Device data extrapolation methods, apparatus, equipment, and media based on digital twins
By acquiring measured values and 3D models of power equipment for data extrapolation, correcting parameters, and performing logical verification, the long-term accuracy and reliability issues of digital twin models in power systems have been resolved, achieving high-fidelity dynamic tracking and ensuring the accuracy and reliability of results.
Patent Information
- Application Number
- CN202611174221.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-04
- Publication Date
- 2026-11-10
AI Technical Summary
In highly reliable power systems, how can we ensure the long-term accuracy and reliability of digital twin models?
By acquiring measured values and a 3D model of the target device, data extrapolation is performed to obtain extrapolated values and predicted events. Model parameters are then corrected, logical verification is conducted, and a numerical-logic dual closed-loop architecture is constructed to ensure the accuracy and reliability of the extrapolation results.
It achieves high-fidelity dynamic tracking of digital twin models in power systems, ensuring the accuracy and logical reliability of output results, and is suitable for industrial scenarios with high requirements for safety and reliability.
Smart Images

Figure CN123069863A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of digital twin technology, specifically relating to a method, apparatus, device, and medium for device data extrapolation based on digital twins. Background Technology
[0002] Digital twins are widely used for condition simulation and predictive maintenance of industrial equipment. In modern power systems, accurate monitoring and predictive maintenance of critical equipment are core requirements for ensuring the safe, stable, and economical operation of the power grid. Digital twin technology can establish multiphysics (electric-thermal-mechanical) models for critical equipment such as converter valves and transformers, and perform real-time simulations, thereby improving the reliability and economy of the equipment. In high-reliability power scenarios, ensuring the long-term accuracy and reliability of digital twin models remains a challenge. Summary of the Invention
[0003] The purpose of this invention is to provide a method, apparatus, device, and medium for device data extrapolation based on digital twins, which can improve the long-term extrapolation accuracy and reliability of extrapolation results of digital twin models.
[0004] In a first aspect, embodiments of the present invention provide a device data extrapolation method based on digital twins, the method comprising: Obtain the measured values of the target device data and the corresponding three-dimensional model of the target device; Based on the three-dimensional model corresponding to the target device, the data of the target device is extrapolated to obtain the extrapolated value and predicted event of the target device data; The deviation trend of the three-dimensional model is obtained based on the inferred value and the measured value of the target device data, and the parameters of the three-dimensional model are corrected according to the deviation trend. The data of the target device is extrapolated using the parameter-corrected 3D model to obtain candidate extrapolation results of the target device data. The predicted event is logically verified according to preset rules to obtain the logical compliance of the target device's data. Based on the logical conformity of the data of the target device, the target inference result of the data of the target device is determined from the candidate inference results.
[0005] Optionally, the target device is a power device, and the three-dimensional model is constructed using an electromagnetic transient model and a thermal field model. The heat source term of the thermal field model is determined based on the power loss of the electromagnetic transient model. The step of extrapolating the data of the target device based on the three-dimensional model corresponding to the target device to obtain the extrapolated values and predicted events of the target device data includes: The feature extraction network of the three-dimensional model obtains the heat source term correction factor corresponding to the thermal field model based on the data of the target device. The power loss of the electromagnetic transient model is updated by the heat source term correction factor to obtain the updated heat source term, and the heat source term is input into the thermal field model. The electromagnetic transient model and the thermal field model are used to extrapolate the data of the target device to obtain the extrapolated values and predicted events of the target device data.
[0006] Optionally, obtaining the deviation trend of the three-dimensional model based on the inferred value and the measured value of the target device data includes: Determine the residual between the extrapolated value and the measured value of the data of the target device; Variational mode decomposition is performed on the residuals to obtain the deviation trend of the three-dimensional model.
[0007] Optionally, performing variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model includes: The residual is subjected to variational mode decomposition to obtain multiple modal components; The modal components with a center frequency lower than a preset frequency threshold are used as the deviation trend; The step of correcting the parameters of the three-dimensional model based on the deviation trend includes: Offline sensitivity analysis is performed on the three-dimensional model to obtain the parameter residual sensitivity mapping, and the target parameters related to the deviation trend in the three-dimensional model are determined based on the parameter residual sensitivity mapping. The correction amount of the target parameter is obtained by an incremental update algorithm, and the target parameter of the three-dimensional model is corrected according to the correction amount.
[0008] Optionally, the step of performing logical verification on the predicted event according to preset rules to obtain the logical compliance of the target device's data includes: The white-box logic verification engine verifies the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules, thereby obtaining the logical compliance of the target device's data.
[0009] Optionally, the white-box logic verification engine is constructed using a fuzzy temporal Petri net, which includes the preset rules; the step of using the white-box logic verification engine to verify the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules, and obtaining the logical compliance of the target device's data, includes: The degree of violation of causal logic, temporal logic, and mutual exclusion logic of the predicted event is determined by the membership function; The logical compliance of the target device's data is obtained based on the degree of violation of predicted events that do not meet the preset rules.
[0010] Optionally, determining the target inference result of the target device's data from the candidate inference results based on the logical conformity of the target device's data includes: Determine the quality score of the target device's data and the stability score of the parameter-corrected 3D model; The overall confidence level of the target device's data is determined based on the logical compliance, the quality score, and the stability score. Based on the comprehensive confidence level, a hierarchical gating strategy is applied to the candidate inference results; the hierarchical gating strategy includes allowing, downgrading, or rolling back. Candidate inference results whose overall confidence level meets the preset confidence threshold are used as the target inference results for the data of the target device.
[0011] Secondly, embodiments of the present invention provide a device for device data extrapolation based on digital twins, the device comprising: The equipment data acquisition module is used to acquire the measured values of the target equipment data and the three-dimensional model corresponding to the target equipment; The equipment data extrapolation module is used to extrapolate the data of the target equipment based on the three-dimensional model corresponding to the target equipment, and obtain the extrapolated value and predicted event of the target equipment data; The model parameter correction module is used to obtain the deviation trend of the three-dimensional model based on the inferred value and the measured value of the target device data, and to correct the parameters of the three-dimensional model according to the deviation trend. The candidate result determination module is used to extrapolate the data of the target device using the parameter-corrected three-dimensional model to obtain candidate extrapolation results of the data of the target device; The event logic verification module is used to perform logical verification on the predicted event according to preset rules to obtain the logical compliance of the data of the target device. The deduction result determination module is used to determine the target deduction result of the target device's data from the candidate deduction results based on the logical conformity of the target device's data.
[0012] Optionally, the target device is a power device, and the three-dimensional model is constructed using an electromagnetic transient model and a thermal field model. The heat source term of the thermal field model is determined based on the power loss of the electromagnetic transient model. The device data extrapolation module includes: The correction factor determination submodule is used to obtain the heat source term correction factor corresponding to the thermal field model based on the data of the target device through the feature extraction network of the three-dimensional model; The heat source term correction submodule is used to update the power loss of the electromagnetic transient model through the heat source term correction factor to obtain the updated heat source term, and input the heat source term into the thermal field model. The equipment data extrapolation submodule is used to extrapolate the data of the target equipment using the electromagnetic transient model and the thermal field model, and to obtain the extrapolated values and predicted events of the target equipment data.
[0013] Optionally, the model parameter correction module includes: The residual determination submodule is used to determine the residual between the extrapolated value and the measured value of the data of the target device. The deviation trend determination submodule is used to perform variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model.
[0014] Optionally, the deviation trend determination submodule includes: A mode decomposition unit is used to perform variational mode decomposition on the residual to obtain multiple mode components; A deviation trend determination unit is used to identify modal components with center frequencies lower than a preset frequency threshold as the deviation trend. The model parameter correction module further includes: The target parameter determination submodule is used to perform offline sensitivity analysis on the three-dimensional model, obtain parameter residual sensitivity mapping, and determine the target parameters in the three-dimensional model that are related to the deviation trend based on the parameter residual sensitivity mapping. The model parameter correction submodule is used to obtain the correction amount of the target parameter through an incremental update algorithm, and to correct the target parameter of the three-dimensional model according to the correction amount.
[0015] Optionally, the event logic verification module includes: The event logic verification submodule is used to verify the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules through a white-box logic verification engine, so as to obtain the logical compliance of the target device's data.
[0016] Optionally, the white-box logic verification engine is constructed using a fuzzy temporal Petri net, which includes the preset rules; the event logic verification submodule includes: The violation determination unit is used to determine the degree of violation of the causal logic, temporal logic, and mutual exclusion logic of the predicted event through a membership function; The logical compliance determination unit is used to obtain the logical compliance of the target device's data based on the degree of violation of the predicted events that do not meet the preset rules.
[0017] Optionally, the deduction result determination module includes: The stability score determination submodule is used to determine the quality score of the target device's data and the stability score of the parameter-corrected 3D model; The comprehensive confidence level determination submodule is used to determine the comprehensive confidence level of the target device's data based on the logical compliance, the quality score, and the stability score. The hierarchical gating strategy execution submodule is used to execute a hierarchical gating strategy on the candidate inference results based on the comprehensive confidence level; the hierarchical gating strategy includes allowing, downgrading, or back off. The deduction result determination submodule is used to take the candidate deduction results whose comprehensive confidence level meets the preset confidence threshold as the target deduction results of the target device's data.
[0018] Thirdly, embodiments of the present invention provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.
[0019] Fourthly, embodiments of the present invention provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.
[0020] The embodiments of the present invention have the following advantages: This invention can acquire measured values of data from a target device and a corresponding 3D model of the target device. The measured data allows for the establishment of a mapping between physical entities and digital space, while the 3D model provides a grid foundation for multiphysics calculations (electromagnetism, heat, and force). Based on the 3D model, the target device's data is extrapolated to obtain extrapolated values and predicted events. Extrapolation using the 3D model captures nonlinear physical processes within the device and outputs predicted events to identify potential risks in advance. The extrapolated and measured values of the target device's data reveal the deviation trend of the 3D model, and the model's parameters are corrected accordingly. The model parameters are then corrected in reverse based on the deviation trend, enabling the digital model to dynamically follow changes in the physical device's health status and maintain high fidelity over the long term. Extrapolation of the target device's data using the parameter-corrected 3D model yields candidate extrapolation results. Further extrapolation based on the corrected parameters of the 3D model ensures that the output candidate results eliminate current systematic biases. Logical verification of predicted events is performed according to preset rules to obtain the logical compliance of the target device's data. Through white-box verification, qualitative logical rules are transformed into quantitative logical compliance indicators to determine the logical credibility of the current inference result. Based on the logical compliance of the target device's data, the target inference result for the target device's data is determined from the candidate inference results. A numerical-logic dual closed-loop architecture is constructed, ensuring both data accuracy and logical credibility in the final determined target inference result. This approach is particularly suitable for industrial scenarios with high security and reliability requirements, such as power systems. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating the steps of a device data extrapolation method based on digital twins according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the steps of another device data extrapolation method based on digital twins according to an embodiment of the present invention. Figure 3 This is a logic diagram of a device data extrapolation method based on digital twins according to an embodiment of the present invention; Figure 4 This is a structural block diagram of a device data extrapolation apparatus based on digital twins according to an embodiment of the present invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0024] The device data extrapolation method based on digital twins provided in this invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.
[0025] Reference Figure 1 The diagram illustrates a flowchart of a device data extrapolation method based on digital twins according to an embodiment of the present invention. The method may specifically include the following steps: Step 101: Obtain the measured values of the target device data and the corresponding 3D model of the target device; In this embodiment of the invention, a data foundation between the physical entity and the digital space can be established first, namely, acquiring the measured data values of the target device and constructing the corresponding three-dimensional model of the device. The system can access data from various sensors deployed on the target power equipment (such as the power switch module in a flexible DC converter valve) and operational data from the station-level monitoring system in real time through a high-precision data acquisition and communication interface, thereby obtaining measured data values that reflect the current real state of the equipment.
[0026] The construction of the 3D model can be a numerical simulation model that carries a physical computational grid. The system can load a pre-built 3D geometric model for the target device. This 3D model is geometrically consistent with the physical entity of the target device and is internally discretized into a computational grid suitable for numerical solutions. This 3D model also carries the material physical properties of each component of the target device and the definitions of contact and boundary conditions between different components, thus serving as the grid basis for multiphysics numerical computation.
[0027] The system can use the measured values at the current moment as boundary and initial conditions to assign to the 3D model, completing the first state synchronization from physical entity to digital twin, thus laying a solid data foundation for subsequent high-fidelity deduction, online parameter correction, logic verification, and numerical logic dual closed-loop collaboration.
[0028] Step 102: Based on the 3D model corresponding to the target device, extrapolate the data of the target device to obtain the extrapolated values and predicted events of the target device data; In this embodiment of the invention, after acquiring measured values and initializing the 3D model, the system can perform real-time deduction of the target device's data based on the 3D model corresponding to the target device, thereby obtaining the deduced values of key physical quantities of the target device and the predicted events generated based on the deduction results. Specifically, the system can use the measured data of the target device acquired at the current moment as input boundary conditions and load it onto a pre-constructed 3D model. This 3D model is geometrically consistent with the actual device and has been discretized into a computational grid suitable for numerical solutions. Driven by these input data, the system simultaneously solves the governing equations of different physical fields within the spatial range defined by the 3D model. During the solution process, the various material properties and boundary conditions pre-assigned in the 3D model jointly determine the evolution law of the physical field in space and time. The system advances the calculation according to a preset time step, updating the physical quantity values at each node of the 3D model within each time step. Through this deduction method, the system can simulate the evolution process of the physical field inside the target device from the current state to the future state, thereby obtaining the physical quantity values at each key location in the future as deduced values.
[0029] After acquiring a continuous sequence of physical quantity projections, the system can perform discretized event identification processing on these continuous values. The system monitors key physical quantities in the projection sequence in real time, checking whether these quantities reach or exceed preset threshold conditions, or whether they exhibit a trend that satisfies a specific pattern. When a predefined event trigger condition is met at a certain moment in the projection sequence, the system generates a corresponding predicted event at that moment. Each predicted event records the event type and the timestamp of its occurrence. In this way, the system transforms the originally continuous physical quantity projections into discrete event information with clear operational semantics, enabling operations personnel to intuitively understand the future state changes predicted by the system.
[0030] Following the aforementioned deduction and event generation process, the system simultaneously obtains two types of output results. The first type is the continuous or discrete time series of key physical quantities of the target equipment within the deduction time domain, i.e., the deduced values. These values quantitatively describe the expected physical state of the equipment at future moments. The second type is a set of predicted events generated based on the deduced value sequence. These events qualitatively summarize possible state changes or anomalies that may occur during the future operation of the equipment. Both types of output results are jointly passed to subsequent processing steps. The deduced values are used to compare with measured values to assess model deviation, while the predicted events are used for logical consistency verification.
[0031] Step 103: Based on the extrapolated and measured values of the target device data, obtain the deviation trend of the three-dimensional model, and perform parameter correction on the three-dimensional model according to the deviation trend; In this embodiment of the invention, the system can compare the projected and measured values of the target device data at the same time, calculate the difference between the two, and form a residual sequence. For each key physical quantity, the system subtracts the projected value from the measured value to obtain the residual at that time. The system organizes these residuals into a sequence in chronological order, which reflects the degree of deviation between the three-dimensional model projection results and the actual state of the physical device at each time.
[0032] The system analyzes and processes the constructed residual sequence to identify systematic patterns of change. The residual sequence may contain rapid, random fluctuations or slowly changing trend components. The system uses signal processing methods to decompose the residual sequence into components with different rates of change, extracting the slowest-changing component as the deviation trend. This deviation trend characterizes the unidirectional change direction of the difference between the projected and measured values over time, such as a gradual increase or decrease in deviation.
[0033] Based on the direction and magnitude of the extracted deviation trend, the system determines which parameters in the 3D model need correction. The direction of the deviation trend indicates the direction of the systematic deviation of the extrapolated values from the measured values, such as the extrapolated values being consistently higher or lower. According to the direction of influence of each model parameter on the extrapolation results, the system selects those parameters that, after adjustment, can bring the extrapolated values closer to the measured values as the objects to be corrected. The magnitude of the deviation trend is used to determine the urgency of the correction and the approximate range of the correction amount.
[0034] Based on the magnitude of the deviation trend and the current value of the parameter to be corrected, the system calculates the relative or absolute amount that the parameter needs to be adjusted. The goal of this correction is to reduce the difference between the 3D model's simulation results and the measured values after parameter adjustment. The system applies the calculated correction to the corresponding parameters of the 3D model, completing the online parameter update. The corrected parameters replace the original parameter values in the 3D model, enabling the model to better match the current actual physical characteristics of the target device and reduce residuals in subsequent simulations.
[0035] Step 104: The data of the target device is extrapolated using the parameter-corrected 3D model to obtain candidate extrapolation results of the target device data; In this embodiment of the invention, the system uses the updated 3D model, after parameter correction, as the basis for the physical calculations of this simulation. The relevant correctable parameters in the 3D model have been replaced with their corrected new values, while the model's geometry, mesh generation, and other uncorrected parameters remain unchanged. After loading, the 3D model is ready to receive input data and perform calculations.
[0036] The system can acquire the measured data of the target device at the current moment and input this measured data as boundary and initial conditions into the calibrated 3D model. This measured data includes, but is not limited to, electrical measurements, thermal measurements, and environmental parameters, which together determine the initial state and external excitation conditions of the simulation. The system ensures that the timestamps of the input data are aligned with the simulation start time to guarantee time consistency during the simulation process.
[0037] The system uses the loaded 3D model and input measured data as a basis, and proceeds with calculations according to a preset time step. During the simulation, each physical field in the 3D model evolves synchronously according to its own governing equations and the coupling relationships between them. At each time step, the system updates the physical quantity values at each node of the 3D model, using these as the initial state for the next time step. This process continues until the preset simulation time domain endpoint is reached.
[0038] The system organizes the key physical quantity values calculated throughout the entire simulation time domain in chronological order to form candidate simulation results. These candidate results reflect the expected physical state change trajectory of the target device over a future period starting from the current moment, driven by the parameter-corrected 3D model. Compared to the simulation values before parameter correction, the candidate simulation results have incorporated parameter correction information obtained based on historical deviations, thus representing an improved simulation output after numerical consistency adjustments. These candidate simulation results will be passed to subsequent steps to combine with logical verification results to ultimately determine whether they should be used as the official output.
[0039] Step 105: Perform logical verification on the predicted event according to preset rules to obtain the logical compliance of the target device's data; In this embodiment of the invention, the system can receive a sequence of predicted events generated from the multiphysics deduction step. This sequence records each discrete event identified by the system based on the deduced values of continuous physical quantities within the deduction time domain. Each event includes an event type and its corresponding timestamp information. These predicted events reflect potential state changes or anomalies that may occur during the future operation of the target device, as predicted by the 3D model.
[0040] The system loads a pre-defined rule base, which stores several preset rules used to verify the logical consistency of predicted events. These rules are derived from the physical laws, operating procedures, and working principles of the target device's domain. Each rule defines the causal, temporal, or mutual exclusion constraints that events should satisfy. The wording of the rules enables the system to determine whether there are any violations of relevant constraints based on the sequence of predicted events.
[0041] The system takes the predicted event sequence as input and calls rules from a pre-defined rule base for logical verification, one by one. For each rule, the system checks whether there are combinations of events in the predicted event sequence that satisfy the rule's preconditions. If so, it further checks whether these events satisfy the constraints defined by the rule. For example, for a rule describing a causal relationship, the system checks whether the event that is the cause is observed as the result within the expected time window after the cause event occurs. For a rule describing a mutually exclusive relationship, the system checks whether two events that should not occur simultaneously appear in the predicted event sequence.
[0042] Based on the verification results of each rule, the system calculates an overall logical compliance score for the predicted event sequence generated in this simulation. Specifically, for each triggered rule, the system evaluates the degree of compliance between the predicted event sequence and the rule's constraints. If the constraints between events are fully satisfied, the compliance score is the highest; if the constraints are partially or completely violated, the compliance score decreases accordingly. The system synthesizes the compliance scores of all triggered rules and obtains a quantitative index between the lowest and highest values through weighted averaging or other fusion methods. This index is the logical compliance score. The logical compliance score represents the extent to which the predicted event sequence conforms to the preset physical laws and operating procedures. A higher logical compliance score indicates a higher logical credibility of the simulation results, while a lower score indicates that there are violations of logical constraints in the simulation results.
[0043] Step 106: Based on the logical conformity of the target device's data, determine the target inference result of the target device's data from the candidate inference results.
[0044] In this embodiment of the invention, the system can receive candidate deduction results obtained from the parameter-corrected 3D model deduction, and simultaneously receive the logical compliance degree of the target device data calculated from the logical verification step. The candidate deduction results describe the expected physical state change trajectory of the target device in the future time domain in the form of a numerical sequence, while the logical compliance degree reflects, in the form of a quantitative index, the extent to which the predicted event sequence corresponding to the candidate deduction result conforms to the preset physical laws and operating procedures.
[0045] The system uses the calculated logical compliance score as the basis for screening candidate prediction results. Specifically, it compares the logical compliance score with a preset threshold to determine whether the predicted event sequence corresponding to the current candidate prediction result violates logical constraints. If the logical compliance score is higher than or equal to the preset threshold, it indicates that the candidate prediction result has good logical consistency, and its prediction values are consistent with physical laws and operating procedures in terms of causality, temporal sequence, and mutual exclusion constraints. If the logical compliance score is lower than the preset threshold, it indicates that the candidate prediction result violates logical constraints, and although its prediction values may meet the expected range at the numerical level, there are contradictions in the logical relationships between events.
[0046] Based on the above screening results, the system can ultimately determine the target deduction result from the candidate deduction results. When a candidate deduction result passes the logical compliance screening, the system directly determines that candidate deduction result as the target deduction result. This result has undergone both numerical consistency adjustment in the parameter correction stage and credibility verification in the logical verification stage. When a candidate deduction result fails the logical compliance screening, the system executes a preset substitution strategy, such as reverting to the deduction result before parameter correction as the target deduction result, or discarding the current candidate deduction result and waiting for the next deduction cycle to generate a new candidate result. Through this decision-making mechanism based on logical compliance, the target deduction result output by the system is guaranteed in both numerical accuracy and logical credibility.
[0047] In this embodiment of the invention, the measured values of data from a target device and a corresponding 3D model of the target device can be obtained. The measured data values allow for the establishment of a mapping between physical entities and digital space, while the 3D model provides a grid foundation for calculations of multiple physics fields (electromagnetism, heat, and force). Based on the 3D model, the data of the target device is extrapolated to obtain extrapolated values and predicted events. Extrapolation using the 3D model captures nonlinear physical processes within the device and outputs predicted events to identify potential risks in advance. The deviation trend of the 3D model is obtained based on the extrapolated and measured values of the target device data, and the parameters of the 3D model are corrected according to this trend. The model parameters are then corrected in reverse based on the deviation trend, enabling the digital model to dynamically follow changes in the health status of the physical device and maintain high fidelity over a long period. Extrapolation of the target device data using the parameter-corrected 3D model yields candidate extrapolation results. Further extrapolation based on the corrected parameters of the 3D model ensures that the output candidate results eliminate the current systematic bias. Logical verification of predicted events is performed according to preset rules to obtain the logical compliance of the target device's data. Through white-box verification, qualitative logical rules are transformed into quantitative logical compliance indicators to determine the logical credibility of the current inference result. Based on the logical compliance of the target device's data, the target inference result for the target device's data is determined from the candidate inference results. A numerical-logic dual closed-loop architecture is constructed, ensuring both data accuracy and logical credibility in the final determined target inference result. This approach is particularly suitable for industrial scenarios with high security and reliability requirements, such as power systems.
[0048] Reference Figure 2 The diagram illustrates a flowchart of another method for device data extrapolation based on digital twins according to an embodiment of the present invention. The method may specifically include the following steps: Step 201: Obtain the measured values of the target device data and the corresponding 3D model of the target device; Step 202: Based on the 3D model corresponding to the target device, extrapolate the data of the target device to obtain the extrapolated values and predicted events of the target device data; In one embodiment, the target device is an electrical device, and the three-dimensional model is constructed using an electromagnetic transient model and a thermal field model. The heat source term of the thermal field model is determined based on the power loss of the electromagnetic transient model. The step of extrapolating the data of the target device based on the three-dimensional model corresponding to the target device to obtain the extrapolated value of the target device data and the predicted event may further include the following sub-steps: Sub-step S11: Based on the data of the target device, the feature extraction network of the three-dimensional model obtains the heat source term correction factor corresponding to the thermal field model. Sub-step S12 involves updating the power loss of the electromagnetic transient model using the heat source term correction factor to obtain the updated heat source term, and then inputting the heat source term into the thermal field model. Sub-step S13 involves using the electromagnetic transient model and the thermal field model to extrapolate the data of the target device, thereby obtaining the extrapolated values and predicted events of the target device's data.
[0049] In this embodiment of the invention, the target device can be an electrical device. The three-dimensional model corresponding to the electrical device is not a single physical field model, but a multi-physics coupled model jointly constructed by an electromagnetic transient model and a thermal field model. The electromagnetic transient model describes the transient changes in electrical quantities such as voltage and current within the device, while the thermal field model describes the internal temperature distribution and its evolution over time. There is a clear coupling relationship between the two models. Specifically, the heat source term required by the thermal field model is determined based on the power loss calculated by the electromagnetic transient model; that is, the power loss output by the electromagnetic transient model is used as a heat source input to the thermal field model to drive the calculation of the temperature field.
[0050] During the simulation based on the 3D model, the system first invokes a pre-built feature extraction network within the 3D model. This feature extraction network takes multi-source data from the target device as input and performs feature extraction and transformation on this data. After processing, the feature extraction network outputs a correction factor for the heat source terms of the thermal field model. This correction factor is a numerical value that adjusts the heat source terms used in subsequent steps; its specific range is determined based on the design of the feature extraction network.
[0051] The system applies a heat source term correction factor to the power loss calculated by the electromagnetic transient model. Specifically, the system uses the power loss output by the electromagnetic transient model as a base value, and applies the correction factor to update this base value, for example, by multiplying or adding the base value with the correction factor, thus obtaining the updated heat source term. After the update is completed, the system uses the updated heat source term as input to the thermal field model, replacing the original uncorrected power loss.
[0052] After updating the heat source terms, the system can initiate a joint simulation of the electromagnetic transient model and the thermal field model. The electromagnetic transient model solves the electromagnetic field equations under given boundary and initial conditions, outputting the time-series changes in electrical quantities and the corresponding power loss. This power loss, updated with a correction factor, is then passed to the thermal field model. Driven by the updated heat source terms, the thermal field model solves the heat conduction equations, outputting the time-series changes in the temperature field. Through the collaborative calculation between the two models, the system ultimately obtains the simulated values of the key physical quantities of the target device. Simultaneously, based on the simulated value sequence, it generates corresponding predicted events according to preset rules, completing this simulation process.
[0053] Step 203: Determine the residual between the extrapolated value and the measured value of the data of the target device; In this embodiment of the invention, the system can compare the projected and measured values of the target device data at the same time, calculate the difference between the two, and thus determine the residual. Specifically, the system can subtract the projected value from the measured value to obtain the residual value at that time. The residual reflects the degree of deviation between the 3D model projection result and the actual state of the physical device at the current time. A positive residual indicates that the projected value is lower than the measured value, and a negative residual indicates that the projected value is higher than the measured value. The system organizes these residuals into a residual sequence in chronological order, which serves as the basic input for subsequent deviation trend extraction.
[0054] Step 204: Perform variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model; In one embodiment, the step of performing variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model may further include the following sub-steps: Sub-step S21: Perform variational mode decomposition on the residual to obtain multiple mode components; Sub-step S22, the modal components with center frequencies lower than a preset frequency threshold are taken as the deviation trend.
[0055] In this embodiment of the invention, the system can input the constructed residual sequence into the variational mode decomposition module. Variational mode decomposition is an adaptive signal decomposition method. Its core idea is to decompose the original signal into several intrinsic mode components with specific sparse bandwidths without pre-specifying basis functions. The system sets the number of decomposition modes K and sets a penalty factor to control the bandwidth of each mode component. By iteratively solving the variational optimization problem using the alternating direction multiplier method, the residual sequence is decomposed into K mode components arranged from high frequency to low frequency. Each mode component has concentrated frequency band energy near its center frequency, and different mode components are separated from each other in the frequency domain. After decomposition, the system obtains a set of mode components that can characterize the different frequency components in the residual sequence.
[0056] After completing variational mode decomposition and obtaining multiple modal components, the system analyzes the frequency characteristics of these K modal components. Specifically, the system calculates the center frequency of each modal component, which reflects the concentration position of the component in the frequency domain. Different modal components have different center frequencies, typically arranged from high to low according to the decomposition order. The system obtains a preset frequency threshold, which is pre-set based on the physical characteristics of the target device and the time scale of model drift. The system compares the center frequency of each modal component with the preset frequency threshold, identifying modal components with center frequencies lower than the preset frequency threshold.
[0057] These low-frequency modal components collectively constitute the bias trend of the 3D model. This bias trend characterizes the slowest-changing component in the residual sequence, reflecting the cumulative change direction of the difference between the projected and measured values over time—for example, whether the bias gradually increases or decreases, and the rate of this change. In contrast to the bias trend are the high-frequency modal components, which mainly reflect random noise or periodic fluctuations in the residual sequence. Through this frequency threshold-based screening method, the system can effectively separate systematic biases from random disturbances, providing targeted bias information for subsequent parameter correction.
[0058] Step 205: Perform parameter correction on the 3D model based on the deviation trend; In one embodiment, the step of correcting the parameters of the three-dimensional model based on the deviation trend may further include the following sub-steps: Sub-step S31: Perform offline sensitivity analysis on the three-dimensional model to obtain parameter residual sensitivity mapping, and determine the target parameters in the three-dimensional model related to the deviation trend based on the parameter residual sensitivity mapping; Sub-step S32: Obtain the correction amount of the target parameter through an incremental update algorithm, and correct the target parameter of the three-dimensional model according to the correction amount.
[0059] In this embodiment of the invention, after extracting the deviation trend of the 3D model through variational mode decomposition, the system can enter the parameter correction stage. This deviation trend reflects the direction and magnitude of the systematic deviation between the extrapolated and measured values accumulated over time. Its root cause is usually that certain key parameters in the 3D model have actually changed due to equipment aging, environmental changes, etc., while the parameter values in the model have not been updated accordingly. Therefore, based on this deviation trend, the system performs targeted online correction on the relevant parameters in the 3D model, enabling the corrected 3D model to better match the current physical characteristics of the target equipment, thereby reducing residuals in subsequent extrapolations.
[0060] The system first performs an offline sensitivity analysis, which can be completed before the system is put into online operation. Specifically, the system performs sensitivity tests on each correctable parameter in the 3D model. For each candidate parameter, the system applies a small perturbation to its nominal value, reruns the multiphysics coupling simulation, and observes the change in key output quantities (such as steady-state junction temperature) relative to the parameter perturbation. The change in output quantity is divided by the input perturbation quantity to obtain the sensitivity coefficient of the parameter. This sensitivity coefficient reflects the degree of influence of a unit change in the parameter on the simulation result; its sign indicates the direction of influence (increasing the parameter leads to an increase or decrease in the simulation result), and its absolute value indicates the intensity of influence. The system summarizes the sensitivity coefficients of all candidate parameters to construct a parameter residual sensitivity mapping. During the online operation phase, after extracting the deviation trend, the system queries the parameter residual sensitivity mapping and, combining the direction and magnitude of the deviation trend, selects the parameter most relevant to the current deviation trend as the target parameter. The selection criteria include whether the sign of the sensitivity coefficient matches the direction of the deviation trend and whether the absolute value of the sensitivity coefficient is sufficiently significant.
[0061] After determining the target parameter to be corrected, the system can use an incremental update algorithm to calculate the required correction amount for that target parameter. The core idea of the incremental update algorithm is to use the parameter value from the previous moment as a benchmark, and based on the current deviation trend and the sensitivity coefficient of the target parameter, calculate the relative or absolute amount of adjustment needed, ensuring that the direction of change in the model output after parameter adjustment precisely offsets the observed deviation trend. The calculated correction amount is added incrementally to the current value of the target parameter, rather than directly replacing it with a completely new value. This incremental update method makes the parameter adjustment process smoother and more stable. After calculating the correction amount, the system applies this correction amount to the target parameter of the 3D model, performing online correction of the target parameter, i.e., replacing the original parameter value with the updated parameter value. After correction, the target parameter in the 3D model reflects the current optimal estimate incorporating historical deviation information, and the model is ready for subsequent extrapolation calculations.
[0062] Step 206: The data of the target device is extrapolated using the parameter-corrected 3D model to obtain candidate extrapolation results of the target device data; In this embodiment of the invention, the system can load preset future load conditions, environmental parameters, or operating instructions as input variables into a parameter-corrected 3D model. Since the model already incorporates correction parameters reflecting the current health status of the equipment, its calculation results are no longer theoretical predictions based on ideal design values, but rather precise extrapolations based on the actual aging degree and current physical characteristics of the equipment. For example, in the electromagnetic and thermal field coupling simulation of power equipment, the corrected model uses updated power loss as a heat source term, thereby more accurately calculating the future temperature rise trend. This extrapolation process can generate multi-dimensional data including temperature field distribution, stress change curves, or failure probabilities, which constitute the basic set of "candidate extrapolation results."
[0063] Step 207: The white-box logic verification engine verifies the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules to obtain the logical compliance of the target device's data. In one embodiment, the white-box logic verification engine is constructed using a fuzzy temporal Petri net, which includes the preset rules. The step of verifying the causal logic, temporal logic, and mutual exclusion logic of the predicted event using the white-box logic verification engine according to the preset rules to obtain the logical compliance of the target device's data may further include the following sub-steps: Sub-step S41: Determine the degree of violation of causal logic, temporal logic, and mutual exclusion logic of the predicted event through the membership function; Sub-step S42: Based on the degree of violation of the predicted event that does not meet the preset rule, obtain the logical compliance degree of the data of the target device.
[0064] In this embodiment of the invention, the system can input the predicted event sequence into a pre-built white-box logic verification engine. The characteristic of this white-box logic verification engine is that its internal rules and reasoning process are completely visible to the outside world, rather than being an uninterpretable black-box model. The engine internally pre-sets several rules for verifying the logical consistency of the predicted events. These rules cover causal logic, temporal logic, and mutual exclusion logic during the operation of the target device. Causal logic is used to verify whether the causal relationship between events holds, temporal logic is used to verify whether the time sequence of events occurs as expected, and mutual exclusion logic is used to verify whether events that should not occur simultaneously do so. The engine applies these pre-set rules one by one to verify the predicted event sequence, ultimately outputting a quantified logical compliance index.
[0065] The white-box logic verification engine first identifies combinations of events in the predicted event sequence that satisfy a preset rule trigger condition. For each triggered rule, the engine does not use a binary "either fully satisfied or not satisfied" approach, but instead uses a membership function to characterize the degree of conformity between the predicted event sequence and the rule requirement. The membership function maps continuous variables such as the time interval between events and the exact time of event occurrence to a membership value between 0 and 1, reflecting the degree to which the constraint is satisfied. For example, for a timing rule requiring event B to occur within a certain time window after event A, the membership function can output a continuous value based on how close the actual time interval is to the expected time window. The engine converts the value output by the membership function into a violation score, typically defined as 1 minus the membership score; the higher the violation score, the closer it is to 1. When the rule is fully satisfied, the violation score is 0. In this way, the engine can provide a quantitative violation score assessment for each triggered rule, thereby finely characterizing the degree of deviation between the predicted event sequence and the ideal logic constraint.
[0066] After calculating the violation rates of all triggered rules, the engine enters the logical compliance synthesis stage. For each predicted event that does not meet the preset rules, the engine has obtained the corresponding violation rate. The engine aggregates these violation rates according to preset weights, where the weights reflect the importance of different rules in the overall logical consistency evaluation. The aggregation method can be weighted average, weighted summation, or other preset fusion functions. After aggregation calculation, the engine obtains an overall quantitative index, which is the logical compliance of the target device data. The logical compliance value is usually set between 0 and 1, where a higher value indicates that the predicted event sequence generally meets the preset rules well, and the logical consistency is good; a lower value indicates that there are more violations in the predicted event sequence, and the logical confidence is low. This logical compliance will be passed to subsequent steps to determine the final target inference result from the candidate inference results.
[0067] Step 208: Based on the logical conformity of the target device's data, determine the target inference result of the target device's data from the candidate inference results.
[0068] In one embodiment, the step of determining the target inference result of the target device's data from the candidate inference results based on the logical conformity of the target device's data may further include the following sub-steps: Sub-step S51: Determine the quality score of the target device's data and the stability score of the parameter-corrected 3D model; Sub-step S52: Determine the overall confidence level of the target device's data based on the logical compliance, the quality score, and the stability score; Sub-step S53: Based on the comprehensive confidence level, a hierarchical gating strategy is executed on the candidate inference results; the hierarchical gating strategy includes allowing, downgrading, or back off. Sub-step S54: The candidate inference results whose comprehensive confidence level meets the preset confidence threshold are used as the target inference results of the data of the target device.
[0069] In this embodiment of the invention, after obtaining the candidate deduction results and their corresponding logical conformity scores, the system enters the final output decision stage. Logical conformity reflects the extent to which the predicted event sequence conforms to preset physical laws and operating procedures; however, logical conformity alone is insufficient to comprehensively evaluate the reliability of the candidate deduction results. Therefore, the system integrates multiple evaluation metrics to perform an overall assessment of the candidate deduction results and decides whether to output the candidate deduction result as the final target deduction result based on the assessment results.
[0070] The system first identifies two auxiliary evaluation metrics. The first is the quality score of the target device data, calculated based on the assessment results of the completeness, accuracy, and timeliness of the input data during the data access phase, reflecting the credibility of the data source driving the current simulation. The second is the stability score of the parameter-corrected 3D model, reflecting the stability of the model output after parameter correction. This can be characterized, for example, by analyzing the energy proportion of noise modes after residual decomposition; lower noise energy indicates a more stable model output. The system uses these two metrics, along with the previously obtained logical compliance score, as input for subsequent confidence level calculations.
[0071] The system integrates three indicators—logical compliance, quality score, and stability score—to calculate a comprehensive confidence level. The integration can be achieved through weighted summation, where each indicator corresponds to a preset weight coefficient. The magnitude of the weight coefficient reflects the importance of that indicator in the overall confidence evaluation. After the integration calculation, the system outputs a comprehensive confidence level value between the lowest and highest values. This value comprehensively reflects the overall credibility of the candidate inference results at both the numerical level (reflected by the quality and stability scores) and the logical level (reflected by the logical compliance).
[0072] The system compares the calculated overall confidence level with multiple preset thresholds and executes a hierarchical gating strategy based on the comparison results. This hierarchical gating strategy includes three levels: allow, downgrade, and rollback. When the overall confidence level is higher than or equal to the first threshold, the system executes the allow strategy, indicating that the candidate inference result has high credibility and can be output normally. When the overall confidence level is between the second and first thresholds, the system executes the downgrade strategy, indicating that the candidate inference result has some uncertainty and can be output, but with an alarm message or marked as low confidence. When the overall confidence level is lower than the second threshold, the system executes the rollback strategy, indicating that the candidate inference result has too low credibility and should not be output. The system then executes a preset alternative, such as outputting the inference result before parameter correction or waiting for the next inference cycle to regenerate candidate results.
[0073] During the execution of the aforementioned hierarchical gating strategy, the system filters out candidate inference results whose overall confidence level meets a preset confidence threshold. Specifically, when the overall confidence level reaches or exceeds the threshold corresponding to the release strategy, the candidate inference result is considered to meet the output conditions. The system determines this candidate inference result as the final target inference result for output display, storage recording, or triggering subsequent operation and maintenance decisions. For candidate results implementing a degradation strategy, the system can also output them as target inference results after annotation, but with corresponding confidence level prompts. For candidate results implementing a rollback strategy, the system does not output them as target inference results and records the rollback event for operation and maintenance personnel to review. Through this hierarchical gating mechanism based on overall confidence level, the system ensures that the final output target inference result meets the preset confidence standards in multiple dimensions.
[0074] In this embodiment of the invention, by calculating the residual between the inferred and measured values, and using variational mode decomposition technology to extract modal components with center frequencies below a preset threshold as "deviation trends," high-frequency random noise can be effectively filtered out, and low-frequency long-term drift caused by changes in the physical characteristics of the equipment (such as insulation aging and mechanical wear) can be accurately captured. By combining offline sensitivity analysis to construct a "parameter residual sensitivity mapping," the system can intelligently identify which model parameters are most sensitive to the current deviation and perform incremental updates only for these "target parameters." This not only improves the targeting of the correction but also avoids destroying the physical basis of the model due to overfitting noise, ensuring that the digital twin can "co-evolve" with the physical equipment throughout its entire lifecycle. A white-box logic verification engine was introduced, which uses fuzzy temporal Petri nets to construct preset rules. It performs triple verification on predicted events based on causal logic (whether cause and effect hold), temporal logic (whether the order of events is reasonable), and mutual exclusion logic (whether contradictory states exist). Membership functions quantify the degree of violation of these logics, thereby calculating the "logic compliance degree." This ensures that the inference results are not only numerically accurate but also reasonable in terms of physical mechanisms and business logic, greatly improving the confidence of operations and maintenance personnel in system alarms and predictions. A comprehensive confidence evaluation system was constructed, including logical compliance degree, data quality score, and model stability score. Based on the comprehensive confidence degree, a hierarchical gating strategy is implemented, effectively preventing misjudgments caused by sensor failure, model inaccuracy, or abnormal operating conditions, providing a final safety line for the high reliability requirements of industrial sites.
[0075] Reference Figure 3 The diagram illustrates a logic diagram of a device data extrapolation method based on digital twins provided by an embodiment of the present invention. To enable those skilled in the art to better understand the embodiments of the present invention, the following explanation is provided... Figure 3 The embodiments of the present invention will be described below.
[0076] Step 301: Multi-source data access and data quality scoring In this embodiment of the invention, the accessed multi-source heterogeneous data may include: structured data, such as valve voltage, valve current, coolant inlet / outlet temperature, coolant flow rate, ambient temperature, power switch surface temperature, and junction temperature estimated from the previous moment; and unstructured data, such as infrared thermal images (used for surface temperature estimation and indirect junction temperature estimation).
[0077] The system can comprehensively score the completeness, accuracy, and timeliness of the input data according to the following formula: S_data = (w1 × completeness + w2 × accuracy + w3 × timeliness) / (w1 + w2 + w3) Where S_data is the overall score of the input data, and w1, w2, and w3 are the weights corresponding to completeness, accuracy, and timeliness, respectively. Example: w1=0.4, w2=0.4, w3=0.2.
[0078] Step 302, Feature Extraction and Heat Source Term Correction In this embodiment of the invention, the Transformer model can be used for feature extraction. The Transformer is a deep learning model based on a self-attention mechanism, which can capture long-range dependencies, and its attention weight matrix has a certain degree of interpretability. The Transformer model can be used as a fine-tuner for feature extraction and physical model input.
[0079] The input information for the Transformer model is as follows: a tensor of shape [B, L, D], where L = 60 (60-second window). After system startup, it enters a stable inference phase after accumulating L sampling points; thereafter, a sliding window is used to update the input sequence and output the results every second. For example, D can be 8 (including valve voltage, valve current, coolant inlet temperature, outlet temperature, flow rate, ambient temperature, IGBT surface temperature, and the junction temperature from the previous inference time step). The Transformer model can output information using the following formula: L_fine=T_j^{meas}-T_j^{sim}(Q_corrected)_2^2+λf_Q(h_L)_2^2,λ=0.01 The system comprises: a driving feature vector, where h_L ∈ R^{d_model} is the output of the last time step (d_model = 64 in this example); a heat source correction factor, which maps h_L to a heat source correction factor via a lightweight network f_Q; f_Q consists of two fully connected layers (64→32→1), activated by tanh and multiplied by a scaling factor of 0.3 (i.e., 0.3×tanh(·)), with an output range of [-0.3, 0.3], and Q_corrected = Q_base×(1+f_Q(h_L)); and interpretability evidence, which outputs an attention weight matrix A ∈ R^{L×L} to explain the historical moments and variable contributions of the model. The system can employ self-supervised pre-training for a masked reconstruction task (masking 15% of time steps) with a loss of MSE.
[0080] Step 303, Multiphysics Coupling Deduction In this embodiment of the invention, the bidirectional coupling between the electromagnetic transient model and the thermal field model can be constructed according to the following formula: dx / dt=A(T_j)x+B(T_j)u+E(T_j)w, ρc* T / t= ·(k T)+Q.
[0081] The parameters for multi-timescale scheduling are as follows: electromagnetic transient step size, 10–100 μs; thermal field step size, 1–60 s; external output frequency, 1 Hz. When the thermal field step size is >1 s, to maintain the external 1 Hz output, zero-order hold, linear / spline interpolation, or compensation for the output within the interval between adjacent thermal field updates can be used. The thermal field step size can be ≤1 s to naturally align with the 1 Hz output.
[0082] Finally, the output includes the extrapolated values of key physical quantities (such as junction temperature) and the predicted event sequence (such as over-temperature alarm, cooling flow boost, etc.). The predicted event sequence can be generated by discretizing a continuous time series according to threshold / rate of change / state machine rules, and the event type and event timestamp are recorded for each event.
[0083] Step 304, numerical consistency closed loop In this embodiment of the invention, VMD (Variational Mode Decomposition) trend modes can be used for targeted correction. VMD can decompose a complex signal into multiple modal components with specific frequency bandwidths, separating the trend modes representing systematic deviations from the residuals. The residual sequence can be constructed according to the following formula: r_T(t) = T_measured(t) - T_inferred(t).
[0084] The VMD decomposition process is as follows: Assuming the number of modes K=5, the penalty factor α=2000, the convergence threshold ε=1e-7 or the upper limit of iteration 500, then the frequency discrimination center frequency <0.01Hz is the trend mode, 0.01–0.1Hz is the periodic mode, and >0.1Hz is the noise mode.
[0085] VMD can output trend, period, noise mode, and noise mode energy percentage S_vmd (used for model stability evaluation). In this embodiment of the invention, only the trend mode is used to correct the model parameters to ensure consistency of the physical mechanism.
[0086] Offline sensitivity analysis (finite difference) is performed according to the following formula: S_i= T_j / θ_i≈[T_j(θ_i+Δθ_i)-T_j(θ_i)] / Δθ_i. Where Δθ_i is taken as 1% of the nominal value (i.e. θ_i×1%), so S_i can be equivalently expressed as °C / %, forming a "parameter-residual sensitivity mapping".
[0087] The mean trend deviation is calculated using the following formula (sliding window T_w = 5 minutes): A_trend = (1 / T_w) × ∫_{t-T_w}^{t}u_trend(τ)dτ.
[0088] The dominant parameter is identified according to the following formula (direction matching + maximum sensitivity): θ_dm=argmax_{θ_i}|S_i|×I(sign(S_i)=sign(A_trend)).
[0089] Incremental correction (driving force) is performed using the following formula: Δθ_dm(n) = η × (A_trend / S_dm) + β × Δθ_dm(n-1). Where Δθ_dm is the percentage correction (%) relative to the nominal value, and S_dm is the sensitivity (°C / %) of the corresponding parameter. For example, η = 0.01, β = 0.9.
[0090] The limit update is performed according to the following formula: θ_dm(n+1)=clip(θ_dm(n)×(1+Δθ_dm(n) / 100), θ_min, θ_max). Here, the variation range of R_th_jc is limited to 70%–130% of the nominal value. This 70%–130% limit can serve as a safety constraint for single-cycle / single-update updates. In multi-cycle continuous correction scenarios, cumulative progress can be achieved through cycle iteration. Alternatively, when the actual aging inference exceeds the upper limit, an alarm is triggered, and a safety mode such as backup simulation strategy / freeze update is switched.
[0091] In this embodiment of the invention, model parameter updates only apply to the next cycle of projection. The output results of the current cycle are still based on the parameter state before the update, and the gating strategy determines whether to allow, downgrade, or roll back. VMD can be replaced by models such as EMD / EEMD, STFT / wavelet packet decomposition, and trend-seasonal-residual decomposition (e.g., STL) that can separate systematic trends and avoid noise overfitting.
[0092] Step 305, Logical Consistency Closed Loop In this embodiment of the invention, FTPN (Fuzzy-Timing Petri Net) can be used for closed-loop verification of logical consistency. FTPN can be replaced by a state machine with time constraints, temporal logic (LTL / MTL) verification, rule engine + fuzzy inference, or other models capable of white-box verification of causal / temporal / mutual exclusion constraints and outputting violations of metric indicators. The FTPN rule verification process is as follows: Sampling is performed according to the following formula: E1={(I(t)-I(t-Δt_s)) / I(t-Δt_s)>δ_I}∧{I(t)>I(t-Δt_s)}. Where the sampling interval Δt_s=1s, and the current change rate threshold δ_I=5%.
[0093] The junction temperature rise event is observed using the following formula: E2={T_j(t)-T_j(t_E1)>δ_T}. Here, E2 represents the junction temperature rise event, with a temperature rise threshold δ_T=2°C within the time window [τ_min, τ_max]=[1s, 5s].
[0094] The violation degree is calculated using the following formula: V_k = 1 - μ(Δt). Wherein, a violation occurs if E2 does not occur within the window, and the triangular membership function parameter is: τ_pt = 2.5s; if E2 does not occur at all, then V_k = 1.
[0095] The logical penalty term P is synthesized according to the following formula: P=Σ_{k=1}^{K}(w_k×V_k) / Σ_{k=1}^{K}w_k. Here, the triggering rules can be weighted and averaged. Example weights: causal rule 1.0, state trigger 0.8, logical association 0.7, mutual exclusion 0.6.
[0096] The overall confidence score is calculated using the following formula: Confidence = a × S_data + b × (1 - S_vmd) + c × (1 - P). Where S_data represents data quality, S_vmd represents model stability, and P represents logical compliance. Example weights: a = 0.4, b = 0.3, c = 0.3. This formula can be replaced with multiplicative fusion, Bayesian confidence, or DS evidence theory models that incorporate data quality, model stability, and logical compliance for gating decisions.
[0097] The hierarchical gating output strategy is as follows: Allow, Confidence ≥ 0.8; Degrade, 0.5 ≤ Confidence < 0.8, for reference only and trigger an alarm; Rollback, Confidence < 0.5, discard the result and trigger the rollback mechanism to restore the parameters to their pre-correction state to prevent error accumulation. The thresholds and levels for allow / degrade / rollback in the gating strategy are adjustable. Rollback can be replaced with safety strategies that prevent the spread of error correction, such as "freezing parameters, downgrading the model, or switching to a backup model."
[0098] Step 306, final trusted output Only high-confidence results that have passed dual closed-loop verification will be output for display, analysis, or triggering control / operation decisions.
[0099] This invention uses the converter valve of a power converter station as an example for illustration: 1. Monitoring Target: The junction temperature of the IGBT module is a key indicator but cannot be directly measured. Infrared surface temperature is used, and the equivalent junction temperature is calculated for verification and loop closure. In one implementation, an equivalent thermal resistance model is used to convert the infrared measured IGBT module case temperature / surface temperature... Converted to equivalent junction temperature :
[0100] in The device power loss is estimated from valve voltage, valve current, and switching status. The junction-shell thermal resistance (can be a nominal value or obtained through online calibration). In another embodiment, an RC thermal network or a manufacturer's thermal model can be used. The equivalent junction temperature is used for residual construction and closed-loop correction, and the specific conversion model form is not limited.
[0101] 2. Data Access: Collect valve voltage / current from the station monitoring system / OWS; collect coolant temperature / flow rate from the valve cooling system; access temperature / alarm data and infrared images from the valve hall infrared thermal imager, etc. Refer to [TR-2] and [TR-3] for the above data sources and their access / security partition specifications; align the sampling and external output frequencies to 1Hz.
[0102] 3. Operation Process: The system receives multi-source data in real time. The Transformer extracts spatiotemporal features and corrects the heat source term, driving the electromagnetic-thermal coupling model to predict the junction temperature trend and event sequence for the next hour. The electrical inputs (valve voltage / valve current, etc.) required for the next hour prediction can be obtained using any of the following methods: 1) Operating condition maintenance assumption: The valve voltage and valve current measured at the current moment are used as constant inputs in the future time domain; 2) Planned operating condition driven: The valve voltage and valve current prediction sequence is generated by the future load curve or power command output by the dispatch / control system; 3) Short-term forecast-driven: Based on historical data of the last L seconds, short-term rolling forecasts of valve voltage / valve current are made and used as future inputs.
[0103] 4. Numerical closed-loop example: In the initial stage of operation, there was a systematic deviation of about 5°C between the estimated junction temperature and the equivalent measured junction temperature; VMD separated the trend mode from the residual and corrected the equivalent thermal resistance parameter. After about 10 minutes, the deviation was eliminated and the error was reduced to within ±2.3°C.
[0104] 5. Logic closed-loop example: Simulate a sensor malfunction causing a sequence of events such as "valve current decreases but junction temperature surges"; FTPN identifies a violation of causal logic, P≈1, the overall confidence level drops to about 0.2, triggering a rollback and alarm to avoid misjudgment.
[0105] 6. VMD Validation: 24-hour residual spectrum analysis showed that the trend mode energy was approximately 62.3% (0.001–0.01Hz), the periodic mode energy was approximately 28.1% (0.01–0.1Hz, of which 0.02–0.05Hz is the main energy concentration band; the energy proportions of 0.01–0.02Hz and 0.05–0.1Hz are relatively low and have been included in the periodic mode statistics), and the noise mode energy was approximately 9.6% (>0.1Hz), proving that systematic deviations can be separated from noise.
[0106] 7. Comparative Experiment (Traditional Correction): Kalman Filter: Convergence time 25 minutes, steady-state error ±4.2°C, noise +50% error amplification +35%; Full Residual Feedback: Convergence time 8 minutes, steady-state error ±5.8°C, noise +50% error amplification +120%, severe overfitting; This Invention: Convergence time approximately 10 minutes, steady-state error ±2.3°C, noise +50% error amplification +8%, no obvious overfitting.
[0107] 8. Comparison with Federated Learning Path (Simulation): In a 5-site drift difference scenario, the MAE of Federated Learning is about 5.1°C with a response time of 30–60 minutes; the MAE of this invention is about 2.3°C with a response time of <10 minutes; and this invention has a logic conflict interception function (>93%).
[0108] This invention can simultaneously include online parameter correction closed loop and logical consistency closed loop in the same inference system; it decomposes residual VMD and uses only trend modes for online model parameter correction (including sensitivity mapping, dominant parameter identification, incremental momentum update and limiting); it uses FTPN for rule triggering, violation calculation and synthesis of penalty term P; it fuses S_data, S_vmd and P to form a comprehensive confidence score and executes release / degradation / rollback; low confidence triggers parameter rollback to the pre-correction state (or equivalent safety recovery strategy); it outputs heat source term correction factor and attention weight interpretation, and does not use the direct output temperature of Transformer as black box prediction; it comprehensively scores the completeness, accuracy and timeliness of input data (S_data) and incorporates it into comprehensive confidence score calculation and gating decision; the electromagnetic transient model and the thermal field model are bidirectionally coupled and inferred using different time steps (e.g. electromagnetic 10-100μs, thermal field 1-60s), and the output frequency is unified. This ensures that the long-term operating error of the system can be stabilized within ±2.3°C, the logic conflict interception rate is >93%, the process is transparent, explainable, and traceable, and convergence occurs within 10 minutes after drift occurs, with a rollback mechanism to prevent the accumulation and spread of errors.
[0109] It should be noted that the device data extrapolation method based on digital twins provided in this embodiment of the invention can be executed by a device data extrapolation device based on digital twins, or by a control module within that device for executing the loading of the device data extrapolation method based on digital twins. This embodiment of the invention uses the execution of the loading of the device data extrapolation method based on digital twins by the device data extrapolation device as an example to illustrate the device data extrapolation method based on digital twins provided in this embodiment of the invention.
[0110] Reference Figure 4 The diagram illustrates a structural block diagram of a device data extrapolation apparatus based on digital twins provided in an embodiment of the present invention. The apparatus includes: The equipment data acquisition module 401 is used to acquire the measured values of the target equipment data and the three-dimensional model corresponding to the target equipment. The equipment data extrapolation module 402 is used to extrapolate the data of the target equipment based on the three-dimensional model corresponding to the target equipment, and obtain the extrapolated value and predicted event of the target equipment data; The model parameter correction module 403 is used to obtain the deviation trend of the three-dimensional model based on the inferred value and the measured value of the target device data, and to correct the parameters of the three-dimensional model based on the deviation trend. The candidate result determination module 404 is used to deduce the data of the target device through the parameter-corrected three-dimensional model to obtain the candidate deduction results of the data of the target device; The event logic verification module 405 is used to perform logical verification on the predicted event according to preset rules to obtain the logical compliance of the target device's data. The deduction result determination module 406 is used to determine the target deduction result of the target device data from the candidate deduction results based on the logical conformity of the target device data.
[0111] In this embodiment of the invention, the target device is a power device, and the three-dimensional model is constructed using an electromagnetic transient model and a thermal field model. The heat source term of the thermal field model is determined based on the power loss of the electromagnetic transient model. The device data extrapolation module 402 includes: The correction factor determination submodule is used to obtain the heat source term correction factor corresponding to the thermal field model based on the data of the target device through the feature extraction network of the three-dimensional model; The heat source term correction submodule is used to update the power loss of the electromagnetic transient model through the heat source term correction factor to obtain the updated heat source term, and input the heat source term into the thermal field model. The equipment data extrapolation submodule is used to extrapolate the data of the target equipment using the electromagnetic transient model and the thermal field model, and to obtain the extrapolated values and predicted events of the target equipment data.
[0112] In this embodiment of the invention, the model parameter correction module 403 includes: The residual determination submodule is used to determine the residual between the extrapolated value and the measured value of the data of the target device. The deviation trend determination submodule is used to perform variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model.
[0113] In this embodiment of the invention, the deviation trend determination submodule includes: A mode decomposition unit is used to perform variational mode decomposition on the residual to obtain multiple mode components; A deviation trend determination unit is used to identify modal components with center frequencies lower than a preset frequency threshold as the deviation trend. The model parameter correction module 403 further includes: The target parameter determination submodule is used to perform offline sensitivity analysis on the three-dimensional model, obtain parameter residual sensitivity mapping, and determine the target parameters in the three-dimensional model that are related to the deviation trend based on the parameter residual sensitivity mapping. The model parameter correction submodule is used to obtain the correction amount of the target parameter through an incremental update algorithm, and to correct the target parameter of the three-dimensional model according to the correction amount.
[0114] In this embodiment of the invention, the event logic verification module 405 includes: The event logic verification submodule is used to verify the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules through a white-box logic verification engine, so as to obtain the logical compliance of the target device's data.
[0115] In this embodiment of the invention, the white-box logic verification engine is constructed using a fuzzy temporal Petri net, which includes the preset rules; the event logic verification submodule includes: The violation determination unit is used to determine the degree of violation of the causal logic, temporal logic, and mutual exclusion logic of the predicted event through a membership function; The logical compliance determination unit is used to obtain the logical compliance of the target device's data based on the degree of violation of the predicted events that do not meet the preset rules.
[0116] In this embodiment of the invention, the deduction result determination module 406 includes: The stability score determination submodule is used to determine the quality score of the target device's data and the stability score of the parameter-corrected 3D model; The comprehensive confidence level determination submodule is used to determine the comprehensive confidence level of the target device's data based on the logical compliance, the quality score, and the stability score. The hierarchical gating strategy execution submodule is used to execute a hierarchical gating strategy on the candidate inference results based on the comprehensive confidence level; the hierarchical gating strategy includes allowing, downgrading, or back off. The deduction result determination submodule is used to take the candidate deduction results whose comprehensive confidence level meets the preset confidence threshold as the target deduction results of the target device's data.
[0117] This invention can acquire measured values of data from a target device and a corresponding 3D model of the target device. The measured data allows for the establishment of a mapping between physical entities and digital space, while the 3D model provides a grid foundation for multiphysics calculations (electromagnetism, heat, and force). Based on the 3D model, the target device's data is extrapolated to obtain extrapolated values and predicted events. Extrapolation using the 3D model captures nonlinear physical processes within the device and outputs predicted events to identify potential risks in advance. The extrapolated and measured values of the target device's data reveal the deviation trend of the 3D model, and the model's parameters are corrected accordingly. The model parameters are then corrected in reverse based on the deviation trend, enabling the digital model to dynamically follow changes in the physical device's health status and maintain high fidelity over the long term. Extrapolation of the target device's data using the parameter-corrected 3D model yields candidate extrapolation results. Further extrapolation based on the corrected parameters of the 3D model ensures that the output candidate results eliminate current systematic biases. Logical verification of predicted events is performed according to preset rules to obtain the logical compliance of the target device's data. Through white-box verification, qualitative logical rules are transformed into quantitative logical compliance indicators to determine the logical credibility of the current inference result. Based on the logical compliance of the target device's data, the target inference result for the target device's data is determined from the candidate inference results. A numerical-logic dual closed-loop architecture is constructed, ensuring both data accuracy and logical credibility in the final determined target inference result. This approach is particularly suitable for industrial scenarios with high security and reliability requirements, such as power systems.
[0118] The device data extrapolation apparatus based on digital twins in this embodiment of the invention can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. This embodiment of the invention does not impose specific limitations.
[0119] The device data extrapolation apparatus based on digital twins in this embodiment of the invention can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this embodiment of the invention does not impose specific limitations.
[0120] The device data extrapolation device based on digital twin provided in this embodiment of the invention can achieve Figures 1 to 3 The various processes implemented by the device data extrapolation device based on digital twins in the method embodiments will not be described again here to avoid repetition.
[0121] Optionally, embodiments of the present invention also provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When executed by the processor, the program or instructions implement the various processes of the above-described embodiments of the device data extrapolation method based on digital twins, and achieve the same technical effects. To avoid repetition, further details are omitted here. It should be noted that the electronic device in the embodiments of the present invention includes the aforementioned mobile electronic device and non-mobile electronic device.
[0122] This invention also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiments of the device data extrapolation method based on digital twins and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0123] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.
[0124] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
[0125] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.
Claims
1. A method for extrapolating device data based on digital twins, characterized in that, The method includes: Obtain the measured values of the target device data and the corresponding three-dimensional model of the target device; Based on the three-dimensional model corresponding to the target device, the data of the target device is extrapolated to obtain the extrapolated value and predicted event of the target device data; The deviation trend of the three-dimensional model is obtained based on the inferred value and the measured value of the target device data, and the parameters of the three-dimensional model are corrected according to the deviation trend. The data of the target device is extrapolated using the parameter-corrected 3D model to obtain candidate extrapolation results of the target device data. The predicted event is logically verified according to preset rules to obtain the logical compliance of the target device's data. Based on the logical conformity of the data of the target device, the target inference result of the data of the target device is determined from the candidate inference results.
2. The device data extrapolation method based on digital twins according to claim 1, characterized in that, The target device is a power device. The three-dimensional model is constructed using an electromagnetic transient model and a thermal field model. The heat source term of the thermal field model is determined based on the power loss of the electromagnetic transient model. The step of extrapolating the data of the target device based on the three-dimensional model corresponding to the target device to obtain the extrapolated values and predicted events of the target device data includes: The feature extraction network of the three-dimensional model obtains the heat source term correction factor corresponding to the thermal field model based on the data of the target device. The power loss of the electromagnetic transient model is updated by the heat source term correction factor to obtain the updated heat source term, and the heat source term is input into the thermal field model. The electromagnetic transient model and the thermal field model are used to extrapolate the data of the target device to obtain the extrapolated values and predicted events of the target device data.
3. The device data extrapolation method based on digital twins according to claim 1, characterized in that, The step of obtaining the deviation trend of the three-dimensional model based on the inferred value and the measured value of the target device data includes: Determine the residual between the extrapolated value and the measured value of the data of the target device; Variational mode decomposition is performed on the residuals to obtain the deviation trend of the three-dimensional model.
4. The device data extrapolation method based on digital twins according to claim 3, characterized in that, The step of performing variational mode decomposition on the residuals to obtain the deviation trend of the three-dimensional model includes: The residual is subjected to variational mode decomposition to obtain multiple modal components; The modal components with a center frequency lower than a preset frequency threshold are used as the deviation trend; The step of correcting the parameters of the three-dimensional model based on the deviation trend includes: Offline sensitivity analysis is performed on the three-dimensional model to obtain the parameter residual sensitivity mapping, and the target parameters related to the deviation trend in the three-dimensional model are determined based on the parameter residual sensitivity mapping. The correction amount of the target parameter is obtained by an incremental update algorithm, and the target parameter of the three-dimensional model is corrected according to the correction amount.
5. The device data extrapolation method based on digital twins according to claim 1, characterized in that, The step of performing logical verification on the predicted event according to preset rules to obtain the logical compliance of the target device's data includes: The white-box logic verification engine verifies the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules, thereby obtaining the logical compliance of the target device's data.
6. The device data extrapolation method based on digital twins according to claim 5, characterized in that, The white-box logic verification engine is constructed using a fuzzy temporal Petri net, which includes the preset rules. The step of using the white-box logic verification engine to verify the causal logic, temporal logic, and mutual exclusion logic of the predicted event according to the preset rules, and obtaining the logical compliance of the target device's data, includes: The degree of violation of causal logic, temporal logic, and mutual exclusion logic of the predicted event is determined by the membership function; The logical compliance of the target device's data is obtained based on the degree of violation of predicted events that do not meet the preset rules.
7. The device data extrapolation method based on digital twins according to claim 1, characterized in that, The step of determining the target inference result of the target device's data from the candidate inference results based on the logical conformity of the target device's data includes: Determine the quality score of the target device's data and the stability score of the parameter-corrected 3D model; The overall confidence level of the target device's data is determined based on the logical compliance, the quality score, and the stability score. Based on the comprehensive confidence level, a hierarchical gating strategy is applied to the candidate inference results; the hierarchical gating strategy includes allowing, downgrading, or rolling back. Candidate inference results whose overall confidence level meets the preset confidence threshold are used as the target inference results for the data of the target device.
8. A device for extrapolating equipment data based on digital twins, characterized in that, The device includes: The equipment data acquisition module is used to acquire the measured values of the target equipment data and the three-dimensional model corresponding to the target equipment; The equipment data extrapolation module is used to extrapolate the data of the target equipment based on the three-dimensional model corresponding to the target equipment, and obtain the extrapolated value and predicted event of the target equipment data; The model parameter correction module is used to obtain the deviation trend of the three-dimensional model based on the inferred value and the measured value of the target device data, and to correct the parameters of the three-dimensional model according to the deviation trend. The candidate result determination module is used to extrapolate the data of the target device using the parameter-corrected three-dimensional model to obtain candidate extrapolation results of the data of the target device; The event logic verification module is used to perform logical verification on the predicted event according to preset rules to obtain the logical compliance of the data of the target device. The deduction result determination module is used to determine the target deduction result of the target device's data from the candidate deduction results based on the logical conformity of the target device's data.
9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the device data extrapolation method based on digital twins as described in any one of claims 1-7.
10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the device data extrapolation method based on digital twins as described in any one of claims 1-7.
