A Dynamic Evolution Method for Digital Twin Model Mechanism Driven by Analog-Digital Fusion

Abstract

This application discloses a method for dynamic evolution of digital twin model mechanisms driven by digital-model fusion, relating to the fields of digital twin and intelligent operation and maintenance technology for industrial equipment. The method includes: constructing a digital twin mechanism evolution model based on mechanism models of various domains and a three-layer mechanism evolution mechanism, including a multi-dimensional model parameter update mechanism, a multi-domain working condition mapping update mechanism, and a multi-physical attribute correlation update mechanism; offline pre-training a generative adversarial network model using complete historical twin data; synchronously inputting complete real-time twin data into the digital twin mechanism evolution model and the pre-trained generative adversarial network model, updating and achieving a closed loop of bidirectional collaborative evolution between the two models based on the deviation between their outputs. This application enables dynamic adaptive evolution of digital twin models for textile equipment, effectively reducing equipment failure rates, extending equipment service life, and supporting the intelligent transformation of complex textile equipment.

Description

Technical Field

[0001] This application relates to the field of digital twin and intelligent operation and maintenance technology of industrial equipment, specifically to a dynamic evolution method of digital twin model mechanism driven by digital-analog fusion. Background Technology

[0002] The statements in this section are merely to provide background information in relation to this application to aid in understanding it, and such background information does not necessarily constitute prior art.

[0003] Textile equipment such as texturing machines, knitting machines, and spinning machines are highly integrated advanced engineering equipment. Their production mode typically involves continuous, 24 / 7 operation, often under heavy load and high-intensity conditions. Wear or performance degradation of key components, such as the false twister friction disc, the winder bearing assembly, and the heating elements of the hot box, not only reduces product quality but can also cause unexpected equipment downtime, resulting in significant economic losses for manufacturing enterprises. The core of using digital twin technology for fault prediction and proactive maintenance of complex textile equipment lies in constructing a high-fidelity digital model that accurately reflects the internal operating mechanisms of the equipment and maintaining long-term consistency between the virtual and real models as the equipment's condition changes.

[0004] However, existing digital twin methods still have significant shortcomings in terms of the dynamic evolution of mechanisms. Most studies focus on local model updates under single physical properties or single operating conditions, lacking a systematic mechanism for the evolution of mechanisms under multi-dimensional, multi-domain, and multi-physical property coupling conditions of complex equipment. Data-driven models and mechanism simulation models often run independently, lacking an effective correlation and collaborative correction mechanism between the two, resulting in lag in model response when data is abnormal, or inability to correct accumulated mechanism deviations in real time. In addition, some methods rely on expert experience for manual parameter tuning, which is inefficient and difficult to scale up. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method for dynamic evolution of digital twin model mechanisms driven by digital-analog fusion. This method involves collecting multi-source heterogeneous operational data from complex textile equipment and fusing it into a complete twin dataset. Based on this dataset, a three-layer mechanism evolution mechanism is constructed, consisting of multi-dimensional parameter iteration, multi-domain working condition mapping, and collaborative updating of multi-physical attribute correlation. Simultaneously, an adversarial pre-training and unsupervised online iterative improvement process for a generative adversarial network (GAN) discriminant model (GAN model) is established. Finally, the deviation between the mechanism simulation output and the data prediction output is used as a closed-loop driving force to continuously correct the digital twin mechanism evolution model.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] According to the first aspect of this application, a method for dynamic evolution of digital twin model mechanisms driven by data-model fusion is provided, comprising: fusing and dimensionality-reducing historical operation process data, historical operation status data, and historical fault data of equipment to form complete twin historical data with fault labels; constructing mechanism models of equipment in various domains, and constructing a digital twin mechanism evolution model based on the mechanism models of each domain and a three-layer mechanism evolution mechanism, wherein the three-layer mechanism evolution mechanism includes a multi-dimensional model parameter update mechanism with data flow coupling and linkage, a multi-domain working condition mapping update mechanism, and a multi-physical attribute correlation update mechanism, wherein the multi-dimensional model parameter update mechanism is used to iteratively update geometry-behavior-rule parameters, the multi-domain working condition mapping update mechanism is used to update the coupling weights corresponding to the mechanism models of each domain according to the updated geometry-behavior-rule parameters, and solve the multi-domain coupling state values ​​of each physical quantity based on the coupling weights, and the multi-physical attribute correlation update mechanism is used to update the coupling weights corresponding to the mechanism models of each domain according to the updated geometry-behavior-rule parameters, and solve the multi-domain coupling state values ​​of each physical quantity based on the preset coupling weights, and the multi-physical attribute correlation update mechanism is used to update the multi-domain coupling state values ​​of each physical quantity according to the updated geometry-behavior-rule parameters. The multi-domain coupling state values ​​of physical quantities determine the changes in corresponding physical characteristics, and the mechanism parameters of the corresponding domain mechanism model are updated according to the changes in physical characteristics. A generative adversarial network (GAN) model is constructed, and the GAN model is pre-trained offline using the complete twin historical data. The GAN model includes a generator and a discriminator. The discriminator is used to characterize the data distribution benchmark of the actual operating state of the equipment, and the generator is used to output the prediction results of the equipment operating state. The real-time operation process data, real-time operating state data, and real-time fault data of the equipment are fused and dimensionality reduced to obtain complete twin real-time data, which is synchronously input into the digital twin mechanism evolution model and the pre-trained GAN model. According to the deviation between the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the output of the generator, the digital twin mechanism evolution model and the generator are updated in reverse, forming a closed loop of bidirectional collaborative evolution between the digital twin mechanism evolution model and the GAN model.

[0008] Preferably, the fusion and dimensionality reduction processing includes: adaptively determining the fusion weight of each modality according to the completeness of each modality data, and performing weighted fusion on each modality data, wherein each modality data is collected by multi-acquisition device collaborative measurement, heterogeneous communication protocol parsing and business system data import, including mechanical modality data, electrical modality data and fault label data; performing kernel mapping dimensionality reduction on the fused high-dimensional heterogeneous data, projecting it to a unified low-dimensional shared latent space to obtain standardized complete twin feature vectors; wherein, for historical operation process data, historical operation status data and historical fault data, before fusion, for scarce fault conditions with sample size below a preset equilibrium threshold, virtual operating condition samples are generated in batches based on a small number of real fault samples of the corresponding operating condition by an operating condition simulation generation operator to supplement the sample distribution of the dataset.

[0009] Preferably, the digital twin mechanism evolution model is constructed by connecting a multi-dimensional model parameter update module, a multi-domain working condition mapping update module, and a multi-physical attribute correlation update module. Specifically: the multi-dimensional model parameter update module is used to run the multi-dimensional model parameter update mechanism; the multi-domain working condition mapping update module is used to run the multi-domain working condition mapping update mechanism, and it includes a mechanical domain mechanism model, an electrical domain mechanism model, an electromagnetic domain mechanism model, and a mechanism simulation solution unit. The mechanism simulation solution unit is used to solve the multi-domain coupling state values ​​of each physical quantity based on the mechanical domain mechanism model, the electrical domain mechanism model, the electromagnetic domain mechanism model, and the coupling weights corresponding to each domain mechanism model; the multi-physical attribute correlation update module is used to run the multi-physical attribute correlation update mechanism.

[0010] Preferably, the digital twin mechanism evolution model is updated in reverse based on the deviation between the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator output. This includes: calculating the L2 deviation between the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator output; when the L2 deviation is greater than a preset deviation threshold, updating the digital twin mechanism evolution model according to the three-layer mechanism evolution mechanism in the following manner: the multi-dimensional model parameter update mechanism includes: firstly, using geometric behavior rule feature matching loss to update the geometric behavior rule parameters within the execution layer. Local iterative updates yield intermediate geometric behavior rule parameters. These parameters are then input into a multi-domain working condition mapping update mechanism and a multi-physical attribute correlation update mechanism for forward inference. Based on the forward inference results, minimizing the L2 deviation is used as the optimization objective. The gradient of the L2 deviation relative to the intermediate geometric behavior rule parameters is calculated using a gradient descent algorithm. Gradient compensation correction is then applied to the intermediate geometric behavior rule parameters along the gradient descent direction to obtain the final effective geometric behavior rule parameters. These final effective geometric behavior rule parameters are then written back to the multi-dimensional model parameter update module as the next... The geometric behavior rule parameters are iterated over time; the multi-domain working condition mapping update mechanism includes: calculating the interaction strength between mechanical, electrical, and electromagnetic components based on the updated geometric-behavior-rule parameter mapping; adaptively updating the coupling weights corresponding to the mechanism models of each domain; solving the multi-domain coupling state values ​​of each physical quantity based on the updated coupling weights and the mechanism models of each domain; if the absolute value of the first error between the multi-domain coupling state value of each physical quantity and the corresponding state variable in the complete twin real-time data is greater than a preset accuracy threshold, then iteratively adjusting the mechanism parameters of the mechanism models of each domain and resolving until the absolute value of the first error is less than or equal to the preset accuracy threshold; the multi-physical attribute correlation update mechanism includes: extracting static feature sets, dynamic feature sets, and elasticity feature sets respectively, each feature set including the corresponding physical features; calculating the nonlinear coupling correlation strength between any two physical features; constructing a multi-physical attribute correlation matrix based on the nonlinear coupling correlation strength of all physical feature pairs; selecting physical quantities with nonlinear coupling strength exceeding a preset correlation strength threshold as strongly correlated features; and updating the mechanism parameters and coupling weights of the domain mechanism models corresponding to the attributes of the strongly correlated features.

[0011] Preferably, the multi-physical attribute correlation update mechanism further includes: after the multi-domain coupling state value of each physical quantity is updated, the physical quantity that has changed is determined, and the corresponding nonlinear coupling correlation strength is recalculated to update the multi-physical attribute correlation matrix; based on the updated multi-physical attribute correlation matrix, new strongly correlated features are selected, and the mechanism parameters and coupling weights of the domain mechanism model corresponding to the attributes of the new strongly correlated features are updated, as are the mechanism parameters and coupling weights of the domain mechanism model corresponding to the attributes of the physical quantity that has changed.

[0012] Preferably, the generative adversarial network model is pre-trained offline using the complete twin historical data, including: performing clustering, regression, and generalization augmentation preprocessing on the complete twin historical data sequentially to form a training dataset; based on the training dataset, using a preset adversarial loss function as the optimization objective, alternatingly iteratively training the generator and the discriminator until convergence, thus completing the offline pre-training of the generative adversarial network model; wherein, the input of the generator is a random noise vector, and the output of the generator is a prediction result of the equipment operating status generated based on the noise; the input of the discriminator is the complete twin historical feature vector in the training dataset, and the output of the discriminator is the probability that its input belongs to a real twin sample.

[0013] Preferably, updating the generator in reverse based on the deviation between the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator output includes: calculating the L2 deviation based on the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator output; when the L2 deviation is greater than a preset deviation threshold, using the L2 deviation as an additional input signal to the generator, and inputting it into the generator along with the complete twin real-time data, with the optimization objective of minimizing the error between the generator output and the corresponding state variables in the complete twin real-time data, obtaining the loss gradient of the error through the backpropagation algorithm, and updating the generator's network weight parameters along the gradient descent direction.

[0014] Preferably, the method further includes: after synchronously inputting the complete twin real-time data into the digital twin mechanism evolution model and the pre-trained generative adversarial network model, using the complete twin real-time data to perform online iterative updates on the discriminator in the following manner: using the complete twin real-time data as positive samples and the output of the digital twin mechanism evolution model as negative samples, constructing an online discriminant loss function; using the maximization of the online discriminant loss function as the optimization objective, iteratively updating the network parameters of the discriminator using the gradient descent method.

[0015] Preferably, the equipment is complex textile equipment, including a texturing machine, a knitting machine, or a spinning machine; the multimodal data, consisting of process data, operating status data, and fault data, includes yarn flow speed, winding tension, hot box temperature, motor speed, current load, false twister operating frequency, and fault label data for yarn breakage, winding abnormalities, and false twister slippage; the geometric-behavior-rule parameters include a set of geometric feature parameters, a set of motion behavior parameters, and a set of operating rule parameters, wherein: the geometric feature parameter set includes the diameter of the false twister friction disc, the size of the hot box heating plate, and the spiral angle of the windinger groove cylinder; the motion behavior parameter set includes kinematic relationship parameters between the rotational speed of the false twister and the twisting angle, and dynamic relationship parameters between the winding tension and the winding speed; the operating rule parameter set includes a yarn breakage alarm threshold, speed matching rules, and temperature gradient control rules.

[0016] According to a second aspect of this application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of the first aspect of this application.

[0017] According to a third aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method of the first aspect of this application.

[0018] Compared with the prior art, this application has the following advantages:

[0019] This application constructs a three-layer mechanistic evolution mechanism, consisting of multi-dimensional model parameter gradient iterative update, multi-domain working condition coupling mapping update, and multi-physical attribute correlation mutual information update, forming a three-dimensional evolution framework of dimension-domain-attribute. By combining offline pre-training of the generative adversarial network model with real-time unsupervised online iterative improvement, a closed-loop correction loop of model-data fusion is established, encompassing mechanistic simulation output, digital-driven prediction output, deviation calculation, and model correction. Using the L2 deviation between the mechanistic simulation output and the data prediction output as the evolutionary driving force, the entire digital twin model, including the mechanistic and data layers, is updated based on deviation gradient back-compensation. The entire process achieves adaptive dynamic evolution of the digital twin mechanistic evolution model without manual parameter tuning. This application organically integrates the data-driven model and the mechanistic simulation model, overcoming the shortcomings of their independent operation and lack of mutual correction mechanism, and realizing the adaptive dynamic evolution of the complex equipment digital twin mechanistic evolution model during long-term operation. Compared with traditional static twin models without correction mechanisms, this application can effectively offset the virtual-to-real drift error of the model under long-term continuous operation conditions, significantly improve the accuracy of equipment fault early warning and equipment service life, and provide effective technical support for the intelligent transformation of complex textile equipment. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0021] Figure 1 This is a flowchart illustrating the dynamic evolution method of digital twin model mechanism driven by digital-analog fusion according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of the operational architecture for the dynamic evolution of the digital twin model mechanism driven by the fusion of digital and analog models according to embodiments of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the technical features involved in the various embodiments described below can be combined with each other as long as they do not conflict with each other.

[0024] Figure 1 This is a flowchart illustrating the dynamic evolution method of digital twin model mechanism driven by digital-analog fusion according to an embodiment of this application. Figure 2 This is a schematic diagram of the operational architecture for the dynamic evolution of the digital twin model mechanism driven by analog-digital fusion according to embodiments of this application. The following will be combined with... Figure 1 and Figure 2 This application details the dynamic evolution method for the mechanism of digital twin models.

[0025] like Figure 1 As shown, the method includes:

[0026] Step 1: Integrate and reduce the historical operation data, historical operation status data, and historical fault data of the equipment to form complete twin historical data with fault labels.

[0027] In some embodiments, the fusion and dimensionality reduction processing specifically includes:

[0028] 1. The collected equipment operation process data, operation status data, and fault data are divided into mechanical modal data, electrical modal data, and fault label data according to data attributes, forming a multimodal raw heterogeneous dataset of the equipment. The fusion weight of each modality is adaptively determined according to the completeness of each modality data, and then the data of each modality is weighted and fused. The modal data is collected using collaborative measurement of multiple acquisition devices, heterogeneous communication protocol parsing, and data import from business systems.

[0029] In some embodiments, the formula for weighted fusion of multimodal data is as follows:

[0030] (1)

[0031] in, This indicates the complete twin characteristics after fusion; Represents mechanical modal data such as mechanical vibration and rotational speed; Represents electrical mode data such as current, voltage, and electromagnetic parameters; This indicates fault label data; , , This represents the corresponding dynamic confidence weight (fusion weight). The weight values ​​are adaptively updated based on the completeness of the corresponding modality data. That is, the weights are allocated according to the proportion of valid samples without missing data in each modality within the time-series sliding window. When a modality sample is missing, the corresponding weight is reset to zero, and the remaining weights are renormalized. The weights are refreshed after each round of twin data collection is completed.

[0032] In some embodiments, for historical operation process data, historical operation status data, and historical fault data, before fusion, simulated virtual sample expansion processing is performed on scarce fault conditions with sample sizes below a preset equilibrium threshold. This involves generating virtual operating condition samples in batches based on a small number of real fault samples from the corresponding operating conditions using a condition simulation generation operator (the type of simulation operator is determined according to the specific application scenario, for example, a simulation operator for fault inference of complex textile equipment), thus supplementing the sample distribution of the dataset. The formula for expanding the simulated virtual sample can be expressed as:

[0033] (2)

[0034] in, This indicates the operator for generating the simulation of the working condition. This represents a small set of real, scarce multimodal fault samples collected. This represents the set of virtual fault condition samples obtained from the output of the simulation generation operator. This represents the control parameters for the simulated operating conditions. In this embodiment, a small number of real fault samples are input to generate virtual operating condition samples in batches, balancing the distribution of the original data samples.

[0035] 2. Utilize kernel mapping functions to perform kernel mapping dimensionality reduction on the fused linear or nonlinear high-dimensional heterogeneous data, project it onto a unified low-dimensional shared latent space, and obtain standardized complete twin feature vectors.

[0036] In some embodiments, the kernel mapping function satisfies formula (3):

[0037] (3)

[0038] in, This represents the covariance operation, which characterizes the degree of statistical correlation between two sets of projected features; This represents the variance operation, characterizing the degree of dispersion of the features themselves after projection. These represent two sets of original heterogeneous datasets of equipment multimodal data from different sources. The different sources refer to the different data sources of the original heterogeneous datasets of equipment multimodal data, which are respectively from mechanical mode, electrical mode, and fault label mode. These correspond to the mechanical vibration and speed data, current, voltage and electromagnetic data, and fault label business data of the equipment. They are heterogeneous datasets obtained from different dimensions and different collection channels of the equipment. These are the projection transformation weight vectors corresponding to the two datasets. This represents the kernel mapping function, used to project original high-dimensional heterogeneous data onto a unified low-dimensional shared latent space, reducing the dimensionality of data computation. This formula achieves alignment and dimensionality reduction of multi-source heterogeneous twin features by maximizing the correlation coefficients of projected features from different modalities.

[0039] Step 2: Construct mechanism models for each domain of the equipment, and build a digital twin mechanism evolution model based on the mechanism models for each domain and the three-layer mechanism evolution mechanism. The three-layer mechanism evolution mechanism includes a multi-dimensional model parameter update mechanism with data flow coupling and linkage, a multi-domain working condition mapping update mechanism, and a multi-physical attribute correlation update mechanism. The multi-dimensional model parameter update mechanism is used to iteratively update the geometry-behavior-rule parameters. The multi-domain working condition mapping update mechanism is used to update the coupling weights corresponding to the mechanism models of each domain according to the updated geometry-behavior-rule parameters, and solve the preset multi-domain coupling state values ​​of each physical quantity based on the coupling weights. The multi-physical attribute correlation update mechanism is used to determine the changes in the corresponding physical characteristics according to the multi-domain coupling state values ​​of each physical quantity, and update the mechanism parameters of the corresponding domain mechanism model according to the changes in the physical characteristics.

[0040] In some embodiments, the digital twin mechanism evolution model is constructed by cascading a multi-dimensional model parameter update module, a multi-domain working condition mapping update module, and a multi-physical attribute correlation update module. Specifically: the multi-dimensional model parameter update module is used to run the multi-dimensional model parameter update mechanism. The multi-domain working condition mapping update module is used to run the multi-domain working condition mapping update mechanism, and it includes a mechanical domain mechanism model, an electrical domain mechanism model, an electromagnetic domain mechanism model, and a mechanism simulation solution unit. The mechanism simulation solution unit is used to solve for the multi-domain coupling state values ​​of each physical quantity based on the mechanical domain mechanism model, the electrical domain mechanism model, the electromagnetic domain mechanism model, and the coupling weights corresponding to each domain mechanism model. The multi-physical attribute correlation update module is used to run the multi-physical attribute correlation update mechanism. For example... Figure 2 As shown, specifically:

[0041] 1. Multi-dimensional model parameter update mechanism:

[0042] The parameters of the time-twin multidimensional model are uniformly represented as ;

[0043] in, express Always equipped with a complete multi-dimensional mechanism parameter set of digital twin (i.e.) The geometric-behavior-rule parameters at any given moment are used to uniformly characterize all the mechanistic information of the equipment's geometric shape, motion behavior, and operational constraints at the current moment. express The set of geometric feature parameters at time t. express The set of motion behavior parameters of the equipment at any given time. express The set of operating rules parameters for the equipment at any given time.

[0044] In some embodiments, the geometric feature parameter set may include geometric dimensional parameters of equipment components, such as the diameter of the false twister friction disc, the size of the heating plate in the hot box, and the helix angle of the winding drum in textile equipment. The motion behavior parameter set may include parameters related to the dynamic relationships of the motion of equipment components, such as the kinematic relationship between the rotational speed of the false twister and the twisting angle, and the dynamic relationship between the winding tension and the winding speed. The operation rule parameter set may include equipment operation constraints and control threshold parameters, such as the yarn breakage alarm threshold, speed matching rules, and temperature gradient control rules.

[0045] In some embodiments, the geometry-behavior-rule parameters can be updated according to formula (4):

[0046] (4)

[0047] in, express Intermediate geometric behavior rule parameters at time +1 (hereinafter referred to as intermediate parameters); Indicates the update step size; express Complete twin characteristics of time; This represents the gradient operator, i.e., the loss function. right Loss reduction obtained by taking partial derivatives The fastest parameter update direction; This represents the constructed geometry-behavior-rule feature matching loss function. Using the mean squared error form, its expression is:

[0048] (5)

[0049] in, This represents the total dimension of the features in the low-dimensional shared latent space; this dimension is obtained through the kernel mapping function described above. The output feature dimension set after projection is a pre-defined hyperparameter and is independent of the original multimodal data dimension. For feature indexes in a low-dimensional shared latent space. The simulation features output from the digital twin mechanism evolution model are projected onto the low-dimensional shared latent space via a kernel mapping function. 3D simulation features; To obtain the complete twin features, the first one is projected onto the low-dimensional shared latent space using the same kernel mapping function. True twin characteristics.

[0050] The loss function calculates the matching error between simulated features and real twin features in the latent feature space, and is used to implement intra-layer local iterative parameter tuning of geometry-behavior-rule parameters in formula (4), outputting the locally updated intermediate parameters. ; This is a local iterative loss, distinct from the global L2 bias. After completing the local parameter tuning, use this intermediate parameter. The global L2 norm deviation is calculated as the starting point, and then the global deviation compensation and correction are carried out based on this L2 norm deviation using formula (12) to obtain the final effective parameters. .

[0051] This loss function is obtained by summing the feature matching errors between the simulation features output by the digital twin mechanism evolution model and the complete twin features fused in real time from the equipment sensors (here, simulation features and complete twin features are feature vectors, not direct physical quantities). The mean squared error is used to measure the overall deviation of geometric, behavioral, and rule-based multi-dimensional features. (Geometric-Behavioral-Rule Feature Matching Loss Function) This characterizes the iterative loss within the multi-dimensional model parameter layer, completing local parameter tuning within the layer; and the global L2 bias. This represents the total deviation of the closed-loop model fusion, completing the global correction of the closed-loop model fusion. Gradient descent is used to synchronously refresh the parameters of three types of models: geometric features, motion behavior, and operating rules. First, formula (4) is executed for local parameter tuning within the layer, and then the global L2 deviation is calculated. If the L2 norm deviation If the deviation exceeds the preset threshold, then formula (12) is executed again, based on the L2 norm deviation. For intermediate parameters Perform global deviation compensation correction and output the final effective parameters. (i.e., the geometric-behavioral-rule parameters at time t+1). If the L2 norm deviation... If the deviation is less than or equal to the preset deviation threshold, then global compensation correction will not be performed, and the value will be directly set to zero. = The intermediate parameters are directly used as the final effective parameters.

[0052] 2. Multi-domain operating condition mapping and update mechanism:

[0053] The mechanism models of the equipment in various fields include mechanical, electrical and electromagnetic mechanism models. The specific methods for establishing these models are well-known technologies in this field and will not be elaborated here.

[0054] In some embodiments, the single-physics elasticity equilibrium control equations satisfy equations (6) and (7):

[0055] (6)

[0056] (7)

[0057] in, Represents the stress tensor. Represents the vector differential operator. Represents the divergence of the stress tensor. This represents the vector of external loads acting on equipment components, including wire tension, friction disc compressive force, and axial mechanical load of the winding mechanism. Represents the stiffness matrix. Indicates total strain. It represents thermal strain.

[0058] The coupling mapping of the multi-domain working conditions is solved according to the coupling weight of each domain. The coupling mapping of the mechanical, electrical and electromagnetic multi-domain working conditions is shown in formula (8):

[0059] (8)

[0060] in, This represents a multi-domain coupled comprehensive working condition mapping function, used to realize the working condition coupling operation between the mechanical, electrical, and electromagnetic domains, and output the comprehensive working condition characteristics, i.e., the preset multi-domain coupling state values ​​of each physical quantity. In some embodiments, in formula (8) The output comprehensive operating condition characteristics include multi-domain coupled state variables (multi-domain coupled state values) such as coupled stress, component deformation, operating current, excitation magnetic field, and mechanism output load. These physical quantities are the basic physical quantities constituting the mechanism simulation sample. By splicing this group of multi-domain coupled state variables at the same moment, a single mechanism simulation sample is obtained. The mechanism simulation samples corresponding to multiple time points are summarized to form a mechanism simulation sample set. . This represents a mechanical domain mechanism model, characterizing the equipment's structure, motion, and related mechanical parameters. This represents the electrical domain mechanism model, characterizing the equipment drive circuit and related electrical control parameters. This represents the electromagnetic domain mechanism model, characterizing the relevant parameters of the equipment's electromagnetic field excitation. Represents the characteristic quantities of the mechanical domain operating conditions. Represents the characteristic quantities of electrical domain operating conditions. Represents the characteristic quantities of electromagnetic domain operating conditions. , , This represents the coupling weight of the corresponding domain mechanism model.

[0061] In some embodiments, the multi-dimensional model parameter update module outputs an updated set of geometric feature parameters. Motion behavior parameter set , Run rule parameter set The input is directly fed into the multi-domain working condition mapping and update module; based on the preset changes in geometric dimensions, motion constraints, and operating rules, the multi-domain working condition mapping and update module maps and calculates the interaction strength between mechanical, electrical, and electromagnetic components, and adaptively updates the coupling weights. , , For example, changing the diameter (geometric parameter) of the false twister friction disc directly alters the mechanical friction load, thereby increasing the coupling weight of the mechanical domain mechanism module.

[0062] The mechanism simulation solver solves the unit solution formula (8), which yields the multi-domain coupled state variables such as equipment coupling stress, component deformation, working current, excitation magnetic field, and mechanism output load. An accuracy threshold δ is set, and if the simulation error satisfies... The mechanistic parameters of the mechanical, electrical, and electromagnetic domain mechanistic models are iteratively adjusted. These mechanistic parameters can refer to internal mechanistic parameters in each domain, such as stiffness and load parameters in the mechanical domain; electrical parameters of the frequency converter and motor in the electrical domain; and magnetic field excitation parameters in the electromagnetic domain. The solution is then recalculated until the accuracy constraints are met. The simulation value representing the comprehensive working condition characteristics of the multi-domain coupled simulation output (obtained from formula (8)); This represents the measured true value of the comprehensive operating condition characteristics obtained from the corresponding data collected during the actual operation of the equipment. The raw data collected by the sensor is then processed with... Dimensionally aligned feature mapping and computational processing yield the measured true values ​​of comprehensive working condition features that are completely consistent with the simulation output dimensions.

[0063] 3. Multi-physical attribute correlation update mechanism:

[0064] Mutual information is used to quantify the coupling strength of static, dynamic, and elastic mechanical characteristics, and the correlation degree is calculated according to the formula:

[0065] (9)

[0066] in, The static characteristic set can include frame static load, component static stress, and assembly preload. The set of dynamic and elastic mechanical features can include the rotational dynamics of the false twister, yarn tension fluctuations, impact loads on the mechanism, elastic deformation of the winding components, yarn elastic modulus, and thermo-elastic deformation. This feature set is obtained from historical data from equipment sensors after feature extraction and preprocessing. Indicates the feature set With feature set The degree of correlation between them; Physical characteristics representing the set of static features. Physical characteristics representing the feature set of dynamics and elasticity; Representing physical characteristics Physical characteristics The joint probability density; , Physical characteristics Physical characteristics Each has its own marginal probability density.

[0067] Calculate the feature set according to formula (9) With feature set Nonlinear coupling correlation strength between any two physical features , A larger numerical value indicates a stronger mutual influence between the two physical quantities. A multi-physical attribute correlation matrix is ​​constructed based on the calculated correlation strength of all physical feature pairs.

[0068] Filtering satisfies Strong correlation features This represents the preset threshold for association strength. Based on the mutual information coupling relationship of strongly associated features, the mechanism parameters and coupling weights of the domain mechanism model corresponding to the attributes of strongly associated features are updated.

[0069] In some embodiments, updates to the comprehensive operating condition characteristics lead to changes in the numerical values ​​of static, dynamic, and elastic physical characteristics. After these physical characteristics (physical quantities) change, in some embodiments, the changes in physical characteristic values ​​can be triggered by two sources: one is from the multi-domain operating condition mapping update mechanism, which causes changes in the numerical values ​​of static, dynamic, and elastic physical characteristics after the comprehensive operating condition characteristics are updated; the other is from changes in physical attribute parameters detected by real-time operational data from the equipment side. It is necessary to recalculate the mutual information (correlation degree) between pairs of physical characteristics and refresh the multi-physical attribute correlation degree matrix. Then, strongly correlated characteristics with coupling strength exceeding a threshold are screened again, triggering synchronous correction of the corresponding domain mechanism model and coupling weights. Simultaneously, based on the changes in physical attribute parameters detected by real-time operational data, the mechanical, electrical, and electromagnetic sub-mechanism models associated with the changed physical characteristics are automatically updated synchronously, ensuring the coupling consistency between the multi-physical attribute models.

[0070] Step 3: Construct a Generative Adversarial Network (GAN) model and pre-train the GAN model offline using complete twin historical data. The GAN model includes a generator and a discriminator. The discriminator is used to characterize the data distribution benchmark of the actual operating state of the equipment, and the generator is used to output the prediction results of the operating state of the equipment.

[0071] In some embodiments, step 3 may include:

[0072] Clustering, regression, and generalization augmentation preprocessing are performed sequentially on the complete twin historical data with fault labels. This preprocessing procedure is a well-known technique and will not be described in detail here.

[0073] Using the standard adversarial loss function as the optimization objective, the generator G and discriminator of the generative adversarial network (GAN) model are trained iteratively and alternately to complete the offline pre-training of the GAN. During pre-training, the input to the discriminator D includes: ① real twin samples. ① The feature vectors in the complete twin historical data after fusion and dimensionality reduction are the multimodal features of the actual equipment operation; ② The simulated twin samples output by the generator G are feature vectors that mimic the actual operating state of the equipment. The input of the generator G is a random noise vector.

[0074] Standard adversarial loss function As shown in formula (10):

[0075] (10)

[0076] Among them, D This represents the output function of the discriminator, which outputs the probability that the input sample belongs to a true twin sample. This represents the generator's output function, which outputs a simulated twin sample generated based on the noise. This represents a noise vector taken from a random noise distribution; This means maximizing the discriminator loss and optimizing the discriminator to distinguish real samples from generated samples as much as possible. This means minimizing the generator loss and optimizing the generator to make the generated samples as close as possible to the real samples. This represents the distribution of truly complete twin data. Indicates to Find the expected value; Represents a random noise distribution. Indicates to Find the expected value.

[0077] Step 4: Fuse and reduce the real-time operation data, real-time operation status data, and real-time fault data of the equipment to obtain complete twin real-time data. Simultaneously input this complete twin real-time data into the digital twin mechanism evolution model and the pre-trained generative adversarial network model. Then, based on the deviation between the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator's output, update the digital twin mechanism evolution model and the generator in reverse, forming a closed loop of bidirectional collaborative evolution between the digital twin mechanism evolution model and the generative adversarial network model.

[0078] During the online operation phase, the input to the digital twin mechanism evolution model is complete real-time twin data, which is obtained by fusing and reducing the dimensions of the equipment's real-time operation process data, real-time operation status data, and real-time fault data. The digital twin mechanism evolution model outputs mechanism simulation samples. The multi-domain coupled state values ​​of each physical quantity serve two purposes: firstly, as negative samples for the online update of the discriminator; and secondly, to calculate the deviation from the generator output, thereby achieving reverse parameter updates. The generator's input is complete real-time twin data, and its output is the generator's predicted output sample. During the online discriminator update phase, the generator parameters remain fixed, and the generator's own parameters are updated only based on the deviation between the generator output and the digital twin mechanism evolution model output. The generator output is not used as a negative sample for the discriminator. The discriminator's input contains two types of sample vectors: one type is real twin samples from the complete real-time twin data. One type serves as a positive sample; the other type consists of mechanism simulation samples output by the digital twin mechanism evolution model. The input sample is used as a negative sample; the discriminator outputs the probability that the input sample belongs to the real twin sample; the discriminator iteratively updates its own parameters based on the distribution difference between the real twin sample and the mechanism simulation sample, maintains the data distribution benchmark, indirectly constrains the generator, and does not directly calculate the L2 bias.

[0079] In step 4, the fusion and dimensionality reduction processes are the same as in step 1, but without virtual condition sample augmentation. During the online operation phase, complete twin real-time data generally does not carry fault labels, enabling unsupervised closed-loop evolution.

[0080] In some embodiments, after the generative adversarial network model has converged through pre-training, the generator G can output prediction results of equipment operating conditions based on complete twin real-time data. This can include the predicted values ​​of each physical quantity and the probability distribution of failure modes.

[0081] In some embodiments, step 4 may include:

[0082] Complete real-time twin data is synchronously input into the digital twin mechanism evolution model and the pre-trained generative adversarial network model. The mechanism simulation results are obtained by solving the digital twin mechanism evolution model. This includes the solved multi-physics simulation vectors of the equipment (i.e., comprehensive operating condition characteristics); and the prediction results driven by the generator G of the generative adversarial network model. This includes the data-driven operating condition prediction vector output by generator G (predicted values ​​of various physical quantities, including stress, deformation, current, magnetic field, etc.). Both are aligned and unified in terms of dimensions and dimensions, serving as equipment multi-operating condition state vectors. If there are inconsistencies in dimensions and dimensions, they are aligned through normalization mapping, and then the L2 norm deviation is calculated.

[0083] Calculate the L2 norm deviation according to formula (10):

[0084] (11)

[0085] in, for Simulation results of the mechanism at time and Prediction results of time The L2 norm deviation.

[0086] When the L2 deviation exceeds a preset deviation threshold, the digital twin mechanism evolution model is updated based on the L2 deviation through reverse compensation. Specifically, the update method is as follows: using the L2 deviation... As a loss function, the evolution model of digital twin mechanisms The gradient can be obtained by direct differentiation. With deviation compensation coefficient As the update step size, the parameters of the digital twin mechanism evolution model are iteratively corrected using gradient descent, that is, the geometry-behavior-rule parameters are updated according to formula (11):

[0087] (12)

[0088] in, This is the deviation compensation coefficient.

[0089] When the L2 deviation exceeds a preset deviation threshold, the generator is updated based on the L2 deviation while updating the digital twin mechanism evolution model. The specific update method is as follows:

[0090] The L2 bias is used as an additional input signal to the generator. The complete twin real-time data is used as the basic input to the generator. The L2 bias is used as an additional input signal to the generator. The two signals are input to the generator together. The optimization objective is to minimize the error between the generator output and the corresponding state variable in the complete twin real-time data. The network weight parameters of the generator are updated along the gradient direction of the L2 bias through the backpropagation algorithm.

[0091] The system continuously collects real-time operational data of the physical equipment, and iteratively executes the entire process of data fusion, mechanism updates, generative adversarial network (GAN) model prediction, and deviation correction, forming a closed loop of simulation output - prediction output - deviation calculation - model correction. This offsets the virtual-real drift error caused by long-term model operation, enabling the digital twin model to continuously adapt to dynamic mechanism evolution. The entire closed-loop evolution process is executed automatically at fixed time steps. At each time step, mechanism simulation output, digital-driven prediction output, and deviation calculation are performed, and only when the calculated L2 deviation is corrected... The digital twin mechanism evolution model correction operation is only performed when the deviation exceeds a preset threshold. If the deviation does not exceed the threshold, model correction is skipped, and the process proceeds directly to the next time step. The closed-loop iteration termination condition only needs to satisfy one of two conditions: one is the L2 deviation. The model achieves two objectives: first, the deviation must be less than or equal to a preset deviation threshold; second, the number of iterations must reach the preset maximum number of iterations. No manual adjustment of model parameters is required throughout the process. After long-term operation, compared with traditional twin models without a correction mechanism, the virtual-real fitting accuracy is significantly improved, and the accuracy of equipment fault warning is greatly increased.

[0092] In some embodiments, during the continuous operation of the physical equipment, after the complete twin real-time data is synchronously input into the digital twin mechanism evolution model and the pre-trained generative adversarial network model, the discriminator is updated online using the complete twin real-time data. The unsupervised online iterative update mechanism includes: fixing the parameters of the generator G, inputting the real-time complete twin data (real twin samples) collected in real time into the discriminator D, and outputting the probability that the sample is a real twin sample. It should be noted that the complete twin real-time data is used as a real positive sample input into the discriminator, and it is expected that the discriminator outputs a high probability, but it is not forced to be 100%; during the online iterative update process, the real sample output is continuously brought closer to 1 and the negative sample output is brought closer to 0 through the loss function constraint, which is the model training optimization goal, and not that the output of each forward inference is strictly equal to 100%. The real twin sample input is used as a positive sample, and the distribution difference between the positive sample and the output result of the digital twin mechanism evolution model is used as a supervision signal, according to the online discriminant loss function shown in formula (13). The parameters of the discriminator D are iteratively updated. During the offline pre-training phase, the negative samples used by the discriminator are the output samples of the generator; during the online iterative update phase, the negative samples of the discriminator are switched to the output samples of the digital twin mechanism evolution model.

[0093] (13)

[0094] in, To complete real-time twin data The real twin samples in the data are obtained by fusion and dimensionality reduction processing of sensor data and are used as positive samples for the discriminator. Indicates from Mid-sample real twin samples Find the expected value; Indicates the set of mechanistic simulation samples Mid-sampling mechanism simulation samples And calculate the expected value. This formula is the discriminator. The value function is maximized, and in actual execution, the negative value of the value function is taken as the discriminator loss. Gradient descent is used to minimize this loss, which is equivalent to maximizing the above value function.

[0095] During the unsupervised online iterative update of the discriminator, the parameters of the generator G are kept constant, and the online discriminative loss function is maximized. The discriminator D's network parameters are updated using gradient descent. This update mechanism requires no sample fault labels, continuously improving the discriminator's discrimination capability and optimizing its discrimination accuracy as the equipment state changes. This provides a high-quality deviation quantification benchmark for subsequent dynamic mechanism evolution driven by digital-analog fusion. Real samples: Real-time complete twin feature vectors obtained through fusion and dimensionality reduction from physical equipment; Simulated twin samples: Mechanism simulation outputs from the digital twin mechanism evolution model, not GAN generator outputs. The discriminator continuously iterates online, improving its ability to distinguish between real samples and mechanism simulation samples, maintaining the data distribution benchmark. This indirectly provides a more reliable data distribution benchmark for closed-loop processing, without directly calculating the L2 deviation. However, the discriminator's distribution baseline constrains the generator's prediction output, indirectly changing the generator G's output. Thus changing the deviation .

[0096] When updating the entire set of mechanism parameters of the digital twin mechanism evolution model and the network weight parameters of the generator G based on the L2 norm deviation, the discriminator D does not participate in the update.

[0097] The following section uses a texturing machine, a typical example of complex textile equipment, as an example to further illustrate the beneficial effects of the embodiments of this application.

[0098] A texturing machine is a complex textile equipment that processes POY (pre-oriented yarn) into DTY (drawn textured yarn), significantly improving the elongation, bulk, and elasticity of synthetic fibers to meet the textile industry's demand for high-quality yarn. A texturing machine mainly consists of multiple components, including a false twister, winder, heating box, and two rollers, as well as thousands of parts, involving knowledge from multiple fields such as mechanics, electrical engineering, and electromagnetism, and disciplines such as dynamics, statics, and elasticity. Developing a digital twin model of the texturing machine's mechanism dynamic evolution requires considering both highly coupled complex modeling and the mechanism evolution of multiple physical attribute relationships, which can fully verify the technical effectiveness of the method proposed in this application.

[0099] The texturing machine's PLC controller and various data acquisition sensors (including temperature, speed, tension, and current sensors) are used to rapidly collect data on the machine's operation (yarn flow speed, winding tension, and hot box temperature profile), operating status (motor speed, current load, and false twister operating frequency), and fault data (yarn breakage, winding abnormalities, and false twister slippage). Mechanical vibration and speed data are used as... Current and voltage data as Fault label as A complete twin feature vector is formed by modal weighted fusion. For rare fault conditions such as high-speed wire breakage, virtual samples are generated through simulation generation operators to improve the balance and completeness of the dataset.

[0100] A three-layer mechanism for the evolution of a texturing machine is constructed. At the multi-dimensional parameter update level, geometric parameters include the diameter of the false twister friction disc, the size of the heating plate in the hot box, and the helix angle of the winder slot; behavioral parameters include the kinematic relationship between the false twister rotation speed and the twisting angle, and the dynamic relationship between the winding tension and the winding speed; and rule parameters include the filament breakage alarm threshold, speed matching rules, and temperature gradient control rules. The three-layer parameter set... The system is updated synchronously according to the gradient iteration formula. At the multi-domain working condition mapping level, a coupled mapping relationship is established between mechanical working conditions (drive motor output torque, transmission gear meshing force, false twister rotational friction torque) and electrical working conditions (motor current-speed curve, inverter output frequency). The models of each domain are fused and solved through coupling weight coefficients, and the accuracy is verified by comparison with measured values. At the multi-physical attribute correlation level, based on the mutual information calculation of statics (frame load, component static load), dynamics (false twister rotation dynamics, yarn tension fluctuation dynamics), and elasticity (winder elastic deformation, yarn elastic modulus change), strongly correlated features are extracted to drive the synchronous correction of the correlated model.

[0101] Fault data such as wire breakage, winding abnormalities, false twister slippage, and hot box temperature abnormalities recorded during the historical operation of the texturing machine were collected. After clustering and classification by fault mode, regression modeling, and generalization expansion, a labeled training set was formed. The generator and discriminator were trained alternately with standard adversarial loss as the optimization objective. The pre-trained generator and discriminator were then connected to the texturing machine's digital twin system. The generator was used to output the operational condition prediction results. The discriminator utilizes real-time data continuously generated during the operation of the physical loading machine for unsupervised online iterative improvement.

[0102] During the operation of the physical texturing machine, complete twin real-time data synchronously drives the three-layer mechanism evolution model and the pre-trained GAN model. The mechanism evolution model, based on the current motor speed, load simulation calculations of the theoretical frictional torque of the false twister, the theoretical winding tension of the winder, and the theoretical influence of the hot box temperature on the yarn elasticity, yields... The GAN generator outputs a fault mode probability distribution and a state deviation metric based on the real-time data feature distribution, thus obtaining... Calculate the L2 norm deviation between the two. The parameter set of the digital twin mechanism evolution model is updated using a deviation gradient reverse compensation method. The above process is continuously iterated as the loading machine runs, and the digital twin model can adaptively track the degradation evolution path and operating condition fluctuations of the physical equipment, maintaining consistency between the virtual and real systems in the long term.

[0103] Experimental results show that the root mean square error (RMSE) of the simulation fitting of the digital twin mechanism evolution model using the method of this application for the key operating parameters of the texturing machine is limited to within 10% in the early stage of evolution (hours 1-120), with an average of about 8%. After continuous closed-loop correction driven by generative adversarial networks, the error further decreases to about 3% in the middle stage of evolution (hours 240-360), and stabilizes below 2% in the late stage of evolution (hours 600-720). Regarding fault prediction accuracy, the method of this application achieves a prediction accuracy of over 93% for wire breakage faults and over 90% for winding anomaly faults, with a comprehensive fault prediction accuracy exceeding 90%. In contrast, the simulation fitting error of the traditional fixed-parameter digital twin model drifts continuously from about 12% to about 18% over 720 hours, with a comprehensive fault prediction accuracy of less than 75%. Experiments show that compared with the traditional digital twin model without a closed-loop correction mechanism, the comprehensive fault prediction accuracy of this application is improved by more than 15 percentage points, and the simulation fitting error is reduced by more than 80% under long-term operating conditions.

[0104] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0105] The embodiments provided in this application can also be computer program products. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application. A computer-readable storage medium may be a tangible device that holds and stores instructions used by an instruction execution device. Computer-readable storage media may include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0106] This application uses specific embodiments to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the solution and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion, characterized in that, include: The historical operation process data, historical operation status data and historical fault data of the equipment are integrated and dimensionality reduced to form complete twin historical data with fault labels. Construct mechanism models for equipment in various fields, and build digital twin mechanism evolution models based on the mechanism models in various fields and the three-layer mechanism evolution mechanism. The three-layer mechanism evolution mechanism includes a multi-dimensional model parameter update mechanism with data flow coupling and linkage, a multi-field working condition mapping update mechanism, and a multi-physical attribute correlation update mechanism. The complete twin historical data is used to pre-train the constructed generative adversarial network model offline. The generative adversarial network model includes a generator, which is used to output the prediction results of the equipment's operating status. The real-time operation data, real-time operation status data, and real-time fault data of the equipment are fused and dimensionality reduced to obtain complete twin real-time data. This data is then synchronously input into the digital twin mechanism evolution model and the pre-trained generative adversarial network model. The digital twin mechanism evolution model and the generator are then updated in reverse based on the deviation between their outputs.

2. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 1, characterized in that, The fusion and dimensionality reduction process includes: The fusion weight of each modality is adaptively determined based on the completeness of each modality data, and the modality data is weighted and fused. The modality data is collected by multi-acquisition device collaborative measurement, heterogeneous communication protocol parsing, and business system data import, including mechanical modality data, electrical modality data, and fault tag data. Kernel mapping is used to reduce the dimensionality of the fused high-dimensional heterogeneous data, which is then projected onto a unified low-dimensional shared latent space to obtain standardized complete twin feature vectors. Specifically, for historical operation process data, historical operation status data, and historical fault data, before fusion, for scarce fault conditions with sample size below the preset equilibrium threshold, virtual fault condition samples are generated in batches based on a small number of real fault samples of the corresponding fault condition by a condition simulation generation operator to supplement the sample distribution of the dataset.

3. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 1, characterized in that, The digital twin mechanism evolution model is constructed by connecting a multi-dimensional model parameter update module, a multi-domain working condition mapping update module, and a multi-physical attribute correlation update module, wherein: The multi-dimensional model parameter update module is used to run the multi-dimensional model parameter update mechanism; The multi-domain working condition mapping and update module is used to run the multi-domain working condition mapping and update mechanism. It includes a mechanical domain mechanism model, an electrical domain mechanism model, an electromagnetic domain mechanism model, and a mechanism simulation and solution unit. The mechanism simulation and solution unit is used to solve the multi-domain coupling state values ​​of each physical quantity based on the mechanical domain mechanism model, the electrical domain mechanism model, the electromagnetic domain mechanism model, and the coupling weights corresponding to each domain mechanism model. The multi-physical attribute correlation update module is used to run the multi-physical attribute correlation update mechanism.

4. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 3, characterized in that, Reverse updating of the digital twin mechanism evolution model includes: Calculate the L2 norm deviation between the output of the digital twin mechanism evolution model and the generator's output. If the L2 norm deviation is greater than a preset deviation threshold, update the digital twin mechanism evolution model according to the three-layer mechanism evolution mechanism. The geometric behavior rule parameters are updated locally within the layer using geometric behavior rule feature matching loss to obtain intermediate geometric behavior rule parameters. Based on the intermediate geometric behavior rule parameters, the intermediate geometric behavior rule parameters are forward inferred and gradient compensation is performed on the intermediate geometric behavior rule parameters based on the forward inference results to obtain the final effective geometric behavior rule parameters, which are used as the geometric behavior rule parameters for the next iteration. The interaction strength between mechanical, electrical, and electromagnetic components is calculated based on the updated geometry-behavior-rule parameter mapping. The coupling weights and mechanism parameters of the mechanism models in each domain are iteratively updated. The multi-domain coupling state values ​​of each physical quantity are solved based on the updated coupling weights and mechanism models in each domain until the absolute value of the first error between the multi-domain coupling state values ​​of each physical quantity and the corresponding state variables in the complete twin real-time data is greater than the preset accuracy threshold. Extract static feature sets, dynamic feature sets, and elasticity feature sets respectively. Calculate the nonlinear coupling correlation strength between any two physical features in each feature set. Select physical quantities whose nonlinear coupling strength exceeds a preset correlation strength threshold as strongly correlated features. Update the mechanism parameters and coupling weights of the domain mechanism model corresponding to the attributes of the strongly correlated features.

5. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 4, characterized in that, Also includes: Construct a multi-physical attribute correlation matrix based on the calculated nonlinear coupling correlation strength of all physical feature pairs; After the multi-domain coupling state values ​​of each physical quantity are updated, the physical quantities that have changed are determined, and the corresponding nonlinear coupling correlation strength is recalculated to update the multi-physical attribute correlation matrix. Based on the updated multi-physical attribute correlation matrix, new strongly correlated features are selected, and the mechanism parameters and coupling weights of the domain mechanism models corresponding to the attributes of the new strongly correlated features are updated, as are the mechanism parameters and coupling weights of the domain mechanism models corresponding to the attributes of the changed physical quantities.

6. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 1, characterized in that, The generative adversarial network model also includes a discriminator, which is used to characterize the data distribution benchmark of the actual operating state of the equipment. The constructed generative adversarial network model is pre-trained offline using the complete twin historical data, including: Clustering, regression, and generalization augmentation preprocessing are performed sequentially on the complete twin historical data to form a training dataset; Based on the training dataset, with a preset adversarial loss function as the optimization objective, the generator and the discriminator are trained alternately and iteratively until convergence, thus completing the offline pre-training of the generative adversarial network model. The generator's input is a random noise vector, and its output is a prediction of the equipment's operating status generated based on the noise. The discriminator's input is a complete twin historical feature vector in the training dataset, and its output is the probability that its input belongs to a real twin sample.

7. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 6, characterized in that, Reverse update of the generator includes: The L2 norm deviation is calculated based on the multi-domain coupling state values ​​of each physical quantity output by the digital twin mechanism evolution model and the generator output. When the L2 deviation is greater than a preset deviation threshold, the L2 deviation is used as an additional input signal to the generator, and is input to the generator together with the complete twin real-time data. The optimization objective is to minimize the error between the generator output and the corresponding state variable in the complete twin real-time data. The loss gradient of the error is obtained through the backpropagation algorithm, and the network weight parameters of the generator are updated along the gradient descent direction.

8. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 6, characterized in that, It also includes synchronously inputting the complete twin real-time data into the digital twin mechanism evolution model and the pre-trained generative adversarial network model, and then using the complete twin real-time data to perform online iterative updates on the discriminator in the following manner: Using the complete real-time twin data as positive samples and the output of the digital twin mechanism evolution model as negative samples, an online discriminative loss function is constructed. With the goal of maximizing the online discriminant loss function, the network parameters of the discriminator are iteratively updated using the gradient descent method.

9. The method for dynamic evolution of digital twin model mechanism driven by digital-analog fusion according to claim 1, characterized in that, The equipment is complex textile equipment, which includes texturing machines, knitting machines, or spinning machines. Multimodal data, consisting of process data, status data, and fault data, includes yarn flow rate, winding tension, hot box temperature, motor speed, current load, false twister operating frequency, and fault tag data such as yarn breakage, winding abnormality, and false twister slippage. The geometry-behavior-rule parameters include a set of geometric feature parameters, a set of motion behavior parameters, and a set of operational rule parameters, wherein: The set of geometric feature parameters includes the diameter of the false twister friction disc, the size of the heating plate in the hot box, and the helix angle of the winder slot; The set of motion behavior parameters includes kinematic parameters relating the rotational speed of the false twister to the twisting angle, and dynamic parameters relating the winding tension to the winding speed. The set of operating rules parameters includes wire breakage alarm threshold, speed matching rules, and temperature gradient control rules.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the dynamic evolution method for digital twin model mechanism driven by digital-analog fusion as described in any one of claims 1-9.

Images