Dynamic data mapping method and system for digital twinning

Abstract

The invention relates to the technical field of digital twinning, and discloses a digital twinning-oriented dynamic data mapping method and system.The digital twinning-oriented dynamic data mapping method comprises the steps that real-time data streams in the equipment operation process are collected and preprocessed; extracting samples and analyzing feature importance, identifying a current operation mode, dynamically selecting a feature subset and mapping parameters according to the mode, and executing dynamic mapping; feature correlation is analyzed and grouped, a mapping structure is constructed, and mapping calculation is executed layer by layer; a mapping residual error is calculated, an error evolution trend is analyzed, an updating process is triggered to re-identify the operation mode, and mapping parameters are updated; statistical analysis is carried out on mapping error characteristics, error characteristic decomposition is carried out, and deviation compensation and smoothing processing are carried out; detecting data abnormity and labeling a quality label, repairing abnormal data, executing mapping and labeling a source, and counting sensor abnormity to generate a maintenance early warning; according to the invention, high-precision real-time mapping of the digital twin system is realized.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and more specifically, to a dynamic data mapping method and system for digital twins. Background Technology

[0002] Digital twin technology, by constructing a virtual mapping model of a physical entity, enables real-time monitoring, simulation analysis, and optimized control of the physical system's operational status. It has become a key enabling technology in fields such as intelligent manufacturing, smart cities, and energy management. The core of a digital twin system lies in establishing a real-time, accurate, and dynamic data mapping relationship between the physical entity and the virtual model, allowing the virtual model to synchronously reflect the real state and behavioral characteristics of the physical entity.

[0003] In the field of intelligent manufacturing, there is a wide variety of production equipment, including CNC machine tools, industrial robots, sensor networks, and vision inspection systems. The data generated by these devices exhibits significant multi-source heterogeneity. The sampling frequencies of the data sources vary greatly, ranging from millisecond-level high-frequency data from machine tool controllers to second-level low-frequency data from production management systems, spanning multiple orders of magnitude in time scale. The time bases of different data sources come from their respective clock systems, resulting in clock drift and synchronization errors. Furthermore, the equipment is physically located in different positions, and their respective local coordinate systems exhibit translational and rotational relationships relative to the global coordinate system of the production line. These spatiotemporal heterogeneous characteristics pose challenges to data mapping.

[0004] Traditional digital twin data mapping methods primarily employ static rule configuration or fixed template-based mapping strategies. During system initialization, the correspondence between data sources and model parameters is established manually, such as multiplying sensor measurements by a fixed coefficient and assigning the result to the model parameters. While this static mapping method can meet basic requirements in the initial stages of system operation, changes in the physical entity's operating conditions—such as process technology switching, equipment status adjustments, or product model changes—alter the relationship between data features and model parameters. The fixed mapping rules become ineffective, causing the virtual model to fail to accurately reflect the true state of the physical entity. Furthermore, static mapping methods lack dynamic data quality assessment and anomaly handling capabilities. When sensors malfunction or data is missing, the mapping process is prone to interruption or erroneous results.

[0005] As digital twin applications become more complex and sophisticated, higher demands are placed on data mapping technologies. It is necessary to address the spatiotemporal alignment problem of multi-source heterogeneous data, ensuring that data with different sampling frequencies and time bases can establish accurate mapping relationships within a unified spatiotemporal coordinate system. Dynamic adaptive adjustment of mapping relationships is also required, enabling mapping strategies to automatically optimize based on changes in the physical entity's state while maintaining mapping accuracy. Simultaneously, in high-dimensional data spaces, it is necessary to reduce the computational complexity of mapping to meet real-time requirements and enhance the robustness of the mapping process to data anomalies and missing data.

[0006] Therefore, there is an urgent need for a dynamic data mapping method that can cope with the temporal inconsistency of multi-source heterogeneous data, adapt to dynamic changes in mapping relationships, reduce the computational complexity of high-dimensional mapping, and ensure the continuity of mapping under abnormal data conditions, so as to support the high-fidelity application of digital twin systems in complex industrial scenarios. Summary of the Invention

[0007] This invention provides a dynamic data mapping method and system for digital twins, which solves the technical problems of inconsistent timing of multi-source heterogeneous data and the lack of dynamic evaluation of data quality and anomaly handling in static mapping methods in related technologies.

[0008] This invention provides a dynamic data mapping method for digital twins, comprising: A standardized data stream is obtained by acquiring real-time data streams from the operation of the equipment using multi-source sensors and preprocessing them. Based on standardized data streams, samples are extracted and analyzed to identify the current operating mode. Feature subsets and mapping parameters are dynamically selected according to the mode, and dynamic mapping is performed to obtain dynamic mapping results. Based on the dynamic mapping results, feature correlation is analyzed and grouped, a mapping structure is constructed and mapping calculations are performed layer by layer to obtain hierarchical mapping results; The mapping residuals are calculated based on the hierarchical mapping results and the error evolution trend is analyzed. The update process is triggered to re-identify the running mode and update the mapping parameters to obtain the optimized and updated mapping results. Based on the optimized and updated mapping results, statistical analysis of mapping error characteristics is performed, error feature decomposition is carried out, and deviation compensation and smoothing are performed to obtain the self-corrected mapping results. Based on the standardized data stream and the mapping results after error self-correction, data anomalies are detected and quality labels are marked. Abnormal data is repaired, mapping is performed and the source is marked. Sensor anomalies are statistically generated to generate maintenance warnings, and anomaly-tolerant mapping results and sensor health management information are obtained.

[0009] In a preferred embodiment, the step of obtaining the standardized data stream includes: Each data source device connects to the clock synchronization network to synchronize the clock with the time server and obtains a global timestamp during data acquisition; The greatest common divisor of the sampling periods of each data source is calculated to obtain the basic time unit. Spatiotemporal anchor points are set with the basic time unit as the interval, and the time axis is divided into time slice sequences. The sliding window method is used to calculate statistical features within time slices, and the statistical features are used as attribute values ​​for time slices. For low-frequency data streams, a state-preservation strategy is used to populate data values ​​into subsequent time slices until the next update; The position data of each device in the local coordinate system is transformed to the global coordinate system using a coordinate transformation matrix; By integrating the results of time alignment and spatial alignment, a unified spatiotemporal data stream is constructed with the time-sliced ​​index as the primary key.

[0010] In a preferred embodiment, the step of obtaining the dynamic mapping result includes: A time window is selected from the standardized data stream, the data sequences of each sensor are extracted to form a feature matrix, the target parameters of the virtual model are extracted to form a target vector, and a mapping learning sample set is obtained. Calculate the mutual information value between each input feature and the target parameter, and select features whose mutual information value exceeds a preset importance threshold as the effective feature subset; Key parameters of the operating modes are extracted from historical data and clustered. The feature vector of each operating mode is calculated and stored in the operating mode library. The key parameters of the current time window are compared with the feature vectors of each mode in the running mode library, and the mode with the highest similarity is selected as the current running mode. Query the mapping model library based on the current running mode label. If the mapping exists, load the mapping parameters. If it does not exist, establish a mapping relationship based on the effective feature subset and store it. The feature matrix is ​​input into the mapping model to calculate the predicted values ​​of the target parameters, and the mapping parameters are fine-tuned based on the mapping residuals.

[0011] In a preferred embodiment, the step of obtaining the hierarchical mapping result includes: Calculate the correlation coefficient between features. If the absolute value of the correlation coefficient is greater than a preset strong correlation threshold, it is considered to be strongly correlated, and the feature correlation matrix is ​​obtained. The feature correlation matrix is ​​treated as an undirected graph, and a graph clustering method is used to divide the nodes into several subgraphs to obtain the feature grouping results. Analyze the mapping relationship between each feature group and the target parameters of the virtual model, and construct a multi-layer mapping structure; The mapping calculation is performed layer by layer according to the hierarchical structure, and the feature groups of each layer are input into the mapping function for weighted summation calculation; The mapping tasks for different feature groups are assigned to different computing cores for parallel execution, and the results are synchronized before proceeding to the next layer. Summarize the final model parameters output from the last layer of mapping and update the model state.

[0012] In a preferred embodiment, the step of obtaining the optimized and updated mapping result includes: The mapping residual is obtained by comparing the predicted output of the virtual model with the actual measured value of the physical entity, and then the mapping error sequence is obtained by organizing it. The slope is calculated after smoothing the mapping error sequence using a moving average method. When the absolute value of the slope is greater than the preset trend judgment threshold or a sudden error is detected, the mapping relationship is determined to have evolved and the update process is triggered. Collect data samples from the latest time window, recalculate the feature importance ranking, re-execute the operation mode recognition, and add any new operation modes to the mode library. Based on the latest samples and the updated feature importance, the original parameters are used as initial values ​​to perform iterative optimization with new samples to obtain the updated mapping parameters. The relevant information on the updated mapping relationship is recorded in the evolutionary history database.

[0013] In a preferred embodiment, the step of obtaining the self-corrected mapping result includes: Extract error data from the stable operation period and calculate the mean and standard deviation of the error; The total error is decomposed into systematic bias and random error. The systematic bias is the mean of the error sequence, and the random error is the difference between the error value at each time step and the systematic bias. For systematic deviations, a deviation compensation function is constructed for correction; for random errors, a filtering method is used for smoothing. The corrected and smoothed mapping output updates the model state, while the residual information is fed back to the mapping relationship learning module.

[0014] In a preferred embodiment, the steps of obtaining the fault-tolerant mapping result and sensor health management information include: Anomaly detection is performed on the data records. When the data value exceeds the normal range, it is marked as a statistical anomaly. When the data value violates the physical constraints, it is marked as a physical anomaly. When the rate of change exceeds the preset threshold, it is marked as a dynamic anomaly. If there are redundant sensors for abnormal data, the data value of the redundant sensor is used as a substitute, and the data source is marked in the quality label; If there are no redundant sensors, the time-series interpolation method is used to calculate the data values ​​at abnormal times; In cases where data is missing for an extended period, estimates of the missing parameters are inferred using measurements from other sensors based on the physical model. Perform a mapping process on the repaired data, and annotate the data source and confidence level in the mapping results; The frequency of data anomalies from each sensor is counted, and a sensor health warning message is generated when the anomaly rate exceeds a preset health warning threshold.

[0015] In a preferred embodiment, the length of the selected time window is set according to the complete processing cycle of the equipment. The key parameters of the operating mode include the spindle speed range, feed rate range, depth of cut, tool type and process type. The historical operating data is divided into roughing high-speed mode, roughing low-speed mode, finishing high-precision mode, drilling mode, tapping mode, no-load mode, fault mode and transition mode by clustering algorithm.

[0016] In a preferred embodiment, the multi-layer mapping structure includes a first-layer mapping, a second-layer mapping, and a third-layer mapping. The first-layer mapping maps the original feature set to intermediate parameters, the second-layer mapping maps the intermediate parameters and feature set to the next-level intermediate parameters, and the third-layer mapping maps the intermediate parameters to the final model parameters. When the absolute value of the slope of the mapping error sequence is greater than the preset trend judgment threshold or an error mutation is detected, feature importance analysis and running pattern recognition are re-executed to update the mapping parameters.

[0017] This invention provides a dynamic data mapping system for digital twins, used to execute the aforementioned dynamic data mapping method for digital twins, comprising: The data acquisition module collects real-time data streams from the multi-source sensors during the operation of the equipment and performs preprocessing to obtain a standardized data stream; The mapping learning module extracts samples and analyzes and identifies the current operating mode based on a standardized data stream. It dynamically selects feature subsets and mapping parameters according to the mode and performs dynamic mapping to obtain dynamic mapping results. The dimensionality reduction calculation module analyzes feature correlation and groups based on dynamic mapping results, constructs a mapping structure, and performs mapping calculations layer by layer to obtain hierarchical mapping results. The evolution tracking module calculates the mapping residuals based on the hierarchical mapping results and analyzes the error evolution trend, triggers the update process to re-identify the operating mode and update the mapping parameters, and obtains the optimized and updated mapping results. The error correction module statistically analyzes the mapping error characteristics based on the optimized and updated mapping results, performs error feature decomposition, and performs deviation compensation and smoothing to obtain the self-corrected mapping results. The robust processing module, based on the standardized data stream and the mapping results after error self-correction, detects data anomalies and marks them with quality labels, repairs abnormal data, performs mapping and marks the source, statistically analyzes sensor anomalies to generate maintenance warnings, and obtains anomaly-tolerant mapping results and sensor health management information.

[0018] The beneficial effects of this invention are as follows: By establishing a unified spatiotemporal anchor reference system, data streams with different sampling frequencies and time bases are aligned to a unified time-sliced ​​sequence, achieving precise correlation between high-frequency and low-frequency data in the time dimension. This eliminates mapping inaccuracies caused by temporal misalignment, improving data utilization and mapping accuracy. Through operational pattern recognition and dynamic feature importance analysis, the system can automatically adjust mapping strategies and parameter weights according to changes in the physical entity's operating conditions. This overcomes the limitations of traditional fixed mapping rules that cannot adapt to state changes, enabling the digital twin model to continuously and accurately reflect the true state of the physical entity, thus enhancing the adaptability and robustness of the mapping.

[0019] By employing feature correlation analysis and hierarchical mapping structure construction, the high-dimensional mapping problem is decomposed into multiple low-dimensional sub-problems. Combined with parallel computing strategies, this significantly shortens the mapping update cycle, meets the response requirements of real-time monitoring, and improves the system's real-time performance. Through mapping error evolution tracking and incremental learning updates, the system can automatically identify changes in mapping relationships and adjust model parameters in a timely manner, avoiding the problem of mapping accuracy degrading over time and ensuring the stability of mapping quality during long-term operation.

[0020] Through residual decomposition and deviation compensation mechanisms, the system can identify and correct systematic deviations, suppress random errors through filtering and smoothing, and form a closed-loop optimization feedback, thereby continuously improving mapping accuracy and enhancing the reliability of the mapping results. Through multi-level anomaly detection and repair strategies, including redundant sensor replacement, temporal interpolation repair, and physical model inference, the system can continue to provide reasonable mapping outputs even in the event of sensor failure or data loss, avoiding virtual model state interruptions or errors, and improving the system's fault tolerance and availability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the main process of a dynamic data mapping method for digital twins according to the present invention; Figure 2 This is a detailed flowchart of a dynamic data mapping method for digital twins according to the present invention; Figure 3 This is a block diagram of a dynamic data mapping system for digital twins according to the present invention. Detailed Implementation

[0022] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a dynamic data mapping method for digital twins, such as... Figures 1 to 2 As shown, it includes: Step 1: Collect real-time data streams from the multi-source sensor acquisition process of the equipment and preprocess them to obtain a standardized data stream; Specifically, the following steps are included: Step 1.1: Deploy a clock synchronization network and establish a global time base; Based on the production line network infrastructure, a network time protocol server is deployed in the production line local area network as a global time reference source.

[0024] The CNC systems, industrial robot controllers, programmable logic controllers, MES workstations, and other data source devices of each CNC machining center are connected to a clock synchronization network. A precision time protocol is used, and each device periodically synchronizes its clock with the time server, with the synchronization cycle set to once every 10 seconds.

[0025] The data collected by each data source device includes: machining parameters such as spindle speed, spindle current, feed rate, tool position, and cutting force collected by CNC machining centers; motion parameters such as joint angle, end effector position, motion speed, and load torque collected by industrial robot controllers; control signals such as equipment start / stop status, process switching signals, and alarm information collected by programmable logic controllers; physical quantities such as vibration amplitude, temperature, pressure, and flow rate collected by various sensors; and production management information such as current process, workpiece number, and production plan collected by MES workstations.

[0026] During data acquisition, each device uses its local clock to obtain the current timestamp. Since the local clock is synchronized with the global time base, the obtained timestamp has global consistency. This results in a raw data stream with global timestamps. Each record in the data stream contains fields such as device identifier, sensor identifier, timestamp, and data value.

[0027] Step 1.2: Analyze the sampling characteristics of the data source and set the spatiotemporal anchor point; The sampling periods for each data source were determined: the CNC system spindle speed data sampling period was 10 milliseconds, the vibration sensor data sampling period was 100 milliseconds, the temperature sensor data sampling period was 1 second, the robot position data sampling period was 50 milliseconds, and the MES production status data update period was 30 seconds. The greatest common divisor of all sampling periods was calculated, resulting in a basic time unit of 10 milliseconds.

[0028] On the global timeline, spatiotemporal anchors are set at intervals based on basic time units, with each anchor corresponding to a time slice. The continuous timeline is divided into a sequence of time slices, with the sequence index increasing from zero, and each slice corresponding to a time length of 10 milliseconds. The resulting time slice sequence serves as a reference framework for data alignment.

[0029] Step 1.3: Perform sliding window aggregation on the high-frequency data stream; For CNC system spindle speed data with a sampling period of 10 milliseconds, since this data is consistent with the time slice period, each sampling point is directly aligned to the corresponding time slice without aggregation. For vibration sensor data with a sampling period of 100 milliseconds, each 100 milliseconds corresponds to 10 time slices. A sliding window method is adopted, with the window length set to 100 milliseconds, and the window contains 10 time slices.

[0030] Statistical features are calculated for the vibration data within the window, including average vibration amplitude, peak vibration value, effective vibration value, and dominant frequency of the vibration spectrum. The calculated statistical features are used as common attribute values ​​for the 10 time slices corresponding to the window. The aligned representation of the vibration data on the time slice sequence is obtained, with each time slice associated with vibration statistical features.

[0031] Step 1.4: Perform state-preserving filling on the low-frequency data stream; For temperature sensor data with a sampling period of 1 second, each second corresponds to 100 time slices. A state preservation strategy is adopted, keeping the temperature sampled values ​​filled into the subsequent 100 time slices until the next temperature sampling update. For MES production status data with an update period of 30 seconds, such as the current process and current workpiece number, the same state preservation strategy is adopted, filling the status values ​​into the subsequent 3000 time slices. This yields an aligned representation of low-frequency data in the time slice sequence, with each time slice associated with the currently valid temperature value and production status.

[0032] Step 1.5: Establish the global coordinate system for the production line and perform spatial coordinate transformation; In the production line space, the starting point of the production line is selected as the origin of the global coordinate system, with the direction along the production line as the positive X-axis and the direction perpendicular to the production line upward as the positive Z-axis, thus establishing a right-handed coordinate system.

[0033] Each piece of equipment in each process is established with its own local coordinate system at the factory to describe the internal motion and measurement of the equipment. For example, the origin of the local coordinate system of a CNC machining center is located at the center of the worktable, and the origin of the local coordinate system of a robot is located at the center of the base. Because the installation positions and orientations of the equipment on the production line vary, the local coordinate systems of each piece of equipment have different spatial relationships relative to the global coordinate system of the production line. The origin positions and axes of these local coordinate systems exhibit translational and rotational relationships relative to the global coordinate system. By measuring the coordinates of the origin of each piece of equipment's local coordinate system in the global coordinate system, and the direction vectors of each axis of the local coordinate system in the global coordinate system, a coordinate transformation matrix is ​​constructed.

[0034] For the position data of industrial robots, which is represented in the robot's local coordinate system, multiplying it by the corresponding coordinate transformation matrix yields the robot's position in the global coordinate system. For other data with spatial position attributes, such as machining point coordinates and inspection point coordinates, the same coordinate transformation method is used to unify them into the global coordinate system. This results in a spatially unified data stream, where all position information is represented in the global coordinate system.

[0035] Step 1.6: Generate a unified spatiotemporal data stream of multi-source heterogeneous data; By integrating the results of time alignment and spatial alignment, a unified spatiotemporal data stream structure is constructed. The data stream uses a time-slice index as the primary key, with each time slice containing data records from all data sources at that moment. Each data record includes the following fields: time-slice index, global timestamp, device identifier, sensor type (including type identifiers for physical quantity measurement devices such as speed sensors, force sensors, vibration sensors, temperature sensors, and position sensors), data value, data quality label, and spatial coordinates. For high-frequency data, each time slice has a corresponding actual sampled value or statistical characteristic value; for low-frequency data, each time slice has a padding hold value.

[0036] Since some older devices cannot access the clock synchronization network, a software clock synchronization method can be used to achieve time alignment. Specifically, a timestamp server is deployed at the data acquisition interface of the older devices. The timestamp server accesses the clock synchronization network to obtain the global time. When the older devices output data, the timestamp server immediately adds a global timestamp to the data. By recording the deviation between the device's local clock and the global clock, a clock drift model is established using a linear regression method. The deviation parameters are periodically calibrated to achieve time alignment of the data from the older devices.

[0037] Data quality labels are used to indicate whether the data source is actual sampling, statistical aggregation, or state preservation, providing a data reliability reference for subsequent mapping. A unified spatiotemporal data stream is stored in a time-series database, supporting fast querying by time-sharded indexes and filtering by device sensor type.

[0038] This step outputs a standardized data stream, which includes time-aligned, spatially uniform, and quality-labeled multi-source heterogeneous data. Step 2: Based on the standardized data stream, extract samples and analyze and identify the current operating mode. Dynamically select feature subsets and mapping parameters according to the mode, and perform dynamic mapping to obtain dynamic mapping results. Specifically, the following steps are included: Step 2.1, extract the mapping learning sample set; From the standardized data stream output in step 1, a time window is selected for sample extraction. The time window length is set to the duration of a complete processing cycle on the production line. For engine block machining, the cycle from loading to unloading a workpiece is approximately 5 minutes, so the time window length is set to 5 minutes, corresponding to 30,000 time slices. Within the time window, data sequences from various sensors are extracted, including spindle speed sequence, feed rate sequence, cutting force sequence, vibration amplitude sequence, temperature sequence, and robot joint angle sequence. The data sequences are organized according to sensor type to form a feature matrix. The rows of the feature matrix correspond to time slices, the columns correspond to sensor channels, and the matrix elements are the data values ​​of the corresponding time slices and sensors. Simultaneously, target parameters of the virtual model are extracted, such as spindle load, tool wear state, and predicted surface quality of the virtual machining center, forming a target vector. The feature matrix and target vector are combined to obtain the mapping learning sample set. In this embodiment, the sample set contains 30,000 samples, corresponding to 30,000 time slices in a 5-minute time window. Each sample contains a 120-dimensional feature vector (corresponding to 120 sensor channels) and a 30-dimensional target vector (corresponding to 30 virtual model parameters).

[0039] Step 2.2: Calculate and rank the feature importance; For the mapping learning sample set, the correlation strength between each input feature and the target parameter is analyzed. Using the mutual information method, for any input feature and any target parameter, the mutual information value is calculated. The mutual information value reflects the information contribution of the input feature to the target parameter. For 120 input features and 30 target parameters, 3600 mutual information values ​​(120 x 30) are calculated, constructing a feature importance matrix. For each target parameter, the mutual information values ​​of its corresponding 120 features are ranked; a larger mutual information value indicates a more important feature. An importance threshold is set, and features with mutual information values ​​exceeding the threshold are selected as the effective feature subset for that target parameter. The feature importance ranking and effective feature subset for each target parameter are obtained, providing a basis for feature selection in subsequent mapping model construction.

[0040] Step 2.3: Identify the equipment operating modes and establish a mode library; This step involves building the pattern library offline. Based on historical operating data, typical operating modes of the equipment are identified. Key parameters characterizing the operating modes are extracted from the historical data, including spindle speed range, feed rate range, depth of cut, tool type, and operation type. The historical data is grouped according to the combination of key parameters, with each group corresponding to one operating mode. The k-means clustering algorithm is used to cluster the key parameters, with the cluster centers representing the parameter configuration of the typical operating mode. The number of clusters is set to 8, resulting in 8 operating mode categories: roughing high-speed mode, roughing low-speed mode, finishing high-precision mode, drilling mode, tapping mode, no-load mode, fault mode, and transition mode. For each operating mode, the statistical characteristics of each sensor data under that mode, such as mean and variance, are calculated as the mode's feature vector. The 8 operating modes and their feature vectors are stored in the operating mode library, which serves as a reference benchmark for pattern recognition.

[0041] Step 2.4: Identify the operating mode of the current time window; This step is an online pattern recognition process. For the data in the current time window, key parameters of the same type as in step 2.3 are extracted, including spindle speed and feed rate, and their average value within the time window is calculated to form the current feature vector. The current feature vector is compared with the feature vectors of eight modes in the operating mode library, using Euclidean distance as the similarity metric; a smaller Euclidean distance indicates greater similarity. The mode with the smallest Euclidean distance is selected as the current operating mode. If the smallest Euclidean distance exceeds a preset new mode determination threshold, the current mode is determined to be a new operating mode, and the current feature vector is added to the operating mode library as a new mode, with a new mode label assigned to it. The operating mode label for the current time window is obtained, indicating the typical operating condition of the current equipment.

[0042] Step 2.5: Load or learn the mapping parameters corresponding to the running mode; A mapping model library is established to store the mapping parameters corresponding to different operating modes. Each record in the mapping model library contains mapping parameters such as operating mode label, feature weight vector, regression coefficient matrix, and bias vector. The library can be indexed and queried by operating mode label.

[0043] Based on the identified operating mode label, the mapping model library is queried to check if the corresponding mapping parameters already exist. If the mapping parameters for this mode exist in the mapping model library, they are loaded directly, including feature weight vectors, regression coefficient matrices, bias vectors, etc.; if the mapping parameters for this mode do not exist in the mapping model library, the mapping model is trained based on the sample set of the current time window.

[0044] Based on the effective feature subset obtained in step 2.2, features from this subset are selected as input to the mapping model. A multiple linear regression method is used to establish a linear mapping relationship between the feature subset and the target parameters, and the regression coefficients are solved using the least squares method. For target parameters with significant nonlinear relationships, support vector regression or neural network methods are used to establish a nonlinear mapping model. After training, the mapping parameters are stored in the mapping model library and associated with the corresponding running mode label. The mapping parameters corresponding to the current running mode are obtained, and these parameters are used to perform the feature-to-target parameter mapping calculation.

[0045] Step 2.6: Perform dynamic mapping and fine-tune parameters in real time; The feature matrix of the current time window is input into the loaded mapping model, and the predicted values ​​of the target parameters are calculated based on the mapping parameters. For each target parameter, the feature values ​​corresponding to its effective feature subset are extracted, and a weighted sum is performed with the regression coefficients. A bias term is added to obtain the mapping output of the target parameter. The mapping output is compared with the actual target value of the virtual model, such as the theoretical value obtained through simulation calculation, and the mapping residual is calculated. Based on the residual, the regression coefficients are fine-tuned using the gradient descent method, with the adjustment magnitude set to the learning rate multiplied by the residual gradient, and the learning rate set to 0.01. After several iterations of fine-tuning, the mapping residual gradually decreases, and the mapping accuracy improves. The fine-tuned mapping parameters are updated in the mapping model library, and the mapping calculation results are summarized.

[0046] Step 2 outputs the dynamic mapping results, including the predicted values ​​of 30 target parameters, the effective feature subset corresponding to the current operating mode, and the mapping parameters after real-time fine-tuning.

[0047] In some embodiments, since historical data to support pattern recognition is lacking in the early stages of production of new product models, transfer learning methods can be used. The aim is to quickly establish an initial mapping model. Specifically, operating pattern features and mapping parameters are extracted from the historical operating data of similar product models. Based on the differences between the new product and similar products in structural and process parameters, the transferred mapping parameters are scaled and offset. The adjusted parameters are used as the initial mapping model of the new product. As new product processing data accumulates, new data is gradually used to fine-tune the model and update the pattern, thereby achieving rapid cold start and continuous optimization of the mapping model.

[0048] Step 3: Analyze the feature correlation and group based on the dynamic mapping results, construct the mapping structure and perform mapping calculations layer by layer to obtain hierarchical mapping results; Specifically, the following steps are included: Step 3.1, calculate the feature correlation matrix; From the mapping learning sample set in step 2, extract the data sequences of all input features and calculate the Pearson correlation coefficients between each pair of features. For the 120 input features, calculate a 120x120 correlation matrix, where each element represents the Pearson correlation coefficient between any two features. The correlation coefficient ranges from -1 to +1, with the absolute value closer to 1 indicating a stronger correlation. Set a strong correlation threshold of 0.8; if the absolute value of the correlation coefficient between two features is greater than 0.8, the two features are considered strongly correlated. This yields the feature correlation matrix, which describes the correlation between features.

[0049] Step 3.2: Group features based on the correlation matrix; The feature correlation matrix is ​​treated as an adjacency matrix of an undirected graph, where nodes correspond to features. If two features are strongly correlated, an edge is connected between their corresponding nodes, with the edge weight being the absolute value of the correlation coefficient. A graph clustering algorithm, such as spectral clustering, is used to divide the nodes in the graph into several subgraphs. Nodes within the same subgraph are tightly connected, while connections between different subgraphs are sparse. This results in 15 feature groups, each containing several strongly correlated features. For example, spindle speed, spindle current, and spindle power are strongly correlated and grouped into the same feature group; cutting force in the X, Y, and Z directions are strongly correlated and grouped into another feature group. The resulting feature groupings are 15 in total, denoted as feature group 1 to feature group 15.

[0050] Step 3.3: Construct the hierarchical structure of feature mapping; The mapping relationship between each feature group and the target parameters of the virtual model is analyzed. If the output of a feature group is not the final model parameter, but an intermediate representation, then the feature group is assigned to the first layer of mapping. For example, the spindle speed group is mapped to the intermediate parameter spindle load. The spindle load is not the final state parameter of the virtual model, but an intermediate quantity used to calculate the final parameters such as equipment energy consumption and heat generation. Therefore, the spindle speed group is assigned to the first layer of mapping. If a feature group is mapped to the final parameter based on the intermediate representation, then it is assigned to the second layer of mapping. For example, the combination of spindle load and ambient temperature is mapped to the final parameter equipment temperature rise. This group is assigned to the second layer of mapping. And so on, a multi-layer mapping structure is constructed. In this embodiment, a three-layer mapping structure is constructed. The first layer contains 4 feature groups, mapped to 6 intermediate parameters; the second layer contains 6 intermediate parameters and 2 feature groups, mapped to 8 intermediate parameters; the third layer contains 8 intermediate parameters, mapped to 12 final model parameters. The hierarchical structure of the feature mapping is obtained, and the input-output relationship of each layer is clarified.

[0051] Step 3.4: Perform mapping calculations according to the hierarchical structure; Following the hierarchical structure, mapping calculations are performed layer by layer. The first layer of mapping involves inputting the four feature groups from the first layer into their respective mapping functions. These functions perform a weighted summation based on the regression coefficients and bias vectors learned in step 2, yielding six intermediate parameters. The second layer of mapping combines the six intermediate parameters from the first layer with the two feature groups from the second layer, inputting this combination into the second layer's mapping function to yield eight intermediate parameters. The third layer of mapping inputs the eight intermediate parameters from the second layer into the third layer's mapping function, yielding twelve final model parameters. The input and output dimensions of each layer are significantly smaller than the original 120-dimensional input to 30-dimensional output mapping, reducing the computational complexity of a single layer. This results in the hierarchical mapping calculations, namely the values ​​of the twelve final model parameters.

[0052] Step 3.5: Optimize mapping efficiency using parallel computing; In each layer of mapping computation, the mapping calculations for different feature groups are independent and can be executed in parallel. Multi-core processors or distributed computing nodes are deployed, distributing mapping tasks for the same layer to different computing cores or nodes. For example, the mapping tasks for the four feature groups in the first layer are distributed to four computing cores, executing the mapping computations simultaneously. After completion, the results are synchronized before proceeding to the next layer. Through parallel computing, the actual computation time for a single layer of mapping is reduced to the computation time of a single task. The overall computation time of the mapping process is the sum of the serial computation times of each layer. Due to the limited number of layers and the short computation time of each layer, the mapping update cycle is reduced from 500 milliseconds in the original method to 50 milliseconds, resulting in a computationally efficient mapping process.

[0053] Step 3.6: Summarize the mapping results and update the virtual model; The 12 final model parameters output from the third-layer mapping are summarized. These parameters include the spindle load, tool wear, predicted surface roughness, equipment temperature rise, energy consumption, vibration intensity, positional deviation, and speed fluctuation of the virtual machining center. These parameters are encapsulated according to the virtual model's data interface format and sent to the digital twin model library via the data bus. Upon receiving the parameter update message, the virtual model writes the new parameter values ​​to the model's state variables, triggering a state update and visualization rendering. In the 3D scene of the virtual production line, the equipment's color, animation, and labels update according to the changes in state parameters, allowing operators to observe the synchronous operation of the virtual and physical production lines in real time.

[0054] Step 3 outputs the hierarchical mapping results, including the updated values ​​of 12 final model parameters and their corresponding hierarchical mapping processes.

[0055] In some embodiments, since the computation of the feature correlation matrix is ​​time-consuming when the number of features reaches hundreds, sparse matrix representation and sampling approximation methods can be used to accelerate the correlation calculation. Specifically, preliminary grouping is performed based on the physical meaning of the features, such as grouping sensor features of the same device into one group, calculating the correlation between representative features of different groups, assuming strong correlation within a group and not requiring calculation, resulting in a sparse inter-group correlation matrix. A random sampling method is used to extract representative samples from the data sequence to calculate the correlation coefficient. The sample size is set to one-tenth of the full data. Through sampling approximation calculation, the computation time of the correlation matrix is ​​shortened by an order of magnitude under the premise that the accuracy loss is acceptable.

[0056] Step 4: Calculate the mapping residuals based on the hierarchical mapping results and analyze the error evolution trend. Trigger the update process to re-identify the operating mode and update the mapping parameters to obtain the optimized and updated mapping results. Specifically, the following steps are included: Step 4.1: Collect the mapping residuals and construct the error sequence; After updating the model parameters in step 3, the virtual model performs simulation calculations to predict the output performance of the equipment. For example, a virtual machining center simulates and calculates the surface roughness of the machined workpiece based on parameters such as spindle load and tool wear. The predicted output of the virtual model is compared with the actual measured values ​​of the physical entity. The actual measured values ​​come from the quality inspection equipment in the production line inspection process, which measures the true values ​​of machining quality parameters such as surface roughness of the workpiece after it leaves the production line. The difference between the predicted value and the true value is calculated to obtain the mapping residual.

[0057] For multiple target parameters, calculate their respective mapping residuals; organize the mapping residuals of multiple consecutive time windows in chronological order to obtain a mapping error sequence, where the horizontal axis of the error sequence is the time window index and the vertical axis is the residual value; the resulting mapping error sequence reflects the evolution of mapping accuracy over time.

[0058] Step 4.2: Analyze the error evolution trend and identify changes in relationships; Trend analysis is performed on the mapping error sequence. A moving average method is used to smooth the error sequence, with a window length of 10 time windows. The slope of the smoothed error sequence is calculated; a positive slope indicates an increasing error trend, while a negative slope indicates a decreasing error trend. A trend judgment threshold is set; if the absolute value of the slope is greater than the threshold, the error trend is considered significant. Abrupt change points in the error sequence are detected using a cumulative sum detection algorithm to identify moments when the mean or variance of the error changes abruptly. If a significant increase in the error trend or a sudden change is detected, the mapping relationship is considered to have evolved, requiring an update to the mapping model; otherwise, the mapping relationship is considered stable, and the current mapping model is maintained. The error evolution characteristics and relationship change judgment results are obtained.

[0059] Step 4.3: Trigger the mapping relationship update process; When the mapping relationship changes, an update process is triggered. The update process includes the following steps: collecting data samples from the latest time window, which reflect the relationship between the current features and the target parameters; re-executing the feature importance analysis in step 2, calculating the mutual information value of each feature with respect to the target parameters in the latest samples, and obtaining the updated feature importance ranking; comparing the updated ranking with the original ranking, and if the ranking changes significantly, such as the top three important features being replaced, then the feature space is determined to have changed; re-executing the running pattern recognition in step 2, comparing the current feature vector with the running pattern library, and if a new running pattern is identified, assigning a label to the new pattern and adding it to the pattern library; obtaining the updated feature importance and running pattern information.

[0060] Step 4.4: Update the mapping parameters using incremental learning; Based on the latest samples and updated feature importance, the mapping model parameters are updated. Incremental learning is employed to avoid complete retraining. The basic idea of ​​incremental learning is to locally adjust the parameters using new samples, based on the original mapping parameters. The specific method includes: adding new samples to the training set; using stochastic gradient descent with the original parameters as initial values; performing several rounds of iterative optimization with the new samples; setting the learning rate to one-tenth of the initial training rate to ensure moderate parameter adjustment and avoid excessive deviation from the original model; after iteration, the updated mapping parameters are obtained, including new regression coefficients and bias terms; the updated parameters are stored in the mapping model library, labeled with the update time and version number; thus, the updated mapping relationship model is obtained.

[0061] Step 4.5: Record the evolution history of the mapping relationship; Each mapping update is recorded in the evolutionary history database. The records include: update trigger time, trigger reason, operating mode before update, operating mode after update, feature importance ranking before update, feature importance ranking after update, mapping parameters before update, mapping parameters after update, and a comparison of mapping errors before and after update. The evolutionary history database provides a complete record of the dynamic changes in mapping relationships, facilitating the analysis of mapping evolution patterns and providing a basis for fault diagnosis and model optimization.

[0062] Step 4 outputs the optimized and updated mapping results, including the updated mapping parameters, running mode information, feature importance ranking, and evolutionary history data recording the update process.

[0063] In some embodiments, frequent updates to the mapping relationship may lead to system instability. An update cooling mechanism and an accumulation threshold determination can be adopted to balance the timeliness and stability of the update. Specifically, the update cooling time is set to 30 minutes. Within 30 minutes after a mapping relationship update, even if an error increase is detected, a new update will not be triggered to avoid frequent model switching in a short period of time. At the same time, an error accumulation threshold is set. An update is only triggered when the accumulated error, i.e., the integral value of the error sequence, exceeds a preset threshold to prevent false triggering caused by random error fluctuations. Through the dual control of the cooling mechanism and the accumulation threshold, the rationality of the mapping relationship update and the stability of the system operation are ensured.

[0064] Step 5: Based on the optimized and updated mapping results, statistically analyze the mapping error characteristics, decompose the error characteristics, and perform deviation compensation and smoothing to obtain the self-corrected mapping results. Specifically, the following steps are included: Step 5.1: Statistically analyze the characteristics of the mapping error; Based on the mapping error sequence obtained in step 4, error data from a stable operating period is extracted. The stable operating period refers to a time frame (at least 100 time windows) during which the mapping relationship remains unchanged and the operating mode is constant. During this period, the error mainly consists of inherent systematic biases and random errors inherent in the mapping model. The statistical characteristics of the error sequence are calculated, including the error mean, standard deviation, skewness, and kurtosis. The error mean reflects the magnitude of the systematic bias, and the standard deviation reflects the fluctuation range of the random error. A histogram of the error probability distribution is plotted to analyze whether the error distribution conforms to a normal distribution. The statistical characteristics of the mapping error, including the mean, standard deviation, skewness, kurtosis, and distribution shape, are obtained, providing a basis for subsequent error decomposition and compensation.

[0065] Step 5.2: Decompose the error into systematic bias and random error; Based on the error statistics obtained in step 5.1, the total error is decomposed into two parts. Systematic bias is defined as the mean of the error sequence, representing the stable bias of the mapping model during long-term operation. Random error is defined as the difference between the error value at each time point and the systematic bias, representing random fluctuations caused by noise and uncertainties. The value of systematic bias is calculated; for example, if the mean mapping error of a certain target parameter is positive 0.5, it indicates that the mapping model systematically underestimates this parameter, exhibiting a bias of positive 0.5. The random error sequence is calculated; its mean is approximately zero, and the standard deviation reflects the fluctuation amplitude. The error decomposition results clarify the source and nature of the bias.

[0066] Step 5.3: Construct the deviation compensation function and correct the mapping output; For the systematic deviations obtained in step 5.2, a deviation compensation function is constructed. The compensation method is selected based on the deviation characteristics: if the deviation is a constant value, a constant compensation function is used, adding the negative of the systematic deviation to the mapped output. In this embodiment, the systematic deviation of a certain target parameter is a constant positive 0.5, which is compensated by subtracting the positive 0.5. If the deviation changes with input characteristics, a variable compensation function is used. For example, when the deviation increases with the spindle speed, a linear compensation function is constructed, and the deviation compensation value equals the spindle speed multiplied by the compensation coefficient. Through regression analysis, the compensation coefficient is determined so that the mean of the compensated residuals approaches zero. The selection of the compensation function type is based on: analyzing the correlation between the systematic deviation and each input feature; if the correlation coefficients are all less than 0.3, it is determined to be a constant deviation, and constant compensation is used; if there are features with correlation coefficients greater than 0.3, it is determined to be a variable deviation, and a linear or nonlinear compensation function is used. The deviation compensation function is applied to the mapped output, and the results of each mapping calculation are corrected to obtain the deviation-corrected mapped output.

[0067] Step 5.4: Use filtering methods to smooth out random errors; To address the random errors obtained from the decomposition in step 5.2, a Kalman filter is used to smooth the mapped output. The Kalman filter establishes a state-space model, using the mapped output as the observed values ​​and the actual model parameters as the state variables. The filter iterates through prediction and update steps. The prediction step predicts the current state based on the previous state, while the update step fuses the predicted and observed values, performing optimal weighted fusion to obtain a smoothed state estimate. Filter parameters include process noise covariance and observation noise covariance, set according to error statistics. Process noise reflects the rate of change of the model parameters, while observation noise reflects the level of random error in the mapped output. After filtering, the random fluctuation amplitude of the mapped output is reduced, and the output curve is smoother, resulting in a smoothed mapped output.

[0068] Step 5.5: Form a closed-loop optimized mapping system; The mapped output, after bias correction and smoothing, is used as the final result to update the state of the digital twin model. Simultaneously, residual information is fed back to the mapping relationship learning module in step 2. In the next round of mapping parameter updates, the residual information is used as part of the optimization objective to adjust the mapping parameters to reduce future residuals. For example, if the weight of a certain feature causes systematic bias, the weight of that feature is reduced and the weights of other features are increased during parameter updates. Through this closed-loop feedback mechanism, the mapping system continuously self-optimizes, and the mapping accuracy is continuously improved.

[0069] Step 5 outputs the self-corrected mapping results, including the model parameter values ​​after bias compensation and smoothing.

[0070] In some embodiments, since it is difficult to establish an explicit mathematical model for the compensation function of systematic deviation, a lookup table compensation method can be used. The purpose is to quickly obtain the compensation value. Specifically, in the offline stage, systematic deviation values ​​under different combinations of input features are obtained through experiments or simulations. The input feature space is discretized into a grid, with each grid point corresponding to a deviation value. A deviation lookup table is constructed. During online mapping, the deviation value of the corresponding grid point is queried in the lookup table according to the current input feature value. If the input value is not on the grid point, the deviation value is calculated using a multilinear interpolation method. The queried deviation value is used to compensate the mapping output. By using the lookup table method, complex function calculations are avoided, and fast compensation is achieved.

[0071] Step 6: Based on the standardized data stream and the mapping results after error self-correction, detect data anomalies and label them with quality tags, repair the abnormal data, perform mapping and label the source, statistically generate maintenance warnings for sensor anomalies, and obtain the mapping results for anomaly tolerance and sensor health management information. Specifically, the following steps are included: Step 6.1: Detect data anomalies and label them with quality tags; Anomaly detection is performed on each data record in the standardized data stream output from step 1. A statistical threshold-based detection method is used to calculate the normal range of values ​​for each sensor data point, set as the historical data mean plus or minus three standard deviations. If the current data value exceeds this range, it is marked as a statistical anomaly. A physical constraint-based detection method is used, setting a physical feasible region based on the sensor's range and the operating characteristics of the physical entity. For example, if the temperature sensor's measurement range is -20°C to 200°C, a data value of 300°C violates the physical constraint and is marked as a physical anomaly. For dynamic parameters, the rate of change constraint is checked; for example, the rate of temperature change within one second should not exceed 10°C per second. If the rate of change exceeds this value, it is marked as a dynamic anomaly. The detection results are written to the quality label field of the data record. Quality labels include categories such as normal, statistical anomaly, physical anomaly, and dynamic anomaly. This results in a data stream labeled with quality labels.

[0072] Step 6.2: Replace abnormal data with redundant sensor data; For data marked as abnormal, check for redundant sensors. In production line design, multiple sensors are typically deployed at critical measurement points to improve reliability. For example, spindle temperature is measured simultaneously by two temperature sensors, sensor A and sensor B. If sensor A's data is marked as abnormal, check the data quality of sensor B. If sensor B's data label is normal, replace the abnormal data value of sensor A with the data value of sensor B. Update the data record with the replaced data value and indicate in the quality label that the data source is a redundant sensor replacement; this yields the corrected data value.

[0073] Step 6.3: Use time-series interpolation to repair short-term anomalies; If no redundant sensors exist, a time-series interpolation method is used. For the abnormal moment, normal data from the sensor before and after the abnormal moment is extracted. For example, if an anomaly is detected at time slice index 1000, normal data from indices 900 to 999 and 1001 to 1100 are extracted. A cubic spline interpolation method is used, based on the trend of the normal data before and after the anomaly, to interpolate and calculate a reasonable data value at index 1000. The interpolation method ensures the smoothness and continuity of the data curve, and the interpolation result conforms to the time-series variation pattern. The interpolated data value is updated in the data record, and the data source is marked as time-series interpolation repair in the quality label; the time-series interpolation repaired data value is obtained.

[0074] Step 6.4: Use a physical model to reason and repair long-term missing data; For situations involving prolonged data loss, such as a sensor malfunction causing 10 consecutive seconds of data absence, the accuracy of time-series interpolation methods decreases. In such cases, a physical model inference method is employed. Based on the kinetic or heat transfer equations of the physical entity, and utilizing measurements from other sensors, estimated values ​​for missing parameters are inferred. For example, when spindle temperature is missing, the spindle's thermal equilibrium state is calculated using known parameters such as spindle power, cooling water flow rate, and ambient temperature, combined with a heat transfer model, to infer the spindle temperature. Physical model inference requires establishing a mathematical model and solving equations, resulting in high computational complexity. To improve efficiency, the physical model response under typical operating conditions is pre-calculated offline, and a response library is established. During online inference, estimated values ​​are quickly obtained by querying the response library and interpolation. The inferred data values ​​are then updated in the data records, and the data source is marked as physical model inference in the quality label; this yields the data values ​​repaired by physical model inference.

[0075] Step 6.5: Execute the mapping process and label the data source; For the repaired data, execute the mapping process from steps 2 to 5 to obtain the mapping results based on the repaired data. In the mapping results, add a data source labeling field and a confidence level labeling field to indicate which target parameters used the repaired data for mapping, and whether the source of the repaired data is redundant sensors, time-series interpolation, or physical model inference. Determine the confidence level based on the data source: mapping results using the original data are labeled "high" confidence level; mapping results using redundant sensors to replace data are labeled "high" confidence level; mapping results using time-series interpolation to repair data are labeled "medium" confidence level; and mapping results using physical model inference data are labeled "low-medium" confidence level. The confidence level reflects the reliability of the mapping results, providing a reference for operator decision-making. In the visualization interface of the digital twin model, parameters using the repaired data are specially marked, such as by displaying different colors or adding warning icons, and the confidence level is displayed, indicating to the operator that the confidence level of this parameter is relatively low and the health status of the original sensor needs to be monitored; thus obtaining the mapping results labeled with data source and confidence level.

[0076] Step 6.6: Generate sensor health management information; The system analyzes the frequency of data anomalies for each sensor and calculates the proportion of abnormal data for each sensor within a specific time window (e.g., one hour). A health warning threshold of 5% is set. If the anomaly rate of a particular sensor exceeds 5%, a sensor health warning message is generated. This message includes the sensor identifier, anomaly rate, anomaly type statistics, and recommended maintenance measures. The warning message is then sent to the maintenance management system, prompting maintenance personnel to inspect, calibrate, or replace the sensor. Through sensor health management, sensor faults can be detected and addressed promptly, improving data quality.

[0077] Step 6 outputs the fault-tolerant mapping results and sensor health management information. The mapping results include the repaired model parameter values, data source labels, and confidence labels. The health management information includes sensor anomaly statistics and maintenance warnings to ensure the continuity of mapping under abnormal data conditions, serving as the final output of the application.

[0078] In some embodiments, due to the computational complexity of physical model inference limiting real-time performance, response library lookup and interpolation methods can be employed to accelerate the inference process. Specifically, in the offline stage, batch simulations of the physical model are performed, representative sampling points in the input parameter space are traversed, and the output response corresponding to each sampling point is calculated. The input sampling points and output responses are stored in a response library. During online inference, based on the current input parameter value, the nearest neighbor sampling point is queried in the response library. Multidimensional linear interpolation or radial basis function interpolation is used, and the output estimate corresponding to the current input is calculated based on the response values ​​of neighboring sampling points. Through the response library method, the computation time of physical model inference is reduced from the second level to the millisecond level, meeting the real-time mapping requirements.

[0079] A dynamic data mapping system for digital twins is provided for executing the aforementioned dynamic data mapping method for digital twins, such as... Figure 3 As shown, it includes: The data acquisition module collects real-time data streams from the multi-source sensors during the operation of the equipment and performs preprocessing to obtain a standardized data stream; The mapping learning module extracts samples and analyzes and identifies the current operating mode based on a standardized data stream. It dynamically selects feature subsets and mapping parameters according to the mode and performs dynamic mapping to obtain dynamic mapping results. The dimensionality reduction calculation module analyzes feature correlation and groups based on dynamic mapping results, constructs a mapping structure, and performs mapping calculations layer by layer to obtain hierarchical mapping results. The evolution tracking module calculates the mapping residuals based on the hierarchical mapping results and analyzes the error evolution trend, triggers the update process to re-identify the operating mode and update the mapping parameters, and obtains the optimized and updated mapping results. The error correction module statistically analyzes the mapping error characteristics based on the optimized and updated mapping results, performs error feature decomposition, and performs deviation compensation and smoothing to obtain the self-corrected mapping results. The robust processing module, based on the standardized data stream and the mapping results after error self-correction, detects data anomalies and marks them with quality labels, repairs abnormal data, performs mapping and marks the source, statistically analyzes sensor anomalies to generate maintenance warnings, and obtains anomaly-tolerant mapping results and sensor health management information.

[0080] In one embodiment of the present invention, a specific example is provided: This invention relates to a digital twin system for an engine block machining production line in an automobile manufacturing company. The production line comprises six main processes: rough milling, finish milling, drilling, tapping, cleaning, and inspection. The total length of the production line is approximately 80 meters, equipped with six CNC machining centers, eight industrial robots, and corresponding conveyor systems, producing approximately 200 engine blocks per day. The system deploys 120 sensors, including spindle speed sensors, feed rate sensors, and cutting force sensors for the CNC machine tools; joint angle sensors and end effector force sensors for the industrial robots; environmental temperature and humidity sensors; and vision inspection cameras, with data sampling frequencies ranging from 10 milliseconds to 30 seconds.

[0081] Using the spatiotemporal anchoring mechanism from step 1, 120 sensor data streams with different sampling frequencies are aligned to a unified 10-millisecond time-slice sequence. The base time unit is determined to be 10 milliseconds by calculating the greatest common divisor of the sampling periods of each data source. Sliding window aggregation is used for high-frequency data, while state-preserving filling is used for low-frequency data. A global coordinate system for the production line is established, and the local coordinates of each device are uniformly transformed to the global coordinate system representation. The aligned, standardized data stream is stored in a time-series database, supporting fast queries by time-slice index and sensor type. An example of the output standardized data stream is shown in Table 1. Table 1: Examples of standardized data flows;

[0082] Note: Spindle speed is in revolutions per minute, feed rate is in millimeters per minute, cutting force is in Newtons, vibration amplitude is in millimeters, temperature is in degrees Celsius, joint angle is in degrees, and spatial coordinates are in meters.

[0083] Using the method in step 2, eight typical operating modes were identified based on historical operating data, including rough milling high-speed mode, rough milling low-speed mode, finish milling precision mode, drilling mode, tapping mode, no-load mode, fault mode, and transition mode. For each mode, the importance of features was calculated through mutual information analysis, and the mapping relationship between features and model parameters was learned. In rough milling high-speed mode, the mutual information values ​​of spindle speed, feed rate, and cutting force were 0.85, 0.78, and 0.82, respectively, which are key features for mapping spindle load. In finish milling precision mode, the mutual information values ​​of vibration amplitude, temperature, and tool wear were 0.91, 0.76, and 0.88, respectively, which are key features for mapping machined surface quality. The system dynamically selects an effective subset of features and mapping parameters according to the current operating mode to achieve adaptive mapping.

[0084] Using the method in step 3, the 120 input features are divided into 15 feature groups through feature correlation analysis, constructing a 3-layer mapping structure. The first layer maps 4 feature groups to 6 intermediate parameters; the second layer maps 6 intermediate parameters and 2 feature groups to 8 intermediate parameters; and the third layer maps 8 intermediate parameters to 12 final model parameters. Parallel computing with 4 cores is employed, with mapping tasks within the same layer executed in parallel. The mapping update cycle is reduced from 500 milliseconds in the original method to 50 milliseconds, meeting real-time requirements.

[0085] Using the method in step 4, the evolution trend of the mapping error was continuously monitored during operation. On day 15, a tool change event was detected, causing a change in the mapping relationship and a sudden change in the error sequence. The system automatically triggered an update process, re-identifying the operating mode and using incremental learning to update the mapping parameters. After the update, the spindle load mapping error decreased from 12% before the change to 5%, and the surface roughness prediction error decreased from 8% to 3.5%. The evolution history database records the triggering reasons, parameter changes, and error comparisons for each update, providing a basis for subsequent optimization.

[0086] Using the method in step 5, the mapping error characteristics during stable operation were statistically analyzed, and the error was decomposed into systematic deviation and random error. A systematic deviation of +2 kW in the spindle load mapping was identified and corrected using a constant compensation function. For the random error, Kalman filtering was used for smoothing, reducing the standard deviation of the mapping output from 5 kW to 2 kW. Through the dual processing of deviation compensation and error smoothing, the mapping accuracy was improved by approximately 20%. The residual information was fed back to the mapping relationship learning module, forming a closed-loop optimization mechanism.

[0087] The method in step 6 was used to detect and repair anomalies in the data stream. During production on day 22, a faulty spindle temperature sensor on the CNC-03 was detected, resulting in a 15-second data loss (1500 time slices). The system first attempted to replace the faulty sensor, but found no redundant sensor. Then, a physical model inference method was used to infer the spindle temperature based on parameters such as spindle power, cooling water flow rate, and ambient temperature through a heat transfer model. The data source was labeled as "physical model inference" and the confidence level was labeled as "medium-low" in the mapping results. The mapping process remained continuous, and the digital twin model status was not interrupted. At the same time, a sensor health warning message was generated to prompt maintenance personnel to repair the sensor. An example of the mapping output results is shown in Table 2. Table 2: Example of mapping output results;

[0088] Note: The time window length is 5 minutes (30,000 time slices), the mapping residual is the predicted value minus the actual measured value, and the confidence level is determined according to the data source label (high for original data, high for redundant substitution, medium for time series interpolation, and medium-low for physical inference).

[0089] The system ran continuously for 30 days, processing approximately 260 million data entries and mapping the machining cycles of about 6,000 workpieces. Statistical results show that the average absolute error of the mapping accuracy was 4.5%, the response latency was 50 milliseconds, the sensor anomaly detection rate was 3.2%, the anomaly handling success rate was 98%, and the system availability reached 99.7%. The digital twin model showed good synchronization with the physical production line, with a virtual model state update latency of less than 100 milliseconds. During the testing period, the system successfully issued warnings for 5 instances of abnormal tool wear, 3 instances of abnormal equipment temperature, and 2 instances of sensor malfunction, preventing potential quality problems and equipment damage.

[0090] By applying the method of this invention, the production line digital twin system achieves real-time alignment and unified representation of multi-source heterogeneous data, increasing data utilization from 60% to 95%. Dynamic mapping based on operating modes improves mapping accuracy by approximately 25% compared to a fixed mapping model. Hierarchical mapping dimensionality reduction improves computational efficiency by approximately 10 times. A continuous optimization mechanism for mapping relationships adapts to changes in operating conditions and equipment aging. An error self-correction mechanism continuously improves mapping accuracy. Anomaly tolerance ensures continuous system operation even under data anomalies. The system provides a reliable digital twin support platform for enterprise production management, quality control, and equipment maintenance.

[0091] The above application examples demonstrate the effectiveness and practicality of the method of the present invention in actual industrial scenarios.

[0092] The system ran continuously for 30 days, processing approximately 260 million data entries and mapping the machining cycles of about 6,000 workpieces. Statistical results show that the average absolute error of the mapping accuracy was 4.5%, the response latency was 50 milliseconds, the sensor anomaly detection rate was 3.2%, the anomaly handling success rate was 98%, and the system availability reached 99.7%. The digital twin model showed good synchronization with the physical production line, with a virtual model state update latency of less than 100 milliseconds. During the testing period, the system successfully issued warnings for 5 instances of abnormal tool wear, 3 instances of abnormal equipment temperature, and 2 instances of sensor malfunction, preventing potential quality problems and equipment damage.

[0093] By applying the method of this invention, the production line digital twin system achieves real-time alignment and unified representation of multi-source heterogeneous data, increasing data utilization from 60% to 95%. Dynamic mapping based on operating modes improves mapping accuracy by approximately 25% compared to a fixed mapping model. Hierarchical mapping dimensionality reduction improves computational efficiency by approximately 10 times. A continuous optimization mechanism for mapping relationships adapts to changes in operating conditions and equipment aging. An error self-correction mechanism continuously improves mapping accuracy. Anomaly tolerance ensures continuous system operation even under data anomalies. The system provides a reliable digital twin support platform for enterprise production management, quality control, and equipment maintenance.

[0094] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A dynamic data mapping method for digital twins, characterized in that, Includes the following steps: A standardized data stream is obtained by acquiring real-time data streams from the operation of the equipment using multi-source sensors and preprocessing them. Based on standardized data streams, samples are extracted and analyzed to identify the current operating mode. Feature subsets and mapping parameters are dynamically selected according to the mode, and dynamic mapping is performed to obtain dynamic mapping results. Based on the dynamic mapping results, feature correlation is analyzed and grouped, a mapping structure is constructed and mapping calculations are performed layer by layer to obtain hierarchical mapping results; The mapping residuals are calculated based on the hierarchical mapping results and the error evolution trend is analyzed. The update process is triggered to re-identify the running mode and update the mapping parameters to obtain the optimized and updated mapping results. Based on the optimized and updated mapping results, statistical analysis of mapping error characteristics is performed, error feature decomposition is carried out, and deviation compensation and smoothing are performed to obtain the self-corrected mapping results. Based on the standardized data stream and the mapping results after error self-correction, data anomalies are detected and quality labels are marked. Abnormal data is repaired, mapping is performed and the source is marked. Sensor anomalies are statistically generated to generate maintenance warnings, and anomaly-tolerant mapping results and sensor health management information are obtained.

2. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the standardized data stream include: Each data source device connects to the clock synchronization network to synchronize the clock with the time server and obtains a global timestamp during data acquisition; The greatest common divisor of the sampling periods of each data source is calculated to obtain the basic time unit. Spatiotemporal anchor points are set with the basic time unit as the interval, and the time axis is divided into time slice sequences. The sliding window method is used to calculate statistical features within time slices, and the statistical features are used as attribute values ​​for time slices. For low-frequency data streams, a state-preservation strategy is used to populate data values ​​into subsequent time slices until the next update; The position data of each device in the local coordinate system is transformed to the global coordinate system using a coordinate transformation matrix; By integrating the results of time alignment and spatial alignment, a unified spatiotemporal data stream is constructed with the time-sliced ​​index as the primary key.

3. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the dynamic mapping result include: A time window is selected from the standardized data stream, the data sequences of each sensor are extracted to form a feature matrix, the target parameters of the virtual model are extracted to form a target vector, and a mapping learning sample set is obtained. Calculate the mutual information value between each input feature and the target parameter, and select features whose mutual information value exceeds a preset importance threshold as the effective feature subset; Key parameters of the operating modes are extracted from historical data and clustered. The feature vector of each operating mode is calculated and stored in the operating mode library. The key parameters of the current time window are compared with the feature vectors of each mode in the running mode library, and the mode with the highest similarity is selected as the current running mode. Query the mapping model library based on the current running mode label. If the mapping exists, load the mapping parameters. If it does not exist, establish a mapping relationship based on the effective feature subset and store it. The feature matrix is ​​input into the mapping model to calculate the predicted values ​​of the target parameters, and the mapping parameters are fine-tuned based on the mapping residuals.

4. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the hierarchical mapping result include: Calculate the correlation coefficient between features. If the absolute value of the correlation coefficient is greater than a preset strong correlation threshold, it is considered to be strongly correlated, and the feature correlation matrix is ​​obtained. The feature correlation matrix is ​​treated as an undirected graph, and a graph clustering method is used to divide the nodes into several subgraphs to obtain the feature grouping results. Analyze the mapping relationship between each feature group and the target parameters of the virtual model, and construct a multi-layer mapping structure; The mapping calculation is performed layer by layer according to the hierarchical structure, and the feature groups of each layer are input into the mapping function for weighted summation calculation; The mapping tasks for different feature groups are assigned to different computing cores for parallel execution, and the results are synchronized before proceeding to the next layer. Summarize the final model parameters output from the last layer of mapping and update the model state.

5. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the optimized and updated mapping results include: The mapping residual is obtained by comparing the predicted output of the virtual model with the actual measured value of the physical entity, and then the mapping error sequence is obtained by organizing it. The slope is calculated after smoothing the mapping error sequence using a moving average method. When the absolute value of the slope is greater than the preset trend judgment threshold or a sudden error is detected, the mapping relationship is determined to have evolved and the update process is triggered. Collect data samples from the latest time window, recalculate the feature importance ranking, re-execute the operation mode recognition, and add any new operation modes to the mode library. Based on the latest samples and the updated feature importance, the original parameters are used as initial values ​​to perform iterative optimization with new samples to obtain the updated mapping parameters. The relevant information on the updated mapping relationship is recorded in the evolutionary history database.

6. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the self-corrected mapping result include: Extract error data from the stable operation period and calculate the mean and standard deviation of the error; The total error is decomposed into systematic bias and random error. The systematic bias is the mean of the error sequence, and the random error is the difference between the error value at each time step and the systematic bias. For systematic deviations, a deviation compensation function is constructed for correction; for random errors, a filtering method is used for smoothing. The corrected and smoothed mapping output updates the model state, while the residual information is fed back to the mapping relationship learning module.

7. The dynamic data mapping method for digital twins according to claim 1, characterized in that, The steps for obtaining the fault-tolerant mapping results and sensor health management information include: Anomaly detection is performed on the data records. When the data value exceeds the normal range, it is marked as a statistical anomaly. When the data value violates the physical constraints, it is marked as a physical anomaly. When the rate of change exceeds a preset threshold, it is marked as a dynamic anomaly. If there are redundant sensors for abnormal data, the data value of the redundant sensor is used as a substitute, and the data source is marked in the quality label; If there are no redundant sensors, the time-series interpolation method is used to calculate the data values ​​at abnormal times; In cases where data is missing for an extended period, estimates of the missing parameters are inferred using measurements from other sensors based on the physical model. Perform a mapping process on the repaired data, and annotate the data source and confidence level in the mapping results; The frequency of data anomalies from each sensor is counted, and a sensor health warning message is generated when the anomaly rate exceeds a preset health warning threshold.

8. A dynamic data mapping method for digital twins according to claim 3, characterized in that, The length of the selected time window is set according to the complete processing cycle of the equipment. The key parameters of the operating mode include the spindle speed range, feed rate range, depth of cut, tool type and process type. The historical operating data is divided into roughing high-speed mode, roughing low-speed mode, finishing high-precision mode, drilling mode, tapping mode, no-load mode, fault mode and transition mode through clustering algorithm.

9. A dynamic data mapping method for digital twins according to claim 4, characterized in that, The multi-layer mapping structure includes a first-layer mapping, a second-layer mapping, and a third-layer mapping. The first-layer mapping maps the original feature set to intermediate parameters. The second-layer mapping maps the intermediate parameters and feature set to the next-level intermediate parameters. The third-layer mapping maps the intermediate parameters to the final model parameters. When the absolute value of the slope of the mapping error sequence is greater than the preset trend judgment threshold or an error mutation is detected, feature importance analysis and running pattern recognition are re-executed to update the mapping parameters.

10. A dynamic data mapping system for digital twins, characterized in that, A method for performing dynamic data mapping for digital twins as described in any one of claims 1-9, comprising: The data acquisition module collects real-time data streams from the multi-source sensors during the operation of the equipment and performs preprocessing to obtain a standardized data stream; The mapping learning module extracts samples and analyzes and identifies the current operating mode based on a standardized data stream. It dynamically selects feature subsets and mapping parameters according to the mode and performs dynamic mapping to obtain dynamic mapping results. The dimensionality reduction calculation module analyzes feature correlation and groups based on dynamic mapping results, constructs a mapping structure, and performs mapping calculations layer by layer to obtain hierarchical mapping results. The evolution tracking module calculates the mapping residuals based on the hierarchical mapping results and analyzes the error evolution trend, triggers the update process to re-identify the operating mode and update the mapping parameters, and obtains the optimized and updated mapping results. The error correction module statistically analyzes the mapping error characteristics based on the optimized and updated mapping results, performs error feature decomposition, and performs deviation compensation and smoothing to obtain the self-corrected mapping results. The robust processing module, based on the standardized data stream and the mapping results after error self-correction, detects data anomalies and marks them with quality labels, repairs abnormal data, performs mapping and marks the source, statistically analyzes sensor anomalies to generate maintenance warnings, and obtains anomaly-tolerant mapping results and sensor health management information.

Images