A method for synchronizing digital twin data between the real and virtual worlds in discrete manufacturing

Abstract

This invention provides a method for synchronizing digital twin data with real-world data in discrete manufacturing, belonging to the field of intelligent manufacturing and industrial information technology. It involves unified preprocessing of multi-source heterogeneous sensor data and using discrete manufacturing equipment data as a time reference to dynamically align the twin model data in time. Based on this, an adaptive correction mechanism based on historical synchronization errors is established to compensate for sensor drift, noise interference, and model prediction errors in real time. Simultaneously, through high- and low-frequency data rate matching and interpolation compensation, high-frequency acquired data and low-frequency simulation models can operate collaboratively on a unified time scale. Through these technical means, the digital twin model can continuously and stably reflect the real operating status of discrete manufacturing equipment, providing a reliable data foundation for production monitoring, status analysis, and intelligent decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial information technology, specifically to a method for synchronizing digital twin data between the physical and digital worlds in discrete manufacturing. Background Technology

[0002] Discrete manufacturing is an important mode of modern industrial production. Its production process typically consists of multiple independent machines, multiple processing and assembly stations, and a complex logistics system. The operating cycle time, data sampling methods, and control logic of each machine are significantly different. With the development of the Internet of Things (IoT), the Industrial Internet, and digital twin technologies, the manufacturing site can collect equipment operating status, workpiece position, process parameters, and quality inspection data in real time. A digital twin model corresponding to the physical workshop can then be built in virtual space to achieve production monitoring, status prediction, and decision support.

[0003] In discrete manufacturing scenarios, digital twin systems need to continuously receive data from multiple sensors and control systems to keep the virtual model synchronized with the discrete manufacturing equipment. However, due to inconsistent sampling frequencies of different devices, network transmission delays, and the computational cycle limitations of the virtual model itself, physical data and virtual model data often cannot be strictly aligned on the timeline, easily leading to a desynchronization between the virtual and physical states. When the virtual model cannot accurately reflect the real-time status of the discrete manufacturing equipment, it directly affects the reliability of production monitoring, fault warning, and scheduling decisions.

[0004] Furthermore, in real-world industrial environments, sensor signals are susceptible to noise interference, long-term drift, and momentary jitter, leading to deviations in the data input to the digital twin model. Simultaneously, a significant time granularity difference exists between high-frequency sensor data and low-frequency model parameter updates. Without effective rate coordination and data compensation, the virtual model's response will lag behind the real device, further amplifying the virtual-real discrepancy. Existing technologies typically employ fixed-time alignment or simple filtering for synchronization, which struggles to simultaneously address time misalignment, data drift, and multi-frequency rate mismatches in complex discrete manufacturing environments, thus failing to meet the application requirements for high-precision digital twin synchronization. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for synchronizing virtual and real data in digital twins for discrete manufacturing. By constructing a collaborative mechanism in the digital twin system for time alignment, dynamic error correction, and multi-frequency rate matching, the data of discrete manufacturing equipment and the twin model data are kept consistent in terms of time dimension, feature dimension, and update rate. This solves the problem of virtual-real discrepancy caused by inconsistent sampling frequencies, transmission delays, data drift, and asynchronous model updates in the prior art.

[0006] To achieve the above technical objectives, the adopted technical solution is: a method for synchronizing digital twin data between the real and virtual worlds in discrete manufacturing, comprising the following steps: The first step is geometric modeling: based on CAD drawings and measurement data, a twin model of the discrete manufacturing equipment is constructed, and a mapping relationship between the spatial position and structural parameters of key components is established; The second step is behavioral modeling: Based on the motion form and driving method of the discrete manufacturing equipment, establish corresponding kinematic or dynamic constraints so that the twin model can calculate the pose, velocity and force state of the discrete manufacturing equipment based on the input data. The third step is data modeling: establishing a data interaction channel between discrete manufacturing equipment and the twin model through an industrial communication interface, so that sensor data and twin model state parameters form a two-way mapping relationship; The fourth step is signal processing: filtering is performed on the collected real signal data of discrete manufacturing equipment to obtain the signal data sequence of discrete manufacturing equipment. The fifth step is signal standardization: the discrete manufacturing equipment signal data sequence obtained in the fourth step is standardized to eliminate differences in dimensions and scales; Step 6, Feature Mapping: The discrete manufacturing equipment signal data sequence obtained in Step 5 is converted into input data that can be received by the twin model through the feature mapping matrix; Step 7, High-dimensional feature projection and sequence alignment: Use CCA to map the real signal data directly collected by the discrete manufacturing equipment and the input data obtained in step 6 to a low-dimensional common subspace; use DTW to calculate the optimal matching path on the low-dimensional sequence; Step 8, Temporal Mapping Regularization: Based on the optimal matching path, reconstruct the temporal mapping relationship between discrete manufacturing equipment data and twin model data, resample and re-index the twin model input data, and realize the regularization and alignment of temporal deviations; Step 9, Anomaly Detection: Principal component analysis is performed on the input data of the twin model after time-series normalization to extract principal component information that reflects normal operation characteristics. The reconstruction error of the principal component information is calculated, and the mean square index is used to identify outliers. When the mean square index identifies outliers that are greater than the set feature range, the data is judged as abnormal. Step 10, Abnormal Data Correction Processing: For data detected as abnormal, nonlinear estimation is performed using an artificial neural network trained based on historical normal operating conditions, and corrected data that conforms to normal operating characteristics is output to replace the original abnormal values. Step 11, Frequency Relationship Determination: Obtain the sampling frequency of the discrete manufacturing equipment data and the update frequency of the twin model state, and determine the frequency relationship between the two. Divide the sampling frequency of the discrete manufacturing equipment data into high-frequency data, low-frequency data, or frequency-matching data. Step 12, Data Classification Processing: The data of different frequency types obtained in Step 11 are processed separately to achieve synchronization of data of different frequencies on a unified time scale; Step 13, Synchronization Control and Maintenance: Based on the degree of difference between the overall operating status of the discrete manufacturing equipment and the overall operating status of the twin model, determine the current synchronization status, and perform synchronization correction operations when the deviation exceeds the limit to achieve continuous and stable operation of the virtual and real systems.

[0007] The data partitioning method in step eleven is as follows: Suppose that the data sampling frequency of a certain sensor on the discrete manufacturing equipment side is . The update frequency of the corresponding state variables in the twin model is The frequency ratio is then defined as: when When, it is high-frequency data; when When, it is low-frequency data; when At that time, it is frequency matching data.

[0008] The method for processing data of different frequency types is as follows: high-frequency data is resampled, low-frequency data is interpolated and extrapolated for compensation, and frequency-matching data is directly and synchronously fused.

[0009] The process of calculating the reconstruction error is as follows: in, Principal component eigenvector matrix, , , This is the input matrix for the twin model.

[0010] The formula for calculating the feature range is: Among them, the mean of the reconstruction error of historical normal operation data in the principal component space is The standard deviation is , and k is an empirical coefficient.

[0011] The specific implementation process of step thirteen is as follows: Within a unified time window, key state variables are extracted from the corresponding parameters of the discrete manufacturing equipment and the twin model. This extraction process is based on the influence of each parameter on the consistency of virtual-real mapping. By analyzing the sensitivity of equipment actions, key process links, and model prediction errors, the set of parameters most sensitive to synchronization deviations is automatically identified and used for the calculation deviation of numerical differences or statistical indicators. When the deviation is within the preset tolerance, the system maintains synchronous operation. When the deviation exceeds the limit, realignment, dynamic correction, or multi-frequency rate matching is automatically triggered to correct the overall state of the twin model and achieve continuous consistency between virtual and real.

[0012] The beneficial effects of this invention are: This invention performs unified preprocessing on multi-source heterogeneous sensor data and uses discrete manufacturing equipment data as a time reference to dynamically align the twin model data in time. Based on this, an adaptive correction mechanism based on historical synchronization errors is established to compensate for sensor drift, noise interference, and model prediction errors in real time. Simultaneously, through high- and low-frequency data rate matching and interpolation compensation, high-frequency acquired data and low-frequency simulation models can operate collaboratively on a unified time scale. Through these technical means, the digital twin model can continuously and stably reflect the real operating status of discrete manufacturing equipment, providing a reliable data foundation for production monitoring, status analysis, and intelligent decision-making.

[0013] The high-dimensional feature projection, sequence alignment, and temporal mapping regularization steps can maintain high alignment accuracy while taking into account computational efficiency and robustness. They can accurately obtain the temporal deviation patterns of multi-source heterogeneous sensors on a digital twin platform, providing a reliable time pairing basis for subsequent dynamic correction and multi-frequency rate adaptation. At the same time, they can provide directly applicable alignment rules for processing signals with irregular or missing data.

[0014] The anomaly detection and anomaly data correction processing steps enable accurate detection and dynamic correction of anomalies in multi-source sensor data. It exhibits significant robustness in suppressing sensor drift, network latency, and random noise, and can complete real-time correction within millisecond response time, providing stable data support and computational assurance for high-precision virtual-real synchronization and state prediction of digital twin systems.

[0015] The frequency relationship determination and data classification processing steps can form a continuous, smooth and consistent model state sequence within the low-frequency model update cycle, providing a stable and reliable time reference for subsequent collaborative matching and unified processing of high-frequency equipment data, while effectively reducing the impact of low-frequency update lag on the accuracy of virtual-real system interaction.

[0016] The synchronous control and maintenance steps enable consistency judgment and feedback mechanisms to set judgment indicators and thresholds according to the characteristics of different discrete manufacturing systems, construct a "judgment-adjustment-operation" closed loop, realize long-term stable synchronization of the digital twin system in complex discrete manufacturing environments, and provide reliable support for production monitoring, status analysis and intelligent decision-making. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The preferred embodiments of the invention are given below with reference to the accompanying drawings to illustrate the technical solution of the invention in detail. The corresponding drawings will be provided for detailed explanation of the invention. It should be particularly noted that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit or restrict the invention.

[0019] like Figure 1 As shown, a method for synchronizing digital twin data between the physical and digital worlds in discrete manufacturing is described, and the specific implementation steps are as follows: The first step is geometric modeling: Based on CAD drawings and measurement data, a three-dimensional geometric model of the discrete manufacturing equipment is constructed, and a mapping relationship between the spatial position and structural parameters of key components is established.

[0020] The second step is behavior modeling: establish a corresponding kinematic or dynamic model based on the motion form and driving method of the equipment, so that the twin model can calculate the posture, velocity and force state of the equipment based on the input variables.

[0021] The third step is data modeling: establishing a data interaction channel between discrete manufacturing equipment and twin models through industrial communication interfaces, so that sensor data and model state parameters form a two-way mapping relationship.

[0022] The fourth step is signal processing: the raw signal data sequence of the discrete manufacturing equipment acquired at the physical end is filtered to obtain the discrete manufacturing equipment signal data sequence.

[0023] The fifth step is signal standardization: the filtered discrete manufacturing equipment signal data sequence is standardized to eliminate differences in dimensions and scales.

[0024] Step 6, Input Data Construction and Interface Adaptation: The standardized discrete manufacturing equipment signal data sequence is constructed into a time series state matrix and converted into an input data sequence that can be received by the twin model data interaction layer.

[0025] Step 7, High-dimensional feature projection and sequence alignment: Use CCA to map the raw signal data sequence directly acquired by discrete manufacturing equipment and the twin model input data sequence containing time distortion transmitted over the network to a low-dimensional common subspace, and use DTW to calculate the optimal matching path on the low-dimensional sequence.

[0026] Step 8, Time Mapping Regularization: Based on the optimal time matching path, reconstruct the time mapping relationship between discrete manufacturing equipment data and twin model data, resample and re-index the twin model data, and realize the regularization and alignment of time deviations.

[0027] Step 9, Anomaly Detection: Perform principal component analysis on the time-ordered input data to extract principal component information that reflects normal operation characteristics. When the data deviates from the specified feature range, it is determined to be abnormal data.

[0028] Step 10: Abnormal data correction processing: For data detected as abnormal, nonlinear estimation is performed using an artificial neural network trained based on historical normal operating conditions, and corrected data that conforms to normal operating characteristics is output to replace the original abnormal values.

[0029] Step 11, Frequency Relationship Determination: Obtain the sampling frequency of the discrete manufacturing equipment data and the update frequency of the digital twin model state, and determine the frequency relationship between the two, dividing the data into high-frequency data, low-frequency data, or frequency-matched data.

[0030] Step 12, Data Classification Processing: For data of different frequency types, high-frequency data is resampled, low-frequency data is interpolated and extrapolated for compensation, and frequency-matching data is directly synchronized and fused to achieve synchronization of data of different frequencies on a unified time scale.

[0031] Step 13, Synchronization Control and Maintenance: Based on the degree of difference between the state of the discrete manufacturing equipment and the state of the twin model, determine the current synchronization state, and perform synchronization correction operations when the deviation exceeds the limit to achieve continuous and stable operation of the virtual and real systems.

[0032] The specific implementation process of this method includes constructing a twin model (steps 1, 2, and 3), twin data preprocessing (steps 4, 5, and 6), multidimensional data temporal regularization (steps 7 and 8), dynamic correction (steps 9 and 10), high- and low-frequency matching (steps 11 and 12), and a virtual-real synchronization consistency determination and operation guarantee mechanism (step 13). Each method processes data sequentially and forms a closed-loop feedback structure to coordinate data at different sampling rates and generate a data stream with a unified time scale. Multi-source operating data of discrete manufacturing equipment is acquired, and the original data is filtered, standardized, and feature-mapped to establish a time mapping relationship between discrete manufacturing equipment data and twin model data. Real-time compensation is performed on the current data based on historical synchronization errors, coordinating data at different sampling rates and generating a data stream with a unified time scale. The virtual-real consistency determination module is used to compare the deviation between the state of the twin model and the state of the discrete manufacturing equipment, and triggers corresponding realignment or correction operations when the deviation exceeds the limit, thereby ensuring the long-term stable synchronous operation of the digital twin system.

[0033] I. Constructing a twin model The twin model construction method proposed in this invention is used to establish a twin model corresponding one-to-one with discrete manufacturing equipment in virtual space, and to provide a unified model carrier for virtual-real data synchronization, state prediction and collaborative control. The twin model adopts a three-layer integrated modeling architecture of "geometric structure-dynamic behavior-data interaction", which achieves high-fidelity mapping of the structural characteristics, motion behavior and operating status of discrete manufacturing equipment by organically combining the physical mechanism model and the data-driven model.

[0034] 1. Geometric Modeling Layer In the geometric structure construction stage of the twin model, it is necessary to perform refined structural modeling of the discrete manufacturing equipment. The external dimensions, assembly relationships, and spatial pose parameters of key components of the equipment are obtained through 3D scanning equipment, industrial cameras, and precision measuring instruments. The acquired 3D data is then corrected and fused with existing CAD design drawings to form a 3D geometric structure dataset of the equipment.

[0035] Based on this, the geometric data is constructed into a computable three-dimensional digital model using 3D modeling software. Unique identifiers and parameter interfaces are set for each key moving part, transmission part and execution unit, so that each part in the twin model can establish a mapping relationship with the corresponding part of the discrete manufacturing equipment, providing a foundation for subsequent dynamic modeling and virtual-real synchronization.

[0036] Dynamic Behavior Layer In the dynamic behavior construction stage, based on the specific structural form and motion mechanism of the discrete manufacturing equipment, corresponding kinematic or dynamic constraints are configured for the twin model. For equipment with multi-degree-of-freedom motion characteristics, such as multi-joint robotic arms, handling robots, or parallel mechanisms, a forward dynamic model is used to describe the mapping relationship from joint inputs to end effector pose, enabling the twin model to calculate the end effector pose in real time based on changes in joint angles, velocities, or driving forces.

[0037] The core idea of ​​forward dynamics is to calculate the spatial position and orientation of the robotic arm's end effector based on the rotational or translational angles of each joint. Its calculation formula can be expressed as: in, For the first The homogeneous transformation matrix of each joint relative to the previous joint, containing rotation and translation information; This represents the pose matrix of the end effector relative to the base coordinate system. By performing successive matrix multiplication, the three-dimensional spatial position and orientation of the end effector can be obtained, enabling dynamic simulation of the robotic arm's motion process using a twin model.

[0038] For equipment involving hydraulic drives, linear actuators, or complex energy exchange relationships, a dynamic modeling approach based on energy balance is adopted. This approach establishes a dynamic mapping between the equipment's state variables and input variables by describing the changes in the system's kinetic and potential energy. The Lagrange equations are applicable to continuous systems with complex constraints or multiple degrees of freedom. The generalized coordinates of the system can be established by analyzing the difference between the system's kinetic energy T and potential energy V (i.e., the Lagrange function L=TV). With generalized force The balance relationship between them is used to derive the dynamic equations of the equipment.

[0039] By employing the above methods, the twin model can maintain consistency with the discrete manufacturing equipment in terms of motion response, force changes, and state evolution, thereby ensuring the comparability and consistency of the virtual and real systems at the dynamic behavior level.

[0040] 3. Data Interaction Layer In the data interaction layer, a unified data acquisition and transmission interface is established for the multi-source sensors and control units of discrete manufacturing equipment to enable real-time information interaction between the discrete manufacturing equipment and the twin model. This layer establishes a connection with the field equipment through industrial communication protocols, transmitting multi-dimensional data such as position, temperature, torque, visual features, and operating status to the twin system in real time.

[0041] Within the twin model, the data is converted into input parameters compatible with both the dynamic and geometric models via an interface mapping module, driving the motion and state updates of the twin model device. Simultaneously, the state results calculated by the twin model can also be fed back to the control or monitoring system through this data interaction layer, enabling bidirectional information transfer between the discrete manufacturing equipment system and the twin model system.

[0042] Through the collaborative construction of the above-mentioned geometric, dynamic behaviors and data interactions, this invention forms a digital twin with physical constraints and data-driven capabilities, enabling the twin device to maintain synchronization with discrete manufacturing equipment at the structural, motion and state levels, providing a unified model basis for subsequent timing alignment, dynamic correction and multi-frequency rate matching.

[0043] II. Twin Data Preprocessing To ensure that the digital twin model accurately reflects the operating status of discrete manufacturing equipment and meets the computational requirements for subsequent timing alignment, dynamic correction, and multi-frequency matching, this invention systematically preprocesses multi-source sensor signals collected at the discrete manufacturing site before the operating data of the discrete manufacturing equipment enters the twin model. This preprocessing process converts unstructured, heterogeneous electrical signals acquired at the industrial site into twin input data with a unified structure, standardized timing, and direct applicability to model calculations.

[0044] The twin data preprocessing includes three stages: signal processing, signal standardization, and feature mapping. Each stage is executed sequentially and works in conjunction with the others to improve the stability, consistency, and computability of the input data.

[0045] 1. Signal Processing In discrete manufacturing environments, the raw signals acquired by sensors are often affected by mechanical vibration, electromagnetic interference, and communication jitter, resulting in random noise and high-frequency disturbances mixed into the data. This reduces the proportion of effective information in the signal and affects the accuracy of subsequent modeling and analysis.

[0046] To reduce the interference of noise and high-frequency disturbances on the effective information, a filtering function is introduced into the original signal for processing.

[0047] Assume there are a total of discrete manufacturing equipment N Each sensor collects data. T Data at each time point can be used to represent the original signal data sequence of the entire device as follows: in , For the first The actual signal from each sensor For noise components, Indicates the time sampling point number.

[0048] Through the filtering function By processing the original signal, we can obtain a filtered discrete device signal data sequence: The filtering function can be one or a combination of low-pass filtering, band-pass filtering, wavelet denoising, or Kalman filtering to suppress high-frequency noise and random fluctuations while retaining the main trend information of the equipment's operating status.

[0049] 2. Signal Standardization Even after filtering, the multi-source sensor signals still exhibit inconsistencies in units, dimensions, and numerical ranges (e.g., significant differences in numerical scale between temperature and vibration signals). To avoid bias in the twin model calculations caused by data of different scales, each signal channel is standardized.

[0050] Let the i-th The signal sequence is Its mean and standard deviation The calculation of the separated and standardized signals can be expressed as: After standardization, the data from each channel have a uniform mean and variance in a statistical sense, which enables data from different types of sensors to be jointly analyzed and calculated in the same model.

[0051] 3. Input data construction and interface adaptation After signal filtering and standardization, the outputs of each sensor yield a standardized signal sequence. This is applicable to a pre-defined discrete manufacturing equipment layout. N The sensor, the first i Each sensor at time t The standardized signal is represented as ,in , t This represents discrete-time sampling points.

[0052] To describe the overall operating status of the device at any given moment, the time values ​​of each sensor are compared. t Standardized signals are combined in channel order to construct instantaneous state vectors. Furthermore, in length of T Within the time window, the state vectors of consecutive sampling moments are stacked to form a time-series state matrix of the multi-source sensors: This matrix, with sensor channels as rows and time as columns, comprehensively describes the multidimensional operating state of the device within the stated time window, serving as the fundamental data representation for subsequent virtual-real synchronization and state calculations. To meet the interface requirements of the digital twin system's data interaction layer, the state matrix is ​​formatted with a length of [missing information - likely a unit of length] before entering the data interaction layer. T The time window is segmented and packaged, and the status column within each time window is combined into an independent data unit with additional time identification information to ensure the timing consistency during data transmission.

[0053] After the above processing, the original signal data sequence is transformed into a standardized discrete device signal data sequence and then processed according to... T The time window segments are treated as a continuous data stream and transmitted to the twin model via an industrial communication interface according to a unified data format. On the twin model side, the received data is parsed into corresponding states according to the interface rules, which are used to drive geometric model updates, dynamic calculations, and equipment state synchronization.

[0054] III. Time Series Regularization of Multidimensional Data Even after data processing, significant timing discrepancies and response delays still exist between the multi-sensor time-series data of discrete manufacturing equipment and its digital twin model. These discrepancies primarily stem from factors such as inconsistent sampling frequencies of the multiple sensors, network communication delays, and differences in equipment response. Without addressing these issues, the digital twin model is prone to lag or misjudgment in prediction, control, or state assessment, directly impacting the system's real-time performance and accuracy.

[0055] Existing technologies still have significant shortcomings in addressing this problem: traditional DTW methods only directly align one-dimensional single-sequence time, without considering the correlation and projection relationships between high-dimensional features, resulting in limited applicability; while multi-scale or multi-dimensional DTW methods improve sequence matching capabilities, they lack complete solutions in terms of real-time performance, robustness, and joint projection optimization; CTW methods enhance nonlinear sequence alignment capabilities, but are mainly applied in vision or motion capture fields and have not yet been optimized for multi-source high-frequency data from industrial heterogeneous sensors and online real-time applications. In summary, existing methods cannot achieve virtual-real alignment of multi-source heterogeneous, high-dimensional digital twin data while ensuring high accuracy, robustness, and linearity requirements.

[0056] To address the aforementioned technical problems, this invention proposes an adaptive time-series warping method for multi-source heterogeneous and multi-dimensional characteristic time-series signals, achieving alignment and time synchronization between digital twin data and discrete manufacturing equipment data at the feature level. The specific strategy is as follows: In this method, raw data from discrete manufacturing equipment is directly acquired from each sensor channel, forming a raw signal data sequence that is not transmitted over a network: It contains real-time sampled values ​​from each channel, representing the actual operating status of discrete manufacturing equipment.

[0057] The data matrix used as input to the twin model, which still contains some temporal distortion after preprocessing and network transmission, is considered as follows: Based on this, the present invention uses X As a reference standard, Y Treating the data as objects to be aligned, an iterative CTW (Canonical Time Warping) optimization framework is constructed. In existing technologies, traditional DTW methods are mainly used for time alignment of low-dimensional sequences. However, in the case of high-dimensional multi-sensor data, direct alignment is easily affected by irrelevant features or noise, resulting in limited accuracy. Therefore, this invention introduces CCA (Canonical Correlation Analysis) to perform dimensionality reduction projection on high-dimensional sequences, mapping the original high-dimensional features to a low-dimensional common subspace. This maximizes the statistical correlation of the mapped sequences, preserving the most useful common trends between the two sequences and providing stable and reliable input for subsequent DTW alignment.

[0058] First, through the linear projection matrix , Mapping the original high-dimensional sequence to a common low-dimensional subspace, such that the mapped sequence... and Maximizing statistical relevance can be formalized as follows: in ， ， These are the sample covariance and cross-covariance matrices, respectively; this step can be transformed into solving a generalized eigenvalue problem to obtain U and V. On the projection subspace, Dynamic Time Warping (DTW) is used to calculate the optimal pairing path. And minimize the cumulative distance: Distance metric Either Euclidean distance or weighted Mahalanobis distance can be used. The DTW subproblem is solved recursively using dynamic programming: Let... D(i,j) for i arrive j The minimum cumulative cost is then: This represents the minimum cumulative cost to reach the current point from the left, top, or top-left direction.

[0059] Dynamic programming can be used to find the globally optimal matching path, which is considered as... in T The data within the time window accurately describes the time deviation between the real signals of discrete manufacturing equipment and the digital twin signals. Simultaneously, the optimal matching path is considered as a set of time index correspondences. These correspondences are used to resample and rearrange the twin model data on the time axis, ensuring a one-to-one correspondence between it and the real data of the discrete manufacturing equipment under a unified time reference. This eliminates time misalignment caused by network latency, interface differences, and sampling asynchrony.

[0060] After the above processing, the original twin model input data with timing discrepancies is transformed into a synchronized sequence that is strictly aligned with the physical device data on the time axis. Let the aligned twin model input data matrix be... It uses the optimal time matching path The original data Y is resampled and indexed and rearranged to obtain: in This indicates a timeline resampling operation based on the matching path. This is the input to the time-normalized twin model, which serves as the standard input for the subsequent dynamic correction module.

[0061] To meet the real-time requirements of online and large-scale scenarios, this invention introduces multi-resolution acceleration and pruning strategies (such as FastDTW and PrunedDTW): first, a rough path is quickly estimated at a coarse-grained level, and then the optimal path is locally searched at a fine-grained level, thereby reducing the original O(NM) complexity to approximately linear level. Combining a sliding window and incremental update strategy, when new data arrives, local CTW optimization is performed only on the data within the window, and the projection and path from the previous iteration are used as initial values ​​to achieve millisecond-level online time-aligned updates.

[0062] Through the above mechanism, the present invention can maintain high alignment accuracy while taking into account computational efficiency and robustness, accurately obtain the time deviation pattern of multi-source heterogeneous sensors on the digital twin platform, provide a reliable time pairing basis for subsequent dynamic correction and multi-frequency rate adaptation, and provide directly applicable alignment rules for processing signals with irregular or missing data.

[0063] IV. Dynamic Correction After completing the multidimensional feature time-series regularization, the input data of the discrete manufacturing equipment and the twin model have been aligned in the time dimension. However, in the actual industrial operating environment, they may still be affected by sensor failures, network jitter and noise interference, causing local data quality degradation, which in turn affects the consistency between the virtual and real worlds and the prediction accuracy of the digital twin system.

[0064] Existing patented methods for real-time digital twin calibration calculate the error and gradient between the digital twin and the physical system state, and then use the gradient to optimize and update model parameters to achieve overall model state alignment. These methods primarily focus on global parameter-level calibration, iteratively optimizing the global error to approximate the physical system state, thus providing a real-time and accurate overall state mapping foundation for the digital twin. However, these methods mainly adjust for the overall error and struggle to effectively address long-term sensor drift, local data anomalies, and nonlinear errors in high-dimensional multi-source data. Consequently, they cannot maintain high-precision, locally sensitive consistency between the digital and physical systems in complex industrial environments.

[0065] To address the shortcomings of the aforementioned methods in handling local sensor quality degradation, noise interference, and nonlinear drift, this invention proposes a dynamic correction method based on the synergy of principal component analysis (PCA) and artificial neural networks (ANN). By performing dimensionality reduction, anomaly detection, and nonlinear correction at the local data feature level, this method achieves refined data consistency correction. The dynamic correction method of this invention includes a four-stage process: "feature dimensionality reduction - error detection - nonlinear compensation - online update".

[0066] In the feature dimensionality reduction stage, the input matrix of the aligned Siamese model is... Principal component analysis (PCA) was used to extract the main variation patterns and separate the noise subspace. This was achieved by solving the covariance matrix. Calculate its eigendecomposition ,in Principal component eigenvector matrix, The corresponding eigenvalues ​​are diagonal matrices. Projecting the input matrix onto a low-dimensional subspace yields the principal component information, represented as follows: : Reconstruction error is defined as This error matrix is ​​used to capture potential abnormal trends or sensor drift characteristics.

[0067] In the error detection phase, outliers are identified by calculating the mean square index of the sample-level reconstruction error. Let the mean of the reconstruction error of historical normal operation data in the principal component space be . The standard deviation is , k If the coefficient is an empirical factor (generally taken as 2~3), then the anomaly detection threshold is defined as follows: when When the data at that moment is deemed abnormal, the threshold is set. This refers to the feature range mentioned in the text. It is obtained by statistically analyzing historical normal operation data in the principal component space and is used to determine whether the current data deviates from the normal operation mode.

[0068] In the nonlinear compensation stage, an artificial neural network (ANN) is introduced to perform nonlinear mapping and dynamic correction on the detected abnormal segments. Let the ANN mapping function be... ,in θ={W,b} Represents network parameters, including the weight matrix. W and bias vector b Reconstruction error E(t) Input, output correction increment : The loss function is defined as: The network employs a two- or three-layer feedforward structure, using ReLU or Tanh activation functions to enhance its ability to approximate nonlinear drift. After training, the output is corrected. This refers to high-quality synchronized data that has undergone dynamic compensation.

[0069] During the online update phase, considering the changing data distribution over time in an industrial environment, a sliding window adaptive update mechanism is designed. Let the window length be L, and calculate the principal component matrix of the latest window when new data arrives. The difference between the parameters is significant. When the difference exceeds a threshold, it triggers the recalculation of principal components and fine-tuning of ANN weights. The update rule is as follows: Where η is the learning rate. This serves as the gradient of the network parameters. This mechanism ensures that the ANN possesses adaptability and stability during long-term operation. By complementing the linear dimensionality reduction of PCA with the nonlinear mapping of ANN, this invention achieves accurate detection and dynamic correction of anomalies in multi-source sensor data. It exhibits significant robustness in suppressing sensor drift, network latency and random noise, and can complete real-time correction within a millisecond response time, providing stable data support and computational assurance for high-precision virtual-real synchronization and state prediction of digital twin systems.

[0070] V. High and Low Frequency Matching After completing multi-dimensional feature time-series regularization and dynamic correction, the data from discrete manufacturing equipment and the digital twin model are aligned on the time axis and corrected at the data quality level. However, in actual industrial systems, different sensors, controllers, and simulation modules typically have different sampling frequencies and update cycles. For example, high-frequency vibration signals, low-frequency process parameters, and simulation states based on physical models are often not updated synchronously. Directly using multi-frequency data for twin modeling and control will lead to information redundancy, information loss, or time-series aliasing, thereby affecting the consistency between virtual and real systems and control stability.

[0071] Compared to the multi-frequency data processing methods commonly used in existing digital twin systems, current technologies mostly employ fixed resampling or single interpolation to perform unified time-base processing on multi-source data. Their core assumption is that there are only constant time scale differences between different data sources. However, in discrete manufacturing environments, digital twin models and discrete manufacturing equipment are often affected by factors such as model complexity, computing resource allocation, and changes in production conditions. They may exhibit different update frequency characteristics at different stages: the sampling frequency of discrete manufacturing equipment may be higher than the update frequency of the twin model, or the update rate of the twin model may be higher than the field sensing rate. Furthermore, these frequency relationships exhibit significant time-varying characteristics with operating time and conditions. Under these circumstances, using a single, fixed frequency processing strategy makes it difficult to simultaneously ensure dynamic response accuracy and state stability.

[0072] To address the issue of inconsistent update frequencies between equipment sensing data and twin model states in discrete manufacturing systems, this invention proposes a matching method for high- and low-frequency data collaboration. This method employs targeted processing strategies for high-frequency and low-frequency data based on the differences in update frequencies from different data sources. This approach suppresses redundant fluctuations while preserving key dynamic characteristics, ensuring temporal consistency and state stability during the interaction between the virtual and physical systems.

[0073] To achieve targeted multi-frequency data coordination, this invention first determines the update rate of the discrete manufacturing equipment data and the twin model data. Let the sampling frequency of a certain sensor data on the discrete manufacturing equipment side be... The update frequency of the corresponding state variables in the twin model is The frequency ratio is then defined as: when When this occurs, it indicates that the physical side data is high-frequency data and there is sampling redundancy; when When this occurs, it indicates that the state update frequency of the twin model is higher than that of the physical data, indicating a time gap; when At that time, it was assumed that the two rates were matched and could be directly synchronized and merged.

[0074] Based on the above frequency difference determination results, the present invention adaptively selects downsampling interpolation or extrapolation compensation strategies for different rate relationships in order to achieve a unified time scale expression of multi-frequency data.

[0075] when To avoid the impact of high-frequency data redundancy on the efficiency of twin modeling and computation, this invention performs downsampling processing on high-frequency sensor data. Specifically, using the update cycle of the twin model as the baseline time window, the continuously acquired high-frequency sensor data is divided into several subsequences, and the data is aggregated within each time window to obtain representative feature values ​​that can characterize the operating status of that time period.

[0076] The downsampling process can be implemented using the mean, weighted mean, or statistical feature extraction methods. For example, for the first... The time window contains High-frequency sample value The downsampling result can be expressed as: After processing, while preserving the key operating characteristics of the equipment, high-frequency noise and transient fluctuations are effectively suppressed.

[0077] when This indicates that there are update gaps in the low-frequency sensor data over time. To address this issue, this invention employs a time compensation processing method combining interpolation and extrapolation for the low-frequency model data, in order to construct a continuous and evolvable sequence of model states within the low-frequency update cycle.

[0078] When the target time point falls between two adjacent low-frequency model updates, an interpolation method is used to estimate the model state at that time point, as shown in the following formula: in, and The first i Second and third i+1 Low-frequency model state values ​​during the next update; and For the corresponding time point; t For the target time point; This represents the estimated low-frequency model state. Interpolation allows for the recovery of continuous state changes within low-frequency update cycles, avoiding gaps and jumps.

[0079] When the target time point exceeds the most recent low-frequency update, extrapolation estimation is introduced to compensate for the state lag: in, This is the state value from the most recent infrequent update. For the corresponding time point, The rate of change of historical states, t For the target time point that needs compensation, This is the state estimate obtained through extrapolation. The rate of state change. It can be calculated from the differences and time intervals of multiple historical update states, reflecting the evolution trend of the model state under the current working conditions.

[0080] By combining interpolation and extrapolation for time compensation, this invention can form a continuous, smooth and consistent model state sequence within the low-frequency model update cycle, providing a stable and reliable time reference for subsequent collaborative matching and unified processing of high-frequency equipment data, while effectively reducing the impact of low-frequency update lag on the accuracy of virtual-real system interaction.

[0081] VI. Mechanism for Determining and Ensuring the Consistency of Virtual and Real-World Synchronization During system operation, the overall operating state of the discrete manufacturing equipment is compared with that of the twin model. The overall operating state of the discrete manufacturing equipment includes real-time data collected by various sensors and control units, such as pose, speed, torque, temperature, process parameters, and equipment motion characteristics. The overall operating state of the twin model is calculated using corresponding parameters and motion characteristics obtained through data preprocessing, time-series normalization, dynamic correction, and high- and low-frequency matching methods. Within a unified time window, key state variables are extracted from the corresponding parameters of the discrete manufacturing equipment and the twin model. This extraction process is based on the influence of each parameter on the consistency of virtual-real mapping. By analyzing the sensitivity of equipment motion, key process links, and model prediction errors, it automatically identifies the set of parameters most sensitive to synchronization deviations, thereby ensuring that the selected key state variables accurately reflect system deviations and are used for the calculation of numerical differences or statistical indicators, achieving quantification of virtual-real mapping consistency. When the deviation is within a preset tolerance, the system maintains synchronous operation; when the deviation exceeds the limit, it automatically triggers realignment, dynamic correction, or multi-frequency rate matching to correct the twin model state and achieve continuous virtual-real consistency.

[0082] This consistency judgment and feedback mechanism can set judgment indicators and thresholds according to the characteristics of different discrete manufacturing systems, construct a "judgment-adjustment-operation" closed loop, realize long-term stable synchronization of digital twin systems in complex discrete manufacturing environments, and provide reliable support for production monitoring, status analysis and intelligent decision-making.

[0083] The above are merely preferred embodiments of the present invention and are not intended to limit or restrict the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection declared by the present invention.

Claims

1. A method for synchronizing virtual and real digital twin data in discrete manufacturing, characterized in that, Includes the following steps: The first step is geometric modeling: based on CAD drawings and measurement data, a twin model of the discrete manufacturing equipment is constructed, and a mapping relationship between the spatial position and structural parameters of key components is established; The second step is behavioral modeling: Based on the motion form and driving method of the discrete manufacturing equipment, establish corresponding kinematic or dynamic constraints so that the twin model can calculate the pose, velocity and force state of the discrete manufacturing equipment based on the input data. The third step is data modeling: establishing a data interaction channel between discrete manufacturing equipment and the twin model through an industrial communication interface, so that sensor data and twin model state parameters form a two-way mapping relationship; The fourth step is signal processing: filtering is performed on the collected real signal data of discrete manufacturing equipment to obtain the signal data sequence of discrete manufacturing equipment. The fifth step is signal standardization: the discrete manufacturing equipment signal data sequence obtained in the fourth step is standardized to eliminate differences in dimensions and scales; Step 6, Feature Mapping: The discrete manufacturing equipment signal data sequence obtained in Step 5 is converted into input data that can be received by the twin model through the feature mapping matrix; Step 7, High-dimensional feature projection and sequence alignment: Use CCA to map the real signal data directly collected by the discrete manufacturing equipment and the input data obtained in step 6 to a low-dimensional common subspace; use DTW to calculate the optimal matching path on the low-dimensional sequence; Step 8, Temporal Mapping Regularization: Based on the optimal matching path, reconstruct the temporal mapping relationship between discrete manufacturing equipment data and twin model data, resample and re-index the twin model input data, and realize the regularization and alignment of temporal deviations; Step 9, Anomaly Detection: Principal component analysis is performed on the input data of the twin model after time-series normalization to extract principal component information that reflects normal operation characteristics. The reconstruction error of the principal component information is calculated, and the mean square index is used to identify outliers. When the mean square index identifies outliers that are greater than the set feature range, the data is judged as abnormal. Step 10, Abnormal Data Correction Processing: For data detected as abnormal, nonlinear estimation is performed using an artificial neural network trained based on historical normal operating conditions, and corrected data that conforms to normal operating characteristics is output to replace the original abnormal values. Step 11, Frequency Relationship Determination: Obtain the sampling frequency of the discrete manufacturing equipment data and the update frequency of the twin model state, and determine the frequency relationship between the two. Divide the sampling frequency of the discrete manufacturing equipment data into high-frequency data, low-frequency data, or frequency-matching data. Step 12, Data Classification Processing: The data of different frequency types obtained in Step 11 are processed separately to achieve synchronization of data of different frequencies on a unified time scale; Step 13, Synchronization Control and Maintenance: Based on the degree of difference between the overall operating status of the discrete manufacturing equipment and the overall operating status of the twin model, determine the current synchronization status, and perform synchronization correction operations when the deviation exceeds the limit to achieve continuous and stable operation of the virtual and real systems.

2. The method for synchronizing virtual and real data in digital twins for discrete manufacturing as described in claim 1, characterized in that, The data partitioning method in step eleven is as follows: Suppose that the data sampling frequency of a certain sensor on the discrete manufacturing equipment side is . The update frequency of the corresponding state variables in the twin model is The frequency ratio is then defined as: when When, it is high-frequency data; when When, it is low-frequency data; when At that time, it is frequency matching data.

3. The method for synchronizing virtual and real data in digital twins for discrete manufacturing as described in claim 1, characterized in that: The method for processing data of different frequency types is as follows: high-frequency data is resampled, low-frequency data is interpolated and extrapolated for compensation, and frequency-matching data is directly and synchronously fused.

4. The method for synchronizing virtual and real data in digital twins for discrete manufacturing as described in claim 1, characterized in that: The process of calculating the reconstruction error is as follows: in, Principal component eigenvector matrix, , , This is the input matrix for the twin model.

5. A method for synchronizing virtual and real data in digital twins for discrete manufacturing as described in claim 1, characterized in that: The formula for calculating the feature range is: Among them, the mean of the reconstruction error of historical normal operation data in the principal component space is The standard deviation is , k This is an empirical coefficient.

6. The method for synchronizing virtual and real data in digital twins for discrete manufacturing as described in claim 1, characterized in that, The specific implementation process of step thirteen is as follows: Within a unified time window, key state variables are extracted from the corresponding parameters of the discrete manufacturing equipment and the twin model. This extraction process is based on the influence of each parameter on the consistency of virtual-real mapping. By analyzing the sensitivity of equipment actions, key process links, and model prediction errors, the set of parameters most sensitive to synchronization deviations is automatically identified and used for the calculation deviation of numerical differences or statistical indicators. When the deviation is within the preset tolerance, the system maintains synchronous operation. When the deviation exceeds the limit, realignment, dynamic correction, or multi-frequency rate matching is automatically triggered to correct the overall state of the twin model and achieve continuous consistency between virtual and real.

Images