A lightweight processing method for digital twin models

Abstract

The present invention relates to the technical field of model lightweight processing, and discloses a digital twin model lightweight processing method, the method comprising: obtaining a digital twin model, constructing a multidimensional twin model parameter importance evaluation objective function, and evaluating the digital twin model parameters obtained by training and solving; lightweight processing of the digital twin model parameters based on the importance evaluation results to obtain a lightweight digital twin model; fine-tuning the lightweight digital twin model to obtain a lightweight tuned digital twin model. The present invention performs gradient evaluation and channel gain evaluation according to the gradient of the digital twin model parameters during the training process and the final training results, selects digital twin model parameters with a higher contribution to the output result, filters redundant parameters, realizes lightweight processing of the digital twin model, and performs fine-tuning processing to retain the performance of the original model, thereby obtaining a digital twin model that reduces the consumption of computing resources and storage costs.

Description

Technical Field

[0001] The present invention relates to the technical field of model lightweight processing, and in particular to a digital twin model lightweight processing method. Background Art

[0002] With the rapid development of technologies such as industrial Internet, Internet of Things, and artificial intelligence, digital twins, as an emerging technical architecture, have been widely used in various fields. Digital twins are used to accurately monitor, simulate, analyze, and predict physical objects or systems through real-time data interaction between physical entities and their virtual models. Its application in manufacturing, aerospace, energy, transportation, and other industries has gradually deepened, and has become one of the key technologies for optimizing production processes, improving system performance, and enhancing equipment predictive maintenance capabilities. The core of digital twin technology is to create a virtual model that accurately reflects the state of physical entities. This virtual model can not only synchronize changes in the physical world in real time, but also predict and optimize based on technologies such as big data, artificial intelligence, and machine learning. However, with the continuous expansion of digital twin applications and the increasing size of models, the computational burden, storage requirements, and communication pressure of digital twin models often reach bottlenecks, resulting in limited response speed and reliability of the system. How to "slim down" the model while ensuring model accuracy has become an important issue that needs to be solved urgently. In response to this problem, the present invention proposes a lightweight processing method for digital twin models, which reduces its computational complexity and resource consumption by lightweight processing of twin models. Summary of the invention

[0003] In view of this, the present invention provides a lightweight processing method for a digital twin model, obtains a digital twin model for simulation and performance optimization, and constructs two parts of the loss function to train the digital twin model parameters respectively. According to the gradient of the digital twin model parameters during the training process and the final training results, gradient evaluation and channel gain evaluation are performed, and digital twin model parameters with higher contribution to the output result are selected, and redundant parameters are filtered to achieve lightweight processing of the digital twin model, so as to obtain a digital twin model with reduced computing resource consumption and storage cost.

[0004] To achieve the above object, the present invention provides a digital twin model lightweight processing method, comprising the following steps:

[0005] S1: Obtain a digital twin model, train and solve the digital twin model, and obtain digital twin model parameters;

[0006] S2: Construct a multi-dimensional twin model parameter importance evaluation objective function to evaluate the digital twin model parameters, where the importance evaluation dimensions of the multi-dimensional twin model parameter importance evaluation objective function include gradient evaluation and channel gain evaluation;

[0007] S3: Based on the importance assessment results, the digital twin model parameters are lightweighted, and a lightweight digital twin model is constructed using the lightweighted digital twin model parameters;

[0008] S4: Fine-tune the lightweight digital twin model to obtain a lightweight tuned digital twin model.

[0009] As a further improvement method of the present invention:

[0010] Optionally, obtaining the digital twin model in step S1 includes:

[0011] The digital twin model uses sensors to collect status data of monitored objects in different scenarios, and uses machine learning algorithms to map the sensor data to obtain monitoring, simulation and optimization results of the monitored objects, where the monitored objects include industrial equipment, industrial machines, vehicle transportation systems, energy systems and human health;

[0012] The monitoring content of the monitored object includes the current state data and historical state data of the monitored object, and early warning of abnormal data to prevent the problem from expanding. The state data is collected by sensors;

[0013] The simulation results of the monitored object include simulating the monitored object using a digital twin model to obtain the behavior and performance of the monitored object in different scenarios, and then predicting the state of the monitored object at future moments, as well as the time and location of potential faults;

[0014] The optimization results of the monitored object include the digital twin model customizing the optimization strategy of the monitored object based on the monitoring content and simulation results to optimize the performance of the monitored object;

[0015] The digital twin model parameters in the digital twin model are divided into two types, namely simulation model parameters and optimization model parameters.

[0016] Optionally, loss functions are constructed for different types of digital twin model parameters for training and solving, including:

[0017] The loss function of the simulation model parameters is:

[0018] ;

[0019] in:

[0020] Represents simulation model parameters The loss function is It represents the true prediction value of the h-th category prediction result of the n-th group of training samples in the simulation model parameter training process. Represents the simulation model parameters based on , the prediction result of the nth group of training samples is the probability value of the hth category prediction result, , H represents the total number of categories of prediction results;

[0021] The loss function of the optimization model parameters is:

[0022] ;

[0023] in:

[0024] Represents the optimization model parameters The loss function is It represents the true optimization value of the mth group of training samples in the dth type of optimization strategy during the training process of the optimization model parameters. Represents the optimization model parameters based on , the prediction result of the mth group of training samples is the probability value of the dth type of optimization strategy, , D represents the total number of categories of optimization strategies, represents the training sample balance coefficient, Set to 2.

[0025] Optionally, the digital twin model parameters obtained by training and solving include:

[0026] The digital twin model parameters obtained by the training and solution include simulation model parameters and optimization model parameters;

[0027] The simulation model parameters obtained by training are:

[0028] ;

[0029] in:

[0030] is the contribution weight parameter of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. , J represents the number of characteristics of the state data input into the digital twin model during the simulation of the monitoring object using the digital twin model;

[0031] The optimized model parameters obtained by training are:

[0032] ;

[0033] in:

[0034] Indicates the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the training process of the optimization model parameters;

[0035] , B represents the monitoring content input into the digital twin model and the number of features of the simulation results in the process of using the digital twin model to customize the optimization strategy for the monitoring object.

[0036] Optionally, in step S2, constructing a multidimensional twin model parameter importance evaluation objective function includes:

[0037] The importance evaluation dimensions of the objective function of multi-dimensional twin model parameter importance evaluation include gradient evaluation and channel gain evaluation. The importance evaluation of digital twin model parameters is performed using the gradient of the digital twin model parameters during the training process. The gradient reflects the impact of the digital twin model parameters on the loss function. The larger the gradient, the greater the contribution of the digital twin model parameters to the output result. After calculating the gradients of all digital twin model parameters, the absolute value of the gradient is used as the gradient evaluation of the digital twin model parameters.

[0038] The strength of the digital twin model parameters is calculated as the channel gain evaluation, where the higher the channel gain evaluation, the stronger the feature extraction capability of the digital twin model parameters;

[0039] The lower the importance assessment result of the digital twin model parameter, the lower the contribution of the digital twin model parameter to the output result, and the more redundant it is in the digital twin model.

[0040] Optionally, the use of the multidimensional twin model parameter importance evaluation objective function to evaluate the importance evaluation result of the digital twin model parameters includes:

[0041] The digital twin model parameters The importance assessment result is :

[0042] ;

[0043] in:

[0044] Represents the parameters of the digital twin model Gradient evaluation of ;

[0045] Represents the parameters of the digital twin model Channel gain evaluation;

[0046] The digital twin model parameters The importance assessment result is :

[0047] ;

[0048] in:

[0049] Represents the parameters of the digital twin model Gradient evaluation of ;

[0050] Represents the parameters of the digital twin model Channel gain evaluation.

[0051] Optionally, in step S3, lightweight processing is performed on the digital twin model parameters based on the importance evaluation result, including:

[0052] Based on the importance evaluation results, the simulation model parameters and the optimization model parameters are ranked respectively. The higher the importance evaluation result, the higher the ranking. The top G% of the simulation model parameters and the optimization model parameters are selected to perform lightweight reconstruction on the digital twin model to obtain a lightweight digital twin model. The unselected simulation model parameters and the optimization model parameters are filtered out. When the digital twin model is used to simulate and optimize the monitored object in the subsequent process, the sample features corresponding to the filtered digital twin model parameters are no longer processed, which reduces the consumption of computing resources and computing costs and improves the portability of the model.

[0053] Optionally, the lightweight digital twin model is fine-tuned in step S4, including:

[0054] R groups of training samples are selected from the training set, and a loss function of the lightweight digital twin model parameters in the lightweight digital twin model is constructed. The lightweight digital twin model parameters are fine-tuned in combination with the loss function, and the lightweight tuned digital twin model is constructed using the fine-tuning results. A compression algorithm is used to compress the lightweight digital twin model parameters after fine-tuning. When the digital twin model does not need to be run, the lightweight digital twin model parameters after fine-tuning are compressed and stored. When the digital twin model needs to be run, the compressed results are decompressed.

[0055] In order to solve the above problem, the present invention provides an electronic device, the electronic device comprising:

[0056] A memory storing at least one instruction;

[0057] Communication interface, enabling electronic equipment to communicate; and

[0058] A processor executes instructions stored in the memory to implement the above-mentioned digital twin model lightweight processing method.

[0059] In order to solve the above problems, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored. The at least one instruction is executed by a processor in an electronic device to implement the above-mentioned digital twin model lightweight processing method.

[0060] Compared with the prior art, the present invention proposes a lightweight processing method for a digital twin model, which has the following advantages:

[0061] First, this scheme proposes a method for evaluating the importance of model parameters, which uses the gradient of the digital twin model parameters during the training process to evaluate the importance of the digital twin model parameters, where the gradient reflects the influence of the digital twin model parameters on the loss function. The larger the gradient, the greater the contribution of the digital twin model parameters to the output result. After calculating the gradient of all digital twin model parameters, the absolute value of the gradient is used as the gradient evaluation of the digital twin model parameters; the strength of the digital twin model parameters is obtained by calculation as the channel gain evaluation, where the higher the channel gain evaluation, the stronger the feature extraction ability of the digital twin model parameters; the lower the importance evaluation result of the digital twin model parameter, the lower the contribution of the digital twin model parameter to the output result, and it is more redundant in the digital twin model.

[0062] At the same time, this scheme proposes a learning rate optimization method. The present invention selects different dynamic learning rate methods in the iteration process of the two digital twin model parameters. In the simulation model parameter training process, a dynamic interval descent strategy is adopted. A higher learning rate is adopted in the early stage of training, which can enable the model to quickly learn the basic pattern of the data and accelerate convergence. The learning rate is gradually reduced to reduce the oscillation caused by the high learning rate, so that the simulation model parameters can converge to the optimal solution more stably; in the training process of optimizing model parameters, a continuous learning rate change strategy is adopted. Smoother learning rate adjustment can make model training more stable and avoid training fluctuations caused by sudden learning rate changes. By periodically resetting the learning rate, the model can be helped to jump out of the local optimal solution and explore a wider parameter space, further improving the convergence performance. The learning rate will be set to a higher value in each iteration cycle, and then gradually reduced, so that the optimization model parameters can be globally searched in the new parameter area, which is helpful to find a better solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic flow chart of a lightweight processing method for a digital twin model provided by one embodiment of the present invention;

[0064] Figure 2 A performance test comparison diagram of a digital twin model before and after lightweight processing provided by an embodiment of the present invention;

[0065] Figure 3A diagram of the importance evaluation results of digital twin model parameters provided in one embodiment of the present invention;

[0066] Figure 4 This is a comparison diagram of an experiment on a model lightweight processing method provided by an embodiment of the present invention;

[0067] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0068] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0069] The embodiment of the present application provides a lightweight processing method for a digital twin model. The execution subject of the lightweight processing method for the digital twin model includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided in the embodiment of the present application. In other words, the lightweight processing method for the digital twin model can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0070] Embodiment 1:

[0071] A digital twin model lightweight processing method includes the following steps:

[0072] S1: Obtain the digital twin model, train and solve the digital twin model, and obtain the digital twin model parameters.

[0073] The digital twin model is obtained in step S1, including:

[0074] The digital twin model uses sensors to collect status data of monitored objects in different scenarios, and uses machine learning algorithms to map the sensor data to obtain monitoring, simulation and optimization results of the monitored objects, where the monitored objects include industrial equipment, industrial machines, vehicle transportation systems, energy systems and human health;

[0075] The monitoring content of the monitored object includes the current state data and historical state data of the monitored object, and early warning of abnormal data to prevent the problem from expanding. The state data is collected by sensors;

[0076] The simulation results of the monitored object include simulating the monitored object using a digital twin model to obtain the behavior and performance of the monitored object in different scenarios, and then predicting the state of the monitored object at future moments, as well as the time and location of potential faults;

[0077] The optimization results of the monitored object include the digital twin model customizing the optimization strategy of the monitored object based on the monitoring content and simulation results to optimize the performance of the monitored object;

[0078] The digital twin model parameters in the digital twin model are divided into two types, namely simulation model parameters and optimization model parameters.

[0079] Construct loss functions for different types of digital twin model parameters for training and solving, including:

[0080] The loss function of the simulation model parameters is:

[0081] ;

[0082] in:

[0083] Represents simulation model parameters The loss function is It represents the true prediction value of the h-th category prediction result of the n-th group of training samples in the simulation model parameter training process. Represents the simulation model parameters based on , the prediction result of the nth group of training samples is the probability value of the hth category prediction result, , H represents the total number of categories of prediction results; in the embodiment of the present invention, Indicates that the prediction result of the nth group of training samples is not the hth category prediction result, Indicates that the prediction result of the nth group of training samples is the hth category prediction result;

[0084] The loss function of the optimization model parameters is:

[0085] ;

[0086] in:

[0087] Represents the optimization model parameters The loss function is It represents the true optimization value of the mth group of training samples in the dth type of optimization strategy during the training process of the optimization model parameters. Represents the optimization model parameters , the prediction result of the mth group of training samples is the probability value of the dth type of optimization strategy, , D represents the total number of categories of optimization strategies, represents the training sample balance coefficient, is set to 2; in the embodiment of the present invention, It means that the optimal optimization strategy for the mth group of training samples is not the dth type of optimization strategy. Indicates that the optimal optimization strategy for the mth group of training samples is the dth type of optimization strategy;

[0088] In the process of training the parameters of the digital twin model, public data sets can be used, such as the Prognostics Data Repository, which contains a fault prediction and health management data set for various devices, the PHM Society Data Challenge, which contains a system fault diagnosis and health prediction data set, and the traffic flow prediction data set METR-LA and PEMS-BAY. In this embodiment, the Prognostics Data Repository is used as the training set, the PHM Society Data Challenge is used as the test set, and the Prognostics Data Repository is split into training samples for simulation model parameter training and training samples for optimization model parameter training.

[0089] In addition, in order to improve the training efficiency of the digital twin model parameters, this embodiment uses the gradient descent method to iteratively solve the simulation model parameters to obtain stable simulation model parameters and simulation results, and then iteratively solve the optimization model parameters. In the iterative solution process, a dynamic learning rate method is introduced to improve the convergence speed and performance of the digital twin model parameters. The iterative formula of the simulation model parameters is:

[0090] ;

[0091] ;

[0092] in:

[0093] It represents the tth iteration result of the weight parameter of the contribution of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. , J represents the number of characteristics of the state data input into the digital twin model during the simulation of the monitoring object using the digital twin model;

[0094] express In the embodiment of the present invention, the gradient The calculation method is: the contribution weight parameter of the jth feature to the hth category prediction result is the variable pair loss function Take the derivative and Substitute the derivative result as the gradient ;

[0095] represents the learning rate of the tth iteration during the simulation model parameter training process, represents the preset initial learning rate, represents the learning rate reduction coefficient, Represents the learning rate decrease interval;

[0096] The iterative formula for optimizing model parameters is:

[0097] ;

[0098] ;

[0099] in:

[0100] It represents the tth iteration result of the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the training process of the optimization model parameters. , B represents the monitoring content input into the digital twin model and the number of features of the simulation results in the process of using the digital twin model to customize the optimization strategy for the monitoring object;

[0101] express In the embodiment of the present invention, the gradient The calculation method is: the contribution weight parameter of the b-th feature to the d-th type of optimization strategy is used as the variable to the loss function Take the derivative and Substitute the derivative result as the gradient ;

[0102] represents the learning rate of the tth iteration during the training process of optimizing model parameters, Represents the preset maximum learning rate, represents the preset minimum learning rate, represents the remainder operator, Indicates the length of an iteration cycle;

[0103] This is a preferred embodiment of the present invention. In the iterative process of the two digital twin model parameters, the present invention selects different dynamic learning rate methods. In the simulation model parameter training process, a dynamic interval descent strategy is adopted. A higher learning rate is adopted in the early stage of training, which can enable the model to quickly learn the basic pattern of the data and accelerate convergence. The learning rate is gradually reduced to reduce the oscillation caused by the high learning rate, so that the simulation model parameters can converge to the optimal solution more stably. In the training process of optimizing the model parameters, a continuous learning rate change strategy is adopted. Smoother learning rate adjustment can make the model training more stable and avoid training fluctuations caused by sudden learning rate changes. By periodically resetting the learning rate, the model can be helped to jump out of the local optimal solution and explore a wider parameter space, further improving the convergence performance. The learning rate will be set to a higher value in each iteration cycle, and then gradually reduced, so that the optimized model parameters can be globally searched in the new parameter area, which is helpful to find a better solution.

[0104] The digital twin model parameters obtained by the training and solution include simulation model parameters and optimization model parameters;

[0105] The simulation model parameters obtained by training are:

[0106] ;

[0107] in:

[0108] is the contribution weight parameter of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. , J represents the number of characteristics of the state data input into the digital twin model during the simulation of the monitoring object using the digital twin model;

[0109] The optimized model parameters obtained by training are:

[0110] ;

[0111] in:

[0112] Indicates the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the training process of the optimization model parameters;

[0113] , B represents the monitoring content input into the digital twin model and the number of features of the simulation results in the process of using the digital twin model to customize the optimization strategy for the monitoring object.

[0114] S2: Construct a multi-dimensional twin model parameter importance evaluation objective function to evaluate the digital twin model parameters.

[0115] Construct a multi-dimensional twin model parameter importance evaluation objective function for evaluating the importance of digital twin model parameters, including:

[0116] The importance evaluation dimensions of the objective function of multi-dimensional twin model parameter importance evaluation include gradient evaluation and channel gain evaluation. The importance evaluation of digital twin model parameters is performed using the gradient of the digital twin model parameters during the training process. The gradient reflects the impact of the digital twin model parameters on the loss function. The larger the gradient, the greater the contribution of the digital twin model parameters to the output result. After calculating the gradients of all digital twin model parameters, the absolute value of the gradient is used as the gradient evaluation of the digital twin model parameters.

[0117] The strength of the digital twin model parameters is calculated as the channel gain evaluation, where the higher the channel gain evaluation, the stronger the feature extraction capability of the digital twin model parameters;

[0118] The lower the importance assessment result of the digital twin model parameter, the lower the contribution of the digital twin model parameter to the output result. It is more redundant in the digital twin model, and filtering can be considered to reduce computational complexity and storage requirements.

[0119] The importance evaluation results of the digital twin model parameters are obtained by evaluating the multi-dimensional twin model parameter importance evaluation objective function, including:

[0120] The digital twin model parameters The importance assessment result is :

[0121] ;

[0122] in:

[0123] Represents the parameters of the digital twin model Gradient evaluation of ;

[0124] Represents the digital twin model parameters Channel gain evaluation;

[0125] The digital twin model parameters The importance assessment result is :

[0126] ;

[0127] in:

[0128] Represents the digital twin model parameters Gradient evaluation of ;

[0129] Represents the digital twin model parameters Channel gain evaluation.

[0130] Specifically, the L2 norm-based method can be used to calculate the strength of the digital twin model parameters as a channel gain evaluation of the digital twin model parameters. The channel gain evaluation expression is:

[0131] ;

[0132] in:

[0133] represents the L2 norm.

[0134] As a preferred embodiment of the present invention, the present invention obtains all gradients of the digital twin model parameters during the training process. Since the gradient of the digital twin model parameters in the later iteration process is small and the corresponding gradient value is not great, the gradient is weighted by a dynamic weight method to increase the gradient weight obtained in the early iteration and reduce the gradient weight obtained in the later iteration. The influence of the gradient on the gradient evaluation in the later iteration process is reduced to obtain a gradient evaluation reflecting the training gradient information of the digital twin model, wherein the digital twin model parameters The gradient of is evaluated to :

[0135] ;

[0136] ;

[0137] in:

[0138] represents the gradient weight of the t-th iteration result of the digital twin model parameters, It represents the tth iteration result of the weight parameter of the contribution of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. for The gradient of

[0139] Indicates the preset maximum number of iterations of simulation model parameters;

[0140] Digital Twin Model Parameters The gradient of is evaluated to :

[0141] ;

[0142] in:

[0143] Indicates the preset maximum number of iterations of optimization model parameters;

[0144] It represents the tth iteration result of the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the training process of the optimization model parameters. express gradient.

[0145] S3: Based on the importance assessment results, the digital twin model parameters are lightweighted, and a lightweight digital twin model is constructed using the lightweighted digital twin model parameters.

[0146] In the step S3, the digital twin model parameters are lightweighted based on the importance evaluation results, including:

[0147] Based on the importance evaluation results, the simulation model parameters and the optimization model parameters are sorted respectively, wherein the higher the importance evaluation result, the higher the ranking. The simulation model parameters and the optimization model parameters with the top G% ranking are selected to perform lightweight reconstruction on the digital twin model to obtain a lightweight digital twin model, and the simulation model parameters and the optimization model parameters that are not selected are filtered. When the digital twin model is subsequently used to simulate and optimize the monitoring object, the sample features corresponding to the filtered digital twin model parameters are no longer processed, thereby reducing the consumption of computing resources and computing costs, improving the portability of the model, improving the simulation and optimization speed of the digital twin model, reducing the number of digital twin model parameters in the digital twin model, reducing storage requirements, and the lightweight digital twin model can be run on devices with limited memory and storage capacity. In an embodiment of the present invention, G is set to 50.

[0148] S4: Fine-tune the lightweight digital twin model to obtain a lightweight tuned digital twin model.

[0149] In the S4 step, the lightweight digital twin model is fine-tuned, including:

[0150] R groups of training samples are selected from the training set, and a loss function of the lightweight digital twin model parameters in the lightweight digital twin model is constructed. The lightweight digital twin model parameters are fine-tuned in combination with the loss function, and the lightweight tuned digital twin model is constructed using the fine-tuning results. A compression algorithm is used to compress the lightweight digital twin model parameters after fine-tuning. When the digital twin model does not need to be run, the lightweight digital twin model parameters after fine-tuning are compressed and stored. When the digital twin model needs to be run, the compressed results are decompressed.

[0151] As an embodiment of the present invention, the loss function of the lightweight digital twin model parameters in the constructed lightweight digital twin model is the sum of the loss function of the lightweight simulation model parameters and the loss function of the optimization model parameters. The lightweight digital twin model parameters to be fine-tuned are the top G% simulation model parameters and the optimization model parameters selected in step S3. The gradient descent method, ADAM and other methods can be used to fine-tune the lightweight digital twin model parameters, and the change rate of the lightweight digital twin model parameters before and after fine-tuning is calculated during the fine-tuning process. If the change rate is higher than the preset threshold, the learning rate of the current iterative step is reduced, and the iterative process is performed again to retain the information of the lightweight digital twin model parameters before the lightweight processing as much as possible and reduce the number of iterations. The lightweight digital twin model parameters before and after fine-tuning are The rate of change is:

[0152] ;

[0153] in:

[0154] represents the parameters of the digital twin model after lightweight processing before fine-tuning, Digital twin model parameters representing lightweight processing The iteration result at the tth iteration step is, Indicates the iterative result during fine-tuning During the fine-tuning process, the preset maximum number of iteration steps is Max.

[0155] Specifically, Huffman coding can be used to compress and store the digital twin model parameters after fine-tuning and lightweight processing, where Huffman coding is a binary tree coding method. According to the frequency of the digital twin model parameters after fine-tuning and lightweight processing, the digital twin model parameters after fine-tuning and lightweight processing with high frequency will be assigned to the upper layer of the binary tree with shorter binary codes, and the digital twin model parameters after fine-tuning and lightweight processing with low frequency will be assigned to the lower layer of the binary tree with longer binary codes. The binary coding results of all the digital twin model parameters after fine-tuning and lightweight processing are obtained, and the binary coding results and the corresponding binary trees are compressed and packaged to compress and store the digital twin model parameters after fine-tuning and lightweight processing.

[0156] Embodiment 2:

[0157] like Figure 2As shown, it is a performance test comparison chart of the digital twin model before and after lightweight processing provided by an embodiment of the present invention. This scheme uses a test set to test the performance of the digital twin model before and after lightweight processing, wherein the selected test set is the system fault diagnosis and health prediction data set PHM Society Data Challenge. The digital twin model adopts a multi-layer neural network architecture, including an input layer, several hidden layers and an output layer. The model output includes simulation results and optimization results. The test content includes the accuracy of the simulation results and the accuracy of the optimization results of the digital twin model before and after lightweight processing, and evaluates the computing performance and storage performance of the model. The digital twin model after lightweight tuning significantly reduces the storage requirements, the number of parameters, the memory requirements and the energy consumption, and speeds up the reasoning speed to meet the real-time application requirements. Although the accuracy rate has decreased, it is within an acceptable range.

[0158] Embodiment 3:

[0159] like Figure 3 As shown, it is a diagram of the digital twin model parameter importance evaluation results provided by one embodiment of the present invention. This scheme applies the proposed model lightweighting method to the industrial manufacturing twin model to achieve model slimming. The lines between the nodes in the diagram are model parameters, and the numerical values ​​are the parameter importance evaluation results.

[0160] Embodiment 4:

[0161] like Figure 4 As shown, it is an experimental comparison diagram of the model lightweight processing method provided by an embodiment of the present invention. The accuracy experimental comparison results of the model lightweight processing method proposed in this application and the original model, random pruning method (random deletion of model parameters) and layer structure pruning method (random deletion of entire model layers) are shown. It can be seen from the figure that the model lightweight processing method proposed in this scheme is equivalent to the accuracy of the original model, and can lightweight the model while ensuring accuracy.

[0162] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0163] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0164] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0165] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A lightweight processing method for a digital twin model, characterized in that: The method comprises: S1: Obtain a digital twin model, train and solve the digital twin model, and obtain digital twin model parameters; In the iterative solution process, a dynamic learning rate method is introduced to improve the convergence speed and performance of the digital twin model parameters. The iterative formula of the simulation model parameters is: ； ； in: It represents the tth iteration result of the weight parameter of the contribution of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. , J represents the number of characteristics of the state data input into the digital twin model during the simulation of the monitoring object using the digital twin model; express The gradient of represents the learning rate of the tth iteration during the simulation model parameter training process, represents the preset initial learning rate, represents the learning rate reduction coefficient, Represents the learning rate decrease interval; The iterative formula for optimizing model parameters is: ； ； in: It represents the tth iteration result of the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the optimization model parameter training process. , B represents the monitoring content input into the digital twin model and the number of features of the simulation results in the process of using the digital twin model to customize the optimization strategy for the monitoring object; express The gradient of represents the learning rate of the tth iteration during the training process of optimizing model parameters, Represents the preset maximum learning rate, represents the preset minimum learning rate, represents the remainder operator, Indicates the length of an iteration cycle; S2: Construct a multi-dimensional twin model parameter importance evaluation objective function to evaluate the digital twin model parameters. The importance evaluation dimensions of the multi-dimensional twin model parameter importance evaluation objective function include gradient evaluation and channel gain evaluation. The gradient of is evaluated to ; ； ； in: represents the gradient weight of the t-th iteration result of the digital twin model parameters, It represents the tth iteration result of the weight parameter of the contribution of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. for The gradient of Indicates the preset maximum number of iterations of simulation model parameters; Digital Twin Model Parameters The gradient of is evaluated to : ； in: Indicates the preset maximum number of iterations of optimization model parameters; It represents the tth iteration result of the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the optimization model parameter training process. express The gradient of S3: Based on the importance assessment results, the digital twin model parameters are lightweighted, and a lightweight digital twin model is constructed using the lightweighted digital twin model parameters; S4: Fine-tune the lightweight digital twin model to obtain a lightweight tuned digital twin model.

2. A digital twin model lightweight processing method according to claim 1, characterized in that: The digital twin model is obtained in step S1, including: The digital twin model uses sensors to collect status data of monitored objects in different scenarios, and uses machine learning algorithms to map the sensor data to obtain monitoring, simulation and optimization results of the monitored objects, where the monitored objects include industrial equipment, industrial machines, vehicle transportation systems, energy systems and human health; The monitoring content of the monitored object includes the current state data and historical state data of the monitored object, and early warning of abnormal data to prevent the problem from expanding. The state data is collected by sensors; The simulation results of the monitored object include simulating the monitored object using a digital twin model to obtain the behavior and performance of the monitored object in different scenarios, and then predicting the state of the monitored object at future moments, as well as the time and location of potential faults; The optimization results of the monitored object include the digital twin model customizing the optimization strategy of the monitored object based on the monitoring content and simulation results to optimize the performance of the monitored object; The digital twin model parameters in the digital twin model are divided into two types, namely simulation model parameters and optimization model parameters.

3. A digital twin model lightweight processing method as claimed in claim 2, characterized in that: Construct loss functions for different types of digital twin model parameters for training and solving, including: The loss function of the simulation model parameters is: ； in: Represents simulation model parameters The loss function is It represents the true prediction value of the h-th category prediction result of the n-th group of training samples in the simulation model parameter training process. Represents the simulation model parameters based on , the prediction result of the nth group of training samples is the probability value of the hth category prediction result, , H represents the total number of categories of prediction results; The loss function of the optimization model parameters is: ； in: Represents the optimization model parameters The loss function is It represents the true optimization value of the mth group of training samples in the dth type of optimization strategy during the training process of the optimization model parameters. Represents the optimization model parameters based on , the prediction result of the mth group of training samples is the probability value of the dth type of optimization strategy, , D represents the total number of categories of optimization strategies, represents the training sample balance coefficient, Set to 2.

4. A digital twin model lightweight processing method as claimed in claim 3, characterized in that: The digital twin model parameters obtained through training include: The digital twin model parameters obtained through training and solution include simulation model parameters and optimization model parameters; The simulation model parameters obtained by training are: ； in: is the contribution weight parameter of the jth feature of the training sample to the hth category prediction result during the simulation model parameter training process. , J represents the number of characteristics of the state data input into the digital twin model during the simulation of the monitoring object using the digital twin model; The optimized model parameters obtained by training are: ； in: Indicates the contribution weight parameter of the bth feature in the training sample to the dth type of optimization strategy during the training process of the optimization model parameters; , B represents the monitoring content input into the digital twin model and the number of features of the simulation results in the process of using the digital twin model to customize the optimization strategy for the monitoring object.

5. The method for lightweight processing of a digital twin model according to claim 1, characterized in that: In the step S2, a multidimensional twin model parameter importance evaluation objective function is constructed, including: The importance evaluation dimensions of the objective function of multi-dimensional twin model parameter importance evaluation include gradient evaluation and channel gain evaluation. The importance evaluation of digital twin model parameters is performed using the gradient of the digital twin model parameters during the training process. The gradient reflects the impact of the digital twin model parameters on the loss function. The larger the gradient, the greater the contribution of the digital twin model parameters to the output result. After calculating the gradients of all digital twin model parameters, the absolute value of the gradient is used as the gradient evaluation of the digital twin model parameters. The strength of the digital twin model parameters is calculated as the channel gain evaluation, where the higher the channel gain evaluation, the stronger the feature extraction capability of the digital twin model parameters; The lower the importance assessment result of the digital twin model parameter, the lower the contribution of the digital twin model parameter to the output result, and the more redundant it is in the digital twin model.

6. A digital twin model lightweight processing method as claimed in claim 5, characterized in that: The importance evaluation results of the digital twin model parameters are obtained by evaluating the multi-dimensional twin model parameter importance evaluation objective function, including: The digital twin model parameters The importance assessment result is : ； in: Represents the digital twin model parameters Gradient evaluation of ; Represents the digital twin model parameters Channel gain evaluation; The digital twin model parameters The importance assessment result is : ； in: Represents the digital twin model parameters Gradient evaluation of ; Represents the digital twin model parameters Channel gain evaluation.

7. A digital twin model lightweight processing method as claimed in claim 6, characterized in that: In the step S3, the digital twin model parameters are lightweighted based on the importance evaluation results, including: Based on the importance evaluation results, the simulation model parameters and the optimization model parameters are ranked respectively. The higher the importance evaluation result, the higher the ranking. The top G% of the simulation model parameters and the optimization model parameters are selected to perform lightweight reconstruction on the digital twin model to obtain a lightweight digital twin model. The unselected simulation model parameters and the optimization model parameters are filtered out. When the digital twin model is used to simulate and optimize the monitored object in the subsequent process, the sample features corresponding to the filtered digital twin model parameters are no longer processed, which reduces the consumption of computing resources and computing costs and improves the portability of the model.

8. The method for lightweight processing of a digital twin model according to claim 1, characterized in that: In the S4 step, the lightweight digital twin model is fine-tuned, including: R groups of training samples are selected from the training set, and a loss function of the lightweight digital twin model parameters in the lightweight digital twin model is constructed. The lightweight digital twin model parameters are fine-tuned in combination with the loss function, and the lightweight tuned digital twin model is constructed using the fine-tuning results. A compression algorithm is used to compress the lightweight digital twin model parameters after fine-tuning. When the digital twin model does not need to be run, the lightweight digital twin model parameters after fine-tuning are compressed and stored. When the digital twin model needs to be run, the compressed results are decompressed.

Images