A digital twin adaptive prediction system based on transfer learning

Abstract

The application discloses a kind of digital twin adaptive prediction system based on transfer learning, including twin model construction module, adaptive update layer;The twin model construction module includes physical entity layer, twin data layer, data connection layer, virtual twin layer;The adaptive update layer includes prediction model construction module, data preprocessing module, model training module, adaptive update module.The application proposes the adaptive migration strategy of combination neural network prediction model in combination with transfer learning, the network structure and parameter trained in source domain are migrated to target domain, and the adaptive network parameter of adjustment is determined according to the prediction threshold value;And the adaptive prediction model is introduced into the five-dimensional twin model constituted by physical entity layer, virtual twin layer, twin data layer, data connection layer and service application layer, the overall framework of adaptive prediction model under twin environment is proposed, realizes the functions such as two-way mapping, real-time transmission, online prediction and feedback control of virtual twin and physical entity, provides method and implementation approach for process industry production line construction driven by digital twin.

Description

Technical Field

[0001] The present invention relates to a digital twin adaptive prediction system based on transfer learning, and belongs to the field of intelligent manufacturing. Background Art

[0002] The process industry plays an important and critical role in the national economy. Ensuring the long-term stable operation of its production process has become a challenge facing the further development of the process industry. Unlike traditional discrete manufacturing, the process production process is composed of multiple production steps, with characteristics such as many segmented production units and frequently changing production requirements for multiple specifications. This also determines that the key to stable and controllable process production quality is how to achieve rapid and accurate prediction of its process quality. In recent years, the use of digital twin technology for equipment and product quality monitoring has become a research hotspot. However, the high-fidelity digital twin model of the production line established for specific scenarios still has difficulty in solving the problem of rapid and accurate prediction of the quality of the production process under variable working conditions. It has become a limitation that restricts the adaptive update of the digital twin workshop to respond to changing working conditions and needs to be solved urgently. Summary of the Invention

[0003] The present invention provides a digital twin adaptive prediction system based on transfer learning, which is used to achieve real-time mapping of digital twin models under variable working conditions.

[0004] The technical solution of the present invention is:

[0005] According to a first aspect of the present invention, there is provided a digital twin adaptive prediction system based on transfer learning, comprising a twin model construction module, an adaptive update layer;

[0006] The twin model building module includes:

[0007] The physical layer is used to configure production equipment and data acquisition devices in the production workshop;

[0008] The twin data layer is used to store the equipment attributes of production equipment and the operating condition data collected by the data acquisition device. The equipment attributes include equipment spatial layout, assembly rules, equipment and material geometric parameters. The operating condition data refers to multiple process parameters with time series characteristics and a quality indicator.

[0009] The data connection layer is used to develop interfaces between the physical entity layer, virtual twin layer, twin data layer, and adaptive update layer based on various communication protocols, and complete data interaction and virtual-real mapping from the physical entity layer to the virtual twin layer;

[0010] The virtual twin layer is used to build virtual models based on the geometric parameters and assembly rules of equipment and materials stored in the twin data layer, and to add additional features to the models; analyze the operating logic of production equipment and build a logical model that reflects the evolution of the production line;

[0011] The adaptive update layer includes:

[0012] A prediction model building module is used to build a combined network model based on a neural combination network, including processing the primary features and time series features of the input data;

[0013] The data preprocessing module is used to preprocess the historical working condition data consisting of process parameters and quality indicators with time series labels stored in the twin data layer to obtain a preprocessed dataset as the source domain dataset; the preprocessed dataset is divided into training set data and test set data for model training and performance evaluation;

[0014] The model training module is used to input the training set data to train the combined network model, adjust the hyperparameters, and obtain a process quality indicator prediction model that is highly consistent with the test set data as the source domain model;

[0015] The adaptive update module compares the prediction accuracy obtained based on the source domain model under real-time working condition data with the set accuracy threshold: if the accuracy threshold is not reached, the adaptive migration strategy proposed based on transfer learning theory is used to adaptively update the source domain model; otherwise, no update is performed.

[0016] Furthermore, the digital twin adaptive prediction system based on transfer learning also includes a service application layer. The service application layer specifically includes: designing and developing a UI interface for user interaction based on Unity's Xcharts plug-in; and dynamically connecting and calling the MySql database and the target domain model based on Socket technology and dynamic link library.

[0017] Furthermore, the virtual model includes: a workshop environment, which is built based on the physical environment of the production workshop; a processing unit, which is established based on the geometric parameters and assembly rules of the processing equipment stored in the twin data layer; a transmission unit, which is established based on the geometric parameters and assembly rules of the transmission device equipment stored in the twin data layer; and a material unit, which is established based on the material geometric parameters stored in the twin data layer. After establishing the above four parts, the positions of the processing unit, transmission unit, and material unit are adjusted according to the spatial layout of the equipment to ensure that the virtual scene is consistent with the physical entity, thereby completing the layout of the visual virtual model.

[0018] Furthermore, the combined network model includes at least an input layer, a primary feature extraction layer, a temporal information processing layer, a fully connected layer, and an output layer.

[0019] Furthermore, the adaptive migration strategy includes:

[0020] In the comprehensive migration step of network structure parameters, the network structure of the source domain model is used as the network structure of the target model, and the parameters of the source domain model are used as the initial values ​​of the target model parameters; then the fully connected layer is unfrozen layer by layer.

[0021] In the step of unfreezing the fully connected layer layer by layer, real-time working condition data consisting of process parameters and quality indicators with time series labels is used as the target domain dataset. The network structure and network parameters of the primary feature extraction layer and the time series information processing layer in the source domain model are shared. The fully connected layer parameters are unfrozen layer by layer. Before the termination of the fully connected layer training, if the prediction accuracy obtained by the unfrozen fully connected layer parameters is better than the prediction accuracy obtained by the parameters before unfreezing, the unfrozen fully connected layer parameters are output and stored as the optimized parameters of the fully connected layer to obtain the first target domain model; otherwise, the retraining step is entered;

[0022] In the retraining step, the network structure of the source domain model is kept unchanged, and the network parameters of the primary feature extraction layer, the temporal information processing layer, and the fully connected layer are randomly initialized. After combining the source domain dataset and the target domain dataset, the model is retrained as the second target domain model.

[0023] According to a second aspect of the present invention, a terminal is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the construction of one or more layers in the digital twin adaptive prediction system based on transfer learning as described above.

[0024] The beneficial effects of the present invention are: in response to the problem that the digital twin workshop lacks adaptive updating capabilities under changing working conditions in the current process manufacturing industry, the present invention introduces neural networks and transfer learning technology, and proposes a digital twin adaptive prediction system based on transfer learning. The framework combines transfer learning to propose an adaptive migration strategy for the combined neural network prediction model, migrates the network structure and parameters trained in the source domain to the target domain, and adaptively determines the adjusted network parameters according to the prediction threshold; and introduces the adaptive prediction model into a five-dimensional twin model composed of a physical entity layer, a virtual twin layer, a twin data layer, a data connection layer and a service application layer, and proposes an overall framework for the adaptive prediction model in a twin environment, which realizes the functions of two-way mapping, real-time transmission, online prediction and feedback control between virtual twins and physical entities, and provides a method and implementation path for the construction of process industry production lines driven by digital twins. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A digital twin adaptive prediction system based on transfer learning;

[0026] Figure 2An adaptive migration strategy for the prediction model of the present invention;

[0027] Figure 3 This is part of the data set of the embodiment of the present invention;

[0028] Figure 4 This is a comparison chart of the predicted value and the actual value of the moisture content of the outlet material under variable working conditions in the embodiment of the present invention;

[0029] Figure 5 This is a diagram of the digital twin system developed based on the digital twin adaptive prediction framework in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other in any way.

[0031] Process manufacturing production lines have the characteristics of large fluctuations in process, operation, and quality data, and strong temporal correlation between data. In this context, according to the first aspect of the embodiment of the present invention, a digital twin adaptive prediction system based on transfer learning is provided, such as Figures 1-5 As shown, the digital twin adaptive prediction system based on transfer learning includes a twin model construction module and an adaptive update layer.

[0032] like Figure 1As shown, the twin model construction module includes a physical entity layer, a twin data layer, a data connection layer, and a virtual twin layer; the physical entity layer is used to configure production operation equipment (production operation equipment includes processing equipment, transmission equipment, storage equipment) and data acquisition devices in the production workshop; the twin data layer is used to store the equipment attributes of the production operation equipment and the working condition data collected by the storage data acquisition device. The equipment attributes include equipment spatial layout, assembly rules, equipment and material geometric parameters. The working condition data refers to multiple process parameters with time series characteristics and a quality indicator; the data connection layer is used to develop interfaces between the physical entity layer, the virtual twin layer, the twin data layer, the adaptive update layer and the service application layer based on various communication protocols to complete data interaction and virtual-real mapping from the physical entity layer to the virtual twin layer; the virtual twin layer is used to construct a virtual model based on the geometric parameters and assembly rules of the equipment and materials stored in the twin data layer, and to give additional features to the model, so as to facilitate direct observation of the process flow and equipment operation status of the physical production line through a visualized virtual model; analyze the operating logic of the production operation equipment, and construct a logical model reflecting the evolution law of the production line based on the C# language. Since processing equipment is the most complex among production equipment, the focus is on analyzing the operating logic of processing equipment.

[0033] like Figure 2 As shown, the adaptive update layer includes:

[0034] The prediction model construction module constructs a combined network model based on a neural network to process the primary features and time series features of the input data; and presets a series of network structures and parameters such as the number of network layers, activation functions, optimizers, and learning rates; the combined network model includes at least an input layer, a primary feature extraction layer, a time series information processing layer, a fully connected layer, and an output layer;

[0035] The data preprocessing module is used to preprocess the historical working condition data composed of process parameters and quality indicators with time series labels stored in the twin data layer, and obtain the preprocessed data set as the source domain data set; the preprocessed data set is divided into training set data and test set data for model training and performance evaluation; among them, the preprocessing includes operations such as outlier removal and normalization; the preprocessing is specifically as follows: first, the material head and tail are removed, then the vacant values ​​in the data are filled with the mean, and then the sample data is eliminated using the 3sigma criterion; the formula is used for the data set judged by the 3sigma criterion Perform normalization to eliminate the dimensionality effect of data; divide the training set and test set into an 8:2 ratio;

[0036] A model training module is used to input training set data to perform combined network model training, adjust hyperparameters, and obtain a process quality indicator prediction model that is highly consistent with the test set data as a source domain model. In the embodiment of the present invention, a prediction accuracy of 95% or more is considered highly consistent.

[0037] The adaptive update module compares the prediction accuracy obtained based on the source domain model under the real-time working condition data of a stage with the set accuracy threshold: if the accuracy threshold is not reached, the adaptive migration strategy proposed based on the transfer learning theory is used to adaptively update the source domain model (that is, used to update the network structure and parameters of the process quality indicator prediction model trained based on historical working condition data); otherwise, no update is performed; the difference in prediction accuracy before and after the model update is used as the evaluation standard, and the working condition data with higher storage performance improvement and the updated algorithm model are considered typical working conditions.

[0038] like Figure 1 As shown, the digital twin adaptive prediction system based on transfer learning also includes a service application layer. The service application layer is specifically: a UI interface for user interaction is designed and developed based on the Xcharts plug-in of Unity, including interfaces such as real-time transmission of working condition data, production scene process monitoring, and online prediction of quality indicators. It is based on Socket technology and dynamic link library to dynamically connect and call the MySql database and the target domain model. For example, Figure 5 Shown is the UI interface developed by the service application layer.

[0039] Furthermore, the physical entity layer is specifically as follows: taking the silk production line of a certain process manufacturing enterprise as an example, the overall process flow is "vacuum moisture recovery → super loosening → leaf moistening and feeding → foreign matter removal → shredding → leaf moisture recovery → heating and humidification → thin plate drying → proportional blending → fragrance addition", and each equipment is composed of a variety of devices. For example, the silk drying machine as a processing equipment is composed of a drum, a steam system, a moisture exhaust system, a discharge port and other devices. Its main function is the thin plate drying process in the process flow; the process parameters of the production process are collected through various data acquisition sensors including temperature sensors, humidity sensors, and pressure sensors, and a required process parameter is selected as a quality indicator.

[0040] In an embodiment of the present invention, knowledge of graphics, mechanics, etc. is used to analyze the equipment properties of the production and operation equipment, and cloud servers and databases are used to store historical working condition data and real-time working condition data consisting of process parameters and quality indicators with time series labels (if used for multiple terminals, cloud servers are used for storage; if used for one terminal, either databases or cloud servers can be used); illustratively, for super loosening, leaf moisture regain, and thin plate drying processes, the working condition data use inlet material moisture, dehumidification damper opening, dehumidification air volume, hot air temperature, HT steam pressure, rear chamber damper opening, inlet thin plate steam pressure, front chamber damper opening, barrel wall temperature pressure, inlet material flow, HT inlet material temperature, outlet material temperature, etc. as process parameters, and the outlet material moisture value is set as the quality indicator.

[0041] Furthermore, the data connection layer is specifically: based on the communication technology framework of OPCUA and Kepware, the real-time working condition data generated by the process manufacturing workshop is collected to the OPCUA server through Kepware and stored in the database, and then the Python script encapsulated based on Socket communication technology uploads the acquired data to the MySql cloud server / database, completing the data interface construction between the physical entity layer and the twin data layer;

[0042] Develop C# code based on Unity's built-in Visual Studio 2019, configure the Mysql.Data.dll file environment, connect to the IP and port of the MySql cloud server using APIs such as MySqlClient, traverse to obtain the database, and complete the data interface between the virtual twin layer and the twin data layer.

[0043] Build a static function in C# to load the Python code process, configure the environment variables for running the Python code, pass input parameters in the form of arrays, lists, strings, etc., and obtain console output results to complete the data interface between the virtual twin layer and the adaptive update layer.

[0044] Develop the interactive interface between the physical entity layer and the virtual twin layer based on the database open API; develop the interactive interface between the service application layer and the twin data layer based on the API of the algorithm model compilation language and the twin system compilation language encapsulated by socket communication technology.

[0045] Furthermore, the virtual model includes: a workshop environment, which is built based on the physical environment of the production workshop; a processing unit, which is established based on the geometric parameters and assembly rules of the processing equipment stored in the twin data layer; a transmission unit, which is established based on the geometric parameters and assembly rules of the transmission device equipment stored in the twin data layer; and a material unit, which is established based on the material geometric parameters stored in the twin data layer. After establishing the above four parts, the positions of the processing unit, transmission unit, and material unit are adjusted according to the spatial layout of the equipment to ensure that the virtual scene is consistent with the physical entity, thereby completing the layout of the visual virtual model.

[0046] For example, 3D modeling software (Soliderworks) is used to create scale models of processing units and transmission units. These are then lightweighted (Polygon Cruncher) and imported into the Unity3D platform. The workshop environment and material units are then built on the Unity3D platform, and additional features are assigned to the models. (For example, in the thin-plate drying process of the silk production line, the additional features assigned include actual materials, texture properties, particle systems, physical collision systems, and lighting rendering. For example, the particle system assignment process involves importing point cloud data of the material and using the particle system to dynamically render the material's operating status in real time.) Analysis of the operating logic of the processing equipment in the production operation reveals that the physical structure of the processing equipment is clearly hierarchical. This allows for the division of different levels according to function and structure, and the corresponding combination and assembly relationships are set for the internal parts of the equipment on the front end of the platform to construct a logical model that reflects the evolution of the production line.

[0047] Furthermore, the adaptive migration strategy is specifically as follows:

[0048] In the comprehensive migration step of network structure parameters, the network structure of the source domain model is used as the network structure of the target model, and the parameters of the source domain model are used as the initial values ​​of the target model parameters; then the fully connected layer is unfrozen layer by layer.

[0049] The step of unfreezing the fully connected layer layer by layer uses the real-time working condition data consisting of process parameters and quality indicators with time series labels as the target domain data set, shares the network structure and network parameters of the primary feature extraction layer and the time series information processing layer in the source domain model, and unfreezes the fully connected layer parameters layer by layer. Before the termination of the fully connected layer training, if the prediction accuracy obtained by the unfrozen fully connected layer parameters is better than the prediction accuracy obtained by the parameters before unfreezing, then the unfrozen fully connected layer parameters are output and stored as the optimized parameters of the fully connected layer to obtain the first target domain model; otherwise, enter the retraining step; the process of unfreezing the fully connected layer parameters layer by layer is as follows: under the number of iterations, if the prediction accuracy obtained by the unfrozen current layer parameters of the fully connected layer is better than the prediction accuracy obtained by the parameters before unfreezing, then the unfrozen current layer parameters of the fully connected layer are output and stored as the optimized parameters of the current layer of the fully connected layer; otherwise, continue to unfreeze the parameters of the next layer of the fully connected layer until the training of the fully connected layer is terminated;

[0050] In the retraining step, the network structure of the source domain model is kept unchanged, and the network parameters of the primary feature extraction layer, the temporal information processing layer, and the fully connected layer are randomly initialized. After combining the source domain dataset and the target domain dataset, the model is retrained as the second target domain model.

[0051] Based on the above, it can be seen that the adaptive migration strategy is used to obtain the first target domain model or the second target domain model.

[0052] Exemplarily, the combined network model takes the CNN-LSTM combined neural network as an example, including an input layer, a CNN convolutional network module as a primary feature extraction layer, an LSTM long short-term memory neural network module as a temporal information processing layer, a fully connected layer, and an output layer; in the above, CNN is used to extract local features of process parameters, and LSTM is used to extract temporal features of process parameters. The advantages of the two are combined to reduce network structure parameters, reduce overfitting risks, integrate features with a fully connected layer, and output prediction results; the final parameters of the source domain model are set to 24 input layer features and 88 / max fully connected layer neurons; the CNN network structure is designed as one convolution layer (Conv1D), 44 convolution kernels, a convolution kernel size of 3×3, a step size of 1, and an activation function of ReLu; the LSTM input feature dimension and hidden layer dimension are both 88; the number of iterations and the batch training number are 40 and 32 respectively, and the Adam optimizer is used.

[0053] Furthermore, the combined network model training and performance evaluation are carried out based on the divided data sets to verify the effectiveness of the combined network model. 2 The size of is used as a measure to evaluate the effectiveness of the model. The three parameters are calculated as follows: , , ; Where: n is the data volume of the predicted moisture value of the export material; y i The actual value of the moisture content of the exported material; is the predicted value of moisture content of export materials; The smaller the RMSE and MAE values ​​are, the higher the R 2 The closer the value is to 1, the more accurate the model prediction results are.

[0054] The following is an explanation based on experimental data:

[0055] The collected production data of the silk production line include three sets of thin plate drying process data sets from October 18, 2022 to October 24, 2022, namely Dataset 1, Dataset 2 and Dataset 3, with a data volume of 20,000 sets each. Part of the original data of Dataset 1 is as shown in the attached figure. Figure 3 The dataset contains 23 process parameters, including actual inlet material moisture, actual inlet material flow rate, and HT inlet material temperature, totaling 22 process parameters, and one quality indicator (actual outlet material moisture). The dataset is divided into a training set (16,000 samples) and a test set (4,000 samples) in an 8:2 ratio. The training set is used for model training, and the test set is used to evaluate the model's generalization ability.

[0056] Taking the CNN-LSTM combined neural network as an example, the prediction model of dataset 1 is pre-trained and the hyperparameters are adjusted to enable it to deeply learn the relationship and pattern between time series and achieve efficient prediction performance. The adaptive migration strategy is introduced to update the trained CNN-LSTM combined neural network prediction model, and predictions are made on datasets 2 and 3. The MAE, RMSE, and R 2 Three evaluation indicators are used to verify the generalization performance of the model, as shown in the attached Figure 4 As shown in Table 1.

[0057] Table 1

[0058]

[0059] The results show that the digital twin adaptive prediction system based on transfer learning of the present invention has demonstrated good prediction effects under different data sets. After transfer adaptive training, the prediction accuracy of the model on data sets 2 and 3 exceeded 98%, indicating that the model has strong generalization ability and good adaptability to different working conditions. In the specific scenario of predicting the moisture content of the material at the outlet of the thin plate drying process, its high-precision prediction ability can provide strong support for refined control in the actual production process.

[0060] According to a second aspect of an embodiment of the present invention, there is provided a terminal comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the construction of one or more layers in any one of the above-mentioned transfer learning-based digital twin adaptive prediction systems. It should be noted that each module / layer in the above-mentioned prediction framework can be implemented by software or hardware. For the latter, it can be implemented in the following manner, but is not limited thereto: the above-mentioned modules / layers are all located in the same processor; or the above-mentioned modules / layers are located in different processors in any combination.

[0061] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A digital twin adaptive prediction system based on transfer learning, characterized in that: Includes twin model building modules and adaptive update layers; The twin model building module includes: The physical layer is used to configure production equipment and data acquisition devices in the production workshop; The twin data layer is used to store the equipment attributes of production equipment and the operating condition data collected by the data acquisition device. The equipment attributes include equipment spatial layout, assembly rules, and equipment and material geometric parameters. The operating condition data refers to multiple process parameters with time series characteristics and a quality indicator. The data connection layer is used to develop interfaces between the physical entity layer, virtual twin layer, twin data layer, and adaptive update layer based on various communication protocols, and complete data interaction and virtual-real mapping from the physical entity layer to the virtual twin layer; The virtual twin layer is used to build virtual models based on the geometric parameters and assembly rules of equipment and materials stored in the twin data layer, and to add additional features to the models; analyze the operating logic of production equipment and build a logical model that reflects the evolution of the production line; The adaptive update layer includes: A prediction model building module is used to build a combined network model based on a combined neural network, including processing primary features and time series features of input data; The data preprocessing module is used to preprocess the historical working condition data consisting of process parameters and quality indicators with time series labels stored in the twin data layer to obtain a preprocessed dataset as the source domain dataset; the preprocessed dataset is divided into training set data and test set data for model training and performance evaluation; The model training module is used to input the training set data to train the combined network model, adjust the hyperparameters, and obtain a process quality indicator prediction model that is highly consistent with the test set data as the source domain model; The adaptive update module compares the prediction accuracy obtained based on the source domain model under real-time working condition data with the set accuracy threshold. If the accuracy threshold is not reached, the adaptive migration strategy proposed based on transfer learning theory is used to adaptively update the source domain model; otherwise, no update is performed. The virtual model includes: a workshop environment, which is built based on the physical environment of the production workshop; a processing unit, which is established based on the geometric parameters and assembly rules of the processing equipment stored in the twin data layer; a transmission unit, which is established based on the geometric parameters and assembly rules of the transmission device stored in the twin data layer; and a material unit, which is established based on the material geometric parameters stored in the twin data layer. After the workshop environment, processing units, transmission units, and material units are established, their positions are adjusted according to the spatial layout of the equipment to ensure that the virtual scene is consistent with the physical entity, thereby completing the layout of the visual virtual model. The adaptive migration strategy includes: In the comprehensive migration step of network structure parameters, the network structure of the source domain model is used as the network structure of the target model, and the parameters of the source domain model are used as the initial values ​​of the target model parameters; then the fully connected layer is unfrozen layer by layer. In the step of unfreezing the fully connected layer layer by layer, the real-time working condition data consisting of process parameters and quality indicators with time series labels is used as the target domain data set. The network structure and network parameters of the primary feature extraction layer and the time series information processing layer in the source domain model are shared. The fully connected layer parameters are unfrozen layer by layer. Before the termination of the fully connected layer training, if the prediction accuracy obtained by the unfrozen fully connected layer parameters is better than the prediction accuracy obtained by the parameters before unfreezing, the unfrozen fully connected layer parameters are output and stored as the optimized parameters of the fully connected layer to obtain the first target domain model; otherwise, the retraining step is entered; In the retraining step, the network structure of the source domain model is kept unchanged, and the network parameters of the primary feature extraction layer, the temporal information processing layer, and the fully connected layer are randomly initialized. After combining the source domain dataset and the target domain dataset, the model is retrained as the second target domain model.

2. The digital twin adaptive prediction system based on transfer learning according to claim 1, characterized in that The digital twin adaptive prediction system based on transfer learning also includes a service application layer. The service application layer specifically includes: designing and developing a UI interface for user interaction based on Unity's Xcharts plug-in; and dynamically connecting and calling the MySql database and the target domain model based on Socket technology and dynamic link library.

3. The digital twin adaptive prediction system based on transfer learning according to claim 1, characterized in that The combined network model includes at least an input layer, a primary feature extraction layer, a temporal information processing layer, a fully connected layer, and an output layer.

4. A terminal, characterized in that: The invention comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the construction of one or more layers in the digital twin adaptive prediction system based on transfer learning according to any one of claims 1 to 3.

Images