Rapid online evaluation method for credibility of equipment digital twinborn integrated evolution model
Through Monte Carlo sampling error aggregation, dynamic trigger evaluation and Bayesian inference correction, the real-time credibility evaluation problem of equipment digital twin ensemble models is solved, and fast online evaluation and accurate correction are achieved, improving the accuracy and credibility of the evaluation.
Patent Information
- Application Number
- CN202510426006.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-07
- Publication Date
- 2025-07-11
AI Technical Summary
The existing equipment digital twin integrated model evaluation method cannot meet the credibility assessment needs of real-time online data evolution, resulting in a large gap between the model's credibility value and the real value, which may cause major failure losses.
Using methods based on Monte Carlo sampling error aggregation, dynamic trigger evaluation and computational tree simplification, combined with Bayesian inference, rapid online correction is performed to improve evaluation accuracy using limited real observation data.
It realizes rapid online aggregation and accurate correction of errors in equipment digital twin integrated models, improves the accuracy and credibility of online evaluation, and meets the needs of real-time applications.
Smart Images
Figure CN120295891A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of digital twin models, and in particular to a rapid online evaluation method for the credibility of an equipment digital twin integrated evolution model. Background Art
[0002] Equipment digital twin refers to a digital twin model built for equipment entities. Its core feature is online self-evolution, that is, online correction of model parameters, structure or mechanism based on real-time collected equipment data, so that the model can maintain a high degree of real-time consistency with the physical object. However, in practice, it is difficult for equipment digital twins to be put into use on a large scale and really play a role. One of the bottlenecks is the trustworthiness of equipment digital twins. Only trustworthy digital twins can accurately and timely reflect the characteristics and status of equipment, and then assist people in making correct decisions. The use of untrustworthy digital twins cannot complete the predetermined tasks and may even bring disastrous consequences. Trustworthy evaluation is an important means to ensure the trustworthiness of equipment digital twins. Due to the dynamic evolution and strong uncertainty characteristics of digital twins, traditional model evaluation methods cannot meet the trustworthy evaluation requirements of equipment digital twins. The overall equipment digital twin model can be regarded as a collection of multi-disciplinary and multi-granular unit models. There are many evaluation methods for unit twin models in various disciplines and granularities, but there is no suitable method for integrated twin models.
[0003] There are two main evaluation methods for integrated twin models: (1) comparing the simulation data and measured data of the system-level integrated evolution model, quantifying them into similarity values, and using them as the credibility values of the integrated evolution model; (2) determining the credibility values of each unit model through a large number of experiments and analyses, and then using methods such as the analytic hierarchy process to aggregate the credibility values as the credibility values of the integrated evolution model. However, neither of these two methods can meet the trustworthiness evaluation requirements of the integrated twin model, mainly because the twin model is continuously evolving online based on real-time data, which will cause the model to undergo unpredictable changes.
[0004] For the first type of method, the integrated twin model involves multiple unit models and therefore has strong uncertainty. Moreover, it is much more difficult to obtain system-level measured data for the integrated evolution model than for the unit model. This requires this type of method to use a small amount of limited real data to evaluate the integrated evolution model with more uncertainty. Obviously, the results obtained are difficult to reflect the true credibility of the integrated evolution model.
[0005] For the second type of method, if the unit model is a basic model, i.e., a model that does not evolve with real-time data, its credibility value can be confirmed offline. However, if the unit model is also an evolving model and its credibility changes with the uncertainty of the demand scenario, it can only be evaluated online. At this time, the integrated evolving model itself is also being applied online, and its real-time credibility directly affects the application results and may even cause significant failure losses. Therefore, it is not possible to wait for the credibility result of the unit evolving model before integration, but only online integrated evaluation can be carried out.
[0006] Therefore, for the equipment digital twin integrated evolving model containing unit evolving models, the evaluation of its credibility can only be carried out online. This will lead to two unsolved problems: (1) At each moment, the integrated evolving model may participate in the application, and its real-time credibility value needs to be updated quickly online to detect newly added untrustworthy situations as early as possible, so as to be able to respond and handle them in a timely manner and avoid significant losses; (2) Due to the errors between the known mechanisms, relationships and the reality, the method of integrating credibility through the correlation relationships between unit models will cause a gap between the credibility value and the true value. Therefore, it is necessary to be able to quickly correct the integrated credibility value through a small amount but continuously accumulating real observation data. Summary of the Invention
[0007] The purpose of the present invention is to provide a method for quickly online evaluating the credibility of an equipment digital twin integrated evolving model, which can realize the rapid online aggregation of integrated model errors, make full use of limited real observation data, and improve the accuracy of online evaluation.
[0008] To achieve the above object, the present invention provides a method for quickly online evaluating the credibility of an equipment digital twin integrated evolving model. The equipment digital twin integrated evolving model is an integrated evolving model with several levels, each level contains several sub-models, the upper-level sub-models are obtained by solving through relationship expressions by several corresponding lower-level sub-models, and the error distribution of the upper-level sub-models is obtained by performing error aggregation operations on the error distributions of the corresponding sub-models. The steps are as follows: S1. Quick online aggregation of unit model error distributions, specifically including: S1.1 Dynamically trigger the aggregation algorithm: Dynamically trigger the error aggregation operation according to the change of the sub-model error distribution in the equipment twin model; S1.2 Simplify the aggregation calculation tree according to the real-time evolution situation: S1.3 Quickly sample and aggregate errors according to the model mechanism: S2. Quick posterior correction of the aggregation result: Use Bayesian inference combined with limited real observation data to correct the error aggregation result and improve the evaluation accuracy; S3. Online score the error distribution of the integrated evolution model according to requirements: Convert user requirements into an error scoring mechanism, and calculate the credibility value of the integrated evolution model according to the error distribution.
[0009] Preferably, the dynamic trigger aggregation algorithm in S1.1 specifically includes: For the upper-level sub-model, calculate the error distribution of the corresponding lower-level sub-model in each evaluation interval, then calculate the total spacing of the sub-model error distributions in adjacent evaluation intervals, and determine whether the total spacing of the error distributions exceeds a preset threshold. When the total spacing of the error distributions exceeds the preset threshold, trigger the error aggregation calculation for the upper-level sub-model.
[0010] Preferably, S1.2 simplifies the aggregation calculation tree according to the real-time evolution situation, which specifically includes: In the evaluation interval, for the sub-models that have not undergone evolution updates, their error distributions adopt the calculation results from the previous aggregation; for the sub-models that have undergone evolution updates, conduct new sampling calculations and update the error distributions; only perform error aggregation calculations on the evolved models and their related upper-level models.
[0011] Preferably, the specific operation of S1.3 for quickly sampling and aggregating errors according to the model mechanism is: For a certain upper-level sub-model, use the Monte Carlo random sampling method to perform a random sampling according to the error distributions of its corresponding sub-models to obtain a possible combination of sub-model errors, and substitute the error values into the relationship expression between the upper-level sub-model and its corresponding sub-models to obtain the error value of the upper-level sub-model under the sub-model error combination.
[0012] Preferably, the quick posterior correction of the aggregation result in S2 specifically includes: First, correct the error aggregation result using the observed data. The core formula of Bayesian calibration is as follows: (1) Among them, is the parameter distribution of the model, D is an observed event, is the prior probability that the model takes this parameter distribution; is the sum of the probabilities of observing this event under various possible parameters, and is a constant used for normalization processing; is the likelihood probability, which is the probability of observing this event when the model conforms to this parameter distribution; is the posterior probability, which is the probability that the model takes this parameter distribution D on the premise of observing event ; The goal is to find the best parameter distribution that maximizes the posterior probability , satisfying: (2) Secondly, for the current error aggregation result, taking the output of the observed integrated evolution model as an event D , a better parameter distribution is searched within its neighborhood such that obtains a larger value. To distinguish various possible parameter distributions, formula (1) is rewritten as: (3) where is one of all possible parameter distributions that the model can take, representing the parameter distribution of the error distribution to be corrected currently aggregated; Finally, for each value of the current parameter distribution, a number of new neighborhood values are randomly generated within its neighborhood, and a number of candidate parameter distributions are generated by randomly combining the neighborhood values of all parameters. For the candidate parameter distributions, the posterior probability is calculated according to formula (3) , and the parameter distribution with the highest posterior probability is selected as the correction result, representing the error distribution of the integrated evolution model closest to the true situation at the current moment.
[0013] Preferably, S3 online scores the error distribution of the integrated evolution model according to requirements, which specifically includes: Firstly, a quantitative scoring mechanism for different error situations of users is established, and the user requirements are transformed into an error scoring function , a critical value is defined. If it exceeds the critical value , the score will drop sharply and quickly tend to 0; Secondly, the error distribution of the integrated evolution model closest to the true situation at the current moment obtained in S2 is scored online against the scoring mechanism as a credibility value, which is expressed as follows: (4) where C represents the credibility value; is the scoring function defined according to user requirements, with the input being the error value , and the output being the score value; represents the probability that the model output falls within the i th error range; is the error probability, representing the probability that the error value appears; n represents the number of probability columns of the error distribution. The probability columns are the intervals into which the error distribution is discretized, and each interval corresponds to an error range and the error probability .
[0014] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention: In response to the need for rapid online evaluation of the credibility value of the equipment digital twin integrated evolution model, a Monte Carlo sampling error aggregation method based on a mechanism model, a method for dynamically triggering evaluation, and a computational tree simplification method are proposed to achieve rapid online aggregation of the integrated model error.
[0015] In response to the problem that there are errors in the aggregated credibility value, a rapid posterior correction method based on Bayesian inference is proposed to make full use of limited real observation data and improve the accuracy of online evaluation.
[0016] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Description of the Drawings
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.
[0018] Figure 1 It is a flowchart of an embodiment of the method for rapidly online evaluating the credibility of the equipment digital twin integrated evolution model of the present invention.
[0019] Figure 2 It is a mechanism diagram of the equipment twin model of an embodiment of the present invention.
[0020] Figure 3 It is a schematic diagram of the model state of the equipment twin model of an embodiment of the present invention.
[0021] Figure 4 It is a distribution diagram of the changed error of an embodiment of the present invention, where (a) is the error distribution before error observation; (b) is the error distribution after error observation.
[0022] Figure 5 It is a simplified error aggregation calculation tree of an embodiment of the present invention.
[0023] Figure 6 It is a schematic diagram of the equipment mechanism model containing multiple types of errors of an embodiment of the present invention.
[0024] Figure 7 It is a schematic diagram of converting the core requirements into a scoring mechanism of an embodiment of the present invention. Detailed Embodiments
[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0026] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Embodiment
[0027] The equipment digital twin integrated evolution model is an integrated model with several levels. Each level contains several sub-models. The sub-models in the upper level are obtained by solving a relational expression with several corresponding sub-models in the lower level. The error distribution of the sub-models in the upper level is obtained by performing an error aggregation operation on the error distributions of the corresponding sub-models.
[0028] As Figure 1 shown, the method for rapid online evaluation of the credibility of the equipment digital twin integrated evolution model is as follows: In Figure 2 the shown equipment mechanism model, A is the system-level output, A1~A5 and A13 are the lower-level sub-models corresponding to A, A11~A13 are the lower-level sub-models corresponding to A1, and A22~A25 and A13 are the lower-level sub-models corresponding to A2. Among them, is an expression composed of the outputs A1, A13, A2, A3, A4, and A5 of the sub-models, from which the output value of A can be obtained. These expressions are generally ordinary differential equations. , and so on.
[0029] S1. Rapid online aggregation of the error distribution of the unit model, specifically including: S1.1 Dynamic trigger aggregation algorithm: Dynamically trigger the error aggregation operation according to the change of the error distribution of the sub-models in the equipment twin model.
[0030] For the actual equipment twin model, as Figure 3 shown, within each evaluation interval, there are three types of model states: the basic model that never participates in evolution, the model that is evolving during evaluation, and the model that has not evolved during evaluation. Except for the model that is evolving, the other two are models that do not need to participate in evolution due to small errors. Therefore, the evaluation conclusions for them can follow the evaluation results of the previous moment. If it is Figure 2 the case, all sub-models are basic models. After confirming the error distribution of the sub-models, only one aggregation calculation in S3 is required to determine the error distribution of the integrated model. But if it isFigure 3 In the case of Figure 3 , there is an evolving sub-model A12, whose error distribution will be continuously updated according to the real-time observed data, resulting in the continuous change of the error distributions of the aggregated ensemble models A1 and A. In theory, whenever the sub-model A12 completes an evolutionary update, the true value of its error distribution will change once, and it is necessary to observe for a period of time to confirm its changed error distribution and then update the error distribution of the ensemble model.
[0031] At this time, it can be noted that the error aggregation algorithm takes a certain amount of time, and the evolution of the model is continuous. If sampling aggregation is performed after each evolution, a large amount of resources will be consumed. However, in fact, although the evolution of the sub-model is continuous, the amount of change in the error of the sub-model is uncertain. Most of the time, the error of the sub-model does not change or changes slightly, and in this case, there is no need to change the previous aggregation calculation result. Only when the error of the sub-model changes greatly is it necessary to update the error of the ensemble model.
[0032] To improve the aggregation speed, this application designs the appropriate timing of dynamic aggregation, which can not only ensure the aggregation accuracy but also greatly improve the calculation speed, and more fully meet the digital twin evaluation requirements. For the upper-layer sub-models, calculate the error distribution of the corresponding lower-layer sub-models within each evaluation interval, and then calculate the total distance between the error distributions of the sub-models in adjacent evaluation intervals. Determine whether the total distance of the error distribution exceeds a preset threshold. When the total distance of the error distribution exceeds the preset threshold, trigger the error aggregation calculation for the upper-layer sub-models.
[0033] As Figure 4 shown, after a period of error observation, the error distribution output by a certain sub-model changes from (a) to (b). Considering the output error of this sub-model, in a specific scenario, although the events encountered by the equipment are uncertain, its frequency distribution is relatively certain. Under each event, the change in the model error caused by the same evolutionary algorithm is also relatively certain. Therefore, if the observed data is sufficient, the overall error distribution obtained by statistics should be relatively stable. It can be considered that if the error distribution hardly changes with the increase of statistical data, it can be explained that this distribution is very close to the true distribution. When the data in the intermediate stage is not yet sufficient, the error distribution obtained by statistics has not covered all possibilities, so there will still be certain changes.
[0034] For intuitive display and computational convenience, the error distribution is uniformly discretized. That is, for the same sub-model, the same error range is specified, and the error intervals are divided at the same interval. Therefore, on the error distribution diagrams of two adjacent versions of the same sub-model, the error meaning represented by each frequency bar is the same. So, the sum of the differences between the frequency bars of adjacent error distribution diagrams can represent the difference between adjacent error distributions. Additionally, the distribution probability of each error distribution diagram is normalized, so this sum of differences can significantly reflect the difference between error distribution diagrams. Each error distribution diagram is evenly divided into 50 frequency bars. Assume Figure 4 Figure (a) in is the error distribution statistically obtained at time . Assume Figure 4 Figure (b) in is the error distribution statistically obtained at time . Then the total interval of the error distributions of the two diagrams is calculated as follows: (5) Thus, the timing for aggregation can be confirmed in combination with the evolution change amount. That is, compare the error distribution at the latest time with the error distribution at the previous update, and calculate the total interval between them. If this value exceeds the preset change amount threshold, trigger the error aggregation algorithm. After using this method, both the accuracy of the aggregation result can be guaranteed, and the limited computing resources can be concentrated on more necessary operations, improving the computing speed and more fully meeting the high-speed requirements of digital twin evaluation.
[0035] S1.2. Simplify the aggregation calculation tree according to the real-time evolution situation.
[0036] Although the timing of error aggregation can be dynamically judged and optimized through S1.1, there may still be evolutions with short intervals and large change amounts between two times, and the aggregation time between the two changes is relatively limited. Therefore, it is necessary to further simplify the unnecessary computational amount during aggregation, thereby improving the speed of error aggregation.
[0037] According to Figure 3It can be seen that within this evaluation interval, since only A12, A1, and A have evolved and been updated, only the output error distributions of these three models have changed. All other models have not evolved due to small errors, so the error distributions of other models can follow the conclusions from the previous aggregation. That is to say, the core of the computational tree simplification lies in that within the evaluation interval, for sub-models that have not undergone evolutionary updates, their error distributions follow the computational results from the previous aggregation; for sub-models that have undergone evolutionary updates, new sampling calculations are performed and the error distributions are updated; error aggregation calculations are only performed on the evolved models and their related upper-layer models.
[0038] According to Figure 5 the evolutionary relationships of each model, the simplified error aggregation computational tree is shown in the figure. A11, A13, A2, A3, A4, and A5 do not need to evolve during the evaluation and can directly use the error distribution conclusions from the previous evaluation. Among them, since A2 can directly use the previous conclusion, the related sub-models no longer need to participate in the aggregation operation. A12, A1, and A are the parts that have evolved during the evaluation and need to update the error distributions. Among them, A12 is an end node and can only update the distribution by observing new data. A1 and A are intermediate nodes and need to update the distribution through error aggregation. Finally, only error aggregation operations need to be performed on A1 and A.
[0039] S1.3. Rapidly sample and aggregate the errors according to the model mechanism.
[0040] For the traditional theoretical aggregation method of dividing intervals and traversing combinations, its complexity is , where k is the number of sub-models and n is the number of samples. It can be seen that its computational complexity is very large. For equipment models composed of dozens or even hundreds of sub-models, the time consumption is excessive and it is not suitable for online aggregation evaluation. Therefore, the sampling method of error points is changed to Monte Carlo random sampling.
[0041] The advantage of the Monte Carlo method is that the calculation of each random sample is independent, so it is very easy to parallelize and is suitable for large-scale calculations, greatly improving the aggregation speed. However, to obtain sufficiently accurate results, a large number of random samples are usually required, which will lead to a large computational overhead, especially when facing high-precision requirements. The accuracy of the Monte Carlo method depends on the number of samples. According to the law of large numbers, as the number of samples increases, the estimated value will approach the true value. Assuming that independent random experiments are carried out, the error usually decreases as the number of samples increases. Specifically, the standard deviation of the error is usually , so increasing the number of samples can reduce the estimation error.
[0042] Considering the online evaluation of the equipment digital twin, when aggregating using this method, generally, smaller-scale unit models are sampled. Their error distribution is relatively simple and there is no multi-peak situation, so the sampling points do not need to be too dense. However, a high aggregation speed is required because it is necessary to quickly know the error situation of the current integrated model in order to understand the degree of satisfaction of the current application requirements. In this regard, the Monte Carlo method is more suitable for this scenario.
[0043] For a certain upper-layer sub-model, the Monte Carlo random sampling method is used to conduct a random sampling once according to the error distribution of its corresponding sub-models, obtaining a possible combination of sub-model errors. According to the Figure 3 correlation mechanism therein, substitute the error values into the relational expression between the upper-layer sub-model and its corresponding sub-models to obtain the error value of the upper-layer sub-model under the sub-model error combination. Taking the error value of the A1 sub-model in the figure as an example, sample the errors of the three sub-models once according to the error frequency distributions of A11, A12, and A13 respectively, and obtain , , respectively. Substitute them into , and it can be obtained that the error presented by A1 under this error combination is: (6) Combining the output error distributions of A1 and A2 with the error distributions of A3, A4, and A5 that have been obtained, an approximate value of the output error distribution of the integrated model A can be further calculated. Then, based on this error distribution and combined with the actual application requirements, the degree of satisfaction of the integrated model A with the application requirements can be analyzed.
[0044] S2. Fast posterior correction of the aggregation result: Use Bayesian inference combined with limited real observation data to correct the error aggregation result and improve the evaluation accuracy.
[0045] There are multiple types of errors that are difficult to eliminate between the equipment mechanism model under the existing technology and the actual physical equipment. As Figure 6 shown, all parts containing deviations are marked in red, and the dotted line is the missing part. Among them, A6 is the missing element. For example, the humidity parameter in the environment has a significant impact on the model performance, but it has not been incorporated into the mechanism model due to insufficient understanding of the developers or limited computing power. As a result, the influence mechanism of A6 on A has not been written into the relational expression. In addition, although A12 has been incorporated into the mechanism model, its direct impact on the integrated model A has not been written into the relational expression, and it only indirectly affects A through A1. In the A2 part, although all the main elements affecting A2 are complete, the characterization of their influence mechanism on A2 There are certain deviations. Therefore, during the simulation process, the value of A2 will continuously accumulate errors, and these errors will further be transmitted into the errors of the integrated model A. The above-mentioned various errors are unavoidable during the model development stage and will cause the error aggregation result in S1 to deviate from the actual error.
[0046] To obtain a more accurate evaluation result of the integrated evolution model, it is necessary to make more full use of the available observation data. Although the observation data of the integrated model is scarce and it is impossible to directly analyze the error distribution of the integrated model from it, these scarce data can be used to correct the existing error aggregation result. Specifically, it includes: First, use the observation data to correct the error aggregation result. The core formula of the Bayesian calibration adopted is as follows: (1) Among them, is the parameter distribution of the model, D is a certain observed event, is the prior probability that the model takes this parameter distribution; is the sum of the probabilities of observing this event under various possible parameters and is a constant used for normalization processing; is the likelihood probability, which is the probability of observing this event when the model conforms to this parameter distribution; is the posterior probability, which is the probability that the model takes this parameter distribution D under the premise of observing the event
[0047] The goal is to find the best parameter distribution that maximizes the posterior probability , satisfying: (2) However, this formula involves many parameters and the expression may be relatively complex. It is difficult to accurately give the exact expression relationship and it is also difficult to obtain the best in a short time. Note that for the equipment digital twin, the observation data output by its integrated model is limited. Therefore, the statistical distribution of the output error of the integrated model has a certain deviation from the true distribution. Therefore, the correction strength based on this data cannot be too large. It should be adjusted gradually and moderately similar to the backpropagation of the neural network. As the observation data of the integrated model increases, its correction accuracy will be higher and the aggregation result can be gradually corrected more effectively.
[0048] Therefore, the core of the correction is to, for the current error aggregation result, take the output of the observed integrated evolution model as the event D , and find a better parameter distribution in its neighborhood to make obtain a larger value.
[0049] To distinguish various possible parameter distributions, formula (1) is rewritten as: (3) where is one of all possible parameter distributions that the model can take, representing the error distribution parameter that needs to be corrected currently aggregated, that is, the values of 50 frequency bins of the error distribution at the latest moment in formula (5) .
[0050] Finally, for each value of the current parameter distribution, several new neighborhood values are randomly generated within its neighborhood, and several candidate parameter distributions are generated by randomly combining the neighborhood values of all parameters. For the candidate parameter distributions, the posterior probability is calculated according to formula (3) , and the parameter distribution with the highest posterior probability is selected as the correction result, representing the error distribution of the integrated evolution model closest to the true situation at the current moment.
[0051] For example, for , 10 new neighborhood values can be randomly taken within the neighborhood of each frequency bin value according to formula (7): (7) where represents a random value within the range of [-1, 1][−1, 1].[[]END]]
[0052] Then, 10 sets of new combinations of frequency values are randomly selected from 50 sets of prepared neighborhood values as candidate parameter distributions within the neighborhood. For each set of parameter distributions, according to formula (6), the probability of obtaining this set of parameter distributions for the output error of the integrated model observed at this time can be calculated. Finally, the set with the highest posterior probability is selected from these 10 sets of frequency value combinations as the corrected parameter distribution. If there is still time remaining for correction, the number of combinations of frequency values extracted within the neighborhood can be increased to find a better parameter distribution.
[0053] S3. Online scoring of the integrated evolution model error distribution according to requirements: A better way to describe the model error is the error distribution, that is, the probability corresponding to the different gaps between the output of the model and the actual data. If the model credibility is to be evaluated according to the error distribution, the user requirements need to be transformed into a scoring mechanism for different error situations. Specifically, it includes: First, establish a quantitative scoring mechanism for different error situations of the user. Transform the user requirements into an error scoring function , define the critical value , if it exceeds the critical value , the score will drop sharply and quickly tend to 0.
[0054] Taking the sorting application of a certain type of robotic arm as an example, the digital twin model of the robotic arm is an integrated model as a whole, including continuously evolving motor unit models, kinematic models, geometric models, etc. Before each sorting application, it is necessary to determine the optimal control signal according to the simulation results of the latest version of the twin model to ensure the sorting accuracy. The core requirement is to accurately transfer the workpiece to be sorted to the central position of the conveyor belt of the lower-level production line. The farther the deviation is, the greater the impact on the next-level processing process. Exceeding a certain deviation threshold will cause a blockage failure of the production line. The scoring mechanism transformed from this core requirement is as Figure 7 shown.
[0055] Among them, the landing point is a region centered on , and the workpiece is an object centered on . is the maximum deviation of the workpiece beyond the landing point area. According to the user's requirement description, The larger it is, the lower the score of each sorting application will be, and there is a critical value . If it exceeds this critical value, it means that the deviation of the workpiece is too large, which will seriously affect the subsequent processing process, and then the score will drop sharply and quickly tend to 0. By determining the critical value and the formulas of the two curves according to the specific requirements of the user, a quantitative scoring mechanism for different error situations can be obtained for the user.
[0056] Then, the error distribution of the integrated evolution model closest to the real situation at the current moment obtained in S2 is used to perform an online score according to the scoring mechanism, which is used as the credibility value and is expressed as follows: (4) Among them, C represents the credibility value; is the scoring function defined according to the user's requirements, with the input being the error value and the output being the score value; represents the probability that the model output falls within the i th error range; is the error probability, indicating the probability that the error value appears; n represents the number of probability columns of the error distribution. The probability column is the interval into which the error distribution is discretized, and each interval corresponds to an error range and the error probability .
[0057] According to the Figure 7 scoring mechanism, the error corresponding to each column is converted into a score , and then multiplied by the probability value , which is the contribution value of each column to the overall credibility. By accumulating the contribution values of each probability column one by one, the overall credibility value of the integrated evolution model can be obtained. As the observed data increases, the resulting error distribution will get closer and closer to the true value. Therefore, the credibility value obtained by scoring will fluctuate and approach the true value. When the fluctuation amplitude remains stable within a small range for a long time, the obtained credibility value can be regarded as the true credibility value of the digital twin integrated evolution model of the equipment in this demand scenario.
[0058] For the remaining technical features in the above embodiments, those skilled in the art can flexibly select them according to the actual situation to meet different specific actual needs. However, it is obvious to those of ordinary skill in the art that these specific details do not have to be adopted to implement the present invention. In other instances, well-known components, structures, or parts are not specifically described in order to avoid obscuring the present invention, and all are within the scope of the technical solutions claimed in the claims of the present invention.
[0059] Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention. In the above description, in order to provide a thorough understanding of the present invention, a large number of specific details are elaborated. However, it is obvious to those of ordinary skill in the art that these specific details do not have to be adopted to implement the present invention. In other instances, well-known technologies, such as specific construction details, working conditions, and other technical conditions, are not specifically described in order to avoid obscuring the present invention.
[0060] In this article, specific examples are used to elaborate on the principles and implementation methods of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.
Claims
1. A rapid online evaluation method for the credibility of an equipment digital twin integrated evolution model, where the equipment digital twin integrated evolution model is an integrated model with several levels, each level contains several sub-models, the sub-models in the upper layer are obtained by solving a relational expression with several corresponding sub-models in the lower layer, and the error distribution of the sub-models in the upper layer is obtained by performing an error aggregation operation on the error distributions of the corresponding sub-models. It is characterized in that, The steps are as follows: S1. Rapid online aggregation of unit model error distribution, specifically including: S1.1 Dynamically trigger the aggregation algorithm: Dynamically trigger the error aggregation operation according to the change of the sub-model error distribution in the equipment twin model; S1.2 Simplify the aggregation calculation tree according to the real-time evolution situation: S1.3 Rapid sampling aggregation of errors according to the model mechanism: S2. Rapid posterior correction of the aggregation result: Use Bayesian inference combined with limited real observation data to correct the error aggregation result and improve the evaluation accuracy; S3. Online scoring of the error distribution of the integrated evolution model according to requirements: Convert the user requirements into an error scoring mechanism and calculate the credibility value of the integrated evolution model according to the error distribution.
2. The rapid online evaluation method for the credibility of the equipment digital twin integrated evolution model according to claim 1, characterized in that: S1.1 The dynamically triggered aggregation algorithm specifically includes: For the upper-layer sub-model, calculate the error distribution of the corresponding lower-layer sub-model in each evaluation interval, and then calculate the total spacing of the sub-model error distributions in adjacent evaluation intervals. Determine whether the total spacing of the error distributions exceeds a preset threshold. When the total spacing of the error distributions exceeds the preset threshold, trigger the error aggregation calculation for the upper-layer sub-model.
3. The rapid online evaluation method for the credibility of the equipment digital twin integrated evolution model according to claim 2, wherein: S1.2 Simplifying the aggregation calculation tree according to the real-time evolution situation specifically includes: In the evaluation interval, for the sub-models that have not undergone evolutionary updates, their error distributions follow the calculation results of the previous aggregation; for the sub-models that have undergone evolutionary updates, new sampling calculations are performed and the error distributions are updated; only the models that have evolved and their related upper-layer models are subjected to error aggregation calculations.
4. The rapid online evaluation method for the credibility of the equipment digital twin integrated evolution model according to claim 3, characterized in that: The specific operation of S1.3 for rapid sampling aggregation of errors according to the model mechanism is: For a certain upper-layer sub-model, use the Monte Carlo random sampling method to perform a random sampling according to the error distributions of its corresponding sub-models to obtain a possible combination of sub-model errors, and substitute the error values into the relationship expression between the upper-layer sub-model and its corresponding sub-models to obtain the error value of the upper-layer sub-model under the sub-model error combination.
5. The rapid online evaluation method for the credibility of the equipment digital twin integrated evolution model according to claim 4, characterized in that: S2 The rapid posterior correction of the aggregation result specifically includes: First, correct the error aggregation result using the observation data. The core formula of the Bayesian calibration used is as follows: （1） Among them, is the parameter distribution of the model, D is a certain observed event, is the prior probability that the model takes this parameter distribution; is the sum of probabilities of observing this event under various possible parameters, and is a constant used for normalization processing; is the likelihood probability, which is the probability of observing this event when the model conforms to this parameter distribution; is the posterior probability, which is the probability that the model takes this parameter distribution D on the premise of observing the event ; The goal is to find the optimal parameter distribution that maximizes the posterior probability and satisfies: （2） Secondly, for the current error aggregation result, taking the output of the observed integrated evolution model as an event D , a better parameter distribution is searched within its neighborhood such that obtains a larger value. To distinguish various possible parameter distributions, Equation (1) is rewritten as: （3） Among them, is one of all possible parameter distributions that the model can obtain, representing the error distribution parameter that needs to be corrected currently aggregated; Finally, for each value of the current parameter distribution, several new neighborhood values are randomly generated within its neighborhood, and several candidate parameter distributions are generated by randomly combining the neighborhood values of all parameters. For the candidate parameter distributions, the posterior probability is calculated according to formula (3). The parameter distribution with the highest posterior probability is selected as the calibration result, representing the error distribution of the integrated evolution model that is closest to the true situation at the current moment.
6. The rapid online evaluation method for the credibility of the equipment digital twin integrated evolution model according to claim 5, characterized in that: S3 Online scoring of the error distribution of the integrated evolution model according to requirements specifically includes: First, establish a quantitative scoring mechanism for users in different error situations, and transform user requirements into an error scoring function , define a critical value . If it exceeds the critical value , the score will drop sharply and quickly tend to 0; Second, the error distribution of the integrated evolution model that is closest to the real situation at the current moment obtained in S2 is scored online according to the scoring mechanism and used as the credibility value, which is expressed as follows: （4） where C represents the credibility value; To define a scoring function according to user requirements, the input is the error value , and the output is the scoring value; Indicates the probability that the model output falls within the i th error range; is the error probability, representing the error value the probability of occurrence; n The number of probability bars representing the error distribution, where the probability bars are the intervals into which the error distribution is discretized, and each interval corresponds to an error range and the error probability .