtain the physical sequence probability distribution. The expression for the physical probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The physical energy value at that location For time frames The total physical energy;

[0022] For the same time frame, the virtual energy values ​​of the virtual time spectrum at each frequency are summed to obtain the virtual total energy. Based on the virtual energy values ​​and the virtual total energy, the virtual probability values ​​of the simulation output data sequence at each frequency are calculated to obtain the virtual sequence probability distribution. The expression for the virtual probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The virtual energy value at that location, For time frames The virtual total energy.

[0023] In one embodiment, sensitivity analysis is performed on key physical parameters corresponding to significant mismatch regions in the digital twin model to obtain a subset of suspected parameters leading to the mismatch, including:

[0024] Multiple key physical parameters associated with the dynamic behavior of the significantly mismatched region are extracted from the digital twin model to form an initial parameter set; the current values ​​of each key physical parameter in the initial parameter set are nominal values.

[0025] Multiple virtual sensors related to the significant mismatch region are selected from the digital twin model, and based on the initial parameter set, the partial derivatives of each key physical parameter at the nominal value are calculated on the simulation output data of each virtual sensor to obtain the parameter sensitivity.

[0026] The parameter sensitivities of all virtual sensors associated with significant mismatch regions are aggregated to generate a sensitivity matrix;

[0027] The covariance matrix is ​​calculated based on the sensitivity matrix, and eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and eigenvectors of each principal component.

[0028] The principal components are sorted from largest to smallest according to their eigenvalues, and the cumulative contribution rate of each principal component is calculated based on the sorted eigenvalues.

[0029] For principal components whose cumulative contribution rate exceeds a preset threshold, the original physical parameters whose absolute load values ​​on each principal component exceed a preset load threshold are identified, and a subset of suspicious parameters is constructed based on the identified original physical parameters.

[0030] In one embodiment, a loss function for a subset of suspicious parameters is constructed based on the differences between the observed data and simulation output data of each physical sensor-virtual sensor pair within a significant mismatch region, including:

[0031] The loss function is obtained using the following formula:

[0032]

[0033]

[0034] in, The loss function; A subset of suspicious parameters; For physical sensor-virtual sensor pair index set; For the first Weighting coefficients for each physical sensor-virtual sensor pair; For the first The mean square error term of a physical sensor-virtual sensor pair; Let be the dynamic characteristic mismatch degree of the i-th physical sensor-virtual sensor pair; The length of the sliding time window; For observation data sequences; This is a sequence of simulation output data obtained by simulating key physical parameters of a subset of suspicious parameters.

[0035] In one embodiment, based on the optimal parameter correction value, the model parameters corresponding to the suspicious parameter subset in the digital twin model are incrementally updated to obtain the corrected digital twin model, including:

[0036] The current parameter vector is obtained based on the nominal values ​​of each key physical parameter within the subset of suspicious parameters;

[0037] The current parameter vector is incrementally updated based on the optimal parameter correction value to obtain a new parameter vector; the incremental update expression is: ,in, The learning rate is preset, and , For the current parameter vector, This is the optimal parameter correction value. For the new parameter vector;

[0038] The corrected digital twin model is obtained by updating the parameter values ​​of the suspicious parameter subset in the digital twin model based on the new parameter vector.

[0039] In one embodiment, the method further includes:

[0040] A complete simulation cycle is performed based on the corrected digital twin model to generate updated simulation output data from the virtual sensor.

[0041] Based on the updated simulation output data of the virtual sensor, combined with the observation data of the physical sensor collected in the corresponding time series, the global mismatch index is calculated to obtain the corrected verification index.

[0042] Determine whether the corrected verification index is lower than the preset threshold, and calculate the percentage decrease of the corrected verification index relative to the global mismatch index when correction is triggered. If the corrected verification index is lower than the preset threshold and the percentage decrease is greater than the preset minimum improvement rate, then output the corrected digital twin model.

[0043] If the corrected verification index exceeds the preset threshold or the decrease rate is less than the preset minimum improvement rate, the current parameter vector is retained and the information of this correction failure is recorded.

[0044] Secondly, this application also provides a real-time monitoring device for intelligent manufacturing production lines, comprising:

[0045] The data module is used to collect observation data from multiple physical sensors on the physical production line and simulation output data from multiple virtual sensors deployed in the digital twin model corresponding to the physical sensors in real time, and to perform time synchronization and alignment processing on the observation data and simulation output data to obtain a physical-virtual sensor synchronized data stream.

[0046] The mismatch detection module is used to calculate the comprehensive mismatch degree between the observation data and simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream based on a preset sliding time window, and to aggregate the comprehensive mismatch degree of all physical sensor-virtual sensor pairs to obtain the global mismatch index.

[0047] The suspicious parameter module is used to identify significant mismatch regions based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair when the global mismatch index exceeds a preset threshold, and to perform sensitivity analysis on the key physical parameters of the corresponding significant mismatch regions in the digital twin model to obtain a subset of suspicious parameters that lead to mismatch.

[0048] The correction module is used to construct a loss function for a subset of suspicious parameters based on the difference between the observation data and simulation output data of each physical sensor-virtual sensor pair in the significant mismatch region. With the goal of minimizing the loss function, an online optimization algorithm is used to solve the subset of suspicious parameters to obtain the optimal parameter correction value.

[0049] The model update module is used to incrementally update the model parameters of the corresponding suspicious parameter subset in the digital twin model based on the optimal parameter correction value, so as to obtain the corrected digital twin model.

[0050] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described intelligent manufacturing production line real-time monitoring methods.

[0051] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-described intelligent manufacturing production line real-time monitoring methods.

[0052] The aforementioned intelligent manufacturing production line real-time monitoring method, device, equipment, and medium collects physical sensor observation data and digital twin virtual sensor simulation data in real time and synchronizes them to generate a synchronous data stream. Then, based on a sliding window, it calculates the comprehensive mismatch degree of each sensor pair and aggregates it into a global mismatch index. When the index exceeds the threshold, it locates the significant mismatch area based on the mismatch degree and performs sensitivity analysis on the key parameters of the digital twin model to identify a subset of suspicious parameters. Based on the difference between the measured and simulated data in this area, it constructs a loss function for this subset and uses an online optimization algorithm to solve for the optimal parameter correction value. Finally, it uses this correction value to incrementally update the corresponding parameters in the model, thereby realizing a fully closed-loop, automated online self-correction of the digital twin model. This effectively solves the model mismatch problem caused by the wear and aging of physical entities and significantly improves the real-time performance, accuracy, and reliability of the digital twin model in production line monitoring, prediction, and decision-making. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the real-time monitoring method for intelligent manufacturing production lines according to the present invention.

[0055] Figure 2 This is a structural diagram of the intelligent manufacturing production line real-time monitoring device of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] In one embodiment, such as Figure 1 As shown, a method for real-time monitoring of an intelligent manufacturing production line is provided. This embodiment illustrates the method applied to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0058] S101. Real-time acquisition of observation data from multiple physical sensors on the physical production line and simulation output data from multiple virtual sensors deployed in the digital twin model corresponding to the physical sensors, and time synchronization and alignment processing of the observation data and simulation output data to obtain a physical-virtual sensor synchronous data stream.

[0059] In a schematic representation, the deployment of sensors on the physical production line is based on the operating characteristics of key equipment and the monitoring needs of process nodes, achieving full-dimensional status perception coverage. The types of physical sensors deployed are directly matched to the monitored physical quantities, including vibration sensors for capturing equipment vibration status and current sensors for monitoring motor operating load. Various sensors are deployed according to the topology and process flow of the production line equipment, and the collected data comprehensively reflects the operating status and equipment interaction behavior of the physical production line. The deployment of virtual sensors in the digital twin model follows the deployment logic of physical sensors. Their type, measurement principle, and virtual space installation coordinates must be completely consistent with the corresponding physical sensors. The simulation output data of the virtual sensors is generated through the core simulation engine of the digital twin model. The geometric model provides accurate topological structure and assembly relationship support, and the physical behavior model calculates the simulated physical quantities corresponding to the objects measured by the physical sensors through multibody dynamics equations, finite element analysis, or data-driven proxy models.

[0060] Optionally, after data acquisition, the physical sensor observation data and virtual sensor simulation output data are aggregated through the data interface between the industrial IoT gateway and the digital twin simulation platform. A unified time reference is established using a high-precision time synchronization protocol. By embedding a time synchronization module in the industrial IoT gateway and the digital twin simulation engine, a unique timestamp is assigned to each piece of acquired physical sensor observation data and virtual sensor simulation output data. The timestamp accuracy must reach the millisecond level to ensure consistent matching in the time dimension. To address potential differences in sampling frequencies between physical and virtual sensors, a data standardization process is adopted. Specifically, outlier removal is performed on the raw observation data and simulation output data. Based on the statistical confidence interval method, abnormal data jumps caused by communication interference, sensor failure, or simulation fluctuations are identified and removed to avoid the impact of abnormal data on synchronization accuracy. Interpolation algorithms are used to unify the time-series reference of data streams with different sampling rates. The choice of interpolation method is determined according to the data type. For continuously changing physical quantities, linear interpolation or cubic spline interpolation is used, while for discrete event data, nearest neighbor interpolation is used. Through interpolation calculation, all data streams are unified to a preset standard sampling period, ultimately forming a physical-virtual sensor synchronized data stream.

[0061] S102. Based on a preset sliding time window, calculate the comprehensive mismatch between the observation data and simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream, and aggregate the comprehensive mismatch of all physical sensor-virtual sensor pairs to obtain the global mismatch index.

[0062] Optionally, the preset sliding time window setting comprehensively considers the production line's operating rhythm and data change characteristics. The window length is determined to capture both the statistical characteristics and dynamic change patterns of the data while ensuring the real-time performance of mismatch calculation. The window adopts a continuous sliding mode, with the sliding step size consistent with the standard sampling period, ensuring a balance between data coverage continuity and computational efficiency. Within each sliding time window, for each pair of physical sensors and virtual sensors, the corresponding physical observation data sequence and virtual simulation data sequence are extracted respectively. The time dimensions of the two sequences are perfectly aligned, ensuring that each time node has a corresponding physical observation value and virtual simulation value.

[0063] Specifically, the comprehensive mismatch is calculated by fusing static feature mismatch and dynamic feature mismatch. The static feature mismatch measures the difference in statistical characteristics between two data sequences, specifically selecting four key indicators: mean, standard deviation, skewness, and kurtosis. The mean reflects the central tendency of the data, the standard deviation characterizes the dispersion, skewness describes the asymmetry of the data distribution, and kurtosis reflects the steepness of the data distribution, thus comprehensively reflecting the static statistical characteristics of the data. Static Feature Mismatch ,in, , , , These are the weighting coefficients for mean, standard deviation, skewness, and kurtosis, respectively. The weighting coefficients are determined through offline calibration based on the importance of the physical quantities monitored by the sensor and their sensitivity to mismatch. For physical quantities that are sensitive to changes in equipment status, such as vibration and current, the weighting coefficient corresponding to the standard deviation is higher. For physical quantities that tend to be stable, such as temperature and pressure, the weighting coefficient corresponding to the mean is higher. The mean of the physical observation data sequence. The mean of the virtual simulation data sequence; The standard deviation of the physical observation data series. The standard deviation of the virtual simulation data sequence; The skewness of the physical observation data sequence. The skewness of the virtual simulation data sequence; The kurtosis of the physical observation data sequence. This represents the kurtosis of the virtual simulation data sequence.

[0064] The dynamic mismatch is quantified using the Kullback-Leibler (KL) divergence method to measure the dynamic difference in the time or frequency domain distribution of two data sequences, i.e., by analyzing the physical observation data sequences. and virtual simulation data sequence Perform short-time Fourier transforms on each to obtain the time spectrum. and ,in For frequency variables, For time frame variables, the Hanning window is chosen as the window function for the short-time Fourier transform. The window length is determined based on the data sampling frequency and the signal characteristic frequency to ensure effective differentiation of the signal's time-frequency variations. Furthermore, for each time frame... The time-frequency spectrum is normalized to obtain the probability distribution in the frequency dimension. and ,in For physical observation data, the spectrum in time frames The normalized probability distribution under the following conditions When the spectrum is used for virtual simulation data in time frames The normalized probability distribution is given by KL divergence. Dynamic characteristic mismatch The mean of the KL divergence of all time frames within the sliding time window is obtained by arithmetically averaging the KL divergence of each time frame, thus providing a quantitative indicator that reflects the dynamic distribution differences of the data within the entire window.

[0065] The overall mismatch degree is obtained by weighted fusion of static feature mismatch degree and dynamic characteristic mismatch degree, that is... ,in, The fusion coefficient has a range of values. The weighting is determined based on the relative importance of static and dynamic characteristics in the monitoring scenario. For example, in equipment wear monitoring, dynamic characteristics have a higher weighting. The value is too small; This is a preset threshold for the static feature mismatch degree. The preset thresholds for dynamic characteristic mismatch are obtained through statistical analysis of historical normal operation data. That is, based on physical-virtual sensor data under normal production line conditions, the statistical upper limit values ​​of static characteristic mismatch and dynamic characteristic mismatch are calculated as thresholds.

[0066] The global mismatch index is calculated by weighted aggregation of the overall mismatch degree of all physical sensor-virtual sensor pairs, i.e. ,in, This represents the total number of physical sensor-virtual sensor pairs. For the first The weighting coefficient of a sensor is determined based on its criticality to the overall performance of the production line from the equipment or process in which it is located. Sensors in key equipment and core parts have higher weighting coefficients. For indicator functions, when When the condition is met, the indicator function takes the value of 1; otherwise, it takes the value of 0. Only sensor pairs with significant mismatch are included in the global mismatch index calculation. For the first The overall mismatch of sensors. The global mismatch index can comprehensively reflect the overall deviation between the digital twin model and the physical entity of the entire production line.

[0067] S103. When the global mismatch index exceeds the preset threshold, the significant mismatch region is identified based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair, and the key physical parameters corresponding to the significant mismatch region in the digital twin model are subjected to sensitivity analysis to obtain a subset of suspicious parameters that lead to mismatch.

[0068] As an illustration, the preset threshold for the global mismatch index is set based on the requirements for production line operation safety and monitoring accuracy, and is determined through historical mismatch case data and simulation verification. When the calculated global mismatch index... When the threshold is exceeded, it indicates that the deviation between the digital twin model and the physical production line has affected monitoring reliability, requiring the initiation of the mismatch root cause localization process. Specifically, the identification of significant mismatch areas is based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair. Based on this, all sensor pairs are sorted in descending order of their overall mismatch degree. The physical regions containing the top-ranked sensor pairs are selected as significant mismatch areas. The number of selected regions needs to be determined based on the production line scale and sensor deployment density to ensure coverage of the main mismatch areas while avoiding excessively large areas that could reduce positioning accuracy. The delineation of significant mismatch areas needs to be associated with the spatial topology of the digital twin model. Each sensor pair corresponds to a specific spatial location and equipment component in the model. Through the mapping relationship between sensor identifiers and model spatial coordinates, the specific range of the significant mismatch area in the digital twin model is clarified, including the corresponding equipment, components, and process steps.

[0069] The selection of key physical parameters focuses on the dynamic behavior and operational characteristics of regions with significant mismatch. The selected key physical parameters are core parameters that directly affect the virtual simulation output of these regions. These include, but are not limited to, the friction coefficient, stiffness coefficient, and damping coefficient of mechanical components; the torque constant and moment of inertia of motors; the elastic modulus and thermal conductivity of materials; and pressure and temperature parameters during the manufacturing process. The selection of key physical parameters is based on the modeling logic of the digital twin model. Specifically, by analyzing the mathematical equations describing the physical behavior of regions with significant mismatch in the model, the independent variables in the equations are extracted as key physical parameters, ensuring that the selected parameters are directly related to the simulation output of these regions.

[0070] Sensitivity analysis aims to identify the key physical parameters that have the most significant impact on the virtual sensor output. It employs a parametric sensitivity analysis technique based on the adjoint method, which quantifies the degree of influence of parameter changes on the simulation output by calculating the partial derivatives of the virtual sensor output with respect to each key physical parameter. For example, a parameterized local digital twin sub-model is constructed for regions with significant mismatch. This sub-model encapsulates a set of key physical parameters related to the region's dynamic behavior. ,in The number of key physical parameters, For the first The modeling accuracy of the sub-model is consistent with that of the global digital twin model, focusing only on areas of significant mismatch to improve computational efficiency.

[0071] The parameter sensitivity was calculated using a combination of perturbation simulation and adjoint equation solving. Specifically, at the nominal values ​​of key physical parameters... Set up small perturbations nearby. For each key physical parameter... After applying a perturbation, the local digital twin sub-model is rerun to obtain the corresponding changes in the virtual sensor output. The approximate sensitivity of the parameters is then calculated using differential methods. Simultaneously, an adjoint equation is constructed based on the physical equations of the sub-model. The accurate analytical solution for the sensitivity is obtained by directly solving the adjoint equation. The results of the two methods are cross-validated to ensure the accuracy of the sensitivity calculation. Sensitivity The definition of virtual sensor output For key physical parameters The partial derivative at the nominal value, i.e. ,in, This indicates that nominal values ​​are taken for key physical parameters. At that time, the virtual sensor output For the Key physical parameters Partial derivatives, sensitivity The larger the absolute value of , the more significant the impact of the parameter change on the virtual sensor output, and the higher the probability that the parameter will become the root cause of mismatch.

[0072] The sensitivity of all key physical parameters constitutes a sensitivity vector. Constructing a sensitivity matrix based on sensitivity vectors ,in The matrix represents the number of virtual sensors within a significant mismatch region. Each row corresponds to the sensitivity of a virtual sensor to each key physical parameter, and each column corresponds to the sensitivity of a key physical parameter to each virtual sensor. To select the smallest subset of parameters from the set of key physical parameters that can explain most of the mismatch phenomenon, principal component analysis (PCA) or sparse regression techniques are used to analyze the sensitivity matrix. PCA extracts the principal components with the largest variance contribution by performing eigenvalue decomposition on the sensitivity matrix; the key physical parameters corresponding to these principal components are the parameters with the most significant impact on the mismatch. Sparse regression constructs an L1-regularized regression model, using the difference between the virtual sensor output and the physical sensor observation data as the dependent variable and the change in key physical parameters as the independent variable. The sparsity of the regression coefficients is used to select the parameters that contribute the most to the mismatch. Finally, the subset of parameters obtained through sensitivity analysis is the suspected parameter subset leading to the mismatch. .

[0073] S104. Based on the differences between the observation data and simulation output data of each physical sensor-virtual sensor pair in the significant mismatch region, construct the loss function of the suspicious parameter subset, and use an online optimization algorithm to solve the suspicious parameter subset with the goal of minimizing the loss function, so as to obtain the optimal parameter correction value.

[0074] Schematic, the loss function is ,in, This is a subset of suspicious parameters, containing several key physical parameters that need to be optimized. This is an index set of physical sensor-virtual sensor pairs within a significant mismatch region, containing only sensor pairs directly related to the significant mismatch region. For the first The weighting coefficients for sensors are determined based on the reliability of the sensor data and its ability to characterize mismatches. Sensors with high data stability and strong characterization ability have higher weighting coefficients. This represents the number of data sampling points within the sliding time window. For the first For the physical sensor in the sensor at time Observational data; For the first For the virtual sensor in the sensor at time Suspicious parameter subset selection Simulation output data at that time; For the first Probability distribution of sensor physical observation data; For the first Extracting sensor virtual simulation data from a subset of suspicious parameters The probability distribution at time; This is the KL divergence between the probability distributions of physical observation data and virtual simulation data, used to quantify the differences between the two types of data distributions.

[0075] Optionally, the online optimization algorithm employs Bayesian optimization, replacing complex loss function calculations with a probabilistic surrogate model. The surrogate model uses a Gaussian process model, which can construct a probability distribution estimate of the loss function using a limited number of sample points and provides quantification of the uncertainty of the prediction results. The kernel function of the Gaussian process model is a squared exponential kernel, balancing function smoothness and local feature capture ability. The kernel function parameters are determined using the maximum likelihood estimation method. The choice of the acquisition function directly affects the optimization efficiency; the Expected Improvement (EI) function is used as the acquisition function. The EI function balances exploring unsampled regions with utilizing known optimal regions, guiding the optimization process to efficiently search for optimal parameters. The Expected Improvement (EI) function is... ,in, A parameter combination within a subset of suspicious parameters; This represents the minimum value of the loss function that has been found so far. For parameter combination The corresponding loss function value; The expectation operator is defined as follows: the larger the value of the expected improvement function, the higher the probability that the parameter combination will obtain a better loss function value.

[0076] Specifically, initial sample points are randomly selected within the parameter space of the subset of suspicious parameters. A Gaussian process surrogate model is constructed based on the loss function values ​​of these initial sample points. The next parameter combination to be evaluated is determined by maximizing the expected improvement function. This parameter combination is then input into the digital twin model to calculate the corresponding loss function value. The new sample points, i.e., the parameter combination and the corresponding loss function value, are added to the sample set, and the Gaussian process surrogate model is updated. The above steps are repeated until the iterative convergence conditions are met. The iterative convergence conditions include the change in the minimum value of the loss function being less than a preset convergence threshold, the maximum value of the expected improvement function being less than a preset exploration threshold, and the number of iterations reaching a preset maximum number of iterations, ensuring that the optimization process is completed within the time constraint. Optimal parameter correction value. The parameter combination that minimizes the loss function after iterative convergence, i.e. ,in, The parameter space is a subset of suspicious parameters. This space is determined by the physical reasonable range of the parameters. That is, the value of each parameter must conform to the actual physical laws to avoid meaningless parameter values. Find the loss function The combination of parameters that achieves the minimum value To ensure the reliability of the optimal parameter correction values, their physical rationality needs to be verified. This involves checking whether the correction value of each parameter is within the preset physical rationality range. If any parameter exceeds the range, the parameter space constraints need to be adjusted and the optimization solution needs to be performed again until the optimal parameter correction values ​​that conform to physical laws are obtained.

[0077] S105. Based on the optimal parameter correction value, the model parameters of the corresponding suspicious parameter subset in the digital twin model are incrementally updated to obtain the corrected digital twin model.

[0078] This illustration demonstrates how incremental parameter updates prevent sudden parameter changes that could lead to instability in the digital twin model simulation, ensuring consistency between model evolution and changes in the physical entity's state. The incremental update formula is as follows: ,in, This is the updated set of model parameters; To update the current values ​​of the subset of suspicious parameters in the digital twin model before updating; The learning rate has a range of values. This is used to control the step size of each parameter update; The optimal parameter correction value is the target parameter value obtained through online optimization algorithms. The learning rate setting needs to comprehensively consider model stability and correction efficiency. The initial learning rate is determined through historical correction data statistics to ensure smooth convergence of the parameter update process. Simultaneously, an adaptive adjustment mechanism is adopted, dynamically adjusting the learning rate based on the global mismatch index change trend during the correction process. When the decrease rate of the global mismatch index after two consecutive updates exceeds a preset threshold, the learning rate is appropriately increased to accelerate the correction speed; when the decrease rate is below the preset threshold or fluctuations occur, the learning rate is decreased to avoid parameter oscillations and ensure the stability of model updates. Optionally, strict physical boundary constraints are applied during the parameter update process to ensure that the updated model parameters always remain within a physically reasonable range, avoiding simulation result distortion due to parameter out-of-bounds errors. After parameter update calculation, boundary checks are performed on the update results of each parameter. If the updated parameter value exceeds the physically reasonable range, it is truncated to the nearest boundary value, and the parameter out-of-bounds event is recorded to provide a reference for subsequent optimization of suspicious parameter subsets.

[0079] For example, parameter updates for digital twin models require a modular update mechanism. The digital twin model is divided into multiple independent parameter modules based on equipment, components, or process steps. Parameters corresponding to a subset of suspicious parameters belong to a specific module. During the update process, only the parameters of that module are modified, while the parameters of other modules remain unchanged, ensuring the stability of the overall model structure. Parameter update operations are implemented through the parameter interface of the digital twin model. This interface supports real-time parameter writing and model re-initialization. After the update is completed, there is no need to restart the entire digital twin model; only local simulation initialization of the corresponding module is required for the new parameters to take effect, ensuring the real-time nature of the correction process.

[0080] In the aforementioned real-time monitoring method for intelligent manufacturing production lines, physical sensor observation data from the physical production line and simulation output data from corresponding virtual sensors deployed in the digital twin model are collected in real time. The two types of data are synchronized in time to form a synchronized physical-virtual sensor data stream. Based on a preset sliding time window, the comprehensive mismatch degree of each physical-virtual sensor pair is calculated and aggregated to obtain a global mismatch index. When the global mismatch index exceeds a preset threshold, significant mismatch regions are identified, and sensitivity analysis is performed on the key physical parameters of the corresponding digital twin model to determine a subset of suspicious parameters causing the mismatch. Then, based on the significant mismatch regions... The difference between physical and virtual data is used to construct a loss function. The optimal parameter correction value is obtained by using an online optimization algorithm with the goal of minimizing the loss function. Based on the optimal parameter correction value, the corresponding suspicious parameter subset in the digital twin model is incrementally updated, thereby fundamentally solving the state mismatch problem between the static model of the digital twin and the dynamic physical entity. This enables the digital twin model to adapt to the physical production line, establishes a closed-loop automatic mismatch correction mechanism, improves the credibility of real-time monitoring, prediction and decision-making of the production line based on digital twin, and ensures that the digital twin model maintains a high-fidelity virtual mapping capability throughout the entire life cycle of the production line.

[0081] In one embodiment, based on a preset sliding time window, the comprehensive mismatch between the observation data and simulation output data of each physical sensor-virtual sensor pair within the physical-virtual sensor synchronous data stream is calculated, including:

[0082] S11. Based on a preset sliding time window, extract the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair from the physical-virtual sensor synchronization data stream.

[0083] Indicatively, under the effect of a preset sliding time window, for each pair of physical-virtual sensors in the physical-virtual sensor synchronous data stream, the corresponding observation data sequence and simulation output data sequence are accurately extracted, ensuring that the two sequences are completely one-to-one in the time dimension, and each time node contains a pair of observation values ​​and simulation values.

[0084] S12. Calculate the mean, standard deviation, skewness, and kurtosis of the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair to obtain the statistical characteristics of each physical sensor-virtual sensor pair.

[0085] Furthermore, for each pair of physical sensors and virtual sensors, statistical characteristics are calculated for their observed data sequences and simulated output data sequences, specifically obtaining the mean, standard deviation, skewness, and kurtosis of the observed data sequences, as well as the mean, standard deviation, skewness, and kurtosis of the simulated output data sequences, forming a unique set of statistical characteristics for each pair of sensors.

[0086] S13. Based on statistical characteristics, the static feature mismatch degree of each physical sensor-virtual sensor pair is calculated using weighted Euclidean distance; the expression for the static feature mismatch degree is: ,in, , , and These are the mean, standard deviation, skewness, and kurtosis of the observed data series, respectively. , , and These represent the mean, standard deviation, skewness, and kurtosis of the simulation output data sequence, respectively. , , and These are weighting coefficients preset according to the sensor type.

[0087] S14. Perform short-time Fourier transform on the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair to obtain the physical time spectrum and virtual time frequency of each physical sensor-virtual sensor pair.

[0088] In this process, short-time Fourier transform is performed on the observation data sequence and simulation output data sequence of each pair of physical sensors and virtual sensors. This transform converts the time-domain data into time-frequency domain data, thereby obtaining the physical time spectrum and virtual time spectrum corresponding to each pair of sensors, and realizing the time-frequency domain characterization of the dynamic change characteristics of the data.

[0089] S15. For each time frame, normalize the physical time spectrum and the virtual time spectrum in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution of each physical sensor-virtual sensor pair.

[0090] Optionally, for the physical time spectrum and virtual time spectrum obtained by the short-time Fourier transform, the frequency dimension spectrum data is normalized in each time frame so that the sum of the frequency dimension spectrum energy in each time frame is 1, thereby obtaining the physical sequence probability distribution and virtual sequence probability distribution of each pair of sensors in each time frame.

[0091] S16. Based on the probability distribution of the physical sequence and the probability distribution of the virtual sequence, calculate the KL divergence on each time frame, and calculate the average value of the KL divergence of all time frames within the entire sliding time window to obtain the dynamic characteristic mismatch of each physical sensor-virtual sensor pair.

[0092] Specifically, based on the physical sequence probability distribution and the virtual sequence probability distribution at each time frame, the KL divergence corresponding to that time frame is calculated. The KL divergence is used to quantify the degree of difference between the two probability distributions. After calculating the KL divergence for all time frames within the sliding time window, the KL divergence values ​​for all time frames are arithmetically averaged to obtain the dynamic characteristic mismatch of each pair of physical and virtual sensors. This indicator can comprehensively reflect the dynamic distribution differences of the data sequence throughout the entire window.

[0093] S17. After normalizing the static feature mismatch and dynamic characteristic mismatch of each physical sensor-virtual sensor pair, the normalized static feature mismatch and dynamic characteristic mismatch are weighted and fused to obtain the comprehensive mismatch of each physical sensor-virtual sensor pair.

[0094] Specifically, the static feature mismatch and dynamic characteristic mismatch of each pair of physical sensors and virtual sensors are normalized. The normalization process uses a preset threshold as a benchmark to convert the mismatch values ​​into standardized values ​​within the 0-1 range, eliminating differences in the dimensions and numerical ranges of different types of mismatch. Based on the relative importance of static and dynamic characteristics in the mismatch assessment, the normalized static feature mismatch and dynamic characteristic mismatch are weighted and fused to obtain the comprehensive mismatch of each pair of physical sensors and virtual sensors. This indicator takes into account both static statistical differences and dynamic distribution differences in the data, and can comprehensively and accurately quantify the degree of mismatch between physical observation and virtual simulation.

[0095] In one embodiment, for each time frame, the physical time spectrum and the virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution for each physical sensor-virtual sensor pair, including:

[0096] S21. For each time frame, sum the physical energy values ​​of the physical time spectrum at each frequency to obtain the total physical energy. Based on the physical energy values ​​and the total physical energy, calculate the physical probability values ​​of the observed data sequence at each frequency to obtain the physical sequence probability distribution. The expression for the physical probability value is: ,in, For time frames, For frequency, For time frames and frequency The physical energy value at that location For time frames The total physical energy.

[0097] Indicative, for the first For physical sensor-virtual sensor, focus on the current time frame. Extract the physical energy values ​​corresponding to all frequency points in the physical time spectrum of that time frame. ,in This is a frequency variable, covering the effective frequency range of the signal monitored by the physical sensor. The physical energy values ​​at all frequency points within this time frame are accumulated to obtain the time frame value. Corresponding total physical energy The total physical energy characterizes the total energy intensity of the observed data sequence within that time frame. At each frequency... Physical energy value at the location For molecules, the total physical energy in the same time frame Using the denominator, the observed data sequence at this frequency is calculated. Physical probability value at location ,in, For the first For the sensor in time frames ,frequency The physical probability value at that point reflects the frequency of the observed data sequence within that time frame. The corresponding energy percentage; For the first For the sensor in time frames ,frequency The physical energy value at a given point is directly obtained from the physical time spectrum. For the first For the sensor in time frames The total physical energy is the sum of the physical energy values ​​at all frequency points within that time frame. (This refers to the sum of the physical energy values ​​at all frequency points within that time frame.) The physical probability values ​​corresponding to all frequency points are integrated to form the physical sequence probability distribution of the observed data sequence within the time frame. This distribution satisfies the probability normalization property that the sum of the physical probability values ​​of all frequency points is 1.

[0098] S22. For the same time frame, sum the virtual energy values ​​of the virtual time spectrum at each frequency to obtain the virtual total energy. Based on the virtual energy values ​​and the virtual total energy, calculate the virtual probability values ​​of the simulation output data sequence at each frequency to obtain the virtual sequence probability distribution; the expression for the virtual probability value is: ,in, For time frames, For frequency, For time frames and frequency The virtual energy value at that location, For time frames The virtual total energy.

[0099] Specifically, for the first Extract the time frame for the physical sensor-virtual sensor. Virtual energy values ​​corresponding to all frequency points in the virtual time spectrum ,frequency The value range is consistent with the frequency range of the physical time spectrum to ensure that the frequency dimensions of the two are comparable. The virtual energy values ​​at all frequency points within this time frame are accumulated to obtain the value for that time frame. Corresponding virtual total energy The calculation logic for virtual total energy is completely consistent with that for physical total energy, representing the total energy intensity of the simulation output data sequence within that time frame. The virtual energy value at each frequency f is... For molecules, the virtual total energy in the same time frame Using the denominator, the simulation output data sequence at this frequency is calculated. Virtual probability value at the location ,in, For the first For the sensor in time frames ,frequency The virtual probability value at that point reflects the frequency of the simulated output data sequence within that time frame. The corresponding energy percentage; For the first For the sensor in time frames ,frequency The virtual energy value at that location is directly obtained from the virtual time spectrum; For the first For the sensor in time frames The virtual total energy is the sum of the virtual energy values ​​at all frequency points within that time frame. (This refers to the sum of virtual energy values ​​at all frequency points within that time frame.) The virtual probability values ​​corresponding to all frequency points are integrated to form a virtual sequence probability distribution of the simulated output data sequence within that time frame.

[0100] In one embodiment, sensitivity analysis is performed on key physical parameters corresponding to significant mismatch regions in the digital twin model to obtain a subset of suspected parameters leading to the mismatch, including:

[0101] S31. Extract multiple key physical parameters related to the dynamic behavior of the significantly mismatched region from the digital twin model to form an initial parameter set; the current value of each key physical parameter in the initial parameter set is the nominal value.

[0102] This schematically focuses on the dynamic behavior characteristics of regions with significant mismatches. Multiple key physical parameters directly related to these regions are systematically extracted from the digital twin model. These parameters directly influence the kinematic properties, dynamic response, or process interactions of the region, thus forming an initial parameter set. The current value of each key physical parameter in the initial parameter set is its nominal value, reflecting the parameter baseline of the physical entity under its current operating state.

[0103] S32. Select multiple virtual sensors related to the significant mismatch region from the digital twin model, and calculate the partial derivatives of each key physical parameter at the nominal value for the simulation output data of each virtual sensor based on the initial parameter set, to obtain the parameter sensitivity.

[0104] Furthermore, based on the spatial extent and monitoring requirements of the significant mismatch area, multiple virtual sensors directly related to this area are precisely selected from the digital twin model. The selected virtual sensors must be able to comprehensively capture the simulation output data of the significant mismatch area, ensuring that the impact of parameter changes on the simulation results can be effectively perceived. Based on the constructed initial parameter set, using the nominal values ​​of each key physical parameter as the calculation benchmark, the partial derivatives of the simulation output data of each virtual sensor with respect to each key physical parameter are calculated. These partial derivatives represent the parameter sensitivity, and their magnitude directly reflects the degree of influence of minute changes in the corresponding key physical parameter on the simulation output data of the virtual sensor.

[0105] S33. Gather the parameter sensitivities of all virtual sensors related to the significant mismatch region and generate a sensitivity matrix.

[0106] Optionally, the parameter sensitivities of all virtual sensors related to the significant mismatch region for each key physical parameter are systematically collected, and a sensitivity matrix is ​​constructed according to preset rules. The row dimension of this matrix corresponds to the number of selected virtual sensors, and the column dimension corresponds to the number of key physical parameters in the initial parameter set. Each element in the matrix represents the parameter sensitivity of the simulation output data of a specific virtual sensor to a specific key physical parameter, comprehensively characterizing the correlation strength between each parameter and the output of each sensor.

[0107] S34. Calculate the covariance matrix based on the sensitivity matrix, and perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues ​​and eigenvectors of each principal component.

[0108] Furthermore, principal component analysis is performed based on the constructed sensitivity matrix. This involves calculating the covariance matrix using statistical methods. The elements of the covariance matrix describe the linear correlation between the sensitivities of different parameters, effectively reflecting the correlation between the influence of each key physical parameter on the virtual sensor output. Eigenvalue decomposition is then performed on the calculated covariance matrix, transforming it into a diagonal matrix form. This yields eigenvalues ​​and eigenvectors corresponding to multiple principal components. The eigenvalues ​​quantify the variance contribution of the corresponding principal component, while the eigenvectors characterize the linear combination relationship between the principal components and the original key physical parameters.

[0109] S35. Sort each principal component in descending order of its eigenvalues, and calculate the cumulative contribution rate of each principal component based on the sorted eigenvalues.

[0110] All principal components are sorted in descending order of their corresponding eigenvalues. Principal components with larger eigenvalues ​​contain richer information about the original parameters and have a stronger explanatory power for mismatch phenomena. Based on the sorted eigenvalues, the contribution rate of each principal component is calculated, which is the ratio of a single eigenvalue to the sum of all eigenvalues. The cumulative contribution rate of each principal component is obtained by summing these values. The cumulative contribution rate reflects the total amount of original parameter information contained in the first k principal components.

[0111] S36. For principal components whose cumulative contribution rate exceeds a preset threshold, identify the original physical parameters whose absolute load values ​​on each principal component exceed a preset load threshold, and construct a subset of suspicious parameters based on the identified original physical parameters.

[0112] Specifically, preset cumulative contribution rate thresholds and loading thresholds are used. The cumulative contribution rate threshold is used to filter principal components that can cover sufficient original information, while the loading threshold is used to identify original physical parameters that significantly contribute to the principal components. For principal components whose cumulative contribution rate exceeds the preset threshold, the corresponding feature vector is extracted. The elements in the feature vector are the loadings of each original physical parameter on that principal component. Original physical parameters whose absolute loading values ​​exceed the preset loading threshold are identified as the core parameters that contribute most significantly to the principal components. All identified original physical parameters of this type are integrated to construct a subset of suspicious parameters. This subset can accurately focus on the core parameters that may lead to mismatch, providing a clear target for subsequent parameter correction.

[0113] In one embodiment, a loss function for a subset of suspicious parameters is constructed based on the differences between the observed data and simulation output data of each physical sensor-virtual sensor pair within a significant mismatch region, including:

[0114] The loss function is obtained using the following formula:

[0115]

[0116]

[0117] in, The loss function; A subset of suspicious parameters; For physical sensor-virtual sensor pair index set; For the first Weighting coefficients for each physical sensor-virtual sensor pair; For the first The mean square error term of a physical sensor-virtual sensor pair; Let be the dynamic characteristic mismatch degree of the i-th physical sensor-virtual sensor pair; The length of the sliding time window; For observation data sequences; This is a sequence of simulation output data obtained by simulating key physical parameters of a subset of suspicious parameters.

[0118] In one embodiment, based on the optimal parameter correction value, the model parameters corresponding to the suspicious parameter subset in the digital twin model are incrementally updated to obtain the corrected digital twin model, including:

[0119] S41. Obtain the current parameter vector based on the nominal values ​​of each key physical parameter within the subset of suspicious parameters.

[0120] This schematically identifies all key physical parameters included in the subset of suspected parameters, with the nominal value of each key physical parameter in the digital twin model representing its value at the current moment. Optionally, according to a preset parameter sorting rule, such as the logical order of parameters in the model's dynamic equations, the order of parameter identification numbers, or the priority order of their impact on the simulation output, the nominal values ​​of each key physical parameter in the subset of suspected parameters are arranged in an ordered manner to construct the current parameter vector. Current parameter vector The dimension of the vector is exactly the same as the number of parameters in the suspicious parameter subset. Each element in the vector uniquely corresponds to a key physical parameter in the suspicious parameter subset, which intuitively and accurately represents the current value state of the parameter in the digital twin model.

[0121] S42. Based on the optimal parameter correction value, incrementally update the current parameter vector to obtain the new parameter vector; the incremental update expression is: ,in, The learning rate is preset, and , For the current parameter vector, This is the optimal parameter correction value. This is the new parameter vector.

[0122] For example, This is the current parameter vector, where each element corresponds to the current nominal value of a key physical parameter within the subset of suspected parameters. The optimal parameter correction value is the parameter combination that minimizes the loss function. Each element corresponds to the target optimization value of the key physical parameters within the subset of suspected parameters, which enables the simulation output of the digital twin model to achieve optimal matching with the physical observation data. The preset learning rate has a strictly limited range of values. To control the step size of each parameter update, the specific value of the learning rate needs to be pre-calibrated based on the parameter type, the simulation sensitivity of the digital twin model, and the stability requirements of the production line operation. For example, for key parameters that have a significant impact on the simulation results, the learning rate should be relatively small to ensure smooth updates. For parameters with a low degree of impact, the learning rate can be appropriately increased to improve the correction efficiency. The new parameter vector is obtained after incremental update. Each element of the vector is the final value of the key physical parameters in the subset of suspicious parameters after smooth adjustment. This vector not only fully incorporates the optimization goal of the optimal parameter correction value, but also retains the reasonable basis of the current parameter vector, realizing a smooth transition of the parameters from the current state to the optimal state.

[0123] S43. Update the parameter values ​​of the suspected parameter subset in the digital twin model based on the new parameter vector to obtain the corrected digital twin model.

[0124] Furthermore, a parameter association mapping is established between the new parameter vector and the suspicious parameter subset in the digital twin model. The mapping is based on the unique identifier information of the parameters, ensuring that each element in the new parameter vector accurately corresponds to the key physical parameter in the model, avoiding parameter matching errors. Through the parameter update interface reserved in the digital twin model, the current values ​​of each key physical parameter belonging to the suspicious parameter subset in the model are replaced one by one with the corresponding element values ​​in the new parameter vector. The update process adopts a modular approach, modifying only the parameters within the suspicious parameter subset, while other parameters in the model unrelated to mismatch correction remain unchanged, preventing unnecessary interference to the overall model structure and the simulation logic of non-mismatch areas. Optionally, after the parameter replacement is completed, a partial simulation initialization is performed on the digital twin model. There is no need to restart the entire model system; only the simulation modules related to the suspicious parameter subset need to be activated, making the new parameters effective immediately. Finally, a corrected digital twin model is generated. This corrected model can accurately match the actual operating state of the physical production line, effectively eliminating the state mismatch between the digital twin and the physical entity, providing high-fidelity virtual mapping support for real-time monitoring, accurate diagnosis, and optimization decisions of the production line.

[0125] In one embodiment, the method further includes:

[0126] S51. Perform a full-cycle simulation calculation based on the corrected digital twin model to generate updated simulation output data of the virtual sensor.

[0127] Indicatively, based on the calibrated digital twin model with completed parameter incremental updates, a complete process cycle corresponding to the actual operation of the physical production line is determined as the simulation cycle. This cycle needs to cover the entire process of the production line from workpiece loading, processing of each step, assembly to unloading, ensuring that the simulation process can completely reproduce a complete operation of the physical production line and avoid distortion of verification results due to incomplete cycle truncation. When the simulation calculation starts, the digital twin model loads the new parameter vector. All corresponding parameter values ​​are set, and the simulation engine operates according to the same timing logic and process constraints as the physical production line. The simulation step size strictly matches the sampling period of the physical sensors, ensuring that the output data of the virtual sensors have a one-to-one correspondence with the physical observation data in the time dimension. During the simulation, each virtual sensor in the model continuously collects various physical quantity data, including equipment operating parameters, workpiece state parameters, and process parameters, according to the preset measurement logic and data output frequency. The output data of all virtual sensors are stored in a unified data format and timestamp specification, and finally, an updated virtual sensor simulation output dataset corresponding to the complete cycle of the physical production line is generated. This dataset must contain the specific output values ​​of each virtual sensor at each time point within the simulation cycle, providing core data support for subsequent verification of calibration effects.

[0128] S52. Based on the updated simulation output data of the virtual sensor, combined with the observation data of the physical sensor collected in the corresponding time series, calculate the global mismatch index to obtain the corrected verification index.

[0129] Based on the updated virtual sensor simulation output data and the corresponding time-series physical sensor observation data, a global mismatch index, i.e., the post-correction verification index, is calculated to verify the correction effect. Specifically, time synchronization verification is performed on the two types of data to ensure that the timestamps of the simulation output data and the physical observation data are completely aligned, eliminating time-series misaligned data caused by data transmission delays or simulation errors, and ensuring the accuracy of comparative analysis. For each pair of physical sensor-virtual sensor data, based on the synchronized observation data and the updated simulation output data, the corresponding comprehensive mismatch degree is calculated, and then the global mismatch index is obtained through weighted aggregation. This index is the post-correction verification index.

[0130] S53. Determine whether the corrected verification index is lower than the preset threshold, and calculate the percentage decrease of the corrected verification index relative to the global mismatch index when the correction is triggered. If the corrected verification index is lower than the preset threshold and the percentage decrease is greater than the preset minimum improvement rate, then output the corrected digital twin model.

[0131] Furthermore, the calibration verification index is checked against a preset threshold, which is consistent with the global mismatch index threshold that triggers the calibration of the digital twin model parameters. This threshold is used to determine whether the matching status between the digital twin model and the physical entity meets the acceptable standard. Simultaneously, the percentage decrease in the calibration verification index relative to the global mismatch index at the time of calibration is calculated. The preset minimum improvement rate is determined based on the minimum requirements for calibration effectiveness in actual engineering projects, combined with the production line monitoring accuracy requirements, equipment operation stability requirements, and historical calibration data statistics to ensure that the calibrated model can effectively improve the fidelity of the virtual mapping. When the calibration verification index is lower than the preset threshold and the percentage decrease is greater than the preset minimum improvement rate, it indicates that the parameter calibration has achieved the expected effect, and the calibrated digital twin model can accurately match the actual operating status of the physical production line. At this point, the model is output as a valid model, and the output includes the model's complete parameter configuration file and the calibrated parameter vector. The calibration effect verification report shows that the model will be directly used for real-time monitoring, fault diagnosis, and process optimization of the subsequent production line.

[0132] S54. If the corrected verification index exceeds the preset threshold or the decrease rate is less than the preset minimum improvement rate, then retain the current parameter vector and record the information of this correction failure.

[0133] If the corrected verification index exceeds the preset threshold, or the decrease rate is less than the preset minimum improvement rate, it indicates that the parameter correction has not achieved the expected effect. In this case, the current parameter vector is retained. All relevant data from this calibration process will be recorded without replacing the original digital twin model's operating parameters to avoid performance degradation due to invalid calibration. Detailed information about calibration failures will be recorded, including the calibration trigger time, global mismatch index values ​​before and after calibration, the rate of decrease, specific parameters of the suspicious parameter subset, optimal parameter correction values, learning rate values, and key data segments from the complete simulation cycle. This information will be categorized and stored in the form of system logs. This failure information will provide important reference for restarting the calibration process, allowing technical personnel to analyze the possible causes of calibration failures based on the recorded information.

[0134] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0135] Based on the same inventive concept, this application also provides a real-time monitoring device for an intelligent manufacturing production line to implement the aforementioned real-time monitoring method. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations in one or more embodiments of the real-time monitoring device for intelligent manufacturing production lines provided below can be found in the limitations of the real-time monitoring method for intelligent manufacturing production lines described above, and will not be repeated here.

[0136] In one exemplary embodiment, such as Figure 2 As shown, a real-time monitoring device for an intelligent manufacturing production line is provided, comprising:

[0137] Data module 201 is used to collect observation data from multiple physical sensors on the physical production line and simulation output data from multiple virtual sensors deployed in the digital twin model corresponding to the physical sensors in real time, and to perform time synchronization and alignment processing on the observation data and simulation output data to obtain a physical-virtual sensor synchronized data stream.

[0138] The mismatch detection module 202 is used to calculate the comprehensive mismatch degree between the observation data and simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream based on a preset sliding time window, and to aggregate the comprehensive mismatch degree of all physical sensor-virtual sensor pairs to obtain the global mismatch index.

[0139] The suspicious parameter module 203 is used to identify significant mismatch regions based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair when the global mismatch index exceeds a preset threshold, and to perform sensitivity analysis on the key physical parameters of the corresponding significant mismatch regions in the digital twin model to obtain a subset of suspicious parameters that lead to mismatch.

[0140] The correction module 204 is used to construct a loss function for a subset of suspicious parameters based on the difference between the observation data and simulation output data of each physical sensor-virtual sensor pair in the mismatched area, and to solve the subset of suspicious parameters by means of minimizing the loss function, so as to obtain the optimal parameter correction value.

[0141] The model update module 205 is used to incrementally update the model parameters of the corresponding suspicious parameter subset in the digital twin model based on the optimal parameter correction value, so as to obtain the corrected digital twin model.

[0142] In one embodiment, the mismatch detection module 202 is further configured to:

[0143] Based on a preset sliding time window, the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair are extracted from the physical-virtual sensor synchronous data stream;

[0144] The mean, standard deviation, skewness, and kurtosis of the observed data sequence and the simulated output data sequence of each physical sensor-virtual sensor pair are calculated respectively to obtain the statistical characteristics of each physical sensor-virtual sensor pair.

[0145] Based on statistical characteristics, the static feature mismatch degree of each physical sensor-virtual sensor pair is calculated using weighted Euclidean distance; the expression for the static feature mismatch degree is: ,in, , , and These are the mean, standard deviation, skewness, and kurtosis of the observed data series, respectively. , , and These represent the mean, standard deviation, skewness, and kurtosis of the simulation output data sequence, respectively. , , and These are weighting coefficients preset according to the sensor type;

[0146] Short-time Fourier transforms are performed on the observation data sequences and simulation output data sequences of each physical sensor-virtual sensor pair to obtain the physical time spectrum and virtual time frequency of each physical sensor-virtual sensor pair.

[0147] For each time frame, the physical time spectrum and the virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution of each physical sensor-virtual sensor pair.

[0148] Based on the probability distribution of the physical sequence and the probability distribution of the virtual sequence, calculate the KL divergence on each time frame, and calculate the average value of the KL divergence of all time frames within the entire sliding time window to obtain the dynamic characteristic mismatch of each physical sensor-virtual sensor pair.

[0149] After normalizing the static feature mismatch and dynamic characteristic mismatch of each physical sensor-virtual sensor pair, the normalized static feature mismatch and dynamic characteristic mismatch are weighted and fused to obtain the comprehensive mismatch of each physical sensor-virtual sensor pair.

[0150] In one embodiment, a probability distribution module is also included, for:

[0151] For each time frame, the physical energy values ​​of the physical time spectrum at each frequency are summed to obtain the total physical energy. Based on the physical energy values ​​and the total physical energy, the physical probability values ​​of the observed data sequence at each frequency are calculated to obtain the physical sequence probability distribution. The expression for the physical probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The physical energy value at that location For time frames The total physical energy;

[0152] For the same time frame, the virtual energy values ​​of the virtual time spectrum at each frequency are summed to obtain the virtual total energy. Based on the virtual energy values ​​and the virtual total energy, the virtual probability values ​​of the simulation output data sequence at each frequency are calculated to obtain the virtual sequence probability distribution. The expression for the virtual probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The virtual energy value at that location, For time frames The virtual total energy.

[0153] In one embodiment, the suspicious parameter module 203 is further configured to:

[0154] Multiple key physical parameters associated with the dynamic behavior of the significantly mismatched region are extracted from the digital twin model to form an initial parameter set; the current values ​​of each key physical parameter in the initial parameter set are nominal values.

[0155] Multiple virtual sensors related to the significant mismatch region are selected from the digital twin model, and based on the initial parameter set, the partial derivatives of each key physical parameter at the nominal value are calculated on the simulation output data of each virtual sensor to obtain the parameter sensitivity.

[0156] The parameter sensitivities of all virtual sensors associated with significant mismatch regions are aggregated to generate a sensitivity matrix;

[0157] The covariance matrix is ​​calculated based on the sensitivity matrix, and eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and eigenvectors of each principal component.

[0158] The principal components are sorted from largest to smallest according to their eigenvalues, and the cumulative contribution rate of each principal component is calculated based on the sorted eigenvalues.

[0159] For principal components whose cumulative contribution rate exceeds a preset threshold, the original physical parameters whose absolute load values ​​on each principal component exceed a preset load threshold are identified, and a subset of suspicious parameters is constructed based on the identified original physical parameters.

[0160] In one embodiment, the model update module 205 is further configured to:

[0161] The current parameter vector is obtained based on the nominal values ​​of each key physical parameter within the subset of suspicious parameters;

[0162] The current parameter vector is incrementally updated based on the optimal parameter correction value to obtain a new parameter vector; the incremental update expression is: ,in, The learning rate is preset, and , For the current parameter vector, This is the optimal parameter correction value. For the new parameter vector;

[0163] The corrected digital twin model is obtained by updating the parameter values ​​of the suspicious parameter subset in the digital twin model based on the new parameter vector.

[0164] In one embodiment, a correction feedback module is also included, for:

[0165] A complete simulation cycle is performed based on the corrected digital twin model to generate updated simulation output data from the virtual sensor.

[0166] Based on the updated simulation output data of the virtual sensor, combined with the observation data of the physical sensor collected in the corresponding time series, the global mismatch index is calculated to obtain the corrected verification index.

[0167] Determine whether the corrected verification index is lower than the preset threshold, and calculate the percentage decrease of the corrected verification index relative to the global mismatch index when correction is triggered. If the corrected verification index is lower than the preset threshold and the percentage decrease is greater than the preset minimum improvement rate, then output the corrected digital twin model.

[0168] If the corrected verification index exceeds the preset threshold or the decrease rate is less than the preset minimum improvement rate, the current parameter vector is retained and the information of this correction failure is recorded.

[0169] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0170] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for real-time monitoring of an intelligent manufacturing production line, characterized in that, The method includes: The observation data from multiple physical sensors on the physical production line and the simulation output data from multiple virtual sensors deployed in the digital twin model corresponding to the physical sensors are collected in real time. The observation data and the simulation output data are then time-synchronized and aligned to obtain a physical-virtual sensor synchronized data stream. Based on a preset sliding time window, the comprehensive mismatch degree between the observation data and the simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream is calculated, and the comprehensive mismatch degree of all physical sensor-virtual sensor pairs is aggregated to obtain the global mismatch index. When the global mismatch index exceeds a preset threshold, a significant mismatch region is identified based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair, and a sensitivity analysis is performed on the key physical parameters in the digital twin model corresponding to the significant mismatch region to obtain a subset of suspicious parameters that lead to mismatch. Based on the differences between the observed data and the simulation output data of each physical sensor-virtual sensor pair within the significant mismatch region, a loss function for the suspicious parameter subset is constructed. With the goal of minimizing the loss function, an online optimization algorithm is used to solve for the suspicious parameter subset to obtain the optimal parameter correction value. Based on the optimal parameter correction value, the model parameters corresponding to the suspicious parameter subset in the digital twin model are incrementally updated to obtain the corrected digital twin model.

2. The method according to claim 1, characterized in that, The calculation of the overall mismatch between the observed data and the simulation output data of each physical sensor-virtual sensor pair within the physical-virtual sensor synchronous data stream, based on a preset sliding time window, includes: Based on the preset sliding time window, the observation data sequence and simulation output data sequence of each physical sensor-virtual sensor pair are extracted from the physical-virtual sensor synchronization data stream; The mean, standard deviation, skewness, and kurtosis of the observed data sequence and the simulated output data sequence of each physical sensor-virtual sensor pair are calculated respectively to obtain the statistical characteristics of each physical sensor-virtual sensor pair. Based on the aforementioned statistical characteristics, the static feature mismatch degree of each physical sensor-virtual sensor pair is calculated using weighted Euclidean distance; the expression for the static feature mismatch degree is as follows: ,in, , , and These are the mean, standard deviation, skewness, and kurtosis of the observed data sequence, respectively. , , and These are the mean, standard deviation, skewness, and kurtosis of the simulation output data sequence, respectively. , , and These are weighting coefficients preset according to the sensor type; Short-time Fourier transforms are performed on the observation data sequence and the simulation output data sequence of each physical sensor-virtual sensor pair to obtain the physical time spectrum and virtual time frequency of each physical sensor-virtual sensor pair. For each time frame, the physical time spectrum and the virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and the virtual sequence probability distribution of each physical sensor-virtual sensor pair. Based on the probability distribution of the physical sequence and the probability distribution of the virtual sequence, calculate the KL divergence on each time frame, and calculate the average value of the KL divergence of all time frames within the entire sliding time window to obtain the dynamic characteristic mismatch of each physical sensor-virtual sensor pair. After normalizing the static feature mismatch and dynamic feature mismatch of each physical sensor-virtual sensor pair, the normalized static feature mismatch and dynamic feature mismatch are weighted and fused to obtain the comprehensive mismatch of each physical sensor-virtual sensor pair.

3. The method according to claim 2, characterized in that, For each time frame, the physical time spectrum and virtual time spectrum are normalized in the frequency dimension to obtain the physical sequence probability distribution and virtual sequence probability distribution for each physical sensor-virtual sensor pair, including: For each time frame, the physical energy values ​​of the physical time spectrum at each frequency are summed to obtain the total physical energy. Based on the physical energy values ​​and the total physical energy, the physical probability values ​​of the observed data sequence at each frequency are calculated to obtain the probability distribution of the physical sequence. The expression for the physical probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The physical energy value at that location For time frames The total physical energy; For the same time frame, the virtual energy values ​​of the virtual time spectrum at each frequency are summed to obtain the virtual total energy. Based on the virtual energy values ​​and the virtual total energy, the virtual probability values ​​of the simulated output data sequence at each frequency are calculated to obtain the virtual sequence probability distribution. The expression for the virtual probability value is as follows: ,in, For time frames, For frequency, For time frames and frequency The virtual energy value at that location, For time frames The virtual total energy.

4. The method according to claim 1, characterized in that, The sensitivity analysis of key physical parameters corresponding to the significant mismatch region in the digital twin model yields a subset of suspected parameters leading to the mismatch, including: Multiple key physical parameters associated with the dynamic behavior of the significant mismatch region are extracted from the digital twin model to form an initial parameter set; the current value of each key physical parameter in the initial parameter set is the nominal value. Multiple virtual sensors related to the significant mismatch region are selected from the digital twin model, and based on the initial parameter set, the partial derivatives of each key physical parameter at the nominal value are calculated on the simulation output data of each virtual sensor to obtain the parameter sensitivity. The parameter sensitivities of all virtual sensors associated with the significant mismatch region are aggregated to generate a sensitivity matrix; The covariance matrix is ​​calculated based on the sensitivity matrix, and eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and eigenvectors of each principal component. The principal components are sorted from largest to smallest according to their eigenvalues, and the cumulative contribution rate of each principal component is calculated based on the sorted eigenvalues. For principal components whose cumulative contribution rate exceeds a preset threshold, the original physical parameters whose absolute load values ​​on each principal component exceed a preset load threshold are identified, and the suspicious parameter subset is constructed based on the identified original physical parameters.

5. The method according to claim 1, characterized in that, The loss function for constructing the subset of suspicious parameters based on the difference between the observed data and the simulation output data for each physical sensor-virtual sensor pair within the significant mismatch region includes: The loss function is obtained using the following formula: in, The loss function is... A subset of suspicious parameters; For physical sensor-virtual sensor pair index set; For the first Weighting coefficients for each physical sensor-virtual sensor pair; For the first The mean square error term of a physical sensor-virtual sensor pair; Let be the dynamic characteristic mismatch degree of the i-th physical sensor-virtual sensor pair; The length of the sliding time window; For observation data sequences; This is a sequence of simulation output data obtained by simulating key physical parameters of a subset of suspicious parameters.

6. The method according to claim 4, characterized in that, The step of incrementally updating the model parameters corresponding to the suspicious parameter subset in the digital twin model based on the optimal parameter correction value to obtain the corrected digital twin model includes: Based on the nominal values ​​of each key physical parameter within the subset of suspected parameters, the current parameter vector is obtained; Based on the optimal parameter correction value, the current parameter vector is incrementally updated to obtain a new parameter vector; the incremental update expression is: ,in, The learning rate is preset, and , For the current parameter vector, This is the optimal parameter correction value. For the new parameter vector; The parameter values ​​belonging to the suspicious parameter subset in the digital twin model are updated based on the new parameter vector to obtain the corrected digital twin model.

7. The method according to claim 6, characterized in that, The method further includes: A complete simulation cycle is performed based on the corrected digital twin model to generate updated simulation output data from the virtual sensor. Based on the updated simulation output data of the virtual sensor, combined with the observation data of the physical sensor collected in the corresponding time series, the global mismatch index is calculated to obtain the corrected verification index. Determine whether the corrected verification index is lower than the preset threshold, and calculate the decrease ratio of the corrected verification index relative to the global mismatch index when the correction is triggered. If the corrected verification index is lower than the preset threshold and the decrease ratio is greater than the preset minimum improvement rate, then output the corrected digital twin model. If the corrected verification index exceeds the preset threshold or the decrease ratio is less than the preset minimum improvement rate, then the current parameter vector is retained and the correction failure information is recorded.

8. A real-time monitoring device for an intelligent manufacturing production line, characterized in that, The device includes: The data module is used to collect observation data from multiple physical sensors on the physical production line and simulation output data from multiple virtual sensors deployed in the digital twin model corresponding to the physical sensors in real time, and to perform time synchronization and alignment processing on the observation data and the simulation output data to obtain a physical-virtual sensor synchronized data stream. The mismatch detection module is used to calculate the comprehensive mismatch degree between the observation data and the simulation output data of each physical sensor-virtual sensor pair in the physical-virtual sensor synchronous data stream based on a preset sliding time window, and to aggregate the comprehensive mismatch degree of all physical sensor-virtual sensor pairs to obtain a global mismatch index. The suspicious parameter module is used to identify significant mismatch regions based on the comprehensive mismatch degree of each physical sensor-virtual sensor pair when the global mismatch index exceeds a preset threshold, and to perform sensitivity analysis on the key physical parameters in the digital twin model corresponding to the significant mismatch regions to obtain a subset of suspicious parameters that lead to mismatch. The correction module is used to construct a loss function for the suspicious parameter subset based on the difference between the observed data and the simulation output data of each physical sensor-virtual sensor pair in the mismatched significant region, and to solve the suspicious parameter subset using an online optimization algorithm with the goal of minimizing the loss function, so as to obtain the optimal parameter correction value. The model update module is used to incrementally update the model parameters corresponding to the suspicious parameter subset in the digital twin model based on the optimal parameter correction value, so as to obtain the corrected digital twin model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of 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 executed by a processor, it implements the method of any one of claims 1 to 7.

Images