Distributed calibration-based historical-data-free digital twin decision model updating method

Abstract

The invention discloses a distributed calibration-based historical-data-free digital twinborn decision model updating method, and belongs to the technical field of deep learning. The method comprises the following steps: constructing a preset digital twinborn decision model, and updating the digital twinborn decision model of a t stage, the updating comprising: obtaining input data of the t stage; based on the fitting distribution of the historical data of the stages from 1 to t-1 stored in the learnable Gaussian mixture model parameter mapping module, sampling the fitting distribution to obtain a target sample, and based on the target sample, obtaining an offset sample of the historical data of the stages from 1 to t-1; correcting the bias sample based on a distribution correction module to obtain correction historical data from 1 to t-1 stages; updating the classifier based on the corrected historical data and the input data features of the t stage to obtain a digital twinborn decision model after the t stage is updated; generating and storing a fitting distribution of the input data of the t phase; according to the method, the dependence on historical data and the storage cost during updating of the digital twin decision model are reduced.

Description

Technical Field

[0001] This application belongs to the technical field of deep learning, and particularly relates to a method for updating a digital twin decision model without historical data based on distribution calibration. Background Art

[0002] In the application scenario of prognostics and health management, as the core technology of intelligent decision-making, the dynamic evolution ability of the digital twin decision model is directly related to the accuracy of equipment health status assessment and the timeliness of operation and maintenance decisions. The asynchronous evolution of the full life cycle data of healthy equipment and the physical model may lead to the lack of accuracy of virtual entities, thereby causing false fault judgments and early warning delays. Therefore, implementing an adaptive calibration mechanism for the model has become the key path to breaking through the technical bottleneck of predictive maintenance.

[0003] Existing state assessment methods mainly rely on artificial experience and signal processing technology, but there are still significant application limitations. First, traditional manual detection methods have insufficient knowledge generalization ability and are difficult to handle the identification of multi-modal fault features under complex working conditions. Second, although the diagnostic technology based on signal processing has physical interpretability, its ability to characterize non-linear degradation processes is limited, and there are inherent defects in the decoupling of feature engineering and diagnostic models. In addition, the deep learning model trained offline adopts an architecture that separates static modeling and dynamic monitoring, lacking a knowledge inheritance mechanism during the model update process, resulting in the problem of catastrophic forgetting in the incremental learning scenario.

[0004] The current technology faces multiple challenges in the application of prognostics and health management. Existing methods are difficult to effectively coordinate the contradiction between model fidelity and computational real-time performance, have weak adaptability to dynamic degradation processes, and generally have problems such as excessive dependence on historical data, fast knowledge forgetting speed, and high storage costs. Summary of the Invention

[0005] This application aims to solve at least one of the technical problems existing in the prior art. For this purpose, this application proposes a method for updating a digital twin decision model without historical data based on distribution calibration, reducing the dependence on historical data and storage costs during the update of the digital twin decision model.

[0006] In a first aspect, this application provides a method for updating a digital twin decision model without historical data based on distribution calibration, the method comprising:

[0007] Construct a preset digital twin decision model, the digital twin decision model includes a feature extractor, a distribution learning unit, and a classifier, the distribution learning unit includes a learnable Gaussian mixture model parameter mapping module and a distribution correction module, the feature extractor is used to extract the features of the input data, the distribution learning unit is used to generate and store the fitted distribution of the input data, and the classifier is used to make a classification decision on the features of the input data;

[0008] Update the digital twin decision model in the t-th stage, where the update includes:

[0009] Obtain the input data in the t-th stage;

[0010] Based on the fitted distribution of the historical data from the 1st to the (t - 1)-th stage stored in the learnable Gaussian mixture model parameter mapping module, sample the fitted distribution to obtain target samples, and obtain the biased samples of the historical data from the 1st to the (t - 1)-th stage based on the target samples;

[0011] Based on the distribution correction module, correct the biased samples to obtain the corrected historical data from the 1st to the (t - 1)-th stage;

[0012] Input the input data in the t-th stage into the feature extractor to obtain the feature of the input data in the t-th stage;

[0013] Based on the corrected historical data and the feature of the input data in the t-th stage, update the classifier to obtain the updated digital twin decision model in the t-th stage;

[0014] Generate and store the fitted distribution of the input data in the t-th stage;

[0015] Where t is a positive integer greater than 1.

[0016] According to an embodiment of the present application, the generating and storing the fitted distribution of the input data in the t-th stage includes:

[0017] Based on the learnable Gaussian mixture model parameter mapping module, generate the fitted distribution of the input data in the t-th stage;

[0018] Based on the fitted distribution of the input data in the t-th stage, update the parameters of the distribution learning unit according to the minimized distribution metric loss function.

[0019] According to an embodiment of the present application, the learnable Gaussian mixture model parameter mapping module includes a mapping function and a multivariate Gaussian mixture distribution sampling module, and the mapping function includes a plurality of fully connected layers.

[0020] According to an embodiment of the present application, the generation process of the fitted distribution of the input data includes:

[0021] Input the input data into the feature extractor to obtain the input data feature;

[0022] Input the input data feature into the mapping function to obtain the mean vector, variance vector and mixing coefficient vector of the input data;

[0023] Based on the mean vector, variance vector, and mixing coefficient vector, obtain the fitted distribution of the input data.

[0024] According to an embodiment of the present application, the distribution correction module is a neural network model constructed based on the U-net network architecture, and the calculation formula for the correction performed by the distribution correction module is as follows:

[0025]

[0026] where g(·) is the calibration process, is the corrected sample, is the biased sample, θ u is the network weight parameter of the distribution learning unit.

[0027] According to an embodiment of the present application, the target sample is obtained by sampling based on the K-means clustering algorithm.

[0028] According to an embodiment of the present application, the calculation formula for minimizing the distribution metric loss function is as follows:

[0029]

[0030] where L d represents the distribution metric loss function, is the sample after correction at stage t, f (t) is the target sample at stage t, θ g is the network weight parameter of the feature extractor, θ u is the network weight parameter of the distribution learning unit, and n is the total number of samples.

[0031] Second, the present application provides a digital twin decision model updating device based on distribution calibration without historical data, and the device includes:

[0032] A construction module, configured to construct a preset digital twin decision model, where the digital twin decision model includes a feature extractor, a distribution learning unit, and a classifier. The distribution learning unit includes a learnable Gaussian mixture model parameter mapping module and a distribution correction module. The feature extractor is used to extract the features of the input data, the distribution learning unit is used to generate and store the fitted distribution of the input data, and the classifier is used to make a classification decision on the features of the input data;

[0033] An update module, configured to update the digital twin decision model at stage t, and the update includes:

[0034] Obtain the input data at stage t;

[0035] Based on the fitted distribution of the historical data from stage 1 to stage t - 1 stored in the learnable Gaussian mixture model parameter mapping module, sample the fitted distribution to obtain target samples, and obtain the biased samples of the historical data from stage 1 to stage t - 1 based on the target samples;

[0036] Based on the distribution correction module, correct the biased samples to obtain the corrected historical data from stage 1 to stage t - 1;

[0037] Input the input data of stage t into the feature extractor to obtain the feature of the input data of stage t;

[0038] Based on the corrected historical data and the feature of the input data of stage t, update the classifier to obtain the updated digital twin decision model at stage t;

[0039] A storage module for generating and storing the fitted distribution of the input data of stage t;

[0040] Where t is a positive integer greater than 1.

[0041] In a third aspect, the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the method for updating the digital twin decision model without historical data based on distribution calibration as described in the first aspect above.

[0042] In a fourth aspect, the present application provides a non - transitory computer - readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method for updating the digital twin decision model without historical data based on distribution calibration as described in the first aspect above.

[0043] In a fifth aspect, the present application provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method for updating the digital twin decision model without historical data based on distribution calibration as described in the first aspect.

[0044] In a sixth aspect, the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the method for updating the digital twin decision model without historical data based on distribution calibration as described in the first aspect above.

[0045] Additional aspects and advantages of the present application will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present application.

[0046] A method for updating a digital twin decision model based on distribution calibration without historical data provided by the present invention has the following beneficial effects compared with the prior art:

[0047] (1) Through the learnable Gaussian mixture model parameter mapping module, the present invention can sample the fitted distribution of the stored historical data to obtain biased samples without historical data, and reverse-infer the biased samples through the distribution calibration module to correct the historical data deviation. In the model update stage, the distribution learning unit can efficiently process the calibration and fusion between the input data and the historical data distribution, reduce the knowledge forgetting of the digital twin decision model, reduce the computational cost, improve the update efficiency while maintaining the model prediction accuracy, and is applicable to the rapid iteration of the model under different application scenarios.

[0048] (2) By generating the fitted distribution of the input data at stage t through the learnable Gaussian mixture model parameter mapping module, the present invention can store the fitted distribution of the input data in the digital twin decision model update stage. By minimizing the distribution metric loss function to update the parameters of the distribution learning unit, the feature storage of the input data at each stage is realized, enabling the digital twin decision model to recover the feature distribution of the historical data without saving historical data, reducing the forgetting rate of the digital twin decision model, improving the diagnostic accuracy of the model under different working conditions, and reducing the model update time and memory consumption.

[0049] (3) By using the mapping function and the multivariate Gaussian mixture distribution sampling module, the present invention can effectively learn and adjust the parameters of the Gaussian mixture model, providing a powerful non-linear modeling ability for the digital twin decision model. By integrating multiple fully connected layers into the mapping function, the complex relationships of the input data can be captured more deeply, reducing the dependence of the digital twin decision model on parameter initialization. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, where:

[0051] Figure 1 is a schematic structural diagram of a digital twin decision model provided by an embodiment of the present application;

[0052] Figure 2 is one of the schematic flowcharts of a method for updating a digital twin decision model based on distribution calibration without historical data provided by an embodiment of the present application;

[0053] Figure 3 is another schematic flowchart of a method for updating a digital twin decision model based on distribution calibration without historical data provided by an embodiment of the present application;

[0054] Figure 4It is a schematic structural diagram of a digital twin decision model updating device based on distribution calibration without historical data provided by an embodiment of the present application;

[0055] Figure 5 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Specific embodiments

[0056] Next, the technical solutions in the embodiments of the present application will be clearly described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0057] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually of the same type, and the number of objects is not limited. For example, the first object can be one or multiple. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / " generally represents an "or" relationship between the associated objects before and after.

[0058] Next, in conjunction with the accompanying drawings, through specific embodiments and their application scenarios, the digital twin decision model updating method based on distribution calibration without historical data, the digital twin decision model updating device based on distribution calibration without historical data, the electronic device, and the readable storage medium provided by the embodiments of the present application will be described in detail.

[0059] Among them, the digital twin decision model updating method based on distribution calibration without historical data can be applied to a terminal, and can be specifically executed by hardware or software in the terminal.

[0060] The terminal includes, but is not limited to, portable communication devices such as mobile phones or tablet computers with a touch-sensitive surface (for example, a touch screen display and / or a touchpad). It should also be understood that in some embodiments, the terminal may not be a portable communication device, but a desktop computer with a touch-sensitive surface (for example, a touch screen display and / or a touchpad).

[0061] In the following various embodiments, a terminal including a display and a touch-sensitive surface is described. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, a mouse, and a joystick.

[0062] The digital twin decision model update method based on distribution calibration without historical data provided by the embodiments of the present application. The execution subject of the digital twin decision model update method based on distribution calibration without historical data can be an electronic device or a functional module or functional entity in the electronic device that can implement the digital twin decision model update method based on distribution calibration without historical data. The electronic devices mentioned in the embodiments of the present application include, but are not limited to, mobile phones, tablet computers, computers, cameras, wearable devices, etc. Hereinafter, taking the electronic device as the execution subject as an example, the digital twin decision model update method based on distribution calibration without historical data provided by the embodiments of the present application will be described.

[0063] With the development of technologies such as big data, cloud computing, the Internet of Things, and artificial intelligence, the trend of automation has become a major development direction in the field of prognostics and health management. The digital twin decision model can integrate the real-time data of sensors and big data analysis to construct an accurate virtual representation of the device for dynamic monitoring and evaluation. This indicates that prognostics and health management technologies can achieve optimal fault diagnosis performance and timely prediction capabilities. Therefore, more and more scholars and engineers have begun to deeply study the theory, methods, and technical applications of digital twins in prognostics and health management.

[0064] Traditional health status assessment methods mainly determine the device status and fault location through manual detection or signal processing techniques. However, traditional methods have high labor costs and rely heavily on the professional knowledge of maintenance personnel. In addition, the diagnostic results of signal processing techniques are often highly professional and have weak interpretability, which greatly limits the wide application of condition monitoring techniques.

[0065] In recent years, with the rapid development of artificial intelligence, end-to-end fault diagnosis methods based on deep learning have received extensive attention, which can achieve automatic diagnosis of device status without relying on manual feature extraction. However, current research still relies on supervised learning of scaled physical prototype data because deep learning-based methods require statistical analysis of large datasets to establish the input-output relationship of the system and rely on large-scale labeled data, while in industrial scenarios, the operating data is scarce, the cost of fault simulation is high, and data annotation is time-consuming and laborious.

[0066] With the development of virtual simulation technology, the concept of digital twins has attracted great attention in both academic and industrial research in the field of prognostics and health management. Since digital twins can construct a high-fidelity virtual representation of physical entities in the digital space, the problem of data scarcity can be effectively solved.

[0067] However, the core of digital twin technology is not limited to high-fidelity modeling and multi-source data integration technologies. The real-time update and calibration of models remain key challenges that need to be addressed urgently. Models trained offline cannot cover all task profiles, and the increasing operational wear and different operating conditions during device use also greatly affect the diagnostic performance of the decision-making model. Currently, most model update methods are based on data replay, that is, a small amount of historical data is added with new data during the update process to prevent catastrophic forgetting. However, the degradation of device components and the changes in task scenarios will lead to frequent updates of the digital twin decision-making model. When the collected data exceeds a certain hardware memory threshold, there will be a problem of insufficient storage space for historical data. Traditional methods have extremely strict requirements for hardware memory and are difficult to be widely applied in actual industrial scenarios. The lack of historical data will cause the model to drift in the direction of the new data distribution, resulting in a deviation between the model and the distribution of historical data.

[0068] Figure 1 is a schematic structural diagram of the digital twin decision-making model provided by an embodiment of the present application. As Figure 1 shown, the digital twin decision-making model includes a feature extractor, a distribution learning unit, and a classifier. The distribution learning unit includes a learnable Gaussian mixture model parameter mapping module and a distribution correction module. The feature extractor is used to extract the features of the input data. The distribution learning unit is used to generate and store the fitted distribution of the input data. The classifier is used to make a classification decision on the features of the input data.

[0069] Exemplarily, first, a preset digital twin decision-making model is constructed. The electronic device inputs the input data under the initial task into the feature extractor to extract general features. The output general features are respectively input into the classifier and the distribution learning unit. When updating the digital twin decision-making model at the t-th stage, Figure 2 is one of the schematic flowcharts of the digital twin decision-making model update method without historical data based on distribution calibration provided by an embodiment of the present application. As Figure 2 shown, the digital twin decision-making model update method without historical data based on distribution calibration includes: step 210, step 220, step 230, step 240, and step 250.

[0070] Step 210: Obtain the input data at the t-th stage;

[0071] It should be noted that when the digital twin decision-making model is updated at the t-th stage, the digital twin decision-making model only obtains the input data at the t-th stage.

[0072] Step 220: Based on the fitted distribution of the historical data from the 1st to the t-1st stage stored in the learnable Gaussian mixture model parameter mapping module, sample the fitted distribution to obtain a target sample, and obtain a biased sample of the historical data from the 1st to the t-1st stage based on the target sample;

[0073] It is easily understandable that the target sample is a sample that can reflect the historical data in the 1 to t-1 stages.

[0074] Step 230: Based on the distribution correction module, correct the biased sample to obtain the corrected historical data in the 1 to t-1 stages;

[0075] Furthermore, the distribution of the biased sample is corrected through the distribution calibration module to align the characteristic distribution of the biased sample with the true historical data in the 1 to t-1 stages.

[0076] Step 240: Input the input data in the t stage into the feature extractor to obtain the feature of the input data in the t stage;

[0077] Step 250: Based on the corrected historical data and the feature of the input data in the t stage, update the classifier to obtain the updated digital twin decision model in the t stage;

[0078] Finally, the updated digital twin decision model in the t stage can adapt to the task scenarios in the 1 to t stages.

[0079] It should be noted that the electronic device simultaneously inputs the input data in the t stage into the distribution learning unit for training, generates and stores the fitted distribution of the input data in the t stage, and serves as the distribution learning unit for reproducing the data in the t stage in the t+1 stage.

[0080] Wherein, t is a positive integer greater than 1.

[0081] According to the method for updating the digital twin decision model based on distribution calibration without historical data provided by the embodiments of the present application, through the learnable Gaussian mixture model parameter mapping module, the fitted distribution of the stored historical data can be sampled to obtain a biased sample without historical data, and the biased sample can be inversely inferred through the distribution calibration module to correct the historical data deviation. In the model update stage, the distribution learning unit can efficiently process the calibration and fusion between the input data and the historical data distribution, reduce the knowledge forgetting of the digital twin decision model, reduce the calculation cost, improve the update efficiency while maintaining the model prediction accuracy, and is applicable to the rapid iteration of the model under different application scenarios.

[0082] In some embodiments, the generating and storing the fitted distribution of the input data in the t stage includes:

[0083] Based on the learnable Gaussian mixture model parameter mapping module, generate the fitted distribution of the input data in the t stage;

[0084] Based on the fitted distribution of the input data in the t stage, update the parameters of the distribution learning unit according to the minimized distribution metric loss function.

[0085] It is easily understood that by minimizing the distribution metric loss, the parameters of the distribution learning unit are updated, so that the sample calibration samples output by the distribution calibration module gradually approach and coincide with the real samples. During the model update process, the deviation distribution of historical data is obtained by sampling the fitting parameters of the Gaussian mixture model, and then corrected by the distribution calibration module to make it similar to the characteristic distribution of real historical data, thus realizing the knowledge replay of historical data.

[0086] In this embodiment, by generating the fitting distribution of the input data at the t stage through the learnable Gaussian mixture model parameter mapping module, the fitting distribution of the input data in the update stage of the digital twin decision model can be stored. By updating the parameters of the distribution learning unit by minimizing the distribution metric loss function, the feature storage of the input data in each stage is realized, enabling the digital twin decision model to recover the characteristic distribution of historical data without saving historical data, reducing the forgetting rate of the digital twin decision model, improving the diagnostic accuracy of the model under different working conditions, and reducing the model update time and memory consumption.

[0087] In some embodiments, the learnable Gaussian mixture model parameter mapping module includes a mapping function and a multivariate Gaussian mixture distribution sampling module, and the mapping function includes a plurality of fully connected layers.

[0088] It should be noted that the learnable Gaussian mixture model parameter mapping module uses the non-linear mapping ability of the mapping function to solve the distribution of the input data. The mapping function maps the input latent feature samples to a higher-dimensional feature space, uses the non-linear mapping ability of the neural network to learn multiple Gaussian components of the data distribution, and can implicitly represent the parameters of the learnable Gaussian mixture model through the mapping function.

[0089] The mapping function includes a plurality of fully connected layers. By configuring the output dimensions of each branch network in the mapping function, and then performing batch normalization and non-linear calculation, a multivariate normal distribution representation can be achieved for each Gaussian component.

[0090] In this embodiment, by using the mapping function and the multivariate Gaussian mixture distribution sampling module, the parameters of the Gaussian mixture model can be effectively learned and adjusted, providing a powerful non-linear modeling ability for the digital twin decision model. By integrating a plurality of fully connected layers into the mapping function, the complex relationships of the input data can be captured more deeply, reducing the dependence of the digital twin decision model on parameter initialization.

[0091] In some embodiments, the generation process of the fitting distribution of the input data includes:

[0092] Input the input data into the feature extractor to obtain the input data features;

[0093] Input the input data features into the mapping function to obtain the mean vector, variance vector, and mixing coefficient vector of the input data;

[0094] Based on the mean vector, variance vector, and mixing coefficient vector, obtain the fitted distribution of the input data.

[0095] It is easy to understand that the parameters of the learnable Gaussian mixture model include mean parameters, variance parameters, and mixing coefficient parameters. Input the input data features into the learnable Gaussian mixture model parameter mapping module, and through the mapping function, output the mean vector μ = {μ (1) , μ (2) , …, μ (k)}, variance vector σ = {σ (1) , σ (2) , …, σ (k)}, and mixing coefficient vector α = {α (1) , α (2) , …, α (k)}, where k is the number of Gaussian distributions. The mean vector, variance vector, and mixing coefficient vector constitute the fitted distribution parameters of the input data.

[0096] In some embodiments, assuming that the dimensions in the latent space are independent of each other, the covariance matrix can be represented by a diagonal matrix, and the diagonal elements represent the variances of each variable in the dimension. The covariance matrix can be expressed as represents the variance of the k-th component in the d-th dimension.

[0097] In this embodiment, by inputting the input data into the feature extractor and obtaining the mean vector, variance vector, and mixing coefficient vector based on the mapping function, the feature information of the input data can be effectively extracted. The mapping function represents the data distributions of different tasks through specific mean and variance matrices, generates the fitted distribution of the input data, reduces the memory overhead and computational cost of the digital twin decision model, can reproduce historical knowledge, and realizes the update of the digital twin decision model without historical data.

[0098] In some embodiments, the distribution correction module is a neural network model constructed based on the U-net network architecture. The calculation formula for the distribution correction by the distribution correction module is as follows:

[0099]

[0100] where g(·) is the calibration process, is the corrected sample, is the biased sample, and θ u is the network weight parameter of the distribution learning unit.

[0101] In some embodiments, when the digital twin decision model is updated at the t-th stage, the distribution calibration module uses the multivariate normal distribution sampling method to resample the fitted distribution of the historical data from stage 1 to stage t-1 stored in the previous learnable Gaussian mixture model parameter mapping module, generating biased samples.

[0102] Furthermore, during the distribution correction process, the electronic device inputs the biased samples into the distribution correction module to extract deep features, concatenates the label information along the channels to the biased samples, and then performs upsampling to reconstruct the corrected samples with the same shape as the biased samples.

[0103] It should be noted that, to improve the model calibration performance, the input of the distribution calibration module is a combination of label information and biased samples, and the label information is the decision result represented by the current data.

[0104] In this embodiment, the distribution correction is performed through a neural network model constructed by the U-net network architecture, which can effectively improve the correction accuracy of the samples. During the correction process, by learning the mapping relationship between the biased samples and the corrected samples, the sample distribution can be automatically adjusted. The distribution calibration module can regenerate a large number of inaccessible historical data by calibrating the biased samples, reducing the catastrophic forgetting during the update of the digital twin decision model.

[0105] In some embodiments, the target samples are sampled based on the K-means clustering algorithm.

[0106] It is easy to understand that representative sample points are selected from the current update stage as target samples through the K-means clustering algorithm. Specifically, the samples closest to each cluster center can be selected as target samples. The fitted distribution parameters of the target samples are stored by the learnable Gaussian mixture model parameter mapping module, and the fitted distribution parameters of the target samples are stored as historical information to update the digital twin decision model, thereby reproducing similar sample points.

[0107] In this embodiment, sampling the target samples through the K-means clustering algorithm can effectively select representative samples, reduce the number of samples and the calculation process, achieve the diversity and uniqueness of the target samples, improve the calculation efficiency of the digital twin decision model, and reduce the storage space of the digital twin decision model.

[0108] In some embodiments, the calculation formula of the minimized distribution metric loss function is as follows:

[0109]

[0110] where L d represents the distribution metric loss function, is the sample after calibration in the t stage, f (t) is the target sample in the t stage, θ g is the network weight parameter of the feature extractor, θ u is the network weight parameter of the distribution learning unit, and n is the total number of samples.

[0111] Figure 3 is the second flow schematic diagram of the digital twin decision model update method based on distribution calibration without historical data provided by the embodiment of the present application. As Figure 3 shown, first, initialize the digital twin decision model, input the labeled initial data set, which contains data samples of multiple categories (such as multiple fault categories like inner ring, outer ring, rolling element, etc.). Input the initial data set into the feature extractor to extract data features and obtain input features. Train the classifier through the input features to learn the classification boundaries of the initial categories, generate the fitting parameters of the input features through the learnable Gaussian mixture model parameter mapping module, and screen and store the fitting parameters of the initial data.

[0112] When the digital twin decision model needs to be updated, input the data samples in the t stage, extract the features of the data in the t stage through the feature extractor. The distribution learning unit restores the sample features from stage 1 to stage t - 1 through the learnable Gaussian mixture model parameter mapping module, combines the data samples in the t stage and the restored sample features from stage 1 to stage t - 1 to update the classifier, making it adapt to the new categories while maintaining the learning ability for historical categories. Finally, update and store the Gaussian mixture model fitting parameters of the data in the t stage through the learnable Gaussian mixture model parameter mapping module.

[0113] In this embodiment, by minimizing the distribution metric loss function, the distribution difference between the calibration sample and the target sample is optimized, the computational cost of the digital twin decision model is reduced, and the efficiency and stability of the digital twin decision model update are improved.

[0114] For the digital twin decision model update method based on distribution calibration without historical data provided by the embodiment of the present application, the execution subject can be the digital twin decision model update device based on distribution calibration without historical data. In the embodiment of the present application, taking the digital twin decision model update device based on distribution calibration without historical data executing the digital twin decision model update method based on distribution calibration without historical data as an example, the digital twin decision model update device based on distribution calibration without historical data provided by the embodiment of the present application is described.

[0115] The embodiment of the present application also provides a digital twin decision model update device based on distribution calibration without historical data. As Figure 4 shown, the digital twin decision model update device based on distribution calibration without historical data includes: a construction module 410, an update module 420, and a storage module 430.

[0116] A construction module 410 for constructing a preset digital twin decision model, the digital twin decision model including a feature extractor, a distribution learning unit, and a classifier, the distribution learning unit including a learnable Gaussian mixture model parameter mapping module and a distribution correction module, the feature extractor being used for extracting features of input data, the distribution learning unit being used for generating and storing a fitted distribution of the input data, and the classifier being used for making a classification decision on the features of the input data;

[0117] An update module 420 for updating the digital twin decision model at stage t, the update including:

[0118] Obtaining the input data at stage t;

[0119] Based on the fitted distribution of the historical data from stage 1 to stage t - 1 stored in the learnable Gaussian mixture model parameter mapping module, sampling the fitted distribution to obtain a target sample, and obtaining a biased sample of the historical data from stage 1 to stage t - 1 based on the target sample;

[0120] Based on the distribution correction module, correcting the biased sample to obtain the corrected historical data from stage 1 to stage t - 1;

[0121] Inputting the input data at stage t into the feature extractor to obtain the feature of the input data at stage t;

[0122] Based on the corrected historical data and the feature of the input data at stage t, updating the classifier to obtain the updated digital twin decision model at stage t;

[0123] A storage module 430 for generating and storing the fitted distribution of the input data at stage t;

[0124] where t is a positive integer greater than 1.

[0125] According to the digital twin decision model update method based on distribution calibration without historical data provided by the embodiments of the present application, through the learnable Gaussian mixture model parameter mapping module, a biased sample can be obtained by sampling the fitted distribution of the stored historical data in the absence of historical data, and through the distribution calibration module, reverse reasoning is performed on the biased sample to correct the historical data deviation. In the model update stage, the distribution learning unit can efficiently process the calibration and fusion between the input data and the historical data distribution, reduce the knowledge forgetting of the digital twin decision model, reduce the calculation cost, improve the update efficiency while maintaining the model prediction accuracy, and is applicable to the rapid iteration of the model in different application scenarios.

[0126] The digital twin decision model update device based on distribution calibration without historical data provided by the embodiments of the present application can achieve Figures 1 to 3For the sake of avoiding repetition, the processes implemented by the embodiments of the method for updating the digital twin decision model based on distribution calibration without historical data are not elaborated here.

[0127] In some embodiments, as Figure 5 shown, an embodiment of the present application also provides an electronic device 500, including a processor 501, a memory 502, and a computer program stored on the memory 502 and executable on the processor 501. When the program is executed by the processor 501, it implements the processes of the above embodiments of the method for updating the digital twin decision model based on distribution calibration without historical data and can achieve the same technical effects. For the sake of avoiding repetition, they are not elaborated here.

[0128] It should be noted that the electronic devices in the embodiments of the present application include the above-mentioned mobile electronic devices and non-mobile electronic devices.

[0129] An embodiment of the present application also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the processes of the above embodiments of the method for updating the digital twin decision model based on distribution calibration without historical data and can achieve the same technical effects. For the sake of avoiding repetition, they are not elaborated here.

[0130] Among them, the processor is the processor in the electronic device 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 disks, or optical discs, etc.

[0131] An embodiment of the present application also provides a computer program product, including a computer program, which implements the above method for updating the digital twin decision model based on distribution calibration without historical data when executed by a processor.

[0132] Among them, the processor is the processor in the electronic device 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 disks, or optical discs, etc.

[0133] Another embodiment of the present application provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the processes of the above embodiments of the method for updating the digital twin decision model based on distribution calibration without historical data and can achieve the same technical effects. For the sake of avoiding repetition, they are not elaborated here.

[0134] It should be understood that the chip mentioned in the embodiments of the present application may also be referred to as a system-on-chip, system chip, chip system, or system-on-chip, etc.

[0135] It should be noted that in this text, the term "including", "comprising", or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article, or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device including that element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in a reverse order according to 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.

[0136] Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the method for updating the digital twin decision model based on distributed calibration without historical data in various embodiments of the present application.

[0137] In the description of the present application, "the first feature", "the second feature" may include one or more of such features.

[0138] In the description of the present application, "a plurality" means two or more.

[0139] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Those of ordinary skill in the art, under the inspiration of the present application and without departing from the spirit and scope protected by the claims of the present application, can also make many forms, all of which fall within the protection scope of the present application.

[0140] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0141] Although the embodiments of this application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and purposes of this application, and the scope of this application is defined by the claims and their equivalents.

Claims

1. A method for updating a digital twin decision model based on distributed calibration without historical data, characterized in that: The method comprises: Constructing a preset digital twin decision model, the digital twin decision model includes a feature extractor, a distribution learning unit and a classifier, the distribution learning unit includes a learnable Gaussian mixture model parameter mapping module and a distribution correction module, the feature extractor is used to extract features of input data, the distribution learning unit is used to generate and store the fitting distribution of the input data, and the classifier is used to make classification decisions on the features of the input data; The digital twin decision model of stage t is updated, and the update includes: Get the input data of stage t; Based on the fitting distribution of the historical data from stage 1 to stage t-1 stored in the learnable Gaussian mixture model parameter mapping module, sampling the fitting distribution to obtain a target sample, and obtaining a bias sample of the historical data from stage 1 to stage t-1 based on the target sample; Correct the biased samples based on the distribution correction module to obtain correction history data from stage 1 to stage t-1; Inputting the input data of the t stage into the feature extractor to obtain the input data features of the t stage; Based on the corrected historical data and the input data features of stage t, the classifier is updated to obtain the updated digital twin decision model of stage t; Generate and store the fitted distribution of the input data at stage t; Wherein, t is a positive integer greater than 1.

2. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 1 is characterized in that: The generating and storing the fitting distribution of the input data of stage t includes: Based on the learnable Gaussian mixture model parameter mapping module, generating a fitting distribution of the input data of the t stage; Based on the fitted distribution of the input data in the t stage, the parameters of the distribution learning unit are updated according to minimizing the distribution metric loss function.

3. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 1 is characterized in that: The learnable Gaussian mixture model parameter mapping module includes a mapping function and a multivariate Gaussian mixture distribution sampling module, and the mapping function includes multiple fully connected layers.

4. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 3 is characterized in that: The generation process of the fitting distribution of the input data includes: Inputting input data into the feature extractor to obtain input data features; Inputting the input data features into the mapping function to obtain a mean vector, a variance vector and a mixing coefficient vector of the input data; Based on the mean vector, variance vector and mixing coefficient vector, a fitting distribution of the input data is obtained.

5. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 1 is characterized in that: The distribution correction module is a neural network model built based on the U-net network architecture. The calculation formula for correction performed by the distribution correction module is as follows: Among them, g(·) is the calibration process, is the corrected sample, is the bias sample, θ u is the network weight parameter of the distribution learning unit.

6. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 1 is characterized in that: The target samples are obtained by sampling based on the K-means clustering algorithm.

7. The method for updating the digital twin decision model based on distributed calibration without historical data according to claim 2 is characterized in that: The calculation formula for minimizing the distribution metric loss function is as follows: Among them, L d represents the distribution metric loss function, is the corrected sample in stage t, f (t) is the target sample at stage t, θ g is the feature extractor network weight parameter, θ u is the network weight parameter of the distribution learning unit, and n is the total number of samples.

8. A digital twin decision model updating device based on distributed calibration without historical data, implemented by the digital twin decision model updating method based on distributed calibration without historical data according to any one of claims 1 to 7, characterized in that: The device comprises: A construction module, used to construct a preset digital twin decision model, the digital twin decision model includes a feature extractor, a distribution learning unit and a classifier, the distribution learning unit includes a learnable Gaussian mixture model parameter mapping module and a distribution correction module, the feature extractor is used to extract features of input data, the distribution learning unit is used to generate and store the fitting distribution of the input data, and the classifier is used to make classification decisions on the features of the input data; An updating module is used to update the digital twin decision model at stage t, wherein the updating includes: Get the input data of stage t; Based on the fitting distribution of the historical data from stage 1 to stage t-1 stored in the learnable Gaussian mixture model parameter mapping module, sampling the fitting distribution to obtain a target sample, and obtaining a bias sample of the historical data from stage 1 to stage t-1 based on the target sample; Correct the biased samples based on the distribution correction module to obtain correction history data from stage 1 to stage t-1; Inputting the input data of the t stage into the feature extractor to obtain the input data features of the t stage; Based on the corrected historical data and the input data features of stage t, the classifier is updated to obtain the updated digital twin decision model of stage t; A storage module, used for generating and storing the fitted distribution of the input data of stage t; Wherein, t is a positive integer greater than 1.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the digital twin decision model updating method based on distributed calibration without historical data as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for updating a digital twin decision model based on distributed calibration without historical data as described in any one of claims 1 to 7 is implemented.

Images